content
stringlengths 1
15.9M
|
|---|
\section{Overview of the idea}
There is strong psychological evidence that people parse visual scenes into part-whole hierarchies and model the viewpoint-invariant spatial relationship between a part and a whole as the coordinate transformation between intrinsic coordinate frames that they assign to the part and the whole~\cite{HintonCube}. If we want to make neural networks that understand images in the same way as people do, we need to figure out how neural networks can represent part-whole hierarchies. This is difficult because a real neural network cannot dynamically allocate a group of neurons to represent a node in a parse tree\footnote{What neurons do is determined by their incoming and outgoing weights and real neurons cannot completely change these weights rapidly.}. The inability of neural nets to dynamically allocate neurons was the motivation for a series of models that used ``capsules''~\cite{capsules2017,capsules2018,capsules2019}. These models made the assumption that a group of neurons called a capsule would be permanently dedicated to a part of a particular type occurring in a particular region of the image. A parse tree could then be created by activating a subset of these pre-existing, type-specific capsules and the appropriate connections between them. This paper describes a very different way of using capsules to represent the part-whole hierarchy in a neural net.
Even though this paper is primarily concerned with the perception of a single static image, GLOM is most easily understood as a pipeline for processing a sequence of frames, so a static image will be treated as a sequence of identical frames.
The GLOM architecture\footnote{The GLOM architecture has some similarity to models that use the errors in top-down predictions as their bottom-up signals~\cite{RaoBallard}, but in a nonlinear system the bottom-up signals cannot just carry the prediction error because the full activity vector is required to select the right operating regime for the non-linear units.} is composed of a large number of columns\footnote{Each level in a column bears some resemblance to a hypercolumn as described by neuroscientists.} which all use exactly the same weights. Each column is a stack of spatially local autoencoders that learn multiple levels of representation for what is happening in a small image patch. Each autoencoder transforms the embedding at one level into the embedding at an adjacent level using a multilayer bottom-up encoder and a multilayer top-down decoder. These levels correspond to the levels in a part-whole hierarchy. When shown an image of a face, for example, a single column might converge on embedding vectors\footnote{An embedding vector is the activity vector of a capsule.} representing a nostril, a nose, a face, and a person. Figure \ref{fig:newglomarchitecture} shows how the embeddings at different levels interact in a single column.
Figure \ref{fig:newglomarchitecture} does not show the interactions between embeddings at the same level in different columns. These are much simpler than the interactions within a column because they do not need to implement part-whole coordinate transforms. They are like the attention-weighted interactions between columns representing different word fragments in a multi-headed transformer~\cite{BERT}, but they are simpler because the query, key and value vectors are all identical to the embedding vector. The role of the inter-column interactions is to produce islands of identical embeddings at a level by making each embedding vector at that level regress towards other similar vectors at nearby locations. This creates multiple local "echo chambers" in which embeddings at a level attend mainly to other like-minded embeddings.
At each discrete time and in each column separately, the embedding at a level is updated to be the weighted average of four contributions:
\begin{enumerate}
\item The prediction produced by the bottom-up neural net acting on the embedding at the level below at the previous time.
\item The prediction produced by the top-down neural net acting on the embedding at the level above at the previous time.
\item The embedding vector at the previous time step.
\item The attention-weighted average of the embeddings at the same level in nearby columns at the previous time.
\end{enumerate}
For a static image, the embeddings at a level should settle down over time to produce distinct islands of nearly identical vectors. These islands should be larger at higher levels as shown in figure \ref{fig:islands}. Using the islands of similarity to represent the parse of an image avoids the need to allocate groups of neurons to represent nodes of the parse tree on the fly, or to set aside groups of neurons for all possible nodes in advance. Instead of allocating neural hardware to represent a node in a parse tree and giving the node pointers to its ancestor and descendants, GLOM allocates an appropriate activity vector to represent the node and uses the same activity vector for all the locations belonging to the node\footnote{The idea of using similarity of vectors to do segmentation has been used in earlier work on directional unit Boltzmann machines~\cite{ZemelWilliamsMozer}}. The ability to access the ancestor and descendants of the node is implemented by the bottom-up and top down neural nets rather than by using RAM to do table look-up.
Like BERT~\cite{BERT}, the whole system can be trained end-to-end to reconstruct images at the final time-step from input images which have missing regions, but the objective function also includes two regularizers that encourage islands of near identical vectors at each level. The regularizers are simply the agreement between the new embedding at a level and the bottom-up and top-down predictions. Increasing this agreement facilitates the formation of local islands.
\newpage
\begin{figure}[h!]
\centerline{\includegraphics[width=5in]{figures/newglomarchitecture.png}}
\caption{Showing the bottom-up, top-down, and same level interactions among three adjacent levels of the proposed GLOM architecture {\it for a single column}. The blue and red arrows representing bottom-up and top-down interactions are implemented by two different neural networks that have several hidden layers. These networks can differ between pairs of levels but they are shared across columns and across time-steps. The top-down net should probably use sinusoidal units\cite{sitzmann2020implicit}. For a static image, the green arrows could simply be scaled residual connections that implement temporal smoothing of the embedding at each level. For video, the green connections could be neural networks that learn temporal dynamics based on several previous states of the capsule. Interactions between the embedding vectors at the same level in different columns are implemented by a non-adaptive, attention-weighted, local smoother which is not shown.
}
\label{fig:newglomarchitecture}
\end{figure}
\newpage
\begin{figure}[h!]
\vspace*{1.5in}
\centerline{\includegraphics[width=5in]{figures/islands.png}}
\caption{A picture of the embeddings {\it at a particular time} in six nearby columns. All of the locations shown belong to the same object and the scene level has not yet settled on a shared vector. The complete embedding vector for each location is shown by dividing the vector into a separate section for each level in the part-whole hierarchy and then showing the high-dimensional embedding vector for a level as a 2-D vector. This makes it easy to illustrate alignment of the embedding vectors of different locations. The islands of identical vectors at the various levels shown in the figure represent a parse tree. But islands of identity are considerably more powerful than phrase structure grammars. They have no difficulty representing disconnected objects as in "Will this slow phrase structure grammarians down?"}
\label{fig:islands}
\end{figure}
\newpage
\section{Introduction}
This paper proposes the idea of using islands of similar vectors to represent the parse tree of an image and then explores some of the many ramifications of this idea by describing an imaginary system called GLOM that implements it. It concludes with some speculations about how the brain might implement some aspects of GLOM. But first some disclaimers:
\noindent {\bf Disclaimer1:} Human vision is a sampling process in which intelligently chosen fixation points are used to acquire the information required to perform a task using retinas that have much higher resolution around the fixation point. The same neural circuitry is reused for each new fixation. For the purposes of this paper, I assume a single retina or camera with uniform resolution and only consider what happens on the first fixation.
\noindent {\bf Disclaimer 2:} To avoid cumbersome terms like ``sub-sub-parts'', I will often talk about parts and wholes as if there were only two levels in the part-whole hierarchy. But a section of the complete embedding vector that is called a whole when considering levels L-1 and L is also called a part when considering levels L and L+1.
In a computer that has general purpose, random access memory, the obvious way to represent the part-whole hierarchy for a specific image is to create a graph structure for that particular image by dynamically allocating pieces of the memory to the nodes in the graph and giving each node pointers to the nodes it is connected to. Combining this type of dynamically created graph with neural network learning techniques has recently shown great promise \cite{PSG}, but if the whole computer is a neural network, it is far less obvious how to represent part-whole hierarchies that are different for every image if we want the structure of the neural net to be identical for all images. If we allow three-way interactions in which the activity of one neuron gates the connection between two other neurons~\cite{HintonCanonical}, it is easy to make the connections dynamic, but it is still unclear how to dynamically create a graph structure without the ability to allocate neurons on the fly. It is especially difficult in a real neural net where the knowledge is in the connection weights, which cannot easily be copied.
One rather cumbersome solution to this problem is to set aside a group of neurons, called a capsule, for each {\it possible} type of object or part in each region of the image\footnote{These regions can be larger for higher level parts which are more diverse but occur more sparsely in any one image.} and to use a routing algorithm to dynamically connect a small subset of active capsules into a graph that represents the parse of the image at hand. The activities of neurons within a capsule can then represent properties of a part such as the pose or deformation of a particular mouth or face.
With considerable effort, models that use capsules have achieved some successes in supervised and unsupervised learning on small datasets \cite{capsules2017,capsules2018,capsules2019}, but they have not scaled well to larger datasets \cite{BarhamIsard}. Capsules do not have the signature of really practical ideas like stochastic gradient descent or transformers which just {\it want} to work. The fundamental weakness of capsules is that they use a mixture to model the set of possible parts. This forces a hard decision about whether a car headlight and an eye are really different parts. If they are modeled by the same capsule, the capsule cannot predict the identity of the whole. If they are modeled by different capsules the similarity in their relationship to their whole cannot be captured.
One way to avoid using a mixture for modeling the different types of part is to have a set of identical, ``universal'' capsules, each of which contains enough knowledge to model any type of part \cite{slots,RussNitish,CanonicalCaps}. This allows part identities to have distributed representations, which allows better sharing of knowledge between similar parts. In neuroscience terminology, identities are value-coded rather than place-coded. However, it creates a symmetry breaking problem in deciding which universal object-level capsule each part should be routed to\footnote{Adam Kosoriek suggested using universal capsules in 2019, but I was put off by the symmetry breaking issue and failed to realise the importance of this approach.}.
A more radical version of universal capsules, which avoids both symmetry breaking and routing, is to pre-assign a universal capsule to every location in the image. These ubiquitous universal capsules can be used to represent whatever happens to be at that location. An even more profligate version is to dedicate several different levels of ubiquitous universal capsule to each location so that a location can belong to a scene, an object, a part and a sub-part simultaneously. This paper explores this profligate way of representing the part-whole hierarchy. It was inspired by a biological analogy, a mathematical analogy, and recent work on neural scene representations \cite{ha2016generating,sitzmann2019NEURIPS}.
\subsection{The biological analogy}
All the cells in the body have a copy of the whole genome. It seems wasteful for brain cells to contain the instructions for behaving like liver cells but it is convenient because it gives every cell its own private access to whatever DNA it might choose to express. Each cell has an expression intensity\footnote{I use the word ``intensity'' rather than the word ``level'' so as not to confuse scalar intensities with discrete levels in a part-whole hierarchy.} for each gene and the vector of expression intensities is similar for cells that form part of the same organ.
The analogy with neural nets goes like this: Each location in the image corresponds to a biological cell. The complete embedding vector for a location is like the vector of gene expression intensities for a cell. The forward pass is like the developmental process that allows a new vector of gene expression intensities to be determined by the previous vectors of expression intensities. Objects are like organs: They are a collection of locations whose embedding vectors are all very similar at a high level. Within an object, the embedding vectors may differ at lower levels that correspond to the parts of the object (see figure \ref{fig:islands}).
\subsection{The mathematical analogy}
The Kolmogorov-Arnold superposition theorem states that every multivariate continuous function can be represented as a superposition of continuous functions of one variable\footnote{This solves a version of Hilbert's 13th problem.}. For example, multiplication can be represented as the sum of the logs of the individual arguments followed by exponentiation. In machine learning terminology, when it comes to multi-argument functions, addition is all you need. This assumes, of course, that you can find the right single-argument functions to encode the arguments of the multivariate function you want to represent and then find the right function to
decode the sum. Kolmogorov proved this can always be done but the encoder functions used for the proof are so bizarre that they are of no practical relevance to neural networks.
The theorem does, however, suggest an interesting approach to combining information coming from many different locations. Perhaps we can learn how to encode the information at each location in such a way that simply averaging the encodings from different locations is the only form of interaction we need\footnote{This has a resemblance to variational learning \cite{radford}, where we start by assuming that the log posterior distribution over the whole set of latent variables is determined by the sum of their individual log posterior distributions and then we try to learn a model for which this additive approximation works well.}. This idea is already used in set transformers \cite{settransformer} for combining information from different members of a set. If we modify this suggestion slightly to use an attention-weighted local average, we get a particularly simple form of transformer in which the key, the query and the value are all the same as the embedding itself and the only interaction between locations is attention-weighted smoothing at each level. All of the adaptation occurs in the bottom-up and top-down neural networks at each location, which are depicted by the blue and red arrows in figure \ref{fig:newglomarchitecture}. These networks are shared across all locations and all time-steps, but possibly not across all levels of the part-whole hierarchy.
\subsection{Neural fields}
Suppose we want to represent the value of a scalar variable, such as the depth or intensity, at every point in an image. A simple way to do this is to quantize the image locations into pixels and use an array that specifies the scalar variable at each pixel. If the values of different pixels are related, it may be more efficient to use a neural network that takes as input a code vector representing the image and outputs an array of pixel values. This is what the decoder of an autoencoder does. Alternatively we could use a neural network that takes as input a code vector representing the image plus an additional input representing an image location and outputs the predicted value at that location. This is called a neural field \footnote{An early example of using neural fields is described in~\cite{Oore1997}. The "image" is always the same, so only the location input is needed. The 12 channels at each "image" location are the depths returned by 12 sonar detectors pointing in different directions. The match between the neural net's prediction for each location and the actual data at the robot's current location is used to perform a Bayesian update of the mobile robot's probability distribution over locations.} and this way of using neural networks has recently become very popular~\cite{ha2016generating,sitzmann2020implicit,NeRF}. Figure \ref{fig:implicit} shows a very simple example in which the intensities at a set of locations can all be reconstructed from the same code, even though the intensities vary.
\begin{figure}[h!]
\centerline{\includegraphics[width=4in]{figures/implicit.png}}
\caption{A very simple example of a neural field using individual pixels as the locations. The intensities of four pixels can all be represented by the same code $(a,b)$ even though their intensities vary according to the function $f(x) = ax +b$. The decoder has an extra input which specifies the location.}
\label{fig:implicit}
\end{figure}
For a complicated image, the neural net could transform a code vector representing the whole image plus a vector representing an image location into the value at that location. But if images are composed of familiar objects and objects are composed of familiar parts it is much more efficient to use a hierarchy of neural fields\footnote{A small step in this direction is simply to have a separate neural field for each type of object. This makes it easy to represent scenes composed of familiar objects in novel arrangements~\cite{giraffe}.}. In GLOM, the scene-level top-down neural network converts the scene vector plus an image location into the appropriate object vector for that location. This vector includes information about the 3-D pose of the object relative to the camera. All of the locations that belong to the same object are given exactly the same object-level vector. The object-level top-down neural network then converts an object vector plus a location into the part vector that is appropriate for that location and so on. For example, exactly the same top-down network acting on exactly the same face vector is able to predict the nose vector in some locations and the mouth vector in other locations.
\subsection{Explicit versus emergent representations of the part-whole hierarchy}
In the symbolic AI representation of the part-whole hierarchy, each node has a unique symbol or memory address, and this symbol or address has an arbitrary relationship to the content. In GLOM, the embedding vector at a particular level that is shared by all the locations in an island at that level plays the same role as the address of a node in a graph, but this vector is not arbitrary. The top-down neural network must predict the embedding vector of a part at level L from the embedding vector of an object at level L+1. This involves using the pose relative to the camera encoded at level L+1 and a representation of the image location to compute where the location is within the intrinsic coordinate frame of the object. This determines which level L part the location belongs to.
There is a very important difference between using an address bus to follow a pointer to the representation of a part of a known object and using a top-down neural network to compute the part vector. Unlike table-lookup, the top-down neural net finds it much easier to deal with shapes in which there are symmetries between the parts. Replicated parts, like the legs of a centipede for example, add very little complexity to the neural net and this remains true even if the legs change along the centipede, so long as they change in a predictable way. Bilateral symmetries that align with the intrinsic coordinate frame of an object reduce the required capacity of the top-down neural net by almost a factor of two.
It is much harder, however, for the neural net to make use of symmetries that do not align with the coordinate frame\footnote{This is why Canonical Capsules~\cite{CanonicalCaps} discover the natural intrinsic coordinate frames.}, and people are generally unaware of such symmetries. Most people, for example, are totally unaware of the threefold rotational symmetry of a cube, despite its name, until they are forced to use a body diagonal through the center of the cube as its intrinsic top-bottom axis~\cite{HintonCube}. They then cease to be aware of any of the right angles in the cube because these angles no longer align with the new intrinsic coordinate frame\footnote{Most people have enormous difficulty imagining a cube when they are forced to use a body diagonal as the top-bottom axis. When asked to point out the corners that are not at the two ends of this axis, they typically point out four corners arranged in a square about halfway up the axis. This structure (two square-based pyramids stuck together base-to-base is actually an octahedron. An octahedron is the dual of a cube with vertices for faces and faces for vertices. So people preserve the fourfold rotational symmetry of a cube relative to its familiar coordinate system. This suggests that the symmetry structure is one of the most important properties encoded in the embedding vector of an object.}.
\section{Some design decisions}
This section discusses some decisions that need to be made when specifying the GLOM architecture.
\subsection{How many levels are there?}
GLOM assumes that the part-whole hierarchy has a fixed depth. People can deal with very deep hierarchies that have stars near the top and atomic nuclei near the bottom. The way to handle such an enormous range is to have a flexible mapping between entities in the world and the levels of GLOM~\cite{HintonPartWhole}. This allows the very same neurons to be used for stars at one time and for atomic nuclei at another, which has the added advantage of facilitating analogies between structures at very different scales like solar systems and atoms. The recursive re-use of the neural hardware raises many interesting issues about temporary storage and control flow~\cite{BaHistory} that will only be fleetingly addressed here.
A reasonable number of embedding levels would be about five. This allows for the pupil and the white of an eye to be the lowest-level sub-sub-parts in a scene composed of objects that are people with parts that are faces and sub-parts that are eyes. If finer details such as the dilation of the iris are required, people probably need to remap the world onto their hardware so that, for example, the face becomes the scene\footnote{The levels in the part-whole hierarchy that are represented in the infero-temporal pathway are probably not the brain's only representation of space. The infero-temporal pathway is used for object recognition and there may well be other representations of the world that are used for other purposes such as detecting ego-motion or visually maintaining balance.}.
One interesting question is whether the bottom-up and top-down neural nets can be shared across levels as well as across locations. This would not work for the lower levels of language processing where entities at different levels, like phonemes or words, have very different properties, but vision is far more fractal. A big advantage of sharing across levels is that the vector representation used for a particular face when it was at the object level would then be forced to be consistent with its representation when it was at the part level. This would make it much easier to remap the visual world onto the hardware by simply copying all of the vectors up or down a few levels. After having used fine details of an eye to extract a highly informed vector representation of the eye when it was at the object level, this same vector could then be used to represent the eye when it was at the sub-part level\footnote{This assumes that the vector can be transported to a different column if the fixation point changes when the face becomes the object of attention rather than the eye.}.
\subsection{How fine-grained are the locations?}
Locations could be as fine-grained as pixels, or they could correspond to larger image patches~\cite{PatchTransformer}. To avoid additional complexity when explaining the basic idea of the paper, I will assume that the grid of locations remains the same at all levels, but this is probably not the best choice.
The granularity could change at different embedding levels. If higher levels in the part-whole hierarchy use a larger stride, the top-down neural net would need to output multiple different predictions for the multiple lower level locations that fall within one higher level location. Similarly, the bottom-up neural net would need to look at all the lower-level locations that get combined at the next level up.
One convenient way to be sensitive to a large spatial context whilst also being able to see fine detail is to have images at several different spatial resolutions all of which have the same number of pixels. The coarsest image conveys a large spatial context but lacks fine detail and the finest image conveys the fine details, but only for a small region. If the visual input is structured into multiple images in this way, it would make sense to make peripheral locations cover larger regions, but this paper will ignore that issue because it makes everything more complicated.
\subsection{Does the bottom-up net look at nearby locations?}
Even if the granularity at different levels remains unchanged, the bottom-up neural net could look at the embedding vectors at nearby locations. This is a less pure version of GLOM which allows the interactions between locations to be more complex than just averaging. The purely bottom-up pathway then resembles a convolutional neural net but with the predictions for the next level up being made by a multi-layer neural net that implements a far more complicated function than just a matrix multiply followed by a scalar non-linearity.
The disadvantage of allowing the bottom-up net to look at other locations is that two locations with identical representations at the part level may have different spatial contexts. We would then lose a very nice property of the pure version of GLOM: locations that have identical representations at the part level make exactly the same bottom-up predictions at the object level.
By looking at other locations, the bottom-up net can reduce the uncertainty before it predicts a distribution at the next level up and this seems like a good thing to do. But it should be possible to get a similar reduction in uncertainty {\it after} making the prediction when the attention-weighted smoothing combines an uncertain bottom-up prediction from one location with the uncertain bottom-up predictions from nearby locations. Of course, this assumes that the bottom-up net can represent the uncertainty in its predictions and that the uncertainties in different locations can be combined correctly by the attention-weighted smoothing. This issue is addressed in section \ref{sec:uncertainty}.
\subsection{How does the attention work?}
One of the contributors to the update of the embedding of level $L$ at location $x$ is the attention-weighted average of the embeddings of level $L$ at nearby locations at the previous time step. GLOM assumes the simplest form of attention weighting in which the weight $w_{xy}$ that location $x$ gives to the embedding at location $y$ is given by
\begin{equation}
w_{xy} = \frac{e^{\beta L_x.L_y}}{\sum_z e^{\beta L_x.L_z}}
\end{equation}
where $.$ is the scalar product of the two embedding vectors, $z$ indexes all the locations that location $x$ attends to at level $L$ and $\beta$ is an "inverse temperature" parameter that determines the sharpness of the attention. $\beta$ could increase as GLOM settles to a firm interpretation of the image. The way attention is intended to work in GLOM has already been used successfully in a system called "ACNe"~\cite{ACNe}.
Pioneering work on using Markov Random Fields for image segmentation \cite{GemanGeman} used the presence of a boundary between pixel $x$ and pixel $y$ to prevent $x$ from attending to $y$. A boundary is more than just a big intensity difference between $x$ and $y$ because its existence depends on the intensities at other locations.
Similarly, early work on learning spatially coherent properties of images used the presence of boundaries to select which expert interpolator to use~\cite{beckerMOE}. Like the seashore, boundaries have a rich life of their own and much more work needs to be done to integrate them into GLOM, especially into its attention mechanism.
\subsection{The visual input}
In most neural networks, the visual input arrives at the bottom layer. In GLOM, a patch of the raw visual input could define the bottom-level embedding at a location by vectorizing the intensities in the image patch, but it is probably more sensible to first apply a convolutional neural net that can see a larger region of the image. The output of this convolutional net would then be the primary, lowest level embedding at each location.
The convolutional net is an open loop way to solve the following inference problem: What lowest-level embedding for that location would be able to reconstruct the pixel intensities using the learned neural field shared by all locations. Once the lowest-level embedding has been initialized, it can be refined in a closed loop by backpropagating the reconstruction error through the neural field~\cite{ckiwold}.
There is no need to confine the direct visual input to the primary embedding layer. A coarser scale convolutional net operating on a lower resolution image could provide useful hints about the higher-level embeddings. For example, a pale vertical oval with a darker horizontal band slightly more than halfway up suggests one kind of face~\cite{ViolaJones} so a convolutional net operating on coarse pixels can provide useful information to directly initialize the higher-level embeddings\footnote{The visual part of the thalamus has direct connections to multiple different levels in the hierarchy of visual areas.}.
\section{Color and texture}
Consider an object whose individual parts are either entirely pale green or entirely mauve. The color of a part is straightforward, but what color is the whole object? One of the motivations for GLOM was the idea that the whole object has a compound color which might be called "pale-green-or-mauve" and {\it at the object level} every location belonging to the object has exactly the same, compound color. The object is pale-green-and-mauve all over. When deciding which other locations at the object level to attend to, preference would be given to locations with a similar compound color.
A similar idea applies to textures. The individual texture elements have their own shapes and poses and spatial relationships, but an object with a textured surface has exactly the same texture everywhere {\it at the object level}. GLOM extends this idea to shapes. An object may have parts that are very different from one another, but at the object level it has exactly the same compound shape in all of the locations that it occupies.
\section{Cluster discovery versus cluster formation}
The EM capsule model~\cite{capsules2018} attempts to activate capsules that represent wholes (e.g. a face) by looking for clusters of similar vector votes for the pose of the whole. These vector votes come from already identified parts (e.g. a nose or mouth) and although the weights on these votes can be modified by an iterative routing procedure the vector votes themselves remain fixed. This is quite problematic if one of the parts has an under-determined pose. For example, a circle representing an eye has no specific orientation and its position in a face depends on whether it is a left or right eye. It does, however, provide some information about the scale of the face and it makes a unimodal prediction for the location of the face in the direction orthogonal to the unknown line between the two eyes\footnote{ The Stacked Capsule Autoencoder model~\cite{capsules2019} deals with this issue by using a set transformer~\cite{settransformer} to allow the parts to interact. This should allow the poses and identities of the parts to be disambiguated before they attempt to activate capsules at the next level up.}.
In GLOM, the embedding vector of a location at level L-1 does not cast an immutable vector vote for the embedding at level L. Instead, it provides a bottom-up vector contribution to this embedding that is combined with the vector contribution coming from level L+1 and the attention-weighted contributions coming from the level L embeddings of other locations to determine the updated level L embedding vector. The bottom-up contribution can start off being quite vague and it can become sharper from time-step to time-step as top-down and lateral contextual information progressively refines the level L-1 embedding of the location. The islands of similar embedding vectors at a level can be viewed as clusters, but these clusters are not {\it discovered} in immutable data. They are {\it formed} by the interaction between an intra-level process that favors islands of similarity and dynamically changing suggestions coming from the location's embeddings at adjacent levels.
\section{Replicating embedding vectors over locations}
At first sight, it seems very inefficient to give a copy of the object-level embedding vector to every location that belongs to an object. Compelling intuitions that stem from programming computers with random access memory suggest that it would be much better to have a single copy of a data-structure for the object. These intuitions are probably misleading for neural nets that do not have RAM, and even if RAM is available there are two good reasons for replicating the embedding vectors over an island.
The island growing process at each level may eventually settle down to several islands of near identical vectors, but the search for these islands needs to be able to consider alternative clusterings of locations into islands and it also needs to allow for negotiations between locations within an island about what identical vector to settle on at each level. These negotiations are non-trivial because each location is also trying to satisfy inter-level constraints that come from its own embedding vectors at the level above and the level below and these embeddings are also being refined at every time-step. During the search, it is very helpful for every location to have its own version of the embedding vector at each level. Uncertainty in the clustering can be represented by making the embedding vector at a location be a blend of the vectors for the different clusters that it might decide to join. This blend can be refined over time and the fact that it lives in a high-dimensional continuous space should make optimization easier.
Intuitively, a blend of two rather different embedding vectors is not similar to either vector. This is true in a low-dimensional vector space, but intuitions derived from low-dimensional spaces cannot be trusted when dealing with high-dimensional spaces. The average of two high-dimensional vectors is much closer to each of those vectors than it is to a random vector. This can be understood by thinking about the correlation between the components of a vector and the components of its average with some other random vector. If the vectors are high-dimensional, this correlation will be very significant~\footnote{ This explains why the first stage of a language model can convert a word like ``bank'' into a single high-dimensional embedding vector rather than needing separate vectors for the ``river'' and the ``money'' senses.}.
A further advantage of islands of near identity is that it allows long range interactions within a level to be sparse. If there is more sparsity at higher levels, the interactions can be longer range without increasing the amount of computation. For locations that belong to an island far away, all the object-level information about that island is contained in each of its locations, so it is only necessary to sample one of those locations for that distant island to compete with other closer islands for a location's attention. Of course, this means that distant islands contribute fewer logits to the attention softmax than closer islands, but the exponential used in the attentional softmax means that one logit from a highly relevant distant island can out-compete multiple logits from a closer but much less relevant island.
A simple way to choose which other locations are allowed to compete for the attention of location {\bf x} is to sample, without replacement, from a Gaussian centered at {\bf x}. Higher level embeddings can sample the same number of other locations but from a larger Gaussian. The sampling could be done only once so it was part of the architecture of the net. Alternatively, lacunae in the sampling could be greatly reduced by sampling independently at each time step.
\section{Learning Islands}
Let us assume that GLOM is trained to reconstruct at its output the uncorrupted version of an image from which some regions have been removed. This objective should ensure that information about the input is preserved during the forward pass and if the regions are sufficiently large, it should also ensure that identifying familiar objects will be helpful for filling in the missing regions. To encourage islands of near identity, we need to add a regularizer and experience shows that a regularizer that simply encourages similarity between the embeddings of nearby locations can cause the representations to collapse: All the embedding vectors may become very small so that they are all very similar and the reconstruction will then use very large weights to deal with the very small scale. To prevent collapse, contrastive learning~\cite{BeckerHinton92,Paccanaro2001,Oord2018} uses negative examples and tries to make representations that should agree be close while maintaining separation between representations which should not agree\footnote{Maintaining separation is quite different from asking representations that should be separate to be far apart. Once two representations are sufficiently different there is no further pressure to push them even further apart.}.
Contrastive learning has been applied very successfully to learn representations of image crops~\cite{ting1, Bachman,moco,ting2,tejankar2020isd} It learns to make the representations of two different crops of the same image agree and the representations of two crops from different images disagree. But this is not a sensible thing to do if our aim is to recognize objects. If crop 1 contains objects A and B and crop 2 from the same image contains objects B and C, it does not make sense to demand that the representations of the two crops be the same {\it at the object level}. It does make sense at the scene level, however. For scenes containing one prominent object, it may be possible to recognize objects using representations that are designed to recognize scenes, but as soon as we distinguish different levels of embedding it becomes clear that it would be better to use a contrastive loss function that encourages very similar representations for two locations at level L only if they belong to the same entity at level L. If they belong to different level L entities their level L embeddings should be significantly different.
From the point of view of a location, at all but the top level it needs to decide which other locations its level L embedding should be similar to. It can then learn to resemble those embeddings and be repelled from the embeddings of locations that belong to different objects in the same or other images. Recent work that uses the similarity of patches along a possible object trajectory to influence whether contrastive learning should try to make them more similar has shown very impressive performance at finding correspondences between patches in video sequences~\cite{JabriOwensEfros}.
The obvious solution is to regularize the bottom-up and top-down neural networks by encouraging each of them to predict the consensus opinion. This is the weighted geometric mean of the predictions coming from the top-down and bottom-up networks, the attention-weighted average of the embeddings at nearby locations at the previous time-step the previous state of the embedding. Training the inter-level predictions to agree with the consensus will clearly make the islands found during feed-forward inference be more coherent.
An important question is whether this type of training will necessarily cause collapse if it is not accompanied by training the inter-level predictions to be different for negative examples that use the consensus opinions for unrelated spatial contexts. Using layer or batch normalization should reduce the tendency to collapse but a more important consideration may be the achievability of the goal.
When the positive examples in contrastive learning are used to try to extract very similar representations for different patches of the same image, the goal is generally not achievable and the large residual errors will always be trying to make the representations collapse. If, however, an embedding at one location is free to choose which embeddings at other locations it should resemble, the goal can be achieved almost perfectly by learning to form islands of identical vectors and attending almost entirely to other locations that are in the same island. This should greatly reduce the tendency towards collapse and when combined with the deep denoising autoencoder objective function and other recent tricks~\cite{grill2020bootstrap,ChenHe2020} it may eliminate the need for negative examples.
\section{Representing coordinate transformations}
When neural networks are used to represent shape, they generally work much better if they represent the details of a shape relative to its intrinsic coordinate frame rather than relative to a frame based on the camera or the world \cite{mocap,NASA}.
Work on the use of neural fields for generating images has established that there are much better ways to represent the location than using two scalars for its x and y coordinates~\cite{sitzmann2020implicit,NeRF}. The product of a delta function at the location with both horizontal and vertical sine and cosine waves of various frequencies works well. A similar representation is used in transformers for the position of a word fragment in a sentence.
The success of highly redundant representations of location suggests that there may also be highly redundant representations of the non-translational degrees of freedom of a coordinate transform that work much better in a neural net than the matrices or quaternions commonly used in computer graphics\footnote{The
standard matrix representation uses the scale of the matrix to represent the change in scale caused by the coordinate transform. Using the scale of the weights to represent scale in this analog way is a particularly bad idea for neural nets.}.
Let us suppose that we would like the pose of a part ({\it i.e.} the coordinate transform between the retina and the intrinsic frame of reference of a part) to be represented by a vector that is a subsection of the embedding vector representing the part. A multi-layer neural network whose weights capture the viewpoint-invariant coordinate transform between a part and a whole can then operate on the pose vector of the part to predict the pose vector of the whole. If we simply flatten the 4x4 matrix representation of a pose into a vector, it is easy to hand-design a multi-layer neural net that takes this vector as input and produces as output a vector that corresponds to the flattened result of a matrix-matrix multiply, {\it provided we know what matrix to multiply by, which depends on the identity of the part}. This dependence on the part identity was the reason for allocating a separate capsule to each type of part in earlier capsule models. Unfortunately, the vector space of flattened 4x4 matrices does not make it easy to represent uncertainty about some aspects of the pose and certainty about others. This may require a much higher-dimensional representation of pose. Designing this representation by hand is probably inferior to using end-to-end learning with stochastic gradient descent. Nevertheless, section \ref{sec:uncertainty} discusses one approach to representing uncertainty in a neural net, just to demonstrate that it is not a major problem.
In a universal capsule the part-identity is represented by an activity vector rather than by the choice of which capsule to activate, so the neural net that implements the appropriate part-whole coordinate transform needs to condition its weights on the part-identity vector~\footnote{In
stacked capsule autoencoders~\cite{capsules2019} the capsule identity determines the default object-part coordinate transform, but the transform can be modulated by a vector that represents the deformation of the object.}.
Consequently, the entire part-level vector of a location needs to be provided as input to the bottom-up neural net that computes the part-whole coordinate transform. This makes the computation much more complicated but it greatly simplifies the design of the architecture. It means that we do not need to designate one part of the embedding vector at a level to represent the pose and the rest to represent other aspects of the entity at that level. All we need to do is to make sure that the neural net that predicts the embedding at one level from the embedding below (or above) has sufficient expressive power to apply a coordinate transform to those components of the embedding vector that represent pose and to make this coordinate transform be contingent on those components of the vector that represent the identity of the part. Since this neural net is going to be learned by stochastic gradient descent, we do not even need to keep components of the embedding vector that represent the pose separate from the components that represent other properties of the entity at that level: individual components can be tuned to combinations of pose, identity, deformation, texture {\it etc}.
Entangling the representations of identity and pose may seem like a bad idea, but how else can a bottom-up prediction from a diagonal line express the opinion that the whole is either a tilted square or an upright diamond? To express this distribution using activities of basis functions, we need basis functions that are tuned to combinations of identity and pose.
Using a small matrix or quaternion to represent pose makes it easy to model the effects of viewpoint changes using {\it linear} operations. At first sight, abandoning these explicit representations of pose seems likely to compromise the ability of a capsule to generalize across viewpoints. This would be true if each capsule only dealt with one type of object, but universal capsules will have seen many different types of object from many different viewpoints and any new type of object will be well approximated by a weighted average of familiar types all of which have learned to model the effects of viewpoint. Moreover, the weights in this average will be the same for all viewpoints. So if a novel object is only seen from a single viewpoint, a universal capsule may well be able to recognize it from radically different viewpoints.
The same argument about generalization can be applied to CNNs, but there is a subtle difference: GLOM is forced to model the coordinate transforms between parts and wholes correctly in order to be able to make use of the spatial relationship between one part and another by using a simple averaging operation at the level of the whole. It is the viewpoint invariance of these part-whole spatial relationships that makes it possible to generalize to radically new viewpoints.
\section{Representing uncertainty}
\label{sec:uncertainty}
It is tempting to imagine that the individual components of an embedding vector correspond to meaningful variables such as the six degrees of freedom of the pose of an object relative to the camera or the class of an object. This would make it easy to understand the representation, but there is a good reason for making the relationship between physically meaningful variables and neural activities a little less direct: To combine multiple sources of information correctly it is essential to take the uncertainty of each source into account.
Suppose we want to represent M-dimensional entities in such a way that different sources of information can contribute {\it probability distributions} over the M-dimensional space rather than just point estimates. We could use a population of $N \gg M$ neurons each of which is tuned to a Gaussian in the $M$-dimensional space~\cite{WilliamsAgakov}. If we take logs, a neuron then corresponds to a parabolic bump in the log probability. This bump could be very wide in some directions and very narrow in others. It could even be a horizontal ridge that is infinitely wide in some of the directions. We treat the activity of a neuron as a vertical scaling of its parabolic bump and simply add up all the scaled bumps to get a parabolic bump which represents the log of the unnormalized Gaussian distribution represented by the population of N neurons.
Sources of information can now contribute probability distributions which will be multiplied together by simply contributing additively to the activities of the N neurons. If we want to keep $N$ relatively small, there will be limitations on the probability distributions that can be represented, but, given a budget of $N$ neurons, learning should be able to make good use of them to approximate the predictive distributions that are justified by the data. If, for example, it is possible for a part to predict the horizontal location of a whole without making this prediction be contingent on other aspects of the pose or identity of the whole, it would be helpful to tune a handful of the $N$ neurons to well-spaced values on the dimension representing the horizontal location of the whole in the underlying $M$-dimensional space. The part can then contribute a Gaussian distribution along this horizontal dimension by making appropriate contributions to this handful of neurons. The relative magnitudes of the contributions determine the mean of the Gaussian and their overall scale determines the inverse variance of the Gaussian.
The assumption that the neurons have Gaussian tuning in the underlying $M$-dimensional space of possible entities was just a simplification to show that neural networks have no problem in representing Gaussian probability distributions and combining them appropriately. A much more flexible way to tune the neurons would be to use a mixture of a Gaussian and a uniform~\cite{ContrastiveDivergence}. The log of this distribution is a {\it localized} bump which will be called a unibump. The sides of a unibump splay out and eventually become horizontal when we are far enough from the mean that the uniform completely dominates the Gaussian. Unlike a parabolic bump which has a quadratically large gradient far from its maximum, a unibump has zero gradient far from its maximum so it makes no contribution to the shape of the unnormalized distribution far from its mean. This allows unibumps to represent multi-modal probability distributions. The sum of one set of nearby unibumps can represent one mode and the sum of another set of unibumps that are close to one another but far from the first set can represent another mode. Using neural activities that correspond to vertical scalings of the unibumps, it is possible to control both the location and the sharpness of each mode separately.
The assumption that individual neurons are tuned to a mixture of a Gaussian and a uniform was just a simplification to show that neural networks can represent multi-modal distributions. The basis functions that neurons actually learn for representing multi-modal log probability distributions in an underlying latent space do not need to be local in that space.
The need to represent uncertainty prevents the simplest kind of representation in which activity in a single neuron represents one dimension of an M-dimensional entity, but it still allows neurons to have tuning curves in the M-dimensional space. Whether it is possible for someone trying to understand the representations to jointly infer both the underlying latent space and the tuning curves of the neurons in that space is a very interesting open problem. But even when it is hard to figure out what the individual neurons are doing it should still be trivial to find islands of nearly identical vectors, so it should be easy to see how GLOM is parsing an image or how a similar model applied to language is parsing a sentence.
When considering how to represent uncertainty about the pose or identity of a part, it is very important to realize that each location assumes that it is only occupied by at most one part at each level of the hierarchy\footnote{We assume the visual world is opaque. Transparency, like reflections in a window of an indoor scene superimposed on an outdoor scene would need to be handled by switching attention between the two different scenes.}. This means that all the neural activities in the embedding vector at a level refer to the same part: there is no binding problem because the binding is done via the location. So a location can use two different neurons whose tuning curves in the underlying M-dimensional space overlap by a lot without causing any confusion. If we do not start by allocating different subsets of the neurons to different locations, the broad tuning curves in the M-dimensional underlying space that are needed for representing uncertainty will cause confusion between the properties of different objects. That is why coarse coding, which uses a single population of broadly tuned neurons to model several different entities at the same time~\cite{HintonCoarse} cannot model uncertainty efficiently.
\subsection{Combining different sources of information when updating the embeddings}
The embedding at each level is updated using information from the previous time-step at adjacent levels and also at other locations on the same level. These sources are far from independent, especially when the image is static so that the visual input is identical at multiple time-steps. The higher-level embeddings obviously depend on the earlier lower-level embeddings. Also, the same-level embeddings that contribute to the attention-weighted local average will have been influenced by early states of the very embedding that the attention-weighted average is trying to update. To avoid becoming over-confident it is better to treat the different sources of information as {\it alternative} paths for computing the embedding vector from the visual input. This justifies taking a weighted geometric mean of the distributions\footnote{When taking the geometric mean of some distributions we assume that the product of the distributions is renormalized to have a probability mass of 1.} predicted by the individual sources rather than a simple product of these distributions which would be appropriate if they were independent.
For interpreting a static image with no temporal context, the weights used for this weighted geometric mean need to change during the iterations that occur after a new fixation. Initially the bottom-up source should be by far the most reliable, but later on, the top-down and lateral sources will improve. Experiments with deep belief nets\cite{HintonNSERC2006} show that gradually increasing the weighting of top-down relative to bottom-up leads to more plausible reconstructions at later times, suggesting that this will be important when GLOM is trained as an end-to-end deep denoising autoencoder.
\section{Comparisons with other neural net models}
This section compares GLOM to some of the neural net models that influenced its design.
\subsection{Comparison with capsule models}
The main advantage of GLOM, compared with capsule models~\footnote{Some capsule models already use universal capsules in which vectors of activity rather than groups of neurons are used to represent the part identity, but they do not replicate these vectors across all locations within the object~\cite{RussNitish}.}, is that it avoids the need to pre-allocate neurons to a discrete set of possible parts at each level. The identity of a part becomes a vector in a continuous space of feature activities. This allows for much more sharing of knowledge between similar parts, like arms and legs, and much more flexibility in the number and type of parts belonging to an object of a particular type.
A second advantage of GLOM is that it does not require dynamic routing. Instead of routing information from a part capsule to a specific capsule that contains knowledge about the relevant type of whole, every location that the part occupies {\it constructs} its own vector representation of the whole. The constraint that a part at one location only belongs to one whole is a necessary consequence of the the fact that the alternative wholes at that location are alternative activity vectors on the same set of neurons. Uncertainty about which of several wholes is the correct parent of a part can still be captured by using blends of activity vectors.
A third advantage of GLOM is that the cluster formation procedure for forming islands is much better than the clustering procedure used in capsule models. To make methods like EM work well when the number of clusters is unknown, it is helpful to introduce split and merge operations~\cite{UedaSMEM} but these operations happen automatically during island formation. Hierarchical Bayesian concerns about finding the correct number of clusters at an embedding level are addressed by starting with one island per location and then reducing the number of distinctly different islands by making embedding vectors agree. This reduction occurs in a continuous space with no need for discrete changes in the number of clusters.
The main disadvantage of GLOM, compared to most capsule models, is that knowledge about the shape of a specific type of object is not localized to a small group of neurons (possibly replicated across quite large regions). Instead, the bottom-up and top-down neural nets (which may be different for every pair of adjacent levels) have to be replicated at every single location. For computer implementations the replication across locations is a big advantage because it allows a weight to be used many times each time it is retrieved from memory, but for biological neural nets it seems very wasteful of synapses. The point of the analogy with genes is that biology can afford to be wasteful so this objection may not be as serious as it seems. There is, however, a more serious issue for a biological version of GLOM: The ubiquitous universal capsules would need to learn the very same knowledge separately at every different location and this is {\it statistically} very inefficient. Fortunately, section \ref{sec:plausible} shows how locations can share what their bottom-up and top-down models have learned without sharing any of their weights.
By allocating neurons to locations rather than to types of object or part, GLOM eliminates a major weakness of capsule models, but it preserves most of the good aspects of those models:
\begin{itemize}
\item {\bf Handling the effects of viewpoint properly:} The weights of the bottom-up and top-down neural networks capture the viewpoint-invariant spatial relationships between parts and wholes and the neural activities capture the viewpoint equivariant information about the pose of an object or part.
\item {\bf Coincidence filtering:} Objects are recognized by using agreement between high-dimensional predictions from their parts. In GLOM, the idea of using agreement is taken even further because it is also used to {\it represent} objects and parts as islands of identity.
\item {\bf No dynamic allocation of neurons:} The part-whole hierarchy can be represented without dynamically allocating neurons to nodes in the parse tree.
\end{itemize}
\subsection{Comparison with transformer models}
The GLOM architecture shown in figure \ref{fig:newglomarchitecture} can be rearranged by viewing the vertical time-slices in figure \ref{fig:newglomarchitecture} as layers in figure \ref{fig:oldglomarchitecture}. This rearrangement of GLOM is then equivalent to a standard version of a transformer~\cite{transformer} but with the following changes:
\begin{itemize}
\item The weights are the same at every layer because GLOM is a recurrent net and we have converted the time slices into layers.
\item The attention mechanism is greatly simplified by using the embedding vector at a level as the query, the key and also the value. The complex interactions between different locations that are normally implemented by attention are thus reduced to a simple, attention-weighted, smoothing operation.
\item The multiple heads used to provide more expressive power in most transformers are re-purposed to implement the multiple levels of a part-whole hierarchy and the interactions between the heads at a location are highly structured so that a level only interacts with the adjacent levels.
\item The bottom-up and top-down neural networks that compute the interactions between adjacent levels perform coordinate transformations between the distributed representations of the poses of parts and wholes and these coordinate transformations depend on the distributed representations of the types of the part and the whole.
\end{itemize}
The justification for eliminating the distinction between the query, the key, the value and the embedding itself is as follows: Consider trying to get a potential mouth to be corroborated by a potential nose in a transformer. The mouth needs to ask "is there anyone in the right spatial relationship to me who could be a nose". If so, please tell me to be more mouth-like. This seems to require the mouth to send out a nose query (that includes the appropriate pose relative to the mouth) that will match the key of the nose. The nose then needs to send back a mouth-like value (that includes the appropriate pose relative to the nose).
But the mouth could also be corroborated by an eye so it needs to send out a different query that will match the key of an eye. This could be handled by using separate heads for a mouth-looking-for-a-nose and a mouth-looking-for-an-eye (as in categorial grammar), but that seems clumsy.
A more elegant solution (inherited from capsule models) is to use a form of the Hough transform. The potential mouth predicts a vector for the face it might be part of. The potential nose and eye do the same. All you need now is agreement of the predictions at the face level so query$=$key$=$value$=$embedding. The face level can then give top-down support to its parts instead of the support coming from a value vector sent by one part to another using a coordinate transform specific to the identities of the two parts.
\newpage
\begin{figure}[h!]
\centerline{\includegraphics[width=5in]{figures/oldglomarchitecture.png}}
\caption{This is a different way of visualizing the architecture shown in figure \ref{fig:newglomarchitecture} which makes the relationship of that architecture to transformers more obvious. The horizontal dimension which represents time in figure \ref{fig:newglomarchitecture} becomes the vertical dimension which represents layers in this figure. At each location, every layer now has embeddings for all of the levels in the part-whole hierarchy. This corresponds to vertically compressing the depiction of the levels within a single time-slice in figure \ref{fig:newglomarchitecture}. A single forward pass through this architecture is all that is required to interpret a static image. All of the level-specific bottom-up and top-down neural nets are shown here as a single neural net. Figure \ref{fig:updown} shows the individual bottom up and top-down neural nets for this alternative way of viewing the GLOM architecture.
}
\label{fig:oldglomarchitecture}
\end{figure}
\newpage
\begin{figure}[h!]
\vspace*{0.5in}
\centerline{\includegraphics[width=5in]{figures/updown.png}}
\caption{A picture of two adjacent layers of GLOM for a single location ({\it i. e.} part of a single column). During the forward pass, the embedding vector at level L receives input from the level L-1 embedding vector in the previous layer via a multi-layer bottom-up neural net. It also receives input from the level L+1 embedding in the previous layer via a multi-layer top-down neural net. The dependence on level L+1 in the previous layer implements top-down effects during the forward pass. The level L embedding in layer $t+1$ also depends on the level L embedding in layer $t$ and an attention-weighted sum of the level L embeddings at other nearby locations in layer $t$. These within-level interactions are not shown.}
\label{fig:updown}
\end{figure}
\newpage
\subsection{Comparison with convolutional neural networks}
Capsules were originally motivated by three perceived deficiencies of CNNs:
\begin{enumerate}
\item The pooling operation in a CNN was designed to achieve local {\it invariance} to translation in the activity vector at the next level up. It seems better to ask for invariance in the weights but equivariance in the activities.
\item CNNs attempt to generalize across viewpoints by using a lot of examples of different viewpoints which may be produced by augmenting the dataset with transformed images. Computer graphics generalizes across viewpoints by having explicit representations of the poses of objects or parts relative to the camera. A change in viewpoint, even a very big one, can be modeled perfectly by {\it linear} operations on these explicit poses. Using the viewpoint invariant relationship between the pose of a part and the pose of the whole seems like a very efficient way to generalize to radically new viewpoints. CNNs do not appear to be doing this, but appearances can be deceptive.
\item In CNNs, the activity of a neuron is determined by the scalar product of a weight vector with an activity vector. This is not a good way to model covariance structure which is very important in vision. Taking the scalar product of an activity vector with another activity vector makes powerful operations like coincidence detection and attention much easier. Coincidences in a high-dimensional embedding space are a good way to filter out noise caused by occlusion or missing parts because, unlike sums, they are very robust to the absence of some of the coinciding predictions.
\end{enumerate}
The {\bf first} deficiency is only apparent. It depends on a common misunderstanding of how CNNs represent the positions of parts. The vector of channel activities at a gridpoint can have a rate-coded representation of the position of a part that is far more accurate than the stride between gridpoints. So when the stride is increased at the next level by pooling it does {\it not} mean that the position of a part is encoded less accurately. A bigger stride does not produce representations that are more translationally invariant. Gridpoints are used to allocate neural hardware not to represent positions. Their spacing is limited by the fact that the neural hardware at a gridpoint assumes it will never be representing more than one of whatever it represents, not by the accuracy with which position needs to be represented.
Attempts to deal with the {\bf second} perceived deficiency led to some interesting models. The transforming autoencoder~\cite{transformingautoencoder} forced an encoder to extract an explicit representation of pose in each capsule by insisting that the reconstructed image should be a transformed version of the original image and specifying this transformation as a matrix that multiplied whatever matrix elements were extracted by the encoder. Similarly, the EM capsule model extrapolated much better to new viewpoints when it was forced to use a matrix to represent the relationship of a part to a whole.
Unfortunately, perception has to deal with uncertainties that are not present in computer graphics\footnote{Even if a generative model is stochastic, it may still be certain about which stochastic choices it made. In some more complex generative models, however, a level only specifies the probability distributions of poses for parts at the level below and an iterative process then reconciles these distributions~\cite{OsinderoMRF}. This kind of generative model may be needed for modelling very precise relationships between highly variable parts, such as two praying hands, and it {\it does} need to be able to represent probability distributions over poses.}
and it needs to be able to represent correlated uncertainties in its pose predictions so that multiple sources of information can be combined properly. This rules out a simple matrix representation of pose. Once we accept that distributions over the possible poses of an entity will be represented by the scales assigned to basis functions in the log probability space, it is quite possible that CNNs actually learn to do something like this. This might allow them to approximate Hough transforms, though this is hard to do without taking scalar products of activity vectors.
The {\bf third} deficiency can be rectified by moving to a transformer-like architecture that uses scalar products of activity vectors to modulate attention.
If you like CNNs, GLOM can be viewed as a special type of CNN that differs from a standard CNN in the following ways:
\begin{itemize}
\item It only uses 1x1 convolutions (except at the front end).
\item Interactions between locations are done by parameter-free averaging that implements a coincidence filter which allows it to use a Hough transform to activate units rather than only using matched filters.
\item Rather than using a single feed-forward pass through the levels of representation, it iterates to allow top-down influences that are implemented by neural fields.
\item It includes contrastive self-supervised learning and performs hierarchical segmentation as a part of recognition rather than as a separate task. No more boxes.
\end{itemize}
\subsection{Representing the ISA hierarchy}
An important idea in Good Old-Fashioned Artificial Intelligence (GOFAI) is property inheritance. It is not necessary to explicitly represent that elephants suckle their young because an elephant ISA mammal and, unless otherwise stated, an elephant inherits this property from its more general type. A simple way to implement property inheritance in a neural network is to make different entities correspond to different vectors of activity on the same set of neurons. Imagine that the components of the vector that represents a concept are ordered from very general to very specific. Mammals all have similar values for the more general components and differ on less general components. Indian and African elephants only differ on fairly specific components. When a neural net learns to make the vectors for concepts have causal effects on other vectors, effects which should be the same for all mammals but not the same for all vertebrates will naturally be implemented by the outgoing weights of the neurons that are active for all mammals but not for all vertebrates. This way of implementing property inheritance makes it easy to add exceptions. The components of a vector that are common to birds will learn weights that capture the knowledge that birds fly and the more specific components that differentiate penguins from other birds will learn stronger weights that overrule the general case~\cite{HintonISA}.
This way of implementing property inheritance has the added advantage that types do not need to form a tree. Dogs inherit many properties from being canines (like wolves) but they also inherit many properties from being pets (like cats). There is no guarantee that properties inherited from these more general, partially overlapping classes will be consistent, but, unlike logic, neural networks have no difficulty dealing with conflicting evidence.
At first sight, the idea of using different sections of the vector representation of a concept to capture different levels in the ISA hierarchy conflicts with the idea of using different sections to capture different levels in the part-whole hierarchy. This seems problematic because {\it hooked beak} is a part of a bird but it also defines a type of bird. The two ideas can be reconciled by first dividing the embedding vector for a location into sections that represent different levels in the part-whole hierarchy and then dividing each section into subsections that represent different levels in the type hierarchy.
\subsection{The relationship to 2-D Ising models}
For each location separately, the embedding vectors at levels L-1 and L+1 on the previous time-step provide input to the neurons that represent the current embedding vector at level L. This acts like the conditioning input in a conditional Markov Random Field: it influences the current step of the iterative, island forming process that tries to make the embedding of the location at level L agree with the embeddings of other locations at level L.
In a 2-D Ising model, a two-dimensional array of binary-valued spins settles into a state in which nearby spins tend to agree so as to minimize an energy function that favors agreement between neighboring spins. The model proposed here resembles the 2-D Ising model because it uses a 2-D grid of image locations but it generalizes the model in the following ways:
\begin{enumerate}
\item It replaces binary spins with high-dimensional real-valued vectors. The fact that these lie in a continuous space should facilitate the search for islands of agreement.
\item it replaces a single field of spins with fields at multiple levels, and allows adjacent level embeddings of the same location to interact \cite{HeZemelPerpinan,Saremi}. The interactions between levels are quite complicated because they involve coordinate transformations between parts and wholes. So for each pair of adjacent embedding levels, the top-down and bottom-up interactions at each location must be computed by a multi-layer neural net rather than a simple weight matrix.
\end{enumerate}
\subsection{Comparison with other methods for removing redundancy}
Methods like principal components analysis remove redundancy in the data by limiting the number of available dimensions in the representation. By contrast, a restricted Boltzmann machine with a large number of hidden units squeezes out redundancy by making nearly all of the exponentially many possible binary configurations of the hidden units have such high energy that they are effectively unavailable. This is a much more flexible way of eliminating redundancy~\cite{Zhu}. It can model multiple fat manifolds\footnote{A manifold is a subset of the points in a space that have lower intrinsic dimensionality than the full space. If we take the points on a manifold and add a small amount of noise that has full dimensionality, the points no longer form a strict manifold, but they will all be close to the manifold. Such a set of points will be said to lie on a fat manifold.} that have different intrinsic dimensionalities and even within a fat manifold it can model variations in the effective dimensionality in different parts of the manifold.
The island forming objective belongs to the second class of methods. At each level, it allows for a large number of small islands if that is what the data requires but strives to use a small number of large islands if that is possible.
\section{Video}
This paper focuses on using the GLOM architecture to process a single fixation of a static image, but the architecture is motivated by the need to deal with video and learning from video is often much easier than learning from static images~\cite{flowcaps}, so I will briefly discuss the simplest temporal extension which is to a single fixation of a time-varying image.
To avoid confusion it may be helpful to distinguish three different types of time:
\begin{itemize}
\item {\bf Event time:} This is the actual time at which an event occurs.
\item {\bf Representation time:} This is the actual time at which a particular representation of an event occurs in the neural network. If the bottom-up neural network uses a predictive model, representations of events could be in synchrony with the events themselves or they could even precede the events which would make catching a ball a lot easier.
\item {\bf Reference time:} This is the actual time that an internal representation refers to. When a memory is retrieved, for example, the reference time of the constructed representation is usually long before the representation time. The reference time can also differ by a lot from the event time if the memory is not veridical.
\end{itemize}
For a sequence of frames representing a static image, multiple time steps can be used to settle on an appropriate set of islands at each level. But in a dynamic image, the very same time-steps must also be used to deal with the fact that the occupants of a location at each level can change with time.
An advantage of using islands of identical vectors to represent an object is that motions between successive frames that are small compared to the size of the object only require large changes to a small subset of the locations at the object level. All of the locations that remain within the object need to change only sightly to represent the slight change in pose of the object relative to the camera.
If the changes in an image are small and predictable, the time-steps immediately following a change of fixation point can be used to allow the embeddings at all levels to settle on slowly changing islands of agreement that track the changes in the dynamic image. The lowest level embeddings may change quite rapidly but they should receive good top-down predictions from the more stable embeddings at the level above. Once the embeddings have formed sensible islands, there is then no problem in using the very same time-step for improving the interpretation of each frame and for keeping the embeddings locked on to the dynamic image.
If the changes are rapid, there is no time available to iteratively settle on a good set of embedding vectors for interpreting a specific frame. This means that the GLOM architecture cannot correctly interpret complicated shapes if the images are changing rapidly. Try taking a irregularly shaped potato and throwing it up in the air in such a way that it rotates at one or two cycles per second. Even if you smoothly track the potato, you cannot see what shape it is.
\section{Is GLOM biologically plausible?}
\label{sec:plausible}
Although GLOM is biologically inspired, it has several features that appear to make it very implausible as a biological model. Three of these features are addressed here.
\begin{itemize}
\item The weight-sharing between the bottom-up or top-down models in different columns.
\item The need to process negative pairs of examples for contrastive learning without interrupting the video pipeline.
\item The use of backpropagation to learn the hidden layers of the top-down and bottom-up models.
\end{itemize}
\subsection{Is the neocortex a giant distillery?}
The replication of DNA in every cell is unproblematic: that is what DNA is good at. But biologists often object to models that use weight-sharing claiming that there is no obvious way to replicate the weights~\cite{naturereviewsBP}. GLOM, however, suggests a fairly simple way to solve this problem by using contextual supervision. In a real brain, what we want is an efficient way of training the bottom-up and top-down nets at a location so that they compute the same function as the corresponding nets at other locations. There is no need for the weights to be identical as long as corresponding nets are functionally identical. We can achieve this using knowledge distillation~\cite{Caruana2006,distillation}. For each level separately, the two students at each location are the bottom-up and top-down neural nets. The teacher is the consensus opinion that is a weighted geometric mean of the opinions of the two students, the previous state of the embedding, and the attention-weighted embeddings at {\it other} locations\footnote{Strictly speaking, this is an example of co-distillation where the ensemble of all the students is used as the teacher. Co-distillation was initially based on an analogy with how scientists in a community learn from each other \cite{codistill}, but the same mechanism could be used in a community of columns. In both cases it would help to explain how a system can win by just replicating a lot of people or columns without requiring any significant architectural innovation.}.
Regressing a student's prediction towards the consensus, allows knowledge in the neural nets at other locations to be transferred to the student via the attention weighted averaging. It is not as effective as sharing weights with those other neural nets, but it works quite well~\cite{distillation} and in the long run all of the networks will converge to very similar functions if the data distribution is translation invariant. In the long run, however, we are all dead\footnote{There are alternative facts.}. So it is interesting to consider what happens long before convergence when the local models are all fairly different.
Suppose all of the locations that form a nose have the same embedding vector at the part level. If they all had exactly the same bottom-up model, they would all make exactly the same prediction for the face at the object level. But if the bottom-up models at different locations are somewhat different, we will get a strong ensemble effect at the object level: The average of all the simultaneous bottom-up predictions for the same object in different locations will be much better than the individual predictions.
One advantage of sharing knowledge between locations via distillation rather than by copying weights is that the inputs to the bottom-up models at different locations do not need to have the same structure. This makes it easy to have a retina whose receptive fields get progressively larger further from the fovea, which is hard to handle using weight-sharing in a convolutional net. Many other aspects, such as the increase in chromatic aberration further from the fovea are also easily handled. Two corresponding nets at different locations should learn to compute the same function of the optic array even though this array is pre-processed differently by the imaging process before being presented to the two nets. Co-distillation also means that the top-down models do not need to receive their location as an input since it is always the same for any given model.
Finally, using distillation to share knowledge between location specific neural networks solves a puzzle about the discrepancy between the number of synapses in the visual system (about $10^{13}$) compared to the number of fixations we make in our first ten years (about $10^9$). Conservative statisticians, concerned about overfitting, would prefer these numbers to be the other way around\footnote{It should help that each example allows contrastive learning at several different levels and the target vectors for contrastive learning are much richer than a single one-of-N label}. If we use, say, $10^4$ columns in different locations, the bottom-up and top-down models at one location only have about $10^9$ synapses between them. Conversely, the number of training examples used to learn the knowledge that is shared across an ensemble of $10^4$ locations is about $10^{13}$, though many of these examples are very highly correlated.
Neural networks that have more training cases than parameters are less magical than some of the highly over-parameterized networks in current use but they may generalize in more predictable ways when presented with data that lies outside their training distribution because the function they compute has been much more highly constrained by the data.
\subsection{A role for sleep in contrastive learning?}
If negative examples are required, GLOM might appear less plausible as a biological model because of the added complexity of finding and processing pairs of images that are similar when they should not be. There is, however, one intriguing possibility that emerged from conversations with Terry Sejnowski in 1983 and 2020.
When using contrastive learning to get representations that are similar for neighbouring video frames, the most effective negative examples are frames in the same video that are nearby but not immediately adjacent. We could avoid compromising the real-time performance of GLOM by taking it offline at night to do the negative learning that prevents the representations from collapsing. If the highest level embeddings have the ability to generate sequences at the highest level, the top-down networks could be used to generate sequences of embeddings at every level in each column. This process does not require any attention between columns because it does not need to perform perceptual inference, so it might be able to generate plausible sequences at a much faster rate than the normal speed. Then we simply do the negative learning for the bottom-up models using the same length of real-time window as is used when awake. There is evidence that high-speed, top-down sequence generation occurs during the spindle stage of sleep~\cite{LeeWilson,BuzReplay}.
The idea that sleep is used to keep apart representations that should not be confused is not new~\cite{CrickMitchison}. Hinton and Sejnowski \cite{BMsleep} even suggested that sleep could be used for following the derivatives of the normalizing term in the negative phase of Boltzmann machine learning. But this reincarnation of the idea has two big advantages over Boltzmann machines. First, contrastive unsupervised learning scales much better than Boltzmann machine learning and second, it is far more tolerant of a temporal separation between the positive and negative phases.
Preliminary experiments using contrastive learning for MNIST digits show that the learning still works if a large number of positive updates are followed by a large number of negative updates. Representation collapse is fairly slow during the positive-only learning and the representations can shrink by a significant factor without much affecting performance. So maybe some pairs of embeddings that ought to be well separated get too close together during the day and are then pushed apart again at night. This would explain why complete sleep deprivation for a few days causes such serious mental confusion\footnote{If sleep is just for doing extra rehearsal or for integrating the day's experiences with older experiences, it is not clear why complete lack of sleep for a few days has such devastating effects.}. The experiments with MNIST also show that after a lot of positive-only learning, performance degrades but is rapidly restored by a small amount of negative learning.
To avoid very long periods of negative-only learning, it might be advisable to start with a negative phase of sleep to push representations apart, and then alternate with a positive phase using input sequences that were generated from the top level or even from a recurrent network close to the sensory input. This conflicts with the Crick-Mitchison theory that REM sleep is for unlearning, but it would still be compatible with our failure to remember almost all of our dreams if episodic memory retrieval depends on the top-level, and the top-level simply does not learn during REM sleep because those episodes simply did not happen.
\subsection{Communicating error derivatives in the brain}
The straightforward way to train GLOM is to ask it to fill in missing regions of images and to backpropagate the reconstruction error through the entire temporal settling process using backpropagation through time. The contrastive representation learning at each level can then be viewed as an additional regularizer. Unfortunately, it is hard to see how a brain could backpropagate through multiple time steps. If, however, the consensus opinion at every level can provide a sufficient teaching signal for the bottom-up and top-down models that predict the embedding vector at that level, implementation in a brain becomes a lot more feasible.
If we could ensure that the representations improved over time, the temporal derivatives of neural activities could represent error derivatives and the local learning procedure would then be spike-time dependent plasticity in which the increase in a synapse strength is proportional to the product of the pre-synaptic activity with the post-synaptic rate of increase of activity.\footnote{This fits very well with the strongly held beliefs of Jeff Hawkins and others that the brain learns by predicting what comes next.} Assuming spikes are caused by an underlying rate variable, we can get a noisy but unbiased estimate\footnote{Stochastic gradient descent is extremely tolerant of noise in the gradient estimates so long as they are unbiased.} of the rate of change of this underlying rate variable by applying a derivative filter to the post-synaptic spike train, which is exactly what STDP does.
A recent review paper~\cite{naturereviewsBP} discusses at great length how temporal derivatives can be used as error derivatives in order to approximate backpropagation in a feedforward network\footnote{This paper summarizes the results of simulations that show that this proposal can be made to work quite well, but not as well as vanilla CNNs. A significant contributor to the performance gap is the statistical inefficiency caused by the lack of weight-sharing and co-distillation should help to fix this.}. The review paper assumes a separate phase in which derivatives, in the form of activity perturbations, are allowed to flow back from the higher levels to the lower levels. This process does not seem plausible for a video pipeline. By contrast, the settling process of GLOM propagates the derivatives required for learning as the temporal derivatives of activity at all levels and the time steps required for this propagation can be the very same time steps as are used for video frames.
For dynamic images, it may seem paradoxical that the representations just keep getting better, but it is no more paradoxical than a surfer who just keeps going downhill without ever changing her elevation. The surface on which the surfer is going downhill is not the same surface. Similarly the time-slice of reality for which the representations are forever improving is not the same time slice. The brain surfs reality.
Unfortunately, this does not explain how to get the derivatives required for learning the hidden layers of the bottom-up and top-down neural networks. Nor does it explain how the derivatives of the error signals at each level are backpropagated through the bottom-up or top-down networks to make the appropriate contributions to the derivatives at adjacent levels. Those thorny issues are addressed in another paper which is in preparation.
\section{Discussion}
This paper started life as a design document for an implementation but it was quickly hijacked by the need to justify the design decisions. I have used the imaginary GLOM architecture as a vehicle for conveying a set of interconnected ideas about how a neural network vision system might be organized. The absence of a working implementation makes it easier to focus on expressing the ideas clearly and it avoids the problem of confounding the quality of the ideas with the quality of the implementation, but it also creates serious credibility concerns. The difference between science and philosophy is that experiments can show that extremely plausible ideas are just wrong and extremely implausible ones, like learning a entire complicated system by end-to-end gradient decent, are just right. I am currently collaborating on a project to test out the ability of the GLOM architecture to generalize shape recognition to radically new viewpoints and I am hoping that other groups will also test out the ideas presented here. This paper has gone on long enough already so I will conclude by making some brief philosophical comments.
The idea that nodes in a parse tree are represented by islands of similar vectors unifies two very different approaches to understanding perception. The first approach is classical Gestalt psychology which tried to model perception by appealing to fields and was obsessed by the idea that the whole is different from the sum of the parts\footnote{Thanks to George Mandler for this more accurate translation.}. In GLOM, a percept really is a field and the shared embedding vector that represents a whole really is very different from the shared embedding vectors that represent the parts. The second approach is classical Artificial Intelligence which models perception by appealing to structural descriptions. GLOM really does have structural descriptions, and each node in the parse tree has its own "address" but the addresses live in the continuous space of possible embeddings, not in the discrete space of hardware locations.
Some critics of deep learning argue that neural nets cannot deal with compositional hierarchies and that there needs to be a "neurosymbolic" interface which allows neural network front- and back-ends to hand over the higher-level reasoning to a more symbolic system\footnote{This is reminiscent of Cartesian dualism which postulated an interface between the body and the mind.}. I believe that our primary mode of reasoning is by using analogies which are made possible by the similarities between learned high-dimensional vectors, and a good analogy for the neurosymbolic interface is a car manufacturer who spends fifty years expounding the deficiencies of electric motors but is eventually willing to use them to inject the gasoline into the engine.
The phenomenal success of BERT~\cite{BERT}, combined with earlier work that demonstrates that neural networks can output parse trees if that is what the task requires~\cite{grammar}, clearly demonstrates that neural networks can parse sentences if they want to. By structuring the interactions between the multiple heads in BERT so that they correspond to levels of representation and by adding a contrastively learned regularizer to encourage local islands of agreement over multiple word fragments at each level, it may be possible to show that GLOMBERT actually does parse sentences.
\subsubsection*{Acknowledgments}
Many people helped me arrive at the set of ideas described in this paper. Terry Sejnowski, Ilya Sutskever, Andrea Tagliasacchi, Jay McClelland, Chris Williams, Rich Zemel, Sue Becker, Ruslan Salakhutdinov, Nitish Srivastava, Tijmen Tieleman, Taco Cohen, Vincent Sitzmann, Adam Kosoriek, Sara Sabour, Simon Kornblith, Ting Chen, Boyang Deng and Lala Li were particularly helpful. People who helped me to improve the presentation of the ideas include David Fleet, David Ha, Michael Isard, Keith Oatley, Simon Kornblith, Lawrence Saul, Tim Shallice, Jon Shlens, Andrea Tagliasacchi, Ashish Vaswani and several others. I would especially like to thank Jeff Dean and David Fleet for creating the environment at Google that made this research possible. There are probably many highly relevant papers that I should have read but didn't and I look forward to learning about them.
\bibliographystyle{apalike}
|
\section{Introduction}\label{sec:intro
A {\em thin} Lie algebra is
a graded Lie algebra $L=\bigoplus_{i=1}^{\infty}L_i$
with $\dim(L_1)=2$ and satisfying the following {\em covering property:}
for each $i$, each nonzero $z\in L_i$ satisfies $[zL_1]=L_{i+1}$.
This implies at once that homogeneous components of a thin Lie algebra are at most two-dimensional.
Those components of dimension two are called {\em diamonds,}
hence $L_1$ is a diamond, and if there are no other diamonds then $L$ is a graded Lie algebra of maximal class, see~\cite{CMN,CN}.
However, we adopt the convention of explicitly excluding graded Lie algebras of maximal class from the definition of thin Lie algebras.
We also require thin Lie algebras to be infinite-dimensional in this paper.
Thus, a thin Lie algebra must have at least one further diamond past $L_1$, and we let $L_k$ be the earliest (the {\em second} diamond).
It turns out that $k$ can only be one of $3$, $5$, $q$, or $2q-1$, where $q$ is a power of the characteristic $p$ when this is positive
(with only $3$ and $2q-1$ occurring in characteristic two).
This fact was proved in~\cite{AviJur}, but a revised proof is given in~\cite{Mat:chain_lengths}.
In particular, only $3$ and $5$ can occur in characteristic zero.
It is then tempting to call the cases where $k$ is $3$ or $5$ the {\em classical} cases,
following a usage in the theory of (finite-dimensional) simple modular Lie algebras, where the classical ones
are those which are analogues of the simple Lie algebras of characteristic zero
(including the exceptional ones,
as is customary in the modular theory).
The main conclusion of this paper, and its interpretation which we give in Section~\ref{sec:loop},
lends further weight to adopting such terminology.
Thin Lie algebras with $k$ equal to $3$ or $5$, subject to a further restriction $\dim(L_4)=1$ in the former case, and excluding some small characteristics,
were shown in~\cite{CMNS} to belong to (up to) three isomorphism types,
associated to $p$-adic Lie groups of the classical types $A_1$ and $A_2$ (see~\cite{Mat:thin-groups} for matrix realizations of those groups).
In contrast, the values $q$ and $2q-1$ for $k$ occur for two broad classes of thin Lie algebras, of which many were built from certain non-classical finite-dimensional
simple modular Lie algebras, and also to other thin Lie algebras obtained from graded Lie algebras of maximal class through various constructions.
General discussions of those two classes of thin Lie algebras can be found in~\cite{AviMat:A-Z} and~\cite{CaMa:Hamiltonian}, respectively.
Our main result is a general fact that was shown so far in {\em ad hoc} manners for various special instances of the class of thin Lie algebras under consideration.
\begin{theorem}\label{thm:sandwich}
Let $L$ be a thin Lie algebra with second diamond $L_k$, where $k>5$.
Then there is a nonzero element $y$ of $L_1$ such that $[Lyy]=0$.
\end{theorem}
We use the left-normed convention
for iterated Lie products, hence $[abc]$ stands for $[[ab]c]$,
and so $[Lyy]=0$ is another way of saying $(\ad y)^2=0$.
According to a definition of Kostrikin, for $p\neq 2$ the conclusion of Theorem~\ref{thm:sandwich}
says that $y$ is a {\em sandwich element} (or simply a {\em sandwich}).
Kostrikin introduced sandwich elements in the context of the Burnside problem, see~\cite{Kostrikin:Burnside}.
Four decades later they played an important role in the classification theory of finite-dimensional simple modular Lie algebras,
especially for small characteristics (five and seven) in~\cite{PreStr:small_I}.
One remarkable property of sandwich elements, noted by Kostrikin and proved by Premet in~\cite{Premet:degeneration},
is that their presence in a finite-dimensional modular Lie algebra characterizes those simple Lie algebras which are not classical.
Although our thin Lie algebras are infinite-dimensional, a connection with finite-dimensional Lie algebras
comes via a {\em loop algebra} construction, which has been used many times to build thin Lie algebras
from finite-dimensional Lie algebras, simple or close to simple, with respect to certain cyclic gradings.
In Section~\ref{sec:loop} we discuss the relevance of the presence of a sandwich element in this context.
\section{A sandwich element in thin Lie algebras}\label{sec:sandwiches
A {\em sandwich element} of a Lie algebra $L$
is a nonzero element $c\in L$ such that $(\ad c)^2=0$ and $(\ad c)(\ad z)(\ad c)=0$ for all $z\in L$.
If the characteristic is different from two then the latter requirement (from which the name originates) is superfluous, as it
follows from the former due to
$0=[u[zcc]]=[uzcc]-2[uczc]+[uccz]$
for $u,z\in L$.
In the context of thin Lie algebras it turns out that a nonzero element $y$ of $C_{L_1}(L_2)$
satisfies $[Lyy]=0$ in most cases, as our Theorem~\ref{thm:sandwich} states,
and so $y$ is a sandwich element if the characteristic is not two.
This fact was crucial in the theory of graded Lie algebras of maximal class, making a theory of {\em constituents} possible,
although the sandwich point of view was not explicitly mentioned in~\cite{CMN} or~\cite{CN}.
In particular, \cite[Lemma~3.3]{CMN} can be restated as follows:
if $L$ is a graded Lie algebra of maximal class,
with the centralizer $C_{L_1}(L_2)$ spanned by an element $y$ as customary,
$[L_{i-1}y]=0$ and $[L_iy]\neq 0$ for some $i$, then $[L_{i+1}y]=0$,
whence $[L_iyy]=0$.
A remarkable consequence of this fact is $[Lyy]=0$, which means $(\ad y)^2=0$, and so $y$ is a sandwich element if the characteristic is not two.
Our main goal in this section is showing that such a nonzero element $y$ of $C_{L_1}(L_2)$
is frequently a sandwich element also in a thin Lie algebra $L$, under certain assumptions which we introduce and justify along the way.
If $L$ has second diamond $L_3$, then $C_{L_1}(L_2)=0$, hence we must take $\dim(L_3)=1$ as minimal assumption for our discussion, and then
$\dim\bigl(C_{L_1}(L_2)\bigr)=1$.
Taking then for $y$ a nonzero element of $C_{L_1}(L_2)$ and extending to a basis $x,y$ of $L_1$,
as we do without further mention in this section, we have $[xyy]=0$.
In the case of characteristic two we restrict ourselves to the following simple observation.
\begin{theorem}\label{thm:sandwich_even}
Let $L$ be a thin Lie algebra of characteristic two, with $\dim(L_3)=1$, and
let $y$ span $C_{L_1}(L_2)$.
Then $[Lyy]=0$.
\end{theorem}
\begin{proof}
Because $(\ad y)^2$ is a derivation of $L$, its kernel is a subalgebra.
However, because $[xyy]=0$ both generators
$x$ and $y$ of $L$ belong to the kernel, and hence that is the whole of $L$.
\end{proof}
We will show that Theorem~\ref{thm:sandwich_even} extends to arbitrary characteristic if we include a further hypothesis $\dim(L_5)=1$, as in Theorem~\ref{thm:sandwich}.
A justification for that hypothesis will emerge as we gradually work towards a proof.
We start with suitably extending the conclusion $[L_iyy]=0$ of~\cite[Lemma~3.3]{CMN}, which was about graded Lie algebras of maximal class,
to certain homogeneous components of any thin Lie algebra $L$ with $\dim(L_3)=1$.
\begin{lemma}\label{lemma:uxyy}
Let $L$ be a thin Lie algebra with $\dim(L_3)=1$, and let $y$ span $C_{L_1}(L_2)$.
Let $L_i$ be a homogeneous component with $[L_iy]\neq 0$, and suppose $L_{i-1}$ has a nonzero element $u$ with $[uy]=0$.
Then $[L_iyy]=0$.
Moreover, $L_i$ and $L_{i+2}$ have dimension one.
\end{lemma}
\begin{proof}
Because of the covering property, $L_i$ is spanned by $[ux]$, and hence has dimension one.
From
\begin{equation}\label{eq:sandwich}
0=[u[xyy]]=[uxyy]-2[uyxy]+[uyyx]=[uxyy]
\end{equation}
it follows that $[L_iyy]=0$.
Because $[uxy]\neq 0$ by hypothesis, the covering property implies that
$L_{i+2}$ is spanned by $[uxyx]$, and hence has dimension one.
\end{proof}
The following immediate consequence of Lemma~\ref{lemma:uxyy} resembles the formulation of~\cite[Lemma~3.3]{CMN} more closely.
\begin{cor}\label{cor:y_centr}
Let $L$ be a thin Lie algebra with $\dim(L_3)=1$, and let $y$ span $C_{L_1}(L_2)$.
If $L_{i-1}$ has a nonzero element centralized by $y$, and $L_i$ has none, then $L_{i+1}$ has such an element as well.
\end{cor}
Equivalently, in a thin Lie algebra with $\dim(L_3)=1$,
at least one of any two consecutive homogeneous components has a nonzero element centralized by $y$.
Corollary~\ref{cor:y_centr} is often useful as simple but weaker replacement for Lemma~\ref{lemma:uxyy}.
However, we need the full strength of Lemma~\ref{lemma:uxyy} in order to deduce the following consequence.
\begin{cor}\label{cor:vyy}
Let $L$ be a thin Lie algebra with $\dim(L_3)=1$,
and let $y$ span $C_{L_1}(L_2)$.
Suppose $\dim(L_j)=2$ for some $j>1$.
Then $\dim(L_{j-1})=1$ and $[L_{j-1}yy]=0$.
In particular, $L_j$ contains a nonzero element centralized by $y$.
\end{cor}
\begin{proof}
Because of the covering property $L_{j-1}$ cannot contain any nonzero element centralized by $y$.
Hence $L_{j-2}$ does contain a nonzero element $u$ with $[uy]=0$.
Consequently, $v=[ux]$ spans $L_{j-1}$,
and $[vyy]=0$ according to Lemma~\ref{lemma:uxyy}.
Because $\dim(L_j)=2$ the element $[vy]$ of $L_j$ is nonzero, and is centralized by $y$ as desired.
\end{proof}
In particular, Corollary~\ref{cor:vyy} implies that two consecutive components in a thin Lie algebra $L$
with $\dim(L_3)=1$ cannot both be diamonds.
Thus, the {\em distance} between any two consecutive diamonds, by which we mean the difference of their degrees, is at least two.
An easily proved consequence of this fact is that $L$ remains thin under base field extensions.
By contrast, this property fails for a thin Lie algebra $L$ with two diamonds occurring as consecutive components, as we show in Remark~\ref{rem:field_extensions}.
In fact, the thin Lie algebras studied in~\cite{GMY,ACGMNO}
where all homogeneous components except $L_2$ are diamonds, require their base field to have a quadratic field extension.
\begin{rem}\label{rem:field_extensions}
Although it is inconsequential for this paper, we explain why in a thin Lie algebra $L$
over an algebraically closed field no two diamonds can occur as consecutive components.
For a graded Lie algebra $L=\bigoplus_{i=1}^{\infty}L_i$ we consider the linear maps
$\psi_j:L_j\to\Hom_F(L_1,L_{j+1})$ obtained by restriction from the adjoint representation.
When $L$ is thin with $L_j$ and $L_{j+1}$ both diamonds,
the covering property for the homogeneous component $L_j$ means that $\psi(L_j)$ is a two-dimensional subspace of $\Hom_F(L_1,L_{j+1})$
where every nonzero element is a surjective linear map.
Upon composing with a linear bijection $L_{j+1}\to L_1$ (whose choice is immaterial) we may view that as a two-dimensional subspace of
$\Hom_F(L_1,L_1)$ where every nonzero element has nonzero determinant.
In terms of the associated projective spaces, $P\bigl(\psi(L_j)\bigr)$ is then
a one-dimensional projective subspace of the three-dimensional projective space $P\bigl(\Hom_F(L_1,L_1)\bigr)$,
disjoint from the non-degenerate quadric given by the zeroes of the determinant map.
That cannot happen if the base field $F$ is algebraically closed.
\end{rem}
Corollary~\ref{cor:vyy} is relevant in assigning {\em types} to certain diamonds of a thin Lie algebra.
This was done in different ways according to whether the second diamond
occurs in degree $q$ or $2q-1$, in~\cite{CaMa:Nottingham} and~\cite{CaMa:thin}, respectively.
However, in both cases necessary conditions for a diamond $L_j$ to be assigned a type
were $\dim(L_{j-1})=1$ and $[L_{j-1}yy]=0$.
Corollary~\ref{cor:vyy} shows that those conditions hold in any thin Lie algebra with $\dim(L_3)=1$.
Now we use Lemma~\ref{lemma:uxyy} to prove that if $L$ is thin with $\dim(L_3)=1$, then
$[L_iyy]=0$ for all one-dimensional components $L_i$.
\begin{lemma}\label{lemma:sandwich}
Let $L$ be a thin Lie algebra with $\dim(L_3)=1$,
and let $y$ span $C_{L_1}(L_2)$.
Suppose $[L_jyy]\neq 0$ for some $j$.
Then $\dim(L_j)=2$.
\end{lemma}
\begin{proof}
Note that $j>2$.
If $L_{j-1}$ had a nonzero element centralized by $y$, then because $[L_jy]\neq 0$ by hypothesis,
Lemma~\ref{lemma:uxyy} would apply with $i=j$ and yield $[L_jyy]=0$, a contradiction.
Therefore, $L_{j-1}$ does not have any nonzero element centralized by $y$.
Then because of Corollary~\ref{cor:y_centr} each of $L_{j-2}$ and $L_{j}$ has such an element.
In particular, because $[L_jy]\neq 0$ we must have $\dim(L_j)=2$.
\end{proof}
Thus, the task of proving $[Lyy]=0$ in a thin Lie algebra $L$ has now been reduced to showing $[L_jyy]$ for the diamonds $L_j$.
We pause to note that $L$ having any two diamonds at distance two implies $[Lyy]\neq 0$.
In fact, if $L_j$ and $L_{j+2}$ are diamonds, in a thin Lie algebra $L$ with $\dim(L_3)=1$,
then the covering property easily implies $[L_jy]\neq 0$, whence $[L_jy]=L_{j+1}$
because $L_{j+1}$ is one-dimensional.
Consequently, $[L_jyy]=[L_{j+1}y]\neq 0$, because $[L_{j+1}y]$ and $[L_{j+1}y]$ span the diamond $L_{j+2}$.
The following remark describes one special thin Lie algebra with $\dim(L_3)=1$
having certain diamonds at distance two.
\begin{rem}\label{rem:diamond_deg_5}
The proof of~\cite[Theorem~2(a)]{CMNS} implies that in characteristic either zero or larger than five there is a unique (infinite-dimensional)
thin Lie algebra $L$ having second diamond $L_5$.
That result, which predates the study of thin Lie algebras {\em per se,}
only claims uniqueness for the graded Lie algebra $L$ associated to the lower central series
of a infinite thin pro-$p$ group with second diamond in weight $5$, for $p>5$.
However, its proof applies to any thin Lie algebra $L$ with second diamond $L_5$, for $p>5$ or $p=0$.
A construction for that thin Lie algebra $L$ was also given in~\cite{CMNS}, and that is valid in every characteristic except for characteristic two.
The diamonds of $L$ occur in each degree congruent to $\pm 1$ modulo $6$, and so as noted above one has
$[L_{i-1}yy]\neq 0$ for each $i$ multiple of six.
Uniqueness of $L$ fails in characteristic three because $5=2q-1$ with $q=3$, and in characteristic five because we may have $5=q$ then,
and so lots of other thin Lie algebras enter those cases (see Remark~\ref{rem:char_5}).
\end{rem}
Our next remark describes how in characteristic three one can produce uncountably many thin Lie algebras with $\dim(L_3)=1$
having diamonds at distance two.
\begin{rem}\label{rem:char_3}
According to~\cite[Section~9]{CMN}, for each power $q$ of the (positive) characteristic
there are uncountably many infinite-dimensional graded Lie algebras of maximal class $M$
with precisely two distinct two-step centralizers
and constituent sequence beginning with $2q,q,q$.
As shown in ~\cite[Section~5]{CaMa:thin},
each such $M$ has a maximal subalgebra $L$ which becomes thin under a new grading, and its diamonds start with $L_1,L_{2q-1},L_{3q-2},L_{4q-3}$.
In characteristic three and taking $q=3$ those diamonds are $L_1,L_5,L_7,L_9$.
Consequently, $[L_5yy]\neq 0$ and $[L_7yy]\neq 0$.
\end{rem}
The thin Lie algebras of the above remarks suggest that we might be able to obtain more information
on the earliest diamond $L_j$ with $[L_jyy]\neq 0$
than on an arbitrary one,
hence we will do that in preparation for a proof of Theorem~\ref{thm:sandwich_odd}.
\begin{lemma}\label{lemma:sandwich_plus}
Let $L$ be a thin Lie algebra with $\dim(L_3)=1$,
and let $y$ span $C_{L_1}(L_2)$.
Suppose $[Lyy]\neq 0$, and let $j$ be minimal such $[L_jyy]\neq 0$.
Then $L_{j-2}$ and $L_{j-3}$ are one-dimensional and centralized by $y$.
\end{lemma}
\begin{proof}
According to Lemma~\ref{lemma:sandwich} we have $\dim(L_j)=2$,
and hence $\dim(L_{j-1})=1$ because of Corollary~\ref{cor:vyy}.
Note that $j>3$.
Let $t$ be any nonzero element of $L_{j-3}$.
Then $[tyy]=0$ and $[txyy]=0$ by minimality of $j$, and hence
\[
0=[t[xyy]]=[txyy]-2[tyxy]+[tyyx]=-2[tyxy].
\]
Because of Theorem~\ref{thm:sandwich_even} we are not in characteristic two, and so we conclude $[tyxy]=0$.
But then $[tyx]=0$, because $L_{j-1}$ contains no nonzero element centralized by $y$.
The covering property now implies $[ty]=0$, otherwise
$[tyx]$ and $[tyy]$, which both vanish, would have to span $L_{j-1}$.
Because $t$ was an arbitrary nonzero element of $L_{j-3}$,
Corollary~\ref{cor:vyy} implies $\dim(L_{j-3})=1$.
Finally, $L_{j-2}$ has dimension one because it is spanned by $[tx]$.
But we know that $L_{j-2}$ contains a nonzero element centralized by $y$, and hence $[L_{j-2}y]=0$.
\end{proof}
We are ready to prove the harder analogue of Theorem~\ref{thm:sandwich_even} for odd characteristics, thus completing a proof of Theorem~\ref{thm:sandwich}.
As our discussion leading to Remark~\ref{rem:diamond_deg_5} shows, we need an additional assumption to ensure that
no diamonds occur at distance two.
As it turns out, assuming $\dim(L_5)=1$ will do, which is another way of asking that the second diamond of $L$ occurs past $L_5$.
\begin{theorem}\label{thm:sandwich_odd}
Let $L$ be a thin Lie algebra of odd characteristic, with
$\dim(L_3)=\dim(L_5)=1$, and
let $y$ span $C_{L_1}(L_2)$.
Then $[Lyy]=0$.
\end{theorem}
\begin{proof}
Besides the relation $[yxy]=0$, which serves to define $y$ up to a scalar, in $L$ we also have $[yxxy]=0$ and $[yxxxy]=0$.
In fact, the former assertion follows from $0=[yx[yx]]=[yxyx]-[yxxy]=[yxxy]$,
and hence $[yxxx]$ spans $L_4$.
To prove the latter, assuming $[yxxxy]\neq 0$ for a contradiction, it will span $L_5$ because of our hypothesis $\dim(L_5)=1$.
Because $0=[yxx[xyy]]=[yxxxyy]$ we then have $[L_5y]=0$, and hence $[yxxxyx]$ spans $L_6$.
Now the generalized Jacobi identity yields the contradiction
$0=[y[yxxxx]]=-4[yxxxyx]$.
Now suppose for a contradiction that $[Lyy]\neq 0$, and let $j$ be minimal such that $[L_jyy]\neq 0$.
We will use the relations $[yxxy]=0$ and $[yxxxy]=0$ to derive a contradiction.
Let $t$ be a nonzero element of $L_{j-3}$.
According to Lemma~\ref{lemma:sandwich} we have $[ty]=0$ and $[txy]=0$,
and $L_{j-1}$ is spanned by $v:=[txx]$.
Furthermore, $\dim(L_j)=2$, and hence $[vyy]=0$ according to Corollary~\ref{cor:vyy}.
The calculations
\[
0=[v[xyy]]=[vxyy]-2[vyxy],
\]
and
\[
0=[t[yxxxy]]=[t[yxxx]y]=-3[txxyxy]+[txxxyy]=[vxyy]-3[vyxy],
\]
taken together imply $[vxyy]=[vyxy]=0$.
Because $[vx]$ and $[vy]$ span $L_j$ we obtain $[L_jyy]=0$,
which gives the desired contradiction.
\end{proof}
We discuss the extent to which hypothesis $\dim(L_5)=1$ of Theorem~\ref{thm:sandwich_odd} is necessary.
As we recalled in Remark~\ref{rem:diamond_deg_5},
over a field of characteristic $p>5$ that hypothesis serves to exclude a single exception,
given by the unique thin Lie algebra with second diamond $L_5$, identified in~\cite[Theorem~2(a)]{CMNS} and described in~Remark~\ref{rem:diamond_deg_5}.
However, if we omit the hypothesis $\dim(L_5)=1$ of Theorem~\ref{thm:sandwich_odd} in characteristic $p=3$,
its conclusion $[Lyy]=0$ is violated by uncountably many thin Lie algebras with second diamond $L_5$ and third diamond $L_7$,
which we described in Remark~\ref{rem:char_3}.
The following remark focuses on the borderline case $p=5$.
\begin{rem}\label{rem:char_5}
We discuss how far Theorem~\ref{thm:sandwich_odd} may be extended to include
thin Lie algebras of characteristic five, with second diamond $L_5$ and an additional assumption.
Countably many thin Lie algebras with second diamond $L_q$ were constructed in~\cite{AviMat:A-Z},
for any power $q$ of the odd characteristic.
In particular, when $q=5$ this gives us a countable family of thin Lie algebras with second diamond $L_5$.
However, all those thin Lie algebras, which have third diamond $L_{2q-1}=L_9$
(possibly {\em fake,} see~\cite{AviMat:A-Z} for what that means), do satisfy the conclusion $[Lyy]=0$ of Theorem~\ref{thm:sandwich_odd}.
In fact, one can prove that $[Lyy]=0$ holds for any thin Lie algebra $L$ of characteristic five satisfying
$\dim(L_3)=1$, $\dim(L_5)=2$, and $[L_7y]=0$.
(Here the hypothesis $[L_7y]=0$ implies that neither $L_7$ nor $L_8$ is a diamond, and hence the third diamond of $L$ does not occur earlier than $L_9$.)
A proof of this fact follows the general inductive strategy employed in~\cite{AviMat:Nottingham_structure},
but is not a formal consequence of~\cite{AviMat:Nottingham_structure},
where some generality was sacrificed in favour of simpler exposition.
In fact, part of that simplification in~\cite{AviMat:Nottingham_structure} relies on using our Theorem~\ref{thm:sandwich_odd}.
\end{rem}
\section{Thin loop algebras and the sandwich element}\label{sec:loop}
We conclude this paper with a discussion of the significance for thin Lie algebras of an important property of sandwich elements.
Classical simple Lie algebras
do not have sandwich elements, see~\cite[p.~124]{Sel}.
By contrast, confirming a conjecture of Kostrikin, Premet proved in~\cite{Premet:degeneration} that
every finite-dimensional simple Lie algebra, over an algebraically closed field of characteristic $p>5$,
which is not classical, must have sandwich elements (that is, have {\em strong degeneration}).
The connection of this characterization of classical Lie algebras with (infinite-dimensional) thin Lie algebras
comes from the fact that several of the latter have been constructed as {\em loop algebras}
of certain finite-dimensional simple Lie algebras, or close to simple.
In the simplest setting, one starts from a finite-dimensional simple Lie algebra $S$ over a field $F$,
with a cyclic grading $S=\bigoplus_{k\in\Z/N\Z} S_k$, and considers the Lie algebra $S\otimes F[t]$ over $F$, where $t$ is an indeterminate.
Its subalgebra
$L=\bigoplus_{k>0}S_{\bar k}\otimes t^k$,
where $\bar k=k+N\Z$,
is naturally graded over the positive integers, and is called a {\em loop algebra} in this context.
In certain cases one needs a slightly more general construction involving also a derivation of $S$
(such as that of~\cite[Definition~2.1]{AviMat:A-Z}),
but that is inconsequential for our present observation.
It was proved in~\cite{CMNS} that a thin Lie algebra $L$ (of characteristic not $2$ or $3$)
having second diamond $L_3$ and $\dim(L_4)=1$ belongs to one of (up to) two isomorphism types,
and a thin Lie algebra $L$ (of characteristic not $2$, $3$ or $5$) with second diamond $L_5$ is unique up to isomorphism
(see also Remark~\ref{rem:diamond_deg_5}).
Each of those Lie algebras can be realized as a loop algebra of a classical simple Lie algebra of type $A_1$ or $A_2$.
Now, the fact that those thin Lie algebras of~\cite{CMNS} are loop algebras of classical simple Lie algebras implies
that there cannot be any nonzero element $y\in L_{\bar 1}$ with $(\ad y)^2=0$, because such $y$ would have the form
$y=c\otimes t$ for some sandwich element $c$ of $S$, which cannot exist.
For the same reason, if a thin Lie algebra $L$ (say of characteristic $p>5$) with second diamond past $L_5$ is a loop algebra of a simple Lie algebra $S$,
then $S$ cannot be classical, because $L$ has a sandwich element $y$ according to Theorem~\ref{thm:sandwich_odd}, and hence so does $S$.
In fact, all constructions of thin Lie algebras with second diamond $q$ or $2q-1$ as loop algebras which were given
in several papers~\cite{Car:Nottingham,AviMat:-1,CaMa:Hamiltonian,AviMat:A-Z,AviMat:mixed_types}
start from non-classical simple Lie algebras of various types.
The occurrence of non-classical algebras in those constructions is hardly surprising,
because the expected structure of such thin Lie algebras $L$ given by corresponding uniqueness results
shows that the dimension of $S$ ought to be a power of $p$ or one or two less in the various cases.
However, the sandwich element $y$ does provide guidance in identifying the appropriate cyclic grading of $S$ employed in those constructions.
|
\section{Introduction}
\IEEEPARstart \noindent Canine stifle joint issues are one of the oldest and most faced issues in the veterinary orthopedics sector \cite{seo2020measurement}. The value of Tibial Plateau Angle (TPA) can be used to identify issues that exist in a canine's leg, how it may react to different forms of injuries or even if there is presence of pre-existing conditions \cite{sorensson2021evaluation}. Injuries such as the cranial cruciate ligament (CCL) rupture (CCLR) are very common, major and progressively deteriorate stifle joints of canines permanently; automated Tibial Plateau Angle (TPA) assessment can help in shortening the process of CCL detection during surgical procedures \cite{pacchiana2003surgical}, additionally, as the TPA value plays an important role in the distribution of force when a dog walks, it can also determine whether or not there is excessive cranial tibial thrust that may predispose canines to CCLR \cite{seo2020measurement, kim2015measurement}.
\indent The utility of automation has become ubiquitous in the modern world where everything is electronically powered and we, humans rely more and more on artificial intelligence for assistance. Automation in image annotation has become a major tool in the medical field, driving patient care decisions \cite{grady2020system}. The latest method of obtaining such decisions in an automated manner involves using a processor for digital image representation acquisition that simultaneously generates annotations and determines the associations between multiple annotations of objects of same class or group \cite{grady2020system}. This same processor simultaneously works on determining and representing the said classes while recording the annotations in its limited memory \cite{grady2020system}. There has been variety of work regarding determination TPA considering multiple factors such as age, sex, breed etc. as seen in studies \cite{kim2015measurement, pacchiana2003surgical, sorensson2021evaluation, moore1995cranial}, however none thus far for the automation of the process. This study delves into this aspect of application and completes the first stage of automation in TPA determination.
\section{TPA Calculation}
\noindent TPA calculation has some prerequisites that must be fulfilled, these are:
\begin{enumerate}
\item The Stifle and Tarsus must be 90\degree flexed;
\item Functional Tibial line must be formed by connecting
\begin{enumerate}
\item Centre of Talus
\item The centre of intercondylar eminences shown in Fig. \ref{fig:tparecs};
\end{enumerate}
\item The Medial Tibial Plateau line should be drawn using the first two.
\end{enumerate}
\begin{figure}[b]
\centering
\begin{subfigure}[b]{0.26\linewidth}
\centering
\includegraphics[width=0.8in, height=1.5in]{Images/TPA_flex.jpg}
\caption{}
\end{subfigure}
\begin{subfigure}[b]{0.26\linewidth}
\centering
\includegraphics[width=0.8in, height=1.5in]{Images/FTL.jpg}
\caption{}
\end{subfigure}
\begin{subfigure}[b]{0.26\linewidth}
\centering
\includegraphics[width=0.8 in, height=1.5in]{Images/mtpl.jpg}
\caption{}
\end{subfigure}
\caption{\small{TPA requirements (a) Stifle (knee joint) and Tarsus (ankle joint) flex, (b) Green line representing the FTL and (c) MTPL pointed with green arrow \cite{MTPAvid}.}}
\label{fig:tparecs}
\end{figure}
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.3\linewidth}
\centering
\includegraphics[width=0.9in,height=1in]{Images/centre_of_IE_point.jpg}
\caption{}
\end{subfigure}
\begin{subfigure}[t]{0.3\linewidth}
\centering
\includegraphics[width=0.9in,height=1in]{Images/centre_of_IE_ref.jpg}
\caption{}
\end{subfigure}
\begin{subfigure}[t]{0.3\linewidth}
\centering
\includegraphics[width=0.9in,height=1in]{Images/centre_of_talus.jpg}
\caption{}
\end{subfigure}
\caption{\small{Functional Tibial line drawing needs; (a) Intercondylar Eminence point, (b) Intercondylar Eminence point reference and (c) Centre of Talus \cite{MTPAvid}.}}
\label{fig:ftlrecs}
\end{figure}
\indent Following the identification of points of interests, and drawing of the lines of interest, i.e. the Functional Tibial Line (FTL) and the Medial Tibial Plateau Line (MTPL), another line is drawn such that the relation of new line and the FTL is 90\degree. Tibial plateau angle is the angle between this new line and the MTPL; this is shown in Fig. \ref{fig:tparef}.
\begin{figure}[b]
\centering
\includegraphics[height=2in]{Images/tparef.jpg}
\caption{\small{Tibial Plateau Angle shown in relation with the new perpendicular line drawn and the MTPL \cite{MTPAvid}.}}
\label{fig:tparef}
\end{figure}
Normal TPA values can range from 18 to 25 degrees \cite{seo2020measurement}; the large range can be attributed to the large range of breed, body weight, age etc.
\subsection{Training Dataset}
\noindent The dataset used for training was collected from various veterinary clinics, as none is available in the public domain, and the objects of interests were manually annotated. The resolution of these images vary largely as they are collected from various sources, thus for the sake of uniformity the images are all scaled to fit the same dimensions. The first part of this project was to develop a lightweight radiograph image sorting algorithm reported in \cite{Tonima_2021}. The images chosen for the task described in this manuscript were all classified to be lateral lower body images by that sorting algorithm. Examples of manual annotations of objects of interests are given in Fig. \ref{fig:tpaann}: here regions A, B and C identify the joints while the point 'e' identifies the centre of Talus and regions d1 and d2 identify the points that form the MTPL.
\begin{figure}[t]
\centering
\includegraphics[width=0.3\columnwidth]{Images/an_ex.jpg}
\caption{\small{Example of manual annotations, highlighting the regions of interests \cite{labelIMg}.}}\label{fig:tpaann}
\end{figure}
\\\indent The images from the source dataset had issues and needed to be consolidated into a more usable framework. These issues were mostly due to inconsistencies in practices of radiographers and movement by the animals during radiography that resulted in radiographs with incomplete data, inconsistent image quality, such as varying contrast, brightness and positioning of point of interest or images that failed to meet the prerequisite conditions for this task. Examples of difficult data are shown in Fig. \ref{fig:difdat}.
\begin{figure}[t]
\centering
\includegraphics[width=.3\columnwidth, height=1.5in]{Images/unsuitable/unsuitable_1.jpg}\hfill
\includegraphics[width=.3\columnwidth, height=1.5in]{Images/unsuitable/unsuitable_2.jpg}\hfill
\includegraphics[width=.3\columnwidth, height=1.5in]{Images/unsuitable/unsuitable_3.jpg}\hfill
\\[\smallskipamount]
\includegraphics[width=.3\columnwidth, height=1.5in]{Images/unsuitable/unsuitable_4.jpg}\hfill
\includegraphics[width=.3\columnwidth, height=1.5in]{Images/unsuitable/unsuitable_5.jpg}\hfill
\includegraphics[width=.3\columnwidth, height=1.5in]{Images/unsuitable/unsuitable_6.jpg}\hfill
\caption{\small{Some examples of difficult data}}\label{fig:difdat}
\end{figure}
\\\indent Most of the source dataset of over a thousand unique knee radiographs were distorted, had poor patient positioning, or were otherwise unfit for annotating. 250 of the original images were of sufficient quality to contribute to this effort. These images were set to have 6 different classes for training purposes, as mentioned and shown in Fig. \ref{fig:tpaann}. These were then trained using YOLOv3 \cite{redmon2018YOLOv3} and the results of the predicted annotations are shown in Table \ref{table:tpares}. From these predictions the centroids were extracted, which is then used to plot the FTL and MTPL. Then using the method mentioned in Section II, TPA is calculated.
Examples of images, the region of interests detection and their respective TPA determination is shown in Fig. \ref{fig:tpapred} and Table \ref{table:tpares}:
\begin{figure}[ht!]
\centering
\begin{subfigure}[]{0.4\linewidth}
\centering
\includegraphics[width=1.2in, height=2in]{test_results/TR1.PNG}\hfill
\caption{}
\end{subfigure}
\begin{subfigure}[]{0.4\linewidth}
\centering
\includegraphics[width=1.2in, height=2in]{test_results/TR4.PNG}\hfill
\caption{}
\end{subfigure}
\\[\smallskipamount]
\begin{subfigure}[]{0.4\linewidth}
\centering
\includegraphics[width=1.2in, height=2in]{test_results/TR5.PNG}\hfill
\caption{}
\end{subfigure}
\begin{subfigure}[]{0.4\linewidth}
\centering
\includegraphics[width=1.2in, height=2in]{test_results/TR7.PNG}\hfill
\caption{}
\end{subfigure}
\\[\smallskipamount]
\begin{subfigure}[]{0.4\linewidth}
\centering
\includegraphics[width=1.2in, height=2in]{test_results/TR8.PNG}\hfill
\caption{}
\end{subfigure}
\begin{subfigure}[]{0.4\linewidth}
\centering
\includegraphics[width=1.2in, height=2in]{test_results/TR9.PNG}\hfill
\caption{}
\end{subfigure}
\caption{\small{Example of algorithm detecting, highlighting the regions of interests.}}\label{fig:tpapred}
\end{figure}
\begin{table}[tb]
\centering
\caption{\small{TPA shown by trained system with calculations from the detected annotations}}
\begin{tabular}{|c | c|}\hline
\textbf{Figure Number} & \textbf{TPA} \\ \hline
\ref{fig:tpapred}(a) & 20.537\degree\\ \hline
\ref{fig:tpapred}(b) & 17.354\degree\\ \hline
\ref{fig:tpapred}(c) & 19.473\degree\\ \hline
\ref{fig:tpapred}(d) & 23.369\degree\\ \hline
\ref{fig:tpapred}(e) & 18.435\degree\\ \hline
\ref{fig:tpapred}(f) & 19.699\degree\\ \hline
\end{tabular}
\label{table:tpares}
\end{table}
\subsection{Activation Functions}
\noindent As activation functions are explored in this paper for the purpose of performance comparison, a small section has been dedicated for understanding of the roles and types of activation function that have been used. Activation function is simply the pathway that allows feeding of the input and output to and from the current neuron. This can range in variety of forms, from as simple as an on/off switch---i.e, Step function to as complex as the Sigmoid. In this section only the functions used for results obtained are discussed:
\begin{enumerate}
\item Linear activation function is exactly what the name suggests. It takes an input, multiples it with the learned weight and produces an output that is a function of the input.
\begin{equation}
f(x)=mx
\end{equation}
\item ReLU is a rectified version of the linear function that does not allow the negative inputs.
\begin{equation}
f(x)=max\begin{cases}x & \text {for $x>0$}\\0 & \text {otherwise}
\end{cases}\end{equation}
The graphical representation of this is shown in Fig. \ref{fig:act}(a).
\item Leaky ReLU is a variation of ReLU that has a small slope on the negative area and thus is more complacent toward negative inputs.
\begin{equation}
f(x)=max\begin{cases}x & \text {for $x>0$}\\ax & \text { $x<=0$}
\end{cases}\end{equation}
The graphical representation of this is shown in Fig. \ref{fig:act}(b).
\item Swish is a gated sigmoid function that has an interesting mathematical form.
\begin{equation}
f(x)=x\times \sigma (\beta (x))
\end{equation}
where the $\beta$ can be a constant or a trainable quantity, depending on which the Swish may act either like a scaled linear (for $\beta=0$) or a ReLU (for $\beta \to \infty$) function. The graphical representation of this is shown in Fig. \ref{fig:act}(c).
\item Mish is an improvement on the existing swish function; it is a smooth and non-monotonic function that can be defined as:
\begin{equation}
f(x)=x\times \tanh (\ln (1+e^x)
\end{equation}
The graphical representation of this is shown in Fig. \ref{fig:act}(d). Since Mish and its predecessor are visibly indistinguishable, a comparison of the two is shown in Fig. \ref{fig:actcomp} that displays the negligible difference between the two.
\end{enumerate}
\begin{figure}[tb!]
\centering
\begin{subfigure}[]{0.4\linewidth}
\centering
\includegraphics[trim={0 0.85cm 4cm 5cm},clip,width=1.2in, height=2in]{Images/ReLU.png}\hfill
\caption{\small{ReLu}}
\end{subfigure}
\begin{subfigure}[]{0.4\linewidth}
\centering
\includegraphics[width=1.2in, height=2in]{Images/Lrelu1.PNG}\hfill
\caption{\small{Leaky Relu}}
\end{subfigure}\\[\smallskipamount]
\begin{subfigure}[t]{0.4\linewidth}
\centering
\includegraphics[trim={0 0 4cm 5cm},clip,width=1.2in, height=2in]{Images/swish.PNG}\hfill
\caption{\small{Swish}}
\end{subfigure}
\begin{subfigure}[t]{0.4\linewidth}
\centering
\includegraphics[trim={0 0.3cm 4cm 5cm},clip,width=1.2in, height=2in]{Images/Mish.PNG}\hfill
\caption{\small{Mish}}
\end{subfigure}
\caption{\small{Graphical form of the activation functions \cite{serengil_2020}.}}
\label{fig:act}
\end{figure}
\begin{figure}[t]{}
\centering
\includegraphics[scale=0.35]{Images/mish-swish.PNG}\hfill
\caption{\small{Mish and Swish comparison for understanding how similar they are. Blue line represent Swish and red represents Mish \cite{serengil_2020}.}}
\label{fig:actcomp}
\end{figure}
\section{Testing and Analysis}
\begin{table}[tb]
\centering
\caption{\small{Comparison of results with variation in algorithm versions}}
\begin{tabular}{|c|c|c|c|c|}\hline
\textbf{Image} & \textbf{YOLOv3} & \textbf{YOLOv4-1} & \textbf{YOLOv4-2} & \textbf{YOLOv4-3} \\ \hline
\ref{fig:tpapred}(a) & 20.537\degree & 22.67\degree & 24.8\degree & 22.9\degree\\ \hline
\ref{fig:tpapred}(b) & 17.354\degree & 17.62\degree & 18.7\degree & 16.22\degree\\ \hline
\ref{fig:tpapred}(c) & 19.473\degree & 19.53\degree & 20.03\degree & 18.24\degree\\ \hline
\ref{fig:tpapred}(d) & 23.369\degree & 27.37\degree & 29.03\degree & 25.8\degree\\ \hline
\ref{fig:tpapred}(e) & 18.435\degree & 15.5\degree & 17.65\degree & 16.8\degree\\ \hline
\ref{fig:tpapred}(f) & 19.699\degree & 18.24\degree & 20.22\degree & 18.59\degree\\ \hline
\end{tabular}
\label{table:comp1}
\end{table}
\noindent For comparison purposes the radiographs have been tested with YOLOv3, original YOLOv4 \cite{bochkovskiy2020yolov4} and custom modifications of the YOLOv4 by changing the activation functions and the results are shown in Table \ref{table:comp1}. The original version of YOLOv4 (addressed as YOLOv4-1 in Table \ref{table:comp1}) combined the Mish, Linear and Leaky activation functions, the versions YOLOv4-2 and YOLOv4-3 used for the purpose of testing in this paper have combinations of Mish, Linear, Swish and Mish, Linear and Relu activation functions respectively. Similarly, comparison of the results that did not fall under the presumed normal range shown in Fig. \ref{fig:tpapredwrng}, for the algorithm, are shown with variations in activation function in Table \ref{table:comp2}. It can be seen here that these images, are giving TPA predictions similar to the original YOLOv3, i.e. outside of presumed range.
\begin{table}[tb]
\centering
\caption{\small{Comparison of `below range' results with variation in algorithm versions}}\label{table:comp2}
\begin{tabular}{|c|c|c|c|c|}\hline
\textbf{Image} & \textbf{YOLOv3} & \textbf{YOLOv4-1} & \textbf{YOLOv4-2} & \textbf{YOLOv4-3} \\ \hline
\ref{fig:tpapredwrng}(a) & 10.4\degree & 11.22\degree & 10.48\degree & 14.91\degree\\ \hline
\ref{fig:tpapredwrng}(b) & 10.3\degree & 8.86\degree & 11.3\degree & 11.11\degree\\ \hline
\ref{fig:tpapredwrng}(c) & 6.53\degree & 8.84\degree & 7.07\degree & 8.18\degree\\ \hline
\end{tabular}
\end{table}
\section{Conclusion and Future Work}
\noindent Automated image annotation has become a major requirement in the medical field, since it can be a great tool to drive quick, intelligent and reliable patient care decisions \cite{grady2020system}. Grady and Schaap \cite{grady2020system}, patented the idea of incorporating user input as part of the learning process, as it is essential that the automated results are validated and appropriately corrected by a user, i.e., a radiographer, when required.
\indent Results that do not fall under the range mentioned by \cite{seo2020measurement}, are shown in Fig. \ref{fig:tpapredwrng} and Table \ref{table:comp2}. These open a possible scope for improvement and future work as these samples might require human expert-intervention in order to correct the annotations and relearn from the said corrections.
\begin{figure}[t]
\centering
\begin{subfigure}[t]{0.26\linewidth}
\centering
\includegraphics[width=0.8in, height=2in]{test_results/TR2.PNG}\hfill
\caption{\small{TPA=10.4\degree}}
\end{subfigure}
\begin{subfigure}[t]{0.26\linewidth}
\centering
\includegraphics[width=0.8in, height=2in]{test_results/TR3.PNG}\hfill
\caption{\small{TPA=10.3\degree}}
\end{subfigure}
\begin{subfigure}[t]{0.26\linewidth}
\centering
\includegraphics[width=0.8in, height=2in]{test_results/TR6.PNG}\hfill
\caption{\small{TPA=6.53\degree}}
\end{subfigure}
\caption{\small{Example of algorithm detecting, highlighting the regions of interests that are below known range of value.}}\label{fig:tpapredwrng}
\end{figure}
\indent For future work, a user based feedback loop will be added to this system that may be used alongside these predictions, as input to train the system further which will result in more accurate TPA calculation. Another degree to this work could be automating this angle value calculation, i.e., using the TPA values as part of the information fed to the system so that it may be able to draw the lines along the bone joint-axes, and calculate the TPA, all within the same deep neural network; simplifying the user interface to just inserting the X-ray and, potentially resulting in a more accurate TPA calculation.\\
\indent The core finding of this paper is that, even without expert intervention, automation of annotations was successfully performed to a significantly accurate degree, 4 out of every 5 images tested in average. With this in mind, in addition to the previously developed lightweight classifier \cite{Tonima_2021}, it can be concluded that the second step of our development, of the automation in diagnostic tool, is complete. This paper confirms that it is possible to automate the system via annotation which can be improved with the formerly mentioned expert feedback in the future.
\section*{Acknowledgment}
\noindent The dataset used in this paper is provided by Innotech Medical Industries Xray. The research work is financially supported in part by a MITACS Accelerate Project (no. FR56849) under the partner organization Innotech Medical Industries Corp. and the Natural Sciences and Engineering Research Council of Canada (NSERC).
\bibliographystyle{unsrt}
|
\subsubsection*{\bibname}}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{url}
\usepackage{booktabs}
\usepackage{amsfonts}
\usepackage{nicefrac}
\usepackage{microtype}
\usepackage{soul}
\usepackage{natbib}
\usepackage{graphicx}
\usepackage{subfigure}
\usepackage{booktabs}
\usepackage{amssymb
\usepackage{pifont
\newcommand{\cmark}{\color{green}\ding{51}}%
\newcommand{\xmark}{\color{red}\ding{55}}%
\usepackage{amsthm}
\usepackage{amsmath}
\usepackage{amsfonts}
\usepackage{bbm}
\usepackage{color}
\usepackage{graphicx}
\usepackage{enumitem}
\usepackage{subfigure}
\usepackage{wrapfig}
\usepackage{svg}
\usepackage{tikz}
\def\checkmark{\tikz\fill[scale=0.4](0,.35) -- (.25,0) -- (1,.7) -- (.25,.15) -- cycle;}
\usepackage{amssymb}
\newcommand{\mathbf{X}}{\mathbf{X}}
\newcommand{\mathbf{Y}}{\mathbf{Y}}
\newcommand{\mathbf{A}}{\mathbf{A}}
\newcommand{\mathbf{B}}{\mathbf{B}}
\newcommand{\mathbf{P}}{\mathbf{P}}
\newcommand{\mathbf{S}}{\mathbf{S}}
\newcommand{\mathbf{Q}}{\mathbf{Q}}
\newcommand{\mathbf{y}}{\mathbf{y}}
\newcommand{\mathbf{p}}{\mathbf{p}}
\newcommand{\mathbf{d}}{\mathbf{d}}
\newcommand{\mathbf{b}}{\mathbf{b}}
\newcommand{\mathbf{e}}{\mathbf{e}}
\newcommand{\mathbf{a}}{\mathbf{a}}
\newcommand{\mathbf{C}}{\mathbf{C}}
\newcommand{\mathbf{w}}{\mathbf{w}}
\newcommand{\mathbf{c}}{\mathbf{c}}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{{\rm Cov}}{{\rm Cov}}
\newcommand{{\rm Var}}{{\rm Var}}
\newcommand{\mathbb{P}}{\mathbb{P}}
\newcommand{\mathbf{x}}{\mathbf{x}}
\newcommand{\mathbf{z}}{\mathbf{z}}
\newcommand{\mathbf{Z}}{\mathbf{Z}}
\newcommand{\mathbf{s}}{\mathbf{s}}
\newcommand{\mathbf{v}}{\mathbf{v}}
\newcommand{\mathbf{u}}{\mathbf{u}}
\newcommand{\boldsymbol{\theta}}{\boldsymbol{\theta}}
\newcommand{\widehat{\boldsymbol{\theta}}}{\widehat{\boldsymbol{\theta}}}
\newcommand{R\'enyi }{R\'enyi }
\DeclareMathOperator{\vect}{vec}
\newcommand{\andy}[1]{{\textbf{\color{violet}andy: #1}}}
\newcommand{\meisam}[1]{{\textbf{\color{red}meisam: #1}}}
\newcommand{\ahmad}[1]{{\textbf{\color{blue}ahmad: #1}}}
\newcommand{\sina}[1]{{\textbf{\color{green}sina: #1}}}
\newcommand{\edit}[1]{{\color{violet} #1}}
\usepackage{amsmath,amsfonts,bm}
\newcommand{{\em (Left)}}{{\em (Left)}}
\newcommand{{\em (Center)}}{{\em (Center)}}
\newcommand{{\em (Right)}}{{\em (Right)}}
\newcommand{{\em (Top)}}{{\em (Top)}}
\newcommand{{\em (Bottom)}}{{\em (Bottom)}}
\newcommand{{\em (a)}}{{\em (a)}}
\newcommand{{\em (b)}}{{\em (b)}}
\newcommand{{\em (c)}}{{\em (c)}}
\newcommand{{\em (d)}}{{\em (d)}}
\newcommand{\newterm}[1]{{\bf #1}}
\def\figref#1{figure~\ref{#1}}
\def\Figref#1{Figure~\ref{#1}}
\def\twofigref#1#2{figures \ref{#1} and \ref{#2}}
\def\quadfigref#1#2#3#4{figures \ref{#1}, \ref{#2}, \ref{#3} and \ref{#4}}
\def\secref#1{section~\ref{#1}}
\def\Secref#1{Section~\ref{#1}}
\def\twosecrefs#1#2{sections \ref{#1} and \ref{#2}}
\def\secrefs#1#2#3{sections \ref{#1}, \ref{#2} and \ref{#3}}
\def\eqref#1{equation~\ref{#1}}
\def\Eqref#1{Equation~\ref{#1}}
\def\plaineqref#1{\ref{#1}}
\def\chapref#1{chapter~\ref{#1}}
\def\Chapref#1{Chapter~\ref{#1}}
\def\rangechapref#1#2{chapters\ref{#1}--\ref{#2}}
\def\algref#1{algorithm~\ref{#1}}
\def\Algref#1{Algorithm~\ref{#1}}
\def\twoalgref#1#2{algorithms \ref{#1} and \ref{#2}}
\def\Twoalgref#1#2{Algorithms \ref{#1} and \ref{#2}}
\def\partref#1{part~\ref{#1}}
\def\Partref#1{Part~\ref{#1}}
\def\twopartref#1#2{parts \ref{#1} and \ref{#2}}
\def\ceil#1{\lceil #1 \rceil}
\def\floor#1{\lfloor #1 \rfloor}
\def\bm{1}{\bm{1}}
\newcommand{\mathcal{D}}{\mathcal{D}}
\newcommand{\mathcal{D_{\mathrm{valid}}}}{\mathcal{D_{\mathrm{valid}}}}
\newcommand{\mathcal{D_{\mathrm{test}}}}{\mathcal{D_{\mathrm{test}}}}
\def{\epsilon}{{\epsilon}}
\def{\textnormal{$\eta$}}{{\textnormal{$\eta$}}}
\def{\textnormal{a}}{{\textnormal{a}}}
\def{\textnormal{b}}{{\textnormal{b}}}
\def{\textnormal{c}}{{\textnormal{c}}}
\def{\textnormal{d}}{{\textnormal{d}}}
\def{\textnormal{e}}{{\textnormal{e}}}
\def{\textnormal{f}}{{\textnormal{f}}}
\def{\textnormal{g}}{{\textnormal{g}}}
\def{\textnormal{h}}{{\textnormal{h}}}
\def{\textnormal{i}}{{\textnormal{i}}}
\def{\textnormal{j}}{{\textnormal{j}}}
\def{\textnormal{k}}{{\textnormal{k}}}
\def{\textnormal{l}}{{\textnormal{l}}}
\def{\textnormal{n}}{{\textnormal{n}}}
\def{\textnormal{o}}{{\textnormal{o}}}
\def{\textnormal{p}}{{\textnormal{p}}}
\def{\textnormal{q}}{{\textnormal{q}}}
\def{\textnormal{r}}{{\textnormal{r}}}
\def{\textnormal{s}}{{\textnormal{s}}}
\def{\textnormal{t}}{{\textnormal{t}}}
\def{\textnormal{u}}{{\textnormal{u}}}
\def{\textnormal{v}}{{\textnormal{v}}}
\def{\textnormal{w}}{{\textnormal{w}}}
\def{\textnormal{x}}{{\textnormal{x}}}
\def{\textnormal{y}}{{\textnormal{y}}}
\def{\textnormal{z}}{{\textnormal{z}}}
\def{\mathbf{\epsilon}}{{\mathbf{\epsilon}}}
\def{\mathbf{\theta}}{{\mathbf{\theta}}}
\def{\mathbf{a}}{{\mathbf{a}}}
\def{\mathbf{b}}{{\mathbf{b}}}
\def{\mathbf{c}}{{\mathbf{c}}}
\def{\mathbf{d}}{{\mathbf{d}}}
\def{\mathbf{e}}{{\mathbf{e}}}
\def{\mathbf{f}}{{\mathbf{f}}}
\def{\mathbf{g}}{{\mathbf{g}}}
\def{\mathbf{h}}{{\mathbf{h}}}
\def{\mathbf{u}}{{\mathbf{i}}}
\def{\mathbf{j}}{{\mathbf{j}}}
\def{\mathbf{k}}{{\mathbf{k}}}
\def{\mathbf{l}}{{\mathbf{l}}}
\def{\mathbf{m}}{{\mathbf{m}}}
\def{\mathbf{n}}{{\mathbf{n}}}
\def{\mathbf{o}}{{\mathbf{o}}}
\def{\mathbf{p}}{{\mathbf{p}}}
\def{\mathbf{q}}{{\mathbf{q}}}
\def{\mathbf{r}}{{\mathbf{r}}}
\def{\mathbf{s}}{{\mathbf{s}}}
\def{\mathbf{t}}{{\mathbf{t}}}
\def{\mathbf{u}}{{\mathbf{u}}}
\def{\mathbf{v}}{{\mathbf{v}}}
\def{\mathbf{w}}{{\mathbf{w}}}
\def{\mathbf{x}}{{\mathbf{x}}}
\def{\mathbf{y}}{{\mathbf{y}}}
\def{\mathbf{z}}{{\mathbf{z}}}
\def{\textnormal{a}}{{\textnormal{a}}}
\def{\textnormal{b}}{{\textnormal{b}}}
\def{\textnormal{c}}{{\textnormal{c}}}
\def{\textnormal{d}}{{\textnormal{d}}}
\def{\textnormal{e}}{{\textnormal{e}}}
\def{\textnormal{f}}{{\textnormal{f}}}
\def{\textnormal{g}}{{\textnormal{g}}}
\def{\textnormal{h}}{{\textnormal{h}}}
\def{\textnormal{i}}{{\textnormal{i}}}
\def{\textnormal{j}}{{\textnormal{j}}}
\def{\textnormal{k}}{{\textnormal{k}}}
\def{\textnormal{l}}{{\textnormal{l}}}
\def{\textnormal{m}}{{\textnormal{m}}}
\def{\textnormal{n}}{{\textnormal{n}}}
\def{\textnormal{o}}{{\textnormal{o}}}
\def{\textnormal{p}}{{\textnormal{p}}}
\def{\textnormal{q}}{{\textnormal{q}}}
\def{\textnormal{r}}{{\textnormal{r}}}
\def{\textnormal{s}}{{\textnormal{s}}}
\def{\textnormal{t}}{{\textnormal{t}}}
\def{\textnormal{u}}{{\textnormal{u}}}
\def{\textnormal{v}}{{\textnormal{v}}}
\def{\textnormal{w}}{{\textnormal{w}}}
\def{\textnormal{x}}{{\textnormal{x}}}
\def{\textnormal{y}}{{\textnormal{y}}}
\def{\textnormal{z}}{{\textnormal{z}}}
\def{\mathbf{A}}{{\mathbf{A}}}
\def{\mathbf{B}}{{\mathbf{B}}}
\def{\mathbf{C}}{{\mathbf{C}}}
\def{\mathbf{D}}{{\mathbf{D}}}
\def{\mathbf{E}}{{\mathbf{E}}}
\def{\mathbf{F}}{{\mathbf{F}}}
\def{\mathbf{G}}{{\mathbf{G}}}
\def{\mathbf{H}}{{\mathbf{H}}}
\def{\mathbf{I}}{{\mathbf{I}}}
\def{\mathbf{J}}{{\mathbf{J}}}
\def{\mathbf{K}}{{\mathbf{K}}}
\def{\mathbf{L}}{{\mathbf{L}}}
\def{\mathbf{M}}{{\mathbf{M}}}
\def{\mathbf{N}}{{\mathbf{N}}}
\def{\mathbf{O}}{{\mathbf{O}}}
\def{\mathbf{P}}{{\mathbf{P}}}
\def{\mathbf{Q}}{{\mathbf{Q}}}
\def{\mathbf{R}}{{\mathbf{R}}}
\def{\mathbf{S}}{{\mathbf{S}}}
\def{\mathbf{T}}{{\mathbf{T}}}
\def{\mathbf{U}}{{\mathbf{U}}}
\def{\mathbf{V}}{{\mathbf{V}}}
\def{\mathbf{W}}{{\mathbf{W}}}
\def{\mathbf{X}}{{\mathbf{X}}}
\def{\mathbf{Y}}{{\mathbf{Y}}}
\def{\mathbf{Z}}{{\mathbf{Z}}}
\def{\textnormal{A}}{{\textnormal{A}}}
\def{\textnormal{B}}{{\textnormal{B}}}
\def{\textnormal{C}}{{\textnormal{C}}}
\def{\textnormal{D}}{{\textnormal{D}}}
\def{\textnormal{E}}{{\textnormal{E}}}
\def{\textnormal{F}}{{\textnormal{F}}}
\def{\textnormal{G}}{{\textnormal{G}}}
\def{\textnormal{H}}{{\textnormal{H}}}
\def{\textnormal{I}}{{\textnormal{I}}}
\def{\textnormal{J}}{{\textnormal{J}}}
\def{\textnormal{K}}{{\textnormal{K}}}
\def{\textnormal{L}}{{\textnormal{L}}}
\def{\textnormal{M}}{{\textnormal{M}}}
\def{\textnormal{N}}{{\textnormal{N}}}
\def{\textnormal{O}}{{\textnormal{O}}}
\def{\textnormal{P}}{{\textnormal{P}}}
\def{\textnormal{Q}}{{\textnormal{Q}}}
\def{\textnormal{R}}{{\textnormal{R}}}
\def{\textnormal{S}}{{\textnormal{S}}}
\def{\textnormal{T}}{{\textnormal{T}}}
\def{\textnormal{U}}{{\textnormal{U}}}
\def{\textnormal{V}}{{\textnormal{V}}}
\def{\textnormal{W}}{{\textnormal{W}}}
\def{\textnormal{X}}{{\textnormal{X}}}
\def{\textnormal{Y}}{{\textnormal{Y}}}
\def{\textnormal{Z}}{{\textnormal{Z}}}
\def{\bm{0}}{{\bm{0}}}
\def{\bm{1}}{{\bm{1}}}
\def{\bm{\mu}}{{\bm{\mu}}}
\def{\bm{\theta}}{{\bm{\theta}}}
\def{\bm{a}}{{\bm{a}}}
\def{\bm{b}}{{\bm{b}}}
\def{\bm{c}}{{\bm{c}}}
\def{\bm{d}}{{\bm{d}}}
\def{\bm{e}}{{\bm{e}}}
\def{\bm{f}}{{\bm{f}}}
\def{\bm{g}}{{\bm{g}}}
\def{\bm{h}}{{\bm{h}}}
\def{\bm{i}}{{\bm{i}}}
\def{\bm{j}}{{\bm{j}}}
\def{\bm{k}}{{\bm{k}}}
\def{\bm{l}}{{\bm{l}}}
\def{\bm{m}}{{\bm{m}}}
\def{\bm{n}}{{\bm{n}}}
\def{\bm{o}}{{\bm{o}}}
\def{\bm{p}}{{\bm{p}}}
\def{\bm{q}}{{\bm{q}}}
\def{\bm{r}}{{\bm{r}}}
\def{\bm{s}}{{\bm{s}}}
\def{\bm{t}}{{\bm{t}}}
\def{\bm{u}}{{\bm{u}}}
\def{\bm{v}}{{\bm{v}}}
\def{\bm{w}}{{\bm{w}}}
\def{\bm{x}}{{\bm{x}}}
\def{\bm{y}}{{\bm{y}}}
\def{\bm{z}}{{\bm{z}}}
\def{\alpha}{{\alpha}}
\def{\beta}{{\beta}}
\def{\epsilon}{{\epsilon}}
\def{\lambda}{{\lambda}}
\def{\omega}{{\omega}}
\def{\mu}{{\mu}}
\def{\psi}{{\psi}}
\def{\sigma}{{\sigma}}
\def{\theta}{{\theta}}
\def{a}{{a}}
\def{b}{{b}}
\def{c}{{c}}
\def{d}{{d}}
\def{e}{{e}}
\def{f}{{f}}
\def{g}{{g}}
\def{h}{{h}}
\def{i}{{i}}
\def{j}{{j}}
\def{k}{{k}}
\def{l}{{l}}
\def{m}{{m}}
\def{n}{{n}}
\def{o}{{o}}
\def{p}{{p}}
\def{q}{{q}}
\def{r}{{r}}
\def{s}{{s}}
\def{t}{{t}}
\def{u}{{u}}
\def{v}{{v}}
\def{w}{{w}}
\def{x}{{x}}
\def{y}{{y}}
\def{z}{{z}}
\def{\bm{A}}{{\bm{A}}}
\def{\bm{B}}{{\bm{B}}}
\def{\bm{C}}{{\bm{C}}}
\def{\bm{D}}{{\bm{D}}}
\def{\bm{E}}{{\bm{E}}}
\def{\bm{F}}{{\bm{F}}}
\def{\bm{G}}{{\bm{G}}}
\def{\bm{H}}{{\bm{H}}}
\def{\bm{I}}{{\bm{I}}}
\def{\bm{J}}{{\bm{J}}}
\def{\bm{K}}{{\bm{K}}}
\def{\bm{L}}{{\bm{L}}}
\def{\bm{M}}{{\bm{M}}}
\def{\bm{N}}{{\bm{N}}}
\def{\bm{O}}{{\bm{O}}}
\def{\bm{P}}{{\bm{P}}}
\def{\bm{Q}}{{\bm{Q}}}
\def{\bm{R}}{{\bm{R}}}
\def{\bm{S}}{{\bm{S}}}
\def{\bm{T}}{{\bm{T}}}
\def{\bm{U}}{{\bm{U}}}
\def{\bm{V}}{{\bm{V}}}
\def{\bm{W}}{{\bm{W}}}
\def{\bm{X}}{{\bm{X}}}
\def{\bm{Y}}{{\bm{Y}}}
\def{\bm{Z}}{{\bm{Z}}}
\def{\bm{\beta}}{{\bm{\beta}}}
\def{\bm{\Phi}}{{\bm{\Phi}}}
\def{\bm{\Lambda}}{{\bm{\Lambda}}}
\def{\bm{\Sigma}}{{\bm{\Sigma}}}
\DeclareMathAlphabet{\mathsfit}{\encodingdefault}{\sfdefault}{m}{sl}
\SetMathAlphabet{\mathsfit}{bold}{\encodingdefault}{\sfdefault}{bx}{n}
\newcommand{\tens}[1]{\bm{\mathsfit{#1}}}
\def{\tens{A}}{{\tens{A}}}
\def{\tens{B}}{{\tens{B}}}
\def{\tens{C}}{{\tens{C}}}
\def{\tens{D}}{{\tens{D}}}
\def{\tens{E}}{{\tens{E}}}
\def{\tens{F}}{{\tens{F}}}
\def{\tens{G}}{{\tens{G}}}
\def{\tens{H}}{{\tens{H}}}
\def{\tens{I}}{{\tens{I}}}
\def{\tens{J}}{{\tens{J}}}
\def{\tens{K}}{{\tens{K}}}
\def{\tens{L}}{{\tens{L}}}
\def{\tens{M}}{{\tens{M}}}
\def{\tens{N}}{{\tens{N}}}
\def{\tens{O}}{{\tens{O}}}
\def{\tens{P}}{{\tens{P}}}
\def{\tens{Q}}{{\tens{Q}}}
\def{\tens{R}}{{\tens{R}}}
\def{\tens{S}}{{\tens{S}}}
\def{\tens{T}}{{\tens{T}}}
\def{\tens{U}}{{\tens{U}}}
\def{\tens{V}}{{\tens{V}}}
\def{\tens{W}}{{\tens{W}}}
\def{\tens{X}}{{\tens{X}}}
\def{\tens{Y}}{{\tens{Y}}}
\def{\tens{Z}}{{\tens{Z}}}
\def{\mathcal{A}}{{\mathcal{A}}}
\def{\mathcal{B}}{{\mathcal{B}}}
\def{\mathcal{C}}{{\mathcal{C}}}
\def{\mathcal{D}}{{\mathcal{D}}}
\def{\mathcal{E}}{{\mathcal{E}}}
\def{\mathcal{F}}{{\mathcal{F}}}
\def{\mathcal{G}}{{\mathcal{G}}}
\def{\mathcal{H}}{{\mathcal{H}}}
\def{\mathcal{I}}{{\mathcal{I}}}
\def{\mathcal{J}}{{\mathcal{J}}}
\def{\mathcal{K}}{{\mathcal{K}}}
\def{\mathcal{L}}{{\mathcal{L}}}
\def{\mathcal{M}}{{\mathcal{M}}}
\def{\mathcal{N}}{{\mathcal{N}}}
\def{\mathcal{O}}{{\mathcal{O}}}
\def{\mathcal{P}}{{\mathcal{P}}}
\def{\mathcal{Q}}{{\mathcal{Q}}}
\def{\mathcal{R}}{{\mathcal{R}}}
\def{\mathcal{S}}{{\mathcal{S}}}
\def{\mathcal{T}}{{\mathcal{T}}}
\def{\mathcal{U}}{{\mathcal{U}}}
\def{\mathcal{V}}{{\mathcal{V}}}
\def{\mathcal{W}}{{\mathcal{W}}}
\def{\mathcal{X}}{{\mathcal{X}}}
\def{\mathcal{Y}}{{\mathcal{Y}}}
\def{\mathcal{Z}}{{\mathcal{Z}}}
\def{\mathbb{A}}{{\mathbb{A}}}
\def{\mathbb{B}}{{\mathbb{B}}}
\def{\mathbb{C}}{{\mathbb{C}}}
\def{\mathbb{D}}{{\mathbb{D}}}
\def{\mathbb{F}}{{\mathbb{F}}}
\def{\mathbb{G}}{{\mathbb{G}}}
\def{\mathbb{H}}{{\mathbb{H}}}
\def{\mathbb{I}}{{\mathbb{I}}}
\def{\mathbb{J}}{{\mathbb{J}}}
\def{\mathbb{K}}{{\mathbb{K}}}
\def{\mathbb{L}}{{\mathbb{L}}}
\def{\mathbb{M}}{{\mathbb{M}}}
\def{\mathbb{N}}{{\mathbb{N}}}
\def{\mathbb{O}}{{\mathbb{O}}}
\def{\mathbb{P}}{{\mathbb{P}}}
\def{\mathbb{Q}}{{\mathbb{Q}}}
\def{\mathbb{R}}{{\mathbb{R}}}
\def{\mathbb{S}}{{\mathbb{S}}}
\def{\mathbb{T}}{{\mathbb{T}}}
\def{\mathbb{U}}{{\mathbb{U}}}
\def{\mathbb{V}}{{\mathbb{V}}}
\def{\mathbb{W}}{{\mathbb{W}}}
\def{\mathbb{X}}{{\mathbb{X}}}
\def{\mathbb{Y}}{{\mathbb{Y}}}
\def{\mathbb{Z}}{{\mathbb{Z}}}
\def{\Lambda}{{\Lambda}}
\def{A}{{A}}
\def{B}{{B}}
\def{C}{{C}}
\def{D}{{D}}
\def{E}{{E}}
\def{F}{{F}}
\def{G}{{G}}
\def{H}{{H}}
\def{I}{{I}}
\def{J}{{J}}
\def{K}{{K}}
\def{L}{{L}}
\def{M}{{M}}
\def{N}{{N}}
\def{O}{{O}}
\def{P}{{P}}
\def{Q}{{Q}}
\def{R}{{R}}
\def{S}{{S}}
\def{T}{{T}}
\def{U}{{U}}
\def{V}{{V}}
\def{W}{{W}}
\def{X}{{X}}
\def{Y}{{Y}}
\def{Z}{{Z}}
\def{\Sigma}{{\Sigma}}
\newcommand{\etens}[1]{\mathsfit{#1}}
\def{\etens{\Lambda}}{{\etens{\Lambda}}}
\def{\etens{A}}{{\etens{A}}}
\def{\etens{B}}{{\etens{B}}}
\def{\etens{C}}{{\etens{C}}}
\def{\etens{D}}{{\etens{D}}}
\def{\etens{E}}{{\etens{E}}}
\def{\etens{F}}{{\etens{F}}}
\def{\etens{G}}{{\etens{G}}}
\def{\etens{H}}{{\etens{H}}}
\def{\etens{I}}{{\etens{I}}}
\def{\etens{J}}{{\etens{J}}}
\def{\etens{K}}{{\etens{K}}}
\def{\etens{L}}{{\etens{L}}}
\def{\etens{M}}{{\etens{M}}}
\def{\etens{N}}{{\etens{N}}}
\def{\etens{O}}{{\etens{O}}}
\def{\etens{P}}{{\etens{P}}}
\def{\etens{Q}}{{\etens{Q}}}
\def{\etens{R}}{{\etens{R}}}
\def{\etens{S}}{{\etens{S}}}
\def{\etens{T}}{{\etens{T}}}
\def{\etens{U}}{{\etens{U}}}
\def{\etens{V}}{{\etens{V}}}
\def{\etens{W}}{{\etens{W}}}
\def{\etens{X}}{{\etens{X}}}
\def{\etens{Y}}{{\etens{Y}}}
\def{\etens{Z}}{{\etens{Z}}}
\newcommand{p_{\rm{data}}}{p_{\rm{data}}}
\newcommand{\hat{p}_{\rm{data}}}{\hat{p}_{\rm{data}}}
\newcommand{\hat{P}_{\rm{data}}}{\hat{P}_{\rm{data}}}
\newcommand{p_{\rm{model}}}{p_{\rm{model}}}
\newcommand{P_{\rm{model}}}{P_{\rm{model}}}
\newcommand{\tilde{p}_{\rm{model}}}{\tilde{p}_{\rm{model}}}
\newcommand{p_{\rm{encoder}}}{p_{\rm{encoder}}}
\newcommand{p_{\rm{decoder}}}{p_{\rm{decoder}}}
\newcommand{p_{\rm{reconstruct}}}{p_{\rm{reconstruct}}}
\newcommand{\laplace}{\mathrm{Laplace}}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{\mathcal{L}}{\mathcal{L}}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\tilde{p}}{\tilde{p}}
\newcommand{\alpha}{\alpha}
\newcommand{\lambda}{\lambda}
\newcommand{\mathrm{rectifier}}{\mathrm{rectifier}}
\newcommand{\mathrm{softmax}}{\mathrm{softmax}}
\newcommand{\sigma}{\sigma}
\newcommand{\zeta}{\zeta}
\newcommand{D_{\mathrm{KL}}}{D_{\mathrm{KL}}}
\newcommand{\mathrm{Var}}{\mathrm{Var}}
\newcommand{\mathrm{SE}}{\mathrm{SE}}
\newcommand{\mathrm{Cov}}{\mathrm{Cov}}
\newcommand{L^0}{L^0}
\newcommand{L^1}{L^1}
\newcommand{L^2}{L^2}
\newcommand{L^p}{L^p}
\newcommand{L^\infty}{L^\infty}
\newcommand{\parents}{Pa}
\newcommand{\mathbb{P}}{\mathbb{P}}
\newcommand{\mathcal{S}}{\mathcal{S}}
\newcommand{\mathcal{X}}{\mathcal{X}}
\newcommand{\mathcal{Z}}{\mathcal{Z}}
\newcommand{\mathcal{R}}{\mathcal{R}}
\newcommand{\text{dp}(\widehat{Y},S)}{\text{dp}(\widehat{Y},S)}
\newcommand{\text{eo}}{\text{eo}}
\newcommand{\eod(\widehat{Y}, S|Y \in \mathcal{A})}{\text{eo}(\widehat{Y}, S|Y \in \mathcal{A})}
\newcommand{\widetilde{\text{eo}}(\widehat{Y}|S, Y \in \mathcal{A})}{\widetilde{\text{eo}}(\widehat{Y}|S, Y \in \mathcal{A})}
\newcommand{\widetilde{\text{eo}}}{\widetilde{\text{eo}}}
\newcommand{\widehat{Y}}{\widehat{Y}}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{p_{\hy}}{p_{\widehat{Y}}}
\newcommand{p_{S}}{p_{S}}
\newcommand{p_{(\hy, S)}}{p_{(\widehat{Y}, S)}}
\newcommand{P_{\hat{y}}}{P_{\hat{y}}}
\newcommand{P_{s}}{P_{s}}
\newcommand{P_{\hat{y}, s}}{P_{\hat{y}, s}}
\newcommand{p_{\hy , S}}{p_{\widehat{Y} , S}}
\newcommand{p_{S|Y}}{p_{S|Y}}
\newcommand{p_{\hy , S| Y}}{p_{\widehat{Y} , S| Y}}
\newcommand{p_{\hy | Y}}{p_{\widehat{Y} | Y}}
\newcommand{e_{\hat{y},s}}{e_{\hat{y},s}}
\newcommand{u_{\hat{y},s}}{u_{\hat{y},s}}
\newcommand{p_{\hat{y},s}}{p_{\hat{y},s}}
\newcommand{p_{(\hy, S | Y)}}{p_{(\widehat{Y}, S | Y)}}
\newcommand{p_{\hy | Y}}{p_{\widehat{Y} | Y}}
\newcommand{p_{S | Y}}{p_{S | Y}}
\newcommand{\mathcal{Y}}{\mathcal{Y}}
\newcommand{\hat{y}}{\hat{y}}
\newcommand{\mathcal{F}(\btheta, \bX)}{\mathcal{F}(\boldsymbol{\theta}, \mathbf{X})}
\newcommand{\mathcal{F}(\btheta^t, \bX_n)}{\mathcal{F}(\boldsymbol{\theta}^t, \mathbf{X}_n)}
\newcommand{\mathcal{W}}{\mathcal{W}}
\newcommand{\Delta_{\Phi}}{\Delta_{\Phi}}
\newcommand{{\cal L}\big({\cal F}(\btheta, \bX_{i}), y_{i}\big)}{{\cal L}\big({\cal F}(\boldsymbol{\theta}, \mathbf{X}_{i}), y_{i}\big)}
\newcommand{{\cal L}\big({\cal F}(\btheta^t, \bX_{n}), y_{n}\big)}{{\cal L}\big({\cal F}(\boldsymbol{\theta}^t, \mathbf{X}_{n}), y_{n}\big)}
\newcommand{\beta_{\mathcal{F}}}{\beta_{\mathcal{F}}}
\newcommand{L_{\mathcal{F}}}{L_{\mathcal{F}}}
\newcommand{\beta_{\mathcal{L}}}{\beta_{\mathcal{L}}}
\newcommand{L_{\mathcal{L}}}{L_{\mathcal{L}}}
\newcommand{\expec[\mathbf{\widehat{y}}(\bx_i, \btheta)\mathbf{\widehat{y}}(\bx_i, \btheta)^T | \bx_i]}{\mathbb{E}[\mathbf{\widehat{y}}(\mathbf{x}_i, \boldsymbol{\theta})\mathbf{\widehat{y}}(\mathbf{x}_i, \boldsymbol{\theta})^T | \mathbf{x}_i]}
\newcommand{\expec[\mathbf{s}_i \mathbf{\widehat{y}}(\bx_i, \btheta)^T | \bx_i, s_i]}{\mathbb{E}[\mathbf{s}_i \mathbf{\widehat{y}}(\mathbf{x}_i, \boldsymbol{\theta})^T | \mathbf{x}_i, s_i]}
\newcommand{\widehat{P}_{\haty}}{\widehat{P}_{\hat{y}}}
\newcommand{\widehat{P}_{s}}{\widehat{P}_{s}}
\newcommand{\widehat{P}_{\haty, s}}{\widehat{P}_{\hat{y}, s}}
\newcommand{\nabla_{\btheta}}{\nabla_{\boldsymbol{\theta}}}
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator*{\argmin}{arg\,min}
\DeclareMathOperator{\sign}{sign}
\DeclareMathOperator{\Tr}{Tr}
\let\ab\allowbreak
\usepackage{mathrsfs}
\usepackage{enumitem}
\usepackage{algorithm}
\usepackage{algcompatible}
\newtheorem{theorem}{Theorem}
\newtheorem{definition}{Definition}
\newtheorem{corollary}{Corollary}
\newtheorem{lemma}{Lemma}
\newtheorem{remark}{Remark}
\newtheorem{assumption}{Assumption}
\newtheorem{proposition}{Proposition}
\usepackage{cleveref}
\crefname{algorithm}{Algorithm}{Algorithms}
\crefname{assumption}{Assumption}{Assumptions}
\crefname{equation}{Eq.}{Eqs.}
\crefname{figure}{Fig.}{Figs.}
\crefname{table}{Table}{Tables}
\crefname{section}{Sec.}{Secs.}
\crefname{theorem}{Theorem}{Theorems}
\crefname{lemma}{Lemma}{Lemmas}
\crefname{proposition}{Proposition}{Propositions}
\crefname{definition}{Definition}{Definitions}
\crefname{corollary}{Corollary}{Corollaries}
\crefname{rem}{Remark}{Remarks}
\crefname{example}{Example}{Examples}
\crefname{appendix}{Appendix}{Appendices}
\title{A Stochastic Optimization Framework for\\
Fair Risk Minimization
}
\author{\name Andrew Lowy\thanks{AL and SB contributed equally to this paper.}~\thanks{The work of AL, SB and MR was supported in part with funding from the NSF CAREER Award 2144985, from a gift from 3M, and from the USC-Meta Center for Research and Education in AI \& Learning.} \email <EMAIL> \\
\addr
University of Southern California
\AND
\name Sina Baharlouei\footnotemark[1]~\footnotemark[2] \email <EMAIL> \\
\addr
University of Southern California
\AND
\name Rakesh Pavan \email [email protected]\\
\addr
University of Washington \\
\AND
\name Meisam Razaviyayn\footnotemark[2] \email [email protected]\\
\addr
University of Southern California \\
\AND \name Ahmad Beirami\thanks{The work of AB was done at Meta AI.} \email <EMAIL> \\
Google Research
}
\newcommand{\marginpar{FIX}}{\marginpar{FIX}}
\newcommand{\marginpar{NEW}}{\marginpar{NEW}}
\def\month{MM}
\def\year{YYYY}
\def\openreview{\url{https://openreview.net/forum?id=P9Cj6RJmN2}}
\begin{document}
\maketitle
\begin{abstract}
Despite the success of large-scale empirical risk minimization (ERM) at achieving high accuracy across a variety of machine learning tasks, \textit{fair} ERM is hindered by the incompatibility of fairness constraints with stochastic optimization. We consider the problem of fair classification with discrete sensitive attributes and potentially large models and data sets, requiring stochastic solvers.
Existing in-processing fairness algorithms are either impractical in the large-scale setting because they require large batches of data at each iteration or they are not guaranteed to converge. In this paper, we develop
\textit{the first stochastic in-processing fairness algorithm with guaranteed convergence}.
For demographic parity, equalized odds, and equal opportunity notions of fairness, we provide slight variations of our algorithm--called FERMI--and prove that each of these variations converges in stochastic optimization with any batch size. Empirically, we show that FERMI is amenable to stochastic solvers with multiple (non-binary) sensitive attributes and non-binary targets, performing well even with minibatch size as small as one. Extensive experiments show that FERMI achieves the most favorable tradeoffs between fairness violation and test accuracy across all tested setups compared with state-of-the-art baselines for demographic parity, equalized odds, equal opportunity. These benefits are especially significant with small batch sizes and for non-binary classification with large number of sensitive attributes, making FERMI a practical, scalable fairness algorithm. The code for all of the experiments in this paper is available at:\\
{\color{blue}\url{https://github.com/optimization-for-data-driven-science/FERMI}}.
\end{abstract}
\section{Introduction}
Ensuring that decisions made using machine learning (ML) algorithms are fair to different subgroups is of utmost importance. Without any mitigation strategy, learning algorithms may result in discrimination against certain subgroups based on sensitive attributes, such as gender or race, even if such discrimination is absent in the training data~\citep{mehrabi2021survey}, and algorithmic fairness literature aims to remedy such discrimination issues~\citep{sweeney2013discrimination, datta2015automated, feldman2015certifying, bolukbasi2016man, machinebias2016, calmon2017optimized, hardt2016equality, fish2016confidence, woodworth2017learning, zafar2017fairness, bechavod2017penalizing, agarwal2018reductions, kearns2018preventing, prost2019toward, lahoti2020fairness}.
Modern ML problems often involve large-scale models with hundreds of millions or even billions of parameters, e.g., BART \citep{lewis2019bart}, ViT \citep{dosovitskiy2020image}, GPT-2 \citep{radford2019language}. In such cases, during fine-tuning, the available memory on a node constrains us to use stochastic optimization with (small) minibatches in each training iteration. In this paper, we address the dual challenges of \textit{fair} and \textit{stochastic} ML, providing \textit{the first stochastic fairness algorithm that provably converges with any batch size}.
A machine learning algorithm satisfies the \emph{demographic parity} fairness notion if the predicted target is independent of the sensitive attributes~\citep{dwork2012fairness}. Promoting demographic parity can lead to poor performance, especially if the true outcome is not independent of the sensitive attributes. To remedy this,
\cite{hardt2016equality} proposed \emph{equalized odds} to ensure that the predicted target is conditionally independent of the sensitive attributes given the true label. A further relaxed version of this notion is \emph{equal opportunity} which is satisfied if predicted target is conditionally independent of sensitive attributes given that the true label is in an advantaged class~\citep{hardt2016equality}. Equal opportunity ensures that false positive rates are equal across different demographics, where negative outcome is considered an advantaged class, e.g., extending a loan. See~\cref{app:existing notions of fairness} for formal definitions of these fairness notions.
\begin{table*}[t]
\centering
\resizebox{0.75\textwidth}{!}{
\begin{tabular}{||l | c | c | c | c | c | c | c | c | c||}
\hline
Reference & NB & NB & NB
& Beyond & Stoch. alg. & Converg. \\
& target & attrib. & code
& logistic& (unbiased$^{**}$) & (stoch.)
\\ [0.5ex]
\hline\hline
{\bf FERMI (this work)}
& \cmark & \cmark & \cmark
& \cmark& {\cmark}~({\cmark}) & {\cmark}~({\cmark})
\\
\hline
\citep{cho2020kde}
& \cmark & \cmark & \cmark
& \cmark & {\cmark}~({\xmark}) & \xmark \\
\hline
\citep{cho2020fair}
&\cmark& \cmark &\xmark & \cmark & {\cmark}~({\cmark}) & \xmark
\\
\hline
\citep{baharlouei2019rnyi}
& \cmark & \cmark & \cmark
&\cmark & \xmark & {\cmark}~({\xmark})
\\
\hline
\citep{rezaei2020fairness}
& \xmark & \xmark & \xmark
& \xmark & \xmark & \xmark \\
\hline
\citep{jiang2020wasserstein}$^*$ &
\xmark & \cmark & \xmark
& \xmark & \xmark & \xmark \\
\hline
\citep{mary19} & \cmark & \cmark & \cmark
&\cmark& {\cmark}~({\xmark}) & \xmark \\
\hline
\citep{prost2019toward}
& \xmark & \xmark & \xmark
&\cmark& {\cmark}~({\xmark}) & \xmark
\\
\hline
\citep{donini2018empirical} &
\xmark & \cmark & \xmark
& \cmark & \xmark & \xmark \\
\hline
\citep{zhang2018mitigating}
& \cmark & \cmark & \xmark
&\cmark& {\cmark}~({\xmark}) & \xmark
\\
\hline
\citep{agarwal2018reductions}
& \xmark & \cmark & \xmark
&\cmark& \xmark & {\cmark}~({\xmark})
\\
\hline
\end{tabular}
}
\caption{
\label{Table: comparison with prior work}\small
Comparison of state-of-the-art in-processing methods ({\bf NB = non-binary}) on whether they (a) handle non-binary targets (beyond binary classification), (b) handle non-binary sensitive attributes, (c) release code that applies to non-binary targets/attributes, (d) extend to arbitrary models, (e) provide code for stochastic optimization (and whether the gradients are unbiased), (f) provide convergence guarantees (for stochastic optimization).
FERMI is the only method compatible with stochastic optimization and guaranteed convergence. The only existing baselines for non-binary classification with non-binary sensitive attributes are \citep{mary19, baharlouei2019rnyi, cho2020kde} (NB code). $^*$We refer to the in-processing method of \citep{jiang2020wasserstein}, not their post-processing method. $^{**}$We use the term ``unbiased'' in statistical estimation sense; not to be confused with bias in the fairness sense.
}
\end{table*}
\noindent \textbf{Measuring fairness violation.}
In practice, the learner only has access to finite samples and cannot verify demographic parity, equalized odds, or equal opportunity. This has led the machine learning community to define several fairness violation metrics that quantify the degree of (conditional) independence between random variables, e.g., $L_\infty$ distance~\citep{dwork2012fairness, hardt2016equality}, mutual information~\citep{kamishima2011fairness, rezaei2020fairness, steinberg2020fast, zhang2018mitigating, cho2020fair, roh2020fr}, Pearson correlation~\citep{zafar2017fairness,beutel2019putting}, false positive/negative rate difference \citep{bechavod2017penalizing}, Hilbert Schmidt independence criterion (HSIC)~\citep{perez2017fair},
kernel-based minimum mean discrepancy (MMD)~\citep{prost2019toward}, R\'enyi correlation~\citep{mary19,baharlouei2019rnyi, grari2019fairness,grari2020learning}, and exponential R\'enyi mutual information (ERMI)~\citep{mary19}. In this paper, we focus on three variants of ERMI specialized to demographic parity, equalized odds, and equal opportunity. The motivation for the use of ERMI is two-fold. First, we will see in \cref{sec:solver} that \textit{ERMI is amenable to stochastic optimization}. Moreover, we observe (\cref{appendix: relations btwn ERMI and fairness}) that ERMI provides an upper bound on several of the above notions of fairness violation. Consequently, a model trained to reduce ERMI will also provide guarantees on these other fairness violations.\footnote{Nevertheless, we use $L_\infty$ distance for measuring fairness violation in our numerical experiments, since $L_\infty$ is broadly used.}
\noindent \textbf{Related work
\& contributions.
}
Fairness-promoting machine learning algorithms can be categorized in three main classes: \emph{pre-processing}, \emph{post-processing}, and \emph{in-processing} methods. Pre-processing algorithms \citep{feldman2015certifying, zemel2013learning,calmon2017optimized} transform the biased data features to a new space in which the labels and sensitive attributes are statistically independent. This transform is oblivious to the training procedure. Post-processing approaches \citep{hardt2016equality, pleiss2017fairness} mitigate the discrimination of the classifier by altering the final decision. In-processing approaches focus on the training procedure and impose the notions of fairness as constraints or regularization terms in the training procedure.
Several regularization-based methods are proposed in the literature to promote fairness~\citep{ristanoski2013discrimination, quadrianto2017recycling} in decision-trees \citep{kamiran2010discrimination,raff2018fair, aghaei2019learning}, support vector machines \citep{donini2018empirical}, boosting~\citep{fish2015fair}, neural networks \citep{grari2020learning, cho2020kde, prost2019toward}, or (logistic) regression models~\citep{zafar2017fairness, berk2017convex, taskesen2020distributionally, chzhen2020minimax, baharlouei2019rnyi, jiang2020wasserstein, grari2019fairness}.
See the recent paper by~\cite{hort2022bia} for a more comprehensive literature survey.
While in-processing approaches generally give rise to better tradeoffs between fairness violation and performance, existing approaches are mostly incompatible with stochastic optimization. This paper addresses this problem in the context of fair (non-binary) classification with discrete (non-binary) sensitive attributes. See \cref{Table: comparison with prior work} for a summary of the main differences between FERMI and existing in-processing methods.
Our main contributions are as follows:
\vspace{-.05in}
\begin{enumerate}[leftmargin=*]
\setlength\itemsep{0.02in}
\item For each given fairness notion (demographic parity, equalized odds, or equal opportunity), we formulate an objective that uses ERMI as a regularizer to balance fairness and accuracy (\cref{eq: FRMI}), and derive an empirical version of this objective (\cref{eq: FERMI}). We propose an algorithm (\cref{alg: SGDA for FERMI}) for solving each of these objectives, which is {\em the first stochastic in-processing fairness algorithm with guaranteed convergence}. The main property needed to obtain a convergent stochastic algorithm is to derive a (stochastically) unbiased estimator of the gradient of the objective function. The existing stochastic fairness algorithms by~\cite{zhang2018mitigating, mary19, prost2019toward, cho2020fair, cho2020kde}
are not guaranteed to converge since there is no straightforward way to obtain such unbiased estimator of the gradients for their fairness regularizers.\footnote{We suspect it might be possible to derive a provably convergent stochastic algorithm from the framework in~\cite{prost2019toward} using our techniques, but their approach is limited to binary classification with binary sensitive attributes. By contrast, we provide (empirical and population-level) convergence guarantees for our algorithm with any number of sensitive attributes and any number of classes.} For any minibatch size (even as small as $1$), we prove (\cref{thm: SGDA for FERMI convergence rate}) that our algorithm converges to an approximate solution of the empirical objective (\cref{eq: FERMI}).
\item We show that if the number of training examples is large enough, then our algorithm (\cref{alg: SGDA for FERMI}) converges to an approximate solution of the \textit{population}-level objective (\cref{thm: SGDA for FRMI convergence rate}).
The proofs of these convergence theorems require the development of novel techniques (see e.g. Proposition~\ref{thm: min-max} and Proposition~\ref{prop: consistency}), and the resourceful application of many classical results from optimization, probability theory, and statistics.
\item We demonstrate through extensive numerical experiments that our stochastic algorithm achieves {superior fairness-accuracy tradeoff curves against all comparable baselines}
for demographic parity, equalized odds, and equal opportunity. In particular, \textit{the performance gap is very large when minibatch size is small} (as is practically necessary for large-scale problems) and the number of sensitive attributes is large.
\vspace{-.05in}
\end{enumerate}
\section{Fair Risk Minimization through ERMI Regularization}
\vspace{-.1in}
\label{sec:solver}
In this section, we propose a fair learning objective (\cref{eq: FRMI}) and derive an empirical variation (\cref{eq: FERMI}) of this objective. We then develop a stochastic optimization algorithm (\cref{alg: SGDA for FERMI}) that we use to solve these objectives, and prove that our algorithm converges to an approximate solution of the two objectives.
Consider a learner who trains a model to make a prediction, $\widehat{Y},$ e.g., whether or not to extend a loan, supported on $[m]:= \{1, \ldots, m\}$.
The prediction is made using a set of features, $\mathbf{X}$, e.g., financial history features. Assume that there is a set of discrete sensitive attributes, $S,$ e.g., race and sex, supported on $[k]$.
We now define the fairness violation notion that we will use to enforce fairness in our model.
\begin{definition}[ERMI -- exponential R\'enyi mutual information]
\label{def: ERMI}
We define the exponential R\'enyi mutual information between random variables $\widehat{Y}$ and $S$ with joint distribution $p_{\widehat{Y},S}$ and marginals $p_{\widehat{Y}}, ~p_S$ by:
\begin{align}
D_R(\widehat{Y}; S)\nonumber := \mathbb{E}
\left\{
\frac{p_{\widehat{Y}, S}(\widehat{Y}, S) }{p_{\widehat{Y}}(\widehat{Y}) p_{S}(S)}
\right\} - 1 = \sum_{j \in [m]} \sum_{r \in [k]} \frac{p_{\widehat{Y}, S}(j, r)^2}{p_{\widehat{Y}}(j) p_S(r)} - 1.
\vspace{-.5in}
\tag{ERMI}
\label{eq: ERMI}
\end{align}
\normalsize
\vspace{-.15in}
\end{definition}
\noindent \cref{def: ERMI} is what we would use if \textit{demographic parity} were the fairness notion of interest. If instead one wanted to promote fairness with respect to equalized odds or equal opportunity, then it is straightforward to modify the definition by substituting appropriate conditional probabilities for $p_{\widehat{Y}, S}, ~p_{\widehat{Y}},$ and $p_S$ in \cref{eq: ERMI}: see~\cref{app:ERMI}. In \cref{app:ERMI}, we
also discuss that ERMI is the $\chi^2$-divergence (which is an $f$-divergence) between the joint distribution, $p_{\widehat{Y}, S},$ and the Kronecker product of marginals, $p_{\widehat{Y}} \otimes p_{S}$~\citep{calmon2017principal}. In particular, ERMI is non-negative, and zero if and only if
demographic parity (or equalized odds or equal opportunity, for the conditional version of ERMI)
is satisfied. Additionally, we show in \cref{appendix: relations btwn ERMI and fairness} that ERMI provides an upper bound on other commonly used measures of fairness violation: Shannon mutual information \citep{cho2020fair}, R\'enyi correlation \citep{baharlouei2019rnyi}, $L_q$ fairness violation \citep{kearns2018preventing, hardt2016equality}. Therefore, any algorithm that makes ERMI small will also have small fairness violation with respect to these other notions.
We can now define our fair risk minimization through exponential R\'enyi mutual information framework:\footnote{In this section, we present all results in the context of
demographic parity,
leaving off all conditional expectations for clarity of presentation. The algorithm/results are readily extended
to equalized odds and equal opportunity
by using
the conditional version of \cref{eq: ERMI} (which is described in \cref{app:ERMI}); we use these resulting algorithms for numerical experiments.
}
\begin{equation}
\label{eq: FRMI}
\min_{\boldsymbol{\theta}}
\left\{\text{FRMI}(\boldsymbol{\theta}) :=
\mathcal{L} (\boldsymbol{\theta}) + \lambda D_R \big(\widehat{Y}_{\boldsymbol{\theta}}(\mathbf{X}); S\big)
\right\} ,
\tag{FRMI obj.}
\end{equation}
\noindent where $\mathcal{L}(\boldsymbol{\theta}):= \mathbb{E}_{(\mathbf{X}, Y)}[\ell(\mathbf{X}, Y; \boldsymbol{\theta})]$ for a given loss function $\ell$ (e.g. $L_2$ loss or cross entropy loss); $\lambda>0$ is a scalar balancing the accuracy versus fairness objectives;
and $\widehat{Y}_{\boldsymbol{\theta}}(\mathbf{X})$ is the output of the learned model (i.e. the predicted label in a classification task). While $\widehat{Y}_{\boldsymbol{\theta}}(\mathbf{X}) = \widehat{Y}(\mathbf{X}; \boldsymbol{\theta})$ inherently depends on $\mathbf{X}$ and $\boldsymbol{\theta}$,
in the rest of this paper, we sometimes leave the dependence of $\widehat{Y}$ on $\mathbf{X}$ and/or $\boldsymbol{\theta}$ implicit for brevity of notation. Notice that we have also left the dependence of the loss on the predicted outcome $\widehat{Y} = \widehat{Y}_{\boldsymbol{\theta}}(\mathbf{X})$ implicit.
As is standard, we assume that the prediction function satisfies $\mathbb{P}(\widehat{Y}(\boldsymbol{\theta}, \mathbf{X}) = j | \mathbf{X}) = {\mathcal{F}}_j(\boldsymbol{\theta}, \mathbf{X})$, where ${\mathcal{F}}(\boldsymbol{\theta}, \mathbf{X}) = ({\mathcal{F}}_1(\boldsymbol{\theta}, \mathbf{X}), \ldots, {\mathcal{F}}_m(\boldsymbol{\theta}, \mathbf{X}))^T \in [0,1]^m$ is differentiable in $\boldsymbol{\theta}$ and $\sum_{j=1}^m {\mathcal{F}}_j(\boldsymbol{\theta}, \mathbf{X}) = 1$.
For example, ${\mathcal{F}}(\boldsymbol{\theta}, \mathbf{X})$ could represent the probability label given by a logistic regression model or the output of a neural network after softmax layer.
Indeed, this assumption is natural for most classifiers. Further, even classifiers, such as SVM, that are not typically expressed using probabilities can often be well approximated by a classifier of the form $\mathbb{P}(\widehat{Y}(\boldsymbol{\theta}, \mathbf{X}) = j | \mathbf{X}) = {\mathcal{F}}_j(\boldsymbol{\theta}, \mathbf{X})$, e.g. by using Platt Scaling~\citep{platt1999probabilistic, niculescu2005predicting}.
The work of~\cite{mary19} considered the same objective \cref{eq: FRMI}, and tried to empirically solve it through a kernel approximation. We propose a different approach to solving this problem, which we shall describe below. Essentially, we express ERMI as a ``max'' function~(\cref{thm: min-max}), which enables us to re-formulate~\cref{eq: FRMI} (and its empirical counterpart~\cref{eq: FERMI}) as a stochastic min-max optimization problem. This allows us to use stochastic gradient descent ascent (SGDA) to solve~\cref{eq: FRMI}. Unlike the algorithm of~\cite{mary19}, our algorithm
provably \textit{converges}. Our algorithm also empirically outperforms the algorithm of \cite{mary19}, as we show in~\cref{sec:experiments} and~\cref{sec:comparison_convergence}.
\subsection{A Convergent Stochastic Algorithm for Fair Empirical Risk Minimization
}
\label{sec: empirical FERMI}
\vspace{-.07in}
In practice, the true joint distribution
of $(\mathbf{X}, S, Y, \widehat{Y})$ is unknown and we only have $N$ samples at our disposal.
Let $\mathbf{D} = \{\mathbf{x}_{i}, s_i, y_{i}, \widehat{y}(\mathbf{x}_i; \boldsymbol{\theta}) \}_{i \in [N]}$ denote the features, sensitive attributes, targets, and the predictions of the model parameterized by $\boldsymbol{\theta}$ for these given samples.
For now, we consider the empirical risk minimization (ERM) problem
and do not require any assumptions on the data set (e.g. we allow for different samples in $\mathbf{D}$ to be drawn from different, heterogeneous distributions).
Consider the empirical objective
\begin{equation}
\min_{\boldsymbol{\theta}}
\left
\{\text{FERMI}(\boldsymbol{\theta}
)
:=
\widehat{\mathcal{L}}(\boldsymbol{\theta}) +\lambda \widehat{D}_R(\widehat{Y}_{\boldsymbol{\theta}}(\mathbf{X}), S)
\right
\},
\tag{FERMI obj.}
\label{eq: FERMI}
\vspace{-0.06in}
\end{equation}
\normalsize
where $\widehat{\mathcal{L}}(\boldsymbol{\theta}):= \frac{1}{N}\sum_{i=1}^N \ell(\mathbf{x}_i, y_i; \boldsymbol{\theta})$ is the empirical loss and\footnote{We overload notation slightly here and use $\mathbb{E}$ to denote expectation with respect to the \textit{empirical} (joint) distribution.}
\small
\begin{equation*}
\small
\widehat{D}_R(\widehat{Y}, S) :=\mathbb{E}\left\{\frac{\hat{p}_{\widehat{Y}, S}(\widehat{Y}, S) }{\hat{p}_{\widehat{Y}}(\widehat{Y}) \hat{p}_{S}(S)} \right\} - 1 = \sum_{j \in [m]} \sum_{r \in [k]} \frac{\hat{p}_{\widehat{Y}, S}(j, r)^2}{\hat{p}_{\widehat{Y}}(j) \hat{p}_S(r)} - 1
\vspace{-0.06in}
\end{equation*}
\normalsize
is \textit{empirical ERMI} with $\hat{p}$ denoting empirical probabilities with respect to $\mathbf{D}$:
$\hat{p}_S(r) = \frac{1}{N}\sum_{i=1}^N \mathbbm{1}_{\{s_i = r\}}$;
$\widehat{p}_{\hat{y}}(j) = \frac{1}{N}\sum_{i=1}^N {\mathcal{F}}_j(\boldsymbol{\theta}, \mathbf{x}_i)$; and
$\hat{p}_{\widehat{Y}, S}(j, r) = \frac{1}{N} \sum_{i=1}^N {\mathcal{F}}_j(\boldsymbol{\theta}, \mathbf{x}_i) \mathbf{s}_{i,r}$ for $j \in [m], r \in [k]$. We shall see (\cref{prop: consistency}) that empirical ERMI is a good approximation of ERMI when $N$ is large. Now, it is straightforward to derive an unbiased estimate for $\widehat{\mathcal{L}}(\boldsymbol{\theta}) $ via $\frac{1}{|B|} \sum_{i \in B} \ell \big(\mathbf{x}_i, y_i; \boldsymbol{\theta} \big)$ where $B\subseteq [N]$ is a random minibatch of data points drawn from $\mathbf{D}$. However, unbiasedly estimating $\widehat{D}_R(\widehat{Y}, S)$ in the objective function~\cref{eq: FERMI} with $|B| < N$ samples is more difficult. In what follows, we present our approach to deriving statistically \textit{unbiased stochastic estimators} of the gradients of $\widehat{D}_R(\widehat{Y}, S)$ given a random batch of data points~$B$. This stochastic estimator is key to developing a stochastic convergent algorithm for solving \cref{eq: FERMI}. The key novel observation that allows us to derive this estimator is that Equation~\ref{eq: FERMI} can be written as a \textit{min-max} optimization problem (see~\cref{coro:MiniBatchUnbiased}). This observation, in turn, follows from the following result:
\begin{proposition}
\label{thm: min-max}
For random variables $\widehat{Y}
$ and $S$ with joint distribution $\hat{p}_{\widehat{Y}, S}$, where $\widehat{Y} \in [m], S \in [k],$ we have
\begin{equation}
\widehat{D}_R(\widehat{Y}; S) = \max_{W \in \mathbb{R}^{k \times m}} \nonumber
\{-\Tr(W \widehat{P}_{\hat{y}} W^T)
+2 \Tr(W \widehat{P}_{\hat{y}, s} \widehat{P}_{s}^{-1/2}) - 1
\},
\label{eq:stochasticDR}
\end{equation}
\noindent if $\widehat{P}_{\hat{y}} = {\rm diag}(\hat{p}_{\widehat{Y}}(1), \ldots, \hat{p}_{\widehat{Y}}(m)),$ $\widehat{P}_{s} = {\rm diag}(\hat{p}_{S}(1), \ldots, \hat{p}_{S}(k)),$ ~and $(\widehat{P}_{\haty, s})_{i,j} =
\hat{p}_{\widehat{Y}, S}(i, j)$ with $\hat{p}_{\widehat{Y}}(i), \hat{p}_{S}(j) > 0$ for $i \in [m], j \in [k].$
\end{proposition}
The proof is a direct calculation, given in~\cref{sec: Appendix Stochastic FERMI}. Let $\widehat{\mathbf{y}}(\mathbf{x}_i, \boldsymbol{\theta})
\in \{0,1\}^m$
and $\mathbf{s}_i = (\mathbf{s}_{i,1}, \ldots, \mathbf{s}_{i, k})^T
\in\{0,1\}^k$
be the one-hot encodings of $\widehat{y}(\mathbf{x}_i, \boldsymbol{\theta})$ and $s_i$, respectively for $i \in [N]$.
Then,~\cref{thm: min-max} provides a useful variational form of \cref{eq: FERMI}, which forms the backbone of our novel algorithmic approach:
\begin{corollary}\label{coro:MiniBatchUnbiased}
Let $(\mathbf{x}_i, s_i, y_i)$ be a random draw from $\mathbf{D}$. Then,
\cref{eq: FERMI} is equivalent to
\begin{align}
\label{eq: empirical minmax}
\min_{\boldsymbol{\theta}} \max_{W \in \mathbb{R}^{k \times m}}
\left\{
\widehat{F}(\boldsymbol{\theta}, W) :=
\widehat{\mathcal{L}}(\boldsymbol{\theta}) + \lambda \widehat{\Psi}(\boldsymbol{\theta}, W)
\right\},
\end{align}
\normalsize
where $\widehat{\Psi}(\boldsymbol{\theta}, W) = - \Tr(W \widehat{P}_{\hat{y}} W^T) + 2 \Tr(W \widehat{P}_{\hat{y}, s} \widehat{P}_{s}^{-1/2}) - 1 = \frac{1}{N}\sum_{i=1}^N \widehat{\psi}_i(\boldsymbol{\theta}, W)$
and
\begin{align*}
\small
\widehat{\psi}_i(\boldsymbol{\theta}, W) &:= -\Tr(W \mathbb{E}[\widehat{\mathbf{y}}(\mathbf{x}_i, \boldsymbol{\theta}) \widehat{\mathbf{y}}(\mathbf{x}_i, \boldsymbol{\theta})^T | \mathbf{x}_i]
W^T)
+ 2 \Tr(W \mathbb{E}[\widehat{\mathbf{y}}(\mathbf{x}_i; \boldsymbol{\theta}) \mathbf{s}_i^T | \mathbf{x}_i, s_i] \widehat{P}_{s}^{-1/2}) - 1 \\
&= -\Tr(W {\rm diag}({\mathcal{F}}_1(\boldsymbol{\theta}, \mathbf{x}_i), \ldots, {\mathcal{F}}_m(\boldsymbol{\theta}, \mathbf{x}_i))
W^T)
+ 2 \Tr(W \mathbb{E}[\widehat{\mathbf{y}}(\mathbf{x}_i; \boldsymbol{\theta}) \mathbf{s}_i^T | \mathbf{x}_i, s_i] \widehat{P}_{s}^{-1/2}) - 1.
\end{align*}
\vspace{-.2in}
\end{corollary}
\cref{coro:MiniBatchUnbiased} implies that for any given data set
$\mathbf{D}$, the quantity $\ell(\mathbf{x}_i, y_i; \boldsymbol{\theta}) + \lambda \widehat{\psi}_i(\boldsymbol{\theta}, W)$ is an unbiased estimator of $\widehat{F}(\boldsymbol{\theta}, W)$
(with respect to the uniformly random draw of $i \in [N]$). Thus, we can use stochastic optimization (e.g. SGDA) to solve \cref{eq: FERMI} with any batch size $1 \leq |B| \leq N$, and the resulting algorithm will be guaranteed to converge since the stochastic gradients are unbiased. We present our proposed algorithm, which we call \textit{FERMI}, for solving \cref{eq: FERMI} in~\cref{alg: SGDA for FERMI}.
\begin{algorithm}
\caption{FERMI Algorithm
}
\label{alg: SGDA for FERMI}
\begin{algorithmic}
\STATE \textbf{Input}: $\boldsymbol{\theta}^0 \in \mathbb{R}^{d_{\theta}}, ~W^0 = 0 \in
\mathbb{R}^{k \times m}$, step-sizes $(\eta_\theta, \eta_w),$
fairness parameter $\lambda \geq 0,$ iteration number $T$, minibatch sizes $|B_t|, t \in \{0, 1, \cdots, T\}$, $\mathcal{W} :=$ Frobenius norm ball of radius $D$ around $0 \in \mathbb{R}^{k \times m}$ for $D$ given in~\cref{sec: Appendix Stochastic FERMI}.
\STATE Compute $\widehat{P}_s^{-1/2} = {\rm{diag}}(\hat{p}_S(1)^{-1/2}, \ldots, \hat{p}_S(k)^{-1/2})$.
\FOR {$t = 0, 1, \ldots, T$}
\STATE Draw a mini-batch $B_t$ of data points $\{(\mathbf{x}_i,s_i, y_i)\}_{i\in B_t}$
\STATE Set $\small \boldsymbol{\theta}^{t+1} \gets \boldsymbol{\theta}^t - \frac{\eta_\theta}{|B_t|} \sum_{i \in B_t} [ \nabla_\theta \ell(\mathbf{x}_i, y_i; \boldsymbol{\theta}^t) + \lambda \nabla_\theta \widehat{\psi}_i(\boldsymbol{\theta}^t, W^t)].$
\STATE Set $\small
W^{t+1} \gets
\Pi_{\mathcal{W}}
\Big(
W^t +
\frac{2 \lambda \eta_w}{|B_t|} \sum_{i\in B_t}\Big[
- W^t
\mathbb{E}[\widehat{\mathbf{y}}(\mathbf{x}_i, \boldsymbol{\theta}) \widehat{\mathbf{y}}(\mathbf{x}_i, \boldsymbol{\theta})^T | \mathbf{x}_i]
+ \widehat{P}_s^{-1/2} \mathbb{E}[{\mathbf{s}}_i\widehat{\mathbf{y}}^T(\mathbf{x}_i; \boldsymbol{\theta}^t)|\mathbf{x}_i, s_i] \Big]
\Big)
$
\ENDFOR
\STATE Pick $\hat{t}$ uniformly at random from $\{1, \ldots, T\}.$\\
\STATE \textbf{Return:} $\boldsymbol{\theta}^{\hat{t}}.$
\end{algorithmic}
\vspace{-.03in}
\end{algorithm}
Note that the matrix $\widehat{P}_s^{-1/2}$ depends only on the full data set of sensitive attributes $\{s_1, \cdots, s_N\}$ and has no dependence on $\boldsymbol{\theta}$, and can therefore be computed just once, in line 2 of~\cref{alg: SGDA for FERMI}. On the other hand, the quantities $\mathbb{E}[\widehat{\mathbf{y}}(\mathbf{x}_i, \boldsymbol{\theta}) \widehat{\mathbf{y}}(\mathbf{x}_i, \boldsymbol{\theta})^T | \mathbf{x}_i]$ and $\mathbb{E}[\widehat{\mathbf{y}}(\mathbf{x}_i; \boldsymbol{\theta}) \mathbf{s}_i^T | \mathbf{x}_i, s_i]$ depend on the sample $(\mathbf{x}_i, s_i, \widehat{y}_i)$ that is drawn in a given iteration and on the model parameters $\boldsymbol{\theta}$, and are therefore computed at each iteration of the algorithm.
Although the min-max problem~\cref{eq: FERMI} that we aim to solve is unconstrained, we project the iterates $W^t$ (in line 5 of~\cref{alg: SGDA for FERMI}) onto a bounded set $\mathcal{W}$ in order to satisfy a technical assumption that is needed to prove convergence of \cref{alg: SGDA for FERMI}\footnote{Namely, bounded $W^t$ ensures that the gradient of $\widehat{F}$ is Lipschitz continuous at every iterate and the variance of the stochastic gradients is bounded.}. We choose $\mathcal{W}$ to be a sufficiently large ball that contains $W^*(\boldsymbol{\theta}) := \argmax_{W} \widehat{F}(\boldsymbol{\theta}, W)$ for every $\boldsymbol{\theta}$ in some neighborhood of $\boldsymbol{\theta}^* \in \argmin_{\boldsymbol{\theta}} \max_{W} \widehat{F}(\boldsymbol{\theta}, W)$, so that~\cref{eq: FERMI} is equivalent to
\vspace{-.04in}
\[\min_{\boldsymbol{\theta}} \max_{W \in \mathcal{W}}
\left\{
\widehat{F}(\boldsymbol{\theta}, W) =
\widehat{\mathcal{L}}(\boldsymbol{\theta}) + \lambda \widehat{\Psi}(\boldsymbol{\theta}, W)
\right\}.
\normalsize
\vspace{-.04in}
\]
See~\cref{sec: Appendix Stochastic FERMI} for details. When applying~\cref{alg: SGDA for FERMI} in practice, it is not necessary to project the iterates; e.g. in~\cref{sec:experiments}, we obtain strong empirical results without projection in~\cref{alg: SGDA for FERMI}.
Since \cref{eq: FERMI} is potentially nonconvex in $\boldsymbol{\theta}$, a global minimum might not exist and even computing a local minimum is NP-hard in general \citep{murty1985}. Thus, as is standard in the nonconvex optimization literature, we aim for the milder goal of finding an approximate \textit{stationary point} of \cref{eq: FERMI}. That is, given any $\epsilon > 0$, we aim to find a point $\boldsymbol{\theta}^*$ such that $\mathbb{E}\|\nabla \text{FERMI}(\boldsymbol{\theta}^*) \| \leq \epsilon,$ where the expectation is solely with respect to the randomness of the algorithm (minibatch sampling). The following theorem guarantees that \cref{alg: SGDA for FERMI} will find such a point efficiently:
\begin{theorem} (Informal statement)
\label{thm: SGDA for FERMI convergence rate}
Let $\epsilon > 0$. Assume that $\ell(\mathbf{x}, y; \cdot)$ and ${\mathcal{F}}(\cdot, \mathbf{x})$ are Lipschitz continuous and differentiable with Lipschitz continuous gradient (see Appendix~\ref{sec: Appendix Stochastic FERMI} for definitions), $\hat{p}_S(j) > 0$ for all sensitive attributes $j \in [k]$ and $\hat{p}_{\hat{Y}}(l) \geq \mu >0$
for all labels $l \in [m]$ and at every iterate $\boldsymbol{\theta}^t$.
Then for any batch sizes $1 \leq |B_t| \leq N$, Algorithm~\ref{alg: SGDA for FERMI} converges to an $\epsilon$-first order stationary point of the
\cref{eq: FERMI}
objective in $\mathcal{O}\left(\frac{1}{\epsilon^5}\right)$ stochastic gradient evaluations.
\vspace{-.03in}
\end{theorem}
The formal statement of \cref{thm: SGDA for FERMI convergence rate} can be found in Theorem~\ref{thm: precise convergence} in Appendix~\ref{sec: Appendix Stochastic FERMI}. \cref{thm: SGDA for FERMI convergence rate} implies that~\cref{alg: SGDA for FERMI} can efficiently achieve any tradeoff between fairness (ERMI) violation and (empirical) accuracy, depending on the choice of $\lambda$.\footnote{This sentence is accurate to the degree that an approximate stationary point of the non-convex objective \cref{eq: FERMI} corresponds to an approximate risk minimizer.} However, if smaller fairness violation is desired (i.e. if larger $\lambda$ is chosen), then the algorithm needs to run for more iterations (see Appendix~\ref{sec: Appendix Stochastic FERMI}). The proof of~\cref{thm: SGDA for FERMI convergence rate} follows from Corollary~\ref{coro:MiniBatchUnbiased} and the observation that $\widehat{\psi}_i$ is strongly concave in $W$ (see \cref{lem: kappa stuff} in~\cref{sec: Appendix Stochastic FERMI}). This implies that \cref{eq: empirical minmax} is a nonconvex-strongly concave min-max problem, so the convergence guarantee of SGDA~\citep{lin20} yields~\cref{thm: SGDA for FERMI convergence rate}.\footnote{A faster convergence rate of $\mathcal{O}(\epsilon^{-3})$ could be obtained by using the (more complicated) SREDA method of \cite{luo20} instead of SGDA to solve FERMI objective. We omit the details here.} The detailed proof of \cref{thm: SGDA for FERMI convergence rate} is given in Appendix~\ref{sec: Appendix Stochastic FERMI}. Increasing the batch size to $\Theta(\epsilon^{-2})$ improves the stochastic gradient complexity to $\mathcal{O}(\epsilon^{-4})$. On the other hand, increasing the batch size further to $|B_t| = N$ results in a deterministic algorithm which is guaranteed to find a point $\boldsymbol{\theta}^*$ such $\|\nabla \text{FERMI}(\boldsymbol{\theta}^*) \| \leq \epsilon$ (no expectation) in $\mathcal{O}(\epsilon^{-2})$ iterations \cite[Theorem 4.4]{lin20},\cite[Remark 4.2]{olr20}; this iteration complexity has the optimal dependence on $\epsilon$ \citep{carmon2020lower, zhang2021complexity}.
However, like existing fairness algorithms in the literature, this full-batch variant is impractical for large-scale problems.
\begin{remark}
\label{rem: strong concavity}
The condition $\hat{p}_{\hat{Y}}(l) \geq \mu$ in~\cref{thm: SGDA for FERMI convergence rate} is assumed in order to ensure strong concavity of $\widehat{F}(\boldsymbol{\theta}^t, \cdot)$ at every iterate $\boldsymbol{\theta}^t$, which leads to the $\mathcal{O}(\epsilon^{-5})$ convergence rate. This assumption is typically satisfied in practice: for example, if the iterates $\boldsymbol{\theta}^t$ remain in a compact region during the algorithm and the classifier uses softmax, then $\hat{p}_{\hat{Y}}(l)\geq \mu > 0$.
Having said that, it is worth noting that this condition
is not absolutely necessary to ensure convergence of~\cref{alg: SGDA for FERMI}. Even if this condition doesn’t hold, then~\cref{eq: empirical minmax} is still a nonconvex-concave min-max problem. Hence SGDA still converges to an $\epsilon$-stationary point, albeit at the slower rate of $\mathcal{O}(\epsilon^{-8})$~\citep{lin20}. Alternatively, one can add a small $\ell_2$ regularization term to the objective to enforce strong concavity and get the fast convergence rate of $\mathcal{O}(\epsilon^{-5})$.
\end{remark}
\subsection{Asymptotic Convergence of~\cref{alg: SGDA for FERMI} for Population-level FRMI Objective}
\label{sec: pop FRMI}
\vspace{-.07in}
So far, we have let $N \geq 1$ be arbitrary and have not made any assumptions on the underlying distribution(s) from which the data was drawn.
Even so, we showed that~\cref{alg: SGDA for FERMI} always converges to a stationary point of~\cref{eq: FERMI}. Now, we will show that if $\mathbf{D}$ contains \textit{i.i.d.} samples from an unknown joint distribution $\mathcal{D}$ and if $N \gg 1$, then~\cref{alg: SGDA for FERMI} converges to an approximate solution of the \textit{population} risk minimization problem~\cref{eq: FRMI}.
Precisely, we will use a one-pass sample-without-replacement (``online'') variant of~\cref{alg: SGDA for FERMI} to obtain this population loss guarantee. The one-pass variant is identical to~\cref{alg: SGDA for FERMI} except that: a) once we draw a batch of samples $B_t$, we remove these samples from the data set so that they are never re-used; and b) the \textbf{for}-loop terminates when we have used all $n$ samples.
\begin{theorem}
\label{thm: SGDA for FRMI convergence rate}
Let $\epsilon > 0$. Assume that $\ell(\mathbf{x}, y; \cdot)$ and ${\mathcal{F}}(\cdot, \mathbf{x})$ are Lipschitz continuous and differentiable with Lipschitz continuous gradient, and that $\min_{r \in [k]} p_S(r) > 0$. Then, there exists $N \in \mathbb{N}$ such that if $n \geq N$ and $\mathbf{D} \sim \mathcal{D}^n$, then a one-pass
sample-without-replacement variant of Algorithm~\ref{alg: SGDA for FERMI} converges to an $\epsilon$-first order stationary point of the \cref{eq: FRMI}
objective in $\mathcal{O}\left(\frac{1}{\epsilon^5}\right)$ stochastic gradient evaluations, for any batch sizes $|B_t|$.
\vspace{-.03in}
\end{theorem}
\cref{thm: SGDA for FRMI convergence rate} provides a guarantee on the fairness/accuracy loss that can be achieved on unseen ``test data.'' This is important because the main goal of (fair) machine learning is to (fairly) give accurate predictions on test data, rather than merely fitting the training data well. Specifically, ~\cref{thm: SGDA for FRMI convergence rate} shows that with enough (i.i.d.) training examples at our disposal, (one-pass) ~\cref{alg: SGDA for FERMI} finds an approximate stationary point of the population-level fairness objective~\cref{eq: FRMI}. Furthermore, the gradient complexity is the same as it was in the empirical case.
The proof of~\cref{thm: SGDA for FRMI convergence rate} will be aided by the following result, which shows that $\widehat{\psi}_i$
is an asymptotically unbiased estimator of $\Psi$,
where $\max_W \Psi(\boldsymbol{\theta}, W)$ equals ERMI:
\begin{proposition}
\label{prop: consistency}
Let $\{z_i\}_{i=1}^n = \{\mathbf{x}_i, s_i, y_i\}_{i=1}^n$ be drawn i.i.d. from an unknown joint distribution $\mathcal{D}$. Denote $\widehat{\psi}_i^{(n)}(\boldsymbol{\theta}, W) = -\Tr(W \mathbb{E}[\widehat{\mathbf{y}}(\mathbf{x}_i, \boldsymbol{\theta}) \widehat{\mathbf{y}}(\mathbf{x}_i, \boldsymbol{\theta})^T | \mathbf{x}_i] W^T) + 2 \Tr\left(W \mathbb{E}[\widehat{\mathbf{y}}(\mathbf{x}_i; \boldsymbol{\theta}) \mathbf{s}_i^T | \mathbf{x}_i, s_i] \left(\widehat{P}_{s}^{(n)}\right)^{-1/2}\right) - 1$, where
$\widehat{P}_{s}^{(n)} = \frac{1}{n} \sum_{i=1}^n {\rm{diag}}(\mathbbm{1}_{\{s_i = 1\}}, \cdots, \mathbbm{1}_{\{s_i = k\}})$. Denote $\Psi(\boldsymbol{\theta}, W) = -\Tr(W P_{\hat{y}} W^T) + 2 \Tr(W P_{\hat{y}, s} P_{s}^{-1/2}) - 1$, where $P_{\hat{y}} = {\rm{diag}}(\mathbb{E} {\mathcal{F}}_1(\boldsymbol{\theta}, \mathbf{x}), \cdots, \mathbb{E}{\mathcal{F}}_m(\boldsymbol{\theta}, \mathbf{x}))$, $(P_{\hat{y}, s})_{j,r} = \mathbb{E}_{\mathbf{x}_i, s_i}[{\mathcal{F}}_j(\boldsymbol{\theta}, \mathbf{x}_i) \mathbf{s}_{i, r}]$ for $j \in [m], r\in [k]$, and $P_s = {\rm{diag}}(P_S(1), \cdots, P_S(k))$. Assume
$p_S(r) > 0$ for all $r \in [k]$. Then,
\[
\vspace{-.04in}
\small
\max_W \Psi(\boldsymbol{\theta}, W) = D_R(\widehat{Y}(\boldsymbol{\theta}); S)
\vspace{-.03in}
\]
\normalsize
and \[
\small
\vspace{-.03in}
\lim_{n \to \infty} \mathbb{E}[\widehat{\psi}_i^{(n)}(\boldsymbol{\theta}, W)] = \Psi(\boldsymbol{\theta}, W).
\vspace{-.004in}
\]
\normalsize
\end{proposition}
\vspace{-.05in}
The proof of~\cref{prop: consistency} is given in~\cref{app: pop frmi}. The first claim is immediate from~\cref{thm: min-max} and its proof, while the second claim is proved using the strong law of large numbers, the continuous mapping theorem, and Lebesgue's dominated convergence theorem.
\cref{prop: consistency}
implies that the empirical stochastic gradients computed in~\cref{alg: SGDA for FERMI} are good approximations of the true gradients of \cref{eq: FRMI}.
Intuitively, this suggests that when we use~\cref{alg: SGDA for FERMI} to solve the fair ERM problem~\cref{eq: FERMI}, the output of~\cref{alg: SGDA for FERMI} will also be an approximate solution of~\cref{eq: FRMI}. While~\cref{thm: SGDA for FRMI convergence rate} shows this intuition does indeed hold, the proof of~\cref{thm: SGDA for FRMI convergence rate} requires additional work. A reasonable first attempt at proving~\cref{thm: SGDA for FRMI convergence rate} might be to try to bound the expected distance between the gradient of $\text{FRMI}$ and the gradient of $\text{FERMI}$ (evaluated at the point $\hat{\boldsymbol{\theta}}$ that is output by~\cref{alg: SGDA for FERMI}) via Danskin's theorem~\citep{danskin1966theory} and strong concavity, and then leverage~\cref{thm: SGDA for FERMI convergence rate} to conclude that the gradient of $\text{FRMI}$ must also be small. However, the dependence of $\hat{\boldsymbol{\theta}}$ on the training data prevents us from obtaining a tight enough bound on the distance between the empirical and population gradients at $\hat{\boldsymbol{\theta}}$. Thus, we take a different approach to proving~\cref{thm: SGDA for FRMI convergence rate}, in which we consider the output of two different algorithms: one is the conceptual algorithm that runs one-pass~\cref{alg: SGDA for FERMI} as if we had access to the true sensitive attributes $P_s$ (``Algorithm A''); the other is the realistic one-pass~\cref{alg: SGDA for FERMI} that only uses the training data (``Algorithm B''). We argue: 1) the output of the conceptual algorithm is a stationary point of the population-level objective; and 2) the distance between the gradients of the population-level objective at $\boldsymbol{\theta}_A$ and $\boldsymbol{\theta}_B$ is small. While 1) follows easily from the proof of~\cref{thm: SGDA for FERMI convergence rate} and the online-to-batch conversion, establishing 2) requires a careful argument. The main tools we use in the proof of~\cref{thm: SGDA for FRMI convergence rate} are~\cref{thm: SGDA for FERMI convergence rate}, ~\cref{prop: consistency}, Danskin's theorem, Lipschitz continuity of the $\argmax$ function for strongly concave objective, the continuous mapping theorem, and Lebesgue's dominated convergence theorem: see~\cref{app: pop frmi} for the detailed proof.
Note that the online-to-batch conversion used to prove~\cref{thm: SGDA for FRMI convergence rate} requires a \textit{convergent stochastic optimization algorithm}; this implies that our arguments could not be used to prove an analogue of~\cref{thm: SGDA for FRMI convergence rate} for existing fair learning algorithms, since existing convergent fairness algorithms are not stochastic. An alternate approach to bounding the ``generalization error'' of our algorithm would be to use a standard covering/uniform convergence argument. However, this approach would not yield as tight a guarantee as~\cref{thm: SGDA for FRMI convergence rate}. Specifically, the accuracy and/or gradient complexity guarantee would depend on the dimension of the space (i.e. the number of model parameters), since the covering number depends (exponentially) on the dimension. For large-scale problems with a huge number of model parameters, such dimension dependence is prohibitive.
As previously mentioned, we can interpret \cref{thm: SGDA for FRMI convergence rate} as providing a guarantee that~\cref{alg: SGDA for FERMI} \textit{generalizes} well, achieving small fairness violation and test error, even on unseen ``test'' examples--as long as the data is i.i.d. and $N$ is sufficiently large. In the next section, we empirically corroborate~\cref{thm: SGDA for FRMI convergence rate}, by evaluating the fairness-accuracy tradeoffs of the FERMI algorithm (\cref{alg: SGDA for FERMI}) in several numerical experiments.
\vspace{-.07in}
\section{Numerical Experiments}
\label{sec:experiments}
\vspace{-.07in}
In this section, we evaluate the performance of FERMI in terms of the fairness violation vs. test error for different notions of fairness (e.g. demographic parity, equalized odds, and equality of opportunity). To this end, we perform diverse experiments comparing FERMI to other state-of-the-art approaches on several benchmarks.
In Section~\ref{sec: experiment binary-binary}, we showcase the performance of FERMI applied to a logistic regression model on binary classification tasks with binary sensitive attributes on Adult, German Credit, and COMPAS datasets. In
Section~\ref{sec: toxic_comment}, we utilize FERMI with a convolutional neural network base model for fair (to different religious groups) toxic comment detection. In Section~\ref{subsection: non-binary non-binary experiment}, we explore fairness in non-binary classification with non-binary sensitive attributes. Finally, Section~\ref{sec:color-MNIST} shows how FERMI may be used beyond fair empirical risk minimization in domain generalization problems to learn a model independent of spurious features.
\subsection{Fair Binary Classification with Binary Sensitive Attributes using Logistic Regression}
\label{sec: experiment binary-binary}
\subsubsection{Benchmarking full-batch performance}
\vspace{-.05in}
In the first set of experiments, we use FERMI to learn a fair logistic regression model on the Adult dataset. With the Adult data set, the task is to predict whether or not a person earns over \$50k annually without discriminating based on the sensitive attribute, gender. We compare FERMI against state-of-the-art in-processing full-batch ($|B| = N$) baselines, including~\citep{zafar2017fairness, feldman2015certifying, kamishima2011fairness, jiang2020wasserstein, hardt2016equality, prost2019toward, baharlouei2019rnyi, rezaei2020fairness, donini2018empirical, cho2020kde}. Since the majority of existing fair learning algorithms cannot be implemented with $|B| < N$, these experiments allow us to benchmark the performance of FERMI against a wider range of baselines. To contextualize the performance of these methods, we also include a {\bf Na\"ive Baseline} that randomly replaces the model output with the majority label ($0$ in Adult dataset), with probability $p$ (independent of the data), and sweep $p$ in $[0,1]$. At one end ($p=1$), the output will be provably fair with performance reaching that of a naive classifier that outputs the majority class. At the other end ($p=0$), the algorithm has no fairness mitigation and obtains the best performance (accuracy). By sweeping $p$, we obtain a tradeoff curve between performance and fairness violation.
\begin{figure*}[h!]
\begin{center}
\subfigure{%
\includegraphics[width=0.315\textwidth]{Figures/FullDP.png}
}
\subfigure{%
\includegraphics[width=0.315\textwidth]{Figures/FullOpportunity.png}
}
\subfigure{%
\includegraphics[width=0.315\textwidth]{Figures/FullEQO.png}
}
\end{center}
\vspace{-0.3in}
\caption{\small
Accuracy/Fairness trade-off of FERMI and several state-of-the-art in-processing approaches on Adult dataset.
FERMI offers the best fairness vs. accuracy tradeoff curve in all experiments against all baselines. \cite{rezaei2020fairness}~only allow for a single output and do not yield a tradeoff curve. Further, the algorithms by~\cite{mary19} and~\cite{baharlouei2019rnyi} are equivalent in this binary setting and shown by the red curve. In the binary/binary setting, FERMI, \cite{mary19} and \cite{baharlouei2019rnyi} all try to solve the same objective~\cref{eq: FRMI}. However, the empirical formulation \cref{eq: FERMI} and FERMI algorithm that we use results in better performance, even though we are using a full-batch for all baselines in this experiment.
}\label{fig: Experiment 1: PIC_EO}
\vspace{-.1in}
\end{figure*}
\vspace{-.05in}
In \cref{fig: Experiment 1: PIC_EO}, we report the fairness violation (demographic parity, equalized odds, and equality of opportunity violations) vs. test error of the aforementioned in-processing approaches on the Adult dataset. The upper left corner of the tradeoff curves coincides with the unmitigated baseline, which only optimizes for performance (smallest test error).
As can be seen, FERMI offers a fairness-accuracy tradeoff curve that dominates all state-of-the-art baselines in each experiment and with respect to each notion of fairness, even in the full batch setting.
Aside from in-processing approaches, we compare FERMI with several pre-processing and post-processing algorithms on Adult, German Credit, and COMPAS datasets in Appendix~\ref{app: complete figure 1}, where we show that the tradeoff curves obtained from FERMI dominate that of all other baselines considered. See Appendix~\ref{appendix: experiment details} for details on the data sets and experiments.
It is noteworthy that the empirical objectives of~\cite{mary19} and~\cite{baharlouei2019rnyi} are exactly the same in the binary/binary setting, and their algorithms also coincide to the red curve in \cref{fig: Experiment 1: PIC_EO}. This is because Exponential R\'enyi mutual information is equal to R\'enyi correlation for binary targets and/or binary sensitive attributes (see~\cref{thm:Renyi-correlation}), which is the setting of all experiments in \cref{sec: experiment binary-binary}. Additionally, like us, in the binary/binary setting these works are trying to empirically solve~\cref{eq: FRMI}, albeit using different estimation techniques; i.e., their empirical objective is different from~\cref{eq: FERMI}.
This demonstrates the effectiveness of our empirical formulation (\ref{eq: FERMI}) and our solver (\cref{alg: SGDA for FERMI}), even though we are using all baselines in full batch mode in this experiment. See \cref{app: complete figure 1} for the complete version of \cref{fig: Experiment 1: PIC_EO} which also includes pre-processing and post-processing baselines.
\cref{fig: outliers} in~\cref{appendix: experiment details} illustrates that FERMI outperforms baselines in the presence of \textit{noisy outliers} and \textit{class imbalance}. Our theory did not consider the role of noisy outliers and class imbalance, so the theoretical investigation of this phenomenon could be an interesting direction for future work.
\vspace{-.08in}
\subsubsection{The effect of batch size on fairness/accuracy tradeoffs}
\label{sec: small batch}
\vspace{-.07in}
Next, we evaluate the performance of FERMI on smaller batch sizes ranging from $1$ to $64$. To this end, we compare FERMI against several state-of-the-art in-processing algorithms that permit stochastic implementation for demographic parity:~\citep{mary19}, ~\citep{baharlouei2019rnyi}, and~\citep{cho2020kde}. Similarly to the full batch setting, for all methods, we train a logistic regression model with a respective regularizer for each method. We use demographic parity $L_\infty$ violation (\cref{def-conditional-DP}) to measure demographic parity violation. More details about the dataset and experiments, and additional experimental results, can be found in Appendix~\ref{appendix: experiment details}.
\begin{figure*}[ht]
\vspace{-.1in}
\begin{center}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/BS1.png}
}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/BS4.png}
}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/BS16.png}
}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/BS64.png}
}
\end{center}
\vspace{-.17in}
\caption{\small
Performance of FERMI,~\citet{cho2020fair},~\citet{mary19},~\cite{baharlouei2019rnyi} with different batch-sizes on Adult dataset. FERMI demonstrates the best accuracy/fainess tradeoff across different batch sizes. \normalsize}
\label{fig: adult_batch}
\vspace{-.1in}
\end{figure*}
\cref{fig: adult_batch} shows that FERMI offers a superior fairness-accuracy tradeoff curve against all baselines, for each tested batch size, empirically confirming~\cref{thm: SGDA for FERMI convergence rate}, as FERMI is the only algorithm that is guaranteed to converge for small minibatches. It is also noteworthy that all other baselines cannot beat \textit{Na\"ive Baseline} when the batch size is very small, e.g., $|B|= 1.$ Furthermore, FERMI with $|B| = 4$ almost achieves the same fairness-accuracy tradeoff as the full batch variant.
\vspace{-.05in}
\subsubsection{The effect of missing sensitive attributes on fairness/accuracy tradeoffs}
\vspace{-.07in}
Sensitive attributes might be partially unavailable in many real-world applications due to legal issues, privacy concerns, and data gathering limitations~\citep{zhao2022towards, coston2019fair}. Missing sensitive attributes make fair learning tasks more challenging in practice.
\begin{wrapfigure}{r}{0.35\textwidth}
\begin{center}
\centerline{\includegraphics[width=0.33\columnwidth]{Figures/Sensitive90.png}}
\vspace{-4mm}
\caption{\small Performance of FERMI and other state-of-the-art approaches on the Adult dataset where $90\%$ of gender entries are missing. Full-sensitive FERMI is obtained by applying FERMI on the data without any missing entries.}
\label{fig: missing}
\end{center}
\end{wrapfigure}
The unbiased nature of the estimator used in FERMI algorithm motivates that it may be able to handle cases where sensitive attributes are \textit{partially} available and are dropped uniformly at random. As a case study on the Adult dataset, we randomly masked $90\%$ of the sensitive attribute (i.e., gender entries). To estimate the fairness regularization term, we rely on the remaining $10\%$ of the training samples ($\approx 3k$) with sensitive attribute information. Figure~\ref{fig: missing} depicts the tradeoff between accuracy and fairness (demographic parity) violation for FERMI and other baselines. We suspect that the superior accuracy-fairness tradeoff of FERMI compared to other approaches is due to the fact that the estimator of the gradient remains unbiased since the missing entries are missing completely at random (MCAR). Note that the \textit{Na\"ive Baseline} is similar to the one implemented in the previous section, and \textit{Full-sensitive FERMI} is an oracle method that applies FERMI to the data with no missing attributes (for comparison purposes only). We observe that FERMI achieves a slightly worse fairness-accuracy tradeoff compared to Full-sensitive FERMI oracle, whereas the other baselines are hurt significantly and are only narrowly outperforming the Na\"ive Baseline.
\vspace{-0.1in}
\subsection{Fair Binary Classification using Neural Models}
\label{sec: toxic_comment}
\begin{figure}[h!]
\vspace{-.25in}
\begin{center}
\subfigure{%
\centering
\includegraphics[width=0.32\textwidth]{Figures/Comment128Acc.png}
} \hspace{0.1\textwidth}
\subfigure{%
\centering
\includegraphics[width=0.32\textwidth]{Figures/Comment16Acc.png}
}
\end{center}
\vspace{-.15in}
\caption{\small Fair toxic comment detection with different batch sizes. For $|B|= 128$, the performance of~\citep{prost2019toward} and FERMI are close to each other, however, when the batch size is reduced to 16, FERMI demonstrates a better fairness/ performance trade-off. The performance and fairness are measured by the test error and the false positive gap between different religious sub-groups (Christians vs Muslim-Jews), respectively.
}%
\label{fig: toxic_comment}
\vspace{.05in}
\end{figure}
\vspace{-0.1in}
In this experiment, our goal is to showcase the efficacy of FERMI in stochastic optimization with neural network function approximation. To this end, we apply FERMI,~\citep{prost2019toward},~\citep{baharlouei2019rnyi}, and~\citep{mary19} (which coincides with~\citep{baharlouei2019rnyi}) to the Toxic Comment Classification dataset where the underlying task is to predict whether a given published comment in social media is toxic. The sensitive attribute is religion that is binarized into two groups: Christians in one group; Muslims and Jews in the other group. Training a neural network without considering fairness leads to higher false positive rate for the Jew-Muslim group. Figure~\ref{fig: toxic_comment} demonstrates the performance of FERMI, MinDiff~\citep{prost2019toward},~\citet{baharlouei2019rnyi}, and na\"ive baseline on two different batch-sizes: $128$ and $16$. Performance is measured by the overall false positive rate of the trained network and fairness violation is measured by the false positive gap between two sensitive groups (Christians and Jews-Muslims). The network structure is exactly same as the one used by MinDiff~\citep{prost2019toward}. We can observe that by decreasing the batch size, FERMI maintains the best fairness-accuracy tradeoff compared to other baselines.
\subsection{Fair Non-binary Classification with Multiple Sensitive Attributes}
\label{subsection: non-binary non-binary experiment}
\vspace{-.06in}
\begin{figure*}[t]
\begin{center}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/fig2_nb_dp02.png}
}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/fig2_nb_dp03.png}
}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/fig2_nb_dp04.png}
}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/fig2_nb_dp05.png}
} \\
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/fig2_nb_eo02.png}
}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/fig2_nb_eo03.png}
}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/fig2_nb_eo04.png}
}
\subfigure{%
\includegraphics[width=0.235\textwidth]{Figures/fig2_nb_eo05.png}
}
\end{center}
\vspace{-.17in}
\caption{\small
Comparison between FERMI, \cite{mary19}, \cite{baharlouei2019rnyi}, and \cite{cho2020kde} on Communities dataset.
\citep{mary19} outperforms \citep{baharlouei2019rnyi, cho2020kde}, which we believe could be attributed to the effectiveness of ERMI as a regularizer. FERMI outperforms \cite{mary19}. This can be attributed to our empirical formulation \cref{eq: FERMI} and unbiased stochastic optimization algorithm.
}
\label{fig:non-binary}
\end{figure*}
In this section, we consider a non-binary classification problem with multiple binary sensitive attributes. In this case, we consider the Communities and Crime dataset,
which has 18 binary sensitive attributes in total. For our experiments, we pick a subset of $1, 2, 3, \ldots, 18$ sensitive attributes, which corresponds to $|\mathcal{S}| \in \{2, 4, 8, \ldots, 2^{18}\}$.
We discretize the target into three classes $\{\text{high}, \text{medium}, \text{low}\}$.
The only baselines that we are aware of that can handle non-binary classification with multiple sensitive attributes are \citep{mary19}, \citep{baharlouei2019rnyi}, \citep{cho2020kde}, \citep{cho2020fair}, and \citep{zhang2018mitigating}. We used the publicly available implementations of~\citep{baharlouei2019rnyi} and \citep{cho2020kde} and extended their binary classification algorithms to the non-binary setting.
The results are presented in~\cref{fig:non-binary}, where we use conditional demographic parity $L_\infty$ violation (\cref{def-conditional-DP}) and conditional equal opportunity $L_\infty$ violation (\cref{def-conditional-EO}) as the fairness violation notions for the two experiments. In each panel, we compare the test error for different number of sensitive attributes for a fixed value of DP violation.
It is expected that test error increases with the number of sensitive attributes, as we will have a more stringent fairness constraint to satisfy.
As can be seen, compared to the baselines, FERMI offers the most favorable test error vs. fairness violation tradeoffs, particularly as the number of sensitive attributes increases and for the more stringent fairness violation levels, e.g., $0.02$.\footnote{\cref{sec:color-MNIST} demonstrated that using smaller batch sizes results in much more pronounced advantages of FERMI over these baselines.}
\subsection{Beyond Fairness: Domain Parity Regularization for Domain Generalization}
\label{sec:color-MNIST}
\vspace{-0.07in}
In this section, we demonstrate that our approach may extend beyond fair empirical risk minimization to other problems, such as domain generalization. In fact, \cite{li2019repair, lahoti2020fairness,creager2021environment} have already established connections between fair ERM and domain generalization.
We consider the Color MNIST dataset~\citep{li2019repair}, where all 60,000 training digits are colored with different colors drawn from a class conditional Gaussian distribution with variance $\sigma^2$ around a certain average color for each digit, while the test set remains black and white. \cite{li2019repair} show that as $\sigma^2 \to 0,$ a convolutional network model overfits significantly to each digit's color on the training set, and achieves vanishing training error. However, the learned representation does not generalize to the black and white test set, due to the spurious correlation between digits and color.
Conceptually, the goal of the classifier in this problem is to achieve \textit{high classification accuracy with predictions that are independent of the color of the digit}. We view color as the sensitive attribute in this experiment and apply fairness baselines for the demographic parity notion of fairness. One would expect that by promoting such independence through a fairness regularizer, the generalization would improve (i.e., lower test error on the black and white test set), at the cost of increased training error (on the colored training set). We compare against \cite{mary19}, \cite{baharlouei2019rnyi}, and \cite{cho2020kde} as baselines in this experiment.
The results of this experiment are illustrated in \cref{fig:color-MNIST-results}.
In the
left panel,
we see that with no regularization ($\lambda = 0$), the test error is around 80\%. As $\lambda$ increases, all methods achieve smaller test errors while training error increases. We also observe that \textit{FERMI offers the best test error} in this setup. In the
right panel,
we observe that decreasing the batch size results in significantly worse generalization for the three baselines considered (due to their biased estimators for the regularizer). However, the negative impact of small batch size is much less severe for FERMI, since FERMI uses unbiased stochastic gradients. In particular, \textit{the performance gap between FERMI and other baselines is more than 20\% for $|B| = 64$}. Moreover, \textit{FERMI with minibatch size $|B|=64$ still outperforms all other baselines with $|B|>1,000$}. Finally, notice that the test error achieved by FERMI when $\sigma = 0$ is $\sim30\%$, as compared to more than $50\%$ obtained using REPAIR~\citep{li2019repair} for $\sigma \leq 0.05$.
\begin{figure}[t]
\vspace{-.1in}
\begin{center}
\subfigure{%
\centering
\includegraphics[width=0.35\textwidth]{Figures/fig3_cmnist_train_test.png}
} \hspace{0.1\textwidth}
\subfigure{%
\centering
\includegraphics[width=0.38\textwidth]{Figures/BatchSize.png}
}
\end{center}
\vspace{-.19in}
\caption{\small Domain generalization on Color MNIST~\citep{li2019repair} using in-process fair algorithms for demographic parity.
{\bf
Left panel:} The dashed line is the training error and the solid line is the test error. As $\lambda$ increases, fairness regularization results in a learned representation that is less dependent on color; hence training error increases while test error decreases (all algorithms reach a plateau around $\lambda = 8$). We use $|B| = 512$ for all baselines. {\bf
Right panel:} We plot test error vs. batch size using an optimized value of $\lambda$ for each algorithm selected via a validation set. The performance of baselines drops 10-20\% as the batch size becomes small, whereas FERMI is less sensitive to batch size.}%
\label{fig:color-MNIST-results}
\vspace{-.2in}
\end{figure}
\section{Discussion and Concluding Remarks}
\label{sec:discussion}
\vspace{-.07in}
In this paper, we tackled the challenge of developing a fairness-promoting algorithm that is amenable to stochastic optimization. As discussed, algorithms for large-scale ML problems are constrained to use stochastic optimization with (small) minibatches of data in each iteration. To this end, we formulated an empirical objective (\ref{eq: FERMI}) using ERMI as a regularizer and derived unbiased stochastic gradient estimators. We proposed the stochastic FERMI algorithm (\cref{alg: SGDA for FERMI}) for solving this objective.
We then provided the \textit{first theoretical convergence guarantees for a stochastic in-processing fairness algorithm}, by showing that FERMI
converges to stationary points of the empirical and population-level objectives (\cref{thm: SGDA for FERMI convergence rate}, \cref{thm: SGDA for FRMI convergence rate}). Further, these convergence results hold even for non-binary sensitive attributes and non-binary target variables, with any minibatch size.
From an experimental perspective, we showed that \textit{FERMI leads to better fairness-accuracy tradeoffs than all of the state-of-the-art baselines} on a wide variety of binary and non-binary classification tasks (for demographic parity, equalized odds, and equal opportunity). We also showed that these benefits are particularly significant when the number of sensitive attributes grows or the batch size is small. In particular, we observed that FERMI consistently outperforms \cite{mary19} (which tries to
solve the same objective~\cref{eq: FRMI}) by up to 20\% when the batch size is small.
This is not surprising since FERMI is the only algorithm that is guaranteed to find an approximate solution of the fair learning objective with any batch size $|B| \geq 1$. Also, we show in Fig.~\ref{fig: convergence_comparison} that the lack of convergence guarantee of \cite{mary19} is not just due to more limited analysis: in fact, their stochastic algorithm does not converge.
Even in full batch mode, FERMI outperforms all baselines, including \citep{mary19} (Fig.~\ref{fig: Experiment 1: PIC_EO}, Fig.~\ref{fig:non-binary}). In full batch mode, all baselines should be expected to converge to an approximate solution of their respective empirical objectives, so this suggests that our empirical objective \cref{eq: FERMI} is fundamentally better, in some sense than the empirical objectives proposed in prior works. In what sense is \cref{eq: FERMI} a better empirical objective (apart from permitting stochastic optimization)? For one, it is an asymptotically unbiased estimator of~\cref{eq: FRMI} (by~\cref{prop: consistency}), and~\cref{thm: SGDA for FRMI convergence rate} suggests that
FERMI algorithm outputs an approximate solution of~\cref{eq: FRMI} for large enough $N$.
By contrast, the empirical objectives considered in prior works do not provably yield an approximate solution to the corresponding population-level objective.
The superior fairness-accuracy tradeoffs of FERMI algorithm over the (full batch) baselines also suggest that the underlying
population-level objective~\cref{eq: FRMI} has benefits over other fairness objectives. What might these benefits be? First, ERMI upper bounds all other fairness violations (e.g., Shannon mutual information, $L_q$, $L_{\infty}$) used in the literature: see~\cref{appendix: relations btwn ERMI and fairness}. This implies that ERMI-regularized training yields a model that has a small fairness violation with respect to these other notions.
Could this also somehow help explain the superior fairness-accuracy tradeoffs achieved by FERMI? Second, the objective function~\cref{eq: FRMI} is easier to optimize than the objectives of competing in-processing methods: ERMI is smooth and is equal to the trace of a matrix (see \cref{lem: D_R = Tr(Q^TQ)} in the Appendix), which is easy to compute. Contrast this with the larger computational overhead of R\'enyi correlation used by \cite{baharlouei2019rnyi}, for example, which requires finding the second singular value of a matrix. Perhaps these computational benefits contribute to the observed performance gains.
We leave it as future work to rigorously understand the factors that are most responsible for the favorable fairness-accuracy tradeoffs observed from FERMI.
\section*{Broader Impact and Limitations}
\vspace{-.07in}
This paper studied the important problem of developing practical machine learning (ML) algorithms that are \textit{fair} towards different demographic groups (e.g. race, gender, age). We hope that the societal impacts of our work will be positive, as the deployment of our FERMI algorithm may enable/help companies, government agencies, and other organizations to train large-scale ML models that are fair to all groups of users. On the other hand, any technology has its limitations, and our algorithm is no exception.
One important limitation of our work is that we have (implicitly) assumed that the data set at hand is labeled accurately and fairly. For example, if race is the sensitive attribute and ``likelihood of default on a loan'' is the target, then we assume that the training data based on past observational data accurately reflects the financial histories of all individuals (and in particular does not disproportionately inflate the financial histories of racial minorities). If this assumption is not satisfied in practice, then the outcomes promoted by our algorithm may not be as fair (in the philosophical sense) as the computed level of fairness violation might suggest.
For example, if racial minorities are identified as higher risk for default on loans, they may be extended loans with higher interest rates and payments, which may in turn increase their likelihood of a default.
Hence, it might be even possible that our mitigation strategy could result in more unfairness than unmitigated ERM in this case. More generally, conditional fairness notions like equalized odds suffer from a potential amplification of the inherent discrimination that may exist in the training data. Tackling such issues is beyond the scope of this work; c.f.~\cite{kilbertus} and \cite{bechavod19}.
Another consideration that was not addressed in this paper is the interplay between fairness and other socially consequential AI metrics, such as privacy and robustness (e.g. to data poisoning). It is possible that our algorithm could increase the impact of data from certain individuals to improve fairness at the risk of leaking private information about individuals in the training data set (e.g. via membership inference attacks or model inversion attacks), even if the data is anonymous~\citep{inversionfred, shokri2017membership, korolova2018facebook, nasr2019comprehensive, carlini2021extracting}. \textit{Differential privacy}~\citep{dwork2006calibrating} ensures that sensitive data cannot be leaked (with high probability), and the interplay between fairness and privacy has been explored (see e.g. \cite{jagielski2019differentially, xu2019achieving, cummings, mozannar2020fair, tranfairnesslens, tran2021differentially}. Developing and analyzing a differentially private version of FERMI could be an interesting direction for future work. Another potential threat to FERMI-trained models is data poisoning attacks. While our experiments demonstrated that FERMI is relatively effective with missing sensitive attributes, we did not investigate its performance in the presence of label flipping or other poisoning attacks. Exploring and improving the robustness of FERMI is another avenue for future research.
\subsection*{Acknowledgment:}
The authors are thankful to James Atwood (Google Research), Alex Beutel (Google Research), Jilin Chen (Google Research), and Flavien Prost (Google Research) for constructive discussions and feedback that helped shape up this paper.
|
\section{Introduction}
In Renyi's {\em A Diary on Information Theory}, \cite{Renyi},
he points out that the entropy formula, i.e.,
$$H(X)=p_1\log\frac{1}{p_1}+p_2\log\frac{1}{p_2}+\cdots+p_N\log\frac{1}{p_N},$$
where logs are to the base~2,
was arrived at, independently, by
Claude Shannon and Norbert Wiener in 1948.
This famous formula was the revolutionary precursor of the information age.
Renyi goes on to say that,
\begin{itemize}
\item[]
this formula had already appeared in the work of
Boltzmann which is why it is also called the
Boltzmann-Shannon Formula.
Boltzmann arrived at this formula in connection
with a completely different problem.
Almost half a century before Shannon
he gave essentially the same formula to describe entropy
in his investigations of statistical mechanics.
He showed that if, in a gas
containing a large number of molecules
the probabilities of the
possible states of the individual molecules are
$p_1,p_2,\ldots,p_N$
then the entropy
of the system is
$H=c(p_1\log\frac{1}{p_1}+p_2\log\frac{1}{p_2}+\cdots+p_N\log\frac{1}{p_N})$
where $c$ is a constant.
(In statistical mechanics the natural logarithm is used
and not base 2 \ldots).
The entropy of a physical system is the measure of its disorder
\ldots
\end{itemize}
Since 1948 there have been many advances in information theory
and entropy. A well-known paper of Dembo, Cover and Thomas
is devoted to inequalities in information theory.
Here, we concentrate on equalities.
We show how a
Shannon function
$H(p,q)$
can be expanded in infinitely many ways in an
infinite series of functions
each of which is a linear combination of
Shannon functions of the type $H\bigl(f(p),g(q)\bigr)$,
where $f,g$ are quotients of polynomials of degree $n$ for any
$n\ge 2$.
Apart from its intrinsic interest,
this new result gives insight into the algorithm
in Section~\ref{section:ShannonLimitAppnsCrypt}
for constructing a common secret key between two
communicating parties ---
see also \cite{Bennett}, \cite{BruenForcinitoMcQuillan}, \cite{BruenForcinitoWehlau}, \cite{MaurerWolf}.
\section{Extensions of a Binary Symmetric Channel}
Recall that a binary symmetric channel
has input and output symbols drawn from
$\{0,1\}$.
We say that there is a common probability
$q=1-p$ of any symbol being transmitted incorrectly,
independently for each transmitted symbol,
$0\le p\le 1$.
We use the channel matrix
$P=\left(
\begin{array}{cc}
p&q\\
q&p
\end{array}\right)$.
Again, $p$ is the {\em probability of success},
i.e., $p$ denotes the probability that 0 is transmitted to 0
and also the probability that 1 gets transmitted to 1.
The {\em second extension} $P^{(2)}$ of $P$
has alphabet
$\{00,01,10,11\}$
and channel matrix
$$P^{(2)}=\left(\begin{array}{cccc}
p^2&pq&qp&q^2\\
pq&p^2&q^2&qp\\
qp&q^2&p^2&pq\\
q^2&qp&pq&p^2
\end{array}\right)
=\left(\begin{array}{cc}
pP&qP\\
qP&pP
\end{array}\right).$$
An alternative way to think of an
$n^{th}$ extension of a channel $C$
(see Welsh~\cite{Welsh})
is to regard it as
$n$ copies of $C$ acting independently and in parallel.
See Figure~\ref{figure:nCopiesOfC}.
\begin{figure}[!ht]
\begin{center}
\includegraphics[trim={20 355 20 160},clip,scale=.75]{nCopiesOfChannel.pdf}
\caption{$n$ copies of $C$ acting independently and in parallel.
\label{figure:nCopiesOfC}}
\end{center}
\end{figure}
Let us assume also that, for $C$, the input probability of 0
and the input probability of 1 are both equal to
$\frac{1}{2}$.
\begin{theoremN}
\label{theorem:XYInputOutput}
Let $\mathbf{X}=(X_1,\ldots,X_n)$,
and $\mathbf{Y}=(Y_1,\ldots,Y_n)$
denote an input-output pair for $C^{(n)}$.
Then
\begin{enumerate}[(a)]
\item
$H(\mathbf{X})=H(X_1,\ldots,X_n)=n$
\item
$H(\mathbf{Y})=H(Y_1,\ldots,Y_n)=n$
\item
$H(\mathbf{X}|\mathbf{Y})$ is equal to $nH(p,q)$.
\item
The capacity of $C^{(n)}$ is $n\bigl(1-H(p,q)\bigr)$.
\end{enumerate}
\end{theoremN}
\begin{proofN}
(a)
Since, by definition, the $X_i$ are independent,
$$H(\mathbf{X})=H(X_1)+H(X_2)+\cdots+H(X_n).$$
We have
\begin{align}
H(X_i)&=-\biggl[Pr(X_i=0)\log Pr(X_i=0)+Pr(X_i=1)\log Pr(X_i=1)\biggr]\nonumber\\
&=-\biggl[\frac{1}{2}\log\bigl(\frac{1}{2}\bigr)+\frac{1}{2}\log\bigl(\frac{1}{2}\bigr)\biggr]\nonumber\\
&=-\log\bigl(\frac{1}{2}\bigr)\nonumber\\
&=-[\log 1-\log 2]=\log 2=1.\nonumber
\end{align}
Thus $H(\mathbf{X})=n$.
\noindent
(b) The value of $Y_i$ only depends on $X_i$.
The $X_i$ are independent.
Thus $Y_1,Y_2,\ldots,Y_n$ are independent.
For any $Y_i$ we have
$$Pr(Y_i=0)=Pr(X_i=0)(p)+Pr(X_i=1)(q)=\frac{1}{2}(p+q)=\frac{1}{2}.$$
Also $Pr(Y_i=1)=\frac{1}{2}$.
Then, as for $X_i$, $H(Y_i)=1$.
Thus $H(\mathbf{Y})=H(Y_1,Y_2,\ldots,Y_n)=\sum_i H(Y_i)=n$.
\noindent
(c)
We have
$H(\mathbf{X})-H(\mathbf{X}|\mathbf{Y})
=H(\mathbf{Y})-H(\mathbf{Y}|\mathbf{X})$.
Since $H(\mathbf{X})=H(\mathbf{Y})=n$
we have $H(\mathbf{X}|\mathbf{Y})=H(\mathbf{Y}|\mathbf{X})$.
Now
$$H(\mathbf{Y}|\mathbf{X})=\sum_{\mathbf{x}}Pr(\mathbf{x})H(\mathbf{Y}\,|\,\mathbf{X}=\mathbf{x}),$$
where $\mathbf{x}$ denotes a given value of the random vector $\mathbf{X}$.
Since the channel is memoryless,
$$H(\mathbf{Y}\,|\, \mathbf{X}=\mathbf{x})
=\sum_i H(Y_i\,|\, \mathbf{X}=\mathbf{x})
=\sum_i H(Y_i\, |\, X_i=x_i).$$
The last step needs a little work ---
see \cite{JJ} Exercise 4.10 or \cite{McEliece}
or \cite{BruenForcinitoMcQuillan} or the example below
for details.
Then
\begin{align}
H(\mathbf{Y}\,| \,\mathbf{X})
&=\sum_{\mathbf{x}}Pr(\mathbf{x})\sum_i H(Y_i\,|\, X_i=x_i)\nonumber\\
&=\sum_i\sum_u H(Y_i\,| \, X_i=u)Pr(X_i=u).\nonumber
\end{align}
Thus
$$H(\mathbf{Y}\,|\,\mathbf{X})=\sum_{i=1}^nH(Y_i\,|\, X_i)=nH(p,q)
=H(\mathbf{X}\,|\,\mathbf{Y}).$$
\medskip\noindent
{\bf Example}\ \
Let $n=2$ and
$\left(\begin{array}{c}x_1\\ x_2\end{array}\right)=
\left(\begin{array}{c}0\\ 1\end{array}\right)$.
Then
$\left(\begin{array}{c}y_1\\ y_2\end{array}\right)=
\left(\begin{array}{c}1\\ 0\end{array}\right)$ or
$\left(\begin{array}{c}0\\ 1\end{array}\right)$ or
$\left(\begin{array}{c}1\\ 1\end{array}\right)$ or
$\left(\begin{array}{c}0\\ 0\end{array}\right)$.
Using independence we get that
the coresponding entropy term is
$-[q^2\log q^2 + p^2\log p^2 + qp\log qp + pq\log pq]$.
This simplifies to
$-2[p\log p + q\log q]=2H(p,q)$.
Note that the probability that
$\left(\begin{array}{c}x_1\\ x_2\end{array}\right)=
\left(\begin{array}{c}0\\ 1\end{array}\right)$
is $\frac{1}{4}$.
\medskip
\noindent
(d)
The capacity of $C$
is the maximum value, over all
inputs, of $H(X)-H(X\,|\, Y)$.
Since $X$ is random, the input probability of a 1 or 0 is 0.5.
This input distribution maximizes $H(X)-H(X\,|\, Y)$ for $C$,
the maximum value being $1-H(p,q)$.
Then the capacity of $C^{(n)}$ is $n(1-H(p,q))$.
It represents the information about
$\mathbf{X}$ conveyed by $\mathbf{Y}$ or the amount of information
about $\mathbf{Y}$ conveyed by $\mathbf{X}$.
\end{proofN}
\section{An Entropy Equality}
\label{section:basicEntropyEquality}
First we need some additional discussion on entropy.
\bigskip\noindent
{\bf A. Extending a basic result.}
\medskip\noindent
A fundamental result for random variables $X,Y$ is that
$H(X)+H(Y|X)=H(Y)+H(X|Y)$.
A corresponding argument may be used to
establish similar identities involving more than
two random variables.
For example,
\begin{align}
H(X,Y,Z)
&=H(X)+H(Y|X)+H(Z\,|\,X,Y)\nonumber\\
&=H(X,Y)+H(Z\, |\, X,Y).\nonumber
\end{align}
Also
$$H(X,Y,Z)=H(X)+H(Y,Z\, |\, X).$$
\bigskip\noindent
{\bf B. From Random Variables to Random Vectors.}
\medskip\noindent
For any random variable $X$ taking only a finite number
of values with probabilities $p_1,p_2,\ldots,\allowbreak p_n$
such that
$$\sum p_i=1\qquad\hbox{and}\qquad p_i>0\quad(1\le i\le n),$$
we define the
entropy
of $X$ using the Shannon formula
$$H(X)=-\sum_{k=1}^{n}p_k\log p_k=\sum_{k=1}^n p_k\log\frac{1}{p_k}.$$
Analogously, if $\mathbf{X}$
is a {\em random vector} which takes only a finite number
of values
$\mathbf{u_1},\mathbf{u_2},\ldots,\allowbreak\mathbf{u_m}$,
we define its
entropy
by the formula
$$H(\mathbf{X})=-\sum_{k=1}^m Pr(\mathbf{u_k})\log Pr(\mathbf{u_k}).$$
For example, when $\mathbf{X}$ is a
two-dimensional random vector, say
$\mathbf{X}=(U,V)$ with
$p_{ij}=Pr(U=a_i, V=b_j)$, then we can write
$$H(\mathbf{X})=H(U,V)=-\sum_{i,j}p_{ij}\log p_{ij}.$$
Note that $\sum_{i,j} p_{ij}=1$.
More generally, if $X_1,X_2,\ldots,X_m$
is a collection of random variables each taking
only a finite set of values,
then we can regard
$\mathbf{X}=(X_1,X_2,\ldots,X_m)$
as a random vector
taking a finite set of values and we define the
{\bf joint entropy}
of $X_1,\ldots,X_m$ by
{\small
\begin{align}
&H(X_1,X_2,\ldots,X_m)\nonumber\\
&=H(\mathbf{X})\nonumber\\
&=-\sum Pr(X_1=x_1,X_2=x_2,\ldots,X_m=x_m)\log Pr(X_1=x_1,X_2=x_2,\ldots,X_m=x_m).
\nonumber
\end{align}
}
Standard results for random variables then carry over to
random vectors ---
see \cite{Ash}, \cite{Welsh}.
\bigskip\noindent
{\bf C. The Grouping Axiom for Entropy.}
\medskip\noindent
This axiom
or identity can shorten calculations.
It reads as follows
(\cite[p. 2]{Welsh}, \cite[p. 8]{Ash},
\cite[Section 9.6]{BruenForcinitoMcQuillan}).
Let $p=p_1+p_2+\cdots+p_m$
and $q=q_1+q_2+\cdots+q_n$
where each $p_i$ and $q_j$ is non-negative.
Assume that $p,q$ are positive with $p+q=1$.
Then
\begin{align}
&
H(p_1,p_2,\ldots,p_m,q_1,q_2,\ldots,q_n)\nonumber\\
&\ \
=H(p,q)+pH\left(\frac{p_1}{p},\frac{p_2}{p},\ldots,\frac{p_m}{p}\right)
+qH\left(\frac{q_1}{q},\frac{q_2}{q},\ldots,\frac{q_n}{q}\right).
\nonumber
\end{align}
For example, suppose $m=1$ so $p_1=p$.
Then we get
$$H(p_1,q_1,q_2,\ldots,q_n)=H(p,q)
+qH\left(\frac{q_1}{q},\ldots,\frac{q_n}{q}\right).$$
This is because
$p_1H\bigl(\frac{p_1}{p_1},0\bigr)=p_1H(1,0)=p_1(1\log 1)=p_1(0)=0$.
\bigskip
\begin{theoremN}
\label{theorem:z;xy}
Let $\mathbf{X},\mathbf{Y},\mathbf{Z}$ be random vectors
such that
$H(\mathbf{Z}\, |\, \mathbf{X},\mathbf{Y})=0$.
Then
\begin{enumerate}[(a)]
\item
$H(\mathbf{X}|\mathbf{Y})=H(\mathbf{X},\mathbf{Z}\, |\, \mathbf{Y})$.
\item
$H(\mathbf{X}|\mathbf{Y})=H(\mathbf{X}\,|\, \mathbf{Y},\mathbf{Z})+
H(\mathbf{Y}\, |\, \mathbf{Z})$.
\end{enumerate}
\end{theoremN}
\begin{proofN}
\begin{align}
H(\mathbf{X}\, |\, \mathbf{Y})
&=H(\mathbf{X},\mathbf{Y})-H(\mathbf{Y})\nonumber\\
&=H(\mathbf{X},\mathbf{Y},\mathbf{Z})-H(\mathbf{Z}\, |\, \mathbf{X},\mathbf{Y})-H(\mathbf{Y})\nonumber\\
&=H(\mathbf{X},\mathbf{Y},\mathbf{Z})-H(\mathbf{Y})
\qquad\hbox{[\,since $H(\mathbf{Z}\,|\,\mathbf{X},\mathbf{Y})=0$\,]}
\nonumber\\
&=H(\mathbf{Y})+H(\mathbf{X},\mathbf{Z}\, |\, \mathbf{Y})-H(\mathbf{Y})\nonumber\\
&=H(\mathbf{X},\mathbf{Z}\, |\, \mathbf{Y}),\qquad\hbox{proving (a).}\nonumber
\end{align}
For (b),
\begin{align}
H(\mathbf{X}\, |\, \mathbf{Y})
&=H(\mathbf{X},\mathbf{Z},\mathbf{Y})-H(\mathbf{Y})\hbox{ from (a)}\nonumber\\
&=H(\mathbf{X},\mathbf{Z},\mathbf{Y})-H(\mathbf{Y},\mathbf{Z})
+H(\mathbf{Y},\mathbf{Z})-H(\mathbf{Y})\nonumber\\
&=H(\mathbf{X}\, |\, \mathbf{Y},\mathbf{Z})+H(\mathbf{Z}\, |\, \mathbf{Y}).\nonumber
\end{align}
\end{proofN}
\section{The New Identities}
\label{section:NewIdentities}
We will use the above identity,
i.e.,
\begin{equation}
\label{equation:BasicIdentity}
H(\mathbf{X}\, |\, \mathbf{Y})
=H(\mathbf{X}\, |\, \mathbf{Y},\mathbf{Z})
+H(\mathbf{Z}\, |\, \mathbf{Y})
\end{equation}
which holds under the assumption that
$H(\mathbf{Z}\, |\, \mathbf{X},\mathbf{Y})=0$.
We begin with arrays
$A=\left(\begin{array}{c}
a_1\\ a_2\\ \vdots \\ a_n
\end{array}\right)$,
$B=\left(\begin{array}{c}
b_1\\ b_2\\ \vdots \\ b_n
\end{array}\right)$,
where $n$ is even.
We assume that $A,B$ are random binary strings
subject to the condition that, for each $i$,
we have
$Pr(a_i=b_i)=p$.
We also assume that the events
$\{(a_i=b_i)\}$
form an independent set.
We divide $A,B$ into blocks of size~2.
To start,
put $\mathbf{X}=\left(\begin{array}{c}x_1\\x_2\end{array}\right)$,
$\mathbf{Y}=\left(\begin{array}{c}y_1\\y_2\end{array}\right)$,
$\mathbf{Z}=x_1+x_2$.
\begin{lemmaN}
$$H(\mathbf{Z}\,|\,\mathbf{X},\mathbf{Y})=0.$$
\label{lemma:HZXY}
\end{lemmaN}
\begin{proofN}
We want to calculate
$\sum_{\mathbf{x},\mathbf{y}}H(\mathbf{Z}\,|\, \mathbf{x},\mathbf{y})Pr(X=\mathbf{x},Y=\mathbf{y})$.
Given $\mathbf{x},\mathbf{y}$, say
$\mathbf{x}=\left(\begin{array}{c}\alpha_1\\ \alpha_2\end{array}\right)$,
$\mathbf{y}=\left(\begin{array}{c}\beta_1\\ \beta_2\end{array}\right)$
the value of $\mathbf{Z}$ is $\alpha_1+\alpha_2$.
There is no uncertainty in the value of $\mathbf{Z}$
given $\mathbf{x},\mathbf{y}$,
i.e., each term in the above sum for $H$ is $H(1,0)=0$.
Therefore $H(\mathbf{Z}\,|\,\mathbf{X},\mathbf{Y})=0$.
\end{proofN}
From this we can use
formula~(\ref{equation:BasicIdentity})
for this block of size two.
We can think of
a channel
from $\mathbf{X}$ to $\mathbf{Y}$ (or from $\mathbf{Y}$ to $\mathbf{X}$)
which is the second extension of a
binary symmetric channel where $p$ is the
probability of success.
We have
$$H(\mathbf{X}\, |\, \mathbf{Y})
=H(\mathbf{X}\, |\, \mathbf{Y},\mathbf{Z})
+H(\mathbf{Z}\, |\, \mathbf{Y}).$$
From Theorem~\ref{theorem:XYInputOutput} part (c)
the left side, i.e., $H(\mathbf{X}\,|\,\mathbf{Y})$ is equal to
$2H(p,q)$.
Next we calculate the right side
beginning with
$H(\mathbf{Z}\,|\,\mathbf{Y})$,
i.e., $H\left(\mathbf{Z}\,\Big|\,\left(\begin{array}{c}y_1\\y_2\end{array}\right)\right)$.
We have
\begin{align}
H(\mathbf{Z}\,|\,\mathbf{Y})
&=H\bigl(\mathbf{Z}\,|\,(y_1+y_2=x_1+x_2)\bigr)Pr(y_1+y_2=x_1+x_2)\nonumber\\
&\qquad +H\bigl(\mathbf{Z}\,|\,(y_1+y_2\ne x_1+x_2)\bigr)Pr(y_1+y_2\ne x_1+x_2).
\nonumber
\end{align}
We know that
$Pr(x_1+x_2=y_1+y_2)=p^2+q^2$
and $Pr(x_1+x_2\ne y_1+y_2)=1-(p^2+q^2)=2pq$
since $p+q=1$.
From the standard formula we have
$H(\mathbf{Z}\,|\,\mathbf{Y})=(p^2+q^2)\log\left(\frac{1}{p^2+q^2}\right)+2pq\log\left(\frac{1}{2pq}\right)$
since
$H(\mathbf{Z}\,|\,\mathbf{Y})=H(p^2+q^2,2pq)$.
Next we calculate
$$H\left(\left(\begin{array}{cc}x_1\\x_2\end{array}\right)\,\Big |\,
\left(\begin{array}{c}y_1\\y_2\end{array}\right), (x_1+x_2)\right)
=H(\mathbf{X}\mid \mathbf{Y},\mathbf{Z}).$$
Again we have two possibilities, i.e.,
$y_1+y_2=x_1+x_2$ and $y_1+y_2\ne x_1+x_2$.
The corresponding probabilities are
$p^2+q^2$ and $2pq$ respectively.
We obtain
$$H(\mathbf{X}|\mathbf{Y},\mathbf{Z})
=(p^2+q^2)H\left(\frac{p^2}{p^2+q^2},\frac{q^2}{p^2+q^2}\right)
+2pqH\left(\frac{pq}{2pq},\frac{pq}{2pq}\right).$$
This comes about from the facts that
\begin{enumerate}[(a)]
\item
If $y_1+y_2=x_1+x_2$ then we either have
$y_1=x_1$ and $y_2=x_2$ or
$y_1=1-x_1$, $y_2=1-x_2$.
\item
If $y_1+y_2\ne x_1+x_2$
then either $y_1=x_1$ and $y_2\ne x_2$
or $y_1\ne x_1$ and $y_2=x_2$.
\item
$H(\frac{1}{2},\frac{1}{2})=1$.
\end{enumerate}
Then from
equation~(\ref{equation:BasicIdentity})
we have
{\em our first identity} as follows
\begin{equation}
\label{equation:FirstNewIdentity}
2H(p,q)=(p^2+q^2)H\left(\frac{p^2}{p^2+q^2},\frac{q^2}{p^2+q^2}\right)
+2pq
+H(p^2+q^2,2pq).
\end{equation}
\bigskip\noindent
{\bf Blocks of Size Three.}
\medskip\noindent
Here
$\mathbf{X}=\left(\begin{array}{c}x_1\\x_2\\x_3\end{array}\right)$,
$\mathbf{Y}=\left(\begin{array}{c}y_1\\y_2\\y_3\end{array}\right)$,
$\mathbf{Z}=x_1+x_2+x_3$.
As in Lemma~\ref{lemma:HZXY}
we have $H(\mathbf{Z}\,|\,\mathbf{X},\mathbf{Y})=0$
so we can use formula~(\ref{equation:BasicIdentity}) again,
i.e.,
$$H(\mathbf{X}\,|\,\mathbf{Y})=H(\mathbf{X}\,|\,\mathbf{Y},\mathbf{Z})
+H(\mathbf{Z}\,|\,\mathbf{Y}).$$
We have a
channel from $\mathbf{X}$ to $\mathbf{Y}$
(or from $\mathbf{Y}$ to $\mathbf{X}$)
which is the third extension $C^{(3)}$
of a binary symmetric channel $C$,
where $p$ is the probability that 0 (or 1) is transmitted
to itself.
From Theorem~\ref{theorem:XYInputOutput} we have
$H(\mathbf{X}|\mathbf{Y})=3H(p,q)$.
Similar to the case of blocks of size~2, we have
$H(\mathbf{Z}|\mathbf{Y})=H(p^3+3pq^2,q^3+3qp^2)$.
This is because the probabilities that
$Z=y_1+y_2+y_3$ or
$Z\ne y_1+y_2+y_3$ are,
respectively, $p^3+3pq^2$ or
$q^3+3qp^2$, as follows.
If $Z=y_1+y_2+y_3$, then either
$x_1=y_1$, $x_2=y_2$, $x_3=y_3$
or else, for some $i$, $1\le i\le 3$ (3 possibilities)
$x_i=y_i$ and, for the other two indices $j,k$,
$x_j\ne y_j$ and $x_k\ne y_k$.
A similar analysis can be carried out for the case
where $Z\ne y_1+y_2+y_3$.
We then get
$H(\mathbf{X}\,|\,\mathbf{Y},\mathbf{Z})=f(p,q)+f(q,p)$
where
$$f(p,q)=(p^3+3pq^2)\left\{H\left(\frac{p^3}{p^3+3pq^2},\frac{pq^2}{p^3+3pq^2},
\frac{pq^2}{p^3+3pq^2},\frac{pq^2}{p^3+3pq^2}\right)\right\}.$$
We now use the grouping axiom for $m=1$.
The $p$ in the formula refers to $\frac{p^3}{p^3+3pq^2}$ here
and the $q$ there is now replaced by
$\frac{3pq^2}{p^3+3pq^2}$.
Then
\begin{align}
f(p,q)
&=(p^3+3pq^2)\left\{
H\left(\frac{p^3}{p^3+3pq^2},\frac{3pq^2}{p^3+3pq^2}\right)
+\frac{3pq^2}{p^3+3pq^2}H(\frac{1}{3},\frac{1}{3},\frac{1}{3})\right\}
\nonumber\\
&=(p^3+3pq^2)H\left(\frac{p^3}{p^3+3pq^2},\frac{3pq^2}{p^3+3pq^2}\right)
+3pq^2\log 3.\nonumber
\end{align}
$f(q,p)$ is obtained by interchanging $p$ with $q$.
We note that, since $p+q=1$,
$3pq^2\log 3+3qp^2\log 3=3pq\log 3$.
From working with blocks of size~3 we get
\begin{align}
\label{equation:3Hpq}
3H(p,q)
&=H(p^3+3pq^2,q^3+3qp^2)
+(p^3+3pq^2)H\left(\frac{p^3}{p^3+3pq^2},\frac{3pq^2}{p^3+3pq^2}\right)
\nonumber\\
&\hphantom{=H(p^3+3pq^2,}
+(q^3+3qp^2)H\left(\frac{q^3}{q^3+3qp^2},\frac{3qp^2}{q^3+3qp^2}\right)
+3pq\log 3.
\end{align}
For blocks of size~2
formula~(\ref{equation:FirstNewIdentity})
can be put in a more compact form in terms of capacities,
namely,
\begin{equation}
\label{equationFirstNewIdentitySize2Blks}
2\bigl(1-H(p,q)\bigr)
=\biggl[1-H(p^2+q^2,2pq)\biggr]
+\left[(p^2+q^2)\left(1-H\left(\frac{p^2}{p^2+q^2},\frac{q^2}{p^2+q^2}\right)\right)\right].
\end{equation}
Using the same method we can find a formula
analogous to formulae~(\ref{equation:FirstNewIdentity}), (\ref{equation:3Hpq})
for obtaining $nH(p,q)$ as a linear combination
of terms of the form
$H(u,v)$ where $u,v$ involve terms in
$p^n,p^{n-2}q^2,\ldots,q^n,q^{n-2}p^2,\ldots$
plus extra terms such as
$3pq\log 3$ as in
formula~(\ref{equation:3Hpq}).
\section{Generalizations, an Addition Formula}
\label{section:GeneralizationsAndAdditionFormula}
The result of Theorem~\ref{theorem:XYInputOutput}
can be extended to the more general case where
we take the product of $n$ binary symmetric channels
even if the channel matrices can be different
corresponding to differing $p$-values.
As an example, suppose we use the product of
2 binary symmetric channels with channel matrices
$$
\left(
\begin{array}{cc}
p_1&q_1\\
q_1&p_1\end{array}\right),
\quad
\left(
\begin{array}{cc}
p_2&q_2\\
q_2&p_2\end{array}\right)
.$$
Then the argument in Section~\ref{section:NewIdentities}
goes through. To avoid being overwhelmed
by symbols we made a provisional notation change.
\begin{notationN}
We denote by $h(p)$ the quantity
$H(p,q)=p\log\frac{1}{p}+q\log\frac{1}{q}$.
\end{notationN}
Then we arrive at the following addition formula
\begin{align}
h(p_1)+&h(p_2)\nonumber\\
&=h(p_1p_2+q_1q_2)
+(p_1p_2+q_1q_2)\,h\left(\frac{p_1p_2}{p_1p_2+q_1q_2}\right)\nonumber\\
&\qquad +(p_1q_2+p_2q_1)\,h\left(\frac{p_1q_2}{p_1q_2+p_2q_1}\right).
\end{align}
Similarly to the above
we can derive a formula for
$h(p_1)+h(p_2)+\cdots+h(p_n)$.
\section{The Shannon Limit and Applications to Cryptography}
\label{section:ShannonLimitAppnsCrypt}
The above method of using blocks of various sizes
is reminiscent of the algorithm for the key exchange in
\cite{BruenForcinitoWehlau}
which relates to earlier work in~\cite{Bennett}, \cite{MaurerWolf}
and others.
Indeed the identities above were informed by the
details of the algorithm.
The algorithm starts with two arrays
$(a_1,\ldots,a_n)$ and $(b_1,\ldots,b_n)$.
We assume that the set of events
$\{a_i=b_i\}$ is an independent set with
$p=Pr(a_i=b_i)$.
We subdivide $A,B$ into corresponding sub-blocks of size~$t$,
where $t$ divides $n$.
Exchanging parities by public discussion we end up with
new shorter sub-arrays $A_1,B_1$, where the probabilities
of corresponding entries being equal are independent
with probability $p_1>p$.
Eventually after $m$ iterations we end up
with a common secret key $A_m=B_m$.
Let us take an example.
Start with two binary arrays
$(a_1,\ldots,a_n)$ and $(b_1,\ldots,b_n)$
of length $n$ with $n$ even, $n=2t$, say.
We subdivide
the arrays into corresponding blocks of size~2.
If the two blocks are
$\left(\begin{array}{c}a_1\\a_2\end{array}\right)$
and
$\left(\begin{array}{c}b_1\\b_2\end{array}\right)$
we discard those blocks if the parities disagree.
If the parities agree, which happens with
probability $p^2+q^2$,
we keep $a_1$ and $b_1$,
discarding $a_2$ and $b_2$.
Thus, on average, we keep
$(p^2+q^2)\frac{n}{2}$ partial blocks
and discard $\bigl[1-(p^2+q^2)\bigr]\frac{n}{2}$ blocks of size~2.
Let us suppose $n=100$ and $p=0.7$.
From Theorem~\ref{theorem:XYInputOutput} part~(d),
the information that $Y$ has about $X$, i.e.,
that $B$ has about $A$ is
$100[1-H(0.7,0.3)]\approx100(1-0.8813)\approx 11.87$~Shannon bits.
We are seeking to find a sub-array of $A,B$ such that
corresponding bits are equal.
Our method is to publicly exchange parities.
The length of this secret key
\index{key length}
will be at most~11.
Back to the algorithm.
$A,B$ keep on average $(50)(p^2+q^2)$ blocks of size~2,
i.e., $(50)(0.58)=29$ blocks of size~2.
$A$ and $B$ remove the bottom element of each block.
We are left with 29 pairs of elements $(a_1,b_1)$.
The probability that $a_1=b_1$ given that
$a_1+a_2=b_1+b_2$ is
$\frac{p^2}{p^2+q^2}$,
i.e.,
$\frac{(0.7)^2}{(0.7)^2+(0.3)^2}=\frac{0.49}{0.58}\approx 0.845$.
Next, $1-H(0.845,0.155)\approx(1-0.6221)\approx 0.3779$.
To summarize, we started with
100 pairs $(a_i,b_i)$ with
$Pr(a_i=b_i)=0.7$.
The information revealed to $B$ by $A$ is
{\em $(100)(1-H(0.7,0.3))=11.87$ Shannon bits of information}.
After the first step of the algorithm
we are left with 29 pairs $(a_j,b_j)$ with
$Pr(a_j=b_j)=0.845$.
The amount of information revealed to the remnant of $B$
by the remnant of $A$ is
{\em $29(0.3779)\approx 10.96$ Shannon bits of information}.
So we have ``wasted'' less than 1 bit,
i.e. the wastage is about 8\%.
Mathematically we have
$100\bigl(1-H(0.7,0.3)\bigr)=29\bigl(1-H(0.845,0.155)\bigr)+$~Wastage.
In general we have
$$n[1-H(p,q)]=
\frac{n}{2}(p^2+q^2)\left[1-H\left(\frac{p^2}{p^2+q^2},\frac{q^2}{p^2+q^2}\right)\right]+W,$$
where $W$ denotes the wastage.
Dividing by $\frac{n}{2}$ we get
$$2[1-H(p,q)]=
(p^2+q^2)\left[1-H\left(\frac{p^2}{p^2+q^2},\frac{q^2}{p^2+q^2}\right)\right]
+\frac{2W}{n}.$$
\noindent
{\em Comparing with formula(\ref{equationFirstNewIdentitySize2Blks})}
we see that
$W=\frac{n}{2}\bigl[1-H(p^2+q^2,2pq)\bigr]$.
In this case $W=50\bigl[1-H(0.58,0.42)\bigr]\approx 50(1-0.9815)
=(50)(0.0185)=0.925$.
{\em To sum up then
the new identities tell us exactly how much information
was wasted and not utilized.}
They also tell us, in conjunction with the algorithm,
the optimum size of the sub-blocks at each stage.
One of the original motivations for work in coding theory
was that
the Shannon fundamental theorem showed
how capacity was the bound for accurate communication
but the problem was to construct linear codes
or other codes such as turbo codes
that came close to the bound.
Here we have an analogous situation.
In the example just considered
the maximum length of a
\index{key length}
cryptographic common secret key
\index{cryptography!secret key@\textsl{secret key}}
obtained as a common subset of $A,B$
is bounded by
$n(1-H(p))$.
The problem is to find algorithms which produce such a
common secret key coming close to this Shannon bound
of $n\bigl(1-H(p)\bigr)$.
This work nicely illustrates
the inter-connections between
codes, cryptography, and information theory.
Information theory tells us the bound.
The identities tell us the size of the sub-blocks
for constructing
a common secret key
which attains, or gets close to,
the information theory bound.
Coding theory is then used to ensure that the
two communicating parties have a common secret key
by using the hash function
described in the algorithm using
a code~$C$.
Error correction can ensure that
the common secret key
can be obtained without using
another round of the algorithm
(thereby shortening the common
key) if the difference between
the keys of $A$ and $B$
is less than the minimum distance
of the dual code of $C$.
This improves on the standard method of
checking parities of random subsets
of the keys $A_m,B_m$ at the last stage.
\section{Concluding Remarks.}
\begin{enumerate}
\item
Please see
Chapter~25 of
the forthcoming
{\em Cryptography, Information Theory, and Error-Correction:
A Handbook for the Twenty-First Century},
by Bruen, Forcinito, and McQuillan, \cite{BruenForcinitoMcQuillan},
for background information, additional details,
and related material.
\item
Standard tables for entropy list values to two decimal places.
When $p$ is close to 1 interpolation
for three decimal places is difficult as
$h(p)$ is very steeply sloped near $p=1$.
Formula~\ref{equation:FirstNewIdentity}
may help since $\frac{p^2}{p^2+q^2}$ is less than $p$,
and the formula can be re-iterated.
\item
In~\cite{BruenForcinitoWehlau}
the emphasis is on the situation where the eavesdropper
has no initial information.
The case where the eavesdropper has initial
information is discussed in~\cite{MaurerWolf}.
In Section 7 of~\cite{BruenForcinitoWehlau}
the quoted result uses Renyi entropy
rather than Shannon entropy.
\item
The methods in this note suggest possible generalizations
which we do not pursue here.
\end{enumerate}
\bigskip\noindent
{\bf
Acknowledgement:}
The author thanks Drs Mario Forcinito, James McQuillan
and David Wehlau for their help and encouragement with
this work.
\bigskip
|
\section{Research Methods}
\end{document}
\endinput
\section{introduction}
Documenting the story behind code and results is critical for data scientists to collabrate effectively with others, as well as \revision{their future selves~\cite{kery_exploring_2017, piorkowski2021ai, kross2021orienting,Muller:2019:DSW:3290605.3300356}}.
The story, code, and computational results together construct a computational narrative.
Unfortunately, data scientists often write messy and drafty analysis code in computational notebooks as they need to quickly test hypotheses and experiment with alternatives.
It is a tedious process for data scientists to then manually document and refactor the raw notebook into a more readable computational narrative, thus many people neglect to do so~\cite{rule2018exploration}.
Many efforts have sought to address the tension between \emph{exploration} and \emph{explanation} in computational notebooks.
For example, researchers have explored the use of code gathering techniques to help data scientists organize cluttered and inconsistent notebooks~\cite{Head2019CHI}, as well as algorithmic and visualization approaches to help data scientists forage past analysis choices~\cite{Kery2019CHI}.
But these efforts focus on the cleaning and organizing of existing notebook content, instead of creating the new content.
Another work developed a chat feature that enables data scientists to have simultaneous discussions while coding in a notebook \cite{wang2020callisto}, and linked their chat messages as documentations to relevant notebook elements as in Google Docs~\cite{wang2016people}.
However, these chat messages are too fragmented and colloquial to be used for documentation; besides, in real practice \revision{data scientists and business analysts rarely work on notebooks at the same time and actively message each other}.
We began our project by asking,
``What makes a well-documented notebook?''
To answer this question, we first conducted an in-depth analysis of how human data scientists document notebooks.
Publicly shared user notebooks on Githubs are often not well documented~\cite{rule2018exploration}, thus we look up to a special set of notebooks -- the highly-voted notebooks users submitted to Kaggle competitions.
We conducted a formative study with a sample of 80 of these notebooks, and our interative indepth coding analysis suggested these 80 notebooks have much better documentations in comparison to the corpus reported in previous literature~\cite{rule2018exploration}.
Thus, we refer to them as ``well-documented'' notebooks.
Our coding process of these 80 notebooks also revealed a taxtonomy of nine categories (e.g., Reason, Process, Result) for the documentation content,
which reflects the thought processes and decisions made by the notebook owner.
These findings together with the insights from related work motivate us to consider AI automation as a potential solution to support the human process of crafting documentation.
We propose Themisto{}, an automated code documentation generation system that integrates into the Jupyter Notebook environment.
To support the diverse types of documentation content and to complement the AI limitations, Themisto{} incorporate three distinct approaches: a deep-learning-based approach to automatically generate new documentation for source code (fully automated); a query-based approach to retrieve existing documentation from online \ac{API} websites for third party packages and libraries (fully automated); and a prompt-based approach to give users a start of the sentence and encourage them to complete the sentence that serves as documentation (semi automated).
We evaluated Themisto{} in a within-subjects experiment with 24 data science practitioners.
We found that Themisto{} reduced the time for data scientists to create documentation, reminded them to document code they would have ignored, and improved their satisfaction with their computational notebooks.
Meanwhile, the quality of the documentation produced with Themisto{} are about the same as what data scientists produced on their own.
\revision{Base on these findings, we re-imagine that the code documentation task can be conducted in a Human-AI Collaboration fasion in the future, where this joint effort may have unique advantages in comparison to the solo effort of a human alone.}
Our paper provides a three-fold contribution to the HCI and data science practitioner communities:
\begin{itemize}
\item providing an empirical understanding of best practices of how human documenting a notebook through an analysis of highly-rated Kaggle notebooks,
\item demonstrating the design of \revision{a human-centered} AI system that can collaborate with human data scientists to create high-quality computational narratives,
\item reporting empirical evidence that Themisto{} can collaborate with data scientists to generate high quality and highly-satisfied computational notebooks in much less time.
\end{itemize}
\section{Related Work}
Our work builds on top of both \ac{HCI} and \ac{ML} fields. Thus, our literature review briefly summarizes the work of both, with a focus on the following three topics: computational notebook management, code documentation supporting systems, and neural-network-based code summarization.
\begin{figure}[t]
\includegraphics[width=\textwidth]{figures/notebooks.pdf}
\caption{Computational notebooks allow data scientists to create (A) markdown cells and (B) code cells, and view (C) code output in the same environment.
Together, the variety of media including text explanations, graphs, forms, interactive visualizations, code segments and their outputs --- weaves into computational narratives.}
\label{fig:notebook}
\end{figure}
\subsection{Computational Notebooks as Computational Narrative}
Computational notebooks allow data scientists to weave together a variety of media, including text explanations, graphs, forms, interactive visualizations, code segments, and their outputs, into computational narratives (as shown in Figure \ref{fig:notebook}).
These computational narratives enable literate programming~\cite{knuth_literate} and allow data scientists to effectively create, communicate, and collaborate on their analysis work.
The data scientist community has widely adopted notebook systems (e.g., Jupyter Notebook~\cite{jupyter_project_2015} and Jupyter Lab~\cite{jupyter_jupyterlab:_2016}) as their main working envrioment~\cite{perkel_why_2018}.
However, it is not easy for data scientists to create a computational narrative while they are coding for rapid exploration.
Data scientists often need to explore diverse sets of hypotheses and theories~\cite{liu2019understanding, rehman2019towards}.
Active exploration of alternatives increases the workload for data scientists to track the history of their experimentation~\cite{Kery:2018:SNE:3173574.3173748}.
Thus, documenting for those alternatives will pose more workload to data scientists, sometimes interference with their cognitive process of coding, and is hardly rewarding because many those alternatives will be discarded in later versions.
Because creating and maintaining a clean computational narrative is often an expensive and tedious process, many computational notebooks shared within open communities are not appropriately documented.
For example, Rule et al. examined 1 million open-source computational notebooks from Github and found that one in four lacked any sort of written documentation~\cite{rule2018exploration}.
In addition, they analyzed a sample of 221 academic computational notebooks, which they considered are higher quality notebooks, and found that academic computational notebooks contained text cells for introduction, describing analytical steps, explaining the reasoning, and discussing results.
Poor documentation hinders the readability and reusability of the notebooks that are shared with other collaborators or even with one's future self~\cite{chattopadhyay2020s}.
Recently, various groups of researchers have developed a wide range of tools to help data scientists to manage their ``messy'' computational notebooks.
\revision{Notably, Lau et al. summarized the design space of computational notebooks which covered an overview for improving explanations in computational notebooks~\cite{lau2020design}.
Many strategies interact with markdown comments.
For example, facilitating cell folding could help surface important markdown cells \cite{rule_aiding_2018};
Kery et al. designed Verdant~\cite{Kery2019CHI}, a lightweight local versioning plugin for Jupyter Lab, that uses algorithmic and visualization techniques for data science workers to better forage their past analysis choices;
Woods et al argued \cite{wood2018design} for simpler and richer narratives;
Head et al. used code gathering tools to help data scientists trace back to the computational code from an end result~\cite{Head2019CHI};
Wenskovitch et al. designed an interactive tool that produced a visual summary of the structure of a computational notebook~\cite{wenskovitch2019albireo};
Wang et al. proposed capturing the contextual connections between notebook content and discussion messages to help data science teams reflect on their decision making process~\cite{wang2020callisto}.
}
However, despite the wide variety of approaches to helping data scientsts manage their notebooks, none of these tools directly aids data scientists in creating new, rich, descriptive contents to document their computational notebooks, and to improve the quality of the computational narrative.
Recent research works have proposed to use AI solutions to automate the various tasks along a data science project, such as the model training, model selection, and feature selection, and these technology are commonly refered as AutoML~\cite{wang2021autods,liu2020admm}.
The research gap and the AutoML techniques motivate us to design and build an AI system to support data scientists to better document their code and to produce higher qualitive computational narratives.
But what makes up a good computational narrative?
Despite the portrait of not-so-good notebooks on Github \cite{rule2018exploration}, we need further understanding and role models for well-documented computational narratives.
Thus, we decided to first conduct an in-depth analysis of some highly-voted notebooks on Kaggle competetion\footnote{https://www.kaggle.com/}.
Kaggle competition provides a platform where organizations post datasets as challenges, and many data scientists submit their notebooks as solutions to a challenge.
If a solution has the highest accuracy, it wins the competition.
But those winning solutions are often not the most voted ones, as community members voted on readability and completeness of the computational narrative.
\subsection{Code Documentation in Software Engineering}
Documentation plays an important role in software programming.
Programmers write comments in their source code to make the code easier for both themselves and others to understand \cite{padioleau2009listening}.
Writing clear and comprehensive documentation is critical to software development and maintenance~\cite{de2005study, kajko2005survey, shi2011empirical, roehm2012professional, maalej2013patterns}.
However, writing documentation itself is a time-consuming task. And that is why documentation practices in open source communities are widely perceived to be of low quality, due in part to low levels of intrinsic enjoyment for doing documentation work~\cite{geiger2018types}.
To save time in creating documentation, template-based approaches are often used to help developers annotate their source code.
For example, tools like JavaDoc\footnote{JavaDoc: \url{https://docs.oracle.com/javase/8/docs/technotes/tools/windows/javadoc.html}} and JSDoc\footnote{JSDoc: \url{https://devdocs.io/jsdoc/}} allow programmers to annotate their code with tags (e.g., @param, @return) and then automatically generate documentation using these tags.
This approach helps programmers create documentation for others and works especially well for documenting \acp{API}, where method signatures and variable types are important pieces of information and can easily be documented from tags.
Although, such methods may not work in the rapid, experimental nature of data science work, because data scientists may be particularly reluctant to create and maintain high-quality documentation of their work.
Furthermore, these methods can not capture other aspects of documentation important in data science, such as how a data set was constructed, the intent behind an analysis, or a description of why an experiment was successful or not.
Recently, some researchers have put forth proposals for better documenting the specific artifacts invovled in a data science workflow, i.e., the data set and the machine learning model~\cite{gebru2018datasheets, holland2018dataset, mitchell2019model, arnold2019factsheets}.
Notably, Gebru et al.~\cite{gebru2018datasheets} and Holland et al.~\cite{holland2018dataset} proposed both qualitative and quantitative \revision{guidelines} for documenting a dataset, so that the dataset creators and maintainers can follow these guidlines to document useful information for the data.
Similarly, Mitchell et al.~\cite{mitchell2019model} and Arnold et al.~\cite{arnold2019factsheets} explored using formulas to document the machine learning model artifacts, and sharing such formulas with others.
These approaches are inline with what is called \textit{provenance}, which refers to tracking what has been done with code and data over time -- typically to aid reproducibility of results -- using applications such as noWorkflow and YesWorkflow \cite{pimentel2016yin}, ReproduceMeGit \cite{samuel2020reproducemegit}, and Provbook \cite{samuel2018provbook}.
However, this approach focus more on the the dataset and the model artifact in the final product of a data science project, and supporting data scientists to create a ``factsheet'' for these artifacts for the non-technical consumers. Instead, we want to support data scientists to better create the documentation during the process of creating models and data science products, and such documentation, together with the code as a computational narrative, is primarily for other technical users to understand and to reuse.
In addition to the various ways of generating new documentation for code, there is another research line that focuses on improving the usability of documentation, as novice programmers may find it difficult to read and use \ac{API} documentation~\cite{horvath2019methods}.
Oney et al. proposed linking interactive documentation and example code in an editor to help novice programmers better understand the external documentation and write code~\cite{oney2012codelets}.
We believe this approach of linking code with external documentation is a promising way to help data scientists to create more usable documentations, and we will also implement this retrieval-based approach in our system.
\subsection{GNN-Based Automatic Code Summarization System}
Automatic code summarization is a rapidly-expanding research topic in the \ac{NLP} and \ac{ML} communities~\cite{eddy2013evaluating, hu2018deep, alon2018code2seq, iyer_summarizing_2016, leclair2020improved}.
The automatic code summarization task can be considered as a translation task, which takes a code snippet as the input sequence, and \say{translating} it into a natural language description of the code as an output sequence.
Early work primarily used predefined templates and heuristics to produce code summaries (e.g.,~\cite{eddy2013evaluating, sridhara2010towards}).
Recent studies have taken the advantage of modern deep neural network architectures to generate the summary for the code (e.g.~\cite{hu2018deep, alon2018code2seq, iyer2016summarizing}).
Motivated by the language translation task (e.g., English to French), most of these learning-based approaches are based on the \ac{NMT} model architecture~\cite{luong2015effective}.
This architecture breaks code into a sequence of input tokens and produce the summarization text as a sequence of output tokens.
However, this sequence-to-sequence approach does not work well in practice because source code is not just a stream of tokens.
There is additional semantic information that is lost when processing source code in this way.
LeClair et al. proposed improving code summarization through the use of \acp{GNN}~\cite{leclair2020improved}.
The GNN model can take in both the code sequence and its \ac{AST} structure (refer to Fig~\ref{fig:model} for an example) as input to generate summary sentences as output.
Their approach achieved better accuracy than the baseline algorithms.
Our work explores neural-network-based automatic code summarization techniques to support document writing in computational notebooks.
To our acknowledge, there has been little discussion in the \ac{HCI} community on leveraging automatic code summarization techniques to improve documentation.
Furthermore, we suspect that the automation approach alone may not work well in the documentation creation task, as data science is a highly interdisciplinary field that requires various human expertises to explain and interpret.
Inspired by prior studies that implement \ac{AI} systems to work together with human~\cite{louie2020novice, wang2020autoai}, we believe the system will work better if it has both the automated documentaiton capability and the capability that allows users to directly manipulate the documentation.
However, what types of documentation may be better suited for \ac{AI} to do, and what works should the system leave to human data scientists?
This is a design question that requires further exploration of the best practices for creating notebooks.
Thus, we start this project with a formative study to fill this research gap.
\section{Formative Study}
In order to build a useful system that can support data scientists to create documentations and to improve their computational narrative's quality, we first need to explore and understand the characteristics of good documentations in high-quality notebooks. \textbf{What does a well-documented computational narrative look like?}
We identify \say{well-documented} computational narratives with ratings from a broader data scientist community (Kaggle), and analyze their characteristics specifically around the documentation.
We consider the community voting number is a good indicator to reflect a computational notebook's quality for our research goal.
Based on this premise, we then conduct a formative study to analyze the characteristics of a set of most voted computational narratives, and explore how the data scientists create documentations for these notebooks.
\subsection{Data Collection}
We collected notebooks from two popular Kaggle competitions --- House Price Prediction\footnote{\url{https://www.kaggle.com/c/house-prices-advanced-regression-techniques}} and Titanic Survival Prediction\footnote{\url{https://www.kaggle.com/c/titanic/}}.
We chose these two competitions because they are the most popular competitions (5280 notebooks submitted for House Price and 6300 notebooks submitted for Titanic Survival) and because many data science courses use these two competitions as a tutorial for beginners~\cite{bai2018problem, fernandez7energy}.
We collected the top 1\% of the submitted notebooks from each competition based on their voting numbers, which resulted in 53 for House Price and 63 for Titanic Survival.
We then filtered out the notebooks that were not written in English and the ones that are not relevant to the particular challenge (e.g., a computational notebook as a tutorial on how to save memories can win lots of votes, but it is not a solution to the challenge), which returned 80 valid notebooks for analysis (39 for House Price and 41 for Titanic Survival).
\begin{figure}[t]
\includegraphics[width=\textwidth]{figures/rule.pdf}
\caption{\revision{We replicated the notebook-level descriptive analysis by Rule et al. \cite{rule2018exploration} to the 80 well-documented notebooks on Kaggle. The left side represents the descriptive visualization of the 80 well-documented computational notebooks from Kaggle (noted as Sample A) and the right side represents the descriptive visualization of the 1 million computational notebooks on Github (noted as Sample B). The highly-voted notebooks on Kaggle are better documented compared to the Github notebooks.}}
\label{fig:rule}
\end{figure}
\subsection{Data Analysis}
Five members of the research team conducted an iterative open coding \revision{process} to analyze the collected notebooks.
Differing from ~\cite{rule2018exploration}, where their qualitative coding \revision{stopped} at the notebook level, our analysis goes deep to the cell granularity: we code each cell's purposes and types of content; and which step (stages) in the data science lifecycle that the cell belongs (e.g., data cleaning or modeling training~\cite{zhang2020data,automationsurvey}).
Our analysis covered 4427 code cells and 3606 markdown cells within the 80 notebooks.
Each notebook took around 1 hour to code as we coded the notebook at the cell level.
Each coder independently analyzed the same six notebooks to develop a codebook.
After discussing and refining the codebook, they again went back to recode the six notebooks and achieved pair-wise inter-rater reliability ranged 0.78-0.95 (Cohen's $\kappa$).
After this step, the five coders divided and coded the remaining notebooks.
\begin{table}[t!]
\caption{We identified 9 categories based on the purpose of markdown cells. Note that a markdown cell may belong to multiple categories of contents or none of the categories.}
\label{tab:category}
\small
\begin{tabular}{p{1.8cm} p{1cm} p{5cm} p{4cm}}
\toprule
\textbf{Category} & \textbf{N} & \textbf{Description} & \textbf{Example} \\
\midrule
Process & 2115 \newline (58.65\%) & The markdown cell describes what the following code cell is doing. This always appears before the relevant code cell. & \mintinline{html}{Transforming Feature X}\newline \mintinline{html}{to a new binary variable} \\
\midrule
Headline & 1167 \newline (32.36\%) & The markdown cell contains a headline in markdown syntax. The cell is used for navigation purposes or marking the structure of the notebook. It may be relevant to a nearby code cell. & \mintinline{html}{# Blending Models}\\
\midrule
Result & 692 \newline (19.19\%) & The markdown cell explains the output. This type always appears after the relevant code cell. & \mintinline{html}{It turns out there is a}\newline \mintinline{html}{long tail of outlying}\newline \mintinline{html}{properties...} \\
\midrule
\revision{Background Knowledge} & 414 \newline (11.48\%) & The markdown cell provides a rich content \revision{for background knowledge}, but may not be relevant to a specific code cell. & \mintinline{html}{Multicollinearity incre-} \newline \mintinline{html}{ases the standard errors} \newline \mintinline{html}{of the coefficients.} \\
\midrule
Reason & 227 \newline (6.30\%) & The markdown cell explains the reasons why certain functions are used or why a task is performed. This may appear before or after the relevant code cell. & \mintinline{html}{We do this manually, be-}\newline\mintinline{html}{cause ML models won't be}\newline\mintinline{html}{able to reliably tell}\newline\mintinline{html}{the differences.} \\
\midrule
Todo & 202 \newline (5.60\%) & The markdown cell describes \revision{a list of actions for upcoming analysis. This normally is not relevant to a specific code cell.} & \mintinline{html}{1. Apply models}\newline\mintinline{html}{2. Get cross validation}\newline\mintinline{html}{scores} \newline \mintinline{html}{3. Calculate the mean} \\
\midrule
Reference & 200 \newline (5.55\%) & The markdown cell contains an external reference. This is also relevant to the adjacent code cell. & \mintinline{html}{Gradient Boosting} \newline\mintinline{html}{Regression Refer}\newline\mintinline{html}{[here](https://...)}\\
\midrule
Meta-Information & 141 \newline (3.91\%) & The markdown cell contains meta-information such as project overview, author's information, and a link to the data sources. This often is not relevant to a specific code. & \mintinline{html}{The purpose of this}\newline \mintinline{html}{notebook is to build a}\newline \mintinline{html}{model with Tensorflow.}\\
\midrule
Summary & 51 \newline (1.41\%) & The markdown cell summarizes what has been done so far for a section or a series of steps. This often is not relevant to a specific code.&\mintinline{html}{**In summary**}\newline\mintinline{html}{By EDA we found a strong}\newline\mintinline{html}{impact of features like}\newline\mintinline{html}{Age, Embarked..} \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Results}
We found that these 80 well-documented computational notebooks all contain rich documentation.
In total, we identified \textbf{nine categories for the content} of the markdown cells.
In addition, we found the markdown cells covered \textbf{four stages and 13 tasks} of the data science workflow~\cite{automationsurvey}.
Note that a markdown cell may belong to multiple categories.
\subsubsection{Descriptive statistics of the notebook.}
We found that on average, each notebook contains 55.3 code cells and 45.1 markdown cells.
We replicated \revision{the} notebook descriptive analysis that Rule et al. used to analyze 1 milion computational notebooks on Github \cite{rule2018exploration}.
As shown in Figure \ref{fig:rule}, the left side represents the descriptive visualization of the 80 well-documented computational notebooks from Kaggle (noted as Sample A) and the right side reprersents the descriptive visualization of the 1 milion computational notebooks on Github (noted as Sample B).
We found that the Sample A has more total cells per notebook (Median = 95) than Sample B (Median = 18).
Sample A has roughly equal ratio of markdown cells and code cells per notebook, while Sample B is unbalanced with majority cells being code cells.
Notably, Sample A has more total words in markdown cells (Median = 1728) than Sample B (Median = 218).
This result indicates that the 80 well-documented computational notebooks are better documented than general Github notebooks.
\subsubsection{Data scientists use markdown cells to document a broad range of topics.}
As shown in Table \ref{tab:category}, our analysis revealed that markdown cells are mostly used to describe what the adjacent code cell is doing (Process, 58.65\%).
Second to the Process category, 32.36\% markdown cells are used to specify a headline for organizing the notebook into separate functional sections and for navigation purposes (Headline).
Markdown cells can also be used to explain beyond the adjacent code cells.
We found that many markdown cells are created to describe the outputs from code execution (Result, 19.19\%), to explain results or critical decisions (Reason, 6.30\%), or to provide an outline for the readers to know what they are going to do in a list of todo actions (Todo, 5.60\%), and/or to recap what has been done so far (Summary, 1.41\%).
We observed that 11.48\% markdown cells explain what a general data science concept means, or how a function works (\revision{Background Knowledge}), while 5.54\% markdown cells are connected with external references for readers to further explore the topics (Reference).
We believe these are the extra efforts that the notebook owners dedicated, to attract a broader audience, especially beginners in the Kaggle community.
In addition, we found that authors approached the story in different styles.
For example, some authors want to leave their own signature, and so they spend spaces at the beginning of the notebooks to debrief the project, to add the author's information, or even to add their mottos (Meta-Information, 3.91\%).
Some authors prefer to use concise and accurate language to convey important information; while others write documentation in more creative and entertaining ways --- for example, making analogies between data science workflow and starting a business.
\begin{table}[t!]
\caption{We coded each markdown cell to which data science stage (or task) they belong. We identified 4 stages with 13 tasks out of the data science lifecycle~\cite{automationsurvey}. Note that a markdown cell may belong to multiple stages or none of the stages.}
\begin{center}
\small
\begin{tabular}{llp{4.5cm}l}
\toprule
\textbf{Stage} & \textbf{Total} & \textbf{Task} & \textbf{N} \\
\midrule
\multirow{2}{*}{Environment Configuration} & \multirow{2}{*}{162 (4.49\%)} & Library Loading & 33 (0.92\%)\\
&&Data Loading & 129 (3.58\%)\\
\midrule
\multirow{3}{*}{Data Preparation and Exploration}&
\multirow{3}{*}{1336 (37.05\%)}&
Data Preparation & 91 (2.52\%)\\
&&Exploratory Data Analysis & 960 (26.62\%)\\
&&Data Cleaning & 285 (7.90\%)\\
\midrule
\multirow{3}{*}{Feature Engineering and Selection}&
\multirow{3}{*}{375 (10.40\%)}&
Feature Engineering & 120 (3.32\%)\\
&& Feature Transformation & 178 (4.94\%)\\
&&Feature Selection & 77 (2.14\%)\\
\midrule
\multirow{5}{*}{Model Building and Selection}&
\multirow{5}{*}{994 (27.57\%)}&
Model Building & 247 (6.85\%)\\
&& Data Sub-Sampling and Train-Test Splitting & 61 (1.69\%)\\
&& Model Training & 377 (10.45\%)\\
&& Model Parameter Tuning & 81 (2.25\%)\\
&& Model Validation and Assembling & 288 (6.32\%)\\
\bottomrule
\end{tabular}
\end{center}
\label{tab:stage}
\end{table}
\subsubsection{Data science stages.}
We coded markdown cells based on where they belong in the data science workflow \cite{wang2019human}.
\revision{As shown in Table \ref{tab:stage},} we identified four stages and 13 tasks.
The four stages include \textbf{environment configuration} (4.50\%), \textbf{data preparation} and exploration (37.05\%), \textbf{feature engineering and selection} (10.40\%), and \textbf{model building and selection} (27.57\%).
At the finer-grained task level, in particular, notebook authors create more markdown cells for documenting exploratory data analysis tasks (26.62\%) and model training tasks (10.45\%).
The rest of the markdown cells are evenly distributed along with other tasks.
\subsection{Design Implications}
In summary, our analysis of markdown cells in well-documented notebooks suggests that data scientists document various types of content in a notebook, and the distribution of these markdown cells generally follows an order of the data science lifecycle, starting with data cleaning, and ending with model building and selection. Based on these findings, we synthesize the following actionable design considerations:
\begin{itemize}
\item \textbf{The system should support more than one type of documentation generation.} Data scientists benefit from documenting not only the behavior of the code, but also interpreting the output, and explaining rationales. Thus, a good system should be flexible to support more than one type of documentation generation.
\item \textbf{Some types of documentations are highly related to the adjacent code cell.} We found at least the Process, Result, Reason, and Reference types of documentations are highly related to the adjacent code cell. To automatically generate interpretations of results or rationale for a decision may be hard, as both involve deep human expertise. But, with the latest neural network algorithms, we believe we can build an automation system to generate Process type of documentation, and we can also retrieve Reference for a given code cell.
\item \textbf{There are certain types of documentations that are irrelevant to the code.} Various types of documentations do not have a relevant code piece upon which the automation algorithm can be trained. Together with the Reason and Result types, the system should also provide a function that the human user can easily switch to the manual creation mode for these types.
\item \textbf{For different types of documentation, it could be at the top or the bottom of the related code cell.} This design insight is particularly important to the Process, Result, and Reason types of documentation. It may be less preferable to put Result documentation before the code cell, where the result is yet to be rendered. The system should be flexible to render documentation at different relative locations to the code cell.
\item \textbf{External resources such as \acp{URL} and the official \ac{API} descriptions may also be useful.} Some types of documentation, such as \revision{Background Knowledge} and Reference, are not easy to be generated with the NN-based models, but they are easy to retrieve from the Internet. So the system should incorporate the capability to fetch relevant web content as candidate documentation.
\item \textbf{There is an ordinality in markdown cells that is aligned with the data science project's lifecycle.} The system should consider that Library Loading types of cells are often at the beginning section of the notebook, and the Model Training type of content may be more likely to appear near the end of the notebook. In our system prototype, though, we did not take this design consideration into account, it will be our future work.
\item \textbf{The notebook would be nice to have documentation with a problem overview at the beginning and a summary at the end.} We considered this design implication not in the system design, but our evaluation study design. For the two barebone notebooks we used in the experiment, we always provide a problem overview as a markdown cell at the top of the notebook.
\end{itemize}
\section{The Themisto System: Design and Implementation}
Based on findings from the formative study and design insights from related works, we design and implement Themisto{}, an automatic documentation generation system that supports data scientists to write better-documented computational narratives.
In this section, we present the system architecture, the user interface design, and the core technical capability of generating documentation.
\subsection{System Architecture}
The Themisto{} system has two components: the client-side \ac{UI} is implemented as a Jupyter Notebook plugin using TypeScript code, and the server-side backend is implemented as a server using Python and Flask.
The client-side program is responsible to render user interface, and also to monitor the user actions on the notebook to edits in code cells.
When the user's cursor is focused on a code cell, the \ac{UI} will send the current code cell content to the server-side program through \ac{HTTP} requests.
The server-side program takes the code content and generates documentation using both the deep-learning-based approach and the query-based approach.
For the deep-learning-based approach, the server-side program first tokenizes the code content and generates the \ac{AST}.
It then generates the prediction with the pre-trained model.
For the query-based approach, the server-side program matches the curated \ac{API} calls with the code snippets and returns the pre-collected descriptions.
For the prompt-based approach, the server-side program sends different prompts (e.g., for interpreting result or for explaining reason) base on the output type of the code cell.
\subsection{User Interface Design}
\begin{figure}[t]
\includegraphics[scale=0.5]{figures/overview.pdf}
\caption{The Themisto{} user interface is implemented as a Jupyter Notebook plugin: (A) When the recommended documentation is ready, a lightbulb icon shows up to the left of the currently-focused code cell. (B -- D) shows the three options in the dropdown menu generated by Themisto{}, (B) A documentation candidate generated for the code with a deep-learning model, (C) A documentation candidate retrieved from the online \ac{API} documentation for the source code, and (D) A prompt message that nudges users to write documentation on a given topic.}
\label{fig:interface}
\end{figure}
Figure \ref{fig:interface} shows the user interface of Themisto{} as a Jupyter Notebook plugin.
Each time the user changes their focus on a code cell, as they may be inspecting or working on the cell, the plugin is triggered.
The plugin sends the user-focused code cell's content to the backend. Using this content, the backend generates a code summarization using the model and retrieves a piece of documentation from the \ac{API} webpage.
When such a documentation generation process is done, the generated documentation is sent from the server-side to the frontend, and a light bulb icon appears next to the code cell, indicating that the there are recommended markdown cells for the selected code cell (as shown in Figure \ref{fig:interface}.A).
When a user clicks on the light bulb icon \revision{which appears next to any selected code cells, Themisto{} render all the three options in the dropdown menu}: (1) a deep-learning-based approach to generate documentation for source code (Figure \ref{fig:interface}.B); (2) a query-based approach to retrieve the online \ac{API} documentation for source code (Figure \ref{fig:interface}.C); and (3) a user prompt approach to nudge users to write more documentation (Figure \ref{fig:interface}.D).
If the user likes one of these three candidates, they can simply click on one of them, and the selected documentation candidate will be inserted into above the code cell (if it is the Process, Reference, or Reason type), or below it (if it is the Result type).
\begin{figure}[t]
\includegraphics[width=\columnwidth]{figures/workflow.pdf}
\caption{An illustration of the three different approaches for documentation generation in Themisto{}.}
\label{fig:workflow}
\end{figure}
\subsection{Three Approaches for Documentation Generation}
In this subsection, we describe the rationale and implementation detail of the three different approaches for documentation generation (Figure \ref{fig:workflow}):
\begin{itemize}
\item Our formative study suggests that the system should be able to \textbf{generate multiple types of documentation} (e.g., Process, Result, \revision{Background Knowledge}, Reason, and Reference).
\item Some types of documentation can be \textbf{directly derived from the code}, thus the automated approaches can help. The Process type of documentation directly describes the coding process, and existing ML literature suggest that the deep-learning-based approach is most suitable for generating it; The Reference type does not need a learning-based approach, it can be achieved with a traditional query-based approach, which locates and retrieves the most relevant online documentation as candidates;
\item Some others types of documentations (e.g., Education, Result, and Reason) are \textbf{not directly related to the code}, thus the fully automated approaches are not capable of generating such contents. We design the prompt-based approach for users to complete the generation process.
\end{itemize}
\subsubsection{Deep-Learning-Based Approach}
We trained a \revision{deep learning model}\footnote{\revision{We release a larger dataset and a refined version of the model in a separate paper \cite{liu2021haconvgnn}.}} using the Graph-Neural-Network architecture based on LeClair et al. \cite{leclair2020improved}.
These \ac{GNN} models can take both the source code's structure (extracted as \ac{AST}) and the source code's content as input.
Thus, it outperforms the traditional sequence-to-sequence model architectures, which only takes the source code's content as an input sequence, in source code summarization tasks for Python code\footnote{All the collected data science notebooks are in Python.}. We did not consider T5, BerT or GPT-3 architectures as these models can take minutes to make one inference (i.e., generate one summary) even with a cluster of GPUs (costing thousands of dollars per hour), whereas our GNN-based model can make an inference within a second with one GPU.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{figures/model.pdf}
\caption{\revision{A code summarization model for the deep-learning-based documentation generation approach via \ac{GNN}. There are three steps of data pre-processing (1) We first extract text code pairs from existing notebooks. (2) We generate \ac{AST} from code. (3) We tokenized each word and translated them into embeddings. And (4), the \ac{GNN} model architecture.}}
\label{fig:model}
\end{figure}
In order to fine-tune the model,
we constructed a training dataset for our particular context.
We collected the top 10\% highly-voted notebooks from two popular Kaggle competitions -- House Price Prediction and Titanic Survival Prediction (N = 1158).
For each of the notebook, we first extracted code cells and the markdown cells adjacent above as a pair of input and output (similar to the data collection approach in \cite{agashe2019juice}).
If there is an inline comment in the first line of the code cell, we replaced the output of the pair using the inline comment.
In total, our dataset has 5,912 pairs of code and its corresponding documentation.
Following the best practice of model training, we split the dataset into training, testing, and validation subsets with an 8 to 1 to 1 ratio.
Before feeding data into the training process, we have a three-step pre-processing stage, as illustrated in Figure~\ref{fig:model}. Step 1 removes the style decoration, formats, and special characters that are not in Python grammar (e.g., Notebook Magics).
We also generate an \ac{AST} for the source code input in step 2 with Python AST library\footnote{https://docs.python.org/3/library/ast.html}. The AST result is equivalent to the source code but with more contextual and relational information.
In step 3, we tokenize source code to a sequence of tokens with an input dictionary, and parse the \ac{AST} nodes as a sequence of tokens with the same input dictionary. We parse the relationship between AST nodes as a matrix of edges.
Finally, we tokenize the output documentation as a sequence of tokens with a separate output dictionary.
After this process, all the tokens are transformed into an array of word embeddings --- vectors of real numbers.
We use these data to train the network for 100 epochs, 30 batch sizes, and 15 early stop points on a two Tesla V100 GPU cluster.
Out of all the epochs, we selected the model with the highest validation accuracy score.
To evaluate our model's performance against baseline models, we conducted both quantitative and qualitative evaluations, as suggested by~\cite{ribeiro2020beyond}.
For the automated quantitative evaluation, we use BLEU scores \cite{papineni2002bleu} as the model performance metric.
BLEU scores are commonly used in the source code summarization tasks.
It evaluates the word similarity between the generated text and the ground truth text.
We selected and trained Code2Seq model~\cite{alon2018code2seq} and Graph2Seq model~\cite{xu2018graph2seq} with the same data split.
Our model achieves 11.41 (BLEU--a), which outperforms the baseline models Code2Seq (BLEU--a = 9.61) and Graph2Seq (BLEU--a = 11.05). These scores suggest that the data science documentation task is more difficult than the benchmark code summarization tasks in the software engineering field.
For example, in data science, a notebook code cell can contain multiple code snippets and functions.
\begin{table}[t!]
\small
\centering
\caption{\revision{Example output from the model. (Example A) The generated text well describes the code. (Example B) The generated text vaguely describes the code. (Example C) The generated text is poorly readable, but still captures the keywords of the descriptions.}}
\label{tab:example_output}
\begin{tabular}{p{2cm}p{7cm}p{3.8cm}}
\toprule
\textbf{Example} & \textbf{Code Cell} & \textbf{Output From the Model}\\
\midrule
Example A &
\begin{tabminted}{python}
train = pd.read_csv('./house-input/train.csv')
test = pd.read_csv('./house-input/test.csv')
\end{tabminted}
& Read the data
\\
\hline
Example B &
\begin{tabminted}{python}
all_data = pd.get_dummies(all_data)
\end{tabminted}
& Convert all the data\\
\hline
Example C &
\begin{tabminted}{python}
pred = Tree_model.predict(x_test)
pred = pd.DataFrame(pred)
pred.columns = ["ConfirmedCases_prediction"]
\end{tabminted}
& Predicate to use a predict function for tests \\
\bottomrule
\end{tabular}
\end{table}
In addition to the automated quantitative evaluation, we also conduct a qualitative analysis of the generated documentation pieces.
We found that despite the word-to-word similarity score is low, the general quality of the content is reasonable and satisfying for building a prototype system.
As an illustration, we provide three examples with both input and model generated outputs, as shown in Table \ref{tab:example_output}.
In the Appendix, we provide full code cells and model-generated outputs for the two experimental notebooks that we used in the user study.
\subsubsection{Query-Based Approach}
Our formative study indicates that the well-documented Kaggle notebooks often have the description of frequently-used data science code functions for educational purposes.
And sometimes data scientists directly paste in a link or a reference to the external \ac{API} documentation for a code function.
Thus, we implement a query-based approach that curates a list of \acp{API} from commonly used data science packages, and the short descriptions from external documentation sites.
In our system, we only cover Pandas\footnote{\url{https://pandas.pydata.org/docs/reference/index.html}}, Numpy\footnote{\url{https://numpy.org/doc/stable/reference/}}, and Scikit-learn\footnote{\url{https://scikit-learn.org/stable/modules/classes.html}} these three libraries as a starting point to explore this approach.
We argue that it can be easily expanded to include other packages in the future.
We collected both the \ac{API} names and the short descriptions by building a crawling script with Python.
When users trigger this query-based approach for a code cell, Themisto{} matches the \ac{API} names with the code snippets and concatenate all the corresponding descriptions.
\subsubsection{Prompt-Based Approach}
Lastly, the system also provides a prompt-based approach that allows users to manually create the documentation.
Because our formative study found that a well-documented notebook not only documents the process of the code, but also interprets the output, and explains rationales.
These types of documentation are hard to generate with automated solutions
To achieve it, we implement a prompt-based approach.
It detects whether the code cell has a cell output or not: if the cell outputs a result, Themisto{} assumes that the user is more likely to add interpretation for the output result, thus the corresponding prompt will be inserted below the code cell.
Otherwise, the system assumes the user may want to insert a reason or some educational types of documentations, thus it changes its prompt message.
\section{User Evaluation of Themisto}
To evaluate the usability of Themisto{} and its effectiveness in supporting data scientists to create documention in notebooks, we conducted a within-subject controlled experiment with 24 data scientists.
The task is to add documentation to the given notebook. And each participant is asked to finish two sessions, one with the Themisto{} support and one without its support.
The evaluation aims to understand (1) how well Themisto can facilitate documenting notebooks and (2) how data scientists perceive the three approaches that are used by Themisto{} for generating documentation.
\subsection{Participants}
\revision{We recruited 24 data science professionals as our evaluation participants in a multinational IT company.}
We used a snowball sampling approach to recruit participants, where we sent recruitment messages to friends and colleagues, various internal mailing-lists, and Slack channels.
We then asked participants to refer their friends and colleagues.
Our recruitment criteria are that the participant must have had experience in data science projects and they are familiar with Python and Jupyter Notebook environment.
As shown in Table \ref{tab:demographic}, participants reported a diverse job role backgrounds, including expert data scientists (N = 8), novice data scientists (N = 9), \acp{AIOP} or \ac{ML} engineers (N = 2), subject matter experts (N = 1), and application developer (N = 4).
\begin{table}[t!]
\caption{Demographics of participants}
\label{tab:demographic}
\begin{tabular}{c c p{4cm}p{6cm}}
\toprule
\textbf{PID}& \textbf{Gender} & \textbf{Job Role} & \textbf{Work Experience in Data Science}\\
\midrule
P1 & M & Expert Data Scientist & 5-10 years \\
P2 & M & Application Developer & 3-5 years \\
P3 & M & Novice Data Scientist & less than 3 years \\
P4 & M & Novice Data Scientist & 0 year (just start learning data science) \\
P5 & M & AI Operator or ML Engineer & 3-5 years \\
P6 & M & Novice Data Scientist & less than 3 years \\
P7 & M & Application Developer & 3-5 years \\
P8 & M & Novice Data Scientist & less than 3 years \\
P9 & F & Expert Data Scientist & 3-5 years \\
P10 & M & Expert Data Scientist & 5-10 years \\
P11 & M & Expert Data Scientist & more than 10 years \\
P12 & F & Novice Data Scientist & 3-5 years \\
P13 & F & Expert Data Scientist & 5-10 years \\
P14 & M & Novice Data Scientist & 0 year (just start learning data science) \\
P15 & M & Expert Data Scientist & more than 10 years \\
P16 & M & AI Operator or ML Engineer & 3-5 years \\
P17 & M & Subject Matter Expert & 3-5 years \\
P18 & M & Expert Data Scientist & more than 10 years \\
P19 & F & Application Developer & 3-5 years \\
P20 & F & Expert Data Scientist & 3-5 years \\
P21 & M & Novice Data Scientist & 3-5 years \\
P22 & M & Novice Data Scientist & less than 3 years \\
P23 & M & Application Developer & less than 3 years \\
P24 & M & Novice Data Scientist & less than 3 years \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Study Protocol}
We conducted a within-subject controlled experiment with 24 data scientist participants.
Their task was to add documentation to a given draft notebook, which only has code and no documentation at all. The participants were told that they were adding documentation for the purpose of sharing those documented notebooks as tutorials for data science students who just started learning data science.
Each participant is asked to finish two sessions, one with the Themisto{} support (Experiment condition) and one without its support (Control Condition).
We prepared two draft notebooks, one for each session, shown in the Appendix.
\revision{The two experiment notebooks are adapted from winning notebooks from two Kaggle challenges, which are not included in the model training dataset: 1) House Price Prediction\footnote{\url{https://www.kaggle.com/c/house-prices-advanced-regression-techniques}}; 2) COVID Case Prediction\footnote{\url{https://www.kaggle.com/c/covid19-global-forecasting-week-1}}.
The two notebooks have the same length (9 code cells) and a similar level of difficulties.
Although the two notebooks are simplified versions from winning notebooks, they cover most stages in data science lifecycles.
In addition, the length of the notebooks falls into the middle range of the notebook length distribution on the GitHub corpus (as refer to Figure \ref{fig:rule}).}
To counterbalance the order effect, we randomized the order of the control condition and the experiment condition for each participant, so some participants encountered Themisto{} in their first session, some others experienced it in their second session.
Each participant was given up to 12 minutes (720 seconds) to finish one session.
\revision{We conducted three pilot run sessions, and all the three pilot participants were able to finish a single task within 10 minutes, with or without the support of Themisto{}.}
Before the experiment condition session, we gave the participant a 1-minute quick demo on the functionality of Themisto{}.
All study sessions were conducted remotely via a teleconferencing tool.
We asked the participants to share their screen and we video recorded the entire session with their permission.
After finishing both sessions, we conducted a post-experiment semi-structured interview session to ask about their experience and feedback.
We had a few pre-defined questions such as \say{How do you compare the experience of the documenting task with or without the support of Themisto{}?} or \say{Did you notice the multiple candidates in the dropdown menu? Which one do you like the most?}
In addition, participants were encouraged to tell their stories and experience outside these structured questions.
The interview sections of the video recordings were transcripted into text.
\subsection{Data Collection and Measurements}
We have three data sources: the observational notes and video recording for each session (N = 48), the final notebook artifact out of each session (N = 48), and the post-task questionnaire and interview transcripts (N = 24).
Our first group of measurements are from coding participants' behavioral data from the session recordings.
In particular, we counted \emph{the task completion time} (in secs) for all sessions.
Then, for experiment condition only, we also counted the followings: how many times a participant clicked on the light bulb icon to check for suggestions (\emph{code cells checked for suggestions}); how many times a participant directly used the generated documentation (\emph{markdown cells created by Themisto{}}); how many times a participant ignored the generated recommendations and manually created the documentation (\emph{markdown cells created by human}); and how many time a markdown cell is co-created by human and Themisto{}(\emph{markdown cells co-created by human and Themisto{}}).
The result of this analysis is reported in Table \ref{tab:usage} and \ref{tab:performance}.
Secondly, to evaluate the quality of the final notebook artifact, we define our second group of measurements by counting: the \emph{number of added markdown cells}, and the \emph{number of added words} as these two are indicators of the quantity and effort each participant spent in a notebook. \revision{Also, we asked the participants to give a self-reported satisfaction score to each of the two documented notebooks.
We considered that score (-2 to 2) as a self-reported subject feeling of the notebook satisfaction.
In parallel, we asked two experts to rate the notebook-level quality (N = 48) with a 3-dimensional rubric (based on \cite{garousi2013evaluating}) to evaluate the documentation's \emph{readability, accuracy, and informativeness} in a notebook.
We considered these three scores (-2 to 2) as an objective quality of the notebook. }
Readability concerns whether the documentation is in readable English grammar and words, while accuracy concerns how the documentation matches the code content, and informativeness evaluates whether the documentation covers more information units.
Two experts iteratively discussed and evaluated the notebooks until the independent ratings achieved an agreement ($\alpha$ = 0.76, Krippendorff's alpha).
The result of this analysis is reported in Table \ref{tab:usage}.
For the experiment session only, we conducted a cell-level expert rating (N = 194) using the same approach as in notebook-level expert rating.
Two experts iteratively discussed and evaluated the notebooks until the independent ratings achieved an agreement ($\alpha$ = 0.88, Krippendorff's alpha).
The result of this analysis is reported in Table \ref{tab:performance}.
In addition, we asked the participants to finish a post-experiment survey (5-point Likert Scale, -2 as strongly disagree and 2 as strongly agree, Figure~\ref{fig:questionnaire}) to collect their feedback specific on the system's \emph{usability, accuracy, trust, satisfaction, and adoption propensity} (based on \cite{weisz2019bigbluebot}).
Lastly, for the interview transcripts, four researchers of this research project conduct an iterative open coding method to get the code, theme, and representative quotes as the third group of data.
They each independently coded a subset of interview transcripts, and discussed the codes and themes together.
After the discussion, they when back and reiterated the coding practice to apply the codes and themes to their assigned notebooks.
Some examples of the identified themes are: pros and cons of Themisto{}; preference of the three document generation approaches; future adoption, and suggestions for design improvement.
We will report the qualitative results as supporting materials together with reporting the quantitative results.
\subsection{Results}
\begin{table}[t]
\caption{\revision{Performance data in two conditions (M: mean, SD: standard deviation): the task completion time (secs), participants' satisfaction with the final notebook (from -2 to 2), graded notebook quality, number of markdown cells, and number of words. In particular, participants spent less time to complete the task in the experimental condition than the control condition (p = .001); participants were more satisfied with the final notebook in the experimental condition than the control condition (p = .04).}}
\label{tab:usage}
\centering
\footnotesize
\begin{tabular}{p{5.5cm}cccl}
\toprule
&\textbf{Condition} & \textbf{M} & \textbf{SD} & \\
\midrule
\vspace{7pt}
\multirow{2}{*}{Number of Added Markdown Cells}
& Experiment & 8.04 & 2.40 &
\multirow{2}{*}{
\includegraphics[width=3.5cm]{figures/plot2_markdown.png}} \\
& Control & 7.79 & 1.91 & \\
\midrule
\vspace{7pt}
\multirow{2}{*}{Number of Added Words}
& Experiment & 95.75 & 50.56 &
\multirow{2}{*}{
\includegraphics[width=3.5cm]{figures/plot2_words.png}} \\
& Control & 100.92 & 53.27 & \\
\midrule
\vspace{7pt}
\multirow{2}{*}{\textbf{**Task Completion Time (secs)}}
& Experiment & 391.12 & 200.15 &
\multirow{2}{*}{
\includegraphics[width=3.5cm]{figures/plot2_time.png}} \\
& Control & 494.75 & 184.28 & \\
\midrule
\vspace{7pt}
\multirow{2}{*}{\textbf{*Satisfaction with the Final Notebook (-2 to 2)}}
& Experiment & 0.96 & 0.69 &
\multirow{2}{*}{
\hspace{-6pt}
\includegraphics[width=3.5cm]{figures/satisfaction.png}} \\
& Control & 0.54 & 0.83 & \\
\midrule
\vspace{7pt}
\multirow{2}{*}{Expert Rating: Accuracy (-2 to 2)}
& Experiment & 1.60 & 0.47 &
\multirow{2}{*}{
\hspace{-4pt}
\includegraphics[width=3.5cm]{figures/accuracy.png}} \\
& Control & 1.62 & 0.52 & \\
\midrule
\vspace{7pt}
\multirow{2}{*}{Expert Rating: Readability (-2 to 2)}
& Experiment & 0.65 & 0.83 &
\multirow{2}{*}{
\hspace{-4pt}
\includegraphics[width=3.5cm]{figures/presentation.png}} \\
& Control & 0.90 & 0.57 &\\
\midrule
\vspace{7pt}
\multirow{2}{*}{Expert Rating: Informativeness (-2 to 2)}
& Experiment & 0.67 & 0.64 &
\multirow{2}{*}{
\hspace{-4pt}
\includegraphics[width=3.5cm]{figures/completeness.png}} \\
& Control & 0.75 & 0.63 & \\
\bottomrule
\end{tabular}
\end{table}
In this section, we present the user study results on: how Themisto{} improved participants' performance on the task, how participants perceived the documentation generation methods in Themisto{}, and how participants described the practical applicability of Themisto{}.
\subsubsection{Themisto{} supports participants to easily add documentations to a notebook. }
Our experiment revealed that Themisto{} improved participants' performance on the task by reducing task completion time and improving the satisfaction with the final notebooks.
We performed a two-way repeated measures ANOVA to examine the effect of the two notebooks and the two conditions (with or without Themisto{}) on task completion time.
As shown in Table \ref{tab:usage}, participants spent significantly \textbf{less time} (p<.001) to complete the task using Themisto{} in the experiment condition (M(SD) = 391.12 (200.15)) than in the control condition (M(SD) = 494.75 (184.28)).
In addition, there was not a statistically significant effect of notebooks on task completion time, nor a statistically significant interaction between the effects of notebooks and conditions on completion time.
The post-experiment survey result supported our findings. Most participants believed it was easier to accomplish the task with Themisto{}'s help (22 out of 24 rated agree or higher), as shown in Figure \ref{fig:questionnaire}.
And the Themisto{} generated recommendation was accurate (20 out of 24 rated agree or higher).
Looking into the qualitative interview data, we can find some potential explanations for why participants believed so.
Participants reported that Themisto{} provided them something to begin with, thus it was easier than starting from scratch: \inlinequote{The plugin makes it easy to just pick it and have something simple. And then I got a couple of times where I went back and said, \say{Oh let me add a few more words.}} (P21).
\begin{figure}[t]
\includegraphics[width=\textwidth]{figures/questionnaire5.pdf}
\caption{\revision{Results of the post-task questionnaire. Note that the disagrees are not from the same participant.}}
\label{fig:questionnaire}
\end{figure}
\subsubsection{Co-creation yields longer documentation and improves accuracy and readability.}
Through coding the video recordings for only the experiment-condition sessions,
we were able to examine the following questions: while the Themisto{} was available, how did the participants use it? Did they check the recommendations it generated? Did they actually use those recommendations in their documentations added into notebooks?
As shown in Table \ref{tab:performance}, we found that while Themisto{} is available, for 86.11\% of code cells, participants checked the recommended documentation by clicking on the light bulb icon to show the dropdown menu.
Then, 46.90\% of the created markdown cells were directly adopted from Themisto{}'s recommendation; while 11.86\% of the created markdown cells were manually crafted by humans alone.
\textbf{The most interesting finding is that 41.24\% markdown cells were co-created by Themisto{} and human participants together}: Themisto{} suggests a markdown cell, human participants take it, and modify on top of it.
This result suggested that most participants used Themisto{} in the creation of documentation, and some of them formed a small collaboration between humans and the \ac{AI}.
This finding inspires us to further explore how participants co-create the documentation with Themisto{} \cite{muller2021hai}.
By looking at the log data, we discover several editing patterns.
For example, many participants added supplemental details (e.g., expanding the steps into substeps) to Themisto's suggested documentation.
Participants also added stylistic edits, including modifying document hierarchies, polishing sentences, and changing conversational tones.
In order to explore the differences among documentation created by three methods (created by Themisto{} only, co-created by human and Themisto{}, created by human only), we conducted a cell-level expert rating (N = 194) along the dimension of accuracy, readability, and informativeness.
We also calculated the word count of the documentation length.
We performed a one-way ANOVA to examine the differences among the three groups.
As shown in Table \ref{tab:performance}, markdown cells that are co-created by humans and Themisto{} have significantly more word count (M(SD) = 15.45 (10.97)) than markdown cells that are manually written by humans alone (M(SD) = 10.26 (7.41)) and the markdown cells that are directly adopted from Themisto{}'s recommandation (M(SD) = 8.88 (7.14)), with F = 11.83, p < 0.001 .
Markdown cells co-created by humans and Themisto{} also yield better results in terms of accuracy (F = 9.43, p < 0.001) and readability (F = 3.28, p = 0.04), while for informativeness, there is no significant differences across three groups.
Our posthoc analysis suggested that no significant differences were found between markdown cells created by Themisto{} and markdown cells created by humans only along all dimensions (including word count, accuracy, readability, and informativeness).
\begin{table}[t]
\caption{Usage data of the plugin in experimental condition. The results indicate that participants used the plugin for recommended documentation on most code cells (86.11\%). For markdown cells in the final notebooks, 46.90\% were directly adopted from the plugin's recommendation, while 41.24\% were modified from the plugin's recommendation and 11.86\% were created by participants from scratch.}
\label{tab:performance}
\scriptsize
\begin{tabular}{p{4.5cm}p{0.2cm}p{0.5cm}p{1.6cm}p{1.25cm}p{1.2cm}p{1.8cm}}
\toprule
& \textbf{$N$} & \textbf{$\%$} & \textbf{Word Count} & \textbf{Accuracy} & \textbf{Readability} & \textbf{Informativeness} \\
\midrule
Code Cells Checked for Suggestions & 186 & 86.11
& - & - & - & - \\
Markdown Cells Created by Themisto{} Only & 91 & 46.90
& 8.88 (7.14) & 1.36 (0.54) & 1.94 (0.23)& 1.34 (0.56) \\
Markdown Cells Co-created by Humans and Themisto{} & 80 & 41.24
& 15.45 (10.97) & 1.68 (0.47) & 1.96 (0.17) & 1.51 (0.55)\\
Markdown Cells Created by Humans Only & 23 & 11.86
& 10.26 (7.41) & 1.28 (0.69) & 1.83 (0.36) & 1.35 (0.65)\\
\bottomrule
\end{tabular}
\end{table}
\subsubsection{Themisto{} increases participant's satisfaction , while maintaining a similar quality of the final notebook.}
The post-task questionnaire revealed that participants were \textbf{more satisfied} with the final notebook after using Themisto{} in the experiment condition than in the control condition (p = .04) (Table \ref{tab:usage}).
The interview results also supported this finding.
P14 believed that Themisto{} helped with wording: \inlinequote{Sometimes I knew what the cells were doing but I did not know how to put things in a really good sentence for others.}
Themisto{} also motivates participants to document the analysis details.
Although we did not see a difference in the number of markdown cells created in two conditions or the number of words in total, Themisto{} helps them overcome the procrastination of writing documentation and reminds them to document things that they might ignore.
\begin{quote}
I think I definitely overlooked some details when I was commenting without the tool, because I just made the assumption that people should know from the code... To be honest, I do not usually follow a good coding practice. My notebooks are really messy and I am the only person who can understand it. I feel sorry for anybody else that has to see it. (P19)
\end{quote}
Moreover, participants believed that Themisto{} can help them form a better documenting practice in the long term: \inlinequote{It very useful to remind me to always put some documentation in a timely manner.} (P13).
The two experts' gradings for the notebook quality suggest that there was not a significant difference for the three dimensions of the quality rubric (\textbf{accuracy, readability, and completeness}).
In the post-task interview, participants mentioned that the accuracy of the generated recommendations plays a role in participant's experience: \inlinequote{My experience with the plugin is definitely better. For the most part, the suggestions are pretty accurate. Although sometimes I did make a few minor changes like rearranging the text.} (P5).
Some participants also mentioned that they needed to edit the format of the generated document to fit their context.
We believe that while Themisto{} offers convenience to improve the data scientists' productivity and saves their time, it may not provide the same level of readability as those notebooks well articulated by humans.
Thus, data scientists may want to further revise the formatting and wording of the Themisto{} generated documentation.
In summary, our experiment indicated that Themisto{} improves participants' productivity for creating documentation.
It also increases their perceived satisfaction with the final notebook, compared to the notebooks written by participants themselves.
\subsubsection{The three approaches of generating documentation are suitable for different scenarios.}
In this section, we have an in-depth analysis of how participants perceived the three different approaches that Themisto{} implemented to generate documentations: the deep-learning-based approach, the query-based approach, and the prompt-based approach.
In the post-experiment interview, we explained how Themisto{} generated the documentation with these three approaches, and asked participants if they like or dislike one particular approach.
Participants reported that they felt the deep-learning-based approach provided concise and general descriptions of the analysis process: \inlinequote{I think the AI suggestion gives me an overview. It is short, and has some useful keywords.} (P12).
Participants also suggested that the deep-learning-based approach sometimes generated inaccurate or very vague documentation: \inlinequote{The first one gives me a very short summary, though it didn't always say what the cell is doing.} (P1).
But the deep-learning-based approach is still perceived useful.
As it is short and with only a couple of keywords, many participants believe it may be more suitable for some quick and simple documentation task, or for the analyst audience who can understand these short keywords.
In terms of the query-based approach, participants believed that the documentation generated from this approach contains has longer and more descriptive information.
This approach is further perceived to be more suitable for educational purposes: \inlinequote{This one gives you really good information. For some specific methods or calls, you don't have to come up with a high-level summary for others and you can directly use it.} (P14).
Participants also acknowledged that such a query-based approach may not work for some scenarios.
For example, participants found that the query-based approach was not useful for summarizing the very fundamental level data manipulations, as there was no core \ac{API} method in it.
Some participants mentioned that the usefulness of this query-based approach depends on the audience.
\begin{quote}
The [deep-learning-based approach] was really useful. The [query-based approach]... it depends on the audience. It is much more appropriate for a novice programmer. (P18)
\end{quote}
We observed in the video recordings that participants rarely used the prompt-based approach in the session.
The interview data confirmed our speculation.
Some participants said that they liked the idea of user prompts, but they did not use it because the deep-learning-based approach and the query-based approach already gave them the actual content.
Other participants pointed out that the prompts were not intelligent enough, so they did not use it: \inlinequote{It always asks the same thing and I just ignored the prompts.} (P18).
Participants suggested that the prompts could be designed to better fit the context.
\begin{quote}
Perhaps the system can infer what the code cell was doing [from the deep-learning-based approach], and show prompts accordingly. Like if I delete a data point from the dataset, there is a prompt asking why I considered it as an outlier or something. (P5)
\end{quote}
Last but not the least, many participants preferred a hybrid approach to combine the deep-learning-based approach and the query-based approach.
For example, P12 mentioned,
\begin{quote}
The first one (deep-learning-based) tells me what the code cell is doing in general and the second one (query-based) tells me the details of the function. I would go with a hybrid approach. (P12)
\end{quote}
\subsubsection{Will participants use Themisto{} in their future data science project?}
Most participants indicated that they would like to use Themisto{} in the future when answering the survey question as shown in Figure \ref{fig:questionnaire}.
The interview data provides more detail and evidence to elaborate on this result.
Participants suggested various scenarios in which the Themisto{} could be useful in their future work, such as they need to add documentation during the exploration process for future selves, or they need to document a notebook in a post-hoc way for sharing it with collaborators, or they need to mentor a team member who is a novice data scientist, or they need to refactor an ill-documented notebook written by others:
\begin{quote}
When I am doing data analysis, I tend to write the code first because there is a flow in my head of what I need to do. And then I will go back afterward to use the plugin and add the comments needed. I will definitely do that before sharing that file or handing it over to others. (P12)
\end{quote}
There was one participant who did not think Themisto{} could fit into his workflow: \inlinequote{I always write documentation before writing code. Maybe Themisto{} does not work for people like me.} (P3)
In our experiment, we provided the scenario as they were documenting the notebook as a tutorial for some data science students.
In the interview, we asked how participants would document a notebook differently if they were created documentation for the notebooks for non-technical domain experts audience.
Some participants suggested that computational notebooks may not be a good medium to present the analysis to non-technical domain experts.
They would prefer to curate all the textual annotations in a standalone report or slide decks.
Some others believed that the notebook could work as the medium but they would change the documentation by using less technical terminology, adding more details on topics that the non-technical domain experts would be interested in (e.g., how data is collected, potential bias in the analysis).
\subsubsection{Participants suggest various design implications for automated code documentation.}
In the interview, participants provided various design suggestions to improve Themisto{} and to design future technologies that can support data scientists to document the notebook.
Participants expected Themisto{} to have more functionalities than simply generating documentation for the code.
For example, P13 proposed maybe Themisto{} can also create a description to document the changes of versions and the editing histories from different team members.
P3 and P4 believed that the automatic generation of Reason is very much needed for explaining decisions such as why selecting a particular algorithm.
P19 wanted the system to automatically add explanations to the execution errors.
Participants also mentioned that Themisto{} should add more varieties into the generated content's formatting.
They would like to see suggested documentation with a better presentation.
And lastly, some participants suggested that maybe such a documentation generation system can take consideration of the purpose of the notebook, the domain-specific terminology, or the indivials' habits for writing documentation.
\subsubsection{Summary of the Results}
In summary, our study found that Themisto{} can support data scientists in generating documentation by significantly reducing their time spent on the task, and improving the perceived satisfaction level of the final notebook.
When Themisto{} is available, participants are very likely to check the generated documentation as a reference.
Many of them directly used the generated documentation, a few of them still prefer to manually type the documentation, while many of them adopted a human-AI co-creation approach that they used the AI-generated one as a baseline and keep improving on top of it.
Participants perceived the documentation generated by the deep-learning-based approach as a short and concise overview, the documentation generated by the query-based approach as descriptive and useful for educational purposes, but they rarely used the prompt-based approach.
Overall, participants enjoyed Themisto{} and would like to use it in the future for various documenting purposes.
\section{Discussion}
\subsection{The Documentation Practices in Data Science is Different from in Software Engineering}
The practice of documentation in data science has both overlaps and strong contrasts in relation to the ones in software engineering in many facets.
Software engineers write inline comments in their work-in-progress code to help collaborators understand the behavior of the code without the burden of going through thousands of lines of code;
they document changes of their code for better version management and improving awareness of their collaborators;
when others need to build upon their work, they write formal documentation and Readme files to describe how to use functions and API in their packages or services~\cite{lowndes2017our,saltz2017comparing,zhang2020data}.
Data scientists write computational narratives as a practice of literate programming~\cite{kery_exploring_2017, Park:2018:PPL:3266037.3266098}, and as a way to think and explore alternatives.
Thus, notebooks often have orphan code cells or out-of-order code snippets, which leads to lower reusability of the notebook and further highlights the importance of documentation in the notebook.
As we found in our formative study, well-documented notebooks explain more than the behavior of the code.
Notebooks cover various topics including describing and interpreting the output of the code, explaining reasons for choosing certain algorithms or models, educating the audience from different levels of expertise, and so on.
Thus, many interventions and lessons learned about documentation in software engineering may not apply in data science context.
For example, how can we evaluate the quality of the documentation?
Software documentation can be assessed based on attributes like completeness, organization, the relevance of content, readability, and accuracy~\cite{garousi2013evaluating}.
Our experiment found that the quality scores assessed by these rubrics does not reflect users' satisfaction with the final notebooks.
Despite many people's efforts to creating a standard documentation practice \cite{rule2019ten, konkol2020publishing}, it remains questionable whether there is a one-size-fits-all solution.
For example, Rule et al. \cite{rule2019ten} suggested ten rules for writing and sharing computational analysis in Jupyter notebooks.
The first rule they proposed is to tell a story to the audience.
However, this description is very general as people may approach storytelling differently.
As we observed in Kaggle notebooks, some notebook authors prefer to use concise and accurate language while others use more colloquial and creative language.
These creative notebooks stand out and receive many votes and compliments from the Kaggle community.
As we recognize documentation in data science as a fluid activity, traditional template-based approaches to aid documentation writing may not work in data science because they can not capture a broader aspects of documentation, and limit the expressiveness of storytelling.
We argue that future work should recognize the difference between data science and software engineering, and tailor the documentation experience for data scientists.
For example, Callisto \cite{wang2020callisto} harnessed the fact that data scientists engage in synchronous work and discussion, and used contextual links between discussion messages and notebook content to aid the explanation of notebooks.
\subsection{Human-AI Collaboration in Code Documentation in Data Science}
\revision{We argue that AI-assisted code documentation process can be viewed as a co-creative process in which machine learning fits into the human workflow and collaborate together to create documents in a notebook. }
The notion of a ``partnership relationship'' between human data scientists and AI has been discussion by Wang et al.~\cite{wang2019human}, and is part of a larger research discussion by many others (e.g.,~\cite{cai2019human,seeber2020machines,malone2018human}).
\revision{We consider this partnership as broadly defined where an AI system does not need an avartar or a conversational interface, but this AI system should be designed to fit into the existing human workflow and assist some parts of the human task to improve the quality or productivity.
Human-AI collaboration, as opposed to human-AI competition (portraited by AlphaGo or DeepBlue), should be the ultimate goal of human-AI interaction research.
Various human-centered AI design principles (such as human-in-the-loop) are means to get to this end goal.
Our study demonstrated another means to achieve human-AI collaboration, where we combined the fully automated neural network approach and the less advanced rule-based or prompt-based approaches.
This design is to acknowledge the limitation of today's neural-network modeling.}
Our result showed that the combined human + AI effort produced a satisfied level of quality at a \emph{more rapid pace} than what human or AI could achieve alone.
We also observed the user interaction pattern in which the AI creates \say{first draft} of the documentation, followed by human review and editing, resulted in a final artifact that not only met the bar for quality, but exceeded it for the level of satisfaction.
Participants were happier with their code documentation when they were assisted by the AI system to create it, rather than when they worked alone.
Thus, we conclude that the benefits of having an AI partner in this task stem from being able to produce the same high level of objective quality, but at a much more rapid pace (20\% faster on average) and with a higher level of satisfaction with the end product.
We speculate that one of the contributing factors for why people were accepting of the AI's suggestions is because the final decision of taking those suggestions was up to the human.
\revision{As an alternative, we could have designed the system to always automatically produce a markdown documentation cell for each identified code cell, but we decided not to.
Because this fully-automated design is an extreme in the framework of automation put forth by Parasuraman et al.~(\cite{parasuraman2000model}; see also \cite{horvitz1999principles}), which people may feel being replaced.
Our results confirmed our assumption --- }participants reported that they enjoyed being able to see multiple suggestions, created using different algorithms, and select the one that was the closest match to their intent in documenting a code cell.
This level of interaction corresponds to \say{AI executes a selection only after a human has approved} in the Parasuraman et al. model \cite{parasuraman2000model}.
Our result also shed light on the research question in~\cite{wang2019human} about the conditions under which human data scientists will enjoy working with AI partnership .
In our case, maintaining control of the initiative and the final decision is an important aspect for people's enjoyment and acceptance of the AI system.
It remains to be studied whether people prefer both to control their own initiative \textit{and} the initiative of a machine teammate, as proposed in Shneiderman's recent two-dimensional model~\cite{shneiderman2020human}.
\revision{Also, we did not focus on the explanability or trust aspect of the designed AI system, such as how to visualize the connection between the generated documentaiton and the original code.
In the future, the explanability and trust aspects of the AI system in the data science context is a very critical research topic (e.g., ~\cite{weidele2020autoaiviz,drozdal2020trust}), and should also be prioritized in the research agenda.}
\revision{There are many other tasks in a data science project's lifecycle that could use AI's help, such as model presentation or feature engineering~\cite{automationsurvey}. In the future, we plan to extend our work to design more human-centered AI systems to support users in these data science tasks as well.}
\revision{In the future, we plan to explore whether the identified benefits and tradeoffs persist or not after a long period of adoption by users.
One of the potential benefits could be: the human-AI collaboration work style helps users to learn more from the AI suggested/reminded documentations, thus they realize more of the value of adding code documentation to notebooks; in contrary, maybe users become over-reliance on AI systems thus they de-skill in this code documentation task, both hypotheses await future research to evaluate.}
\subsection{Design Implications}
We offer designers and tool builders the following suggestions to encourage data scientists to write better documentation:
\subsubsection{Towards Hybrid and Adapted Code Summarization}
Our evaluation of Themisto{} indicates that instead of a fully automatic approach, data scientists prefer to use a hybrid method for helping them write documentation.
We argue that future work for code summarization should investigate a hybrid and adapted approach.
We suggest that \textit{adaptive interactive prompting} may be a worthwhile research topic.
For example, prompts could be based on the contents of the code cell which the user was trying to document.
Another possibility is that prompts be based on the user's own history of writing markdown cells, and could either appeal to the user's strengths, or could anticipate and accommodate the user's weaknesses.
In a more socially-oriented approach, users within an organization might rate the initial set of prompts, voting some prompts up or down depending on their usefulness. An evolution of this idea might allow users to propose new prompts for use by selves and others (e.g., \cite{dimicco2009people}).
Furthermore, we argue that future code summarization tools would benefit from a reinforcement learning approach which learns from users' modifications to the original proposed texts, and could anticipate the users' preference in subsequent documentation.
\subsubsection{Customizing the Recommendations based on Usage Scenarios}
As prior work stated \cite{zhang2020data}, data science workers engage in various collaborations during different stages of the data science lifecycle. Documentation plays an important role in many scenarios.
For example, handing off work between data engineers, communicating results with stakeholders, or informal notes to future self.
Data science workers may have different needs of the documentation for different usage scenarios.
Designers and tool builders should take a user-centered approach to understand the purpose of documentation, the appropriate level of details, and the best way to present the documentation.
For example, participants suggested that future versions of Themisto{} being able to document the changes of versions, errors, and related online forum posts.
Participants also suggested that they would like to see more varieties into the generated narrative's formatting.
\subsubsection{Inverting Themisto{} -- Automatic Code Generation from Documentation.}
The premise of Themisto{} was to generate descriptive material based on program code.
Following some of the ideas in Seeber et al.~\cite{seeber2020machines}, we might invert this strategy.
We recall that P3 told us that he wrote documentation in advance of writing the code itself.
If there are other people who use the same discipline as P3, could we generate code from descriptive text?
We suspect that this idea would not work for just \textit{any} textual description.
However, there could be certain stylized ways of writing descriptions that might be translatable into code; pseudocode could provide a starting point for the design of such a stylized type of description.
We recognize that this kind of approach would need to have a representation of code packages and libraries, so that it could generate code that was appropriately structured for those packages.
Of course, package documentation could be used to construct such a representation.
\subsection{Limitation}
\revision{Our formative study only explores notebooks from the Kaggle corpus, which may leave out some varieties of markdown cells that only exist in messy notebooks that can benefit from the support of documentation generation.
However, notebooks on the Kaggle platforms are based on real data and real problems, and they aim for rich explanations and narratives, where other places do not have high-quality notebooks with rich documentation.
Future work should expand the exploration on the other notebook corpus, for example, notebooks published with scientific paper which contain fine-grained documentation.}
Our experiment has several limitations: it focuses only on the documenting (instead of coding) process, it is a controled experiment study, and participants did not work on notebooks created by themselves.
Thus, for example, we do not know how participants would perceive the usefulness of the tool in \revision{realistic notebooks}, which may be longer and more complicated (e.g., having out-of-order cells) than the notebooks we provided.
However, we believe the result is still promising to shed light for future research and future system design.
Future work can explore the generalizability of Themisto{} through a long-term deployment study.
As for the Human-AI Collaboration research initiative, our work only reports the findings on how human and AI collaborated at a coarse-grained level (Table~\ref{tab:performance}).
In the future work, we will have an indepth analysis to break down to the level of individual cells, and further analyze the difference between automatically-generated, co-edited, or manually-produced cells.
This detailed analysis will help us to understand how human behave and perceive the fine-grained collaboraiton and interaction with the AI partner.
And such findings and its derived design insights could also help researchers who are studying Human-AI Collaborations in other usage scenarios (e.g., in Healthcare or in Educational settings~\cite{xu161same}) beyond the nodebook documentation context in this paper.
\section{Conclusion}
In this paper, we have designed and built Themisto{} to support human data scientists in the notebook documentation task.
This researhc prototype also serves as a prompt to explore the human-AI collaboration research agenda within the automated notebook documentation user scenario.
The system design is driven by insights from previous literature, and also by a formative study that analyzed 80 highly-voted Kaggle notebooks to understand how human data scientists document notebooks.
The follow-up user evaluation suggested that the collaboration between data scientists and Themisto{} significantly reduced task completion time and resulted in a final artifact that not only met the bar of quality, but also exceed it for the level of satisfaction.
\section{Example of Documentation Generation in Themisto}
\begin{landscape}
\begin{table}[h]
\footnotesize
\centering
\caption{\revision{Example notebook (House Prediction), documentation generation in Themisto, and user submitted documentation (T - Markdown Cells Created by Themisto Only, C - Markdown Ceells Co-created by Humans and Themisto, H - Markdown Cells Created by Humans Only).}}
\label{tab:notebook-house}
\begin{tabular}{p{5.5cm}p{2cm}p{3.5cm}p{2cm}p{3.8cm}}
\toprule
\textbf{Source Code} & \textbf{DL-Based} & \textbf{Query-Based} & \textbf{Prompt-Based} & \textbf{P2}\\
\midrule
\begin{tabminted}{python}
import pandas as pd
import numpy as np
from sklearn.linear_model import LassoCV
from sklearn.model_selection
import cross_val_score
\end{tabminted}
& Importing libraries
& Pandas is for data manipulation and analysis; NumPy is a library for ...
& This code cell is for \_ \_ \_ \_ \_
& Importing libraries (T)\\
\midrule
\begin{tabminted}{python}
train = pd.read_csv('train.csv')
test = pd.read_csv('test.csv')
\end{tabminted}
& Read the data
& Read a comma-separated values (csv) file into DataFrame; Return the first 5 rows.
& This code cell is for \_ \_ \_ \_ \_
& Read the data (T)\\
\midrule
\begin{tabminted}{python}
train.head()
\end{tabminted}
& Let's see the values
& Return the first 5 rows
& The table shows \_ \_ \_ \_ \_
& Return the first 5 rows. (defValue=5) (C)\\
\midrule
\begin{tabminted}{python}
all_data = pd.concat((
train.loc[:,'SubClass':'SaleCond'],
test.loc[:,'SubClass':'SaleCond']))
\end{tabminted}
& A generator for feature
& Concatenate pandas objects along a particular axis with optional set logic along the other axes.
& This code cell is for \_ \_ \_ \_ \_
& Concat train and test col "SaleCondition" (C)\\
\midrule
\begin{tabminted}{python}
all_data = pd.get_dummies(all_data)
\end{tabminted}
& Convert all the data
& Convert categorical variable into dummy/indicator variables
& This code cell is for \_ \_ \_ \_ \_
& Convert categorical variable into dummy/indicator variables. (T)\\
\midrule
\begin{tabminted}{python}
all_data = all_data
.fillna(all_data.mean())
\end{tabminted}
& Check the missing values
& Fill NA/NaN values using the specified method
& This code cell is for \_ \_ \_ \_ \_
& \\
\midrule
\begin{tabminted}{python}
X_train = all_data[:train.shape[0]]
X_test = all_data[train.shape[0]:]
y = train.SalePrice
\end{tabminted}
& Create the target and the test data
& Slice string
& This code cell is for \_ \_ \_ \_ \_
& \\
\midrule
\begin{tabminted}{python}
model_lasso = LassoCV(
alphas = [1, 0.1, 0.001, 0.0005])
.fit(X_train, y)
\end{tabminted}
& Model
& Lasso linear model with iterative fitting along a regularization path.
& This code cell is for \_ \_ \_ \_ \_
& Fit regression model (H)\\
\midrule
\begin{tabminted}{python}
def rmse_cv(model):
rmse= np.sqrt(-cross_val_score(
model, X_train, y, scoring=
"neg_mean_squared_error", cv = 5))
return(rmse)
rmse_cv(model_lasso).mean()
\end{tabminted}
& A simple example model with the lasso
& Evaluate a score by cross-validation
& The result indicates that \_ \_ \_ \_ \_
& Define score function and evaluate (H)\\
\bottomrule
\end{tabular}
\end{table}
\end{landscape}
\begin{landscape}
\begin{table}[ht!]
\footnotesize
\centering
\caption{\revision{Example notebook (Covid Prediction), documentation generation in Themisto, and user submitted documentation (T - Markdown Cells Created by Themisto Only, C - Markdown Ceells Co-created by Humans and Themisto, H - Markdown Cells Created by Humans Only).}}
\label{tab:notebook-covid}
\begin{tabular}{p{5.5cm}p{2cm}p{3.5cm}p{2cm}p{3.8cm}}
\toprule
\textbf{Source Code} & \textbf{DL-Based} & \textbf{Query-Based} & \textbf{Prompt-Based} & \textbf{P5}\\
\midrule
\begin{tabminted}{python}
import numpy as np
import pandas as pd
from sklearn.ensemble
import RandomForestClassifier
\end{tabminted}
& Importing libraries
& Pandas is for data manipulation and analysis; NumPy is a library for ...
& This code cell is for \_ \_ \_ \_ \_
& Importing libraries (T)\\
\midrule
\begin{tabminted}{python}
train = pd.read_csv("train.csv")
test = pd.read_csv("test.csv")
train.head()
\end{tabminted}
& Read the data
& Read a comma-separated values (csv) file into DataFrame; Return the first 5 rows.
& The table shows \_ \_ \_ \_ \_
& Read and sanity check the data (C)\\
\midrule
\begin{tabminted}{python}
train.describe()
\end{tabminted}
& Let's see the values
& Generate descriptive statistics. Descriptive statistics include ...
& The table shows \_ \_ \_ \_ \_
& \\
\midrule
\begin{tabminted}{python}
train["Date"] = train["Date"]
.apply(lambda x: x.replace("-",""))
train["Date"] = train["Date"]
.astype(int)
train.head()
\end{tabminted}
& Convert all the data
& Replace a specified phrase with another specified phrase
& The table shows \_ \_ \_ \_ \_
& Preprocess the data (C)\\
\midrule
\begin{tabminted}{python}
train.isnull().sum()
\end{tabminted}
& Check the missing values
& Detect missing values for an array-like object
& The result indicates that \_ \_ \_ \_ \_
& Check the missing values (T)\\
\midrule
\begin{tabminted}{python}
test["Date"] = test["Date"]
.apply(lambda x: x.replace("-",""))
test["Date"] = test["Date"]
.astype(int) \end{tabminted}
& Convert all the data
& Replace a specified phrase with another specified phrase
& This code cell is for \_ \_ \_ \_ \_
& Preprocess the date column (C)\\
\midrule
\begin{tabminted}{python}
x = train[['Lat', 'Long', 'Date']]
y = train[['ConfirmedCases']]
x_test = test[['Lat', 'Long', 'Date']]
\end{tabminted}
& Create the target and the test data
& Select subsets of data
& This code cell is for \_ \_ \_ \_ \_
& Create the train/test data and the target (C)\\
\midrule
\begin{tabminted}{python}
Tree_model = RandomForestClassifier(
max_depth=200,
random_state=0)
Tree_model.fit(x,y)
\end{tabminted}
& Model
& A random forest is a meta estimator that fits a number of decision tree classifiers on ...
& This code cell is for \_ \_ \_ \_ \_
& Define and configure the model
A random forest is a meta ... We also train the model with `.fit()` (C)\\
\midrule
\begin{tabminted}{python}
pred = Tree_model.predict(x_test)
pred = pd.DataFrame(pred)
pred.columns =
["ConfirmedCases_prediction"]
\end{tabminted}
& Predicate to use a predicate function for tests
& A random forest is a meta estimator that fits a number of decision tree classifiers on ...
& This code cell is for \_ \_ \_ \_ \_
& Run the model to generate predictions on the test data and store them as a `DataFrame`(H) \\
\bottomrule
\vspace{1mm}
\end{tabular}
\end{table}
\end{landscape}
\pagebreak
\section{Coding Book for the Interview Transcripts}
\begin{table}[h!]
\centering
\caption{Coding Book for the Interview Transcripts}
\label{tab:codebook}
\small
\begin{tabular}{lp{6.5cm}}
\toprule
\textbf{Theme} & \textbf{Code} \\
\midrule
\multirow{8}{*}{Pros of Themisto}
& Easy to Use \\
& Provie Inspirations \\
& Improve Content \\
& Efficiency \\
& Hybrid Approach \\
& Useful for Long Term \\
& Prefer the Plugin \\
\midrule
\multirow{2}{*}{Cons of Themisto}
& Inaccurate \\
& Not Useful \\
\midrule
\multirow{6}{*}{Perceptions of the Deep-Learning-Based Approach}
& Concise \\
& Useful \\
& Accurate \\
& Inaccurate \\
& For Own Use \\
& For Collaboration Use \\
\midrule
\multirow{5}{*}{Perceptions of the Query-Based Approach}
& Descriptive \\
& Too Long \\
& Useful \\
& Confusing \\
& Instructive \\
\midrule
\multirow{3}{*}{Perceptions of the Prompt-Based Approach}
& Tedious \\
& Easy to Use \\
& Inspiring \\
\midrule
\multirow{3}{*}{Future Adoption}
& Positive Adoption Propensity \\
& Scenarios for Future Adoption \\
& Negative Adoption Propensity \\
\midrule
\multirow{6}{*}{Design Improvements}
& More Options Generated by AI \\
& Handle Presentation and Formatting \\
& Summarize Other Information (e.g., Reasons, Summary, Errors) \\
& Custimization \\
& Optimize UI \\
& Adaptive Prompts \\
\bottomrule
\end{tabular}
\end{table}
\section{Example of Documentation Generation in Themisto}
\begin{table}[h]
\small
\centering
\caption{\revision{Example notebook (House Prediction), documentation generation in Themisto, and user submitted documentation (T - Markdown Cells Created by Themisto Only, C - Markdown Ceells Co-created by Humans and Themisto, H - Markdown Cells Created by Humans Only).}}
\label{tab:notebook-house}
\begin{tabular}{p{5.5cm}p{2cm}p{3.5cm}p{2cm}p{3.8cm}}
\toprule
\textbf{Source Code} & \textbf{DL-Based} & \textbf{Query-Based} & \textbf{Prompt-Based} & \textbf{P2}\\
\midrule
\begin{tabminted}{python}
import pandas as pd
import numpy as np
from sklearn.linear_model import LassoCV
from sklearn.model_selection
import cross_val_score
\end{tabminted}
& Importing libraries
& Pandas is for data manipulation and analysis; NumPy is a library for ...
& This code cell is for \_ \_ \_ \_ \_
& Importing libraries (T)\\
\midrule
\begin{tabminted}{python}
train = pd.read_csv('train.csv')
test = pd.read_csv('test.csv')
\end{tabminted}
& Read the data
& Read a comma-separated values (csv) file into DataFrame; Return the first 5 rows.
& This code cell is for \_ \_ \_ \_ \_
& Read the data (T)\\
\midrule
\begin{tabminted}{python}
train.head()
\end{tabminted}
& Let's see the values
& Return the first 5 rows
& The table shows \_ \_ \_ \_ \_
& Return the first 5 rows. (defValue=5) (C)\\
\midrule
\begin{tabminted}{python}
all_data = pd.concat((
train.loc[:,'SubClass':'SaleCond'],
test.loc[:,'SubClass':'SaleCond']))
\end{tabminted}
& A generator for feature
& Concatenate pandas objects along a particular axis with optional set logic along the other axes.
& This code cell is for \_ \_ \_ \_ \_
& Concat train and test col "SaleCondition" (C)\\
\midrule
\begin{tabminted}{python}
all_data = pd.get_dummies(all_data)
\end{tabminted}
& Convert all the data
& Convert categorical variable into dummy/indicator variables
& This code cell is for \_ \_ \_ \_ \_
& Convert categorical variable into dummy/indicator variables. (T)\\
\midrule
\begin{tabminted}{python}
all_data = all_data
.fillna(all_data.mean())
\end{tabminted}
& Check the missing values
& Fill NA/NaN values using the specified method
& This code cell is for \_ \_ \_ \_ \_
& \\
\midrule
\begin{tabminted}{python}
X_train = all_data[:train.shape[0]]
X_test = all_data[train.shape[0]:]
y = train.SalePrice
\end{tabminted}
& Create the target and the test data
& Slice string
& This code cell is for \_ \_ \_ \_ \_
& \\
\midrule
\begin{tabminted}{python}
model_lasso = LassoCV(
alphas = [1, 0.1, 0.001, 0.0005])
.fit(X_train, y)
\end{tabminted}
& Model
& Lasso linear model with iterative fitting along a regularization path.
& This code cell is for \_ \_ \_ \_ \_
& Fit regression model (H)\\
\midrule
\begin{tabminted}{python}
def rmse_cv(model):
rmse= np.sqrt(-cross_val_score(
model, X_train, y, scoring=
"neg_mean_squared_error", cv = 5))
return(rmse)
rmse_cv(model_lasso).mean()
\end{tabminted}
& A simple example model with the lasso
& Evaluate a score by cross-validation
& The result indicates that \_ \_ \_ \_ \_
& Define score function and evaluate (H)\\
\bottomrule
\end{tabular}
\end{table}
\end{landscape}
|
\section{Introduction}
The beginning of the new millennium marked an important point in the history of geometric analysis.
In a series of three very condensed papers \cite{Perelman1,Perelman2,Perelman3}, Perelman presented a solution of the Poincar\'e and Geometrization Conjectures.
These conjectures --- the former of which was 100 years old and a Millennium Problem --- were fundamental problems concerning the topology of 3-manifolds, yet their solution employed Ricci flow, a technique introduced by Hamilton in 1982, that combines methods of PDE and Riemannian geometry (see below for more details).
These spectacular applications were far from coincidental as they provided a new perspective on 3-manifold topology using the geometric-analytic language of Ricci flows.
It took several years for the scientific community to digest Perelman's arguments and verify his proof.
This was an exciting time for geometric analysts, since the new techniques generated a large number of interesting questions and potential applications.
Amongst these, perhaps the most intriguing goal was to find further applications of Ricci flow to problems in topology.
This goal will form the central theme of this article.
I will first summarize the important results pertaining to Perelman's work.
Next, I will discuss more recent research that stands in the same spirit; first in dimension 3, leading, amongst others, to the resolution of the Generalized Smale Conjecture and, second, work in higher dimensions, where topological applications may be forthcoming.
There are also many other interesting recent applications of Ricci flows, which are more of geometric or analytic nature, and which I won't have space to cover here, unfortunately.
Among these are, for example, work on manifolds with positive isotropic curvature \cite{Brendle_PIC_2018}, smoothing constructions for limit spaces of lower curvature bounds \cite{Simon_Topping_local_moll, Bamler_CR_Wilking_2019,Lai_local_nn_2019}, lower scalar curvature bounds for $C^0$-metrics \cite{Bamler_RF_proof_Gromov, Burkhardt_2019}, as well as work on K\"ahler-Ricci flow.
\section{Riemannian geometry and Ricci flow}
A Riemannian metric $g$ on a smooth manifold $M$ is given by a family of inner products on its tangent spaces $T_p M$, $p \in M$, that can be expressed in local coordinates $(x^1, \ldots, x^n)$ as
\[ g = \sum_{i,j} g_{ij} \, dx^i \, dx^j, \]
where $g_{ij} = g_{ji}$ are smooth coefficient functions.
A metric allows us to define notions such as angles, lengths and distances on $M$.
One can show that the metric $g$ near any point $p \in M$ looks Euclidean up to order 2, that is we can find suitable local coordinates around $p$ such that
\[ g_{ij} = \delta_{ij} + O(r^2). \]
The second order terms can be described by an invariant called the \emph{Riemann curvature tensor} $\Rm_{iklj}$, which is independent of the choice of coordinates up to a simple transformation rule.
Various components of this tensor have local geometric interpretations.
A particularly interesting component is called the \emph{Ricci tensor} $\Ric_{ij}$, which arises from tracing the $k,l$ indices of $\Rm_{iklj}$.
This tensor roughly describes the second variation of areas of hypersurfaces under normal deformations.
A Ricci flow on a manifold $M$ is given by a smooth family $g(t)$, $t \in [0,T)$, of Riemannian metrics satisfying the evolution equation
\begin{equation} \label{eq_RF}
\partial_t g(t) = - 2 \Ric_{g(t)}.
\end{equation}
Expressed in suitable local coordinates, this equation roughly takes the form of a non-linear heat equation in the coefficient functions:
\begin{equation} \label{eq_dtgij_RF}
\partial_t g_{ij} = \triangle g_{ij} + \ldots
\end{equation}
In addition, if we compute the time derivative of the Riemannian curvature tensor $\Rm_{g(t)}$, we obtain
\begin{equation} \label{eq_dtRm_RF}
\partial_t \Rm_{g(t)} = \triangle \Rm_{g(t)} + Q(\Rm_{g(t)}),
\end{equation}
where the last term denotes a quadratic term; its exact form will not be important in the sequel.
Equations (\ref{eq_dtgij_RF}), (\ref{eq_dtRm_RF}) suggest that the metric $g(t)$ becomes smoother or more homogeneous as time moves on, similar to the evolution of a heat equation, under which heat is distributed more evenly across its domain.
On the other hand, the last term in (\ref{eq_dtRm_RF}) seems to indicate that --- possibly at larger scales or in regions of large curvature --- this diffusion property may be outweighed by some other non-linear effects, which could lead to singularities.
But we are getting ahead of ourselves.
Let us first state the following existence and uniqueness theorem, which was established by Hamilton \cite{Hamilton_3_manifolds} in the same paper in which he introduced the Ricci flow equation:
\begin{Theorem}
Suppose that $M$ is compact and let $g_0$ be an arbitrary Riemannian metric on $M$ (called the initial condition). Then:
\begin{enumerate}
\item The evolution equation (\ref{eq_RF}) combined with the initial condition $g(0) = g_0$ has a unique solution $(g(t))_{t \in [0,T)}$ for some maximal $T \in (0, \infty]$.
\item If $T < \infty$, then we say that the flow \emph{develops a singularity at time $T$} and the curvature blows up:
$\max_M |{\Rm}_{g(t)}| \xrightarrow[t \nearrow T]{} \infty $.
\end{enumerate}
\end{Theorem}
The most basic examples of Ricci flows are those in which $g_0$ is Einstein, i.e. $\Ric_{g_0} = \lambda g_0$.
In this case the flow simply evolves by rescaling:
\begin{equation} \label{eq_homotheticsol}
g(t) = (1-2\lambda t) g_0
\end{equation}
So for example, a round sphere ($\lambda > 0$) shrinks under the flow and develops a singularity in finite time, where the diameter goes to 0.
On the other hand, if we start with a hyperbolic metric ($\lambda < 0$), then the flow is immortal, meaning that it exists for all times, and the metric expands linearly.
In the following we will consider more general initial metrics $g_0$ and hope that --- at least in some cases --- the flow is asymptotic to a solution of the form (\ref{eq_homotheticsol}).
This will then allow us to understand the topology of the underlying manifold in terms of the limiting geometry.
\section{Dimension 2}
In dimension 2, Ricci flows are very well behaved:
\begin{Theorem} \label{Thm_RF_2d}
Any Ricci flow on a compact 2-dimensional manifold converges, modulo rescaling, to a metric of constant curvature.
\end{Theorem}
In addition, one can show that the flow in dimension 2 preserves the conformal class, i.e. for all times $t$ we have $g(t) = f(t) g_0$ for some smooth positive function $f(t)$ on $M$.
This observation, combined with Theorem~\ref{Thm_RF_2d} can in fact be used to prove of the Uniformization Theorem:\footnote{The proof of Theorem~\ref{Thm_RF_2d}, which was established by Chow and Hamilton \cite{Hamilton_RF_surfaces, Chow_RF_2sphere}, relied on the Uniformization Theorem.
This dependence was later removed by Chen, Lu, Tian \cite{Chen_Lu_Tian_2006}.
}
\begin{Theorem}
Each compact surface $M$ admits a metric of constant curvature in each conformal class.
\end{Theorem}
This is our first topological consequence of Ricci flow.
Of course, the Uniformization Theorem has already been known for about 100 years.
So in order to obtain any \emph{new} topological results, we will need to study the flow in higher dimensions.
\begin{figure*}[!ht]
\centering
\includegraphics[width=118mm]{rot-sym-step-8}
\caption{Different singularity formations in the rotationally symmetric case, depending on the choice of the radii $r_1, r_2, r_3$.
The flows depicted on the top are the corresponding singularity models.
These turn out to be the only singularity models, even in the non-rotationally symmetric case (see Subsection~\ref{subsec_sing_models}).
\label{fig_rot_sym}}
\end{figure*}
\section{Dimension 3}
In dimension 3, the behavior of the flow --- and its singularity formation --- becomes far more complicated.
In this section, I will first discuss an example and briefly review Perelman's analysis of singularity formation and the construction of Ricci flows with surgery.
I will keep this part short and only focus on aspects that will become important later; for a more in-depth discussion and a more complete list of references see the earlier Notices article \cite{Anderson_Notices_04}.
Next, I will focus on more recent work by Kleiner, Lott and the author on singular Ricci flows and their uniqueness and continuous dependence, which led to the resolution of several longstanding topological conjectures.
\subsection{Singularity formation --- an example} \label{subsec_dumbbell} \label{subsec_dumbell}
To get an idea of possible singularity formation of 3-dimensional Ricci flows, it is useful to consider the famous dumbbell example (see Figure~\ref{fig_rot_sym}).
In this example, the initial manifold $(M, g_0)$ is the result of connecting two round spheres of radii $r_1, r_3$ by a certain type of rotationally symmetric neck of radius $r_2$ (see Figure~\ref{fig_rot_sym}).
So $M \approx S^3$ and $g_0 = f^2(s) g_{S^2} + ds^2$ is a warped product away from the two endpoints.
It can be shown that any flow starting from a metric of this form must develop a singularity in finite time.
The singularity \emph{type,} however, depends on the choice of the radii $r_1, r_2, r_3$.
More specifically, if the radii $r_1, r_2, r_3$ are comparable (Figure~\ref{fig_rot_sym}, left), then the diameter of the manifold converges to zero and, after rescaling, the flow
becomes asymptotically round --- just as in Theorem~\ref{Thm_RF_2d}.
This case is called \emph{extinction}.
On the other hand, if $r_2 \ll r_1, r_3$ (Figure~\ref{fig_rot_sym}, right), then the flow develops a \emph{neck singularity}, which looks like a round cylinder ($S^2 \times \mathbb{R}$) at small scales.
Note that in this case the singularity only occurs in a certain region of the manifold, while the metric converges to a smooth metric everywhere else.
Lastly, there is also an intermediate case (Figure~\ref{fig_rot_sym}, center),
in which the flow develops a singularity that is modeled on the Bryant soliton --- a one-ended paraboloid-like model.
\begin{figure}
\centering
\includegraphics[width=60mm]{convergent-parabolic-nbhds}
\caption{A Ricci flow $M \times [0,T)$ that develops a singularity at time $T$ and a sequence of points $(x_i,t_i)$ that ``run into a singularity''.
The geometry in the parabolic neighborhoods around $(x_i,t_i)$ (rectangles) is close to the singularity model modulo rescaling if $i \gg 1$.
\label{fig_blowup}}
\end{figure}
\subsection{Blow-up analysis} \label{subsec_blowup}
Perelman's work implied that the previous example is in fact prototypical for the singularity formation of
3-dimensional Ricci flows starting from \emph{general,} not necessarily rotationally symmetric, initial data.
In order to make this statement more precise, let us first recall a method called \emph{blow-up analysis,} which is used frequently to study singularities in geometric analysis.
Suppose that $(M, (g(t))_{t \in [0,T)})$ is a Ricci flow that develops a singularity at time $T < \infty$ (see Figure~\ref{fig_blowup}).
Then we can find a sequence of spacetime points $(x_i, t_i) \in M \times [0,T)$ such that $\lambda_i := |{\Rm}|(x_i, t_i) \to \infty$ and $t_i \nearrow T$; we will say that the points $(x_i, t_i)$ ``run into a singularity''.
Our goal will be to understand the local geometry at small scales near $(x_i, t_i)$, for large $i$.
For this purpose, we
consider the sequence of pointed, parabolically rescaled flows
\[ \big(M, (g'_i(t) := \lambda_i g(\lambda_i^{-1} t + t_i))_{t \in [-\lambda_i t_i, 0]}, (x_i,0) \big). \]
Geometrically, the flow $(g'_i(t))$ is the result of rescaling distances by $\lambda_i^{1/2}$, times by $\lambda_i$ and an application of a time-shift such that the point $(x_i,0)$ in the new flow corresponds the point $(x_i, t_i)$ in the old flow.
The new flow $(g'_i(t))$ still satisfies the Ricci flow equation and it is defined on larger and larger backwards time-intervals of size $\lambda_i t_i \to \infty$.
Moreover, we have $|{\Rm}|(x_i,0)=1$ on this new flow.
Observe also that the geometry of the original flow near $(x_i,t_i)$ at scale $\lambda_i^{-1/2} \ll 1$ is a rescaling of the geometry of $(g'_i(t))$ near $(x_i,0)$ at scale $1$.
The hope is now to apply a compactness theorem (\`a la Arzela-Ascoli) such that, after passing to a subsequence, we have convergence
\begin{multline} \label{eq_RF_convergence}
\big( M, (g'_i(t))_{t \in [-\lambda_i^{2} t_i, 0]}, (x_i,0) \big) \\ \xrightarrow[i \to \infty]{} \big( M_\infty, (g_\infty(t))_{t \leq 0}, (x_\infty, 0) \big).
\end{multline}
The limit is called a \emph{blow-up} or \emph{singularity model}, as it gives valuable information on the singularity near the points $(x_i, t_i)$.
This model is a Ricci flow that is defined for all times $t \leq 0$; it is therefore called \emph{ancient.}
So in summary, a blow-up analysis reduces the study of singularity \emph{formation} to the classification of ancient singularity \emph{models.}
The notion of convergence in (\ref{eq_RF_convergence}) is a generalization of Cheeger-Gromov convergence to Ricci flows, due to Hamilton.
Instead of demanding global convergence of the metric tensors, as in Theorem~\ref{Thm_RF_2d}, we only require convergence up to diffeomorphisms here.
We roughly require that we have convergence
\begin{equation} \label{eq_Hamilton_phi_Cinfty}
\phi_i^* g'_i (t) \xrightarrow[i \to \infty]{C^\infty_{\loc}} g_\infty
\end{equation}
on $M_\infty \times (-\infty, 0]$ of the pull backs of $g'_i (t)$ via (time-independent) diffeomorphisms $\phi_i : U_i \to V_i \subset M$ that are defined over larger and larger subsets $U_i \subset M_\infty$ and satisfy $\phi_i(x_\infty) = x_i$.
We will see later (in Section~\ref{sec_dim_n_geq_4}) that this notion of convergence is too strong to capture the more subtle singularity formation of higher dimensional Ricci flows and we will discuss necessary refinements.
Luckily, in dimension 3 the current notion is still sufficient for our purposes, though.
\subsection{Singularity models and canonical neighborhoods} \label{subsec_sing_models}
Arguably one of the most groundbreaking discoveries of Perelman's work was the classification of singularity models of 3-dimensional Ricci flows and the resulting structural description of the flow near a singularity.
The following theorem\footnote{Perelman proved a version of Theorem~\ref{Thm_classificaton_3d} that contained a more qualitative characterization in Case~3., which was sufficient for most applications.
Later, Brendle \cite{Brendle_Bryant_2020} showed that the only possibility in Case~3. is the Bryant soliton.} summarizes this classification
\begin{Theorem} \label{Thm_classificaton_3d}
Any singularity model $(M_\infty, (g_\infty(t))_{t \leq 0})$ obtained as in (\ref{eq_RF_convergence}) is isometric, modulo rescaling, to one the following:
\begin{enumerate}
\item A quotient of the round shrinking sphere $(S^3, (1-4t)g_{S^3})$.
\item The round shrinking cylinder $(S^2 \times \mathbb{R}, (1 - 2t)g_{S^2} + g_{\mathbb{R}})$ or its quotient $(S^2 \times \mathbb{R})/\mathbb{Z}_2$.
\item The Bryant soliton $(M_{\Bry}, (g_{\Bry}(t)))$
\end{enumerate}
\end{Theorem}
Note that these three models correspond to the three cases in the rotationally symmetric dumbbell example from Subsection~\ref{subsec_dumbbell} (see Figure~\ref{fig_rot_sym}).
The Bryant soliton in 3. is a rotationally symmetric solution to the Ricci flow on $\mathbb{R}^3$ with the property that all its time-slices are isometric to a metric of the form
\[ g_{\Bry} = f^2(r) g_{S^2} + dr^2, \qquad f(r) \sim \sqrt{r}. \]
The name \emph{soliton} refers to the fact that all time-slices of the flow are isometric, so the flow merely evolves by pullbacks of a family of diffeomorphisms (we will encounter this soliton again at the end of this article).
It turns out that we can do even better than Theorem~\ref{Thm_classificaton_3d}: we can describe the structure of the flow near \emph{any} point of the flow that is close to a singularity --- not just near the points used to construct the blow-ups.
In order to state the next result efficiently, we will need to consider the class of \emph{$\kappa$-solutions.}
This class simply consists of all solutions listed in Theorem~\ref{Thm_classificaton_3d}, plus an additional compact, ellipsoidal solution (the details of this solution won't be important here\footnote{This solution does not occur as a singularity of a single flow, but can be observed as some kind of transitional model in families of flows that interpolate between two different singularity models.
}).
Then we have:
\begin{Theorem}[Canonical Neighborhood Theorem] \label{Thm_CNT}
If $(M, (g(t))_{t \in [0,T)})$, $T < \infty$, is a 3-dimensional Ricci flow and $\varepsilon >0$, then there is a constant $r_{\can} (g(0), T, \varepsilon) > 0$ such that for any $(x,t) \in M \times [0,T)$ with the property that
\[ r := |{\Rm}|^{-1/2}(x,t) \leq r_{\can} \]
the geometry of the metric $g(t)$ restricted to the ball $B_{g(t)}(x, \varepsilon^{-1} r)$ is $\varepsilon$-close\footnote{Similar to (\ref{eq_Hamilton_phi_Cinfty}), this roughly means that there is a diffeomorphism between an $\varepsilon^{-1}$-ball in a $\kappa$-solution and this ball such that the pullback of $r^{-2} g(t)$ is $\varepsilon$-close in the $C
^{[\varepsilon^{-1}]}$-sense to the metric on the $\kappa$-solution.} to a time-slice of a $\kappa$-solution.
\end{Theorem}
Let us digest the content of this theorem.
Recall that the norm of the curvature tensor $|{\Rm}|$, which can be viewed as a measure of ``how singular the flow is near a point'', has the dimension of $\length^{-2}$, so $r = |{\Rm}|^{-1/2}$ can be viewed as a ``curvature scale''.
So, in other words, Theorem~\ref{Thm_CNT} states that regions of high curvature locally look very much like cylinders, Bryant solitons, round spheres etc.
\begin{figure}
\centering
\includegraphics[width=60mm]{RF-w-surg}
\caption{A schematic depiction of a Ricci flow with surgery. The almost-singular parts $M_{\text{almost-sing}}$, i.e. the parts that are discarded under each surgery construction, are hatched.\label{fig_RF_surgery}}
\end{figure}
\subsection{Ricci flow with surgery} \label{subsec_RFsurg}
Our understanding of the structure of the flow near a singularity now allows us to carry out a so-called \emph{surgery construction.}
Under this construction (almost) singularities of the flow are removed, resulting in a ``less singular'' geometry, from which the flow can be restarted.
This will lead to a new type of flow that is defined beyond its singularities and which will provide important information on the underlying manifold.
Let me be more precise.
A \emph{(3-dimensional) Ricci flow with surgery} (see Figure~\ref{fig_RF_surgery}) consists of a sequence of Ricci flows
\begin{align*}
&(M_1, (g_1(t))_{t \in [0,T_1]}), \quad (M_2, (g_2(t))_{t \in [T_1,T_2]}), \quad \\
&\qquad (M_3, (g_3(t))_{t \in [T_2,T_3]}), \quad \ldots,
\end{align*}
which live on manifolds $M_1, M_2, \ldots$ of possibly different topology and are parameterized by consecutive time-intervals of the form $[0,T_1], [T_1, T_2], \ldots$ whose union equals $[0, \infty)$.
The time-slices $(M_i, g_i(T_i))$ and $(M_{i+1}, g_{i+1}(T_i))$ are related by a \emph{surgery process,} which can be roughly summarized as follows.
Consider the set $M_{\almostsing} \subset M_i$ of all points of high enough curvature, such that they have a canonical neighborhood as in Theorem~\ref{Thm_CNT}.
Cut $M_i$ open along approximate cross-sectional 2-spheres of diameter $r_{\surg}(T_i) \ll 1$ near the cylindrical ends of $M_{\almostsing}$, discard most of the high-curvature components (including the closed, spherical components of $M_{\almostsing}$), and glue in cap-shaped 3-disks to the cutting surfaces.
In doing so we have constructed a new, ``less singular'', Riemannian manifold $(M_{i+1}, g_{i+1}(T_i))$, from which we can restart the flow.
Stop at some time $T_{i+1} > T_i$, shortly before another singularity occurs and repeat the process.
The precise surgery construction is quite technical and more delicate than presented here.
The main difficulty in this construction is to ensure that the surgery times $T_i$ do not accumulate, i.e. that the flow can be extended for all times.
It was shown by Perelman that this and other difficulties can indeed be overcome:
\begin{Theorem}
Let $(M,g)$ be a closed, 3-dimensional Riemannian manifold.
If the surgery scales $r_{\surg}(T_i) > 0$ are chosen sufficiently small (depending on $(M,g)$ and $T_i$), then a Ricci flow with surgery with initial condition $(M_1, g_1(0)) = (M,g)$ can be constructed.
\end{Theorem}
Note that the topology of the underlying manifold $M_i$ may change in the course of a surgery, but only in a controlled way.
In particular, it is possible to show that for any $i$ the initial manifold $M_1$ is diffeomorphic to a connected sum of components of $M_i$ and copies of spherical space forms $S^3/\Gamma$ and $S^2 \times S^1$.
So if the flow goes \emph{extinct} in finite time, meaning that $M_i = \emptyset$ for some large $i$, then
\begin{equation} \label{eq_sph_prime_dec}
M_1 \approx \#_{j=1}^k (S^3/\Gamma_j) \# m (S^2 \times S^1).
\end{equation}
Perelman moreover showed that if $M_1$ is simply connected, then the flow \emph{has} to go extinct and therefore $M_1$ must be of the form (\ref{eq_sph_prime_dec}).
This immediately implies the Poincar\'e Conjecture --- our first true topological application of Ricci flow:
\begin{Theorem}[Poincar\'e Conjecture]
Any simply connected, closed 3-manifold is diffeomorphic to $S^3$.
\end{Theorem}
On the other hand, Perelman showed that if the Ricci flow with surgery does not go extinct, meaning if it exists for all times, then for large times $t \gg 1$ the flow decomposes the manifold (at time $t$) into a thick and a thin part:
\begin{equation} \label{eq_thick_thin_3d}
M_{\thick}(t) \,\, \ensuremath{\mathaccent\cdot\cup} \,\, M_{\thin}(t),
\end{equation}
such that the metric on $M_{\thick}(t)$ is asymptotic to a hyperbolic metric and the metric on $M_{\thin}(t)$ is collapsed along fibers\footnote{The precise description of these fibrations is a bit more complex and omitted here. In addition, $M_{\thick}(t)$ and $M_{\thin}(t)$ are separated by embedded 2-tori that are incompressible, meaning that the inclusion into the ambient manifold is an injection on $\pi_1$.
} of type $S^1, S^2$ or $T^2$.
A further topological analysis of these induced fibrations implied the Geometrization Conjecture --- our second topological application of Ricci flow:
\begin{Theorem}[Geometrization Conjecture]
Every closed 3-manifold is a connected sum of manifolds that can be cut along embedded, incompressible copies of $T^2$ into pieces which each admit a locally homogeneous geometry.
\end{Theorem}
\begin{figure}
\hspace{-3mm}
\includegraphics[width=70mm]{sing-RF-2}
\caption{Illustration of a singular Ricci flow given by a Ricci flow spacetime.
The arrows indicate the time-vector field $\partial_\mathfrak{t}$.\label{fig_sing_RF}}
\end{figure}
\subsection{Ricci flows through singularities} \label{subsec_RF_through_sing}
Despite their spectacular applications, Ricci flows with surgery have one major drawback: their construction is not canonical.
In other words, each surgery step depends on a number of auxiliary parameters, for which there does not seem to be a canonical choice, such as:
\begin{itemize}
\item The surgery scales $r_{\surg}(T_i)$, i.e. the diameters of the cross-sectional spheres along which the manifold is cut open.
These scales need to be positive and small.
\item The precise locations of these surgery spheres.
\end{itemize}
Different choices of these parameters may influence the future development of the flow significantly (as well as the space of future surgery parameters).
Hence a Ricci flow with surgery is not \emph{uniquely} determined by its initial metric.
This disadvantage was already recognized in Perelman's work, where he conjectured that there should be another flow, in which surgeries are effectively carried out automatically at an infinitesimal scale (think ``$r_{\surg} = 0$''), or which, in other words, ``flows through singularities''.
Perelman's conjecture was recently resolved by Kleiner, Lott and the author \cite{Kleiner_Lott_singular, bamler_kleiner_uniqueness_stability}:\footnote{The ``Existence'' part is due to Kleiner and Lott and the ``Uniqueness'' part is due to Kleiner and the author. Part 2. of Theorem~\ref{Thm_sing_RF} follows from a combination of both papers.}
\begin{Theorem} \label{Thm_sing_RF}
There is a notion of singular Ricci flow (through singularities) such that:
\begin{enumerate}
\item For any compact, 3-dimensional Riemannian manifold $(M,g)$ there is a unique singular Ricci flow $\mathcal{M}$ whose initial time-slice $(\mathcal{M}_0, g_0)$ is $(M,g)$.
\item Any Ricci flow with surgery starting from $(M,g)$ can be viewed as an approximation of $\mathcal{M}$.
More specifically, if we consider a sequence of Ricci flows with surgery starting from $(M,g)$ with surgery scales $\max_t r_{\surg}(t) \to 0$, then these flows converge to $\mathcal{M}$ in a certain sense.
\end{enumerate}
\end{Theorem}
In addition, the concept of a singular Ricci flow is far less technical than that of a Ricci flow with surgery ---
in fact, I will be able to state its full definition here.
To do this, I will first define the concept of a Ricci flow spacetime.
In short, this is a smooth 4-manifold that locally looks like a Ricci flow, but which may have non-trivial \emph{global} topology (see Figure~\ref{fig_sing_RF}).
\begin{Definition} \label{Def_RF_spacetime}
A \emph{Ricci flow spacetime} consists of:
\begin{enumerate}
\item A smooth 4-dimensional manifold $\mathcal{M}$ with boundary, called \emph{spacetime.}
\item A \emph{time-function} $\mathfrak{t} : \mathcal{M} \to [0, \infty)$.
Its level sets $\mathcal{M}_t := \mathfrak{t}^{-1}(t)$ are called \emph{time-slices} and we require that $\mathcal{M}_0 = \partial \mathcal{M}$.
\item A \emph{time-vector field} $\partial_{\mathfrak{t}}$ on $\mathcal{M}$ with $\partial_{\mathfrak{t}} \cdot \mathfrak{t} \equiv 1$.
Trajectories of $\partial_{\mathfrak{t}}$ are called \emph{worldlines.}
\item A family $g$ of inner products on $\ker d\mathfrak{t} \subset T \mathcal{M}$, which induce a Riemannian metric $g_t$ on each time-slice $\mathcal{M}_t$.
We require that the Ricci flow equation holds:
\[ \mathcal{L}_{\partial_{\mathfrak{t}}} g_t = - 2\Ric_{g_t}. \]
\end{enumerate}
By abuse of notation, we will often write $\mathcal{M}$ instead of $(\mathcal{M}, \mathfrak{t}, \partial_{\mathfrak{t}}, g)$.
\end{Definition}
A classical, 3-dimensional Ricci flow $(M, (g(t))_{t \in [0,T)})$ can be converted into a Ricci flow spacetime by setting $\mathcal{M} := M \times [0,T)$, letting $\mathfrak{t}, \partial_{\mathfrak{t}}$ be the projection onto the second factor and the pullback of the unit vector field on the second factor, respectively, and letting $g_t$ be the metric corresponding to $g(t)$ on $M \times \{ t \} \approx M$.
Hence worldlines correspond to curves of the form $t \mapsto (x,t)$.
Likewise, a Ricci flow with surgery, given by flows $(M_1, \linebreak[1] (g_1(t))_{t \in [0,T_1]}), \linebreak[0] (M_2,\linebreak[1] (g_2(t))_{t \in [T_1,T_2]}), \ldots$ can be converted into a Ricci flow spacetime as follows.
Consider first the Ricci flow spacetimes $M_1 \times [0,T_1], M_2 \times [T_1,T_2], \ldots$ arising from each single flow.
We can now glue these flows together by identifying the set of points $U_i^- \subset M_i \times \{ T_i \}$ and $U_i^+ \subset M_{i+1} \times \{ T_i \}$ that survive each surgery step via maps $\phi_i : U_i^- \to U_i^+$.
The resulting space has a boundary that consists of the time-$0$-slice $M_1 \times \{ 0 \}$ and the points
\[ \mathcal{S}_i = (M_i \times \{ T_i \} \setminus U_i^- ) \cup ( M_{i+1} \times \{ T_{i} \} \setminus U_i^+ ), \]
which were removed and added during each surgery step.
After removing these points, we obtain a Ricci flow spacetime of the form:
\begin{multline} \label{eq_MM_from_RF_surg}
\mathcal{M} = (M_1 \times [0,T_1] \cup_{\phi_1} M_2 \times [T_1, T_2] \cup_{\phi_2} \ldots) \\ \setminus (\mathcal{S}_1 \cup \mathcal{S}_2 \cup \ldots).
\end{multline}
Note that for any regular time $t \in (T_{i-1}, T_i)$ the time-slice $\mathcal{M}_t$ is isometric to $(M_{i}, g_i(t))$.
On the other hand, the time-slices $\mathcal{M}_{T_i}$ corresponding to surgery times are incomplete; they have cylindrical open ends of scale $\approx r_{\surg}(T_i)$.
The following definition captures this incompleteness:
\begin{Definition}
A Ricci flow spacetime is \emph{$r$-complete,} for some $r \geq 0$, if the following holds.
Consider a smooth path $\gamma : [0,s_0) \to \mathcal{M}$ with the property that
\[ \inf_{s \in [0,s_0)} |{\Rm}|^{-1/2} (\gamma(s)) > r \]
and:
\begin{enumerate}
\item $\gamma([0,l)) \subset \mathcal{M}_t$ is contained in a single time-slice and its length measured with respect to the metric $g_t$ is finite, or
\item $\gamma$ is a worldline, i.e. a trajectory of $\pm\partial_{\mathfrak{t}}$.
\end{enumerate}
Then the limit $\lim_{s \nearrow s_0} \gamma(s)$ exists.
\end{Definition}
So $\mathcal{M}$ being $r$-complete roughly means that it has only ``holes'' of scale $\lesssim r$.
For example, the flow from (\ref{eq_MM_from_RF_surg}) is $C\max_t r_{surg}(t)$-complete for some universal $C < \infty$.
In addition, Theorem~\ref{Thm_CNT} motivates the following definition:
\begin{Definition}
A Ricci flow spacetime is said to satisfy the \emph{$\epsilon$-canonical neighborhood assumption at scales $(r_1, r_2)$} if for any point $x \in \mathcal{M}_t$ with $r := |{\Rm}|^{-1/2}(x) \in (r_1, r_2)$ the metric $g_t$ restricted to the ball $B_{g_t} (x, \varepsilon^{-1} r)$ is $\varepsilon$-close, after rescaling by $r^{-2}$, to a time-slice of a $\kappa$-solution.
\end{Definition}
We can finally define singular Ricci flows (through singularities), as used in Theorem~\ref{Thm_sing_RF}:
\begin{Definition}
A \emph{singular Ricci flow} is a Ricci flow spacetime $\mathcal{M}$ with the following two properties:
\begin{enumerate}
\item It is $0$-complete.
\item For any $\varepsilon > 0$ and $T < \infty$ there is an $r(\varepsilon, T) > 0$ such that the flow $\mathcal{M}$ restricted to $[0,T)$ satisfies the $\varepsilon$-canonical neighborhood assumption at scales $(0, r)$.
\end{enumerate}
\end{Definition}
See again Figure~\ref{fig_sing_RF} for a depiction of a singular Ricci flow.
The time-slices $\mathcal{M}_t$ for $t < T_{\sing}$ develop a cylindrical region, which collapses to some sort of topological double cone singularity in the time-$T_{\sing}$-slice $\mathcal{M}_{T_{\sing}}$.
This singularity is immediately resolved and the flow is smooth for all $t > T_{\sing}$.
Let us digest the definition of a singular Ricci flow a bit more.
It is tempting to think of the time function $\mathfrak{t}$ as a Morse function and compare critical points with infinitesimal surgeries.
However, this comparison is flawed:
First, by definition $\mathfrak{t}$ cannot have critical points since $\partial_\mathfrak{t} \, \mathfrak{t} = 1$.
In fact, a singular Ricci flow is a completely smooth object.
The ``singular points'' of the flow are not part of $\mathcal{M}$, but can be obtained after metrically completing each time-slice by adding a discrete set of points.
Second, it is currently unknown whether the set of singular times, i.e. the set of times whose time-slices are incomplete, is discrete.
Similar notions of singular flows have been developed for the mean curvature flow (a close cousin of the Ricci flow).
These are called level set flows and Brakke flows.
However, their definitions differ from singular Ricci flows in that they characterize the flow equation at singular points via barrier and weak integral conditions, respectively.
This is possible, in part, because a mean curvature flow is an embedded object and its singular set has an analytic meaning.
By contrast, the definition of a singular Ricci flow only characterizes the flow on its regular part.
In lieu of a weak formulation of the Ricci flow equation on the singular set, we have to impose the canonical neighborhood assumption, which serves as an asymptotic characterization near the incomplete ends.
Finally, I will briefly explain how singular Ricci flows are constructed and convey the meaning of Part 2. of Theorem~\ref{Thm_sing_RF}.
Fix an initial time-slice $(M,g)$ and consider a sequence of Ricci flow spacetimes $\mathcal{M}^j$ that correspond to Ricci flows with surgery starting from $(M,g)$, with surgery scale $\max_t r_{\surg,j}(t) \to 0$.
It can be shown that these flows are $C \max_t r_{\surg,j}(t)$-complete and satisfy the $\varepsilon$-canonical neighborhood assumption at scales $(C_\varepsilon \max_t r_{\surg,j}(t), r_\varepsilon)$, where $C, C_\varepsilon, r_{\varepsilon}$ do not depend on $j$.
A compactness theorem implies that a subsequence of the spacetimes $\mathcal{M}^j$ converges to a spacetime $\mathcal{M}$, which is a singular Ricci flow.
This implies the existence of $\mathcal{M}$; the proof of uniqueness uses other techniques, which are outside the scope of this article.
\begin{figure*}
\centering
\includegraphics[width=120mm]{rot-sym-family-2}
\caption{A family of singular Ricci flows starting from a continuous family of initial conditions.\label{fig_rot_sym_family}}
\end{figure*}
\subsection{Continuous dependence}
The proof of the uniqueness property in Theorem~\ref{Thm_sing_RF}, due to Kleiner and the author, implies an important continuity property, which will lead to further topological applications.
To state this property, let $M$ be a compact 3-manifold and for every Riemannian metric $g$ on $M$ let $\mathcal{M}^g$ be the singular Ricci flow with initial condition $(\mathcal{M}^g_0, g) = (M,g)$.
\begin{Theorem} \label{Thm_cont_dep}
The flow $\mathcal{M}^g$ depends continuously on $g$.
\end{Theorem}
Recall that the topology of the flow $\mathcal{M}^g$ may change as we vary $g$.
We therefore have to choose an appropriate sense of continuity in Theorem~\ref{Thm_cont_dep} that allows such a topological change.
This is roughly done via a topology and lamination structure on the disjoint union $\bigsqcup_g \mathcal{M}^g$, transverse to which the variation of the flow can be studied locally.
Instead of diving into these technicalities, let us discuss the example illustrated in Figure~\ref{fig_rot_sym_family}.
In this example $(g_s)_{s \in [0,1]}$ denotes a continuous family of metrics on $S^3$ such that the corresponding flows $\mathcal{M}^s := \mathcal{M}^{g_s}$ interpolate between a round and a cylindrical singularity.
For $s \in [0,\frac12)$ the flow $\mathcal{M}^s$ can be described in terms of a conventional, non-singular Ricci flow $(g^s_t)$ on $M$ and the continuity statement in Theorem~\ref{Thm_cont_dep} is equivalent to continuous dependence of this flow on $s$.
Likewise, the flows $\mathcal{M}^s$ restricted to $[0,T_{\sing})$ can again be described by a continuous family of conventional Ricci flows.
The question is now what happens at the critical parameter $s = \frac12$, where the type of singularity changes.
The uniqueness property guarantees that the flows $\mathcal{M}^s$ for $s \nearrow \frac12$ and $s \searrow \frac12$ must limit to the same flow $\mathcal{M}^{\frac12}$.
The convergence is locally smooth, but the topology of the spacetime manifold $\mathcal{M}^s$ may still change.
\subsection{Topological Applications}
Theorem~\ref{Thm_cont_dep} provides us a tool to deduce the first topological applications of Ricci flow since Perelman's work.
The first example of such an application concerns the space of metrics of positive scalar curvature $\Met_{\PSC}(M) \subset \Met(M)$ on a manifold $M$, which is a subset of the space of all Riemannian metrics on $M$ (both spaces are equipped with the $C^\infty$-topology).
Since the positive scalar curvature condition is preserved by Ricci flow, Theorem~\ref{Thm_cont_dep} roughly implies that --- modulo singularities and the associated topological changes --- Ricci flow is a ``continuous deformation retraction'' of $\Met_{\PSC}(M)$ to the space of round metrics on $M$.
This heuristic was made rigorous by Kleiner and the author \cite{BamlerKleiner-PSC} and implied:
\begin{Theorem} \label{Thm_PSC}
For any closed 3-manifold $M$ the space $\Met_{\PSC}(M)$ is either contractible or empty.
\end{Theorem}
The study of the spaces $\Met_{\PSC}(M)$ was initiated by Hitchin in the 70s and has led to many interesting results --- based on index theory --- which show that these spaces have non-trivial topology when $M$ is high dimensional.
Theorem~\ref{Thm_PSC} provides first examples of manifolds of dimension $\geq 3$ for which the homotopy type of $\Met_{\PSC}(M)$ is completely understood; see also prior work by Marques \cite{Marques_2012}.
A second topological application concerns the diffeomorphism group $\Diff(M)$ of a manifold $M$, i.e. the space of all diffeomorphisms $\phi : M \to M$ (again equipped with the $C^\infty$-topology).
The study of these spaces was initiated by Smale, who showed that $\Diff(S^2)$ is homotopy equivalent to the orthogonal group $O(3)$, i.e. the set of isometries of $S^2$.
More generally, we can fix an arbitrary closed manifold $M$, pick a Riemannian metric $g$ and consider the natural injection of the isometry group
\begin{equation} \label{eq_Isom_2_Diff}
\Isom(M,g) \longrightarrow \Diff(M).
\end{equation}
The following conjecture allows us to understand the homotopy type of $\Diff(M)$ for many important 3-manifolds.
\begin{Conjecture}[Generalized Smale Conjecture] \label{GSC}
Suppose that $(M^3,g)$ is closed and has constant curvature $K \equiv \pm 1$.
Then (\ref{eq_Isom_2_Diff}) is a homotopy equivalence.
\end{Conjecture}
This conjecture has had a long history and many interesting special cases were established using topological methods, including the case $M = S^3$ by Hatcher and the hyperbolic case by Gabai.
For more background see the first chapter of \cite{Rubinstein_book}.
An equivalent version of Conjecture~\ref{GSC} is that Theorem~\ref{Thm_PSC} remains true if we replace $\Met_{\PSC}(M)$ by the space $\Met_{K \equiv \pm 1}(M)$ of constant curvature metrics.
This was verified by Kleiner and the author \cite{BamlerKleiner2017b, BamlerKleiner-PSC}, which led to:
\begin{Theorem} \label{Thm_GSC}
The Generalized Smale Conjecture is true.
\end{Theorem}
The proof of Theorem~\ref{Thm_GSC} provides a unified treatment of all possible topological cases and it can also be extended to other manifolds $M$.
In addition it is independent of Hatcher's proof, so it gives an alternative proof in the $S^3$-case.
\section{Dimensions $n \geq 4$} \label{sec_dim_n_geq_4}
For a long time, most of the known results of Ricci flows in higher dimensions concerned special cases, such as K\"ahler-Ricci flows or flows that satisfy certain preserved curvature conditions.
\emph{General} flows, on the other hand, were relatively poorly understood.
Recently, however, there has been some movement on this topic --- in part, thanks to a slightly different geometric perspective on Ricci flows \cite{Bamler_HK_RF_partial_regularity,Bamler_HK_entropy_estimates,Bamler_RF_compactness}.
The goal of this section is to convey some of these new ideas and to provide an outlook on possible geometric and topological applications.
\subsection{Gradient shrinking solitons}
Let us start with the following basic question:
What would be reasonable singularity models in higher dimensions?
One important class of such models are gradient shrinking solitons (GSS).
The GSS equation concerns Riemannian manifolds $(M,g)$ equipped with a potential function $f \in C^\infty (M)$ and reads:
\[ \Ric + \nabla^2 f - \frac12 g = 0. \]
This generalization of the Einstein equation is interesting, because it gives rise to an associated selfsimilar Ricci flow:
\[ g(t) := |t| \phi_t^* g, \quad t < 0, \]
where $(\phi_t : M \to M)_{t < 0}$ is the flow of the vector field $|t| \nabla f$.
A basic example of a GSS is an Einstein metric ($\Ric = \frac12 g$), for example a round sphere.
In this case $g(t)$ just evolves by rescaling and becomes singular at time $0$.
A more interesting class of examples are round cylinders $S^{k \geq 2} \times \mathbb{R}^{n-k}$, where
\[ g = 2(k-1) g_{S^k} + g_{\mathbb{R}^{n-k}}, \quad
f = \frac14 \sum_{i=k+1}^n x_i^2. \]
In this case $|t| \nabla f$ generates a family of dilations on the $\mathbb{R}^{n-k}$ factor and
\[ g(t) = 2(k-1) |t| g_{S^k} + g_{\mathbb{R}^{n-k}}, \]
which is isometric to $|t| g$.
In dimensions $n \leq 3$, all (non-trivial\footnote{Euclidean space $\mathbb{R}^n$ equipped with $f = \frac14 r^2$ is called a trivial GSS.}) GSS are quotients of round spheres or cylinders.
However, more complicated GSS exist in dimensions $n \geq 4$.
By construction, GSS (or their associated flows, to be precise) are invariant under parabolic rescaling.
So the blow-up singularity model of the singularity at time $0$ (taken along an appropriately chosen sequence of basepoints) is equal to the flow itself.
Therefore every GSS does indeed occur as a singularity model, at least of its own flow.
Vice versa, the following conjecture, which I will keep vague for now, predicts that the converse should also be true in a certain sense.
\begin{Conjecture} \label{Conj_Folklore}
For any Ricci flow ``most'' singularity models are gradient shrinking solitons.
\end{Conjecture}
This conjecture has been implicit in Hamilton's work from the 90s and a similar result is known to be true for mean curvature flow.
In the remainder of this section I will present a resolution of a version of this conjecture.
\subsection{Examples of singularity formation} \label{subsec_Appleton}
Let us first discuss an example in order to adjust our expectations in regards to Conjecture~\ref{Conj_Folklore}.
In \cite{Appleton:2019ws}, Appleton constructs a class of 4-dimensional Ricci flows\footnote{The flows are defined on non-compact manifolds, but the geometry at infinity is well controlled.}
that develop a singularity in finite time, which can be studied via the blow-up technique from Subsection~\ref{subsec_blowup} --- this time we even allow the rescaling factors to be \emph{any} sequence of numbers $\lambda_i \to \infty$, not just $\lambda_i = |{\Rm}|^{1/2}(x_i,t_i)$.
Appleton obtains the following classification of all non-trivial blow-up singularity models:
\begin{enumerate}
\item The Eguchi-Hanson metric, which is Ricci flat and asymptotic to the flat cone $\mathbb{R}^4/\mathbb{Z}_2$.
\item The flat cone $\mathbb{R}^4 / \mathbb{Z}_2$, which has an isolated orbifold singularity at the origin.
\item The quotient $M_{\Bry} / \mathbb{Z}_2$ of the Bryant soliton, which also has an isolated orbifold singularity at its tip.
\item The cylinder $\mathbb{R} P^3 \times \mathbb{R}$.
\end{enumerate}
Here the models 1., 2. \emph{have} to occur as singularity models, and it is unknown whether the models 3., 4. actually do show up.
The only gradient shrinking solitons in this list are 2., 4.
Note that the flow on $\mathbb{R}^4 / \mathbb{Z}_2$ is constant, but each time-slice is a metric cone, and therefore invariant under rescaling.
So we may also view this model as a (degenerate) gradient shrinking soliton (in this case $f = \frac14 r^2$).
It is conceivable that there are Ricci flow singularities whose only blow-up models are of type 1., 2.
In addition, there are further examples in higher dimensions \cite{Stolarski:2019uk} whose only blow-up models that are gradient shrinking solitons must be singular and possibly degenerate.
This motivates the following revision of Conjecture~\ref{Conj_Folklore}.
\begin{Conjecture} \label{Conj_Folklore_revised}
For any Ricci flow ``most'' singularity models are gradient shrinking solitons that may be degenerate and may have a singular set of codimension $\geq 4$.
\end{Conjecture}
\subsection{A compactness and partial regularity theory}
The previous example taught us that in higher dimensions it becomes necessary to consider \emph{non-smooth} blow-up limits.
The usual convergence and compactness theory of Ricci flows due to Hamilton from Subsection~\ref{subsec_blowup}, which relies on curvature bounds and only produces smooth limits, becomes too restrictive for such purposes.
Instead, we need a fundamentally new \emph{compactness and partial regularity theory} for Ricci flows, which will enable us to take limits of arbitrary Ricci flows and study their structural properties.
This theory was recently found by the author and will lie at the heart of a resolution of Conjecture~\ref{Conj_Folklore_revised}.
Compactness and partial regularity theories are an important feature in many subfields of geometric analysis.
To gain a better sense for such theories, let us first review the compactness and partial regularity theory for Einstein metrics\footnote{This theory is due to Cheeger, Colding, Gromov, Naber and Tian.} --- an important special case, since every Einstein metric corresponds to a Ricci flow.
Consider a sequence $(M_i, g_i, x_i)$ of pointed, complete $n$-dimensional Einstein manifolds, $\Ric_{g_i} = \lambda_i g_i$, $|\lambda_i| \leq 1$.
Then, after passing to a subsequence, these manifolds (or their metric length spaces, to be precise) converge in the Gromov-Hausdorff sense to a pointed metric space
\[ (M_i, d_{g_i}, x_i) \xrightarrow[i \to \infty]{GH} (X, d, x_\infty). \]
If we now impose the following \emph{non-collapsing condition:}
\begin{equation} \label{eq_noncollapsing_Einstein}
\vol B(x,r) \geq v > 0,
\end{equation}
for some uniform $r, v > 0$, then the limit admits is a regular-singular decomposition
\[ X = \mathcal{R} \,\, \ensuremath{\mathaccent\cdot\cup} \,\, \SS \]
such that the following holds:
\begin{enumerate}
\item $\mathcal{R}$ can be equipped with the structure of a Riemannian Einstein manifold $(\mathcal{R}, g_\infty)$ in such a way that the restriction $d|_{\mathcal{R}}$ equals the length metric $d_{g_\infty}$.
In other words, $(X,d)$ is isometric to the metric completion of $(\mathcal{R}, d_{g_\infty})$.
\item We have the following estimate on the Minkowski dimension of the singular set:
\[ \dim_{\mathcal{M}} \SS \leq n-4 \]
\item Every tangent cone (i.e. blow-up of $(X,d)$ pointed at the same point) is in fact a metric cone.
\item We have a filtration of the singular set
\[ \SS^0 \subset \SS^1 \subset \ldots \subset \SS^{n-4} = \SS \]
such that $\dim_{\mathcal{H}} \SS^k \leq k$ and such that every $x \in \SS^k \setminus \SS^{k-1}$ has a tangent cone that splits off an $\mathbb{R}^k$-factor.
\end{enumerate}
Interestingly, compactness and partial regularity theories for other geometric equations (e.g. minimal surfaces, harmonic maps, mean curvature flow, \ldots) take a similar form.
The reason for this is that these theories rely on only a few basic ingredients (e.g. a monotonicity formula, an almost cone rigidity theorem and an $\varepsilon$-regularity theorem), which can be verified in each setting.
A theory for Ricci flows, however, did not exist for a long time, because these basic ingredients are --- at least \emph{a priori} --- not available for Ricci flows.
So this setting required a different approach.
Let us now discuss the new compactness and partial regularity theory for Ricci flows.
Before we begin, note that there is an additional complication:
Parabolic versions of notions like ``metric space'', ``Gromov-Hausdorff convergence'', etc. didn't exist until recently, so they --- and a theory surrounding them --- first had to be developed.
I will discuss these new notions in more detail in Subsection~\ref{subsec_met_flow}.
For now, let us try to get by with some more vague explanations.
The first result is a compactness theorem.
Consider a sequence of pointed, $n$-dimensional Ricci flows
\[ (M_i, (g_i(t))_{t \in (-T_i,0]},(x_i,0)), \]
where we imagine the basepoints $(x_i,0)$ to live in the final time-slices, and suppose that $T_\infty := \lim_{i \to \infty} T_i > 0$.
Then we have:
\begin{Theorem} \label{Thm_compactness_RF}
After passing to a subsequence, these flows $\mathbb{F}$-converge to a pointed metric flow:
\[ (M_i, (g_i(t)), (x_i,0)) \xrightarrow[i \to \infty]{\mathbb{F}} (\mathcal{X}, (\nu_{x_\infty;t})). \]
\end{Theorem}
Here a ``metric flow'' can be thought of as a parabolic version of a ``metric space''.
It is some sort of Ricci flow spacetime (as in Definition~\ref{Def_RF_spacetime}) that is allowed to have singular points; think, for example, of isolated orbifold singularities in every time-slice as in Appleton's example (see Subsection~\ref{subsec_Appleton}), or a singular point where a round shrinking sphere goes extinct.
The term ``$\mathbb{F}$-convergence'' can be thought of as a parabolic version of ``Gromov-Hausdorff convergence''.
The next theorem concerns the partial regularity of the limit
in the non-collapsed case, which we define as the case in which
\begin{equation} \label{eq_NN_lower_bound}
\mathcal{N}_{x_i,0}(r_*^2) \geq -Y_*,
\end{equation}
for some uniform constants $r_* > 0$ and $Y_* < \infty$.
This condition is the parabolic analogue to (\ref{eq_noncollapsing_Einstein}).
The quantity $\mathcal{N}_{x,t}(r^2)$ is the pointed Nash-entropy, which is related to Perelman's $\mathcal{W}$-functional; it was rediscovered by work of Hein and Naber.
Assuming (\ref{eq_NN_lower_bound}), we have:
\begin{Theorem} \label{Thm_part_reg_RF}
There is a regular-singular decomposition
\[ \mathcal{X} = \mathcal{R} \, \ensuremath{\mathaccent\cdot\cup} \, \SS \]
such that:
\begin{enumerate}
\item The flow on $\mathcal{R}$ can be described by a smooth Ricci flow spacetime structure.
Moreover, the entire flow $\mathcal{X}$ is uniquely determined by this structure.
\item We have the following dimensional estimate
\[ \dim_{\mathcal{M}^*} \SS \leq (n+2)-4. \]
\item Tangent flows (i.e. blow-ups based at a fixed point of $\mathcal{X}$) are (possibly singular) gradient shrinking solitons.
\item There is a filtration $\SS^0 \subset \ldots \subset \SS^{n-2}= \SS$ with similar properties as in the Einstein case.
\end{enumerate}
\end{Theorem}
A few comments are in order here.
First, note that the fact that $\mathcal{X}$ is uniquely determined by the smooth Ricci flow spacetime structure on $\mathcal{R}$ is comparable to what we have observed in dimension 3 (see Subsection~\ref{subsec_RF_through_sing}), where we didn't even \emph{consider} the entire flow $\mathcal{X}$.
Second, Property 2. involves a parabolic version of the Minkowski dimension that is natural for Ricci flows; a precise definition would be beyond the scope of this article.
Note that the time direction accounts for 2 dimensions, which is natural.
An interesting case is dimension $n=3$, in which we obtain that the set of singular times has dimension $\leq \frac12$; this is in line with what is known in this dimension.
Lastly, in Property~3. the role of metric cones is now taken by gradient shrinking solitons; these are analogues of metric cones, as both are invariant under rescaling.
The dimensional bounds in Theorem~\ref{Thm_part_reg_RF} are optimal.
In Appleton's example, the singular set $\SS$ may consist of an isolated orbifold point in every time-slice; so its parabolic dimension is $2 = (4+2)-4$.
On the other hand, a flow on $S^2 \times T^2$ develops a singularity at a single time and collapses to the 2-torus $T^2$, which again has parabolic dimension $2$.
\subsection{Applications}
Theorems~\ref{Thm_compactness_RF} and \ref{Thm_part_reg_RF} finally enable us to study the finite-time singularity formation and long-time behavior of Ricci flows in higher dimensions.
Regarding Conjecture~\ref{Conj_Folklore_revised}, we roughly obtain:
\begin{Theorem} \label{Thm_finite_time}
Suppose that $(M, (g(t))_{t \in [0,T)})$ develops a singularity at time $T < \infty$.
Then we can extend this flow by a ``singular time-$T$-slice'' $(M_T, d_T)$ such that the tangent flows at any $(x,T) \in M_T$ are (possibly singular) gradient shrinking solitons.
\end{Theorem}
Regarding the long-time asymptotics, we obtain the following picture, which closely resembles that in dimension~3 compare with (\ref{eq_thick_thin_3d}) in Subsection~\ref{subsec_RFsurg}:
\begin{Theorem} \label{Thm_longtime}
Suppose that $(M, (g(t))_{t \geq 0})$ is immortal.
Then for $t \gg 1$ we have a thick-thin decomposition
\[ M = M_{\thick} (t) \,\, \ensuremath{\mathaccent\cdot\cup} \,\, M_{\thin}(t) \]
such that the flow on $M_{\thick}(t)$ converges, after rescaling, to a singular Einstein metric $(\Ric_{g_\infty} = -g_{\infty}$) and the flow on $M_{\thin}(t)$ is collapsed in the opposite sense of (\ref{eq_NN_lower_bound}).
\end{Theorem}
Theorems~\ref{Thm_finite_time} and \ref{Thm_longtime} essentially generalize Perelman's results to higher dimensions.
\subsection{Metric flows} \label{subsec_met_flow}
So what precisely is a metric flow?
To answer this question, we will imitate the process of passing from a (smooth) Riemannian manifold $(M,g)$ to its metric length space $(M, d_g)$.
Here a new perspective on the geometry of Ricci flows will be key.
So our goal will be to turn a Ricci flow $(M, (g(t))_{t \in I})$ into a synthetic object, which we call ``metric flow''.
To do this, let us first consider the spacetime $\mathcal{X} := X \times I$ and the time-slices $\mathcal{X}_t := X \times \{ t \}$ equipped with the length metrics $d_t := d_{g(t)}$.
It may be tempting to retain the product structure $X \times I$ on $\mathcal{X}$, i.e. to record the set of worldlines $t \mapsto (x,t)$.
However, this turns out to be unnatural.
Instead, we will view the time-slices $(\mathcal{X}_t, d_t)$ as separate, metric spaces, whose points may not even be in 1-1 correspondence to some given space $X$.
It remains to record some other relation between these metric spaces $(\mathcal{X}_t, d_t)$.
This will be done via the conjugate heat kernel $K(x,t;y,s)$ --- an important object in the study of Ricci flows.
For fixed $(x,t) \in M \times I$ and $s < t$ this kernel satisfies the backwards conjugate\footnote{(\ref{eq_conj_HK}) is the $L^2$-conjugate of the standard (forward) heat equation and $K(\cdot,\cdot;y,t)$ is a heat kernel placed at $(y,t)$.} heat equation on a Ricci flow background:
\begin{equation} \label{eq_conj_HK}
(- \partial_s - \triangle_{g(s)} + R_{g(s)}) K(x,t;\cdot,s) = 0,
\end{equation}
centered at $(x,t)$.
This kernel has the property that for any $(x,t)$ and $s < t$
\[ \int_M K(x,t;\cdot, s) dg(s) = 1, \]
which motivates the definition of the following probability measures:
\[ d\nu_{(x,t);s} := K(x,t;\cdot,s) dg(s), \quad \nu_{(x,t);t} = \delta_x. \]
This is the additional information that we will record.
So we define:
\begin{Definition}
A \emph{metric flow} is (essentially\footnote{This is a simplified definition.}) given by a pair
\[ \big( (\mathcal{X}_t, d_t)_{t \in I}, (\nu_{x;s})_{x \in \mathcal{X}_t, s < t, s \in I} \big) \]
consisting of a family of metric spaces $(\mathcal{X}_t, d_t)$ and probability measures $\nu_{x;s}$ on $\mathcal{X}_s$, which satisfy certain basic compatibility relations.
\end{Definition}
So given points $x \in \mathcal{X}_t$, $y \in \mathcal{X}_s$ at two times $s < t$, it is not possible to say whether ``$y$ corresponds to $x$''.
Instead, we only know that ``$y$ belongs to the past of $x$ with a probability density of $d\nu_{x;s}(y)$''.
This definition is surprisingly fruitful.
For example, it is possible to use the measures $\nu_{x;s}$ to define a natural topology on $\mathcal{X}$ and to understand when and in what sense the geometry of time-slices $\mathcal{X}_t$ depends continuously on $t$.
The concept of metric flows also allows the definition of a natural notion of geometric convergence --- $\mathbb{F}$-convergence --- which is similar to Gromov-Hausdorff convergence.
Even better, this notion can be phrased on terms of a certain $d_{\mathbb{F}}$-distance, which is similar to the Gromov-Hausdorff distance, and the Compactness Theorem~\ref{Thm_compactness_RF} can be expressed as a statement on the compactness of a certain subset of metric flow (pairs),\footnote{Strictly speaking, $\mathbb{F}$-convergence and $d_{\mathbb{F}}$-distance concern metric flow \emph{pairs}, $(\mathcal{X}, (\nu_{x;t}))$, where the second entry serves as some kind of substitute of a basedpoint.} just as in the case of Gromov-Hausdorff compactness.
\begin{figure}
\centering
\includegraphics[width=65mm]{bryant-evolving-basepoint}
\caption{The Bryant soliton and a conjugate heat kernel starting at $(x_{\Bry},0)$.\label{fig_Bryant_evolution}}
\end{figure}
Lastly, I will sketch an example that illustrates why it was so important that we have divorced ourselves from the concept of worldlines.
Consider the Bryant soliton $(M_{\Bry}, (g_{\Bry}(t))_{t \leq 0})$ (see Figure~\ref{fig_Bryant_evolution}) from Subsection~\ref{subsec_dumbell}.
Recall that every time-slice $(M_{\Bry}, g_{\Bry}(t))$ is isometric to the same rotationally symmetric model with center $x_{\Bry}$.
By Theorems~\ref{Thm_compactness_RF} and \ref{Thm_part_reg_RF} any pointed sequence of blow-\emph{downs} ($\lambda_i \to 0$),
\[ (M_{\Bry}, (\lambda_i^2 g_{\Bry}( \lambda_i^{-2} t))_{t \leq 0}, (x_{\Bry},0)), \]
$\mathbb{F}$-converges to a pointed metric flow $\mathcal{X}$ that is regular on a large set.
What is this $\mathbb{F}$-limit $\mathcal{X}$?
For any fixed time $t < 0$, the sequence of pointed Riemannian manifolds $(M_{\Bry}, \lambda_i^2 g_{\Bry}( \lambda_i^{-2} t), x_{\Bry})$ converges to a pointed ray of the form $([0, \infty), 0)$.
This seems to contradict Theorem~\ref{Thm_part_reg_RF}.
However, here we have implicitly used the concept of worldlines, because we have used the point $(x_{\Bry}, t)$ corresponding to the ``official'' basepoint $(x_{\Bry}, 0)$ at time $t$.
Instead, we have to focus on the ``past'' of $(x_{\Bry}, 0)$, i.e. the region of $(M_{\Bry}, \lambda_i^2 g_{\Bry}( \lambda_i^{-2} t))$ where the conjugate heat kernel $\nu_{(x_{\Bry},0);\lambda_i^{-2}t}$ is concentrated.
This region is cylindrical of scale $\sim \sqrt{|t|}$, because the conjugate heat kernel ``drifts away from the tip'' at an approximate linear rate.
In fact, one can show that the blow-down limit $\mathcal{X}$ is isometric to a round shrinking cylinder that develops a singularity at time $0$.
While this may seem slightly less intuitive at first, it turns out to be a much more natural way of looking at it.
\subsection{What's next?}
This new theory demonstrates that, at least on an analytical level, Ricci flows behave similarly in higher dimension as they do in dimension 3.
However, while there are only a handful of possible singularity models in dimension 3, gaining a full understanding of all such models in higher dimensions (e.g. classifying gradient shrinking solitons) seems like an intimidating task.
Some past work in dimension 4 (e.g. by Munteanu and Wang) has demonstrated that most \emph{non-compact} gradient shrinking solitons have ends that are either cylindrical or conical.
This motivates the following conjecture:
\begin{Conjecture}
Given a closed Riemannian 4-manifold $(M,g)$ there is a certain kind of ``Ricci flow through singularities'' in which topological change occurs along cylinders or cones and in which time-slices are allowed to have isolated orbifold singularities.
\end{Conjecture}
Showing the existence of such a flow would be an analytical challenge, given that the construction in dimension 3 already filled several hundred pages.
However, I currently don't see a reason why such a flow should not exist.
Even more exciting would be the question what such a flow (or an analogue in higher dimensions) would accomplish on a topological level.
It is unlikely that it would allow us to prove the smooth Poincar\'e Conjecture in dimension 4, because even in the best possible case, such a flow seems to provide insufficient topological information.
A more feasible application would the $\frac{11}{8}$-Conjecture, which states that for every closed spin 4-manifold we have
\begin{equation} \label{eq_11_8}
b_2(M) \geq \tfrac{11}8 |\sigma(M)|.
\end{equation}
This conjecture is the missing piece in the classification of closed, simply connected, smooth 4-manifolds up to homeomorphy (due to Donaldson, Freedman and Kirby).
At least on a heuristic level, it would be suited for a Ricci flow approach since (closed) Einstein manifolds and spin gradient shrinking solitons automatically satisfy (\ref{eq_11_8}) --- due to the Hitchin-Thorpe inequality in the Einstein case or the fact that gradient shrinking solitons have positive scalar curvature.
Other potential applications would be questions concerning the topology of 4-manifolds that admit metrics of positive scalar curvature.
Lastly, there also seems to be potential in K\"ahler geometry, for example towards the Minimal Model Program and the Abundance Conjecture.
Time will tell how far Ricci flow methods will take us precisely.
Almost 20 years ago, geometric analysts and topologists were busy digesting Perelman's work on the Poincar\'e and Geometrization Conjectures, thus closing an important chapter in the field.
Today, we have good reasons to be optimistic that further topological applications are on the horizon.
The field certainly still has an exciting future ahead and hopefully encourages further research and collaboration between geometers, analysts, topologists and complex geometers.
|
\section{Boolean Circuits}
\subsection{Exact Circuit Synthesis}
A~Boolean \emph{straight line program}
of size~$r$ for input variables $(x_1, \dotsc, x_n)$
is a~sequence of~$r$~instructions where each
instruction $g \gets h \circ k$
applies a~binary Boolean operation~$\circ$ to
two operands $h,k$ each of which is either an~input bit
or the result of a~previous instruction.
If $m$~instructions are designated as outputs,
the straight line program computes a~function
$\{0,1\}^n \to \{0,1\}^m$ in a~natural way.
We denote the set of all such functions by $B_{n,m}$ and we let $B_n=B_{n,1}$.
For
a~Boolean function $f \colon \{0,1\}^n \to \{0,1\}^m$,
by $\size(f)$ we denote the minimum size of
a~straight line program
computing~$f$. A~Boolean \emph{circuit}
shows a~flow graph of a~program.
Figure~\ref{figure:sum23} gives an~example for
the
$\SUM_n \colon \{0,1\}^n \to \{0,1\}^l$ function
that computes the binary representation of~the sum of~$n$~bits:
\[\SUM_n(x_1, \dotsc, x_n)=(w_0, w_1, \dotsc, w_{l-1})\colon \sum_{i=1}^{n}x_i=\sum_{i=0}^{l-1}2^iw_i \text{, \, where } l=\lceil \log_2(n+1)\rceil \, .\]
This function transforms $n$~bits
of weight~0 into $l$~bits
of~weights $(0,1,\dotsc,l-1)$.
\begin{figure}[!ht]
\begin{minipage}{.28\textwidth}
\inputminted[firstline=15,lastline=18]{python}{../tutorial.py}
\end{minipage}
\begin{minipage}{.18\textwidth}
\begin{tikzpicture}[label distance=-.9mm,scale=.9]
\foreach \n/\x/\y in {1/0/1, 2/1/1}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate, label=left:$w_1$] (g1) at (0,0) {$\land$};
\node[gate, label=right:$w_0$] (g2) at (1,0) {$\oplus$};
\foreach \f/\t in {x1/g1, x1/g2, x2/g1, x2/g2}
\draw[->] (\f) -- (\t);
\end{tikzpicture}
\end{minipage}
\begin{minipage}{.33\textwidth}
\inputminted[firstline=22,lastline=28]{python}{../tutorial.py}
\end{minipage}
\begin{minipage}{.18\textwidth}
~
\begin{tikzpicture}[label distance=-.9mm,scale=.9]
\foreach \n/\x/\y in {1/0/3, 2/1/3, 3/2/3}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate,label=left:$a$] (g1) at (0.5,2) {$\oplus$};
\node[gate,label=left:$b$] (g2) at (1.5,2) {$\oplus$};
\node[gate,label=left:$c$] (g3) at (0.5,1) {$\lor$};
\node[gate, label=right:$w_0$] (g4) at (1.5,1) {$\oplus$};
\node[gate, label=right:$w_1$] (g5) at (0.5,0) {$\oplus$};
\foreach \f/\t in {x1/g1, x2/g1, x2/g2, x3/g2, g1/g3, g2/g3, g1/g4, g3/g5, g4/g5}
\draw[->] (\f) -- (\t);
\path (x3) edge[bend left,->] (g4);
\end{tikzpicture}
\end{minipage}
\caption{Optimal size straight line programs and circuits for $\SUM_2$ and $\SUM_3$. These two circuits are known as~\emph{half adder} and \emph{full adder}.}
\label{figure:sum23}
\end{figure}
The straight line
programs are given
in~\texttt{Python} programming language
so~that
it~is particularly easy to~verify their correctness.
For example, the program for $\SUM_3$ can be verified
with just three lines of code:
\inputminted[firstline=32,lastline=36]{python}{../tutorial.py}
Determining $\size(f)$ requires
proving lower bounds:
to~show that $\size(f)> \alpha$,
one needs to~prove that \emph{every} circuit
of~size at~most~$\alpha$ does not compute~$f$.
Known lower bounds are far from being satisfactory:
the strongest known lower bound for a~function family
in~NP is~$(3+1/86)n-o(n)$~\cite{DBLP:conf/focs/FindGHK16}.
Here, by a~\emph{function family} we mean an~infinite sequence of~functions $\{f_n\}_{n=1}^{\infty}$ where $f_n \in B_n$.
Even proving lower bounds for specific functions (rather than function families) is~difficult. Brute force approaches become impractical quickly: $|B_n|=2^{2^n}$, hence already for $n=6$, one cannot just enumerate all
functions from~$B_n$; also, the number of~circuits
of~size~$s$ is $s^{\Theta(s)}$, hence checking all circuits of size~$s$ takes reasonable time for small values of~$s$ only.
Knuth~\cite{Knuth:2008:ACP:1377542} found the exact circuit size of all functions from~$B_4$ and~$B_5$.
Finding the exact value of $\size(f)$ for $f \in B_6$
is~already a~difficult computational task for modern
computers and techniques. One approach
is~to~translate a~statement ``there exists a~circuit
of~size~$s$ computing~$f$'' to a~Boolean formula and
to~pass~it to~a~SAT-solver. Then, if the formula
is~satisfiable, one decodes a~circuit from its satisfying
assignment; otherwise, one gets a~(computer generated) proof of a~lower bound $\size(f) > s$. This circuit synthesis approach was proposed
by~Kojevnikov et al.~\cite{DBLP:conf/sat/KojevnikovKY09}
and, since then, has been used in various circuit synthesis programs (\texttt{abc}~\cite{abc}, \texttt{mockturtle}~\cite{EPFLLibraries}, \texttt{sat-chains}~\cite{knuthreduction}).
The state-of-the-art SAT-solvers are surprisingly efficient and allow to~handle
various practically important problems (with millions
of~variables) and even help to~resolve open
problems~\cite{DBLP:conf/cade/BrakensiekHMN20}.
Still, already for small values of $n$~and~$s$ the problem
of~finding a~circuit of~size~$s$ for a~function from~$B_n$
is~difficult for SAT-solvers.
We demonstrate the limits
of~this approach on~\emph{counting} functions:
\begin{equation}\label{eq:mod}
\MOD_n^{m,r}(x_1, \dotsc, x_n)=[x_1+\dotsb+x_n \equiv r \bmod m]
\end{equation}
(here, $[\cdot]$ is the Iverson bracket: $[S]$~is equal to~$1$ if $S$~is true and is equal to~$0$ otherwise).
Using SAT-solvers, Knuth~\cite[solution to exercise~$480$]{Knuth:2015:ACP:2898950}
found $\size(\MOD_n^{3,r})$ for all $3 \le n \le 5$ and all $0 \le r \le 2$:
\begin{center}
\begin{tabular}{lccc}
\toprule
& $r=0$ & $r=1$ & $r=2$\\
\midrule
$n=3$ & $3$ & $4$ & $4$\\
$n=4$ & $7$ & $7$ & $6$\\
$n=5$ & $10$ & $9$ & $10$\\
\bottomrule
\end{tabular}
\end{center}
Generalizing this pattern, he~made the following conjecture:
\begin{equation}\label{conjecture}
\size(\MOD_n^{3,r})=3n-5-[(n+r) \equiv 0\bmod 3] \text{ for all $n \ge 3$ and $r$.}
\end{equation}
He was also able to~prove (using SAT-solvers)
that $\size(\MOD_6^{3,0})=12$
and wrote:
``The case $n=6$ and $r \neq 0$, which lies tantalizingly close to the limits of
today's solvers, is still unknown.''
To~summarize, our current abilities for checking whether there exists a~Boolean circuit of~size~$s$ are roughly the following:
\begin{itemize}
\item for $s \le 6$, this can be~done in a~few seconds;
\item for $7 \le s \le 12$, this can (sometimes)
be~done in a~few days;
\item for $s \ge 13$, this is~out of~reach.
\end{itemize}
\subsection{New Results}
In~this paper, we~explore the limits of~the following natural idea: given a~circuit, try to~improve its size by~improving (using SAT-solvers, for example) the size
of~its subcircuit of~size seven. This is a~kind of a~local search approach: we~have no~possibility to~go through the whole space of all circuits, but we can at~least
search in a~neighborhood of a~given circuit.
This allows~us to~work with circuits consisting
of many gates.
As the results of experiments, we~show several circuits
for which the approach described above leads to improved upper bounds.
\begin{itemize}
\item We support Knuth's conjecture~\eqref{conjecture} for $\MOD_n^{3,r}$ by~proving the matching upper bound:
\[\size(\MOD_n^{3,r}) \le 3n-5-[(n+r) \equiv 0\bmod 3] \text{ for all $n \ge 3$ and $r$.}\]
This improves slightly the previously known upper bound $\size(\MOD_n^{3,r}) \le 3n-4$ by~Demenkov et al.~\cite{DBLP:journals/ipl/DemenkovKKY10}.
\item We~present improvements for $\size(\SUM_n)$ for various small~$n$ and show that some of~these circuits and their parts can be~used as~building blocks to~design efficient circuits for other functions in~semiautomatic fashion. In~particular, we~show that a~part of an~optimal circuit for $\SUM_5$ can be~used to~build optimal circuits of~size~$2.5n$ for $\MOD_n^{4,r}$~\cite{DBLP:journals/mst/Stockmeyer77} and best known circuits of~size~$4.5n+o(n)$ for $\SUM_n$~\cite{DBLP:journals/ipl/DemenkovKKY10}.
In~turn, an~efficient circuit for $\SUM_5$ can be~found
in a~few seconds if~one starts from a~standard circuit for $\SUM_5$ composed
out of~two full adders and one half adder.
\item We~design new circuits for the threshold function defined as~follows:
\[\THR_n^k(x_1, \dotsc, x_n)=[x_1+\dotsb+x_n \ge k] \, .\]
The best known upper bounds for $\THR$ are the following:
\begin{align*}
\size(\THR_n^k) &\le kn+o(n) \text{ for $2 \le k \le 4$~\cite{Dunne84} (see also~\cite[6.2, Theorem~2.3]{DBLP:books/teu/Wegener87}),}\\
\size(\THR_n^k) &\le 4.5n+o(n) \text{ for $5 \le k$~\cite{DBLP:journals/ipl/DemenkovKKY10}.}
\end{align*}
We~get the following improvement:
\begin{align*}
\size(\THR_n^k) &\le (4.5-2^{2-\lceil\log_2k\rceil})n+o(n) \text{ for $4 \le k = O(1)$.}
\end{align*}
In~particular, $\size(\THR_n^4) \le 3.5n+o(n)$ and $\size(\THR_n^k) \le 4n+o(n)$ for $5 \le k \le 8$.
\end{itemize}
The improved upper bounds are obtained in a~semiautomatic fashion:
\begin{enumerate}
\item first, we~automatically improve a~given small circuit with a~fixed number of~inputs using SAT-solvers;
\item then, we~generalize~it to~every input size.
\end{enumerate}
For some function families,
the second step is~already known (for example, given a~small circuit
for $\SUM_5$, it is not difficult to~use~it as a~building block
to~design an~efficient circuit for $\SUM_n$ for every~$n$; see Section~\ref{section:sum}), though in~general this still needs to~be
done manually.
\subsection{Related work}
The approach we~use in this paper follows the SAT-based
local improvement method (SLIM): to~improve an~existing
discrete structure one goes through all its substructures
of~size accessible to a~SAT-solver.
SLIM has been applied successfully to~the following
structures:
branchwidth~\cite{DBLP:journals/tocl/LodhaOS19},
treewidth~\cite{DBLP:conf/sat/FichteLS17},
treedepth~\cite{DBLP:conf/cp/RamaswamyS20},
Bayesian network structure learning~\cite{DBLP:conf/aaai/RamaswamyS21},
decision tree learning~\cite{DBLP:conf/aaai/SchidlerS21}.
\section{Program: Feature Overview and Evaluation}
The program is~implemented in~\texttt{Python}.
We give a~high-level overview of~its main features below.
All the code shown below
can be~found in~the
file \texttt{tutorial.py} at~\cite{git-improvement}.
One may run~it after installing a~few \texttt{Python} modules. Alternatively, one may run the Jupyter notebook
\texttt{tutorial.ipynb} in~the cloud (without installing
anything) by pressing the badge ``Colab''
at~the repository page~\cite{git-improvement}.
\subsection{Manipulating Circuits}
This is done through the \mintinline{python}{Circuit}
class. One can load and save circuits as~well~as
print and draw them. A~nicely looking layout of
a~circuit is produced by the \texttt{pygraphviz} module~\cite{pygraphviz}. The program also contains some built-in
circuits that can be used as~building blocks.
The following sample code constructs a~circuit
for $\SUM_5$ out of two full adders and
one half adder. This construction is~shown
in~Figure~\ref{figure:sumfive}(a). Then,
the circuit is verified via the
\mintinline{python}{check_sum_circuit} method.
Finally, the circuit is drawn. As a~result, one gets
a~picture similar to~the one in~Figure~\ref{figure:sumfive}(b).
\inputminted[firstline=40,lastline=47]{python}{../tutorial.py}
\begin{figure}[t]
\begin{mypic}
\begin{scope}[scale=.9]
\begin{scope}[yshift=-10mm]
\foreach \n in {1,...,5}
\node[input] (\n) at (\n,6) {$x_{\n}$};
\draw (0.5, 5.5) rectangle (3.5, 4.5); \node at (2, 5) {$\SUM_3$};
\foreach \n in {1, 2, 3}
\draw[->] (\n) -- (\n, 5.5);
\draw (2.5, 3.5) rectangle (5.5, 2.5); \node at (4, 3) {$\SUM_3$};
\path (3, 4.5) edge[->] node[l] {0} (3, 3.5);
\foreach \n in {4, 5}
\draw[->] (\n) -- (\n, 3.5);
\draw (1.5, 1.5) rectangle (3.5, 0.5); \node at (2.5, 1) {$\SUM_2$};
\path (2, 4.5) edge[->] node[l] {1} (2, 1.5);
\path (3, 2.5) edge[->] node[l] {1} (3, 1.5);
\node[input] (w2) at (2,-1) {$w_2$};
\node[input] (w1) at (3,-1) {$w_1$};
\node[input] (w0) at (4.5,-1) {$w_0$};
\path (2, 0.5) edge[->] node[l] {1} (w2);
\path (3, 0.5) edge[->] node[l] {0} (w1);
\path (4.5, 2.5) edge[->] node[l] {0} (w0);
\end{scope}
\begin{scope}[label distance=-1mm, xshift=70mm, yshift=20mm]
\foreach \n/\x/\y in {1/0/3, 2/1/3, 3/2/3, 4/2.5/1, 5/3.5/1}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate,label=left:$g_1$] (g1) at (0.5,2) {$\oplus$};
\node[gate,label=left:$g_2$] (g2) at (1.5,2) {$\oplus$};
\node[gate,label=left:$g_3$] (g3) at (0.5,1) {$\lor$};
\node[gate,label=left:$g_4$] (g4) at (1.5,1) {$\oplus$};
\node[gate,label=left:$g_5$] (g5) at (0.5,0) {$\oplus$};
\node[gate,label=left:$g_6$] (g6) at (2,-1) {$\oplus$};
\node[gate,label=right:$g_7$] (g7) at (3,-1) {$\oplus$};
\node[gate,label=right:$g_8$] (g8) at (2,-2) {$\lor$};
\node[gate, label=right:$w_0$] (g9) at (3,-2) {$\oplus$};
\node[gate, label=right:$g_9$] (g10) at (2,-3) {$\oplus$};
\node[gate, label=right:$w_1$] (g11) at (2,-4) {$\oplus$};
\node[gate, label=left:$w_2$] (g12) at (1,-4) {$\land$};
\foreach \f/\t in {x1/g1, x2/g1, x2/g2, x3/g2, g1/g3, g2/g3, g1/g4, g3/g5, g4/g5, g4/g6, x4/g6, x4/g7, x5/g7, g6/g8, g7/g8, g8/g10, g6/g9, g9/g10, g10/g11, g10/g12}
\draw[->] (\f) -- (\t);
\path (x3) edge[->,bend left] (g4);
\path (x5) edge[->,bend left=35] (g9);
\path (g5) edge[->,bend right=25] (g11);
\path (g5) edge[->,bend right=15] (g12);
\draw[dashed] (-0.5,-0.25) rectangle (2,2.5);
\draw[dashed] (1.25,-3.25) rectangle (4,-0.5);
\draw[dashed] (0,-3.5) rectangle (3,-4.5);
\end{scope}
\begin{scope}[label distance=-1mm, xshift=120mm, yshift=20mm]
\foreach \n/\x/\y in {1/0/3, 2/1/3, 3/2/3, 4/2.5/1, 5/3.5/1}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate,label=left:$g_1$] (g1) at (0.5,2) {$\oplus$};
\node[gate,label=left:$g_2$] (g2) at (1.5,2) {$\oplus$};
\node[gate,label=left:$g_3$] (g3) at (0.5,1) {$\lor$};
\node[gate,label=left:$g_4$] (g4) at (1.5,1) {$\oplus$};
\node[gate,label=left:$g_5$] (g5) at (0.5,0) {$\oplus$};
\node[gate,label=left:$g_6$] (g6) at (2,0) {$\oplus$};
\node[gate,label=right:$g_7$] (g7) at (3,0) {$\oplus$};
\node[gate,label=right:$g_8$] (g8) at (2,-1) {$>$};
\node[gate, label=right:$w_0$] (g9) at (3,-1) {$\oplus$};
\node[gate, label=right:$w_1$] (g10) at (2,-2) {$\oplus$};
\node[gate, label=right:$w_2$] (g11) at (1.5,-3) {$>$};
\foreach \f/\t in {x1/g1, x2/g1, x2/g2, x3/g2, g1/g3, g2/g3, g1/g4, g3/g5, g4/g5, x4/g6, g4/g6, x4/g7, x5/g7, g6/g8, g7/g8, g7/g9, g3/g10, g8/g10, g10/g11, g5/g11}
\draw[->] (\f) -- (\t);
\path (x3) edge[->,bend left] (g4);
\path (g4) edge[->,bend left=20] (g9);
\end{scope}
\foreach \x/\n in {3/a, 9/b, 14/c}
\node at (\x,-3) {(\n)};
\end{scope}
\end{mypic}
\caption{(a)~A~schematic circuit for $\SUM_5$ composed out of two full adders and one half adder. (b)~The corresponding circuit of size~$12$. (c)~An~improved circuit of size~$11$.}
\label{figure:sumfive}
\end{figure}
\subsection{Finding Efficient Circuits}
The class \mintinline{python}{CircuitFinder}
allows to~check whether there exists a~circuit
of~the required
size for a~given Boolean function. For example,
one may discover the full adder as follows. (The function
\mintinline{python}{sum_n} returns the list of $\lceil \log_2(n+1) \rceil$ bits of~the binary representation of~the sum of~$n$~bits.)
\inputminted[firstline=51,lastline=58]{python}{../tutorial.py}
This is~done by~encoding the task as~a~CNF formula
and invoking a~SAT-solver
(via the \texttt{pysat} module~\cite{DBLP:conf/sat/IgnatievMM18}).
The reduction to~SAT is~described
in~\cite{DBLP:conf/sat/KojevnikovKY09}. Basically, one translates
a~statement ``there exists a~circuit of~size~$s$ comuting a~given function~$f \colon \{0,1\}^n \to \{0,1\}^m$'' to CNF-SAT. To~do this, one introduces many
auxiliary variables: for example, for every $x \in \{0,1\}^n$ and every $1 \le i \le r$, one uses a~variable that is~responsible for the value of the $i$-th gate
on an~input~$x$.
As~mentioned in~the introduction, the limits
of~applicability of~this approach (for finding a~circuit
of~size~$s$) are roughly the following:
for $s \le 6$, it~usually works in~less than a~minute;
for $7 \le s \le 12$, it~may already take up~to
several hours or days; for $s \ge 13$, it~becomes almost impractical. The running time may vary a~lot
for inputs of the same length. In~particular,
it usually takes much longer to~prove that
the required circuit does not exist (by~proving that the corresponding formula is~unsatisfiable). Table~\ref{table:runningtimes} reports the running time
of~this approach on several datasets.
\begin{table}[!ht]
\begin{center}
\begin{tabular}{lllr}
\toprule
function & circuit size & status & time (sec.)\\
\midrule
$\SUM_5$ & 12 & SAT & 141.4\\
$\SUM_5$ & 11 & SAT & 337.8\\
$\MOD_4^{3,0}$ & $7$ & SAT & 0.2\\
$\MOD_4^{3,0}$ & $6$ & UNSAT & 1178.8\\
$\MOD_4^{3,1}$ & $7$ & SAT & 0.2\\
$\MOD_4^{3,1}$ & $6$ & UNSAT & 1756.5\\
$\MOD_4^{3,2}$ & $6$ & SAT & 0.2\\
$\MOD_4^{3,2}$ & $5$ & UNSAT & 12.6\\
$\MOD_5^{3,0}$ & $10$ & SAT & 90.1\\
$\MOD_5^{3,1}$ & $9$ & SAT & 50.1\\
$\MOD_5^{3,2}$ & $10$ & SAT & 74.3\\
\bottomrule
\end{tabular}
\end{center}
\caption{The running time of \texttt{CircuitFinder} on~various Boolean functions.} \label{table:runningtimes}
\end{table}
\subsection{Improving Circuits}
The method \mintinline{python}{improve_circuit}
goes through all subcircuits of a~given size
of a~given circuit and checks whether any
of~them can be~replaced by a~smaller subcircuit
(computing the same function) via \mintinline{python}{find_circuit}. For example, applying this method
to~the circuit from Figure~\ref{figure:sumfive}(b)
gives the circuit from Figure~\ref{figure:sumfive}(c)
in~a~few seconds.
\inputminted[firstline=62,lastline=67]{python}{../tutorial.py}
Table~\ref{table:improvementrunningtimes} shows the time
taken by~\texttt{improve\_circuit} to~improve some
of~the known circuits for $\SUM$, $\MOD^3$, and $\THR^4$.
For $\SUM$, we~start from known circuits of size about $5n$
(composed out of~full adders and half adders). For $\MOD^3$,
we start from circuits of size $3n-4$ presented by~Demenkov et al.~\cite{DBLP:journals/ipl/DemenkovKKY10}. For $\THR^4$, we start from circuits of
size about~$5n$ (the start by~computing $\SUM_n$ and then compare the
resulting $\log n$-bit integer to~$4$).
\begin{table}[ht]
\begin{center}
\begin{tabular}{llr}
\toprule
function & circuit size & time (sec.)\\
\midrule
$\SUM_5$ & $12 \to 11$ & 6.7\\
$\SUM_7$ & $20 \to 19$ & 5.8\\
\midrule
$\MOD_6^{3,0}$ & $15 \to 14$ & 17.0\\
$\MOD_6^{3,1}$ & $15 \to 14$ & 17.2\\
$\MOD_6^{3,2}$ & $14 \to 13$ & 16.7\\
$\MOD_7^{3,0}$ & $17 \to 16$ & 31.3\\
$\MOD_7^{3,1}$ & $17 \to 16$ & 33.6\\
$\MOD_7^{3,2}$ & $16 \to 15$ & 30.5\\
\midrule
$\THR^4_5$ & $23 \to 10$ & $38.6$\\
$\THR^4_6$ & $28 \to 14$ & $42.1$\\
$\THR^4_7$ & $31 \to 17$ & $43.8$\\
$\THR^4_8$ & $40 \to 22$ & $55.1$\\
\bottomrule
\end{tabular}
\end{center}
\caption{The running time of \texttt{improve\_circuit} on~various Boolean functions.} \label{table:improvementrunningtimes}
\end{table}
\section{New Circuits}
In~this section, we~present new circuits for symmetric functions found with the help of~the program. A~function $f(x_1,\dotsc,x_n)$ is called \emph{symmetric} if~its value depends on~$\sum_{i=1}^nx_i$ only. They are among the most basic Boolean functions:
\begin{itemize}
\item to~specify an~arbitrary Boolean function
from~$B_n$, one needs
to~write down its truth table of length~$2^n$; symmetric functions allow for more compact representation: it is enough to specify $n+1$ bits (for each of $n+1$ values
of~$\sum_{i=1}^nx_i$);
\item circuit complexity of almost all functions
of~$n$ variables is exponential ($\Theta(2^n/n)$), whereas any symmetric function can~be computed by a~linear size circuit ($O(n)$).
\end{itemize}
Despite simplicity of~symmetric functions, we still do~not
know how optimal circuits look like for most of~them. Below, we~present new circuits for some of these functions.
\subsection{Sum Function}\label{section:sum}
The $\SUM$ function is a~fundamental symmetric function: for any symmetric $f \in B_n$, $\size(f)\le \size(\SUM_n)+o(n)$. The reason for this is that any function from~$B_n$ can be~computed by a~circuit of size $O(2^n/n)$ by the results of Muller~\cite{M56} and Lupanov~\cite{Lup59}. This allows to compute any symmetric $f(x_1, \dotsc, x_n) \in B_n$ as follows: first, compute $\SUM_n(x_1, \dotsc, x_n)$ using $\size(\SUM_n)$ gates; then, compute the resulting bit using at most $O(2^{\log n}/\log n)=o(n)$ gates. For the same reason, any lower bound $\size(f) \ge \alpha$ for
a~symmetric function~$f \in B_n$ implies a~lower bound $\size(\SUM_n) \ge \alpha-o(n)$. Currently, we know the following bounds for $\SUM_n$:
\[2.5n-O(1) \le \size(\SUM_n) \le 4.5n+o(n) \, .\]
The lower bound is~due to~Stockmeyer~\cite{DBLP:journals/mst/Stockmeyer77}, the upper bound is~due to~Demenkov et al.~\cite{DBLP:journals/ipl/DemenkovKKY10}.
A~circuit for $\SUM_n$ can~be constructed from circuits for $\SUM_k$ for some small~$k$. For example,
using full and half adders as~building blocks, one can compute $\SUM_n$ (for any~$n$) by a~circuit of size $5n$ as~follows. Start from $n$~bits $(x_1, \dotsc, x_n)$ of~weight~$0$. While there are three bits of the same weight~$k$, replace them by~two bits of weights~$k$ and~$k+1$ using a~full adder. This way, one gets at~most two bits of~each weight $0,1,\dotsc,l-1$ ($l=\lceil \log_2(n+1)\rceil$) in at most $5(n-l)$ gates (as each full adder reduces the number of~bits). To~leave exactly one bit
of~each weight, it suffices to use at~most~$l$ half
or~full adders ($o(n)$ gates). Let~us denote the
size of the resulting circuit by~$s(n)$. The second row
of~Table~\ref{table:sum} shows the values of~$s(n)$ for
some $n \le 15$ (see~(28) in~\cite{Knuth:2008:ACP:1377542}).
\begin{table}
\begin{center}
\begin{tabular}{lrrrrrrrrrrrrrrrrrrrrrr}
\toprule
$n$ & $2$ & $3$ & $4$ & $5$ & $6$ & $7$ & $8$ & $9$ & $10$ & $15$
\\
\midrule
$s(n)$ &$2$ & $5$ & $9$ & $12$ & $17$ & $20$ & $26$ & $29$& $34$&
$55$
\\
$\size(\SUM_n)$ & $2$ & $5$ & $9$ & $11$
& $\le 16$ & $\le 19$ & $\le 25$ & $\le 27$
& $\le 32$ & $\le 53$
\\
\bottomrule
\end{tabular}
\end{center}
\caption{The first line shows the value of~$n$. The second line gives the size $s(n)$ of~a~circuit for $\SUM_n$
composed out~of half and full adders. The third row shows known bounds for $\size(\SUM_n)$.}
\label{table:sum}
\end{table}
In a~similar fashion, one can get an~upper bound (see Theorem~1 in~\cite{DBLP:conf/date/Kulikov18})
\begin{equation}\label{eq:sumupper}
\size(\SUM_n) \le \frac{\size(\SUM_k)}{k-\lceil \log_2(k+1) \rceil} \cdot n + o(n) \, .
\end{equation}
This motivates the search for efficient circuits
for $\SUM_k$ for small values of~$k$. The bottom row
of~Table~\ref{table:sum} gives upper bounds that
we~were able to~find using the program
(the upper bounds for $n \le 7$ were found
by~Knuth~\cite{Knuth:2008:ACP:1377542}).
The table shows that the first value where $s(n)$ is not
optimal is $n=5$. The best upper bound for $\SUM_n$ given
by~\eqref{eq:sumupper} is $4.75n+o(n)$ for $n=7$. The upper
bound for $n=15$ is $53n/11+o(n)$ which is worse than the
previous upper bound. But if it~turned out that
$\size(\SUM_{15}) \le 52$,
it~would give a~better upper bound.
The found circuits for $\SUM_n$ for $n \le 15$
do~not allow to~improve the strongest known upper bound
$\size(\SUM_n) \le 4.5n+o(n)$ due to~Demenkov et al.~\cite{DBLP:journals/ipl/DemenkovKKY10}. Below,
we~present several interesting observations on~the found circuits.
\subsubsection{Best Known Upper Bound for the SUM Function}
The optimal circuit of size~$11$ for $\SUM_5$ shown in~Figure~\ref{figure:sumfive}(c) can be~used to get an~upper bound $4.5n+o(n)$ for $\size(\SUM_n)$
(though not through~\eqref{eq:sumupper} directly).
To~do this,
consider two consecutive $\SUM_3$ circuits shown in~Figure~\ref{figure:mdfa}(a).
\begin{figure}
\begin{center}
\begin{tikzpicture}
\begin{scope}[scale=.7]
\draw (1,0) rectangle (3,2); \node at (2,1) {$\SUM_3$};
\draw (5,0) rectangle (7,2); \node at (6,1) {$\SUM_3$};
\foreach \n/\x/\y in {3/0/1, 2/1.5/3, 1/2.5/3, 4/5.5/3, 5/6.5/3}
\node[input] (\n) at (\x,\y) {$x_{\n}$};
\foreach \n/\t/\x/\y in {a1/a_1/2/-1, b1/b_1/6/-1, b0/b_0/8/1}
\node[input] (\n) at (\x,\y) {$\t$};
\draw[->] (3)--(1,1);
\draw[->] (2)--(1.5,2);
\draw[->] (1)--(2.5,2);
\draw[->] (4)--(5.5,2);
\draw[->] (5)--(6.5,2);
\draw[->] (3,1)--(5,1);
\draw[->] (7,1)--(b0);
\draw[->] (2,0)--(a1);
\draw[->] (6,0)--(b1);
\node at (-1,1) {(a)};
\end{scope}
\begin{scope}[scale=.7,yshift=-60mm]
\draw (1,0) rectangle (7,2); \node at (4,1) {MDFA};
\foreach \n/\x/\y in {3/0/1, 2/1.5/4, 1/2.5/4, 4/5.5/4, 5/6.5/4}
\node[input] (\n) at (\x,\y) {$x_{\n}$};
\node[gate] (xor1) at (2.5,3) {$\oplus$};
\node[gate] (xor2) at (6.5,3) {$\oplus$};
\foreach \n/\t/\x/\y in {a1/a_1/2/-1, b1/{a_1 \oplus b_1}/6/-1, b0/b_0/8/1}
\node[input] (\n) at (\x,\y) {$\t$};
\draw[->] (3)--(1,1);
\draw[->] (2)--(1.5,2);
\draw[->] (1) -- (xor1); \draw[->] (2) -- (xor1);
\draw[->] (xor1) -- (2.5,2);
\draw[->] (4)--(5.5,2);
\draw[->] (5)-- (xor2); \draw[->] (xor2) -- (6.5,2); \draw[->] (4) -- (xor2);
\draw[->] (7,1)--(b0);
\draw[->] (2,0)--(a1);
\draw[->] (6,0)--(b1);
\node at (-1,1) {(b)};
\end{scope}
\begin{scope}[yscale=.8, xshift=80mm, yshift=-20mm]
\draw[draw=none, rounded corners=0,fill=gray!20] (0,1.5)--(1,1.5)--(1,2.5)--(2,2.5)--(2,0.5)--(2.5,0.5)--(2.5,-0.5)--(3.5,-0.5)--
(3.5,-2.5)--(0,-2.5)--(0,1.5);
\foreach \n/\x/\y in {1/0/3, 2/1/3, 3/2/3, 4/2.5/1, 5/3.5/1}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate] (g1) at (0.5,2) {$\oplus$};
\node[gate] (g2) at (1.5,2) {$\oplus$};
\node[gate] (g3) at (0.5,1) {$\lor$};
\node[gate] (g4) at (1.5,1) {$\oplus$};
\node[gate, label=left:$a_1$] (g5) at (0.5,0) {$\oplus$};
\node[gate] (g6) at (2,0) {$\oplus$};
\node[gate] (g7) at (3,0) {$\oplus$};
\node[gate] (g8) at (2,-1) {$>$};
\node[gate, label=right:$b_0$] (g9) at (3,-1) {$\oplus$};
\node[gate, label=right:$a_1 \oplus b_1$] (g10) at (2,-2) {$\oplus$};
\node[gate] (g11) at (1.5,-3) {$>$};
\foreach \f/\t in {x1/g1, x2/g1, x2/g2, x3/g2, g1/g3, g2/g3, g1/g4, g3/g5, g4/g5, x4/g6, g4/g6, x4/g7, x5/g7, g6/g8, g7/g8, g7/g9, g3/g10, g8/g10, g10/g11, g5/g11}
\draw[->] (\f) -- (\t);
\path (x3) edge[->,bend left] (g4);
\path (g4) edge[->,bend left=20] (g9);
\node at (1.5,-4) {(c)};
\end{scope}
\end{tikzpicture}
\end{center}
\caption{(a)~Two consecutive $\SUM_3$ blocks. (b)~The $\MDFA$ block. (c)~The highlighted part of the optimal circuit
for $\SUM_5$ computes $\MDFA$.}
\label{figure:mdfa}
\end{figure}
Its specification~is: $x_1+\dotsb+x_5=b_0+2(a_1+b_1)$, its size is~equal to~10. One can construct a~similar block, called MDFA (for modified double full adder), of size~8, whose specification is
\[\MDFA(x_1 \oplus x_2, x_2, x_3, x_4, x_4 \oplus x_5)=(b_0, a_1, a_1 \oplus b_1) \, ,\]
see Figure~\ref{figure:mdfa}(b).
The fact that $\MDFA$ uses the encoding $(p, p \oplus q)$ for pairs of bits $(p,q)$, allows to use~it recursively
to~compute $\SUM_n$. As~the original construction is~presented in~\cite{DBLP:journals/ipl/DemenkovKKY10}, below we~give a~sketch
only.
\begin{enumerate}
\item Compute $x_2 \oplus x_3, x_4 \oplus x_5, \dotsc, x_{n-1} \oplus x_n$ ($n/2$ gates).
\item Apply at~most~$n/2$ $\MDFA$ blocks (no~more than $4n$~gates).
\item The last MDFA block outputs two bits: $a$~and~$a\oplus b$. Instead of~them, one needs to~compute $a \oplus b$ and $a \land b$. To~achieve this,
it~suffices to apply $x>y=(x \land \overline{y})$ operation:
\[a \land b = a>(a \oplus b)\, .\]
\end{enumerate}
Figure~\ref{figure:sum17} shows an~example for $n=17$.
\begin{figure}[!ht]
\begin{center}
\begin{tikzpicture}[xscale=0.75, yscale=.6]
\foreach \n in {2,...,17}
\node[input] (\n) at (\n,6) {$x_{\n}$};
\foreach \i in {1,...,8} {
\tikzmath{\k=int(2*\i); \j=int(2*\i+1);}
\node[gate] (a) at (\j,5) {$\oplus$};
\draw[->] (\k) -- (a); \draw[->] (\j) -- (a);
\draw[->] (a) -- (\j,4); \draw[->] (\k) -- (\k,4);
}
\foreach \x/\y/\w in {2/4/3, 6/4/3, 10/4/3, 14/4/3, 2/2.5/7, 10/2.5/7, 2/1/15} {
\draw (\x-0.15,\y) rectangle (\x+\w+0.15,\y-1);
\node at (\x+\w/2,\y-0.5) {MDFA};
}
\foreach \y/\l/\w in {3.5/x_1/w_0, 2/0/w_1, 0.5/0/w_2} {
\node[input] (b) at (1,\y) {$\l$};
\draw[->] (b) -- (1.85,\y);
\node[input] (c) at (18,\y) {$\w$};
\draw[->] (17.15,\y) -- (c);
}
\foreach \x in {3, 4, 7, 8, 11, 12, 15, 16}
\draw[->] (\x,3) -- (\x,2.5);
\foreach \x in {4, 7, 12, 15}
\draw[->] (\x,1.5) -- (\x,1);
\foreach \x/\y in {5/3.5, 9/3.5, 13/3.5, 9/2}
\draw[->] (\x+0.15,\y) -- (\x+0.85,\y);
\node[input] (w3) at (12,-2) {$w_3$};
\draw[->] (12,0) -- (w3);
\node[gate] (x) at (7,-1) {$>$};
\draw[->] (7,0) -- (x);
\node[input] (w4) at (7,-2) {$w_4$};
\draw[->] (x) -- (w4);
\path (12,0) edge[->,out=-135,in=0] (x);
\end{tikzpicture}
\end{center}
\caption{A~circuit computing $\SUM_{17}$ composed out of~MDFA blocks.}
\label{figure:sum17}
\end{figure}
The $\MDFA$ block was constructed by~Demenkov et al.~\cite{DBLP:journals/ipl/DemenkovKKY10} in~a~semiautomatic
manner. And it turns out that $\MDFA$ is just a~subcircuit of the optimal circuit for $\SUM_5$! See Figure~\ref{figure:mdfa}(c).
\subsubsection{Best Known Circuits for SUM with New Structure}
For many upper bounds from the bottom row of~Table~\ref{table:sum}, we found circuits with the following interesting structure: the first thing the circuit computes is $x_1 \oplus x_2 \oplus \dotsb \oplus x_n$; moreover the variables $x_2, \dotsc, x_n$ are used for this only. This is best illustrated by an~example~--- see Figure~\ref{figure:xorsum}.
\begin{figure}[!ht]
\begin{mypic}
\begin{scope}[scale=.9]
\begin{scope}[xshift=20mm, yshift=30mm]
\draw[draw=none, rounded corners=0,fill=gray!20] (0.6,2.5)--(0.6,3.5)--(2.4,3.5)--(2.4,2.5);
\foreach \n/\x/\y in {1/0/4, 2/1/4, 3/2/4}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate] (g1) at (1,3) {$\oplus$};
\node[gate,label=right:$w_0$] (g2) at (2,3) {$\oplus$};
\node[gate] (g3) at (0,2) {$\oplus$};
\node[gate] (g4) at (1,1.5) {$\lor$};
\node[gate, label=right:$w_1$] (g5) at (2,1) {$\oplus$};
\foreach \f/\t in {x1/g1, x2/g1, g1/g2, x3/g2, x1/g3, g2/g3, g1/g4, g3/g4, g2/g5, g4/g5}
\draw[->] (\f) -- (\t);
\end{scope}
\begin{scope}[label distance=-1mm, xshift=70mm, yshift=20mm]
\draw[draw=none, rounded corners=0,fill=gray!20] (0.6,3.5)--(0.6,4.5)--(3.4,4.5)--(3.4,3.5);
\foreach \n/\x/\y in {1/0/5, 2/1/5, 3/2/5, 4/3/5}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate] (g1) at (1,4) {$\oplus$};
\node[gate] (g2) at (2,4) {$\oplus$};
\node[gate,label=right:$w_0$] (g3) at (3,4) {$\oplus$};
\node[gate] (g4) at (0.5,3) {$\oplus$};
\node[gate] (g5) at (3,3) {$>$};
\node[gate] (g6) at (1,2) {$\lor$};
\node[gate] (g7) at (2,1.5) {$\oplus$};
\node[gate,label=right:$w_1$] (g8) at (3,1) {$\oplus$};
\node[gate,label=left:$w_2$] (g9) at (2,0.5) {$>$};
\foreach \f/\t in {x1/g1, x2/g1, g1/g2, x3/g2, g2/g3, x4/g3, x1/g4, g2/g4, g2/g5, g3/g5, g4/g6, g1/g6, g6/g7, g2/g7, g7/g8, g5/g8, g8/g9, g7/g9}
\draw[->] (\f) -- (\t);
\end{scope}
\begin{scope}[label distance=-1mm, xshift=120mm, yshift=20mm]
\draw[draw=none, rounded corners=0,fill=gray!20] (0.6,3.5)--(0.6,4.5)--(4.4,4.5)--(4.4,3.5);
\foreach \n/\x/\y in {1/0/5, 2/1/5, 3/2/5, 4/3/5, 5/4/5}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate] (g1) at (1,4) {$\oplus$};
\node[gate] (g2) at (2,4) {$\oplus$};
\node[gate] (g3) at (3,4) {$\oplus$};
\node[gate,label=right:$w_0$] (g4) at (4,4) {$\oplus$};
\node[gate] (g5) at (0.5,3) {$\oplus$};
\node[gate] (g6) at (4,3) {$\oplus$};
\node[gate] (g7) at (1,2) {$\lor$};
\node[gate] (g8) at (3.5,2) {$>$};
\node[gate] (g9) at (2,1) {$\oplus$};
\node[gate,label=right:$w_1$] (g10) at (3,1) {$\oplus$};
\node[gate,label=right:$w_2$] (g11) at (2.5,0) {$>$};
\foreach \f/\t in {x1/g1, x2/g1, g1/g2, x3/g2, g2/g3, x4/g3, g3/g4, x5/g4, x1/g5, g2/g5, g2/g6, g4/g6, g1/g7, g5/g7, g3/g8, g6/g8, g2/g9, g7/g9, g8/g10, g7/g10, g9/g11, g10/g11}
\draw[->] (\f) -- (\t);
\end{scope}
\end{scope}
\end{mypic}
\caption{Optimal circuits computing $\SUM_n$
for $n=3,4,5$ with a~specific structure: every input,
except for~$x_1$, has out-degree~one.}
\label{figure:xorsum}
\end{figure}
These circuits can be~found using the following code.
It~demonstrates two new useful features: fixing gates and forbidding wires between some pairs of~gates.
\inputminted[firstline=71,lastline=86]{python}{../tutorial.py}
\subsubsection{Optimal Circuits for Counting Modulo~4}
The optimal circuit for $\SUM_5$ can be~used to~construct
an~optimal circuit of~size $2.5n+O(1)$ for $\MOD_{4,r}^n$ due
to~Stockmeyer~\cite{DBLP:journals/mst/Stockmeyer77}
(recall that $\MOD_n^{4,r}(x_1, \dotsc, x_n)=[x_1+\dotsb+x_n \equiv r \bmod 4]$).
To~do this, note that
there is a~subcircuit (of the circuit at~Figure~\ref{figure:sumfive}(c)) of size~9 that computes the two least significant bits ($w_0,w_1$) of $x_1+\dotsb+x_5$ (one removes the gates $g_5, w_2$). To~compute $x_1+\dotsb+x_n \bmod 4$, one first applies $\frac n4$ such blocks and then computes the parity of~the resulting bits of weight~$1$
(every block takes four fresh inputs as~well as~one bit of~weight~$0$ from the previous block).
The total size is~$9 \cdot \frac n4 + \frac n4=2.5n$.
Thus,
the circuit that Stockmeyer constructed in~1977 by~hand,
nowadays can be~found automatically in~a~few seconds.
\subsection{Modulo-3 Function}
In~\cite{DBLP:conf/sat/KojevnikovKY09}, Kojevnikov et~al.
presented circuits of~size $3n+O(1)$ for $\MOD_n^{3,r}$ (for any~$r$). Later,
Knuth~\cite[solution to exercise~$480$]{Knuth:2015:ACP:2898950} analyzed their construction and proved an~upper
bound $3n-4$. Also, by~finding the exact values
for $\size(\MOD_n^{3,r})$ for all $3 \le n \le 5$ and all $0 \le r \le 2$, he~made the conjecture~\eqref{conjecture}.
Using our program, we~proved the conjectured upper bound for all~$n$.
\begin{theorem}\label{theorem:mod3upper}
For all $n \ge 3$ and all $r \in \{0,1,2\}$,
\[\size(\MOD_n^{3,r}) \le 3n-5-[(n+r) \equiv 0\bmod 3] \,. \]
\end{theorem}
To~prove Knuth's conjecture, one also needs to~prove a~lower bound on $\size(\MOD_n^{3,r})$. The currently strongest known lower bound for $\size(\MOD_n^{3,r})$
is~$2.5n-O(1)$ due to~Stockmeyer~\cite{DBLP:journals/mst/Stockmeyer77}
(and no~stronger lower bound is known for any other symmetric function).
\begin{proof}
As~in~\cite{DBLP:conf/sat/KojevnikovKY09},
we~construct the required circuit out of~constant size blocks. Schematically, the circuit looks as~follows.
\begin{center}
\begin{tikzpicture}[scale=.8]
\foreach \x/\k in {0.5/1, 2.5/k, 4.5/{k+1}, 5.5/{k+2}, 6.5/{k+3}, 13.5/{n-l+1}, 15.5/n}
\node[above] at (\x,2.5) {$x_{\k}$};
\foreach \x in {0, 4, 9, 13}
\draw (\x,0) rectangle (\x+3,2);
\foreach \x in {0.5, 2.5, 4.5, 5.5, 6.5, 9.5, 10.5, 11.5, 13.5, 15.5}
\draw[->] (\x,2.5) -- (\x,2);
\foreach \x/\y in {1.5/2.25, 14.5/2.25, 8/1}
\node at (\x,\y) {$\dotsb$};
\foreach \a/\b in {3/4, 12/13, 7/7.5, 8.5/9} {
\draw[->] (\a,1.5) -- (\b,1.5);
\draw[->] (\a,0.5) -- (\b,0.5);
}
\draw[->] (16,1) -- (16.5,1);
\node at (1.5, 1) {$\IN_k$};
\node at (5.5, 1) {$\MID_3$};
\node at (10.5, 1) {$\MID_3$};
\node at (14.5, 1) {$\OUT_l^r$};
\end{tikzpicture}
\end{center}
Here, the $n$~input bits are passed from above.
What is passed from block to block (from left to~right)
is the pair of bits $(r_0, r_1)$ encoding the current remainder~$r$ modulo~$3$ as~follows: if $r=0$, then $(r_0,r_1)=(0,0)$; if $r=1$, then $(r_0,r_1)=(0,1)$; if $r=2$, then $r_0=1$. The first block $\IN_k$ takes the first $k$~input bits and computes the remainder of their sum modulo~$3$. It is followed by a~number of~$\MID_3$ blocks each of which takes the current remainder and three new input bits and computes the new remainder. Finally, the block~$\OUT_l^r$ takes the remainder and the last~$l$ input bits and outputs $\MOD_n^{3,r}$. The integers $k,l$ take values in~$\{2,3,4\}$ and $\{1,2,3\}$, respectively. Their exact values depend on~$r$ and $n \bmod 3$ as~described below.
The theorem follows from the following upper bounds
on~the circuit size of the just introduced functions:
$\size(\IN_2) \le 2$,
$\size(\IN_3) \le 5$,
$\size(\IN_4) \le 7$,
$\size(\MID_3) \le 9$,
$\size(\OUT_2^0) \le 5$,
$\size(\OUT_1^1) \le 2$,
$\size(\OUT_3^2) \le 8$.
The corresponding circuits are presented in the Appendix
by a~straightforward \texttt{Python} code that verifies their correctness. (The presented code proves the mentioned upper bounds by~providing explicit circuits. We~have also verified that no~smaller circuits exist meaning that the
inequalities above are in fact equalities.)
Table~\ref{table:parameters} shows how to~combine
the blocks to~get a~circuit computing~$\MOD_n^{3,r}$ of the required size.
\begin{table}[!ht]
\begin{center}
\begin{tabular}{cccc}
\toprule
& $n=3t$ & $n=3t+1$ & $n=3t+2$\\
\midrule
$r=0$
& $(4, t-2, 2)$, $(7, 5)$, $3n-6$
& $(2, t-1, 2)$, $(2, 5)$, $3n-5$
& $(3, t-1, 2)$, $(5, 5)$, $3n-5$
\\
$r=1$
& $(2, t-1, 1)$, $(2, 2)$, $3n-5$
& $(3, t-1, 1)$, $(5, 2)$, $3n-5$
& $(4, t-1, 1)$, $(7, 2)$, $3n-6$
\\
$r=2$
& $(3, t-2, 3)$, $(5, 8)$, $3n-5$
& $(4, t-2, 3)$, $(7, 8)$, $3n-6$
& $(2, t-1, 3)$, $(2, 8)$, $3n-5$
\\
\bottomrule
\end{tabular}
\end{center}
\caption{Choosing parameters $k, m, l$ depending on $n \bmod 3$ and~$r$. The circuit is composed out of~blocks
as~follows: $\IN_k + m \times \MID_3 + \OUT_r^l$. For each
pair $(n \bmod 3, r)$ we show three things: the triple
$(k, m, l)$; the sizes of~two blocks: $\size(\IN_k)$ and $\size(\OUT_r^l)$; the size of~the resulting circuit
computed as $s=\size(\IN_k)+9m+\size(\OUT_r^l)$. For example, the top left cell is read as follows: when $r=0$ and $n=3t$, we set $k=4,m=t-2,l=2$; the resulting circuit
is~then $\IN_4 + (t-2) \times \MID_3 + \OUT_0^2$; since $\size(\IN_4)=7$ and $\size(\OUT_0^2)=5$, the size of~the circuit is $7+9(t-2)+5=9t-6=3n-6$.
There are three corners cases that are not well-defined as~they require the
number of~$\MID$ blocks to~be negative ($k=t-2$): $(n=3,r=0)$, $(n=3,r=2)$, and $(n=4, r=2)$. The corresponding circuits are
given in the~Appendix.}
\label{table:parameters}
\end{table}
\end{proof}
\subsection{Threshold Function}
\begin{theorem}
For any $4 \le k = O(1)$,
\[
\size(\THR_n^k) \le (4.5-2^{2 - \lceil \log_2k\rceil})n+o(n) \, .
\]
\end{theorem}
\begin{proof}
For a~sequence of~$2m$ formal variables $y_1, z_1, \dotsc, y_m, z_m$,
consider a~function $g \in B_{2m}$ that takes
\[y_1, y_1 \oplus z_1, y_2, y_2 \oplus z_2, \dotsc, y_m, y_m \oplus z_m\]
as~input and outputs $\THR_{2m}^2(y_1, z_1, \dotsc, y_m, z_m)$.
Note that $\THR_{2m}^2(y_1, z_1, \dotsc, y_m, z_m)=1$ iff there is
a~pair containing two~$1$'s or there are two pairs each containing at~least one~$1$: $\THR_{2m}^2(y_1, z_1, \dotsc, y_m, z_m)=1$ iff there exists
$1 \le i \le m$ such that $y_i=z_i=1$ or $\THR_m^2(y_1 \oplus z_1, \dotsc, y_m \oplus z_m)=1$. The condition $y_i=z_i=1$ can be~computed
through $y_i$ and $y_i \oplus z_i$ using a~single binary gate:
\[(y_i \land z_i)=(y_i \land \overline{(y_i \oplus z_i)}) \, .\]
Thus,
\[g(y_1, y_1 \oplus z_1,\dotsc, y_m, y_m \oplus z_m)=\THR_m^2(y_1 \oplus z_1, \dotsc, y_m \oplus z_m) \lor \bigvee_{i=1}^{m}(y_i \land \overline{(y_i \oplus z_i)}) \, .\]
Now, $\size(\THR_m^2) \le 2m+o(m)$ as~shown by~Dunne~\cite{Dunne84}.
Also, clearly,
\[\size\left(\bigvee_{i=1}^{m}(y_i \land \overline{(y_i \oplus z_i)}) \right)\le 2m-1 \, .\]
Thus,
\begin{equation}\label{eq:g}
\size(g) \le 4m+o(m) \, .
\end{equation}
First, consider the case $k=2^t$ where $t \ge 2$ is an~integer.
To~construct a~circuit for $\THR_n^k$, we~first apply $t-1$
layers of~$\MDFA$'s (as~in Figure~\ref{figure:sum17}). It~takes
\[\frac{n}{2} + n\sum\limits_{i=1}^{t-1} 2^{2-i}=(4.5 - 2^{3-t})n\] gates.
As a~result, we get bits \[w_0, \dotsc, w_{t-2}, a_1, a_1 \oplus b_1, \dotsc, a_m, a_m \oplus b_m \,, \]
where $m=n/2^t$,
such that
\[x_1+\dotsb+x_n=w_0+2w_1+\dotsb+2^{t-2}w_{t-2}+2^{t-1}(a_1+b_1+\dotsb+a_m+b_m) \, .\]
Note that $w_0+2w_1+\dotsb+2^{t-2}w_{t-2} < 2^{t-1}$. Hence,
\[[x_1+\dotsb+x_n \ge 2^t]=[a_1+b_1+\dotsb+a_m+b_m \ge 2]\,.\]
Thus, it~remains to~compute the function~$g$
given $2m$~bits $a_1, a_1 \oplus b_1, \dotsc, a_m, a_m \oplus b_m$.
By~\eqref{eq:g}, it takes $4m+o(m)=2^{2-t}n+o(n)$ gates.
The total size of the constructed circuit~is
\[(4.5-2^{3-t}+2^{2-t})n+o(n)=(4.5-2^{2-t})n+o(n) \, .\]
Now, assume that $2^{t-1}<k<2^t$. Then, $\lceil \log_2k \rceil = t$
and $2^t<2k=O(1)$. Clearly,
\[[x_1+\dotsb+x_n \ge k]=[(2^t-k)+x_1+\dotsb+x_n \ge 2^t] \, .\]
By~the previous argument, there exists a~circuit~$C$ computing $\THR_{n+(2^t-k)}^{2^t}$ of~size
\[(4.5-2^{2-t})(n+(2^t-k))+o(n)=(4.5-2^{2-t})n+o(n) \, .\]
By~replacing arbitrary $(2^t-k)$ inputs of~$C$ by~$1$'s,
one gets a~circuit of~the required size computing~$\THR_n^k$.
\end{proof}
We~conclude by~presenting an~example of a~reasonably small circuit
that our program fails to~improve though a~better circuit
is~known.
Figures~\ref{figure:thr31} and~\ref{figure:thr29} show circuits of~size~$31$ and~$29$ for $\THR_{12}^2$.
They are
quite different and our program is~not able to~find out
that the circuit
of~size~$31$ is~suboptimal.
One can construct the two circuits in~the program as~follows.
\inputminted[firstline=90,lastline=96]{python}{../tutorial.py}
\begin{figure}[t]
\begin{mypic}
\begin{scope}[scale=.6]
\begin{scope}[xshift=20mm, yshift=30mm]
\foreach \n/\x/\y in {1/0/11, 2/1/12, 3/2/12, 4/3/12, 5/4/12, 6/5/12, 7/6/12, 8/7/12, 9/8/12, 10/9/12, 11/10/12, 12/11/12}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\foreach \n/\x/\y in {2/1/11, 4/2/11, 6/3/11, 8/4/11, 10/5/11, 12/6/11, 14/7/11, 16/8/11, 18/9/11, 20/10/11}
\node[gate] (g\n) at (\x, \y) {$\lor$};
\foreach \n/\x/\y in {1/0/10, 3/1/10, 5/2/10, 7/3/10, 9/4/10, 11/5/10, 13/6/10, 15/7/10, 17/8/10, 19/9/10}
\node[gate] (g\n) at (\x, \y) {$\land$};
\foreach \n/\x/\y in {21/0.3/9, 22/1.3/9, 23/2.3/9, 24/3.3/9, 25/4.3/9, 26/5.3/9, 27/6.3/9, 28/7.3/9, 29/8.3/9}
\node[gate] (g\n) at (\x, \y) {$\lor$};
\node[gate] (g30) at (10,10) {$\land$};
\node[gate, label=right:output] (g31) at (9.3,9) {$\lor$};
\foreach \f/\t in {x1/g1, x1/g2, x2/g1, x2/g2,
g2/g3, g2/g4, x3/g3, x3/g4,
g4/g5, g4/g6, x4/g5, x4/g6,
g6/g7, g6/g8, x5/g7, x5/g8,
g8/g9, g8/g10, x6/g9, x6/g10,
g10/g11, g10/g12, x7/g11, x7/g12,
g12/g13, g12/g14, x8/g13, x8/g14,
g14/g15, g14/g16, x9/g15, x9/g16,
g16/g17, g16/g18, x10/g17, x10/g18,
g18/g19, g18/g20, x11/g19, x11/g20,
g1/g21, g3/g21, g21/g22, g5/g22,
g22/g23, g7/g23, g23/g24, g9/g24,
g24/g25, g11/g25, g25/g26, g13/g26,
g26/g27, g15/g27, g27/g28, g17/g28,
g28/g29, g19/g29,
x12/g30, g20/g30, g29/g31, g30/g31}
\draw[->] (\f) -- (\t);
\end{scope}
\end{scope}
\begin{scope}[scale=.6,xshift=-90mm,yshift=100mm]
\node[input] (x1) at (0, 3.5) {$x_{1}$};
\foreach \n/\x/\y in {2/1.5/5, 3/3/5, 4/4.5/5, 11/7.5/5, 12/9/5}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node at (6,3.5) {$\dotsb$};
\foreach \x/\y/\z in {1/2/1.5, 2.5/3.5/3, 4/5/4.5, 7/8/7.5} {
\draw (\x,2.8) rectangle (\y,4.2);
\node[rotate=90] at (\z,3.5) {sort};
}
\node[gate] (g1) at (9,3.5) {$\land$};
\node[gate] (g2) at (7.5,2) {$\lor$};
\draw[->] (x1) -- (1,3.5);
\draw[->] (x2) -- (1.5,4.2);
\draw[->] (x3) -- (3,4.2);
\draw[->] (x4) -- (4.5,4.2);
\draw[->] (x11) -- (7.5,4.2);
\draw[->] (x12) -- (g1);
\draw[->] (2,3.5) -- (2.5,3.5);
\draw[->] (3.5,3.5) -- (4,3.5);
\draw[->] (5,3.5) -- (5.5,3.5);
\draw[->] (6.5,3.5) -- (7,3.5);
\draw[->] (8,3.5) -- (g1);
\draw[->] (1.5,2.8) -- (g2);
\draw[->] (3,2.8) -- (g2);
\draw[->] (4.5,2.8) -- (g2);
\draw[->] (7.5,2.8) -- (g2);
\draw[->] (g1) -- (g2);
\foreach \f/\t in {}
\draw[->] (\f) -- (\t);
\end{scope}
\end{mypic}
\caption{A~circuit of~size~$31$ for~$\THR_{12}^2$:
(a)~block structure and (b)~gate structure.
The $\operatorname{SORT}(u,v)$ block sorts two input bits
as~follows: $\operatorname{SORT}(u,v)=(\min\{u,v\}, \max\{u,v\})=(u \land v, u \lor v)$. The circuit performs
one and a~half iterations of~the bubble sort algorithm:
one first finds the maximum bit among $n$~input bits; then,
it~remains to~compute the disjunction of the remaining $n-1$ bits to~check whether there~is at~least one~$1$ among them.
In~general, this leads to~a~circuit of~size $3n-5$.}
\label{figure:thr31}
\end{figure}
\begin{figure}[t]
\begin{mypic}
\begin{scope}[xscale=.75,yscale=.7]
\begin{scope}[xshift=20mm, yshift=40mm]
\node[input] (x5) at (2, 7) {$x_5$};
\node[input] (x1) at (4, 7) {$x_1$};
\node[input] (x6) at (6, 7) {$x_6$};
\node[input] (x2) at (8, 7) {$x_2$};
\node[gate] (g9) at (1.5,6) {$\lor$};
\node[input] (x9) at (2.5, 6) {$x_9$};
\node[input] (x10) at (3.5, 6) {$x_{10}$};
\node[gate] (g3) at (4.5,6) {$\lor$};
\node[gate] (g11) at (5.5,6) {$\lor$};
\node[input] (x7) at (6.5, 6) {$x_7$};
\node[input] (x3) at (7.5, 6) {$x_3$};
\node[gate] (g0) at (8.5,6) {$\lor$};
\node[gate, label={[shift={(-0.4,-0.4)}]$c_1$}] (g10) at (1,5) {$\lor$};
\node[gate] (g6) at (2,5) {$\lor$};
\node[gate, label={[shift={(0.4,-0.74)}]$c_2$}] (g12) at (3,5) {$\lor$};
\node[input] (x11) at (4, 5) {$x_{11}$};
\node[gate] (g4) at (5,5) {$\lor$};
\node[gate] (g13) at (6,5) {$\lor$};
\node[input] (x8) at (7, 5) {$x_8$};
\node[input] (x4) at (8, 5) {$x_4$};
\node[gate] (g1) at (9,5) {$\lor$};
\node[gate] (g22) at (1.5,4) {$\land$};
\node[gate] (g21) at (2.5,4) {$\lor$};
\node[gate] (g7) at (3.5,4) {$\lor$};
\node[gate, label={[shift={(0.08,-0.18)}]$c_3$}] (g14) at (4.5,4) {$\lor$};
\node[input] (x12) at (5.5, 4) {$x_{12}$};
\node[gate, label={[shift={(0.45,-0.50)}]$r_2$}] (g5) at (6.5,4) {$\lor$};
\node[gate] (g15) at (8,4) {$\lor$};
\node[gate, label=right:$r_1$] (g2) at (9,4) {$\lor$};
\node[gate] (g24) at (2,3) {$\land$};
\node[gate] (g23) at (3,2.6) {$\lor$};
\node[gate, label={[shift={(-0.45,-0.5)}]$r_3$}] (g8) at (4.5,3) {$\lor$};
\node[gate, label={[shift={(0.28,-0.2)}]$c_4$}] (g16) at (5.5,3) {$\lor$};
\node[gate] (g17) at (7,3) {$\lor$};
\node[gate] (g18) at (8,3) {$\land$};
\node[gate] (g25) at (1.5,2) {$\lor$};
\node[gate] (g26) at (4,2) {$\land$};
\node[gate] (g19) at (5,2) {$\land$};
\node[gate] (g27) at (4,1) {$\lor$};
\node[gate] (g20) at (5,1) {$\lor$};
\node[gate, label=right:output] (g28) at (4.5,0) {$\lor$};
\foreach \f/\t in {x5/g9, x1/g9, x5/g3, x6/g3, x6/g11, x2/g11, x2/g0, x1/g0,
g9/g10, x9/g10, x9/g6, x10/g6, x10/g12, g11/g12, g3/g4, x7/g4, x7/g13, x3/g13, x3/g1, g0/g1,
g10/g22, g10/g21, g6/g7, g12/g22, g12/g21, x11/g7, x11/g14, g4/g5, g13/g14, x8/g5, x8/g15, x4/g15, x4/g2, g1/g2,
g22/g25, g21/g24, g21/g23, g7/g8, g14/g24, g14/g23, x12/g8, x12/g16, g5/g17, g5/g18, g15/g16, g2/g17, g2/g18,
g24/g25, g23/g26, g8/g19, g16/g26, g17/g19, g18/g20,
g25/g27, g26/g27, g19/g20,
g27/g28, g20/g28}
\draw[->] (\f) -- (\t);
\end{scope}
\end{scope}
\begin{scope}[scale=.8,xshift=-60mm,yshift=25mm]
\draw (0.5,4.5) rectangle (4.5,7.5); \node at (0,0) {};
\foreach \n/\x/\y in {1/1/7, 2/2/7, 3/3/7, 4/4/7, 5/1/6, 6/2/6, 7/3/6, 8/4/6, 9/1/5, 10/2/5, 11/3/5, 12/4/5}
\node[input] (x\n) at (\x,\y) {$x_{\n}$};
\foreach \n/\x/\y in {1/5.5/7, 2/5.5/6, 3/5.5/5}
\node[gate, label={[shift={(0.3,-0.2)}]$r_\n$}] (r\n) at (\x, \y) {$\lor$};
\draw[->] (4.5,7)--(r1);
\draw[->] (4.5,6)--(r2);
\draw[->] (4.5,5)--(r3);
\foreach \n/\x/\y in {1/1/3.5, 2/2/3.5, 3/3/3.5, 4/4/3.5}
\node[gate, label={[shift={(0.3,-0.2)}]$c_\n$}] (c\n) at (\x, \y) {$\lor$};
\draw[->] (1,4.5)--(c1);
\draw[->] (2,4.5)--(c2);
\draw[->] (3,4.5)--(c3);
\draw[->] (4,4.5)--(c4);
\draw (0.5,1.5) rectangle (4.5,2.5);
\node at (2.5,2) {$\THR_2^4$};
\draw[->] (c1) -- (1,2.5);
\draw[->] (c2) -- (2,2.5);
\draw[->] (c3) -- (3,2.5);
\draw[->] (c4) -- (4,2.5);
\draw (6.5,4.5) rectangle (7.5,7.5);
\node[rotate=90] at (7,6) {$\THR_2^3$};
\draw[->] (r1) -- (6.5,7);
\draw[->] (r2) -- (6.5,6);
\draw[->] (r3) -- (6.5,5);
\node[gate] (g) at (7, 2) {$\lor$};
\draw[->] (4.5,2)--(g);
\draw[->] (7,4.5)--(g);
\end{scope}
\end{mypic}
\caption{A~circuit of~size~$29$ for $\THR_{12}^2$: (a)~block structure and (b)~gate structure. It~implements a~clever trick by~Dunne~\cite{Dunne84}. Organize $12$~input bits
into a~$3 \times 4$ table. Compute disjunctions $r_1,r_2,r_3$ of~the rows and disjunctions $c_1,c_2,c_3,c_4$ of~the columns. Then, there are at least two~$1$'s among
$x_1, \dotsc, x_{12}$ if~and only if~there are at~least
two~$1$'s among either $r_1,r_2,r_3$ or $c_1,c_2,c_3,c_4$.
This allows to~proceed recursively. In~general, it~leads
to~a~circuit of~size $2n+o(n)$. (Sergeev~\cite{Sergeev2020} showed recently that the \emph{monotone} circuit size of $\THR_n^2$ is $2n+\Theta(\sqrt n)$.)}
\label{figure:thr29}
\end{figure}
\section{Further Directions}
In~the paper, we~focus mainly on~proving asymptotic upper bounds for function families (that is, upper bounds that hold for every input size).
A~natural further step is~to~apply the program
to~specific circuits that are used in~practice.
It~would also be~interesting to~extend the program
so~that it~is able to~discover the circuit from Figure~\ref{figure:thr29}.
\section{Overview}
Let $l=\lceil \log_2(n+1)\rceil$. The function $\SUM_n \colon \{0,1\}^n \to \{0,1\}^{l}$ computes the binary representation of~the sum of~$n$~bits:
$\SUM_n(x_1, \dotsc, x_n)=(w_0, w_1, \dotsc, w_{l-1})$ such that
\[\sum_{i=1}^{n}x_i=\sum_{i=0}^{l-1}2^iw_i \, .\]
This way, the function transform $n$~bits of weight~0 into $l$~bits
of~weights $(0,1,\dotsc,l-1)$.
This is~a~fundamental symmetric function. In~particular, for any symmetric predicate~$f \colon \{0,1\}^n \to \{0,1\}$,
\[\size(f) \le \size(\SUM_n)+o(n) \, ,\]
where $\size(\cdot)$ is~the circuit size (in~this text, our default computational model is Boolean circuits over the full binary basis).
\section{Circuit Size of $\SUM_n$
\subsection{Three Inputs}
The exact circuit size of $\SUM_n$ is known for all~$n$ up to~$n=5$. The table below shows upper bounds for some other values of~$n$:
\begin{center}
\begin{tabular}{lccccccccccc}
\toprule
$n$ & $2$ & $3$ & $4$ & $5$ & $6$ & $7$ & $8$ & $9$ & $10$ & $15$ & $31$\\
\midrule
$\size(\SUM_n)$ & $2$ & $5$ & $9$ & $11$ & $16$ & $19$ & $25$ & $ 27$ & $33$ & $53$ & $125$\\
\bottomrule
\end{tabular}
\end{center}
Below, we show optimal circuits for $n=2,3$. They are known as half adder and full adder, respectively.
\begin{mypic}
\begin{scope}[yscale=.8]
\begin{scope}[yshift=2cm]
\foreach \n/\x/\y in {1/0/1, 2/1/1}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate, label=left:$w_1$] (g1) at (0,0) {$\land$};
\node[gate, label=right:$w_0$] (g2) at (1,0) {$\oplus$};
\foreach \f/\t in {x1/g1, x1/g2, x2/g1, x2/g2}
\draw[->] (\f) -- (\t);
\end{scope}
\begin{scope}[xshift=4cm]
\foreach \n/\x/\y in {1/0/3, 2/1/3, 3/2/3}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate] (g1) at (0.5,2) {$\oplus$};
\node[gate] (g2) at (1.5,2) {$\oplus$};
\node[gate] (g3) at (0.5,1) {$\lor$};
\node[gate, label=right:$w_0$] (g4) at (1.5,1) {$\oplus$};
\node[gate, label=right:$w_1$] (g5) at (0.5,0) {$\oplus$};
\foreach \f/\t in {x1/g1, x2/g1, x2/g2, x3/g2, g1/g3, g2/g3, g1/g4, g3/g5, g4/g5}
\draw[->] (\f) -- (\t);
\path (x3) edge[bend left,->] (g4);
\end{scope}
\end{scope}
\end{mypic}
\subsection{From Three Inputs to~the General Case}
These two basic blocks can already be~used to~compute $\SUM_n$ for any~$n$.
E.g., one can compute $\SUM_5$ by a~circuit of~size~12 as~follows:
\begin{mypic}
\begin{scope}[yscale=.7]
\foreach \n in {1,...,5}
\node[input] (\n) at (\n,6) {$x_{\n}$};
\draw (0.5, 5.5) rectangle (3.5, 4.5); \node at (2, 5) {$\SUM_3$};
\foreach \n in {1, 2, 3}
\draw[->] (\n) -- (\n, 5.5);
\draw (2.5, 3.5) rectangle (5.5, 2.5); \node at (4, 3) {$\SUM_3$};
\path (3, 4.5) edge[->] node[l] {0} (3, 3.5);
\foreach \n in {4, 5}
\draw[->] (\n) -- (\n, 3.5);
\draw (1.5, 1.5) rectangle (3.5, 0.5); \node at (2.5, 1) {$\SUM_2$};
\path (2, 4.5) edge[->] node[l] {1} (2, 1.5);
\path (3, 2.5) edge[->] node[l] {1} (3, 1.5);
\node[input] (w2) at (2,-1) {$w_2$};
\node[input] (w1) at (3,-1) {$w_1$};
\node[input] (w0) at (4.5,-1) {$w_0$};
\path (2, 0.5) edge[->] node[l] {1} (w2);
\path (3, 0.5) edge[->] node[l] {0} (w1);
\path (4.5, 2.5) edge[->] node[l] {0} (w0);
\end{scope}
\end{mypic}
The same trick shows that $\size(\SUM_n) \le 5n+o(n)$. Indeed,
the $\SUM_3$ circuit replaces three bits of weight~$i$ by two bits of weight~$i$ and $i+1$. Thus, to~get from $n$~bits of weight~$0$ to $l$~bits of weight $(0,\dotsc, l-1)$, one may
repeatedly apply $\SUM_3$ to three bits of the same weight.
An~application of~$\SUM_3$ reduces the number of~bits by~one, hence
the number of such applications is~at~most~$n$. What will remain after this step is at most two bits on each of $l=o(n)$~levels. Pictorially, this looks as~follows.
\begin{mypic}
\foreach \n/\t/\x in {{x1/x_1/0.5}, {x2/x_2/1}, {x3/x_3/2}, {x4/x_4/2.5}, {x5/x_5/3.5}, {x6/x_6/4}, {x7/x_7/5}, {x8/x_8/5.5}, {x10/x_n/8.5}}
{
\node[] (\n) at (\x,6) {$\t$};
\draw[->] (\n) -- (\x,5.25);
}
\foreach \n/\t/\y in {{y1/w_0/4.75}, {y2/w_1/3}, {y3/w_{l-1}/0.5}}
\node[] (\n) at (9.5,\y) {$\t$};
\foreach \l/\r/\x/\y in {{0.25/5.25/1.25/4.25}, {1.75/5.25/2.75/4.25}, {3.25/5.25/4.25/4.25}, {4.75/5.25/5.75/4.25}, {7.75/5.25/8.75/4.25}, {1/3.5/2/2.5}, {4/2.5/5/3.5}, {4/0/5/1}}
\draw (\l,\r) rectangle (\x,\y);
\foreach \x/\y in {{0.75/4.75}, {2.25/4.75}, {3.75/4.75}, {5.25/4.75}, {8.25/4.75}, {1.5/3}, {4.5/3}, {4.5/0.5}}
\node[] at (\x,\y) {\small $\SUM_3$};
\foreach \x/\y/\l/\r in {{1.25/4.75/1.75/4.75}, {2.75/4.75/3.25/4.75}, {4.25/4.75/4.75/4.75}, {5.75/4.75/6.25/4.75}, {2/3/4/3}, {5/3/7/3}, {0.75/4.25/1.25/3.5}, {2.25/4.25/1.75/3.5}, {3.75/4.25/4.25/3.5}, {5.25/4.25/4.75/3.5}, {5.25/1.75/4.75/1.0}, {3.75/1.75/4.25/1.0}, {7.25/4.75/7.75/4.75}}
\draw[->] (\x,\y) -- (\l,\r);
\foreach \x/\y/\t in {{6.75/4.75/\cdots}, {7.74/3/\cdots}, {4.5/2/\cdots}}
\node[] at (\x,\y) {$\t$};
\foreach \x/\y/\v in {{8.75/4.75/y1}, {8.25/3/y2}, {5/0.5/y3}}
\draw[->] (\x,\y) -- (\v);
\end{mypic}
In~general, this gives the following upper bound:
if $\size(\SUM_k) \le r$, then for all~$n$,
\begin{equation}\label{eq:master}
\size(\SUM_n) \le \frac{rn}{k-\lceil \log_2(k+1) \rceil}+o(n) \, .
\end{equation}
\subsection{Five Inputs and Improved General Case}
It~turns out that
there exists a~smaller circuit for $\SUM_5$
and that it can be used in two different ways
to~improve the general upper bound! The following circuit of size~11 was found by Knuth. This is how he describes it: ``But $s(5)=12$ is \emph{not} optimum, despite the beauty of~7.1.2-(29)!'' (Here, 7.1.2-(29) refers to~the circuit of size~12.)
\begin{mypic}
\begin{scope}[yscale=.8, label distance=-1mm]
\foreach \n/\x/\y in {1/0/3, 2/1/3, 3/2/3, 4/2.5/1, 5/3.5/1}
\node[input] (x\n) at (\x, \y) {$x_{\n}$};
\node[gate,label=left:$g_1$] (g1) at (0.5,2) {$\oplus$};
\node[gate,label=left:$g_2$] (g2) at (1.5,2) {$\oplus$};
\node[gate,label=left:$g_3$] (g3) at (0.5,1) {$\lor$};
\node[gate,label=left:$g_4$] (g4) at (1.5,1) {$\oplus$};
\node[gate,label=left:$g_5$] (g5) at (0.5,0) {$\oplus$};
\node[gate,label=left:$g_6$] (g6) at (2,0) {$\oplus$};
\node[gate,label=left:$g_7$] (g7) at (3,0) {$\oplus$};
\node[gate,label=left:$g_8$] (g8) at (2,-1) {$>$};
\node[gate, label=right:$w_0$] (g9) at (3,-1) {$\oplus$};
\node[gate, label=right:$w_1$] (g10) at (2,-2) {$\oplus$};
\node[gate, label=right:$w_2$] (g11) at (1.5,-3) {$>$};
\foreach \f/\t in {x1/g1, x2/g1, x2/g2, x3/g2, g1/g3, g2/g3, g1/g4, g3/g5, g4/g5, x4/g6, g4/g6, x4/g7, x5/g7, g6/g8, g7/g8, g7/g9, g3/g10, g8/g10, g10/g11, g5/g11}
\draw[->] (\f) -- (\t);
\path (x3) edge[->,bend left] (g4);
\path (g4) edge[->,bend left=20] (g9);
\end{scope}
\end{mypic}
Using this circuit, one can build a~circuit of size $5+11+2+1=19$ for $\SUM_7$:
\begin{mypic}
\begin{scope}[yscale=.7]
\foreach \n in {1,...,7}
\node[input] (\n) at (\n,6) {$x_{\n}$};
\draw (0.5, 5.5) rectangle (3.5, 4.5); \node at (2, 5) {$\SUM_3$};
\foreach \n in {1, 2, 3}
\draw[->] (\n) -- (\n, 5.5);
\draw (2.5, 3.5) rectangle (7.5, 2.5); \node at (5, 3) {$\SUM_5$};
\path (3, 4.5) edge[->] node[l] {0} (3, 3.5);
\foreach \n in {4, 5, 6, 7}
\draw[->] (\n) -- (\n, 3.5);
\draw (1.5, 1.5) rectangle (3.5, 0.5); \node at (2.5, 1) {$\SUM_2$};
\path (2, 4.5) edge[->] node[l] {1} (2, 1.5);
\path (3, 2.5) edge[->] node[l] {1} (3, 1.5);
\draw (2.5, -0.5) rectangle (4.5, -1.5); \node at (3.5, -1) {XOR};
\path (4, 2.5) edge[->] node[l] {2} (4, -0.5);
\path (3, 0.5) edge[->] node[l] {1} (3, -0.5);
\node[input] (w0) at (6,-3) {$w_0$};
\node[input] (w2) at (3.5,-3) {$w_2$};
\node[input] (w1) at (2,-3) {$w_1$};
\path (6, 2.5) edge[->] node[l] {0} (w0);
\path (2, 0.5) edge[->] node[l] {0} (w1);
\path (3.5, -1.5) edge[->] (w2);
\end{scope}
\end{mypic}
Plugging $\size(\SUM_7) \le 19$ into~\eqref{eq:master}, gives an~upper bound $\size(\SUM_n) \le 4.75n$.
Interestingly, the same $\SUM_5$ circuit can be used to~get a~$4.5n$ upper bound!
For this,
Using this block, one can compute $\SUM_n$ by a~circuit of size $4.5n$ as~follows:
\begin{mypic}
\foreach \n/\t/\x in {{x1/x_1/0.5}, {x2/x_2/1}, {x3/x_3/1.5}, {x4/x_4/2}, {x5/x_5/3}, {x6/x_6/3.5}, {x7/x_7/4}, {x8/x_8/4.5}, {x9/x_{n-1}/7.75}, {x10/x_n/8.5}}
\node[] (\n) at (\x,6.5) {$\t$};
\foreach \n/\x/\a/\b in {{xor1/1/x1/x2}, {xor2/2/x3/x4}, {xor3/3.5/x5/x6}, {xor4/4.5/x7/x8}, {xor5/8.5/x9/x10}}
{
\node[gate] (\n) at (\x, 5.75) {$\oplus$};
\draw[->] (\a) -- (\n);
\draw[->] (\b) -- (\n);
\path[draw,->] (\n) -- (\x,5.25);
}
\foreach \n/\x in {x1/0.5, x3/1.5, x5/3, x7/4, x9/7.75}
\path[draw,->] (\n) -- (\x,5.25);
\foreach \n/\t/\y in {{y1/w_0/4.75}, {y2/w_1/3}, {y3/w_{l-1}/0.5}}
\node[] (\n) at (9.5,\y) {$\t$};
\foreach \l/\r/\x/\y in {{0.25/5.25/2.25/4.25}, {0.25+2.5/5.25/2.25+2.5/4.25}, {0.25+6.5/5.25/2.25+6.5/4.25}, {1.5/2.5/3.5/3.5}, {3.5/0/5.5/1}
\draw (\l,\r) rectangle (\x,\y);
\foreach \x/\y in {{1.25/4.75}, {3.75/4.75}, {7.75/4.75}, {2.5/3}, {4.5/0.5}}
\node[] at (\x,\y) {MDFA};
\foreach \xa/\xb in {0.75/1.75, 1.25/2.25, 3.75/2.75, 4.25/3.25}
\draw[->] (\xa,4.25) -- (\xb,3.5);
\foreach \x/\y/\t in {{5.75/4.75/\cdots}, {6.75/3/\cdots}, {4.5/1.75/\cdots}}
\node[] at (\x,\y) {$\t$};
\foreach \x/\y/\v in {{8.75/4.75/y1}, {8.25/3/y2}, {5.5/0.5/y3}}
\draw[->] (\x,\y) -- (\v);
\foreach \xa/\xb/\y in {2.25/2.75/4.75, 4.75/5.25/4.75, 6.25/6.75/4.75, 3.5/4.5/3}
\draw[->] (\xa,\y) -- (\xb,\y);
\end{mypic}
Finally, the block MDFA is contained in the circuit of size~11 for $\SUM_5$!
\end{document} |
\section{INTRODUCTION}
Automated vehicles rely on an accurate model of their environment for planning safe and efficient behavior. Depending on the chosen representation of the environment in this model, different perception algorithms and different sensor modalities are best suited, each coming with corresponding advantages and disadvantages. Often, several different approaches are combined to compensate for the disadvantages of one perception algorithm with the advantages of another.
One common representation of the dynamic environment are object lists. They contain the state of all objects detected and tracked by the vehicle. Several methods for detection and tracking of objects in camera, radar and lidar data have been published during the past years \cite{Aeberhard.2017, Zhou.2018, Lang.2019b}. A drawback of object lists is that the perception algorithms generating them can often only detect a fixed set of predefined object classes. Objects of these classes need to be explicitly contained in the perception algorithms' training data. Since there exist extremely many classes of objects in the world, it is difficult to ensure that all relevant objects are accounted for.
Occupancy grid mapping algorithms, which are agnostic to the specific class of an object, can compensate for this disadvantage by reducing their task to simply assigning an occupancy state to each cell in a grid, which describes a defined area around a vehicle \cite{Elfes.1989, Thrun.2005}. They usually take distance measurements, e.g. from a lidar sensor, as input. The resulting occupancy grid maps (OGMs) are for example used in the automated vehicles developed in the UNICARagil project \cite{Woopen.2018, Buchholz.2020}, from which this paper also originates.
To determine cell occupancy states from sensor data, an inverse sensor model (ISM) is required. In the past, geometric models have mostly been used \cite{Thrun.2005}. These approaches are often suitable for static and flat environments, but fail in dynamic and non-flat environments. Recent works also propose deep neural networks for this task as so-called deep ISMs \cite{Bauer.2019, Sless.2019}. As usual with supervised learning, the gathering of training data poses a challenge here. Both \cite{Bauer.2019} and \cite{Sless.2019} use cross-modal training. They make use of a lidar-based geometric ISM to generate training data for a radar-based deep ISM. This approach enables the learned model to infer occupancy information which cannot be deduced with the geometrical approach alone. Since the model is trained with data that does not constitute ground-truth data but only an estimation from the lidar-based geometric ISM, the trained models suffer from the restrictions of the lidar-based ISM.
\begin{figure}[!t]
\centering
\includegraphics[width=\linewidth]{img/Real-Prediction-Rviz.png}
\caption{\label{fig_prediction}A deep learning-based inverse sensor model predicts an evidential occupancy grid map (right) from a real-world lidar point cloud (left). Grid cells with a large belief mass for the state \textit{Free} are colored green, those with a large belief mass for the state \textit{Occupied} are colored red. The bigger the belief mass, the larger the respective value in the color channel. Black represents the maximum of epistemic uncertainty.}
\end{figure}
If cross-modal training is discarded, one can make use of manually labeled data for training deep ISMs. This approach is infeasible for many though, because of the associated amount of manual labeling work. Fortunately, simulation software for automated driving is rapidly evolving. Better sensor models and the possibility to automatically generate ground-truth data make this approach viable. With more realistic synthetic data, the reality gap is closing and generalization of neural networks from synthetic to real-world data becomes possible.
In any event, there is a need for the quantification of uncertainty in the estimates of perception algorithms. The capability of neural networks to output a measure of their confidence in a prediction is not yet reflected in many deep learning-based approaches.
Our work takes into account all of the aforementioned challenges and contributes a framework in which they are dealt with.
\section{CONTRIBUTION}
We present an end-to-end learning framework for training deep learning-based inverse lidar sensor models using synthetic data. First, a method for generating training samples consisting of lidar point clouds as input data and evidential ground-truth OGMs as labels is presented. Second, we propose a suitable neural network architecture that extends the popular PointPillars architecture \cite{Lang.2019b} with an evidential prediction head (EPH). The EPH is capable of estimating an evidential OGM. Finally, we evaluate the performance of our approach both on synthetic data and on real-world data. We show that the trained model is able to generalize to different synthetic environments and also produces promising results on real-world data.
\section{BACKGROUND}
In the following, the concept of OGMs and an overview of current approaches for computing OGMs through inverse sensor models (ISMs) is presented. A geometric ISM will be the baseline for the evaluation of the deep ISM proposed in this work. Additionally, a brief introduction into evidence theory and subjective logic is given as this is the basis for the loss function that is used with the proposed deep neural network.
\subsection{Occupancy Grid Maps}\label{ogm}
OGMs as introduced in \cite{Elfes.1989} divide the vehicle's environment into discrete cells containing occupancy information. The occupancy state of each cell at time $k$ can be represented as a binomial random variable $ o_k \in \{ O,F \} $ ($O$: \textit{Occupied}; $F$: \textit{Free}). The probability of each cell being occupied at time $k$ can be derived from the current measurement $z_{k}$ using an inverse sensor model (ISM) $p_{z_{k}}(o_{k}|z_{k})$. Often, a binary Bayes filter is used to create an OGM from a set of distance measurements \cite{Thrun.2005}. This approach has some deficiencies. First, the binary Bayes filter relies on the Markov assumption that there is no temporal correlation of the occupancy state, and that the state does not change in time, which is not adequate for a dynamically changing environment. Additionally, by representing the occupancy as a probability value $p_k(o_k)$, it cannot be distinguished between cells which are uncertain because they cannot be observed, e.g. due to occlusions, and because of conflicting evidence as both cases are described with $p_k(o_k) \approx 0.5$. This challenge can be tackled with evidence theory as described in Section \ref{evidence_theory}.
\subsection{Inverse Sensor Models}
While a sensor model describes how the environment is represented in sensor data, an inverse sensor model (ISM) is required to reconstruct information about the environment from sensor data, as done in e.g. grid mapping.
\textbf{Geometric ISMs} process distance measurements and calculate a probability distribution for the occupancy of cells at the location of the measurement as well as for cells in between the sensor and the measurement. These models consider the sensor's inaccuracy, which is usually hand-crafted, but can also be learned \cite{Thrun.2005}. Hand-designed methods assume a ground model to separate ground from obstacles in the measurement. By filtering techniques, such as Bayesian filtering, the information gathered from each data point in the measurement can be combined into one measurement OGM. By combining the latter with a previous estimate of the OGM, a new estimate with reduced uncertainty can be found \cite{Thrun.2005}.
\textbf{Deep ISMs} are deep learning-based inverse sensor models, which take distance measurements as input for a deep neural network, which is used to predict an OGM. For this purpose, tensor-like representations for the sensor data as input and the OGM as output must be found such that they can be processed.
In \cite{Sless.2019}, a deep radar ISM is introduced. Multiple radar measurements are combined into one bird's-eye-view image serving as input for a semantic segmentation task with classes \textit{Occupied}, \textit{Free} and \textit{Unobserved}. Training data is created from OGMs calculated from lidar measurements using a geometric ISM.
The model presented in \cite{Dequaire.2018} predicts future OGMs from lidar measurements. It is trained using unsupervised learning by creating training samples of OGMs with previous measurements. They use a naive model to create the ground-truth OGM by solely treating all reflection points in a height of $0.6$ to $1.5$ meters as obstacles. This only works in flat environments and with small pitch and roll angles of the ego-vehicle. They encode the lidar measurements into two matrices covering the vehicle's environment with binary values describing the visibility and the occupancy of each cell. The OGM is encoded as a binary matrix which does not sufficiently allow describing uncertainty.
In \cite{Bauer.2019}, radar data is transformed into a two-channel bird's-eye-view image with one channel containing static and the other containing dynamic detections. The OGM is represented as a three-channel image containing belief masses (cf. Section \ref{evidence_theory}) for the states \textit{Occupied}, \textit{Free} and \textit{Unknown}. The task is treated as image segmentation problem. In \cite{Bauer.2020}, they combine a deep radar ISM with geometrical lidar and radar ISMs to increase the perception field and to reduce the time needed to populate the occupancy grid.
The deep neural network presented in \cite{Wu.2020} predicts a bird's-eye-view image containing occupancy, object class and motion information in one shot. A sequence of lidar point clouds encoded as binary matrices which are stacked along a third dimension are used as input data. The model is trained using labeled data from the nuScenes data set \cite{Caesar.2020}.
A methodology capable of transforming segmented images from vehicle cameras to a semantic OGM is presented in~\cite{Reiher.2020}. They also use synthetic data and try to bridge the reality gap by using segmented images as an intermediate representation between real-world and synthetic sensor data.
\subsection{Evidence Theory} \label{evidence_theory}
\textbf{Evidence Theory} as introduced by Dempster and Shafer (DST) \cite{Shafer.1976} can be understood as a generalization of Bayesian probability theory \cite{Dempster.1968}. It allows the explicit consideration of epistemic uncertainty and has also been used with OGMs \cite{Nuss.2018}. Using DST, belief masses are assigned to all subsets of the frame of discernment $\Theta$. For cells in an evidential OGM, this can consist of all possible and mutually exclusive cell states \textit{Free} ($F$) and \textit{Occupied} ($O$): $\Theta = \{ F,O \}$. The power set $2^\Theta = \{ \emptyset, \{ F \}, \{ O \}, \Theta \}$ contains all possible subsets of $\Theta$ to which belief masses $m$ can be assigned.
\begin{eqnarray}
m: 2^\Theta & \rightarrow & [0,1] \\
m(\emptyset) & = & 0 \\
\sum_{A \in 2^\Theta} m(A) & = & 1
\end{eqnarray}
In this example, $m(O)$ constitutes evidence for a cell being occupied and $m(F)$ for a cell being free. Additionally, a state for which no evidence is available can be addressed with $m(\Theta)$ while the empty set is no possible outcome.
\textbf{Subjective Logic} (SL) is a mathematical framework for reasoning under uncertainty. It explicitly distinguishes between epistemic opinions and aleatoric opinions. A direct bijective mapping between the belief mass distribution $m(A)$ in DST and a subjective opinion, i.e. a belief mass distribution $b_A$ and an uncertainty mass $u$ in SL is given by \cite{Jsang.2016}.
\begin{eqnarray}
b_A & = & m(A) , \qquad A \in \Theta \\
u & = & m(\Theta) \\
\sum_{A \in \Theta} b_A + u & = & 1
\end{eqnarray}
Subjective opinions are equivalent to a Dirichlet probability density function (PDF)
\begin{eqnarray}
\text{Dir}(\boldsymbol{p}, \boldsymbol{\alpha}) & = & \frac{1}{B(\boldsymbol{\alpha})} \prod_{A \in \Theta} p_A^{\alpha_A - 1} \label{eq:dirichlet_pdf}
\end{eqnarray}
with prior probabilities $\boldsymbol{p}$ and parameters $\boldsymbol{\alpha}$
\begin{equation}
\begin{split}
\boldsymbol{p}& = \left\{ p_A \quad | \quad A \in \Theta, \quad \sum_{A \in \Theta} p_A = 1, \quad 0 \leq p_A \leq 1 \right\} \\
\boldsymbol{\alpha}& = \left\{ \alpha_A \quad | \quad A \in \Theta \right\}
\end{split}
\end{equation}
and the multivariate beta function $B$ in terms of the gamma function $\Gamma$
\begin{eqnarray}
B(\boldsymbol{\alpha}) & = & \frac{\prod_{A \in \theta} \Gamma(\alpha_A)}{\Gamma ( \sum_{A \in \Theta} \alpha_A )}.
\end{eqnarray}
Evidence for the singletons in the FOD $e_{A} \geq 1, A \in \Theta$ can be converted to parameters of a Dirichlet PDF and to a subjective opinion $(\boldsymbol{b}, u)$ with the number of classes $K = \left| \Theta \right|$ and the Dirichlet strength $S = \sum_{A \in \Theta} \alpha_{A}$:
\begin{eqnarray}
\alpha_{A} & = & e_{A} + 1 \;, \qquad A \in \Theta \label{eq:evidence_to_alpha} \\
b_A & = & \frac{e_A}{S} \\
u & = & \frac{K}{S}
\end{eqnarray}
The authors of \cite{Sensoy.2018} show that a deep neural network can be trained to predict the parameters $\boldsymbol{\alpha}$ of a Dirichlet PDF to express uncertainty in a classification task on the MNIST data set \cite{LeCun.2010}. They propose three loss functions while the approach using the sum of squares led to the best results and is used in this work. The loss for one grid cell $i$ with network parameters $\boldsymbol{w}$, expected class probabilities $\boldsymbol{\hat{p}_i}$ and the true state $\boldsymbol{y_i} = \{y_{i,A} \in \{0,1\} \; | \; A \in \Theta\}$ with $y_{i,A}$ being one if state $A$ is true and zero if false or unknown is given:
\begin{equation}
\mathcal{L}_i(\boldsymbol{w}) = \left\| \boldsymbol{y_i} - \boldsymbol{\hat{p}_i} \right\|_2^2 \text{Dir}(\boldsymbol{\hat{p}_i}, \boldsymbol{\hat{\alpha}_i}) \label{eq:dir_loss}
\end{equation}
This motivates us to create an evidential deep neural network for the task of occupancy grid mapping. We train a model to predict the parameters of a Dirichlet PDF describing the states of cells in an OGM.
\section{LEARNING FRAMEWORK}
Our learning framework processes lidar point clouds as input data and predicts evidential OGMs. In the following, the network architecture and a simulation-based method for generating and augmenting training data is presented.
\subsection{Network Architecture} \label{network_architecture}
The architecture of our deep neural network is based on the popular PointPillars architecture \cite{Lang.2019b}, which is capable of accurately detecting objects in lidar point clouds, while providing a relatively low execution time. The network is divided into three parts. First, there is the Pillar Feature Net, which encodes the reflection points from the lidar measurement into denser features. Second, there is a 2D CNN backbone, which transforms the features into a high-level representation. Last, there are detection heads estimating bounding boxes and motion states for the measured objects.
\begin{figure*}[!t]
\centerline{
\includegraphics[width=\textwidth]{img/training_data.png}}
\caption{The left image shows the point cloud of a simulated VLP32C lidar sensor where the point color indicates the intensity of reflection. In the middle, the corresponding high-definition (HD) point cloud with 3000 layers and points colored by the material causing the reflection is shown. This is used to create the ground-truth occupancy grid map that is shown in the right picture. Free cells are colored green, occupied cells are colored red and unknown cells are black. In addition to the information gained from the HD point cloud, all cells covered by other traffic participants are marked as occupied. A training sample consists of one point cloud as shown in the left image and a corresponding ground truth OGM as shown in the right image.}
\label{fig_training_data}
\end{figure*}
In this work, the detection heads used in \cite{Lang.2019b} are replaced with a 2D convolutional layer with two output channels and ReLU activation. Each pixel represents a cell $i$ in the predicted OGM where the channels contain evidence $e_{i,A}$ for both singletons in the frame of discernment $\Theta = \{F,O\}$, i.e. the cell being occupied or free. The predicted evidence can be converted into estimated parameters of a Dirichlet PDF $\hat{\alpha}_{i,A}$ with Equation (\ref{eq:evidence_to_alpha}). The expected probability values of a cell being occupied or free can be derived from the Dirichlet PDF as $\hat{p}_{i,A} = \hat{\alpha}_{i,A} / S_i$. Thus, following the simplification presented in \cite{Sensoy.2018}, the loss function from Equation (\ref{eq:dir_loss}) in our application reduces to
\begin{equation}
\begin{split}
\mathcal{L}_i(w)& = \mathbb{E} \left[ y_{i,F}^2 - 2 y_{i,F} p_{i,F} + p_{i,F}^2 \right] \\
& \quad + \mathbb{E} \left[ y_{i,O}^2 - 2 y_{i,O}p_{i,O} + p_{i,O}^2 \right] \\
& = (y_{i,F} - \hat{p}_{i,F})^2 + \frac{\hat{p}_{i,F} (1 - \hat{p}_{i,F})}{S_i + 1} \\
& \quad + (y_{i,O} - \hat{p}_{i,O})^2 + \frac{\hat{p}_{i,O} (1 - \hat{p}_{i,O})}{S_i + 1}.
\end{split}
\end{equation}
We extend the loss function with a Kullback-Leibler divergence term with an impact factor $\lambda_t = \min(1.0, t/10)$ that increases from zero to one with the epoch number $t$ as proposed in \cite{Sensoy.2018}. This regularization term penalizes a divergence of the Dirichlet parameters resulting from conflicting evidence $\boldsymbol{\Tilde{\alpha}_i} = \boldsymbol{y_i}+(1-\boldsymbol{y}) \odot \boldsymbol{\alpha_i}$ from a distribution resulting from no evidence, i.e. $\boldsymbol{\alpha_i} = \boldsymbol{1}$. Hence, it encourages the network to reduce conflicting evidence in its output. This leads to the total loss function for all $N$ cells of the OGM in one sample.
\begin{equation}
\mathcal{L}(w) = \sum_{i=1}^N \mathcal{L}_i(w) + \lambda_t \text{KL}\left[ \text{Dir}(\boldsymbol{p_i}|\boldsymbol{\Tilde{\alpha}_i}) || \text{Dir}(\boldsymbol{p_i} | \boldsymbol{1}) \right]
\end{equation}
with the Kullback-Leibler divergence in terms of the gamma function $\Gamma$ and the digamma function $\psi$.
\begin{equation}
\begin{split}
&\text{KL}\left[ \text{Dir}(\boldsymbol{p_i}|\boldsymbol{\Tilde{\alpha}_i}) || \text{Dir}(\boldsymbol{p_i} | \boldsymbol{1}) \right] = \\
& \quad \text{log}\left( \frac{\Gamma(\Tilde{\alpha}_{iF} + \Tilde{\alpha}_{iO})}
{\Gamma(2) \Gamma(\Tilde{\alpha}_{iF}) \Gamma(\Tilde{\alpha}_{iO})}
\right) \\
& \qquad + (\Tilde{\alpha}_{iF} - 1) \left[ \psi(\Tilde{\alpha}_{iF}) - \psi\left(\Tilde{\alpha}_{iF} + \Tilde{\alpha}_{iO} \right) \right] \\
& \qquad + (\Tilde{\alpha}_{iO} - 1) \left[ \psi(\Tilde{\alpha}_{iO}) - \psi\left(\Tilde{\alpha}_{iF} + \Tilde{\alpha}_{iO}\right) \right]
\end{split}
\end{equation}
This loss function will be used to train the model in the experiments such that it approximates the evidence masses in the grid cells while loss for occupied cells is weighted $100$ times compared to the other cells to compensate for their under-representation.
\subsection{Training Data}\label{training_data}
To create training data from the simulation, we use a method that we have already presented in \cite{Lampe.2020}. The simulation environment \cite{vonNeumannCosel.2009} provides a lidar plugin that supports advanced ray tracing and physically-based rendering, i.e. it considers the physical properties of materials to generate more realistic point clouds. We have modeled the sensor setup of one of our research vehicles, which has a Velodyne VLP32C lidar sensor mounted in the middle of its roof. In addition to this 32-layer lidar sensor, another lidar sensor with 3000 layers, which will be called high-definition (HD) lidar from now on, was added to the simulated vehicle. As the lidar plugin provides information about the material type that caused a reflection, it is possible to derive a dense ground-truth OGM from the simulated HD lidar data. The HD lidar covers the same field of view and is exposed to the same occlusions as the real sensor to ensure that the OGMs in the training data will only contain information which can theoretically be derived from the input point cloud. In addition, we augment the ground-truth OGMs based on a ground-truth object list that is also provided by the simulation and contains information about the position and shape of other traffic participants. As we expect the deep ISM to learn to recognize the shape of e.g. cars and trucks, we mark all cells covered by them as occupied in the ground-truth OGM if a minimum of 50 reflections on the vehicle is present in the input point cloud. Figure \ref{fig_training_data} shows that cells with reflections on obstacles such as cars and trees are marked as occupied in the ground-truth OGM, whereas the ground is marked as free. We decided to mark also sidewalks as free space as precise information on the lane geometry can be obtained from a static map. In the event of a serious obstruction of road traffic, the sidewalk could also be an option to pass by.
\subsection{Augmentation}
During the training, we augment the training data by applying a random rotation to the input point cloud and the label OGM. Both are rotated around a vertical axis at the origin of the lidar sensor. This turned out to be important, as without augmentation, cells on the far left and right side of the OGM are almost always occupied, which makes it difficult for the deep ISM to classify those cells correctly.
\section{EXPERIMENTAL SETUP AND RESEARCH QUESTION}
As explained in Section \ref{training_data}, training data for the deep learning-based inverse lidar model is created using a simulation environment. We have selected a model of an urban area and created ten scenarios, each containing random variations of the dynamic environment. In half of the scenarios, the ego-vehicle takes a clockwise route, in the other half, it takes a counter-clockwise route. The randomly generated pulk traffic contains cars, trucks and motorcycles. The type of each of these objects is also chosen randomly from a larger catalogue of possibilities. Additionally, pedestrians are placed at various locations on the sidewalks. Parked vehicles are randomly put on parking lanes. With each scenario, $1.000$ training samples were created at a sampling rate of one per second. In total, $10.000$ samples were generated for training. For the validation data set, another $1000$ samples were created in the same static, but a different dynamic environment. Another test dataset consists of $100$ samples from a different scenario in another part of the simulated urban area and will be used for evaluation in Section \ref{eval_syn}.
The neural network is trained using the Adam optimizer. The Pillar Feature Net \cite{Lang.2019b} creates $10.000$ pillars with a maximum of $100$ reflection points per pillar. The intensity of the reflection points in the simulated lidar point cloud is normalized such that a distribution similar to one observed in real-world lidar point clouds is achieved. The output grid maps have a length of $81.92$ and a width of $56.32$ meters. A cell's side length is $16$ centimeters, resulting in a $512$ by $352$ cell grid map. The sensor origin is at the center of the grid map. The map dimensions and cell size correspond to the detection area and step size of the Pillar Feature Net. After training for $100$ epochs with a batch size of $5$, a minimum loss of $0.104$ and a Kullback-Leibler divergence of $0.357$ on the validation data was achieved.
In the following, we want to answer the following research questions: How well does a deep convolutional neural network that is based on our presented methodology perform when predicting dense OGMs from lidar measurements? In particular, we want to analyze whether it is capable of capturing the epistemic uncertainty for cells where no reflection point is located. Then, how well does the network perform when presented with real-world sensor data?
\section{RESULTS AND DISCUSSION}
First, we evaluate the performance of the trained model on a synthetic test dataset to analyze how well the trained ISM predicts OGMs in a scenario that was not contained in the training data.
Afterwards, we test our model on real-world sensor data that was recorded with one of our research vehicles.
\begin{figure}[!h]
\center
\includegraphics[width=\linewidth]{img/Naive-vs-Prediction.png}
\caption{\label{fig_naive}The geometric ISM (middle) is only able to determine occupied cells containing reflection points of the measurement (left). The deep learning-based ISM (right) has learned to derive more information, e.g. the whole space occupied by vehicles.}
\end{figure}
\subsection{Evaluation on Synthetic Data} \label{eval_syn}
\input{evaluation_plot}
We compare the OGMs that are created using our proposed deep ISM to OGMs created using a geometric ISM where all cells containing reflection points in a height of $0.5$ to $2.0$ meters above ground are marked as occupied while all cells between the sensor origin and these cells are marked as free. Figure \ref{fig_evaluation_masses} shows the mean belief masses in both OGMs during the test scenario. It is apparent and confirmed by Figure \ref{fig_naive} that the geometric ISM creates a considerably higher proportion of cells with an unknown state, hence, less cells are classified as free or occupied. Figure \ref{fig_evaluation_kld} shows the Kullback-Leibler divergence of both Dirichlet PDFs, the one predicted from the deep ISM $\text{Dir}(\boldsymbol{p}|\boldsymbol{\hat{\alpha}})$ and the one from the geometric ISM $\text{Dir}(\boldsymbol{p}|\boldsymbol{\hat{\alpha}_G})$, from the true PDF $\text{Dir}(\boldsymbol{p}|\boldsymbol{\alpha})$. It is evident that the deep ISM estimates the cell states better than the geometric ISM at any time.
\subsection{Evaluation on Real-World Data}
\begin{figure}[!ht]
\center
\includegraphics[width=0.9\linewidth]{img/Real-Data.png}
\caption{\label{fig_real}Two representative samples showing the predicted OGMs (right) using the deep ISM trained with synthetic data on real-world measurements (left) compared to a geometric ISM (middle).}
\end{figure}
Finally, we want to analyze whether the deep ISM trained with synthetic data can also be successfully applied to real-world data. Thus, we test the model on real lidar measurements from one of our research vehicles that has a similar shape and sensor setup as the simulated vehicle. Figure \ref{fig_real} shows two representative samples of the model's performance on a dataset recorded in an urban environment. It is apparent that the performance cannot keep up with the evaluation on synthetic data. Nevertheless, it creates promising results in an environment that is considerably different from the synthetic environment. In particular, the occupancy states of cells that do not contain a reflection point are mostly rated correctly. A higher real-world performance is expected when the simulation scenarios would take place in a model of the actual static environment of the real-world scenarios. More samples and a higher diversity of the training data are relatively easy to achieve in the simulation and are also expected to further increase performance.
\section{CONCLUSION}
We presented a new methodology to train neural networks for the task of occupancy grid mapping using lidar point clouds. The trained network performs considerably better than a classical approach when presented with synthetic data. It also shows advantages when presented with real-world data, even though many options to increase the network's generalization capabilities remain. In contrast to many other deep learning-based approaches, our network is trained such that it can give an estimate of its first- and second-order uncertainty. Our methodology is especially promising because it does not rely on manually labeled data. Future research will investigate methods that further increase the network's performance on real-world data.
\vfill\null
\addtolength{\textheight}{-4cm}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sect:intro}
Relativistic jet outflows from active galactic nuclei can transport significant energy from the central supermassive black hole to the surrounding intercluster medium \citep[ICM;][]{Scheuer1974,Scheuer1982,Begelman1984}. Interactions between the jet and ICM will create radio lobes at kiloparsec-scale distances, making radio observations well suited for extragalactic jet studies \citep[e.g.,][]{Blandford1974,Hargrave1974,Perley1984}. Jets are responsible for feedback processes that prevent central gas from cooling in clusters of galaxies \citep{Birzan2004, Rafferty2006, McNamara2007}, consequently reducing star formation at the cluster centers \citep{Fabian1994, Fabian2012}. Thus, interactions between jets and the surrounding medium govern the overall evolution of these extragalactic systems.
\begin{table*}
\caption{\chandra{} Observations of Quasar Sample}
\begin{tightcenter}
\label{table:obs}
\begin{tabular}{ c c c c c c c c c D }
\hline \hline
Object & $z$\tablenotemark{a} & R.A.\tablenotemark{b} & Decl.\tablenotemark{b} & ObsID\tablenotemark{c} & Observation & $t_{\rm exp}$\tablenotemark{d} & log($M_{\rm BH}$)\tablenotemark{e} & log($L_{\rm bol}$)\tablenotemark{e} & \multicolumn{2}{c}{log($\lambda_{\rm Edd}$)\tablenotemark{f}}\\
& & [J2000] & [J2000] & & Date & [ks] & [log($M_{\odot})$] & [log($\rm erg\,s^{-1}$)]& & \\
\hline
\decimals
J0801$+$4725 & 3.256 & 08:01:37.682 & +47:25:28.24 & 20405 & 2018 Jan 19 & 9.40 & $8.53\pm0.47$ & 46.97 & $0.34$ \\
J0805$+$6144 & 3.033 & 08:05:18.180 & +61:44:23.70 & 20399 & 2018 Jun 07 & 9.75 & -- & -- & \multicolumn{2}{c}{--} \\
J0833$+$0959 & 3.713 & 08:33:22.514 & +09:59:41.14 & 20401 & 2019 Feb 04 & 9.57 & $9.58\pm0.09$ & 47.10 & $-0.58$ \\
J0909$+$0354 & 3.288 & 09:09:15.915 & +03:54:42.98 & 20404 & 2018 Mar 03 & 9.57 & $9.55\pm0.05$ & 46.88 & $-0.77$ \\
J0933$+$2845 & 3.431 & 09:33:37.298 & +28:45:32.24 & 20403 & 2018 Jun 10 & 9.57 & $9.56\pm0.01$ & 47.27 & $-0.39$ \\
J1016$+$2037 & 3.114 & 10:16:44.322 & +20:37:47.30 & 20411 & 2018 Jan 24 & 9.57 & $8.91\pm0.03$ & 47.08 & $0.07$ \\
J1128$+$2326 & 3.042 & 11:28:51.701 & +23:26:17.35 & 20412 & 2019 Mar 10 & 9.57 & $9.08\pm0.04$ & 47.07 & $-0.11$ \\
J1223$+$5038 & 3.491 & 12:23:43.169 & +50:37:53.40 & 20402 & 2018 Jul 31 & 9.57 & $9.99\pm0.02$ & 47.86 & $-0.23$ \\
J1405$+$0415 & 3.215 & 14:05:01.120 & +04:15:35.82 & 20408 & 2018 May 08 & 9.57 & $8.86\pm0.04$ & 46.97 & $-0.01$ \\
J1435$+$5435 & 3.810 & 14:35:33.779 & +54:35:59.31 & 20400 & 2018 Sep 07 & 9.57 & $8.72\pm0.06$ & 46.63 & $-0.19$ \\
J1610$+$1811 & 3.122 & 16:10:05.289 & +18:11:43.47 & 20410 & 2018 May 24 & 9.09 & $9.94\pm0.03$ & 47.31 & $-0.73$ \\
J1616$+$0459 & 3.212 & 16:16:37.557 & +04:59:32.74 & 20407 & 2018 May 08 & 9.48 & -- & -- & \multicolumn{2}{c}{--} \\
J1655$+$3242 & 3.181 & 16:55:19.225 & +32:42:41.13 & 20409 & 2018 Nov 26 & 9.57 & $8.69\pm0.03$ & 46.80 & $0.01$ \\
J1655$+$1948 & 3.262 & 16:55:43.568 & +19:48:47.12 & 20406 & 2018 Jun 17 & 9.57 & $9.06\pm0.04$ & 46.78 & $-0.38$ \\
\hline
\end{tabular}
\end{tightcenter}
\tablenotemark{a}{Redshift measurements from \cite{Sowards-Emmerd2005}, \cite{Husband2015}, and \cite{Paris2018}.}
\tablenotemark{b}{Radio centroid coordinates from VLA positions reported in \cite{Gobeille2014}.}
\tablenotemark{c}{Observations performed using \chandra\ ACIS-S instrument with the aimpoint on the S3 chip.}
\tablenotemark{d}{Total exposure time after flare removal reprocessing and dead time correction.}
\tablenotemark{e}{Measurements from \cite{Rakshit2020}.}
\tablenotemark{f}{Eddington ratios for the quasar sample.}
\end{table*}
X-ray observations of extragalactic radio sources provide insights into fundamental physical processes present within these systems \citep[e.g.,][]{Heinz1998, Harris2006, Stawarz2008, Worrall2008}. Although the origin of the X-ray jet emission mechanism is not uniquely defined, one probable mechanism is that the X-rays are generated from inverse Compton upscattering of the cosmic microwave background radiation \citep[IC/CMB;][]{Tavecchio2000, Celotti2001}. Under such an assumption, X-ray observations of jets may be used to measure the enthalpy flux, or ``power," transported by the jets to the radio lobes and the ICM \citep{Scheuer1974, Heinz1998, Reynolds2001}. IC/CMB is predicted to be the dominant X-ray emission mechanism for high-redshift radio jets as the cosmological diminution of surface brightness by the factor $(1+z)^{-4}$ is offset by the $(1+z)^{4}$ increase in the CMB energy density. IC/CMB is additionally bolstered by the longer lifetimes of the 100\,MeV electrons, which generate such emission, relative to the 10\,GeV electrons required for radio synchrotron radiation. These emission properties make X-ray observations well suited for detecting high-redshift jets.
Although X-rays are uniquely suitable for investigating high-redshift jets, the only telescope presently capable of resolving extended X-ray structures from high-redshift radio sources is the \chandra{} X-ray Observatory. Previous \chandra{} observations have demonstrated this capability by resolving jets in luminous radio sources at redshifts up to $z = 4.7$ \citep[e.g.,][]{Siemiginowska2003,Cheung2012, McKeough2016, Simionescu2016}. However, high-redshift X-ray jets remain vastly undersampled in comparison to nearby sources \citep{Worrall2020}. It is therefore the focus of this work to examine \chandra{} observations of additional high-redshift radio sources for evidence of X-ray structure(s) and to quantify the emission properties of these features, particularly in the context of IC/CMB radiation.
This paper is one in a series on \chandra{} observations of 14 radio-luminous quasars selected from GHz surveys where all targets are in the redshift range $3.0 < z < 4.0$ \citep{Schwartz2020}. Here, we investigated for X-ray emission from the sources while also analyzing the spectroscopic properties of both the quasar core and the extended features. The remainder of the paper is arranged as follows. Section~\ref{sect:observation} describes the sample selection criteria and X-ray data reprocessing. Section~\ref{sect:spec} describes the X-ray spectroscopic analysis of the sample, while Section~\ref{sect:extend} covers the X-ray morphological analysis. Section~\ref{sect:outliers} details the measured properties of the five sources from our sample where extended X-ray structure is detected, including relevant flux and surface brightness results. Optical and radio properties of the quasar cores with regard to both the sample and other quasar surveys are provided in Sections~\ref{sect:ox} and \ref{sect:radio}, respectively. Our concluding remarks are provided in Section~\ref{sect:conclude}.
For this paper, we adopted the cosmological parameters $H_{0} = 70\rm\,km\,s^{-1}\,Mpc^{-1}$, $\Omega_{\Lambda} =0.7$, and $\Omega_{M} = 0.3$ \citep{Hinshaw2013}.
\begin{figure*}
\begin{tightcenter}
\includegraphics[width=0.996\linewidth]{samplev3.eps}
\end{tightcenter}
\caption{\chandra{} 0.5--7.0\,keV images of the fourteen targets binned in 0\farcs25 pixels. The X-ray images are overlaid with radio map contours (green) from VLA observed at 6.2\,GHz, except for J0833$+$0959 which is overlaid with a 1.4\,GHz VLA radio map. The restoring beam for each radio map is shown as a black ellipse.}
\label{fig:sample}
\end{figure*}
\section{Sample Selection and Data Reduction}
\label{sect:observation}
Targets for our analysis were selected from a catalog of 123 radio-bright quasars at redshifts $z>2.5$ \citep{Gobeille2011,Gobeille2014} constructed from an overlapping region of the VLA-FIRST radio survey \citep{Becker1995} and the Sloan Digital Sky Survey \citep{Abazajian2003}. All sources in the catalog have a spectroscopically measured redshift and a flux density in excess of 70\,mJy at either 1.4 or 5\,GHz. From this catalog, we focused on the 61 sources that possess resolved radio features at 1\arcsec{} or better resolution. We also prioritized sources where the separation distance between the radio features exceeded 1\arcsec{} as such spatial separations are resolvable with the \chandra\ X-ray Observatory, assuming comparable sizes for any X-ray counterparts. Lastly, we eliminated radio objects identified as triples because their jets are less likely to be at a line of sight that will achieve the relativistic beaming required for detection in X-rays. Of the remaining 31 sources that satisfied these criteria, 14 objects at redshifts $z > 3$ with resolved radio structure that had previously not been observed in X-rays were selected for our survey.
Each target selected for our survey was observed with \chandra\ using the Advanced CCD Imaging Spectrometer (ACIS) with the aimpoint centered on the S3 chip. The instrument was placed in the 1/4 subarray timed exposure mode with {\tt vfaint} telemetry enabled in an effort to mitigate pileup. Roll direction for each observation was defined such that the radio features did not coincide with the readout direction of the chip. We also confirmed that the known \chandra\ PSF artifact\footnote{See the PSF artifact caveat in the \ciao{} User Guide: \\ \url{https://cxc.harvard.edu/ciao/caveats/psf_artifact.html}}, which can create nonphysical X-ray features on a subarcsecond scale, was not coincident with any radio feature in our sample.
\begin{table*}
\caption{VLA Observations of Quasar Sample}
\label{table:radio_obs}
\begin{tightcenter}
\begin{tabular}{ccccccDc}
\hline \hline
Object & Exposure & Frequency & Observation & Beam & Beam & \multicolumn2c{Beam Position} & Minimum Contour \\
& [s] & [GHz] & Date & Maximum & Minimum & \multicolumn2c{Angle} & Intensity \\
& & & & [arcsec] & [arcsec] & \multicolumn2c{[deg]} & [mJy beam$^{-1}$] \\
\hline
\decimals
J0801$+$4725 & 378.9 & 6.2 & 2012 Nov 18 & 0.36 & 0.31 & \hspace{1.3em}-16.0 & 0.45 \\
J0805$+$6144 & 319.1 & 6.2 & 2012 Nov 18 & 0.40 & 0.32 & -11.5 & 0.68 \\
J0833$+$0959 & 580.0 & 1.4 & 2008 Oct 23 & 1.95 & 1.41 & 2.3 & 0.37 \\
J0909$+$0354 & 319.1 & 6.2 & 2012 Nov 18 & 0.44 & 0.32 & -23.5 & 0.10 \\
J0933$+$2845 & 319.1 & 6.2 & 2012 Nov 18 & 0.37 & 0.34 & -57.9 & 0.52 \\
J1016$+$2037 & 398.9 & 6.2 & 2012 Nov 12 & 0.35 & 0.32 & 11.4 & 0.13 \\
J1128$+$2326 & 319.1 & 6.2 & 2012 Nov 12 & 0.34 & 0.32 & -7.7 & 0.15 \\
J1223$+$5038 & 299.2 & 6.2 & 2012 Nov 12 & 0.41 & 0.33 & 72.3 & 0.12 \\
J1405$+$0415 & 319.1 & 6.2 & 2012 Nov 18 & 0.39 & 0.32 & 7.5 & 0.62 \\
J1435$+$5435 & 319.1 & 6.2 & 2012 Nov 18 & 0.41 & 0.32 & -61.4 & 0.40 \\
J1610$+$1811 & 319.1 & 6.2 & 2012 Nov 08 & 0.34 & 0.31 & -16.6 & 0.18 \\
J1616$+$0459 & 259.2 & 6.2 & 2012 Nov 08 & 0.39 & 0.31 & -19.7 & 0.85 \\
J1655$+$3242 & 378.9 & 6.2 & 2012 Nov 08 & 0.24 & 0.21 & 63.8 & 0.68 \\
J1655$+$1948 & 319.1 & 6.2 & 2012 Nov 08 & 0.35 & 0.31 & -28.9 & 0.34 \\
\hline
\end{tabular}
\end{tightcenter}
\end{table*}
All X-ray observations were analyzed using the level 2 data products from the standard \chandra\ data processing pipeline together with the software analysis package \ciao{}v4.12 with \caldb{}v4.9.2.1. Each observation was reprocessed using the routine {\tt deflare} to remove background flaring periods from the data, and the average cleaned exposure time per target is 9.53\,ks. Pileup was estimated for each source using \pimms{}v4.11a, and all sources had $< 3\%$ pileup over the 0.5--7.0\,keV energy band. Thus, pileup was ignored for our X-ray analysis. Details on our \chandra\ observations are provided in Table~\ref{table:obs}.
In addition to the X-ray data, radio observations of the sample were included in our analysis. We obtained new Karl G. Jansky Very Large Array (VLA) A-array observations of these 14 quasars (see Table~\ref{table:radio_obs} for a summary) as part of a more extensive radio imaging and follow-up program of this sample \citep{Gobeille2014}. The data were calibrated and imaged using standard procedures in CASA and AIPS. The majority of the VLA maps (13/14) were obtained as single $\sim$\,4--6 minute snapshot observations in November 2012 (program 12B-230) using two 1\,GHz wide intermediate-frequency bands centered at 4.9 and 7.4\,GHz (effective center frequency of 6.2 GHz). The maps of J1405$+$0415 and J1610$+$1811 were previously published in \cite{Schwartz2020}. For J0833$+$0959, we used an earlier VLA A-array 1.4\,GHz map we obtained in October 2008 (program AW748) and published in \cite{Gobeille2014}.
The X-ray and radio observations were aligned based on the centroid position of the core. Centroid position was measured in each observation using the {\tt dmstat} routine in \ciao, and the 0.5--7.0\,keV X-ray observation coordinates were shifted with {\tt wcs\_update} to agree with the radio data. The average astrometric translations for the sample were $\Delta x_{\rm avg} = 0\farcs62$ and $\Delta y_{\rm avg} = 0\farcs45$, which are consistent with the overall 90\% absolute position uncertainty of 0\farcs8 for \chandra. We also generated radio contours from the available maps. The minimum contour level was defined as $\sim$\,3--5 times the rms noise measured in an off-source region on the image, and the contour levels were drawn at increasing factors of four intervals for all sources. The resulting X-ray--radio overlays are shown in Figure~\ref{fig:sample}, where the 0.5--7.0\,keV \chandra\ images are subpixel binned in 0\farcs25 pixels. Additional details on the radio maps and contours used in our analysis are provided in Table~\ref{table:radio_obs}.
\section{X-ray Emission Spectra}
\label{sect:spec}
\begin{table*}
\caption{X-ray Properties of the Quasar Sample}
\label{table:xray}
\begin{tightcenter}
\begin{tabular}{ c c c c D c c c D D D }
\hline \hline
Object & $C_{\rm obs}$ & Model & $N_{\rm H}$ & \multicolumn2c{$N_{\rm H}^i$} & $\Gamma$ & $E_{\rm line}$ & $I_{\rm line}$ & \multicolumn2c{$f_{\rm 0.5-7.0\,keV}$} & \multicolumn2c{$L_{\rm 2-10\,keV}$} & \multicolumn2c{$\ell_{\rm 2\,keV}$} \\
(1) & (2) & (3) & (4) & \multicolumn2c{(5)} & (6) & (7) & (8) & \multicolumn2c{(9)} & \multicolumn2c{(10)} & \multicolumn2c{(11)} \\
\hline
\decimals
J0801$+$4725 & 48 & A & 0.0454 & $<15.5$ & $2.01\substack{+0.30\\-0.29}$ & -- & -- & $5.8\substack{+1.0\\-0.9}$ & $3.4\substack{+1.2\\-1.0}$ & $4.3\substack{+3.1\\-1.9}$ \\
J0805$+$6144 & 475 & A & 0.0452 & $<4.7$ & $1.24\substack{+0.09\\-0.09}$ & -- & -- & $66.1\substack{+3.4\\-3.4}$ & $19.4\substack{+2.4\\-2.2}$ & $12.7\substack{+2.8\\-2.3}$ \\
J0833$+$0959 & 173 & A & 0.0394 & $<11.5$ & $1.60\substack{+0.16\\-0.16}$ & -- & -- & $23.3\substack{+2.0\\-1.9}$ & $13.7\substack{+3.1\\-2.7}$ & $12.5\substack{+5.1\\-3.8}$ \\
J0909$+$0354 & 797 & A & 0.0347 & $<3.5$ & $1.16\substack{+0.07\\-0.07}$ & -- & -- & $116.4\substack{+4.4\\-4.6}$ & $37.0\substack{+3.6\\-3.4}$ & $22.4\substack{+3.8\\-3.4}$ \\
J0933$+$2845 & 80 & A & 0.0191 & $<10.7$ & $1.65\substack{+0.22\\-0.22}$ & -- & -- & $10.4\substack{+1.2\\-1.2}$ & $5.3\substack{+1.6\\-1.3}$ & $5.1\substack{+2.9\\-1.9}$ \\
J1016$+$2037 & 196 & A & 0.0242 & $<6.2$ & $1.67\substack{+0.14\\-0.14}$ & -- & -- & $25.6\substack{+2.1\\-2.0}$ & $10.9\substack{+2.0\\-1.8}$ & $10.6\substack{+3.6\\-2.8}$ \\
J1128$+$2326 & 91 & A & 0.0134 & $<9.3$ & $1.48\substack{+0.20\\-0.20}$ & -- & -- & $12.7\substack{+1.6\\-1.3}$ & $4.5\substack{+1.3\\-1.0}$ & $3.7\substack{+2.0\\-1.3}$ \\
J1223$+$5038 & 464 & A & 0.0169 & $<4.3$ & $1.62\substack{+0.08\\-0.09}$ & -- & -- & $62.4\substack{+3.3\\-3.1}$ & $32.7\substack{+4.1\\-3.7}$ & $30.6\substack{+6.6\\-5.5}$ \\
... & ... & B & ... & $<4.3$ & $1.50\substack{+0.12\\-0.06}$ & $6.28\substack{+0.05\\-0.09}$ & $5.44\substack{+1.99\\-1.83}$ & $60.5\substack{+3.2\\-3.1}$ & $28.8\substack{+4.5\\-2.8}$ & $24.1\substack{+7.0\\-3.5}$ \\
J1405$+$0415 & 280 & A & 0.0217 & $<4.9$ & $1.49\substack{+0.11\\-0.11}$ & -- & -- & $37.3\substack{+2.7\\-2.3}$ & $14.8\substack{+2.4\\-2.1}$ & $12.3\substack{+3.5\\-2.7}$ \\
J1435$+$5435 & 38 & A & 0.0127 & $<34.0$ & $1.32\substack{+0.32\\-0.32}$ & -- & -- & $5.3\substack{+1.1\\-0.9}$ & $2.5\substack{+1.5\\-1.0}$ & $1.8\substack{+2.0\\-1.0}$ \\
J1610$+$1811 & 397 & A & 0.0362 & $<4.5$ & $1.64\substack{+0.10\\-0.10}$ & -- & -- & $53.4\substack{+3.0\\-2.8}$ & $22.2\substack{+2.8\\-2.5}$ & $21.1\substack{+4.8\\-4.0}$ \\
J1616$+$0459 & 386 & A & 0.0475 & $<6.0$ & $1.48\substack{+0.10\\-0.10}$ & -- & -- & $51.1\substack{+3.3\\-3.0}$ & $20.1\substack{+2.9\\-2.5}$ & $16.5\substack{+4.1\\-3.3}$ \\
J1655$+$3242 & 41 & A & 0.0222 & $<19.3$ & $1.51\substack{+0.31\\-0.31}$ & -- & -- & $5.5\substack{+1.0\\-0.9}$ & $2.2\substack{+1.0\\-0.7}$ & $1.8\substack{+1.7\\-0.9}$ \\
J1655$+$1948 & 47 & A & 0.0549 & $<14.1$ & $1.91\substack{+0.29\\-0.29}$ & -- & -- & $5.8\substack{+1.0\\-0.8}$ & $3.2\substack{+1.2\\-0.9}$ & $3.8\substack{+2.7\\-1.7}$ \\
\hline
\end{tabular}
\end{tightcenter}
(1) Object name. (2) Observed counts over the 0.5--7.0\,keV band from 1\farcs5 radius circle centered on the quasar, where the average background contribution is 0.17\% of the observed counts. (3) Spectral model. (4) Galactic column density extrapolated from \cite{Dickey1990}, in units of $10^{22}\rm\,cm^{-2}$. (5) Intrinsic column density 3$\sigma$ upper limits, in units of $10^{22}\rm\,cm^{-2}$. (6) Photon index estimated from 0.5--7.0\,keV best-fit spectral model. (7) Rest-frame iron emission line energy, in unit of keV. (8) Observed line intensity, in units of $10^{-6}\rm\,photons\,cm^{-2}\,s^{-1}$. (9) Observed 0.5--7.0\,keV flux, in units of $10^{-14}\rm\,erg\,cm^{-2}\,s^{-1}$. (10) Rest-frame 2--10\,keV luminosity, in units of $10^{45}\rm\,erg\,s^{-1}$. (11) Rest-frame, monochromatic luminosity at 2\,keV, in units of $10^{27}\rm\,erg\,s^{-1}\,Hz^{-1}$.
\end{table*}
The unabsorbed X-ray fluxes and luminosities for the sample were determined by modeling the spectrum from each source. An emission spectrum from each observation was extracted using the {\tt specextract} routine in \ciao. We defined the source region as a 1\farcs5 radius circle centered on the quasar, while the background region was defined as an annulus, also centered on the quasar, with an inner radius of 15\arcsec{} and an outer radius of 30\arcsec. We masked any point sources in the background region, if present, to avoid count biases. We note that the average percent difference between the total and background-corrected counts is 0.17\% for our extracted spectra, so the background contribution is minor. Each spectrum was binned at 1 count per bin over the 0.5--7.0\,keV band.
While performing our spectral analysis of the sample, we additionally tested for the presence of spectral lines, such as an Fe\,K line. We defined two different spectral models, one with an emission line and one without, and assessed both models with a likelihood ratio test for each observation. The line emission component was included in our model in cases where its introduction improved the $p$-value of the likelihood ratio test to a value less than 0.001. The models are as follows:
Model A: {\tt phabs$\cdot$powerlaw}. This is the default model for X-ray emission from the quasar core that was used for all 14 sources.
Model B: {\tt phabs$\cdot$(powerlaw\,$+$\,zgauss)}. This model includes the primary X-ray emission and a Gaussian emission line, where we fixed the line width to 0.001\,keV. This model was tried for all sources but preferred only for J1223$+$5038.
Each extracted spectrum was fit over the 0.5--7.0\,keV energy band using WStat in \ciao's modeling and fitting package \sherpa. Galactic hydrogen column density $N_{\rm H}$ was fixed to extrapolated values from \cite{Dickey1990}, while the photon index $\Gamma$ ($dN/dE \propto E^{-\Gamma}$) and the normalization(s) were allowed to vary for all models. Once a best-fit model was realized, an additional intrinsic absorption component (i.e., {\tt zphabs}) was added to the model to determine if the intrinsic absorption column density $N_{\rm H}^i$ could be constrained. The revised model was fit to the data with $N_{\rm H}^i$, normalization, and $\Gamma$ as free parameters. In all cases, only an upper limit on the intrinsic absorption could be established as its addition did not improve the overall fit statistics. Model parameter best-fit results from the spectral analysis and their respective $1\sigma$ confidence intervals are provided in Table~\ref{table:xray}. We note that the upper limits on $N_{\rm H}^i$ are reported at $3\sigma$ confidence.
We found from our spectral analysis that all 14 sources in the sample could be individually fit with Model A, providing constraints on both $N_{\rm H}^i$ and $\Gamma$. Additionally, we determined from our likelihood ratio tests that J1223$+$5038 was best fit with Model B due to the presence of an emission line in its spectrum. The measured emission line is consistent with neutral Fe emission at 6.4\,keV, while the $N_{\rm H}^i$ and $\Gamma$ best-fit parameters agree with the Model A results. The observed spectrum and model fit for J1223$+$5038 are shown in Figure~\ref{fig:spectrum}, and the best-fit results are in Table~\ref{table:xray}.
Once the best-fit spectral models were obtained, we measured the observed Galactic absorption-corrected 0.5--7.0\,keV flux $f_{\rm 0.5-7.0\,keV}$ for each source. The rest-frame 2--10\,keV luminosity $L_{\rm 2-10\,keV}$ and monochromatic 2\,keV luminosity $\ell_{\rm 2\,keV}$ were additionally determined from each best-fit model. We applied aperture correction for all measured fluxes and luminosities, where the correction factor was derived from our encircled counts fraction (ECF) analysis of the sources (see Section~\ref{sect:extend}). The average correction factor for our sample is 1.075, and the range is 1.056--1.098. All measured X-ray fluxes and luminosities for the sample are shown in Table~\ref{table:xray}.
\section{X-ray Morphology}
\label{sect:extend}
As discussed in Section~\ref{sect:observation}, our sample of high-redshift quasars has known radio features at distances of 1\arcsec{} or greater from the quasar core. Existing X-ray counterparts to these radio features may be resolvable with our \chandra\ observations, assuming that both a satisfactory signal-to-noise ratio (S/N) is achieved and the radio-/X-ray-emitting regions are comparable in size. Thus, we investigated our sample for evidence of resolved X-ray structures.
The asymmetric point-spread function (PSF) for \chandra\footnote{See `Understanding the Chandra PSF' thread in the \ciao{} guide: \url{https://cxc.cfa.harvard.edu/ciao/PSFs/psf_central.html}}
must be considered when investigating features on scales of $\sim$\,1\arcsec, which we suspected for our sources. As there is presently no analytic PSF for \chandra, we generated synthetic PSF images for our observations. We began by simulating 500 ray-tracing files for each X-ray observation using ChaRT\,v2, a web interface to the SAOsac ray-trace code. The ray-tracing files and various physical parameters of the observation, such as exposure time and detector orientation, were input into the {\tt simulate\_psf} script in \ciao\ to accurately generate simulated PSF images. All ray-tracing files were projected using MARX\,v5.5.0, and the extracted spectra from Section~\ref{sect:spec} were used to reproduce the spectral response of our observations. Since the default value for the aspect blur parameter in {\tt simulate\_psf} has previously been shown to generate a simulated PSF profile narrower than what is observed,\footnote{See `The AspectBlur Parameter in MARX' thread in the \ciao{} guide: \url{https://cxc.cfa.harvard.edu/ciao/why/aspectblur.html}}, we generated multiple simulated PSF with different aspect blur values. The ECF was calculated for each simulation and compared with the observed data. See Figure~\ref{fig:ecf} for an example of this comparison for J1223$+$5038. From our analysis, we found an aspect blur value of 0\farcs28 to best fit all sources in our survey.
\begin{figure}
\begin{tightcenter}
\includegraphics[width=0.99\linewidth]{spectrumv2.eps}
\end{tightcenter}
\caption{\chandra\ X-ray spectrum of J1223$+$5038. The spectrum is binned by 5 counts per bin (solely for illustrative purposes) and is fitted over the 0.5--7.0\,keV energy range with an absorbed power-law model ($\Gamma = 1.50\substack{+0.12\\-0.06}$) and an Fe\,K emission line at the rest energy of 6.28\,keV, or observed energy of 1.39\,keV. The lower panel shows the residuals from the best-fit model.}
\label{fig:spectrum}
\end{figure}
After verifying that all PSF simulations were caligned with their respective \chandra\ observations to subpixel accuracy, we defined an annulus region for each quasar. The outer annular radius was set such that it encompassed the observed radio features plus an additional 1\arcsec{} to account for \chandra\ PSF blurring. Because the radio map for J0833$+$0959 has a poorer resolution than the remainder of our sample, we defined its outer annular radius as 1\arcsec{} from the centroid of the external radio feature. The inner annular radius was set equal to the 95\% ECF radius of the core, as measured from our simulated ECF profiles, in cases where the core separation distance for the nearest radio feature is $>2\arcsec{}$. In cases where the core separation distance is $\leq 2\arcsec$, the inner radius was set equal to 85\% ECF radius. The resulting inner radii for the sample ranged between 1\farcs0--1\farcs7, and the outer radii were between 2\farcs0--11\farcs0. Each annulus was additionally divided into twelve $30{^\circ}$ sectors, where we defined our coordinate system as $0{^\circ}$ West rotating counterclockwise. The sectors used for J1610$+$1811 are shown in Figure~\ref{fig:sectors}.
Total counts for each annular sector of the observations and simulations were measured using {\tt dmextract} in \ciao. Counts were also measured for a circular region 1\arcsec{} in radius surrounding the core in each X-ray image, and the simulated leakage counts into each annular sector were scaled based on the ratio of the observed-to-simulated core counts. The average observed background counts per sector were estimated using the background regions discussed in Section~\ref{sect:spec}.
\begin{figure}
\begin{tightcenter}
\includegraphics[width=0.99\linewidth]{ecf_figv3.eps}
\end{tightcenter}
\caption{A comparison of the background-subtracted encircled counts fraction (ECF) for the quasar J1223$+$5038 versus simulated PSFs generated from the {\tt simulate\_psf} routine in \ciao. For each simulation, a different aspect blur parameter was used. We found that a blur parameter of 0\farcs28 is best for reproducing the sources in our sample.}
\label{fig:ecf}
\end{figure}
\subsection{Statistical Assessment of X-ray Features}
\label{sect:annulus}
Having measured the counts per annular sector for both observations and simulations, we examined which sectors, if any, have elevated observed counts relative to the simulated PSF as this would indicate the presence of extended X-ray emission. To ensure a rigorous detection criterion, we derived a counts probability for each sector based on the simulated and field background emissions in order to find sectors that are statistically significant outliers.
To begin, we inferred that the simulated count probability distribution of each sector in our analysis could be accurately modeled with Poisson statistics. Thus, the simulated counts per sector may be represented as a Poisson distribution where the average is calculated from the 500 simulations. We note that the simulations do not include background emission, which must be accounted for when comparing to the observed counts. The background counts may similarly be represented as another Poisson distribution where the mean value corresponds to the predicted background counts per sector. Because the sum of independent Poisson random variables is itself Poisson \citep{Grimmett1986}, we may define the sum of simulated and background distributions as $P(\lambda_{1}+\lambda_{2}$), where $\lambda_{1}$ is the mean simulated counts per sector and $\lambda_{2}$ is the expected background counts per sector. We may therefore accurately determine the probability for the total observed counts per annular sector for each source.
\begin{figure}
\begin{tightcenter}
\includegraphics[width=0.999\linewidth]{sectors.eps}
\end{tightcenter}
\caption{The annular sectors defined in Section~\ref{sect:extend} for 1 of the 14 sources in our quasar sample. Each annulus was divided into twelve 30$^{\circ}$ sectors, beginning West and rotating counterclockwise. In the source shown, the inner and outer radii correspond to 1\farcs5 and 6\arcsec, respectively. For convenience, sectors 1 and 12 have been labeled.}
\label{fig:sectors}
\end{figure}
Using the predicted probability distributions, we calculated the cumulative probability of obtaining the observed count rate or higher for each sector of each source. Due to our method of defining sectors, it may be possible that extended X-ray features lie in multiple sectors. We therefore also assessed probabilities for the sum of adjacent sectors, up to a total sector size of $90^{\circ}$, in order to address these edge cases. We defined a probability $p \leq 0.005$ as the detection threshold for the sample. Based on our defined significance threshold, we found five sectors from five unique sources that have significant evidence of X-ray emission. The X-ray features and their respective probabilities are provided in Table~\ref{table:extend}, and images of the five sources with their corresponding emission sectors are shown in Figure~\ref{fig:extended}. A machine-readable table of the complete sector analysis for all sources is available as an accompanying online resource to this article.
\section{Resolved X-ray Features}
\label{sect:outliers}
The morphological analysis in Section~\ref{sect:extend} identified five quasars with extended X-ray emission at distances of 1--12\arcsec\ from the quasar core, where the projected jet length range is 20--80\,kpc. These jet lengths are consistent with those observed for nearby, high-luminosity jets \citep{Marshall2018}. Examination of the X-ray properties of the quasar cores, as well as the multiwavelength properties of the cores (see Sections~\ref{sect:ox} and \ref{sect:radio}), show no identifiable characteristics that would isolate these sources from the remainder of the sample. We therefore sought to quantify the X-ray properties of the five extended X-ray features.
\begin{table}
\caption{X-ray Morphological Analysis
\label{table:extend}}
\begin{tightcenter}
\footnotesize
\begin{tabular}{c c D c c c D}
\hline \hline
Object & $r_{\rm in}$ & \multicolumn2c{$r_{\rm out}$} & $C_{\rm obs}$ & $C_{\rm PSF}$ & $C_{\rm bg}$ & \multicolumn2c{$p$} \\
(1) & (2) & \multicolumn2c{(3)} & (4) & (5) & (6) & \multicolumn2c{(7)} \\
\hline
J0833$+$0959 & 1.5 & 11.0 & 8 & 0.84 & 0.60 & $<0.0001$ \\
J0909$+$0354 & 1.7 & 3.5 & 12 & 2.42 & 0.07 & $<0.0001$ \\
J1016$+$2037 & 1.0 & 3.0 & 8 & 1.67 & 0.04 & $0.0004$ \\
J1405$+$0415 & 1.0 & 2.5 & 8 & 2.19 & 0.03 & $0.0021$ \\
J1610$+$1811 & 1.6 & 6.0 & 7 & 1.68 & 0.21 & $0.0033$ \\
\hline
\end{tabular}
\end{tightcenter}
{(1) Object name. (2) Inner annular radius, in units of arcseconds. (3) Outer annular radius, in units of arcseconds. (4) Observed 0.5--7.0\,keV counts. (5) Mean 0.5--7.0\,keV counts from simulated PSF. (6) Expected background 0.5--7.0\,keV counts. (7) Cumulative Poisson probability of detecting counts $\geq C_{\rm obs}$. The results listed in the table are those sectors where the probability was $\leq$ 0.005, indicating strong evidence of a resolved X-ray feature. (A machine-readable table of the complete sector analysis for all sources is available.)}
\end{table}
\subsection{Flux and Surface Brightness}
\label{sect:flux_extend}
\begin{figure*}
\begin{tightcenter}
\includegraphics[width=0.98\linewidth]{extendedv3.eps}
\end{tightcenter}
\caption{\chandra{} 0.5--7.0\,keV images of the five sources where resolved X-ray structure is detected from our statistical analysis. Each image is binned in 0\farcs25 pixels and overlaid with radio map contours (green). The regions shown are the sectors where X-ray features were detected, and the restoring beam for each radio map is shown as a black ellipse.}
\label{fig:extended}
\end{figure*}
The flux and the surface brightness of each X-ray feature detected in Section~\ref{sect:annulus} may be measured, given some assumptions on the X-ray emission. Due to the limited X-ray counts in our observations, identification of edges for the X-ray extension could not be performed using either surface brightness profiles or contour mapping, as is standard in jet and hotspot analyses. We instead assumed that the X-rays were generated from jets that extended from the quasar core. We additionally assumed that the annular sectors in Figure~\ref{fig:sectors} were sufficient in size to encompass the extended X-ray emission. The annular sectors were verified to reproduce the same measured count rates as other region shapes, such as rectangular and elliptical regions \citep[e.g.,][]{Schwartz2020}. Furthermore, we found the annular sectors to minimize the overall systematic error introduced from the PSF counts, with a 20\% reduction in $C_{\rm PSF}$ relative to similarly sized rectangles and ellipses, while also improving the overall reproducibility of our flux measurements. Thus, we used the count rates measured with the annular sectors (Table~\ref{table:extend}) for our flux analysis.
The source counts of each region were calculated by subtracting the mean counts from the simulated PSF and the expected background counts from the total observed counts. We then modeled each region with a {\tt phabs$\cdot$powerlaw} expression in \pimms{}v4.11a, using the latest \chandra\ response files, and set the spectral normalization equal to our X-ray jet count rates. The Galactic column density was fixed equal to values from \cite{Dickey1990}. We also fixed the photon index to $\Gamma = 1.9$, which we inferred to be a standard slope for X-ray jet and knot spectra \citep[e.g.,][]{Marshall2002,Siemiginowska2002,Schwartz2006,Schwartz2006b,Goodger2010,Zhang2018}. The observed 0.5--7.0\,keV X-ray jet fluxes are shown in Table~\ref{table:flux}. We note that the flux values for J1405$+$0415 and J1610$+$1811 agree with previous measurements from \cite{Schwartz2020}. Rest-frame 2--10\,keV luminosities were also measured for each source and are provided in Table~\ref{table:flux}. We found the fluxes and luminosities to be similar for the five jets.
We additionally calculated the surface brightness values for each source. We assumed the emitting area originated at the quasar core and extends to the outer radius of our defined regions. The width of the emitting region was set equal to the FWHM of \chandra, which is 0\farcs5. See Table~\ref{table:flux} for the flux, luminosity, and surface brightness measurements of the X-ray jets. Given the limited count statistics for these extended X-ray features, the measurements in Table~\ref{table:flux} have an approximate factor of 2 error.
\subsection{IC/CMB Emission}
\label{sect:iccmb}
The IC/CMB in a jet with bulk relativistic motion is a known X-ray emission mechanism in quasar jets, where the broadband emission is attributed to a single spectrum of relativistic electrons \citep{Tavecchio2000,Celotti2001,Schwartz2002,Siemiginowska2002,Schwartz2006,Schwartz2006b,Worrall2009,Marshall2011,Marshall2018, Worrall2020}. However, recent studies have demonstrated that IC/CMB from a single electron population does not reproduce the observed spectral properties of several lower-redshift jets \citep[e.g.,][]{Jester2006,Siemiginowska2007,Cara2010,Meyer2015,Clautice2016}. In addition, upper limits on 0.1--100\,GeV $\gamma$-ray production in jets measured from Fermi observations imply that a single electron population is insufficient to produce the observed X-ray emission in several jets at $z < 1$ \citep{Breiding2017,Meyer2017b}, though this issue may be alleviated with increased electron cooling \citep{Lucchini2016}. Overall, these results refute the single electron population required for standard IC/CMB jet models and instead require multiple electron populations and/or different emission mechanisms to reproduce the observed multiwavelength radiation.
Despite these results for low-redshift ($z<3$) systems, IC/CMB is theorized to be the predominant emission mechanism at $z \gtrsim 3$ as the energy density of the CMB exceeds the magnetic energy density for the quasar jets \citep{Schwartz2020}. This effect increases the X-ray intensity of the high-redshift jet as well as the X-ray-to-radio intensity ratio, which has been observed in a number of systems \citep{Siemiginowska2003,Cheung2004,Cheung2006,Cheung2012,McKeough2016,Simionescu2016}. Although our sample lacks the required count statistics necessary to directly confirm or deny the presence of IC/CMB via spectroscopic modeling (see Section~\ref{sect:flux_extend}), a predicted effect of the elevated IC/CMB jet emission is an increase in the jet flux $f_{\rm jet}$ relative to the quasar core flux $f_{\rm core}$. Thus, we examined the jet-to-core flux ratios from our sample for evidence of redshift dependence in the context of IC/CMB jet emission.
\bibliographystyle{aasjournal}
We estimated the jet-to-core flux ratios $f_{\rm jet}/f_{\rm core}$ for our sample using the observed 0.5--7.0\,keV fluxes in Tables~\ref{table:xray} and \ref{table:flux}, and the ratio results are shown in Table~\ref{table:flux}. The measured flux ratios are between 1.0--3.6\%, with a mean value of 2.2\%. These results are in excellent agreement with the median 2\% $f_{\rm jet}$/$f_{\rm core}$ found by \cite{Marshall2018} for quasars at $z < 2$. The X-ray flux densities we derived consequently do not follow the expected $(1 + z)^4$ dependence for IC/CMB, which is consistent with the conclusion of \cite{Marshall2011}. Our sample also agrees with the $f_{\rm jet}/f_{\rm core}$ results from \cite{Worrall2020} that were based on the few known X-ray jets at $z>3.5$ with good S/N.
Although we did not observe an elevation in X-ray jet flux as a function of redshift, we stress that this result is indicative. It is possible that the quasar core is also dominated by beaming \citep{Worrall1987}, biasing our X-ray ratios. Additionally, the shallow line-of-sight expected for our quasars will cause the innermost region of the jet to appear as part of the core due to the limited spatial resolution of \chandra. This innermost region of the jet has been shown to produce the brightest features in some nearby sources \citep[e.g.,][]{Snios2019b,Snios2019}, which may bias our findings. Our comparison also requires that the high- and low-redshift quasar cores be physically consistent with one another, which may not be valid. Deeper \chandra\ observations are necessary to investigate the spectrum and spatial structure of these systems. For now, we can only reiterate that the flux density jet-to-core ratio derived from our high-redshift sample is consistent with measurements from low-redshift quasars, suggesting that it is independent of redshift.
\subsection{Flux Limits for Undetected X-ray Jets}
\label{sect:undected}
Despite X-ray features being detected from five sources in our quasar sample, the remaining nine sources show no evidence of resolved X-ray structure(s). However, X-ray jets should be cospatial with the radio features observed at GHz frequencies, so it is possible that X-ray emission is present but resides below our detection threshold. We therefore stacked the nine sources with no detections in an effort to measure the average intensity of their X-ray jets.
\begin{table}
\caption{X-ray Jet Properties
\label{table:flux}}
\begin{tightcenter}
\begin{tabular}{c D c D c c c }
\hline \hline
Object & \multicolumn2c{$d_{\rm jet}$} & $C_{\rm src}$ & \multicolumn2c{$f_{\rm jet}$} & $S_{\rm jet}$ & $L_{\rm jet}$ & $f_{\rm jet}$/$f_{\rm core}$\\
(1) & \multicolumn2c{(2)} & (3) & \multicolumn2c{(4)} & (5) & (6) & (7)\\
\hline
J0833$+$0959 & 11.0 & 6.56 & 8.3 & 1.5 & 6.2 & 0.036 \\
J0909$+$0354 & 3.5 & 9.51 & 12.1 & 6.9 & 6.7 & 0.010 \\
J1016$+$2037 & 3.0 & 6.29 & 8.0 & 5.4 & 3.9 & 0.031 \\
J1405$+$0415 & 2.5 & 5.78 & 7.4 & 5.9 & 3.9 & 0.020 \\
J1610$+$1811 & 6.0 & 5.11 & 6.8 & 2.3 & 3.3 & 0.013 \\
\hline
\end{tabular}
\end{tightcenter}
{\footnotesize (1) Object name. (2) Projected jet length, in units of arcseconds. (3) Source 0.5--7.0\,keV counts from the jet. (4) Observed 0.5--7.0\,keV jet flux, in units of $10^{-15}\rm\,erg\,cm^{-2}\,s^{-1}$. (5) Observed 0.5--7.0\,keV jet surface brightness, in units of $10^{-15}\rm\,erg\,cm^{-2}\,s^{-1}\,arcsec^{-2}$ (6) Rest-frame 2--10\,keV jet luminosity, in units of $10^{44}\rm\,erg\,s^{-1}$. (7) Ratio of X-ray fluxes for the jet to core. All flux-related measurements have an approximate factor of two uncertainty.}
\end{table}
Using the annular sectors from Section~\ref{sect:annulus}, we determined for each source which X-ray sector(s) was (were) coaligned with a radio feature. In total, we defined 11 sectors from the nine sources that were coaligned with radio features. Counts from the 11 sectors were summed, giving us values for the observed, simulated PSF, and background counts. In total, we measured 16 observed counts and expect 14.37 simulated PSF counts plus 0.87 background counts. Based on our detection criterion from Section~\ref{sect:annulus}, this gives a false detection probability of 0.456. As this does not satisfy our detection threshold of $p\leq$\,0.005, extended X-rays are not detected from the stacked image.
Given the lack of detection from the stacked analysis, we instead determined the upper limit of the average X-ray jet flux for the nine quasars. We began by estimating an upper limit count rate using the method described in \cite{Kashyap2010}. To detect a source at our defined 0.005 probability for a total background of 15.24 counts (simulated PSF plus field background), we need a minimum of 27 observed counts. We therefore must expect 36.08 counts to ensure that we obtain $\geq 27$ counts to 95\% probability. In total, we estimated a required net of 20.84 counts over the background, or a rate of $1.99\times10^{-4}\rm\,cts\,s^{-1}$.
Having determined an upper limit count rate, we modeled the jet with a {\tt phabs$\cdot$powerlaw} expression. The Galactic column density was fixed to the average value of the nine sources, giving us $N_{\rm H} = 3.02\times10^{20}\rm\,cm^{-2}$. Consistent with Section~\ref{sect:flux_extend}, we fixed the jet photon index to \mbox{$\Gamma=1.9$}. Using \pimms{}v4.11a with our count rate limit of $1.99\times10^{-4}\rm\,cts\,s^{-1}$, we found an upper limit for the average unabsorbed jet flux of $2.26\times10^{-15}\rm\,erg\,cm^{-2}\,s^{-1}$ over the observed 0.5--7.0\,keV band. This estimated flux is consistent with typical background limits for \chandra\ ACIS observations.
With the X-ray jet flux limit now available, we tested for evidence of IC/CMB emission by calculating the average X-ray jet-to-core ratio limit for the nine quasars in our sample with no detected X-ray jets. We measured the jet-to-core flux ratio limit for each quasar and then averaged the values, which we found to be 2.4\%. These results are consistent with the mean 2.2\% flux ratio of the five quasars with X-ray features. The flux ratio is also in agreement with the distribution of jet-to-core ratios found by \cite{Marshall2018}. Overall, these results reinforce the finding of Section~\ref{sect:iccmb} that there is no observed redshift dependence in the X-ray jet emission.
\subsection{Non-Coincident Radio and X-ray Features in J1016$+$2037}
\label{sect:J1016}
Examination of our quasar sample shows that the majority of X-ray features, when detected, are spatially coincident with radio features. However, J1016$+$2037 stands out amongst our sample due to its non-coincident X-ray and radio features, where the measured offset between the X-ray emission and radio feature centroid is $195^{\circ}$. This offset is despite the excellent alignment of the X-rays and radio from its quasar core. Furthermore, the detection probability for the extended X-rays from J1016 is amongst the highest rated for our sample (Table~\ref{table:extend}), suggesting that this feature is indeed real. Thus, the origin of the X-ray feature and its misalignment with the radio merits further discussion.
We began by verifying that the known \chandra\ PSF artifact (see Section~\ref{sect:observation}) was not aligned with the observed X-ray feature in J1016, confirming that the extended X-rays are not due to a systematic effect from \chandra. We then investigated the possibility of an unassociated X-ray source being the origin of the extended X-rays observed in J1016. Our measured X-ray flux for the extended source is $\sim$\,$10^{-14}\rm\,erg\,cm^{-2}\,s^{-1}$ (Table~\ref{table:flux}), and there are $\sim$\,100 X-ray sources\,deg$^{-2}$ above such a flux limit \citep{Civano2016}. As a result, there is only a $\sim$\,1\% probability of a chance association within 5\arcsec\ of any quasar in our sample. We also confirmed that the X-ray feature is not coincident with any optical source in the Pan-STARRS 3Pi survey to a minimum brightness limit of 21.3\,mag in the $grizy$ bands \citep{Tonry2012}. It is therefore likely that the extended X-rays are associated with J1016.
Given that the offset between the radio and X-ray features is $\sim$\,$180^{\circ}$, it is possible that the different emissions may originate from jet/counterjet features. One-sided radio emission from quasars is normally interpreted as jet emission, meaning that the X-rays would be from the counterjet side. However, this scenario is unlikely as the beamed jet will generally have a higher X-ray flux than the counterjet region, and so X-rays should be detected coincident with the observed radio feature. In addition, the measured jet-to-core X-ray flux ratio for J1016 (Table~\ref{table:flux}) is consistent with flux ratios from a beamed jet, suggesting that the observed extended X-rays are due to emission from the near jet. It is therefore unclear from the multiwavelength data what is the orientation of the jetted outflow in J1016, assuming jets are present.
Despite concluding that the extended X-rays in J1016 are associated with the system, we lacked the X-ray count statistics required to determine the physical origin of the observed misalignment. Follow-up, deep-exposure observations with \chandra\ will permit a spectroscopic and morphological analysis of the X-ray structure where different emission models may be investigated to determine the physical origin for this irregular emission feature.
\section{Optical Properties of Quasar Sample}
\label{sect:ox}
\begin{table}
\caption{Optical Properties of the Quasar Sample}
\label{table:alpha}
\begin{tightcenter}
\begin{tabular}{ c c D c }
\hline \hline
Object & $m_{1450\rm\,\textup{\AA}}$ & \multicolumn2c{$\ell_{2500\rm\,\textup{\AA}}$} & $\alpha_{\rm ox}$ \\
(1) & (2) & \multicolumn2c{(3)} & (4) \\
\hline
\decimals
J0801$+$4725 & 19.58 & 6.7 & $-1.61\substack{+0.09\\-0.10}$ \\
J0805$+$6144 & 19.92 & 4.1 & $-1.35\substack{+0.03\\-0.03}$ \\
J0833$+$0959 & 21.09 & 2.3 & $-1.25\substack{+0.06\\-0.06}$ \\
J0909$+$0354 & 19.94 & 4.9 & $-1.28\substack{+0.03\\-0.03}$ \\
J0933$+$2845 & 17.92 & 34.5 & $-1.86\substack{+0.08\\-0.08}$ \\
J1016$+$2037 & 19.12 & 9.1 & $-1.51\substack{+0.05\\-0.05}$ \\
J1128$+$2326 & 18.56 & 14.5 & $-1.76\substack{+0.07\\-0.07}$ \\
J1223$+$5038 & 17.47 & 54.7 & $-1.63\substack{+0.03\\-0.03}$ \\
J1405$+$0415 & 19.97 & 4.4 & $-1.37\substack{+0.04\\-0.04}$ \\
J1435$+$5435 & 20.19 & 5.5 & $-1.72\substack{+0.13\\-0.14}$ \\
J1610$+$1811 & 18.33 & 18.9 & $-1.52\substack{+0.03\\-0.03}$ \\
J1616$+$0459 & 19.19 & 9.1 & $-1.44\substack{+0.04\\-0.04}$ \\
J1655$+$3242 & 19.58 & 6.3 & $-1.74\substack{+0.11\\-0.12}$ \\
J1655$+$1948 & 20.02 & 4.4 & $-1.56\substack{+0.09\\-0.10}$ \\
\hline
\end{tabular}
\end{tightcenter}
(1) Object name. (2) Rest-frame, monochromatic AB apparent magnitude at 1450\,\textup{\AA}, as measured from Pan-STARRS $r$ band. (3) Rest-frame, monochromatic luminosity at 2500\,\textup{\AA}, in units of $10^{31}\rm\,erg\,s^{-1}\,Hz^{-1}$. (4) Optical-to-X-ray power-law slope. Fluxes are assumed to have a 0.1\,mag uncertainty, and $\alphaox$ errors include both optical and X-ray uncertainties.
\end{table}
Previous quasar studies have demonstrated an inverse relationship between the optical-to-X-ray flux ratio and optical luminosity \citep[e.g.,][]{Avni1982,Tananbaum1986,Wilkes1994}. We therefore compiled optical properties of our sample in an effort to investigate the optical and X-ray relationship amongst our sources.
In keeping with our previous quasar analysis \citep{Snios2020b}, we obtained optical fluxes at the rest-frame wavelength of 1450\,\textup{\AA}. For our current sample, this corresponds to observed wavelengths between 5850--7000\,\textup{\AA}. We consequently utilized the monochromatic AB apparent magnitude from the Pan-STARRS $r$ band as it has an effective wavelength of 6241\,\textup{\AA}{} \citep{Tonry2012}. All sources were identified in the Pan-STARRS 3Pi Survey catalog, and no Galactic extinction corrections were applied to our measurements. We additionally extrapolated the rest-frame $m_{1450\rm\,\textup{\AA}}$ results to determine the rest-frame 2500\,\textup{\AA}{} luminosity $\ell_{2500\rm\,\textup{\AA}}$. A UV spectral index of $\alpha = -0.5$ was assumed for our extrapolation ($f_{\nu} \propto \nu^{\alpha}$), which is consistent with prior quasar studies \citep[i.e.][]{Shemmer2006,Nanni2017,Snios2020b}. The optical fluxes and luminosities are shown in Table~\ref{table:alpha}.
Having determined both the optical and X-ray luminosities for the sample, we calculated the optical-to-X-ray power-law index $\alphaox$ for each source. We defined $\alphaox$ the same as \cite{Tananbaum1979}:
\begin{equation}
\frac{{\rm log}(\ell_{\rm 2\,keV}/\ell_{\rm 2500\,\textup{\AA}})}{{\rm log}(\nu_{\rm 2\,keV}/\nu_{\rm 2500\,\textup{\AA}})} = 0.3838\cdot\rm log(\ell_{2\,keV}/\ell_{\rm 2500\,\textup{\AA}}),
\end{equation}
where $\ell_{\rm 2\,keV}$ and $\ell_{\rm 2500\textup{\AA}}$ are the monochromatic luminosities at 2\,keV and 2500\,\textup{\AA}, respectively. Consistent with \cite{Snios2020b}, the $\alphaox$ error was derived by summing the X-ray luminosity uncertainty in quadrature with an approximated 10\% optical uncertainty, or $\sim$0.1 mag. Results for $\alphaox$ are provided in Table~\ref{table:alpha}.
\subsection{Optical and X-Ray Relationship}
\begin{figure}
\begin{tightcenter}
\includegraphics[width=0.99\linewidth]{jet_alpha_ox.eps}
\end{tightcenter}
\caption{Optical-to-X-ray power-law slope $\alphaox$ vs. UV luminosity $\ell_{2500\rm\,\textup{\AA}}$. Results from this work (black) are compared against other quasar samples from the literature. The black dotted line is the best-fit model, while the gray region is the $3\sigma$ confidence level. Previous measurements are taken from \cite{Shemmer2006,Just2007,Lusso2016,Siemiginowska2016,Nanni2017,Martocchia2017,Zhu2019,Vito2019b, Snios2020b}.}
\label{fig:alpha}
\end{figure}
\begin{table}
\caption{Radio Properties of the Sample}
\label{table:radio}
\begin{tightcenter}
\begin{tabular}{ c D D c }
\hline \hline
Object & \multicolumn2c{$f_{\rm5\,GHz}$} & \multicolumn2c{$\ell_{\rm5\,GHz}$} & log($\ell_{\rm5\,GHz}$/$\ell_{\rm2\,keV}$)\\
(1) & \multicolumn2c{(2)} & \multicolumn2c{(3)} & (4) \\
\hline
J0801$+$4725 & 89~ & 6.7 & $6.2\pm0.3$ \\
J0805$+$6144 & $798^{*}$ & 50.4 & $6.6\pm0.2$ \\
J0833$+$0959 & 150~ & 15.5 & $6.1\pm0.2$ \\
J0909$+$0354 & 125$^{*}$ & 9.6 & $5.6\pm0.2$ \\
J0933$+$2845 & $76^{*}$ & 6.4 & $6.1\pm0.3$ \\
J1016$+$2037 & 808~ & 54.4 & $6.7\pm0.2$ \\
J1128$+$2326 & 174~ & 11.1 & $6.5\pm0.2$ \\
J1223$+$5038 & 265~ & 23.7 & $5.9\pm0.2$ \\
J1405$+$0415 & 1187~ & 85.9 & $6.8\pm0.2$ \\
J1435$+$5435 & 115~ & 12.5 & $6.8\pm0.5$ \\
J1610$+$1811 & 239~ & 16.2 & $5.9\pm0.2$ \\
J1616$+$0459 & $393^{*}$ & 28.6 & $6.2\pm0.2$ \\
J1655$+$3242 & 201~ & 14.4 & $6.9\pm0.4$ \\
J1655$+$1948 & $194^{*}$ & 14.6 & $6.6\pm0.3$ \\
\hline
\end{tabular}
\end{tightcenter}
(1) Object name. (2) Rest-frame 5\,GHz flux density from the FIRST VLA survey \citep{White1997}, in units of $\rm mJy$. In cases where no FIRST detection was found (denoted with a ${}^{*}$), flux densities were measured from 1.4\,GHz VLA observations. (3) Monochromatic luminosity at 5\,GHz, in units of $10^{33}\rm\,erg\,s^{-1}\,Hz^{-1}$. (4) Rest-frame radio-to-X-ray ratio. Flux densities and luminosities are assumed to have a 15\% uncertainty. Radio-to-X-ray ratio errors include both radio and X-ray uncertainties.
\end{table}
Prior studies of quasars have shown an anticorrelation between $\alphaox$ and $\ell_{2500\textup{\AA}}$ that is independent of redshift \citep{Bechtold1994,Vignali2003, Steffen2006, Kelly2007, Nanni2017,Snios2020b}, so we investigated if our radio-loud quasar sample was consistent with this relationship. In keeping with the method from \cite{Snios2020b}, we selected quasars from the literature with known optical and X-ray properties. We included 16 targets from \cite{Shemmer2006}, 34 from \cite{Just2007}, 2153 from \cite{Lusso2016}, 18 from \cite{Nanni2017}, 35 from \cite{Martocchia2017}, 15 from \cite{Zhu2019}, 7 from \cite{Vito2019b}, and 15 from \cite{Snios2020b}. In total, we compiled a sample of 2307 quasars with known $\alphaox$ and $\ell_{2500\textup{\AA}}$ parameters for our analysis.
The resulting $\alphaox$\,-\,$\ell_{2500\textup{\AA}}$ relationship from the complete dataset is shown in Figure~\ref{fig:alpha}. Using a linear regression with the {\tt scipy} Python package \citep{Virtanen2020}, we found a best-fit relation of
\begin{equation}
\label{eq:alpha1}
\alphaox = (-0.150\pm0.005)\,{\rm log}(\ell_{2500\textup{\AA}}) + (3.0\pm0.1),
\end{equation}
for the total sample, where the errors are reported to 1$\sigma$. Our best fit is consistent with previous studies to within 1$\sigma$ \citep{Nanni2017,Snios2020b}. Additionally, we verified that the best-fit relationship is independent of redshift, which is consistent with previous works \citep[i.e.,][]{Just2007, Snios2020b}. We repeated the analysis using only the 14 quasars from this work, finding a slope of $-0.3\pm0.1$ and an intercept of $9\pm4$, where the best fit was determined using an orthogonal distance regression in {\tt scipy}. Thus, the quasar sample discussed in this paper is consistent within $2\sigma$ to the overall trend for the broader quasar population.
\subsection{Impact of Radio Loudness on Optical Properties}
\label{sect:op-rl}
\begin{figure}
\begin{tightcenter}
\includegraphics[width=0.99\linewidth]{radio_xray.eps}
\end{tightcenter}
\caption{Comparison of rest-frame radio ($L_{\rm5\,GHz}$) and X-ray ($L_{\rm 2-10\,keV}$) luminosities. The black dotted line is the best-fit relation, while the gray region is the $1\sigma$ confidence level. Radio-to-X-ray ratios are broadly consistent for our sample.}
\label{fig:radio}
\end{figure}
Radio-loud quasars generally have a higher X-ray luminosity than radio-quiet quasars for a given optical luminosity \citep[e.g.,][]{Worrall1987, Wu2013, Zhu2019}. Given that our $\alphaox$\,-\,$\ell_{2500\textup{\AA}}$ analysis includes both radio-loud and radio-quiet quasars, we investigated the impact of radio loudness on the $\alphaox$\,-\,$\ell_{2500\textup{\AA}}$ relationship. Our sample of 2307 quasars was separated into radio-loud and radio-quiet populations, where the radio-loud dataset includes all sources from \cite{Zhu2019}, \cite{Snios2020b}, and this work. This gave us a total 44 radio-loud sources and 2263 radio-quiet sources for our analysis.
Using the radio-loud dataset, we found a linear relationship between $\alphaox$\,-\,$\ell_{2500\textup{\AA}}$ with a best-fit slope of $-0.33\pm0.06$ and an intercept of $9.1\pm1.8$. Repeating our analysis with the radio-quiet sample, we found a slope of $-0.157\pm0.005$ and an intercept of $3.2\pm0.1$. The radio-loud sample therefore diverges from the radio-quiet sample by $\sim$\,$3\sigma$.
Because we previously verified that our best fit was independent of redshift, the observed discrepancy between the radio-loud and radio-quiet sample is not due to differences in redshift selection criteria.
Despite the observed impact of radio-loudness on the $\alphaox$\,-\,$\ell_{2500\textup{\AA}}$ relationship, we stress that this result may be biased and/or incomplete due to our small sample size of radio-loud quasars. The available radio-loud quasar sample lacks the comprehensive coverage of X-ray and UV luminosities needed to accurately fit a relationship for a broad range of sources. Further study with a larger radio-loud sample is required in order to verify the dependence of radio loudness on $\alphaox$ properties.
\section{Radio Properties of Quasar Sample}
\label{sect:radio}
Given that our quasar sample is comprises of sources detected in both X-rays and radio, we investigated the relation between their radio and X-ray luminosities. Radio flux densities for quasars discussed in this work were provided from the VLA-FIRST survey \citep{White1997}, which observed at 1.4\,GHz. In cases where no FIRST detection was found, flux densities were directly measured from archival 1.4\,GHz VLA observations. For our sample, these radio measurements correspond to rest-frame frequencies between 5.6--6.7\,GHz, which we assumed to be a reasonable approximation of the rest-frame 5\,GHz flux density. Radio flux densities and luminosities for the sample are provided in Table~\ref{table:radio}. We note that the reported radio luminosities represent the core emission for each source.
The quasars from this sample were added together with the radio-loud sources from \cite{Snios2020b}, giving us 29 radio-loud quasars at $z>3$ with known radio and X-ray fluxes. The radio--X-ray luminosity relationship for these sources is shown in Figure~\ref{fig:radio}. We fit the data using linear orthogonal data regression in {\tt scipy}, and we found a best-fit relation of
\begin{equation}
{\rm log}(L_{\rm5\,GHz}) = (0.73\pm0.22){\rm log}(L_{\rm2-10\,keV}) + (10.3\pm10.3),
\end{equation}
where the reported errors are $1\sigma$. We note that outliers from \cite{Snios2020b} that reside above the best fit are known to be either Compton-thick and/or possess unresolved X-ray structure and are likely not representative of the remaining sample. In comparison, all sources from our current work agree with the measured relationship within 3$\sigma$.
Our measured radio--X-ray luminosity correlation is consistent with the relationship for radio-loud quasars at $z < 2$ \citep{Fan2016}, further reinforcing the lack of spectral evolution over redshift. Additional, in-depth analysis of the radio and X-ray properties of these quasars and fundamental plane results, which is beyond the scope of this work, will be discussed in a forthcoming paper.
\section{Conclusions}
\label{sect:conclude}
We analyzed \chandra\ observations of 14 radio-loud quasars at redshifts $3 < z < 4$, each with radio features in addition to core emission, to measure their X-ray spectral properties and search for evidence of resolved structure. We detected all quasars in the 0.5--7.0\,keV band and extracted emission spectra of the quasar cores. Each spectrum was fit with an absorbed power-law model, where our mean best-fit photon index is \mbox{$1.6\pm0.2$}. Observed X-ray fluxes and rest-frame luminosities were also determined from the spectral best fits. Additionally, we detected an Fe\,K emission line from the quasar J1223$+$5038 at high significance.
We performed a morphological analysis of each X-ray source using \chandra\ observations, and we detected X-ray features at distances up to 12\arcsec{} from the quasar core in five of the sources. The X-ray features are spatially coincident with existing radio features for four of the five sources, suggesting that the majority of the X-ray features are jets. J1016$+$2037 stands out amongst our sample due to a $\sim$\,$180^{\circ}$ misalignment between its X-ray and radio features. We speculated on the cause this observed misalignment, but the available X-ray counts were insufficient to conclusively determine its origin.
Rest-frame 2--10\,keV luminosities of the X-ray jets were estimated for the five quasars with extended X-rays, and their X-ray jet-to-core flux ratios were measured to be up to 3.6\%. We also estimated an upper limit on the average X-ray jet flux for the remaining nine quasars from a stacked image analysis, finding a limit of $2.3\times10^{-15}\rm\,erg\,cm^{-2}\,s^{-1}$. This flux limit corresponds to an average jet-to-core X-ray flux ratio upper limit of 2.4\%. Our measured jet-to-core flux ratios, both for those directly measured and the upper limits, agree well with measurements from low-redshift quasars, suggesting that the observed X-ray jet emission mechanism is independent of redshift. Deeper \chandra\ observations are required to investigate the spectrum and spatial structure of the detected X-ray features.
Beyond our morphological study, we determined the optical-to-X-ray power-law slope $\alphaox$ for each quasar core using optical/UV data available in the literature. We observed a clear anticorrelation trend between $\alphaox$ and $\ell_{2500\textup{\AA}}$, where our derived best-fit relationship is consistent with other quasar surveys. We also measured radio-to-X-ray luminosity ratios for our sources, and our results are broadly consistent with other radio-loud quasar surveys regardless of redshift. These multiwavelength results reinforce that the spectral evolution of quasars is independent of redshift.
Our results demonstrate the strength of high-resolution X-ray imaging in studying quasars at high redshifts. Further sampling of the high-redshift quasar population with high-resolution instruments, such as \chandra\ or the proposed missions AXIS \citep{Mushotzky2019} and Lynx \citep{Gaskin2019}, will allow us to investigate jet properties over a broad redshift range while also identifying extended X-ray features ideal for follow-up morphological and spectroscopic analyses.
\acknowledgements{
We thank the referee for their comments and suggestions. B.S., D.A.S., A.S., and M.S. were supported by NASA contract NAS8-03060 (\chandra\ X-ray Center). B.S. and D.A.S. were also supported by CXC grant GO8-19077X. Work by C.C.C. at the Naval Research Laboratory is supported by NASA DPR S-15633-Y. Radio observations were provided by the National Radio Astronomy Observatory, a facility of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
\software{\ciao v4.12, \caldb v4.9.2.1 \citep{Fruscione2006}, Sherpa \citep{Freeman2001}, {\tt scipy} \citep{Virtanen2020}}
}
\bibliographystyle{aasjournal}
|
\section{Introduction}
\label{S1}
Bound states in the continuum (BICs) are trapped or guided modes with
their frequencies in the radiation continua~\cite{neumann29,hsu16}.
They exist in a variety of photonic structures including periodic
structures sandwiched between two homogeneous
media~\cite{bonnet94,padd00,ochiai01,tikh02,shipman03,lee12,port05,mari08,hsu13_2,bulg14b,gan16},
waveguides with local defects~\cite{evans94,bulg08}, waveguides with
lateral leakage channels~\cite{zou15,bezus18,nguyen19,yu19,byk20},
etc.
In structures that are invariant or periodic in one or two spatial
directions, a BIC is a special point with an infinite quality factor
($Q$ factor) in a band of resonant
states~\cite{hsu13_2,yuan17_2,yuan18,jin19}, and it becomes a
high-$Q$ resonance if the structure is perturbed generically \cite{kosh18,yuan20}.
High-$Q$ resonances lead to strong local field
enhancement~\cite{mocella15,yoon15,huzhen20} and abrupt features in
reflection and transmission spectra~\cite{yuan17}, and are essential
for sensing, lasing, switching and nonlinear optics applications.
For theoretical interest and practical applications, it is important
to understand how a BIC is affected by a perturbation of the
structure.
If the BIC is protected by a symmetry \cite{bonnet94,padd00,ochiai01,tikh02,shipman03,lee12,evans94,bulg08}, i.e., there is a symmetry mismatch between the BIC and the compatible radiation modes, it naturally continues its existence if the perturbation preserves the symmetry \cite{bonnet94,evans94,shipman07}. Therefore, a symmetry-protected BIC is robust with respect to symmetry-preserving perturbations.
If a perturbation breaks the symmetry, a symmetry-protected BIC
typically, but not always, becomes a resonant state with a finite $Q$
factor~\cite{kosh18,yuan20,yuan20b}. In periodic structures sandwiched
between two homogeneous media, there are also BICs with a nonzero
Bloch wavevector, and they propagate in the periodic directions \cite{port05,mari08,hsu13_2,bulg14b,yuan17,maksim,hu18}.
Such a propagating BIC is not protected by symmetry in the usual sense, but can also be robust with respect to symmetry-preserving perturbations~\cite{zhen14,yuan17ol,conrob}.
More precisely, in a periodic structure with an up-down mirror
symmetry and an in-plane inversion symmetry, a generic low-frequency
propagating BIC (with only one radiation channel) continues its
existence
if the structure is perturbed by a perturbation
preserving these two symmetries~\cite{yuan17ol,conrob}.
The BICs in periodic structures exhibit interesting topological
properties. Zhen {\it el al.}~\cite{zhen14} realized that a BIC in a
biperiodic structure is a polarization singularity in momentum
space (the plane of two wavevector components), and defined a
topological charge based on the winding number of the far-field
polarization vector. Since the
far field of a resonant state is typically elliptically polarized,
Bulgakov and Makismov
refined the definition using the major polarization vector~\cite{maksim}. The topological charge can be used to classify the BICs and illustrate their generation, interaction and annihilation processes when structural parameters are tuned \cite{zhen14,maksim}. Importantly, the topological charge
is a conserved quantity that cannot be changed by small structural
perturbations. However, this does not imply that the BICs (with a
nonzero topological charge) are robust with respect to arbitrary
structural perturbations, because a resonant state can be circularly
polarized and also has a nonzero topological charge. In a recent work,
Liu {\it el al.}~\cite{fudan} showed that symmetry-protected BICs in a
photonic crystal slab, protected by the in-plane inversion
symmetry, turn to pairs of circularly polarized resonant states (CPSs)
when the structure is perturbed breaking the symmetry.
Most propagating BICs are found in periodic structures with both the
up-down mirror symmetry and the in-plane inversion symmetry.
It is known that if one of these two symmetries is broken, a
propagating BIC is usually, but not always, destroyed
\cite{hu18,yuan20b}. It has been shown that CPSs can exist in structures
without the up-down mirror symmetry and in-plane inversion
symmetry~\cite{yin20}. In this paper, we show that CPSs emerge
when propagating BICs are destroyed by generic and arbitrarily small
perturbations that break only the in-plane inversion symmetry.
We consider vectorial BICs in a periodic array of of
dielectric cylinders, introduce a small deformation of the cylinder
boundary, and show that pairs of CPSs emerge when propagating BICs
with topological charge $\pm 1$ are destroyed. In addition, we follow
the CPSs as the deformation parameter is further increased, and show
that two CPSs with opposite topological charges and same handedness may merge, and their dependence on the deformation parameter may be multi-valued with self-generation and annihilation points.
The rest of this paper is organized as follows. In Sec.~\ref{S2}, we
recall the mathematical formulation for vectorial eigenmodes
in two-dimensional (2D) periodic structures and the definition of
topological charge. In Sec.~\ref{S2a}, we consider a periodic array of
circular cylinders and briefly describe our computational method.
In Sec.~\ref{S3}, we introduce boundary deformations to the cylinders
and show the generation and annihilation of CPSs. The paper is
concluded with a brief discussion in Sec.~\ref{S4}.
\section{Eigenmodes in periodic structures}
\label{S2}
We consider a 2D structure that is invariant in $x$, periodic in $y$ with
period $L$, bounded in $z$ by $\abs{z}<D$ for some $D>0$,
and surrounded by air (for $|z| > D$), where
$\{x,y,z \}$ is a Cartesian coordinate system.
Let $\varepsilon = \varepsilon(y,z)$ be the relative permittivity of this
structure and its surrounding medium, then
$\varepsilon({\bf r}) =\varepsilon(y+L,z)$ for all ${\bf r}=(y,z)$ and
$\varepsilon({\bf r}) = 1$ for $\abs{z}>D$. We study time-harmonic
electromagnetic waves that depend on time $t$ and variable $x$
as $\exp[ i(\alpha x -\omega t)]$, where $\omega$ is the angular
frequency and $\alpha$ is a real wavenumber in the $x$
direction. From the frequency-domain Maxwell's equations, it
is easy to obtain the following system
\begin{eqnarray}
\label{Eq1}
&& \nabla\cdot\cbrc{ \frac{\varepsilon}{\eta}\nabla E_x} +
\nabla\cdot\cbrc{ \frac{\alpha}{k\eta} P \cdot\nabla
\widetilde{H}_x}+\varepsilon E_x = 0, \quad \\
\label{Eq2}
&& \nabla\cdot\cbrc{ \frac{1}{\eta}\nabla \widetilde{H}_x} -
\nabla\cdot\cbrc{ \frac{\alpha }{k\eta} P\cdot\nabla E_x}+\widetilde{H}_x = 0,
\end{eqnarray}
where $E_x$ is the $x$ component of the electric field,
$\widetilde{H}_x$ is the $x$ component of a scaled magnetic field
(magnetic field multiplied by free space impedance), $k = \omega/c$
is the free space wavenumber, $c$ is the speed of
light in vacuum, and
\begin{equation}
\label{defeta}
\eta = k^2\varepsilon({\bf r}) -\alpha^2, \quad
\nabla = \insMat{ \partial_y \\ \partial_z}, \quad P = \insMat{0& 1\\-1 & 0}.
\end{equation}
The other four field components can be obtained from the following equations:
\begin{eqnarray}
\label{Eq3}
&& \insMat{E_y\\E_z}=\frac{i}{\eta}\cbrc{ \alpha \nabla E_x + k\,
P \cdot \nabla \widetilde{H}_x}, \\
\label{Eq4}
&& \insMat{\widetilde{H}_y\\\widetilde{H}_z}=\frac{i}{\eta}\cbrc{
\alpha \nabla \widetilde{H}_x - k\,\varepsilon P \cdot \nabla E_x}.
\end{eqnarray}
If $\alpha = 0$, the equations for $E_x$ and $\widetilde{H}_z$ are
decoupled, we say the corresponding wave is scalar. The case $\alpha
\ne 0$ is referred to as vectorial.
Due to the periodicity in $y$, any eigenmode of the structure is a
Bloch mode with an electric field given by
\begin{equation}
\label{bloch}
{\bf E}({\bf r}) = {\bf F}({\bf r}) e^ {i\beta y},
\end{equation}
where $\beta$ is a real Bloch wavenumber satisfying $\abs{\beta}\leq
\pi/L$, and ${\bf F}$ is a periodic in $y$ with period $L$. For
$\abs{z}>D$, the field can be expanded in plane waves as
\begin{equation}
\label{planewave}
{\bf E}({\bf r}) = \sum_{m= -\infty }^{\infty} {\bf c}_m^{\pm} e^{i
(\beta_m y\pm \gamma_m z)},\quad \pm z> D,
\end{equation}
where
$\beta_0 = \beta$, and
\begin{equation}
\label{defbeta}
\beta_m = \beta+ \frac{2\pi m}{L}, \quad \gamma_m = \sqrt{ k^2 -\alpha^2
-\beta_m^2}.
\end{equation}
A guided mode satisfies the condition ${\bf E} \to {\bf 0}$ as $z\to \pm
\infty$. For a positive $k$, guided modes usually
exist when $ k < \sqrt{ \alpha^2 + \beta^2}$, so that all $\gamma_m$ are
pure imaginary and all plane waves in the right hand side of Eq.~(\ref{planewave}) are evanescent
in $z$. A BIC is also a guided mode, but it satisfies the condition
$k > \sqrt{\alpha^2 + \beta^2}$, thus $\gamma_0$ and probably a few
other
$\gamma_m$ are real positive. We study BICs with a real
$k$, a real $\alpha$, and a real $\beta$ such that
\begin{equation}
\label{onerc}
\sqrt{ \alpha^2 + \beta^2} < k <
\sqrt{ \alpha^2 + \left( \frac{2\pi}{L} -|\beta| \right)^2}.
\end{equation}
The above condition implies that $\gamma_0$ is positive and all other
$\gamma_m$ are pure imaginary. Since the
plane waves $\exp[ i (\beta_0 y \pm \gamma_0 z)]$ can propagate to
infinity, the coefficients ${\bf c}_0^{\pm}$ of the BIC must vanish.
A resonant state is also an eigenmode, but it satisfies an outgoing
radiation condition as $z \to \pm \infty$. Because of the radiation
loss, the amplitude of a resonant state decays in time, thus the
frequency $\omega$ or $k$ must be complex with a negative imaginary
part. This implies that the real and imaginary parts of $\gamma_0$ are
positive and negative, respectively, and $\exp[ i (\beta_0 y +
\gamma_0 z)] $ is an amplifying outgoing plane wave as $z \to +
\infty$. We are concerned with resonant states with only one radiating
plane wave for either $z > D$ or $z < -D$. Therefore, it is assumed
that $\mbox{Im}(\gamma_m) > 0$ for all $m \ne 0$.
Resonant states form bands where each band corresponds to $k$ being a
complex-valued function of $\alpha$ and $\beta$.
The $Q$ factor of a resonant mode is given by $Q =
-0.5\mathrm{Re}(k)/\mathrm{Im}(k)$. A BIC is a special
point (with a real $k$) in a band of resonant states.
We assume the periodic structure has an up-down mirror symmetry, i.e.,
$\varepsilon({\bf r}) = \varepsilon(y,-z)$ for all ${\bf r}$, then
it is sufficient to study the field of any eigenmode in the upper half
space ($z>0$), because the field components are either even in $z$ or
odd in $z$. For a BIC with a frequency and wavevector satisfying Eq.~(\ref{onerc}),
it is possible to define a topological charge based on the far field
polarization vector of the surrounding resonant states
\cite{zhen14,maksim}. Let
$\mathcal{C}$ be a closed contour
in the $\alpha$-$\beta$ plane. Each point on
$\mathcal{C}$ corresponds to a resonant
state in a band that contains the BIC. The resonant state contains a
far-field outgoing plane wave (for $z \to +\infty$) with a vector
amplitude ${\bf c}_0^+$. Its projection on the $x$-$y$ plane is
\begin{equation}
\label{projected}
\widehat{\bfm{E}} =
\insMat{ c^+_{0x} \\ c^+_{0y} \\ 0}
e^{ i (\alpha x + \beta y + \gamma_0 z)},
\end{equation}
where $c^+_{0x}$ and $c^+_{0y}$ are the $x$ and $y$ components of
${\bf c}_0^+$. For any fixed $z> D$, the real projected electric field
$\mbox{Re} ( \widehat{\bf E} e^{-i \omega t})$ is typically elliptically
polarized,
and the major polarization vector (along the major axis of the
polarization ellipse) forms an angle
$\theta$ with the $x$ axis. If it is possible to define $\theta$
continuously as $(\alpha, \beta)$ traverses along $C$ in
counterclockwise direction from a starting
point back to the same point (the ending point), and $\theta|_{\rm end} - \theta|_{\rm
start} = 2\pi q$ for some $q$, where $\theta|_{\rm start}$ and $\theta|_{\rm end}$
are the values of $\theta$ at the starting point and the ending point
respectively, then $q$ is the winding number (of the projected major polarization
vector) on $\mathcal{C}$. Alternatively, $q$ can be evaluated by
the integral formula
\begin{equation}
\label{Eq9}
q = \frac{1}{2\pi}\oint_\mathcal{C} d {\bm \alpha} \cdot
\nabla_{\bm \alpha} \theta({\bm \alpha}),
\end{equation}
where ${\bm \alpha} = (\alpha,\beta)$ and $\nabla_{\bm \alpha} =
(\partial_\alpha, \partial_\beta)$.
A BIC is a polarization singularity, since it does not have a far
field (${\bf c}_0^\pm = 0$). The
topological charge of a BIC is defined as the winding number $q$, if
$\mathcal{C}$ is sufficiently close to the BIC and encloses the BIC in
the $\alpha$-$\beta$ plane \cite{zhen14,maksim}.
It is important to note that the major polarization vector or the
angle $\theta$ is undefined, if the projected far-field plane wave is circularly
polarized. This implies that a circularly polarized state (CPS),
i.e., a resonant state with a circularly polarized far field, is also
a polarization singularity. If $\mathcal{C}$ contains the wavevector
of a CPS, the winding number on $\mathcal{C}$ is undefined. The
definition of topological charge requires the underlying
assumption that no CPSs exist in a small neighborhood (in the
$\alpha$-$\beta$ plane) of the BIC.
Finally, we recall that the polarization state of the plane wave given
by Eq.~(\ref{projected}) can be characterized by the Stokes
parameters~\cite{Stokes}
\begin{align}
\mathcal{S}_0 & = |c_{0x}^+|^2+ |c^+_{0y}|^2,\quad
\mathcal{S}_1 = |c_{0y}^+|^2- |c^+_{0x}|^2,\\
\mathcal{S}_2 &= 2 |c^+_{0x} c^+_{0y}| \cos(\varphi_y-\varphi_x),\\
\mathcal{S}_3 & = 2 |c^+_{0x} c^+_{0y}| \sin(\varphi_y-\varphi_x),
\end{align}
where $\varphi_x = \arg(c^+_{0x} )$, $\varphi_y = \arg(c^+_{0y} )$ are
the phases of $c_{0x}^+$ and $c_{0y}^+$, respectively. Since a BIC has
no radiation, it corresponds to the point $\cbrc{\mathcal{S}_0,\mathcal{S}_1,\mathcal{S}_2,\mathcal{S}_3=0}$. A
CPS satisfies the condition $(\mathcal{S}_1,
\mathcal{S}_2 , \mathcal{S}_3/\mathcal{S}_0)= (0,0, \pm 1)$, where the
ellipticity $ \mathcal{S}_3/\mathcal{S}_0 = +1$ for the left CPS
and $-1$ for right CPS. A linearly polarized state is any point with
$\mathcal{S}_3=0$. The principle value of angle $\theta$
is given by $\theta = \arg (\mathcal{S}_1 +i \mathcal{S}_2)/2$.
\section{Circular cylinders}
\label{S2a}
A periodic array of circular dielectric cylinders surrounded by air,
as shown in Fig.~\ref{fig_carray},
\begin{figure}[htb]
\centering
\includegraphics[width=1\linewidth]{figures/Fig1}
\caption{A periodic array of circular dielectric cylinders
surrounded by air. The array is periodic in $y$ with period
$L$. The cylinders are parallel to the $x$ axis. The radius
and dielectric constant of the cylinders are $a$ and
$\varepsilon_c$, respectively.}
\label{fig_carray}
\end{figure}
is a simple structure supporting many different BICs~\cite{shipman03,shipman07,bulg14b,yuan17,maksim}.
In general, a BIC may be a standing wave ($\alpha=\beta=0$), may
propagate along the $y$ axis ($\alpha=0$ and $\beta \ne 0$),
along the $x$ axis ($\alpha \ne 0$ and $\beta = 0$), or in both $x$
and $y$ directions ($\alpha \ne 0$ and $\beta \ne 0$). Those BICs with
$\alpha=0$ are scalar ones with either the $H$ or $E$
polarization. The BICs with $\alpha\ne 0$ are vectorial ones with
nonzero $E_x$ and $\widetilde{H}_x$.
In Table~\ref{Tab1},
\begingroup
\squeezetable
\begin{table}[htb]
\caption{A few BICs in periodic arrays of circular cylinders
with radius $a$ and dielectric constant $\varepsilon_c = 15$.}
\centering
\begin{ruledtabular}
\begin{tabular}{llcccc }
& $a/L$ & $\alpha L$ & $\beta L$ &$k L$& $q$\\\hline
BIC$_1$ & $0.45042$ & $0$ & $0$ &$3.512894$&$-1$\\
BIC$_2$ & $ 0.45042$& $0.584168$& $0$ &$3.523918$&$+1$\\
BIC$_3$ & $ 0.45$ & $0$ & $0.479401$ &$3.471973$&$+1$\\
BIC$_4$ &$ 0.3 $ & $1.190401$& $0.96714$ &$3.196556$&$+1$\\
\end{tabular}
\end{ruledtabular}
\label{Tab1}
\end{table}
\endgroup
we list four BICs for periodic arrays with a fixed dielectric constant $\varepsilon_c = 15$. The topological charges of
the BICs are listed in the last column. More BICs in this periodic
array can be found in Ref.~\cite{maksim}.
Normally, the resonant states and the BICs are computed by solving an
eigenvalue problem for the Maxwell's equations. To take advantage of
the special geometry of the circular cylinders, we use a semi-analytic
method based on cylindrical wave expansions. Since the structure is
invariant in $x$ and periodic in $y$, it is sufficient to solve the
eigenmodes in one period of the structure, i.e., a 2D domain
$\Omega_{\infty}$ given by $|y| < L/2$ and
$|z| < \infty$. Inside $\Omega_\infty$, there is a square $\Omega$
given by $|y| < L/2$ and $|z| < L/2$. We assume one cylinder is
located at the center of $\Omega$. For given $\alpha$ and $\omega$,
the electromagnetic field in $\Omega$ can be expanded in vectorial cylindrical
waves with unknown coefficients~\cite{yumao}. If $\beta$ is also specified, we can
expand the electromagnetic field in plane waves (also with unknown
coefficients) for $|z| > L/2$. Relating the field at $y=\pm L/2$
by the quasi-periodic condition, assuming even or odd symmetry in $y$,
and imposing continuity conditions at $z= L/2$, we can obtain an
operator $\mathscr{A}$, such that
\begin{equation}
\label{EqOP}
\mathscr{A}(k, \alpha,\beta) \, \mathbf{u}|_{z=L/2} = {\bf 0},
\end{equation}
where ${\bf u}$ is a column vector for $E_x$ and
$\widetilde{H}_x$, and ${\bf u}|_{z=L/2}$ denotes ${\bf u}$ at $z=L/2$
for $|y| < L/2$. Notice that $\mathscr{A}$ is an operator that
acts on a vector of two single-variable functions.
Since $\mathscr{A}$ depends on $k$, Eq.~(\ref{EqOP}) is a nonlinear
eigenvalue problem. In practice, $y \in (-L/2, L/2)$ is sampled by $N$ points,
${\bf u}|_{z=L/2}$ becomes a vector of length $2N$, and $\mathscr{A}$ is
approximated by a $(2N) \times (2N)$ matrix. The method can be
extended to the case where the boundary of the cylinders are slightly
and smoothly deformed. Details are given in Appendix.
For computing resonant states, $\alpha$ and $\beta$ are given, we look
for a complex $k$ such that Eq.~(\ref{EqOP}) has a nontrivial
solution. One possible approach is to solve $k$ from
\begin{equation}
\label{EqLam}
\lambda_1(\mathscr{A}) = 0,
\end{equation}
where $\lambda_1$ is the eigenvalue of $\mathscr{A}$ with the smallest
magnitude. A BIC is a special resonant state with $\mathrm{Im}(k)= 0$. For
propagating BICs, it is more efficient to treat $\alpha$ and/or $\beta$
also as unknowns. We can solve Eq.~(\ref{EqLam}) for a real $k$, a
real $\alpha$, and/or a real $\beta$.
\section{Slightly noncircular cylinders}
\label{S3}
In this section, we consider a periodic array of slightly
noncircular cylinders, study the emergence of CPSs when propagating
BICs are destroyed, and also the annihilation and generation of
CPSs. We assume
the boundary of the cylinder centered at the
origin is given by
\begin{equation}\label{EqX}
y = - a\sin(\tau)+\delta \cos(g\tau), \quad
z = a\cos(\tau),
\end{equation}
for $0\leq \tau < 2\pi$ and $g = 2$ or $4$, where $\delta>0$ is a
small deformation parameter. The dielectric constant of the cylinders
is fixed at $\varepsilon_c = 15$. The deformed cylinders for $\delta = 0.1a$
are shown in Figs.~\ref{Fig2}(a) and \ref{Fig2}(b)
\begin{figure}[htb]
\centering
\includegraphics[width=1\linewidth]{figures/Fig2}
\caption{Cross-sections $\Omega_1$ of two deformed cylinders
with a boundary given by Eq.~\eqref{EqX} for $\delta =
0.1a$ and the cases (a) $g = 2$ and (b) $g=4$.
$\Omega_2$ is the exterior domain outside the cylinder and
inside the square $\Omega = (-L/2,L/2) \times (-L/2,L/2)$.}
\label{Fig2}
\end{figure}
for $g=2$ and $g=4$, respectively. The small boundary deformation is a
perturbation that breaks the
reflection symmetry in $y$ (the in-plane inversion
symmetry), but preserves the reflection symmetry in $z$ (the up-down
mirror symmetry).
To calculate vectorial resonant states, we extend the method described
in the previous section. Importantly, a general electromagnetic field
in $\Omega$
(the square domain containing one cylinder centered at the origin) can
still be expanded in cylindrical waves, although the expansions are
more complicated due to the deformation of the cylinder
boundary. As shown in Appendix, the eigenvalue problem for resonant
states is reduced to Eq.~(\ref{EqOP}), where $\mathscr{A}$ is a $(2N)
\times (2N)$ matrix depending on $k$, $\alpha$ and $\beta$, and $N$
is the number of sampling points for an interval of length $L$.
It is highly desirable to compute a CPS without calculating all nearby resonant
states. To find a left or right CPS, we solve a real $\alpha$, a real
$\beta$ and a complex $k$, from Eq.~(\ref{EqLam}) and
\begin{equation}
\label{Eq15}
\mathcal{S}_1 = 0, \quad \mathcal{S}_3/\mathcal{S}_0 = \pm 1,
\end{equation}
where $\mathcal{S}_0$, $\mathcal{S}_1$ and $\mathcal{S}_3$ are the
Stokes parameters.
First, we show that when the deformation parameter $\delta$ is
increased from zero, all four BICs listed in Table~\ref{Tab1} are
destroyed and pairs of CPSs emerge. If the topological charge of the
BIC is 1 (or $-1$), a pair of CPSs with topological charge $1/2$ (or
$-1/2$) and different handedness emerge. The net topological
charge is conserved.
In Fig.~\ref{Fig6},
\begin{figure}[htb]
\centering
\includegraphics[width =\linewidth]{figures/Fig6}
\caption{A pair of CPSs with topological charge $1/2$
emerged from BIC$_3$ for a periodic array of slightly
noncircular cylinders with $g = 4$.}
\label{Fig6}
\end{figure}
we show the emergence of CPSs from BIC$_3$ for a periodic array with
$a=0.45L$ and $g = 4$.
The purple arrows indicate the direction of increasing $\delta$. The
curve shows the wavevector $(\alpha, \beta)$ for a pair of left and
right CPSs, and it is shown in color to indicate the value of
$\delta$.
These two CPSs exhibit a symmetry with respect to the $\beta$
axis. For each left CPS with wavevector $(\alpha, \beta)$, there is
also a right CPS with wavevector $(-\alpha, \beta)$. The complex
frequencies of these two CPSs are exactly the same.
In Figs.~\ref{Fig7}(a) and \ref{Fig7}(b),
\begin{figure}[htb]
\centering
\includegraphics[width =\linewidth]{figures/Fig7}
\caption{Pairs of CPSs emerged from BIC$_4$ for periodic
arrays of slightly noncircular cylinders with (a) $g = 2$
and (b) $g = 4$.}
\label{Fig7}
\end{figure}
we show CPSs emerged from BIC$_4$ for periodic arrays with $g = 2$ and
$g=4$, respectively. The radius of the original cylinders is $a =
0.3L$.
There is no apparent symmetry between the two CPSs emerged from the
BIC, but there is still a symmetry with respect to the $\beta$ axis.
Since the structure is invariant in $x$, the resonant states (or BICs) for
$(\alpha, \beta)$ and
$(-\alpha, \beta)$ are reflections of each other.
The CPSs with wavevectors $(\alpha, \beta)$ and $(-\alpha, \beta)$
have the opposite handedness.
Next, we show that as $\delta$ is increased, the CPSs emerged from
BICs of opposite topological charge may collapse to a CPS with a zero
charge. BIC$_1$ and BIC$_2$ in Table~\ref{Tab1} are found in the same
periodic array of circular cylinders with radius $a= 0.45042L$, but
their topological charges are $-1$ and $+1$, respectively.
As $\delta$ is increased from zero, both BIC$_1$ and BIC$_2$ are
destroyed. In Figs.~\ref{Fig3}(a) and \ref{Fig3}(b),
\begin{figure}[htb]
\centering
\includegraphics[width =\linewidth]{figures/Fig3}
\caption{Emergence of CPSs from BIC$_1$ and BIC$_2$, and
annihilation of CPSs of opposite charges for periodic arrays
of slightly noncircular cylinders with (a) $g = 2$ and (b) $g = 4$.}
\label{Fig3}
\end{figure}
we show the emergence of CPS pairs from BIC$_1$ and BIC$_2$ for
periodic arrays of deformed cylinders with $g =2$ and $g=4$, respectively.
The CPSs emerged from BIC$_1$ and BIC$_2$ carry
topological charges $-1/2$ and $1/2$, respectively. As $\delta$
reaches a critical value $\delta_{\rm AP}$
($\delta_{\rm AP} \approx 7.586 \times 10^{-5}L$ for $g=2$ and $\delta_{\rm
AP} \approx 3.736 \times 10^{-5}L$ for $g=4$), two CPSs with the same
handedness, one from BIC$_1$ and the other from BIC$_2$, collapse to a
CPS with a zero topological charge. Since these two CPSs cease to
exist for $\delta > \delta_{\rm AP}$, we call $\delta_{\rm AP}$
an annihilation point (AP).
Finally, we show that CPSs connected to a single BIC may encounter
self-generation and self-annihilation points as $\delta$ is increased.
For a periodic array with $a = 0.45L$ and $g=2$, a continuous branch
of CPSs emerging from BIC$_3$ is shown in Fig.~\ref{Fig4}(a).
\begin{figure}[htb]
\centering
\includegraphics[width =\linewidth]{figures/Fig45}
\caption{(a) A continuous right CPS emerged from BIC$_3$ for a
periodic array with $g = 2$, with $\delta$ increased from $0$ to
$\delta_{AP}$, decreased to $\delta_{GP}$, and increased
again. The topological charge of the CPS has changed from $1/2$ to
$-1/2$ and back to $1/2$ accordingly. (b) The $Q$ factor, (c) $\alpha$ and (d) $\beta$ of the right CPS as multivalued functions of $\delta$. The blue and red curves
correspond to topological
charge $1/2$ and $-1/2$, respectively. }
\label{Fig4}
\end{figure}
The curve in Fig.~\ref{Fig4}(a) depicts the wavevector
$(\alpha, \beta)$ of a CPS when $\delta$ is first increased from zero
to $\delta_{\rm AP} \approx 0.0077L$, then decreased
to $\delta_{\rm GP} \approx 0.0070L$, finally increased again. The
topological charge of the CPS has changed from $1/2$ to $-1/2$, and
back to $1/2$ accordingly.
In the process of an increasing $\delta$, the critical value
$\delta_{\rm AP}$ is annihilation point where two CPSs
of opposite topological charge and same handedness collapse to a CPS
with a zero charge, but these two CPSs are not connected
to different BICs. In fact, the CPS with topological charge $-1/2$
and another CPS with topological charge $1/2$ emerge from a CPS of
zero topological charge at $\delta_{\rm GP}$. We call $\delta_{\rm
AP}$ a self-annihilation point and $\delta_{\rm GP}$ a
self-generation point. However, annihilation and generation are
relative terms depending on how the structure is tuned. For example,
$\delta_{\rm GP}$ can also be
regarded as a self-annihilation point, if the structure is tuned by a
decreasing $\delta$. It is clear that the continuous branch of CPSs
emerged from BIC$_3$, as shown in Fig.~\ref{Fig4}(a), exhibits a multi-valued
dependence on $\delta$. In
Figs.~\ref{Fig4}(b), (c) and (d),
we respectively show the $Q$ factor, and wavevector components $\alpha$ and $\beta$
as multivalued functions of $\delta$. Despite the generation and
annihilation of CPSs as $\delta$ is varied, the net topological
charge is conserved and remains at $1/2$.
\section{Conclusion}
\label{S4}
Many applications of BICs are realized in periodic structures
sandwiched between two homogeneous media. It is known
that the existence and robustness of BICs, including the
propagating BICs in periodic structures, depend crucially on
symmetry \cite{hsu13_2,zhen14}. More precisely, it has been proved
that some propagating BICs
are robust with respect to structural perturbations that preserve the
in-plane inversion symmetry and the up-down reflection symmetry, even
though these BICs do not have a symmetry mismatch with compatible
radiating waves \cite{yuan17ol,conrob}. If the perturbation breaks one of these two
symmetries, the propagating BICs are typically destroyed and become
resonant states with a finite $Q$ factor \cite{hu18}.
The topological charge of a BIC, defined using the polarization vector
of the resonant states surrounding the BIC in momentum space, is an
interesting concept and is useful for understanding the evolution of BICs as
structural parameters are varied \cite{zhen14,maksim}.
The definition of topological charge does not require the in-plane inversion symmetry. When structural parameters are
varied, the topological charge is always conserved, but this
does not imply that the BICs are robust with respect to arbitrary
structural perturbations, because a CPS is also a polarization
singularity in momentum space and it can have a nonzero topological
charge. Therefore, the conservation of topological charge is only
valid when both BICs and CPSs are included. For
propagating BICs, the connection between symmetry-breaking
perturbations and the emergence of CPSs has not been clearly
established in existing literature. Using a periodic array of slightly
noncircular cylinders as an example, we show that pairs of CPSs emerge when
propagating BICs are destroyed by arbitrarily small perturbations that
break only the in-plane inversion symmetry. We also study the
generation and annihilation of CPSs when structural parameters are
varied. It is shown that pairs of CPSs of opposite topological charge
can collapse at special CPSs with a zero topological charge, but the net
topological charge is still conserved.
\section*{Acknowledgments}
The authors acknowledge support from the Research Grants Council of
Hong Kong Special Administrative Region, China (Grant No. CityU 11305518).
|
\section{Introduction}
Railway companies have been working on defect detection of railway equipment. However the equipment is replaced early to ensure the smooth operation of the railway with safety as a priority, leading to railway companies facing a common issue of having a lack of images of defective equipment.\\
We developed a defect detection model that is able to identify railway defects by utilizing the images captured by the railway monitoring system\cite{jreMonitoring} installed on the commercial trains of East Japan Railway Company. However, we faced a problem in which the detection accuracy could not be improved due to the lack of training data for a certain type of defects.
\section{Defect detection pipeline for multiple types of track defects using ResNet\cite{he2015deep}}
The sample images of scratch, inside-scratch and shelling defect are shown in Fig.\ref{fig:scratches}. The number of images used for training the defect detection model are shown in Table\ref{tab:defects}. It is evident that the data of shelling defect is considerably less compared to the scratch and inner side scratch defect.
\begin{figure}[ht]
\includegraphics[width=\columnwidth]{figures/scratches.png}
\caption{Railway scratches (L to R: normal, scratch, inside-scratch and shelling defects)}
\label{fig:scratches}
\end{figure}
\begin{table}[ht]
\begin{tabularx}{\columnwidth}{|X|X|X|X|X|}
\hline
\textbf{} & \textbf{Normal} & \textbf{Scratch Defect} & \textbf{Inside-scratch Defect} & \textbf{Shelling Defect}\\
\hline Training data & 4500 & 992 & 1160 & 82 \\
\hline Test data & 100 & 50 & 50 & 10 \\
\hline
\end{tabularx}
\caption{Number of images used for defect detection model}
\label{tab:defects}
\end{table}
Fine-tuning is a method of relearning the weights of a neural network by replacing the final output layer with an existing model trained in a different domain. The structure of the anomaly detection model is shown in Fig.\ref{fig:resModel}.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/Resnet_model_JEIS_02.png}
\caption{The structure of the anomaly detection model}
\label{fig:resModel}
\end{figure}
The defect detection model was unable to identify 1 incident of a superficial shelling defect. The accuracy of the defect detection model is shown in Fig.\ref{fig:beforeAcc} while the image of the shelling that was not detected is shown in Fig.\ref{fig:shellingImage}.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/beforeAcc.png}
\caption{Accuracy of defect detection model\\(0: normal, 1:scratch, 2: inside-scratch, 3: shelling defects)}
\label{fig:beforeAcc}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=0.25\columnwidth]{figures/shelling.png}
\caption{Image of shelling that was not detected}
\label{fig:shellingImage}
\end{figure}
\section{Railway defect generation model using CycleGAN}
Due to the lack of adequate training data in order to generate images of railway defects, we have chosen an algorithm known as CycleGAN. CycleGAN is a type of style transfer method where an image is generated based on original input with its characteristics altered. The examples of images generated by CycleGAN and the algorithm used are shown in Fig.\ref{fig:cycle_examples} and the configuration of the Generator and Discriminator used in this research is shown in Fig.\ref{fig:cycle_model}.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/cycle_examples.png}
\caption{Examples of images generated by CycleGAN\cite{zhu2020unpaired}}
\label{fig:cycle_examples}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/cycle_model.png}
\caption{Configuration of Generator and Discriminator}
\label{fig:cycle_model}
\end{figure}
The image conversion using CycleGAN model is shown in Fig.\ref{fig:cycle_results}. The images converted by CycleGAN were difficult for human eyes to distinguish from actual images on rails.
From the 1000 images generated by CycleGAN model, 70 of them which were similar to the undetected shelling defect, were chosen and added to the defect detection model as training data. The accuracy of detecting shelling defects was successfully improved and the test data showed that the average area under the curve was 1, higher compared with 0.993 of Fig.\ref{fig:beforeAcc}. The test results are shown in Fig.\ref{fig:after_acc}.
\begin{figure}[ht]
\includegraphics[width=\columnwidth]{figures/cycle_results.png}
\caption{Image conversion using CycleGAN}
\label{fig:cycle_results}
\end{figure}
\begin{figure}[ht]
\includegraphics[width=\columnwidth]{figures/afterAcc.png}
\caption{Accuracy of defect detection model after addition of CycleGAN generated images.}
\label{fig:after_acc}
\end{figure}
\section{Visualization class activation mapping using CAM\cite{Selvaraju_2019} and t-SNE\cite{vanDerMaaten2008}}
In order to visually confirm the changes that occurred following the addition of training data to the defect detection model, we have utilized a method known as Class Activation Model (CAM) to create a heat map. This is a visual representation of which part of the image the defect detection model focused on when classifying the images. The parts that were used to classify the images are shown in red while the parts that were not are shown in blue. The results are shown in Fig.\ref{fig:heatmap}.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/heatmap.png}
\caption{Heat map changes before and after addition of images generated by CycleGAN}
\label{fig:heatmap}
\end{figure}
As shown in Fig.\ref{fig:heatmap}, the defect detection model before the input of images generated by CycleGAN reacted strongly towards the linear parts other than the defects on the rails. It can be thought that small shelling defects were not detected as such due to the lack of data on images of defective equipment. It is evident that adding the generated images of shelling defects as training data enables the neural network to focus on the shelling defects and identify them as such.
t-SNE was used to perform classification analysis (clustering) on the data of 2048-dimensions immediately before the generation of final results by the defect detection model. The results are shown in Fig.\ref{fig:afterTsne}.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{figures/after_tsne.png}
\caption{t-SNE graph input of images generated by CycleGAN\\(0: normal, 1:scratch, 2: inside-scratch, 3: shelling defects)}
\label{fig:afterTsne}
\end{figure}
After the addition of images of shelling defect as training data, the aforementioned image was categorized in the same group as other images of shelling defect, showing that the defect detection model identified the image as similar to other images of shelling defect.
\section{Conclusion and future implications}
In this research, the accuracy of the defect detection model was successfully improved by using CycleGAN to generate images of defective equipment that were indistinguishable from real images and using them as training data. Detecting defects on rails were the subject of this study but we think these findings are applicable to defect detection of various equipment. We plan to conduct further studies with the aim of improving defect detection accuracy using CycleGAN not only for railway tracks but railway equipment in general.
|
\section*{\large{Summary}}
{\small The Hosmer-Lemeshow (HL) test is a commonly used global goodness-of-fit (GOF) test that assesses the quality of the overall fit of a logistic regression model. In this paper, we give results from simulations showing that the type 1 error rate (and hence power) of the HL test decreases as model complexity grows, provided that the sample size remains fixed and binary replicates are present in the data. We demonstrate that the generalized version of the HL test by \cite{surjanovic2020generalized} can offer some protection against this power loss. We conclude with a brief discussion explaining the behaviour of the HL test, along with some guidance on how to choose between the two tests.
\noindent \textit{Key Words: } chi-squared test, generalized linear model, goodness-of-fit test, Hosmer-Lemeshow test, logistic regression}
\section{Introduction}
Logistic regression models have gained a considerable amount of attention as a tool for estimating the probability of success of a binary response variable, conditioning on several explanatory variables. Researchers in health and medicine have used these models in a wide range of applications. One of many examples includes estimating the probability of hospital mortality for patients in intensive care units as a function of various covariates \citep{lemeshow1988predicting}.
Regardless of the application, it is always desirable to construct a model that fits the observed data well. One of several ways of assessing the quality of the fit of a model is with goodness-of-fit (GOF) tests \citep{bilder2014analysis}. In general, GOF tests examine a null hypothesis that the structure of the fitted model is correct. They may additionally identify specific alternative models or deviations to test against, but this is not required. Global (omnibus) GOF tests are useful tools that allow one to assess the validity of a model without restricting the alternative hypothesis to a specific type of deviation.
An example of a well-known global GOF test for logistic regression is the Hosmer-Lemeshow (HL) test, introduced by \cite{hosmer1980goodness}. The test statistic is a Pearson statistic that compares observed and expected event counts from data grouped according to ordered fitted values from the model. The decision rule for the test is based on comparing the test statistic to a chi-squared distribution with degrees of freedom that depend on the number of groups used to create the test statistic. The HL test is relatively easy to implement in statistical software, and since its creation, the HL test has become quite popular, particularly in fields such as biostatistics and the health sciences.
Despite its popularity, the HL test is known to have some drawbacks. In both experimental and observational studies, it is possible to have data for which binary observations have the same explanatory variable patterns (EVPs). In this case, responses can be aggregated into binomial counts and trials. When there are observations with the same EVP present in the data, it is possible to obtain many different p-values depending on how the data are sorted \citep{bertolini2000one}. A related disadvantage of HL-type tests is that the test statistic depends on the way in which groups are formed, as remarked upon by \cite{hosmer1997comparison}. In this paper we will highlight a related but different problem with the HL test that does not appear to be well known.
Models for logistic regression with binary responses normally assume a Bernoulli model where the probability parameter is related to explanatory variables through a logit link. As mentioned, when several observations have the same EVP, responses can be summed into binomial counts. Other times, the joint distribution of covariates may cause observed values to be clustered into near-replicates, so that the responses might be viewed as being approximately binomially distributed. These cases present no problem for model fitting and estimating probabilities. However, it turns out that this clustering in the covariate space may materially impact the validity of the HL test applied to the fitted models.
For such data structures, a chi-squared distribution does not represent the null distribution of the HL test statistic well in finite samples, as suggested by simulation results in Section \ref{sec:sim_results}. This, in turn, adversely affects both the type 1 error rate and the power of the HL test. \cite{bertolini2000one} demonstrated that it is possible to obtain a wide variety of p-values and test statistics when there are replicates in the data, simply by reordering the data. Our analysis also deals with replicates in the data. However, we find a different phenomenon: as the model size grows for a fixed sample size, the type 1 error rate tends to decrease.
In this paper we show that the HL test can be improved upon by using another existing global GOF test, the generalized HL (GHL) test from \cite{surjanovic2020generalized}. Empirical results suggest that the GHL test performs reasonably well even when applied to moderately large models fit on data with exact replicates or clusters in the covariate space. We offer a brief discussion as to why one might expect clustering in the covariate space to affect the regular HL test. A simple decision tree is offered to summarize when each test is most appropriate.
An overview of the HL and GHL tests is given in Section \ref{sec:methods}. The design of the simulation study comparing the performance of these two tests is outlined in Section \ref{sec:sim_design}, with the results given in Section \ref{sec:sim_results}. We end with a discussion of the implications of these results and offer some guidance on how to choose between the two tests in Section \ref{sec:discussion}.
\section{Methods}
\label{sec:methods}
In what follows, we use the notation of \cite{surjanovic2020generalized}. We let $Y \in \{0,1\}$ denote a binary response variable associated with a $d$-dimensional covariate vector, $X \in \mathbb{R}^d$, where the first element of $X$ is equal to one. The pairs $(X_i, Y_i)$, $i=1,\ldots,n$ denote a random sample, with each pair being distributed according to the joint distribution of $(X,Y)$. The observed values of $(X_i, Y_i)$ are denoted using lowercase letters as $(x_i, y_i)$.
In a logistic regression model with binary responses, one assumes that
$$\Exp(Y|X=x) = \pi(\beta^\top x) = \frac{\exp(\beta^\top x)}{1+\exp(\beta^\top x)},$$
for some $\beta \in \mathbb{R}^d$. The likelihood function is
$$\mathcal{L}(\beta) = \prod_{i=1}^n \pi(\beta^\top x_i)^{y_i} (1-\pi(\beta^\top x_i))^{1-y_i}.$$
From this likelihood, a maximum likelihood estimate (MLE), $\beta_n$, of $\beta$ is obtained.
\bigskip \noindent \textbf{The HL Test Statistic} \\
To compute the HL test statistic, one partitions the observed data, $(x_i, y_i)$, into $G$ groups. Typically, the groups are created so that fitted values are similar within each group and the groups are approximately of equal size. To achieve this, a partition is defined by a collection of $G+1$ interval endpoints, $-\infty = k_0 < k_1 < \cdots < k_{G-1} < k_G = \infty$. The $k_g$ often depend on the data, usually being set equal to the logits of equally-spaced quantiles of the fitted values, $\hat\pi_i = \pi(\beta_n^\top x_i)$. We define $I_i^{(g)} = \mathbbm{1}(k_{g-1} < \beta_n^\top x_i \leq k_g)$, $O_g = \sum_{i=1}^n y_i I_i^{(g)}$, $E_g = \sum_{i=1}^n \hat\pi_i I_i^{(g)}$, $n_g = \sum_{i=1}^n I_i^{(g)}$, and $\bar{\pi}_g = E_g/n_g$, where $\mathbbm{1}(A)$ is the indicator function on a set $A$. With this notation, the number of observations in the $g$th group is represented by $n_g$, and $\bar{\pi}_g$ denotes the mean of the fitted values in this group. The HL test statistic is a quadratic form that is commonly written in summation form as
$$\widehat{C}_G = \sum_{g=1}^G \frac{(O_g - E_g)^2}{n_g \bar{\pi}_g (1-\bar{\pi}_g)}.$$
When $G>d$, \cite{hosmer1980goodness} find that the HL test statistic is asymptotically distributed as a weighted sum of chi-squared random variables under the null hypothesis, after checking certain conditions of Theorem 5.1 in \cite{moore1975unified}. Precisely,
\begin{equation}
\label{eq:HL_limit}
\widehat{C}_G \xrightarrow{d} \chi^2_{G-d} + \sum_{j=1}^d \lambda_j \chi_{1j}^2,
\end{equation}
with each $\chi_{1j}^2$ being a chi-squared random variable with 1 degree of freedom, and each $\lambda_j$ an eigenvalue of a certain matrix that depends on both the distribution of $X$ and the vector $\beta_0$, the true value of $\beta$ under the null hypothesis. \cite{hosmer1980goodness} conclude through simulations that the right side of (\ref{eq:HL_limit}) is well approximated by a $\chi^2_{G-2}$ distribution in various settings.
The HL test statistic and the corresponding p-value both depend on the chosen number of groups, $G$. Typically, $G=10$ groups are used, so that observations are partitioned into groups that are associated with ``deciles of risk''. Throughout this paper we use 10 groups, and therefore compare the HL test statistic to a chi-squared distribution with 8 degrees of freedom, but the results hold for more general choices of $G$.
\bigskip \noindent \textbf{The GHL Test Statistic} \\
The GHL test introduced by \cite{surjanovic2020generalized} generalizes several GOF tests, allowing them to be applied to other generalized linear models. Tests that are generalized by the GHL test include the HL test \citep{hosmer1980goodness}, the Tsiatis \citep{tsiatis1980note} and generalized Tsiatis tests \citep{canary2016summary}, and a version of the ``full grouped chi-square'' from \cite{hosmer2002goodness} with all weights equal to one. The test statistic is a quadratic form like $\widehat{C}_G$, but with important changes to the central matrix. The theory behind this test depends on the residual process, $R_n^1(u)$, $u \in \mathbb{R}$, defined in \cite{stute2002model}. In the case of logistic regression,
$$ R_n^1(u) = \frac{1}{\sqrt{n}} \sum_{i=1}^n [Y_i - \pi(\beta_n^\top X_i)] \mathbbm{1}(\beta_n^\top X_i \leq u) , $$
a cumulative sum of residuals that are ordered according to the size of their corresponding fitted values. This process is transformed into a $G$-dimensional vector, $S_n^1$, which forms the basis of the HL and GHL test statistics, with
$$ S_n^1 = (R_n^1(k_1) - R_n^1(k_0), \ldots, R_n^1(k_G) - R_n^1(k_{G-1}) )^\top. $$
In order to approximate the variance of $S_n^1$, we need to define several matrices. Let
\begin{align*}
\left(G_n^{*}\right)_{gi} &= I_i^{(g)}, \\
V^{*1/2} &= \diag\left( [\pi(\beta_0^\top x_i) (1-\pi(\beta_0^\top x_i))]^{1/2} \right),
\end{align*}
for $i = 1, \ldots, n$, and $g = 1, \ldots, G$. Also, define $X^*$ to be the $n \times d$ matrix with $i$th row given by $x_i^\top$, and let $V_n^{*1/2}$ be the same as $V^{*1/2}$, but evaluated at the estimate $\beta_n$ of $\beta_0$. Finally, define
\begin{align}
\label{eq:Sigma_n}
\Sigma_n
&= \frac{1}{n} G_n^* \left(V_n^* - V_n^* X^* (X^{*\top} V_n^* X^*)^{-1} X^{*\top} V_n^{*}\right) G_n^{*\top} \nonumber \\
&= \frac{1}{n} G_n^* V_n^{*1/2} \left(I_n - V_n^{*1/2} X^* (X^{*\top} V_n^* X^*)^{-1} X^{*\top} V_n^{*1/2}\right) V_n^{*1/2} G_n^{*\top},
\end{align}
where $I_n$ is the $n \times n$ identity matrix.
For logistic regression models, the GHL test statistic is then
$$ X^2_{\text{GHL}} = S_n^{1\top} \Sigma_n^{+} S_n^1, $$
where $\Sigma_n^{+}$ is the Moore-Penrose pseudoinverse of $\Sigma_n$. Under certain conditions given by \cite{surjanovic2020generalized},
$$ S_n^{1\top} \Sigma_n^{+} S_n^1 \xrightarrow{d} \chi^2_\nu, $$
where $\nu = \rank(\Sigma)$, with $\Sigma$ a matrix defined in their paper. Since the rank of $\Sigma$ might be unknown, they use the rank of $\Sigma_n$ as an estimate. We use the same approach to estimating $\nu$, empirically finding that the estimated rank of $\Sigma_n$ is often equal to $G-1$ for logistic regression models.
The GHL test statistic for logistic regression models is equivalent to the Tsiatis GOF test statistic \citep{tsiatis1980note} and the $X^2_w$ statistic from \cite{hosmer2002goodness} with all weights set equal to 1, when a particular grouping method is used---that is, when $G_n^*$ is the same for all methods. However, use of the GHL test is justified for a wide variety of GLMs and grouping procedures with a proper modification of $\Sigma_n$, as described by \cite{surjanovic2020generalized}. For both the HL and GHL tests, we use the grouping method proposed by \cite{surjanovic2020generalized}, which uses random interval endpoints, $k_g$, so that $\sum_{i=1}^n \hat\pi_i (1-\hat\pi_i) I_i^{(g)}$ is roughly equal between groups. Further details of the implementation are provided in the supplementary material of their paper.
It is important to note that $\Sigma_n$ is a non-diagonal matrix that standardizes and accounts for correlations between the grouped residuals in the vector $S_n^1$. This can be seen from (\ref{eq:Sigma_n}), which shows that $\Sigma_n$ contains a generalized hat matrix for logistic regression. In contrast, when written as a quadratic form, the central matrix of the HL test statistic is diagonal and does not account for the number of parameters in the model, $d$, when standardizing the grouped residuals. We expect this standardization to be very important when exact replicates are present, as the binomial responses might be more influential than sparse, individual binary responses.
It is extremely common to fit logistic regression models to data where multiple Bernoulli trials are observed at some or all EVPs, even when the underlying explanatory variables are continuous. As with any fitted model, a test of model fit would be appropriate, and the HL test would likely be a candidate in a typical problem. It is therefore important to explore how the HL and GHL tests behave with large models when exact or near-replicates are present in the data.
\section{Simulation Study Design}
\label{sec:sim_design}
We compare the performance of the HL and GHL tests by performing a simulation study. Of particular interest is the rejection rate under the null, when the tests are applied to moderately large models that are fit to data with clusters or exact replicates in the covariate space.
In all settings, the true regression model is
\begin{equation}
\label{eq:sim_prob}
\Exp(Y|X=x) = \logit(\beta_0 + \beta_1 x_1 + \ldots + \beta_{d-1} x_{d-1}),
\end{equation}
with $d$ in $\{2, 3, \ldots, 25\}$. Here, $\beta_0$ represents the intercept term. To produce replicates in the covariate space, $m \leq n$ unique EVPs are drawn randomly from a $(d-1)$-dimensional spherical normal distribution with marginal mean 0 and marginal variance $\sigma^2=1$ for each simulation realization. At each EVP, $n/m$ replicate Bernoulli trials are then created, with probabilities determined by (\ref{eq:sim_prob}). In our simulation study, we fix $n=500$ and select $m \in \{50, 100, 500\}$ so that the number of replicates at each EVP, $n/m$, is 10, 5, or 1, respectively.
We set $\beta_0 = 0.1$ and $\beta_1 = \ldots = \beta_{d-1} = 0.535 / \sqrt{d-1}$. This results in fitted values that rarely fall outside the interval $[0.1, 0.9]$, regardless of the number of parameters in the model, so that the number of expected counts in each group is sufficiently large for the use of the Pearson-based test statistics.
We also perform some simulations with $n=100$, using smaller values of $d$ and $m$ than for $n=500$. However, we focus on results for $n=500$ because we are then able to increase the number of replicates per EVP, $n/m$, while still maintaining large enough $m$ so that it is possible to create ten viable groupings. In each simulation setting, 10,000 realizations are produced. All simulations are performed using \textsf{R}.
\section{Simulation Results}
\label{sec:sim_results}
Figure \ref{fig:null_results} presents plots of the sample mean and variance of the HL and GHL test statistics against the number of variables in the model, separately for each $m$. An analogous plot of the estimated type 1 error rate of the tests against the number of variables is also presented. For the HL test, all three statistics show a clear decreasing pattern with increasing model size when replicates are present, with a sharper decrease when the number of replicates per EVP is larger. Since the estimated variance is not always twice the size of the mean, we can infer that the chi-squared approximation to the null distribution of the HL test statistic is not adequate in finite samples for these data structures. Simulation results with a sample size of $n=100$ are not displayed, but are quite similar.
From the same figure, we see that the GHL test performs well in the settings considered. The estimated mean and variance of the test statistic stay close to the desired values of $G-1=9$ and $2(G-1)=18$. We note that the GHL test can have an inflated type 1 error rate, particularly when it is applied to highly complex models. The models considered here are only moderately large, with $d \leq \min \{ n/20, m/2 \}$. If one wishes to use the GHL test to assess the quality of fit of larger models with only a moderate sample size, one should be wary of an inflated type 1 error rate that can become considerably large for complex models. A possible explanation for this is that estimating the off-diagonal elements of the matrix $\Sigma_n$ can potentially introduce a considerable amount of variance into the test statistic in small samples.
Recall from (\ref{eq:HL_limit}) that the asymptotic $\chi^2_{G-2}$ distribution for the HL test proposed by \cite{hosmer1980goodness} is based on a sum of chi-squares, where one has $G-d$ degrees of freedom. We investigated whether maintaining $G=10$ while increasing $d$ contributes to the phenomena we have observed. We set $G=26$ and performed a similar simulation study. The adverse behavior of the HL statistic still persists despite this modification.
We also investigated the effect of near-replicate clustering in the covariate space. We fixed $n$ and $m$ as in Section \ref{sec:sim_design}, but added a small amount of random noise with marginal variance $\sigma^2_e$ to each replicate within the $m$ sampled vectors. The amount of clustering was controlled by varying $\sigma^2_e$, as shown in Figure \ref{fig:clustering}. As expected, increasing $\sigma^2_e$ reduces the severity of the decreasing mean, variance and type 1 error rate for the HL test statistic. However, the pattern remains evident while $\sigma^2_e/\sigma^2$ remains small.
\section{Discussion}
\label{sec:discussion}
The original HL test, developed by \cite{hosmer1980goodness}, is a commonly-used test for logistic regression among researchers in biostatistics and the health sciences. Although its performance is well documented \citep{lemeshow1982review, hosmer1997comparison, hosmer2002goodness}, we have identified an issue that does not seem to be well known. For moderately large logistic regression models fit to data with clusters or exact replicates in the covariate space, the null distribution of the HL test statistic can fail to be adequately represented by a chi-squared distribution in finite samples. Using the original chi-squared distribution with $G-2$ degrees of freedom can result in a reduced type 1 error rate, and hence lower power to detect model misspecifications. Based on the results of the simulation study, the GHL test can perform noticeably better in such settings, albeit with a potentially inflated type 1 error rate.
Similar behaviour of the HL test was observed in \cite{surjanovic2020generalized}, where the regular HL test was naively generalized to allow for it to be used with Poisson regression models. In their setup, even without the presence of clusters or exact replicates in the covariate space, as the number of model parameters increased for a fixed sample size, the estimated type 1 error rate decreased. The central matrix in the GHL test statistic, $\Sigma_n$, makes a form of correction to the HL test statistic by standardizing and by accounting for correlations between the grouped residuals that comprise the quadratic form in both the HL and GHL tests. This is evident from (\ref{eq:Sigma_n}), which shows that $\Sigma_n$ contains the generalized hat matrix subtracted from an identity matrix. To empirically assess the behaviour of $\Sigma_n$, we varied $\sigma^2_e$ in the setup with replicates and added noise, described at the end of Section~\ref{sec:sim_results}. For large $d$ and moderate $n$, both fixed, we found that the diagonal elements of $\Sigma_n$ tend to shrink, on average, as $\sigma^2_e$ decreases. In contrast, the elements of the HL central matrix remain roughly constant. Therefore, the GHL statistic seems to adapt to clustering or replicates in $X$, whereas the HL test statistic does not.
In logistic regression with exact replicates, grouped binary responses can be viewed as binomial responses that can be more influential. In this scenario, as $d$ increases for a fixed sample size $n$, the distribution of the regular HL test statistic diverges from a single chi-squared distribution, suggesting that the standardization offered by the central GHL matrix becomes increasingly important.
The real-life implications of the reduced type 1 error rate and power of the regular HL test are that in models with a considerable number of variables---provided that the data contains clusters or exact replicates---the HL test has limited ability to detect model misspecifications. Failure to detect model misspecification can result in retaining an inadequate model, which is arguably worse than rejecting an adequate model due to an inflated type 1 error rate, particularly when logistic regression models are used to estimate probabilities from which life-and-death decisions might be made.
Our advice for choosing between the two GOF tests is displayed as a simple decision tree in Figure~\ref{fig:decision_tree}. The advice should be interpreted for $G=10$ groups, the most commonly used number of groups. With large samples, provided that $m$ is sufficiently large compared to $d$, it should generally be safe to use the GHL test. Our simulations explored models with $d \leq 25$, so some caution should be exercised if the GHL test is to be used with larger models. For small or moderate samples, such as when $n=100$ or $500$, it is important to identify whether there are clusters or exact replicates in the covariate space. One can compute the number of unique EVPs, $m$, and compare this number to the sample size, $n$. If $n/m \geq 5$, say, then there is a considerable amount of ``clustering''. For data without exact replicates, clusters can still be detected using one of many existing clustering algorithms, and the average distances between and within clusters can be compared. Informal plots of the $x_i$ projected onto a two- or three-dimensional space can also be used as an aid in this process.
If there is no evidence of clustering or replicates, the HL test should not be disturbed by this phenomenon. On the other hand, if there is a noticeable amount of clustering, and the regression model is not too large, say $d \leq \min \{ n/20, m/2 \}$, where $m$ also represents the number of estimated clusters, then one can use the GHL test. In the worst-case scenario with a small sample size, clustering, and a large regression model, one can use both tests as an informal aid in assessing the quality of the fit of the model, recognizing that GHL may overstate the lack of fit, while HL may understate it. If the two tests agree, then this suggests that the decision is not influenced by the properties of the tests. When they disagree, conclusions should be drawn more tentatively.
\section*{Acknowledgements}
We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada (NSERC), [funding reference number RGPIN-2018-04868]. Cette recherche a été financée par le Conseil de recherches en sciences naturelles et en génie du Canada (CRSNG), [numéro de référence RGPIN-2018-04868].
\setlength{\bibsep}{0pt plus 0.3ex}
|
\section{Introduction}
\label{sec:intro}
Automatic speech recognition (ASR) has achieved rapid progress with the development of deep learning. Advanced models such as DNN~\cite{dahl2011context}, CNN~\cite{sainath2013deep}, RNN~\cite{graves2013speech} and end-to-end models~\cite{chan2016listen,mohamed2019transformers} result in better recognition accuracy compared with conventional hybrid models~\cite{trentin2001survey}. However, as a side effect, deep learning-based models require a large amount of labeled training data to combat overfitting and ensure high accuracy, especially for speech recognition tasks with few training data.
Therefore, a lot of data augmentation methods for ASR~\cite{ko2015audio,shahnawazuddin2016pitch,toth2018perceptually,park2019specaugment,ren2019almost,xu2020lrspeech} were proposed, mainly on augmenting speech data. For example, speed perturbation \cite{ko2015audio}, pitch adjust~\cite{shahnawazuddin2016pitch}, adding noise~\cite{toth2018perceptually} and vocal tract length perturbation increases the quantity of speech data by adjusting the speed or pitch of the audio, or by adding noisy audio on the original clean audio, or by transforming spectrograms. Recently, SpecAugment~\cite{park2019specaugment} was proposed to mask the mel-spectrogram along the time and frequency axes, and achieve good improvements on recognition accuracy. Furthermore, \cite{wang2019semantic} masks the speech sequence in the time domain according to the alignment with text to explore the semantical relationship. As can be seen, most previous methods focus on augmenting the speech input while not changing the corresponding label (text), which needs careful tuning on the augmentation policy. For example, SpecAugment needs a lot of hyper-parameters (the time warp parameter $W$, the time and frequency mask parameters $T$ and $F$, a time mask upper bound $p$, and the number of time and frequency mask $m_F$ and $m_T$) to determine how to perform speech augmentation, where improper parameters may cause much information loss and thus cannot generate text correctly, or may have small changes on the speech and thus have no augmentation effect.
Recently, mixup technique~\cite{zhang2017mixup} is proposed to improve the generalization limitation of empirical risk minimization (ERM)~\cite{vapnik1999overview}. Vanilla mixup randomly chooses a weight from a distribution and combines two samples and corresponding labels with the same weight. Recent works apply mixup on different tasks, including image classification~\cite{zhang2018deep}, and sequence classification~\cite{guo2019augmenting}. Different from classification task with one-hot labels where mixup can be easily incorporated, conditional sequence generation tasks such as automatic speech recognition, image captioning, and handwritten text recognition cannot directly apply mixup on the target sequence due to different lengths.
\cite{medennikov2018investigation} directly apply mixup to train a neural acoustic model like LF-MMI \cite{povey2016purely}, which simply integrates inputs and targets frame-by-frame with shared mixture weight because the input feature and label are aligned frame-wisely.
In this paper, we propose MixSpeech, a simple yet effective data augmentation method for automatic speech recognition. MixSpeech trains an ASR model by taking a weighted combination of two different speech sequences (e.g., mel-spectrograms or MFCC) as input, and recognizing both text sequences. Different from mixup~\cite{zhang2017mixup} that uses the weighted combination of two labels as the new label for the mixed input, MixSpeech uses each label to calculate the recognition loss and combines the two losses using the same weight as in the speech input. MixSpeech is much simple with only a single hyper-parameter (the combination weight $\lambda$), unlike the complicated hyper-parameters used in SpecAugment. Meanwhile, MixSpeech augments the speech input by introducing another speech, which acts like a contrastive signal to force the ASR model to better recognize the correct text of the corresponding speech instead of misled by the other speech signal. Therefore, MixSpeech is more effective than previous augmentation methods that only change the speech input with masking, wrapping, pitch, and duration adjusting, without introducing contrastive signals.
\section{Method}
In this section, we first briefly recap the concept of mixup~\cite{zhang2017mixup}, which is an efficient augmentation approach for the single-label task. In order to handle sequence generation tasks (e.g., ASR), then we introduce MixSpeech. At the last, we describe two popular models on which MixSpeech is implemented to verify the effectiveness of MixSpeech.
\subsection{Mixup Recap}
Mixup~\cite{zhang2017mixup} is an effective data augmentation method for supervised learning tasks. It trains a model on a convex combination of pairs of inputs and their targets to make the model more robust to adversarial samples. The mixup procedure can be demonstrated as:
\begin{equation}
\begin{aligned}
\label{eq:data_mix}
X_{mix} &= \lambda X_i + (1-\lambda) X_j,\\
Y_{mix} &= \lambda Y_i + (1-\lambda) Y_j,
\end{aligned}
\end{equation}
where $X_i$ and $Y_i$ are the $i$-th input and target of the data sample, $X_{mix}$ and $Y_{mix}$ represent the mixup data by combining a pair of data samples ($i$ and $j$), and $\lambda \sim Beta(\alpha, \alpha)$ with $\alpha\in(0, \infty)$, is the combination weight.
\subsection{MixSpeech}
Mixup can be easily applied to classification tasks where the target is a single label. However, it is difficult to apply into sequence generation tasks such as ASR due to the following reasons: 1) Two text sequences ($Y_i$, $Y_j$) may have different lengths and thus cannot be directly mixed up as in Eq.~\ref{eq:data_mix}; 2) Text sequences are discrete and cannot be directly added; 3) If adding two text sequences in the embedding space, the model will learn to predict a mixture embedding of two text sequences, which will confuse the ASR model and may hurt the performance since its final goal is to recognize a single text from the speech input.
To ensure effective mixup for the sequential data (speech and text) in ASR, we propose MixSpeech, which mixes two speech sequences in the input and mixes two loss functions regarding the text output, as shown in Figure \ref{fig:model_structure}. The formulation of MixSpeech is as follows:
\begin{equation}
\begin{aligned}\label{eq:mixspeech}
& X_{mix} = \lambda X_i + (1-\lambda) X_j, \\
&\mathcal{L}_{i} = \mathcal{L}(X_{mix},Y_i) \quad \mathcal{L}_{j} = \mathcal{L}(X_{mix},Y_j), \\
&\mathcal{L}_{mix}=\lambda\mathcal{L}_{i} + (1-\lambda)\mathcal{L}_{j},
\end{aligned}
\end{equation}
where $X_i$ and $Y_i$ are the input speech sequence and target text sequence of $i$-th sample, $X_{mix}$ is the mixup speech sequence by adding the two speech sequences frame-wisely with weight $\lambda$. $\mathcal{L}(\cdot, \cdot)$ calculates the ASR loss (which also includes the recognition process), and $\mathcal{L}_{mix}$ combines the two losses with the same weight $\lambda$ as in the speech input during the training phase. Following the original mixup, we choose $\lambda \sim Beta(\alpha, \alpha)$ where $\alpha\in(0, \infty)$.
\begin{figure}[t]
\centering
\includegraphics[scale=0.3]{mixup_procedure_4icassp-final-2.pdf}
\caption{The pipeline of MixSpeech. $X_{mix}$ is the mixture of two mel-spectrograms, $X_i$ and $X_j$. $Loss_i$ is calculated from the output of $X_{mix}$ and the corresponding label $Y_i$. $Loss_{mix}$ is the weighted combination of two losses $Loss_i$ and $Loss_j$.}
\label{fig:model_structure}
\end{figure}
\subsection{Model Structure}
LAS~\cite{chan2016listen} and Transformer~\cite{vaswani2017attention} architectures have been widely used and achieved great performance on ASR tasks. In this paper, we implement MixSpeech on these two popular models to demonstrate its effectiveness. The two models both leverage the joint CTC-attention~\cite{kim2017joint} structure to train the model, which consists of two loss functions: 1) CTC (Connectionist Temporal Classification) \cite{graves2006connectionist} loss on the encoder output and 2) Cross-Entropy (CE) loss at the output of decoder at each timestep. To train the CTC part, a sequence of labels, denoted as $l$, are used to compute loss with an effective algorithm such as the Baum-Welch algorithm~\cite{baum1970maximization}. The CTC loss can be written as:
\begin{equation}
\mathcal{L}_{CTC}(\hat{y},l) = -\log\left(\sum_{\pi\in\beta^{-1}(l)}\prod_{t=1}^T y_{\pi_t}^t\right),
\end{equation}
where $\beta$ is the function that removes repeated labels and blank labels, $\pi_t$ represents intermediate label sequence containing blank label and $l$ is the target sequence without blank. The Cross-Entropy loss can be written as:
\begin{equation}
\begin{aligned}
\mathcal{L}_{CE} &=-\sum_{u}\log(P(y_u|x, y_{1:u-1})),\\
\end{aligned}
\end{equation}
where $y_u$ is the $u$-th target token. In LAS~\cite{chan2016listen}, the encoder and decoder are the stacked BiLSTM and LSTM respectively. In Transformer based structure~\cite{vaswani2017attention}, they are stacked of multi-head attention and feed-forward network.
During training, following~\cite{kim2017joint}, we leverage the multi-task learning by combining CTC and Cross-Entropy loss as follows:
\begin{equation}
\begin{aligned}
\label{eq:multi-task-learning}
\mathcal{L}_{MTL} = \beta \mathcal {L}_{CTC} + &(1-\beta) \mathcal{L}_{CE},
\end{aligned}
\end{equation}
where $\beta \in [0,1]$ is a tunable hyper-parameter. By combining with MixSpeech in Eq.~\ref{eq:mixspeech}, the final training objective can be written as:
\begin{equation}
\label{eq:mixup_para}
\mathcal{L}_{mix} = \lambda\mathcal{L}_{MTL}(X_{mix},Y_i) + (1-\lambda)\mathcal{L}_{MTL}(X_{mix},Y_j).
\end{equation}
\section{Experimental results}
\subsection{Datasets}
We evaluate our proposed MixSpeech on several datasets including TIMIT~\cite{garofolo1993darpa}, WSJ~\cite{paul1992design} and the Mandarin ASR corpus HKUST~\cite{liu2006hkust}. For TIMIT dataset, we use the standard train/test split, where we randomly sample 10$\%$ of the training set for validation and regard the left as the training set. WSJ~\cite{paul1992design} is a corpus of read news articles with approximately 81 hours of clean data. Following the standard process, we set the Dev93 for development and Eval92 for evaluation. For HKUST dataset, the development set contains 4,000 utterances (5 hours) extracted from the original training set with 197,387 utterances (173 hours) and the rest are used for training, besides 5,413 utterances (5 hours) for evaluation. The speech sentences of these corpora with 16k sampling rate extract the fbank features as the model input. For TIMIT dataset, fbank features with 23 dimensions are extracted at each frame, while for the other two datasets 80 dimensions fbank features are extracted at each frame following the common practice.
\subsection{Model Settings} \label{sec:settings}
We implement MixSpeech based on the ESPnet codebase\footnote{https://github.com/espnet/espnet}. MixSpeech is designed for extracting and splitting targets from mixed inputs to train a more generalizable model, which is a more difficult task than training without MixSpeech. Thus we give more training time to the model enhanced with MixSpeech\footnote{More training time for the baseline model cannot get better results according to our preliminary experiments.}.
The baselines of LAS~\cite{chan2016listen} and Transformer~\cite{vaswani2017attention} are implemented as the following settings. For the LAS (Listen Attend and Spell), the encoder has 4-layer Bi-LSTM with width 320 and the decoder has a single layer LSTM with width 300. For the Transformer based architecture, the encoder has 12 layers and the decoder has 6 layers with width 256, where each layer is composed of multi-head attention and fully connected layer. CTC and decoder cross-entropy loss are used for jointly decoding using beam search with beam size 20. The default $\alpha$ for Beta distribution is set to 0.5 following~\cite{zhang2017mixup}. The multi-task learning parameter $\beta$ is set to 0.3. In the practice, we randomly select part of paired data within one batch to train the model with MixSpeech and other data within one batch to train the model as usual. The proportion of one batch data to train using MixSpeech is denoted as $\tau$, which is set to 15\% as default and achieve good performance according to our experiments in Table~\ref{tab:mixup_prob_comparison}.
\subsection{Results of MixSpeech} \label{exp:mixspeech_main}
We compare MixSpeech with two settings: 1) models trained without MixSpeech (Baseline) and 2) models trained with SpecAugment~\cite{park2019specaugment} (SpecAugment), an effective data augmentation technique. Table \ref{tab:preliminary_comparison} shows the performance of MixSpeech on TIMIT dataset with the LAS model, comparing with Baseline and SpecAugment. Furthermore, Table \ref{tab:preliminary_comparison_trans} shows the comparison of our method with Baseline and SpecAugment on three different low resource datasets by different metrics with Transformer. We can find that MixSpeech improves the baseline model on TIMIT dataset relatively by 10.6\% on LAS and 6.9\% on Transformer. As shown in the Table \ref{tab:preliminary_comparison_trans}, MixSpeech achieves 4.7\% WER on WSJ, outperforming the baseline model alone and the baseline model with SpecAugment. MixSpeech also achieves better performance on HKUST~\cite{liu2006hkust}, which is a large Mandarin corpus. The results demonstrate that MixSpeech can consistently improve the recognition accuracy across datasets with different languages and data sizes.
\begin{table}[th]
\caption{The phone error rate (PER) results of LAS (Listen, Attend and Spell), and LAS enhanced with SpecAugment or MixSpeech on TIMIT dataset.}
\label{tab:preliminary_comparison}
\centering
\vspace{3mm}
\begin{tabular}{@{}llllc@{}}
\toprule
\textbf{Method} & & & & \textbf{PER} \\ \midrule
Baseline & & & & 21.8\% \\
+SpecAugment & & & & 20.5\% \\ \midrule
+MixSpeech & & & & \textbf{19.5\%} \\ \bottomrule
\end{tabular}
\end{table}
\begin{table}[th]
\caption{The results of Transformer, and Transformer enhanced with SpecAugment or MixSpeech on TIMIT, WSJ, and HKUST dataset. The metrics of the three datasets are phone error rate (PER) and word error rate (WER) respectively.}
\label{tab:preliminary_comparison_trans}
\centering
\vspace{3mm}
\begin{tabular}{@{}lccc@{}}
\toprule
$\textbf{Method}$ & $\textbf{TIMIT}_{PER}$ & $\textbf{WSJ}_{WER}$ & $\textbf{HKUST}_{WER}$ \\ \midrule
Baseline & 23.1\% & 5.9\% & 23.9\% \\
+SpecAugment & 22.5\% & 5.4\% & 23.5\% \\ \midrule
+MixSpeech & \textbf{21.8\%} & \textbf{4.7\%} & \textbf{22.9\%} \\ \bottomrule
\end{tabular}
\end{table}
\subsection{Method Analysis}
Both LAS and Transformer based models leverage multi-task learning to boost the performance and $\beta$ in Equation~\ref{eq:multi-task-learning} is a hyper-parameter to adjust the weight of them. When $\beta$ is set to 0 or 1, there is only Cross-Entropy loss or CTC loss. We vary $\beta$ and get the results of the baseline as shown in Table~\ref{tab:mtl_comparison}. The results show that the model cannot perform well with only one target and requires alignment information guided by CTC loss to help the attention decoder. In Table~\ref{tab:mtl_comparison}, $\beta=0.3$ gets the lowest PER score, and thus is set as default in the baselines and models enhanced with MixSpeech.
\begin{table}[th]
\caption{The phone error rate (PER) results by varying the multi-task learning parameters $\beta$ on TIMIT dataset.}
\label{tab:mtl_comparison}
\centering
\vspace{3mm}
\begin{tabular}{@{}ccccc@{}}
\toprule
$\beta$ & 0 & 0.3 & 0.5 & 0.7 \\ \midrule
\textbf{PER} & 26.6\% & \textbf{23.0\%} & 23.1\% & 23.7\% \\ \bottomrule
\end{tabular}
\vspace{1mm}
\end{table}
As described in Sec.~\ref{sec:settings}, we randomly select part of paired data within one batch (the proportion is denoted as $\tau$) to train the model with MixSpeech and other data within one batch to train the model as usual. We further study whether the proportion $\tau$ of a batch data has much influence on the results and the results are shown in Table~\ref{tab:mixup_prob_comparison}. we can see that the model enhanced with MixSpeech can consistently outperform the baseline ($\tau=0$) by varying $\tau$, which shows that MixSpeech is not sensitive to $\tau$.
\begin{table}[th]
\caption{The phone error rate (PER) results by varying proportion $\tau$ of a batch data to conduct MixSpeech on TIMIT dataset.}
\label{tab:mixup_prob_comparison}
\centering
\vspace{3mm}
\begin{tabular}{@{}ccccc@{}}
\toprule
$\tau$ & 0\% & 15\% & 20\% & 30\% \\ \midrule
\textbf{PER} & 23.1\% & \textbf{21.8\%} & 21.9\% & 22.0\% \\ \bottomrule
\end{tabular}
\end{table}
Rather than mixup on two inputs, we can extend mixup to more inputs. Here we randomly select three inputs, named Tri-mixup, and set the mixup weight $\lambda$ to $1/3$\footnote{\cite{zhang2017mixup} points out that complex prior distribution like the Dirichlet distribution to generate $\lambda$ fails to provide further gain but increase the computation cost. Thus we choose $\lambda=$1/3.}. Similar to MixSpeech, the proportion for mixup $\tau$ is also set to 15\%. Table \ref{tab:tri_mixup} shows that the performance of Tri-mixup drops dramatically, comparing with the baseline and MixSpeech. Tri-mixup may introduce too much information beyond the original ones, which hinders the model from learning useful patterns from the data and leads to the performance drop.
\begin{table}[th]
\caption{The phone error rate (PER) results of different data augmentation methods including Tri-mixup, Noise Regularization and MixSpeech on TIMIT dataset. The Baseline refers to the Transformer model without data augmentation. Tri-mixup means that we select three inputs and conduct mixup with weight $\lambda=$1/3. The noise regularization is to add Gaussian noise to make the model more robust to noise.}
\label{tab:tri_mixup}
\centering
\vspace{3mm}
\begin{tabular}{@{}ccccc@{}}
\toprule
& \multirow{2}{*}{Baseline} & \multirow{2}{*}{Tri-mixup} & Noise & \multirow{2}{*}{MixSpeech} \\
& & & Regularization & \\ \midrule
\textbf{PER} & 23.1\% & 35.8\% & 22.0\% & \textbf{21.8\%} \\ \bottomrule
\end{tabular}
\end{table}
Noise regularization \cite{seltzer2013investigation, yu2008minimum, maas2012recurrent}, which incorporates noise on the inputs, is a common data augmentation method to make the model more robust to noise and more generalizable to unseen data. To compare it with MixSpeech, we add Gaussian noise with SNR 5dB on the training set and report the results in Table~\ref{tab:tri_mixup}. We can see that MixSpeech outperforms the noise regularization.
\section{Conclusion and future work}
In this paper, we proposed MixSpeech, a new data augmentation method to apply the mixup technique on ASR tasks for low-resource scenarios. Experimental results show that our method improves the recognition accuracy compared with the baseline model and previous data augmentation method SpecAgument, and achieves competitive WER performance on WSJ dataset. In the future, we will consider another mixup strategy by concatenating different segments of speech and also their corresponding text sequences, which can create more data samples with novel distributions.
\iffalse
\begin{figure}[htb]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics[width=8.5cm]{image1}}
\centerline{(a) Result 1}\medskip
\end{minipage}
\begin{minipage}[b]{.48\linewidth}
\centering
\centerline{\includegraphics[width=4.0cm]{image3}}
\centerline{(b) Results 3}\medskip
\end{minipage}
\hfill
\begin{minipage}[b]{0.48\linewidth}
\centering
\centerline{\includegraphics[width=4.0cm]{image4}}
\centerline{(c) Result 4}\medskip
\end{minipage}
\caption{Example of placing a figure with experimental results.}
\label{fig:res}
\end{figure}
\fi
\iffalse
\vfill\pagebreak
\section{REFERENCES}
\label{sec:refs}
\fi
\ninept
\bibliographystyle{IEEEtran}
|
\section{Experiments with pre-trained models}
\label{sec:linear}
Recent works have shown that pre-training the models on unlabeled data can be beneficial for subsequent learning tasks \citep{chen2020big,he2020momentum}. \cite{tramer2020differentially} demonstrate that differentially private linear classifier can achieve high accuracy using the features produced by those per-trained models. We show that GEP can also benefit from such pre-trained models.
Inspired by \cite{tramer2020differentially}, we use the output of the penultimate
layer of a pre-trained ResNet152 model as feature to train a private linear classifier. The ResNet152 model is pre-trained on unlabeled ImageNet using SimCLR \citep{chen2020big}. The feature dimension is 4096.
\textbf{Implementation Details} We choose the privacy parameter $\epsilon$ from $[0.1, 0.5, 1, 2]$. The privacy parameter $\delta$ is $1\times 10^{-5}$. We run all experiments for 5 times and report the average accuracy. The clipping threshold of residual gradients is still one-fifth of the clipping threshold of the original gradients. The dimension of anchor subspace is set as $200 \simeq\sqrt{p}$ where $p=40960$ is the model dimension. We randomly sample $500$ samples from the test set as auxiliary data and evaluate performance on the rest test samples. The optimizer is Adam with default momentum coefficients. Other hyper-parameters are listed in Table~\ref{tbl:hyper}.
\begin{table} [h]
\centering
\begin{tabular}{|c|c|}
\hline \hline
Hyperparameter & Values \\ \hline
Learning rate & 0.01, 0.05, 0.1 \\ \hline
Running steps & 50, 100, 400 \\ \hline
Clipping threshold & 0.01, 0.1, 1 \\ \hline
\hline
\hline
\end{tabular}
\caption{Hyperparameter values used in Appendix~\ref{sec:linear}. }
\label{tbl:hyper}
\end{table}
\textbf{Results} The experiment results are shown in Table~\ref{tbl:five_eps0.5}. GEP outperforms GP on all values of $\epsilon$. With privacy bound $\epsilon=2$, GEP achieves 94.8\% validation accuracy on CIFAR10 dataset, improving over the GP baseline by 1.4\%. For very strong privacy guarantee ($\epsilon=0.1$), B-GEP performs on par with GEP because strong privacy guarantee requires large noise and the useful signal in residual gradient is submerged in the added noise. B-GEP benefits less from larger $\epsilon$ compared to GP or GEP. For $\epsilon=1$ and $2$, the performance of B-GEP is worse than the performance of GP. This is because larger $\epsilon$ can not reduce the systematic error of B-GEP (see Remark~\ref{thm:error_cp} in Section~\ref{sec:ppgr_error}).
\begin{table} [h]
\caption{Validation accuracy (in \%) on CIFAR10 with varying choices of $\epsilon$. We train a private linear model on top of the features from a ResNet152 model, which is pre-trained on unlabeled ImageNet. }
\label{tbl:five_eps0.5}
\centering
\begin{tabular}{ p{1.75cm}p{2.25cm}p{2.25cm}p{2.25cm} p{2.25cm}}
\hline \hline
& $\epsilon=0.1$ & $\epsilon=0.5$ & $\epsilon=1$ & $\epsilon=2$ \\[0.4ex]
\hline
Non private & 96.3 & 96.3 & 96.3 & 96.3 \\[0.4ex]
GP & 88.2 ($\pm$0.16) & 91.1 ($\pm$0.17) & 93.2 ($\pm$0.19) & 93.4 ($\pm$0.12) \\[0.4ex]
B-GEP & \textbf{91.0} ($\pm$0.07) & 92.9 ($\pm$0.03) & 93.1 ($\pm$0.10) & 93.2 ($\pm$0.08) \\[0.4ex]
GEP & 90.9 ($\pm$0.19) & \textbf{93.5} ($\pm$0.06) & \textbf{94.3} ($\pm$0.09) & \textbf{94.8} ($\pm$0.06) \\[0.4ex]
\hline
\hline
\end{tabular}
\end{table}
\section{Complexity Analysis}
\label{sec:complexity}
We provide an analysis of the computational and memory costs of the construction of anchor subspace. The computation of the anchor subspace is the dominant additional cost of GEP compared to conventional gradient perturbation. Notations: $k$, $m$, $n$, and $p$ are the dimension of anchor subspace, number of anchor gradients, number of private gradients, and the model dimension, respectively. In order to reduce the computational and memory costs, we divide the parameters into $g$ groups and compute one orthonormal basis for each group. We refer to this approach as `parameter grouping'. In this section, we assume the parameters and the dimension of the anchor subspace are both divided evenly. Table~\ref{tbl:cost} summarizes the additional costs of GEP with/without parameter grouping. Using parameter grouping can reduce the computational/memory cost significantly.
\begin{table}
\caption{Computational and memory costs of a single power iteration in Algorithm~\ref{alg:ppg}. The computation cost is measured by the number of floating point operations. The memory cost is measured by the number of floating-point numbers we need to store. `GEP+PG' denotes GEP with parameter grouping and $g$ denotes the number of groups. Notations: $k$, $m$, $n$, and $p$ are the dimension of anchor subspace, number of anchor gradients, number of private gradients, and the model dimension, respectively. }
\label{tbl:cost}
\centering
\begin{tabular}{ p{1.75cm}p{5cm}p{5cm} }
\hline \hline
& Computational Cost & Memory Cost \\[0.4ex]
\hline
GEP & $2mkp+pk^2$ & $\max\left(0,\left(m-n+k\right)p + mk\right)$ \\[0.4ex]
GEP+PG & $2mkp/g+pk^2/g^2$ & $\max\left(0,\left(m-n+\frac{k}{g}\right)p + mk\right)$ \\[0.4ex]
\hline
\hline
\end{tabular}
\end{table}
\section{Ablation Study}
\label{apdx:abla}
\textbf{The influence of choosing different auxiliary datasets.} We conduct experiments with different choices of auxiliary datasets. For CIFAR10, we try 2000 random test samples from CIFAR10, 2000 random samples from CIFAR100, and 2000 random samples from ImageNet. When the auxiliary dataset is CIFAR10, we try both correct labels and random labels. For all choices of auxiliary datasets, the test accuracy is evaluated on 8000 test samples of CIFAR10 that are not used as auxiliary data. Other implementation details are the same as in Section~\ref{sec:exp}. The results are shown in Table~\ref{tbl:aux_data}. Surprisingly, using samples from CIFAR10 with correct labels yields the worst accuracy. This may because the model `overfits' the auxiliary data when it has access to correct labels, which makes the anchor subspace contains less information about the private gradients. The best accuracy is achieved using samples from CIFAR10 with random labels, this makes sense because in this case the features of auxiliary data and private data have the same distribution. Using samples from CIFAR100 or ImageNet as auxiliary data has a small influence on the test accuracy.
\begin{table} [h]
\caption{Test accuracy on CIFAR10 with different choices of auxiliary datasets. The privacy guarantee is $(8,10^{-5})$-DP. We report the average accuracy of five runs with standard deviations in brackets. }
\label{tbl:aux_data}
\centering
\begin{tabular}{ p{3.5cm}p{3.5cm}p{3.5cm}}
\hline \hline
Auxiliary Data & Random Label? & Test Accuracy \\[0.4ex]
\hline
CIFAR10 & No & 72.9 ($\pm$0.31) \\[0.4ex]
CIFAR10 & Yes & \textbf{75.1} ($\pm$0.42) \\[0.4ex]
CIFAR100 & Yes & 74.7 ($\pm$0.46)\\[0.4ex]
ImageNet & Yes & 74.8 ($\pm$0.39) \\[0.4ex]
\hline
\hline
\end{tabular}
\end{table}
\textbf{The influence of the number of anchor gradients.} In the main text, the size of auxiliary dataset is $m=2000$. We conduct more experiments with different sizes of auxiliary dataset to examine the influence of $m$. The auxiliary data is randomly sampled from ImageNet. Table~\ref{tbl:aux_data_size} reports the test accuracy on CIFAR10 with different choices of $m$. For both B-GEP and GEP, increasing $m$ leads to slightly improved performance.
\begin{table} [h]
\caption{Test accuracy on CIFAR10 with different sizes of auxiliary dataset. The privacy guarantee is $(8,10^{-5})$-DP. We report the average accuracy of five runs with standard deviations in brackets. }
\label{tbl:aux_data_size}
\centering
\begin{tabular}{ p{2.5cm}p{2.5cm}p{2.5cm}p{2.5cm}}
\hline \hline
Algorithm & $m=1000$ & $m=2000$ & $m=4000$ \\[0.4ex]
\hline
B-GEP & 62.2 ($\pm$0.26) & 62.6 ($\pm$0.24) & 63.3 ($\pm$0.27) \\[0.4ex]
GEP & 74.6 ($\pm$0.41) & 74.8 ($\pm$0.39) & \textbf{75.2} ($\pm$0.34) \\[0.4ex]
\hline
\hline
\end{tabular}
\end{table}
\textbf{The projection error of random basis vectors.} It is tempting to construct the anchor subspace using random basis vectors because Johnson–Lindenstrauss Lemma \citep{dasgupta2003elementary} guarantees that one can preserve the pairwise distance between any two points after projecting into a random subspace of much lower dimension. We empirically verify the projection error of Gaussian random basis vectors on CIFAR10 and SVHN. The experiment settings are the same as in Section~\ref{sec:exp}. The projection errors over the training process are plotted in Figure~\ref{fig:rp_projection_error}. The projection error of random basis vectors is very high ($>95\%$) throughout training. This is because preserving the pairwise distance is not sufficient for high quality gradient reconstruction, which requires one to preserve the average ‘distance’ between any individual gradient and all other gradients.
\begin{figure}
\centering
\includegraphics[width=0.6\linewidth]{imgs/rp_approx_error_anchor.pdf}
\caption{Projection error rate of random basis vectors. The dimension of subspace is denoted by $k$. }
\label{fig:rp_projection_error}
\end{figure}
\section{Missing Proofs}
\label{sec:proof_dp}
\lowrankerror*
\begin{proof}
We extend the Theorem 3.2 in \cite{eaton1973non} to the low-rank case.
\begin{theorem}[Theorem 3.2 in \cite{eaton1973non}] \label{thm:non-singularity}
Let ${\bm{X}}=({\bm{x}}_1, ..., {\bm{x}}_n)$ where the ${\bm{x}}_i$ are i.i.d. random vectors in ${\mathbb{R}}^p$, $n\ge p$. If ${\mathbb{P}}\{{\bm{x}}_1 \in {\mathbb{M}}\}=0$ for all proper manifolds ${\mathbb{M}}\subset {\mathbb{R}}^p$, then ${\mathbb{P}}\{{\bm{X}} \text{ is non-singular}\}$=1.
\end{theorem}
We note that the subspace spanned by $\hat{{\bm{V}}}_{k'}$ is in the space spanned by ${\bm{V}}_{k}$ by definition. Hence $k'\le k$.
Let $\hat{{\bm{x}}}_i := {\bm{V}}_k^T \hat{{\bm{\xi}}}_i \in {\mathbb{R}}^k$ for $i\in [m]$. Then $\hat{{\bm{X}}} := (\hat{\bm{x}}_1, ..., \hat{\bm{x}}_m)$ is non-singular because of the assumption and Theorem \ref{thm:non-singularity}. That is $rank(\hat{{\bm{X}}}) =k$. Therefore $rank((\hat{\bm{\xi}}_1, ..., \hat{\bm{\xi}}_m))\ge k$, $rank(\hat{\bm{S}})\ge k$ and $k'\ge k$. Therefore $k'=k$ and the subspace spanned by $\hat{{\bm{V}}}_{k'}$ and the subspace spanned by ${\bm{V}}_{k}$ are identical.
\end{proof}
\privacyppgr*
\begin{proof}[Proof of Theorem~\ref{thm:privacy_ppgr}]
We first introduce some background knowledge of Rényi differential privacy (RDP) \citep{mironov2017renyi}. RDP measures the Rényi divergence between two output distributions.
\begin{definition}[$(\lambda,\gamma)$-RDP]
A randomized mechanism $f$ is said to guarantee $(\lambda,\gamma)$-RDP if for any neighboring datasets ${\mathbb{D}},{\mathbb{D}}'$ and $\lambda>1$ it holds that \[D_{\lambda}(f({\mathbb{D}})||f({\mathbb{D}}'))\leq\gamma,\]
where $D_{\lambda}(\cdot||\cdot)$ denotes the Rényi divergence of order $\lambda$.
\end{definition}
We next introduce some useful properties of RDP.
\begin{lemma} [Gaussian mechanism of RDP]
\label{lma:gm_rdp}
Let $S=\max_{{\mathbb{D}}\sim {\mathbb{D}}'}\norm{f({\mathbb{D}})-f({\mathbb{D}}')}$ be the $l_{2}$ sensitivity, then Gaussian mechanism $\mathcal{M}=f({\mathbb{D}})+{\bm{z}}$ satisfies $(\lambda,\frac{\lambda S^{2}}{2\sigma^{2}})$-RDP, where ${\bm{z}}\sim\mathcal{N}(0,\sigma^{2}I_{p\times p})$.
\end{lemma}
\begin{lemma} [Composition of RDP]
\label{lma:composition}
If $M_{1}$, $M_{2}$ satisfy $(\lambda,\gamma_{1})$-RDP and $(\lambda,\gamma_{2})$-RDP respectively, then their composition satisfies $(\lambda,\gamma_{1}+\gamma_{2})$-RDP.
\end{lemma}
\begin{lemma} [Conversion from RDP to $(\epsilon,\delta)$-DP]
\label{lma:conversion}
If $\mathcal{M}$ obeys $(\lambda,\gamma)$-RDP, then $\mathcal{M}$ obeys $(\gamma+\log(1/\delta)/(\lambda-1), \delta)$-DP for all $0<\delta<1$.
\end{lemma}
\medskip
Now we proof Theorem~\ref{thm:privacy_ppgr}. Let ${\bm{W}},{\bm{W}}'$ be the gradient embeddings of two neighboring datasets ${\mathbb{D}}\sim {\mathbb{D}}'$ and ${\bm{R}},{\bm{R}}'$ be corresponding residual gradients. Without loss of generality, suppose ${\bm{W}}$ (${\bm{R}}$) has one more row than ${\bm{W}}'$ (${\bm{R}}'$). For given sensitivity $S_{1}, S_{2}$,
\[\max_{{\mathbb{D}}\sim {\mathbb{D}}'}\norm{{\bm{w}}-{\bm{w}}'}=\max_{{\bm{W}}\sim {\bm{W}}'}\norm{ {\bm{W}}_{n,:}}\leq S_{1}, \quad \max_{{\mathbb{D}}\sim {\mathbb{D}}'}\norm{{\bm{r}}-{\bm{r}}'}=\max_{ {\bm{R}}\sim {\bm{R}}'}\norm{ {\bm{R}}_{n,:}}\leq S_{2}.\]
If we set $\sigma_{1}=S_{1}\sigma$ and $\sigma_{2}=S_{2}\sigma$ for some $\sigma$, then Algorithm~\ref{alg:ppg} satisfies $(\lambda,\frac{\lambda}{\sigma^{2}})$-RDP because of Lemma~\ref{lma:gm_rdp} and~\ref{lma:composition}. In order to guarantee $(\epsilon,\delta)$-DP, we need
\begin{equation}
\begin{aligned}
\label{eq:privacyppg0}
\frac{\lambda }{\sigma^{2}} + \frac{\log(1/\delta)}{\lambda-1} \leq \epsilon.
\end{aligned}
\end{equation}
Choose $\lambda=1+\frac{2\log(1/\delta)}{\epsilon}$ and rearrange Eq~(\ref{eq:privacyppg0}), we need
\begin{equation}
\begin{aligned}
\label{eq:privacyppg1}
\sigma^{2}\geq \frac{2\left(\epsilon+2\log(1/\delta)\right)}{\epsilon^2}.
\end{aligned}
\end{equation}
Then using the constraint on $\epsilon$ concludes the proof.
\end{proof}
\privacygd*
\begin{proof}[Proof of Theorem~\ref{thm:privacy_gd}]
From the proof of Theorem~\ref{thm:privacy_ppgr}, we have each call of GEP satisfies $(\lambda,\frac{\lambda}{\sigma^{2}})$-RDP. Then by the composition property of RDP (Lemma~\ref{lma:composition}), the output of Algorithm~\ref{alg:dp_gd} satisfies $(\lambda,\frac{T\lambda}{\sigma^{2}})$-RDP. Plugging $\frac{T\lambda}{\sigma^{2}}$ into Equation~\ref{eq:privacyppg0} and~\ref{eq:privacyppg1} concludes the proof.
\end{proof}
\convergence*
\begin{proof} [Proof of Theorem~\ref{thm:convergence}]
The $\beta$-smooth condition gives
\begin{equation}
\begin{aligned}
\label{eq:convergence_eq0}
L(\boldsymbol\theta_{t+1})\leq L(\boldsymbol\theta_{t})+\idot{\nabla L(\boldsymbol\theta_{t}),\boldsymbol\theta_{t+1}-\boldsymbol\theta_{t}}+\frac{\beta}{2}\norm{\boldsymbol\theta_{t+1}-\boldsymbol\theta_{t}}^{2}.
\end{aligned}
\end{equation}
\newcommand{\boldsymbol\theta}{\boldsymbol\theta}
Based on the update rule of GEP we have
\begin{equation}
\begin{aligned}
\label{eq:convergence_eq1}
\boldsymbol\theta_{t+1}-\boldsymbol\theta_{t}=-\eta\tilde{\bm{v}}=-\eta\nabla L(\boldsymbol\theta_{t})-\frac{\eta}{n}({\bm{z}}^{(1)}_{t}{\bm{B}}+{\bm{z}}^{(2)}_{t}),
\end{aligned}
\end{equation}
where ${\bm{z}}^{(1)}_{t}\sim\mathcal{N}(0,\sigma^{2}{\bm{I}}_{k\times k})$, ${\bm{z}}^{(2)}_{t}\sim\mathcal{N}(0,\sigma^{2}r_{t}^{2}{\bm{I}}_{p\times p})$ are the perturbation noises and $r_{t}=\max_{i} \norm{({\bm{R}}_{t})_{i,:}}$ is the sensitivity of residual gradients at step $t$.
Take expectation on Eq~(\ref{eq:convergence_eq0}) with respect to the perturbation noises.
\begin{equation}
\begin{aligned}
\label{eq:convergence_eq2}
\mathbb{E}[L(\boldsymbol\theta_{t+1})]\leq \mathbb{E}[L(\boldsymbol\theta_{t})]-(\eta-\beta\eta^{2}/2)\mathbb{E}[\norm{\nabla L(\boldsymbol\theta_{t})}^{2}]+\frac{\beta\eta^{2}\sigma^{2}}{2 n^2}\left(k+pr_{t}^{2}\right).
\end{aligned}
\end{equation}
Subtract $L(\boldsymbol\theta_{*})$ from both sides, we have
\begin{equation}
\begin{aligned}
\label{eq:convergence_eq3}
\mathbb{E}[L(\boldsymbol\theta_{t+1})] - L(\boldsymbol\theta_{*})&\leq \mathbb{E}[L(\boldsymbol\theta_{t})] - L(\boldsymbol\theta_{*})-(\eta-\beta\eta^{2}/2)\mathbb{E}[\norm{\nabla L(\boldsymbol\theta_{t})}^{2}]+\frac{\beta\eta^{2}\sigma^{2}}{2 n^2}\left(k+pr_{t}^{2}\right)\\
&\leq \mathbb{E}[\idot{\nabla L(\boldsymbol\theta_{t}),\boldsymbol\theta_{t}-\boldsymbol\theta_{*}}]-(\eta-\beta\eta^{2}/2)\mathbb{E}[\norm{\nabla L(\boldsymbol\theta_{t})}^{2}]+\frac{\beta\eta^{2}\sigma^{2}}{2 n^2}\left(k+pr_{t}^{2}\right).
\end{aligned}
\end{equation}
The second inequality holds because $L$ is convex. Then choose $\eta=\frac{1}{\beta}$ and plug $\nabla L(\boldsymbol\theta_{t})=(\boldsymbol\theta_{t}-\boldsymbol\theta_{t+1})/\eta- ({\bm{z}}^{t}_{1}{\bm{B}}+{\bm{z}}^{t}_{2})/n$ into Eq~(\ref{eq:convergence_eq3}).
\begin{equation}
\begin{aligned}
\label{eq:convergence_eq4}
\mathbb{E}[L(\boldsymbol\theta_{t+1})] - L(\boldsymbol\theta_{*})&\leq \beta\mathbb{E}[\idot{\boldsymbol\theta_{t}-\boldsymbol\theta_{t+1},\boldsymbol\theta_{t}-\boldsymbol\theta_{*}}]-\frac{\beta}{2}\mathbb{E}[\norm{\boldsymbol\theta_{t}-\boldsymbol\theta_{t+1}}^{2}]+\frac{\sigma^{2}}{\beta n^2}\left(k+pr_{t}^{2}\right)\\
&=\frac{\beta}{2}\left(\mathbb{E}[\norm{\boldsymbol\theta_{t}-\boldsymbol\theta_{*}}^2]-\mathbb{E}[\norm{\boldsymbol\theta_{t+1}-\boldsymbol\theta_{*}}^{2}]\right)+\frac{\sigma^{2}}{\beta n^2}\left(k+pr_{t}^{2}\right).
\end{aligned}
\end{equation}
Sum over $t=0,\ldots,T-1$ and use convexity, we have
\begin{equation}
\begin{aligned}
\label{eq:convergence_eq5}
\mathbb{E}[L(\boldsymbol{\bar\theta})]-L(\boldsymbol{\theta_{*}})\leq
\frac{\beta}{2T}\norm{\boldsymbol\theta_{0}-\boldsymbol\theta_{*}} + \frac{\sigma^{2}}{\beta n^{2}}(k+\frac{p}{T}\sum_{t=0}^{T-1}r_{t}^{2}).
\end{aligned}
\end{equation}
Then substituting $T=\frac{n\beta\epsilon}{\sqrt{p}}$ and $\sigma=\mathcal{O}(\sqrt{T\log(1/\delta)}/\epsilon)$ yields the desired bound.
\end{proof}
\section{Preliminaries}
\label{sec:pre}
\vspace{-2mm}
We introduce some notations and definitions. We use bold lowercase letters, e.g., ${\bm{v}}$, and bold capital letters, e.g., ${\bm{M}}$, to denote vectors and matrices, respectively. The $L^2$ norm of a vector ${\bm{v}}$ is denoted by $\|{\bm{v}}\|$. The spectral norm and the Frobenius norm of a matrix ${\bm{M}}$ are denoted by $\|{\bm{M}}\|$ and $\|{\bm{M}}\|_F$, respectively. A sample $d=({\bm{x}},y)$ consists of feature ${\bm{x}}$ and label $y$. A dataset ${\mathbb{D}}$ is a collection of individual samples. A dataset ${\mathbb{D}}'$ is said to be a neighboring dataset of ${\mathbb{D}}$ if they differ in a single sample, denoted as ${\mathbb{D}}\sim {\mathbb{D}}'$. Differential privacy ensures that the outputs of an algorithm on neighboring datasets have approximately indistinguishable distributions.
\begin{definition}[$(\epsilon,\delta)$-DP \citep{dwork2006our, dwork2006calibrating}]
A randomized mechanism $\mathcal{M}$ guarantees $(\epsilon,\delta)$-differential privacy if for any two neighboring input datasets ${\mathbb{D}}\sim {\mathbb{D}}^{'}$ and for any subset of outputs $S$ it holds that $\text{Pr}[\mathcal{M}({\mathbb{D}})\in S]\leq e^{\epsilon}\text{Pr}[\mathcal{M}({\mathbb{D}}^{'})\in S]+\delta$.
\end{definition}
By its definition, $(\epsilon, \delta)$-DP controls the maximum influence that any individual sample can produce. One can adjust the privacy parameters to trade off between privacy and utility. Differential privacy is immune to \emph{post-processing} \citep{algofound}, i.e., any function applied on the output of a differentially private algorithm would not increase the privacy loss as long as it does not have new interaction with the private dataset. Differential privacy also allows \emph{composition}, i.e., the composition of a series of differentially private mechanisms is also differentially private but with different parameters. Several variants of $(\epsilon, \delta)$-DP have been proposed \citep{bun2016concentrated,dong2019gaussian} to address certain weakness of $(\epsilon, \delta)$-DP, e.g., they achieve better composition property. In this work, we use \emph{Rényi differential privacy} \citep{mironov2017renyi} to track the privacy loss and then convert it to $(\epsilon, \delta)$-DP.
Suppose that there is a private dataset ${\mathbb{D}} = \{({\bm{x}}_i, y_i)\}_{i=1}^n$ with $n$ samples. We want to train a model $f$ to learn the mapping in ${\mathbb{D}}$. Specifically, $f$ takes ${\bm{x}}$ as input and outputs a label $y$, and $f$ has parameter $\theta\in {\mathbb{R}}^p$. The training objective is to minimize an empirical risk $\frac{1}{n}\sum_{i=1}^n \ell(f({\bm{x}}_i), y_i)$, where $\ell(\cdot, \cdot)$ is a loss function. We further assume that there is an auxiliary dataset ${\mathbb{D}}^{(a)}= \{(\tilde{{\bm{x}}}_j, \tilde{y}_j)\}_{j=1}^m$ that $\tilde{{\bm{x}}}$ shares similar features as ${\bm{x}}$ in ${\mathbb{D}}$ while $\tilde{y}$ could be random.
\section{Experiments}
\label{sec:exp}
\vspace{-2mm}
We conduct experiments on MNIST, extended SVHN, and CIFAR-10 datasets. Our implementation is publicly available\footnote{\url{https://github.com/dayu11/Gradient-Embedding-Perturbation}}. The model for MNIST has two convolutional layers with max-pooling and one fully connected layer. The model for SVHN and CIFAR-10 is ResNet20 in \cite{he2016deep}. We replace all batch normalization \citep{ioffe2015batch} layers with group normalization \citep{wu2018group} layers because batch normalization mixes the representations of different samples and makes the privacy loss cannot be analyzed accurately. The non-private accuracy for MNIST, SVHN, and CIFAR-10 is 99.1\%, 95.9\%, and 90.4\%, respectively.
We also provide experiments with pre-trained models in Appendix~\ref{sec:linear}. \cite{tramer2020differentially} show that differentially private linear classifier can achieve high accuracy using the features produced by pre-trained models. We examine whether GEP can improve the performance of such private linear classifiers. Notably, using the features produced by a model pre-trained on unlabeled ImageNet, GEP achieves 94.8\% validation accuracy on CIFAR10 with $\epsilon=2$.
\textbf{Evaluated algorithms} We use the algorithm in \cite{abadi2016deep} as benchmark gradient perturbation approach, referred to as ``GP''. We also compare GEP with PATE \citep{papernot2016semi}. We run the experiments for PATE using the official implementation. The privacy parameter $\epsilon$ of PATE is data-dependent and hence cannot be released directly (see Section 3.3 in \cite{papernot2016semi}). Nonetheless, we report the results of PATE anyway.
\textbf{Implementation details} At each step, GEP needs to release two vectors: the noisy gradient embedding and the noisy residual gradient. The gradient embeddings have a sensitivity of $S_{1}$ and the residual gradients have a sensitivity of $S_{2}$ because of the clipping. The output of GEP can be constructed as follows: (1) normalize the gradient embeddings and residual gradients by $1/S_{1}$ and $1/S_{2}$, respectively, (2) concatenate the rescaled vectors, (3) release the concatenated vector via gaussian mechanism with sensitivity $\sqrt{2}$, (4) rescale the two components by $S_{1}$ and $S_{2}$. B-GEP only needs to release the normalized noisy gradient embedding. We use the numerical tool in \cite{mironov2019renyi} to compute the privacy loss. For given privacy budget and sampling probability, $\sigma$ is set to be the smallest value such that the privacy budget is allowable to run desired epochs.
All experiments are run on a single Tesla V100 GPU with 16G memory. For ResNet20, the parameters are divided into five groups: input layer, output layer, and three intermediate stages. For a given quota of basis vectors, we allocate it to each group according to the square root of the number of parameters in each group. We compute an orthonormal subspace basis on each group separately. Then we concatenate the projections of all groups to construct gradient embeddings. The number of power iterations $t$ is set as $1$ as empirical evaluations suggest more iterations do not improve the performance for GEP and B-GEP.
For all datasets, the anchor gradients are computed on $2000$ random samples from ImageNet. In Appendix~\ref{apdx:abla}, we examine the influence of choosing different numbers of anchor gradients and different sources of auxiliary data. The selected images are downsampled into size of $32\times 32$ ($28\times 28$ for MNIST) and we label them randomly at each update. For SVHN and CIFAR-10, $k$ is chosen from $[500, 1000, 1500, 2000]$. For MNIST, we halve the size of $k$. We use SGD with momentum 0.9 as the optimizer. Initial learning rate and batchsize are $0.1$ and $1000$, respectively. The learning rate is divided by $10$ at middle of training. Weight decay is set as $1\times 10^{-4}$. The clipping threshold for is $10$ for original gradients and $2$ for residual gradients. The number of training epochs for CIFAR-10 and MNIST is 50, 100, 200 for privacy parameter $\epsilon=2,5,8$, respectively. The number of training epochs for SVHN is 5, 10, 20 for privacy parameter $\epsilon=2,5,8$, respectively. Privacy parameter $\delta$ is $1\times 10^{-6}$ for SVHN and $1\times 10^{-5}$ for CIFAR-10 and MNIST.
\textbf{Results} The best accuracy with given $\epsilon$ is in Table~\ref{tbl:no_fine-tune_result}. For all datasets, GEP achieves considerable improvement over GP in \cite{abadi2016deep}. Specifically, GEP achieves $74.9\%$ test accuracy on CIFAR-10 with $(8,10^{-5})$-DP, outperforming GP by $18.5\%$. PATE achieves best accuracy on MNIST
but its performance drops as the dataset becomes more complex.
We also plot the relation between accuracy and $k$ in Figure~\ref{fig:influence_k}. GEP is less sensitive to the choice of $k$ and outperforms B-GEP for all choices of $k$. The improvement of increasing $k$ becomes smaller as $k$ becomes larger. We note that the memory cost of choosing large $k$ is high because we need to store at least $k$ individual gradients to compute anchor subspace.
\begin{table}
\label{tbl:no_fine-tune_result}
\small
\centering
\caption{Test accuracy (in \%) with varying choices of privacy bound $\epsilon$. The numbers under symbol $\Delta$ denote the improvement over GP baseline. }
\begin{tabular}{ll|l|l|l|l|l|l}
\hline
\hline
Dataset & Algorithm & $\epsilon=2$ & $\Delta$ & $\epsilon=5$ & $\Delta$ & $\epsilon=8$ & $\Delta$ \\
\hline
\multirow{4}{*}{MNIST} & GP & 94.7 & +0.0 & 96.8 & +0.0 & 97.2 & +0.0 \\\cline{2-8}
& PATE & 98.5 & \textbf{+3.8} & 98.5 & \textbf{+1.7} & 98.6 & \textbf{+1.4} \\\cline{2-8}
& B-GEP & 93.1 & -1.6 & 94.5 & -2.3 & 95.9 & -1.3 \\\cline{2-8}
& GEP & 96.3 & +1.6 & 97.9 & \textbf{+1.1} & 98.4 & \textbf{+1.2} \\\hline \hline
\multirow{4}{*}{SVHN} & GP & 87.1 & +0.0 & 91.3 & +0.0 & 91.6 & +0.0 \\\cline{2-8}
& PATE & 80.7 & -6.4 & 91.6 & +0.3 & 91.6 & +0.0 \\\cline{2-8}
& B-GEP & 88.5 & +1.4 & 91.8 & +0.5 & 92.3 & +0.7 \\\cline{2-8}
& GEP & 92.3 & \textbf{+5.2} & 94.7 & \textbf{+3.4} & 95.1 & \textbf{+3.5} \\\hline \hline
\multirow{3}{*}{CIFAR-10} & GP & 43.6 & +0.0 & 52.2 & +0.0 & 56.4\tablefootnote{The test accuracy of DP-SGD can be improved to $\sim$ 62\% by tuning the hyperparameters. See the implementation in \url{https://github.com/dayu11/Differentially-Private-Deep-Learning}.} & +0.0 \\\cline{2-8}
& PATE & 34.2 & -9.4 & 41.9 & -10.3 & 43.6 & -12.8 \\\cline{2-8}
& B-GEP & 50.3 & +6.7 & 59.5 & +7.3 & 63.0 & +6.6 \\\cline{2-8}
& GEP & 59.7 & \textbf{+16.1} & 70.1 & \textbf{+17.9} & 74.9 & \textbf{+18.5} \\\hline
\hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{imgs/accuracy_vary_k.pdf}
\caption{Test accuracy when varying the dimension of anchor subspace. GEP significantly outperforms B-GEP for all $k$. Moreover, the performance of GEP is not that sensitive to $k$. }
\label{fig:influence_k}
\end{figure}
\vspace{-2mm}
\section{Conclusion}
\vspace{-2mm}
In this paper, we propose Gradient Embedding Perturbation (GEP) for learning with differential privacy. GEP leverages the gradient redundancy to reduce the added noise and outputs an unbiased estimator of target gradient. The several key designs of GEP significantly boost the applicability of GEP. Extensive experiments on real world datasets demonstrate the superior utility of GEP.
\section{Appendix}
\end{document}
\section{Introduction}
Recent works have shown that the trained model may leak/memorize the information of its training set \citep{fredrikson2015model, wu2016methodology, shokri2017membership, hitaj2017deep}, which raises privacy issue when the models are trained with sensitive data. \emph{Differential privacy} (DP) mechanism provides a way to quantitatively measure and upper bound such information leakage. It theoretically ensures that the influence of any individual sample is negligible with the DP parameter $\epsilon$ or $(\epsilon, \delta)$. Moreover, it has been observed that differentially private models can also resist model inversion attack \citep{carlini2019secret}, membership inference attack \citep{rahman2018membership,bernau2019assessing,sablayrolles2019white,yu2021how}, gradient matching attack \citep{zhu2019deep}, and data poisoning attack \citep{ma2019data}.
One popular way to achieve differentially private machine learning is to perturb the training process with noise \citep{song2013stochastic,bassily2014differentially,shokri2015privacy,wu2017bolt,fukuchi2017differentially,iyengar2019towards,phanscalable}. Specifically, \emph{gradient perturbation} perturbs the gradient at each iteration of (stochastic) gradient descent algorithm and guarantees the privacy of the final model via \emph{composition property} of DP. It is worthy to note that gradient perturbation does not assume (strongly) convex objective and hence is applicable to various settings \citep{abadi2016deep,wang2017differentially,lee2018concentrated,jayaraman2018distributed,wang2019differentially,yu2020gradient}. Specifically, for given gradient sensitivity $S$, a general form of gradient perturbation is to add an isotropic Gaussian noise ${\bm{z}}$ to the gradient ${\bm{g}} \in\mathbb{R}^{p}$ independently for each step,
\begin{equation}
\begin{aligned}
\label{eq:general_gp}
\tilde {\bm{g}}={\bm{g}}+{\bm{z}}, \;\; \text{where} \;\; {\bm{z}}\sim \mathcal{N}(0,\sigma^{2}S^{2}{\bm{I}}_{p\times p}).
\end{aligned}
\end{equation}
One can set proper variance $\sigma^2$ to make each update differentially private with parameter $(\epsilon,\delta)$. It is easy to see that the intensity of the added noise $\mathbb{E}[\|{\bm{z}}\|^2]$ scales linearly with the model dimension $p$. This indicates that as the model becomes larger, the useful signal, i.e., gradient, would be submerged in the added noise (see Figure~\ref{fig:grad_noise}). This dimensional barrier restricts the utility of deep learning models trained with gradient perturbation.
\begin{figure}
\begin{minipage}{0.4\textwidth}
\centering
\includegraphics[width=0.7\linewidth]{imgs/noise_signal.pdf}
\caption{Noise norm vs gradient norm of ResNet20 at initialization. The noise variance is chosen such that SGD satisfies $(5,10^{-5})$-DP after $90$ epochs in \cite{abadi2016deep}. }
\label{fig:grad_noise}
\end{minipage}
\hfill
\begin{minipage}{0.4\textwidth}
\centering
\includegraphics[width=0.65\linewidth]{imgs/stable_rank.pdf}
\caption{Stable rank $\|\cdot\|_F^2/\|\cdot\|^2$ \citep{tropp2015introduction} of batch gradient matrix of given groups (with $p$ parameters). The setting is ResNet20 on CIFAR-10. The stable rank is small throughout training. }
\label{fig:stbl_rank}
\end{minipage}
\end{figure}
The dimensional barrier is attributed to the fact that the added noise is isotropic while the gradients live on a very low dimensional manifold, which has been observed in \citep{gur2018gradient, vogels2019powersgd,gooneratne2020low,li2020hessian} and is also verified in Figure~\ref{fig:stbl_rank} for the gradients of a 20-layer ResNet \citep{he2016deep}. Hence to limit the noise energy, it is natural to think
``Can we reduce the dimension of gradients first and then add the isotropic noise onto a low-dimensional gradient embedding?"
The answer is affirmative. We propose a new algorithm \emph{Gradient Embedding Perturbation (GEP)}, illustrated in Figure~\ref{fig:overview}. Specifically, we first compute \emph{anchor gradients} on some non-sensitive auxiliary data, and identify an \emph{anchor subspace} that is spanned by several top principal components of the anchor gradient matrix. Then we project the private gradients into the anchor subspace and obtain low-dimensional \emph{gradient embeddings} and small-norm \emph{residual gradients}. Finally, we perturb the gradient embedding and residual gradient separately according to the sensitivities and privacy budget.
We intuitively argue why GEP could reduce the perturbation variance and achieve good utility for large models. First, because the gradient embedding has a very low dimension, the added isotropic noise on embedding has small energy that scales linearly only with the subspace dimension. Second, if the anchor subspace can cover most of the gradient information, the residual gradient, though high dimensional, should have small magnitude, which permits smaller added noise to guarantee the same level privacy because of the reduced sensitivity. Overall, we can use a much lower perturbation compared with the original gradient perturbation to guarantee the same level of privacy.
We emphasize several properties of GEP. First, the non-sensitive auxiliary data assumption is weak. In fact, GEP only requires a small number of non-sensitive unlabeled data following a similar feature distribution as the private data, which often exist even for learning on sensitive data. In our experiments, we use a few unlabeled samples from ImageNet to serve as auxiliary data for MNIST, SVHN, and CIFAR-10. This assumption is much weaker than the public data assumption in previous works \citep{papernot2016semi, papernot2018scalable,alon2019limits,wang2020differentially}, where the public data should follow exactly the same distribution as the private data. Second, GEP produces an unbiased estimator of the target gradient because of releasing both the perturbed gradient embedding and the perturbed residual gradient, which turns out to be critical for good utility. Third, we use \emph{power method} to estimate the principal components of anchor gradients, achievable with a few matrix multiplications. The fact that GEP is not sensitive to the choices of subspace dimension further allows a very efficient implementation.
Compared with existing works of differentially private machine learning, our contribution can be summarized as follows: (1) we propose a novel algorithm GEP that achieves good utility for large models with modest differential privacy guarantee; (2) we show that GEP returns an unbiased estimator of target private gradient with much lower perturbation variance than original gradient perturbation; (3) we demonstrate that GEP achieves state-of-the-art utility in differentially private learning with three benchmark datasets. Specifically, for $\epsilon=8$, GEP achieves $74.9\%$ test accuracy on CIFAR-10 with a ResNet20 model. To the best of our knowledge, GEP is the first algorithm that can achieve such utility with training deep models from scratch for a ``single-digit" privacy budget\footnote{\cite{abadi2016deep} achieve $73\%$ accuracy on CIFAR-10 but they need to pre-train the model on CIFAR-100.}.
\begin{figure}
\centering
\includegraphics[width=0.9\linewidth]{imgs/overview_ppg.png}
\caption{Overview of the proposed GEP approach. 1) We estimate an anchor subspace on some non-sensitive data; 2) We project the private gradients into the anchor subspace, producing low-dimensional embeddings and residual gradients; 3) We perturb the gradient embedding and residual gradient separately to guarantee differential privacy. The auxiliary data are only required to share similar features as the private data. In our experiments, we use $2000$ images from ImageNet as auxiliary data for MNIST, SVHN, and CIFAR-10 datasets. }
\label{fig:overview}
\end{figure}
\vspace{-1mm}
\subsection{Related work}
\label{sec:related_work}
\vspace{-2mm}
Existing works studying differentially private machine learning in high-dimensional setting can be roughly categorized into two sets. One is treating the optimization of the machine learning objective as a whole mechanism and adding noise into this process. The other one is based on the knowledge transfer of machine learning models, which trains a differentially private publishable student model with private signals from teacher models. We review them one by one.
Differentially private convex optimization in high-dimensional setting has been studied extensively over the years \citep{kifer2012private,thakurta2013differentially,talwar2015nearly,wang2019sparse,wang2019differentially}. Although these methods demonstrate good utility on some convex settings, their analyses can not be directly applied to non-convex setting. Right before the submission, we note two independent and concurrent works \citep{zhou2020bypassing,kairouz2020dimension} that also leverage the gradient redundancy to reduce the added noise. Specifically, \cite{kairouz2020dimension} track historical gradients to do dimension reduction for private AdaGrad. \cite{zhou2020bypassing} requires gradients on some public data and then project the noisy gradients into a public subspace at each update. One core difference between these two works and GEP is that we introduce residual gradient perturbation and GEP produces an unbiased estimator of the private gradients, which is essential for achieving the superior utility. Moreover, we weaken the auxiliary data assumption and introduce several designs that significantly boost the efficiency and applicability of GEP.
One recent progress towards training arbitrary models with differential privacy is \emph{Private Aggregation of Teacher Ensembles (PATE)} \citep{papernot2016semi, papernot2018scalable, jordon2019pate}. PATE first trains independent teacher models on disjoint shards of private data. Then it trains a student model with privacy guarantee by distilling noisy predictions of teacher models on some public samples. In comparison, GEP only requires some non-sensitive data that have similar natural features as the private data while PATE requires the public data follow exactly the same distribution as the private data and in practice it uses a portion of the test data to serve as public data. Moreover, GEP demonstrates better performance than PATE especially for complex datasets, e.g., CIFAR-10, because GEP can train the model with the whole private data rather than a small shard of data.
\section{Gradient embedding perturbation}
\vspace{-1mm}
An overview of GEP is given in Figure~\ref{fig:overview}. GEP has three major ingredients: 1) first, estimate an anchor subspace that contains the principal components of some non-sensitive anchor gradients via power method; 2) then, project private gradients into the anchor subspace and produce low-dimensional embeddings of private gradients and residual gradients; 3) finally, perturb gradient embedding and residual gradient separately to establish differential privacy guarantee. In Section~\ref{sec:ppg_algo}, we present the GEP algorithm in detail. In Section~\ref{sec:ppgr_error}, we given an analysis on the residual gradients. In Section~\ref{sec:learning_ppg}, we give a differentially private learning algorithm that updates the model with the output of GEP.
\vspace{-1mm}
\subsection{The GEP algorithm and its privacy analysis}
\label{sec:ppg_algo}
\vspace{-1mm}
The pseudocode of GEP is presented in Algorithm~\ref{alg:ppg}. For convenience, we write a set of gradients and a set of basis vectors as matrices with each row being one gradient/basis vector.
The anchor subspace is constructed as follows. We first compute the gradients of the model on an auxiliary dataset ${\mathbb{D}}^{(a)}$ with $m$ samples, which is referred to as the anchor gradients ${\bm{G}}^{(a)}\in {\mathbb{R}}^{m\times p}$. We then use the power method to estimate the principal components of ${\bm{G}}^{(a)}$ to construct a subspace basis ${\bm{B}}\in {\mathbb{R}}^{k\times p}$, which is referred to as the anchor subspace. All these matrices are publishable because ${\mathbb{D}}^{(a)}$ is non-sensitive. We expect that the anchor subspace ${\bm{B}}$ can cover most energy of private gradients when the auxiliary data are not far from private data and $m,k$ are reasonably large.
Suppose that the private gradients are ${\bm{G}}\in {\mathbb{R}}^{n\times p}$. Then, we project the private gradients into the anchor subspace ${\bm{B}}$. The projection produces low-dimensional embeddings ${\bm{W}} = {\bm{G}} {\bm{B}}^T$ and residual gradients ${\bm{R}} ={\bm{G}} - {\bm{G}} {\bm{B}}^T{\bm{B}}$. The magnitude of residual gradients is usually much smaller than original gradient even when $k$ is small because of the gradient redundancy.
Then, we aggregate the gradient embeddings and the residual gradients, respectively. We perturb the aggregated embedding and the aggregated residual gradient respectively to guarantee certain differential privacy. Finally, we release the perturbed embedding and the perturbed residual gradient and construct an unbiased estimator of the private gradient: $\tilde{{\bm{v}}} := (\tilde {\bm{w}}^{T}{\bm{B}}+\tilde {\bm{r}})/n$. This construction process does not resulting in additional privacy loss because of DP's post-processing property. The privacy analysis of the whole process of GEP is given in Theorem~\ref{thm:privacy_ppgr}.
\begin{algorithm}
\caption{Gradient embedding perturbation}
\label{alg:ppg}
\begin{algorithmic} [1]
\STATE {\bfseries Input:} anchor gradients ${\bm{G}}^{(a)}\in\mathbb{R}^{m\times p}$; number of basis vectors $k$; private gradients ${\bm{G}}\in\mathbb{R}^{n\times p}$; clipping thresholds $S_{1},S_{2}$; standard deviations $\sigma_{1}, \sigma_{2}$; number of power iterations $t$.
\medskip
\STATE //\textsl{First stage: Compute an orthonormal basis for the anchor subspace.}
\STATE Initialize ${\bm{B}}\in\mathbb{R}^{k\times p}$ randomly.
\FOR{$i=1$ {\bfseries to} $t$}
\STATE Compute ${\bm{A}}={\bm{G}}^{(a)}{\bm{B}}^{T}$ and ${\bm{B}}={\bm{A}}^{T}{\bm{G}}^{(a)}$.
\STATE Orthogonalize ${\bm{B}}$ and normalize row vectors.
\ENDFOR
\STATE Delete ${\bm{G}}^{(a)}$ to free memory.
\medskip
\STATE //\textsl{Second stage: project the private gradients ${\bm{G}}$ into anchor subspace ${\bm{B}}$}
\STATE Compute gradient embeddings ${\bm{W}}={\bm{G}}{\bm{B}}^{T}$ and clip its rows with $S_{1}$ to obtain $\hat {\bm{W}}$.
\STATE Compute residual gradients ${\bm{R}}={\bm{G}}-{\bm{W}}{\bm{B}}$ and clip its rows with $S_{2}$ to obtain $\hat {\bm{R}}$.
\medskip
\STATE //\textsl{Third stage: perturb gradient embedding and residual gradient separately}
\STATE Perturb embedding with noise ${\bm{z}}^{(1)} \sim \mathcal{N}(0, \sigma_1^2 {\bm{I}}_{k\times k})$: $\;{\bm{w}} := \sum_{i}\hat{\bm{W}}_{i,:},\;\; \tilde {\bm{w}} := {\bm{w}}+{\bm{z}}^{(1)}.$
\STATE Perturb residual gradient with noise ${\bm{z}}^{(2)} \sim \mathcal{N}(0, \sigma_2^2{\bm{I}}_{p\times p})$:
$\;{\bm{r}}:=\sum_{i} \hat{\bm{R}}_{i,:}, \;\; \tilde {\bm{r}} := {\bm{r}} + {\bm{z}}^{(2)}.$
\STATE Return $\tilde{{\bm{v}}} := (\tilde {\bm{w}}^{T}{\bm{B}}+\tilde {\bm{r}})/n$.
\end{algorithmic}
\end{algorithm}
\begin{restatable}{thm}{privacyppgr}
\label{thm:privacy_ppgr}
Let $S_{1}$ and $S_{2}$ be the sensitivity of ${\bm{w}}$ and ${\bm{r}}$, respectively, the output of Algorithm~\ref{alg:ppg} satisfies $(\epsilon,\delta)$-DP for any $\delta\in (0,1)$ and $\epsilon\leq 2\log(1/\delta)$ if we choose $\sigma_{1}\geq 2S_{1}\sqrt{2\log(1/\delta)}/\epsilon$ and $\sigma_{2}\geq 2S_{2}\sqrt{2\log(1/\delta)}/\epsilon$.
\end{restatable}
A common practice to control sensitivity is to clip the output with a pre-defined threshold. In our experiments, we use different thresholds $S_{1}$ and $S_{2}$ to clip the gradient embeddings and residual gradients, respectively. The privacy loss of GEP consists of two parts: the privacy loss incurred by releasing the perturbed embedding and the privacy loss incurred by releasing the perturbed residual gradient. We compose these two parts via the Rényi differential privacy and convert it to $(\epsilon,\delta)$-DP.
We highlight several implementation techniques that make GEP widely applicable and implementable with reasonable computational cost. Firstly, auxiliary non-sensitive data do not have to be the same source as the private data and the auxiliary data can be randomly labeled. This non-sensitive data assumption is very weak and easy to satisfy in practical scenarios. To understand why random label works, a quick example is that for the least squares regression problem the individual gradient is aligned with the feature vector while the label only scales the length but does not change the direction. This auxiliary data assumption avoids conducting principal component analysis (PCA) on private gradients, which requires releasing private high-dimensional basis vectors and hence introduces large privacy loss. Secondly, we use \emph{power method} \citep{panju2011iterative, vogels2019powersgd} to approximately estimate the principal components. The new operation we introduce is standard matrix multiplication that enjoys efficient implementation on GPU. The computational complexity of each power iteration is $2mkp$, where $p$ is the number of model parameters, $m$ is the number of anchor gradients and $k$ is the number of subspace basis vectors. Thirdly, we divide the parameters into different groups and compute one orthonormal basis for each group. This further reduces the computational cost. For example, suppose the parameters are divided into two groups with size $p_{1},p_{2}$ and the numbers of basis vectors are $k_{1}, k_{2}$, the computational complexity of each power iteration is $2m(k_{1}p_{1}+k_{2}p_{2})$, which is smaller than $2m(k_{1}+k_{2})(p_{1}+p_{2})$. In Appendix~\ref{sec:complexity}, we analyze the additional computational and memory costs of GEP compared to standard gradient perturbation.
Curious readers may wonder if we can use random projection to reduce the dimensionality as Johnson–Lindenstrauss Lemma \citep{dasgupta2003elementary} guarantees that one can preserve the pairwise distance between any two points after projecting into a random subspace of much lower dimension. However, preserving the pairwise distance is not sufficient for high quality gradient reconstruction, which is verified by the empirical observation in Appendix~\ref{apdx:abla}.
\vspace{-1mm}
\subsection{An analysis on the residual gradients of GEP}
\label{sec:ppgr_error}
\vspace{-1mm}
Let ${\bm{g}} :=\frac{1}{n} \sum_{i}{\bm{G}}_{i,:}$ be the target private gradient. For a given anchor subspace ${\bm{B}}$, the residual gradients are defined as ${\bm{R}}:={\bm{G}}-{\bm{G}}{\bm{B}}^{T}{\bm{B}}$. We then analyze how large the residual gradients could be. The following argument holds for all time steps and we ignore the time step index for simplicity.
For the ease of discussion, we introduce ${\bm{\xi}}_i := ({\bm{G}}_{i,:})^T$ for $i \in [n]$ to denote the the private gradients and the $\hat{{\bm{\xi}}}_j := ({\bm{G}}^{(a)}_{j,:})^T$ for $j \in [m]$ to denote the anchor gradients. We use $\lambda_{k}(\cdot)$ to denote the $k_{th}$ largest eigenvalue of a given matrix. We assume that the private gradients ${\bm{\xi}}_1, ..., {\bm{\xi}}_n$ and the anchor gradients $\hat{{\bm{\xi}}}_1, ..., \hat{{\bm{\xi}}}_m$ are sampled independently from a distribution ${\mathcal{P}}$. We assume ${\bm{\Sigma}} := \mathbb{E}_{{\bm{\xi}}\sim {\mathcal{P}}} {\bm{\xi}} {\bm{\xi}}^T \in {\mathbb{R}}^{p\times p}$ to be the population gradient (uncentered) covariance matrix. We also consider the (uncentered) empirical gradient covariance matrix $\hat{{\bm{S}}} := \frac{1}{m}\sum_{i=1}^m\hat{{\bm{\xi}}}_i\hat{{\bm{\xi}}}_i^T$.
One case is that the population gradient covariance matrix ${\bm{\Sigma}}$ is low-rank $k$. In this case we can argue that the residual gradients are $0$ once the number of anchor gradients $m>k$.
\begin{restatable}{lemma}{lowrankerror}
\label{lma:lowrank_error}
Assume that the population covariance matrix ${\bm{\Sigma}}$ is with rank $k$ and the distribution ${\mathcal{P}}$ satisfies ${\mathbb{P}}({\bm{\xi}} \in {\mathbb{F}}_s) =0$ for all $s$-flats ${\mathbb{F}}_s$ in ${\mathbb{R}}^p$ with $0\le s < k$. Let ${\bm{\Sigma}} = {\bm{V}}_k \Lambda {\bm{V}}_k^T$ and $\hat{{\bm{S}}} = \hat{{\bm{V}}}_{k'} \hat{\Lambda} \hat{{\bm{V}}}_{k'}^T$ be the eigendecompositions of ${\bm{\Sigma}}$ and the empirical covariance matrix $\hat{{\bm{S}}}$, respectively, such that $\lambda_{k'}(\hat{{\bm{S}}}) >0$ and $\lambda_{k'+1}(\hat{{\bm{S}}})=0$. Then if $m\ge k$, we have with probability 1,
\begin{flalign}
k'= k \quad \text{and} \quad \|{\bm{V}}_k{\bm{V}}_k^T - \hat{{\bm{V}}}_k \hat{{\bm{V}}}_k^T\|_2 = 0.
\end{flalign}
\end{restatable}
\begin{proof}
The proof is based on the non-singularity of covariance matrix. See Appendix \ref{sec:proof_dp}.
\end{proof}
We note that $s$-flat is the translate ${\mathbb{F}}_s = {{\bm{x}}} + {\mathbb{F}}_{s(0)}$ of an $s$-dimensional linear subspace ${\mathbb{F}}_{s(0)}$ in ${\mathbb{R}}^p$ and the normal distribution satisfies such condition \citep{eaton1973non,muirhead2009aspects}. Therefore, we have seen that for low-rank case of population covariance matrix, the residual gradients are 0 once $m>k$. In the general case, we measure the expected norm of the residual gradients.
\begin{lemma}
\label{lma:approx_error}
Assume that ${\bm{\xi}} \sim {\mathcal{P}}$ and $\|{\bm{\xi}}\|^2<T$ almost surely. Let ${\bm{\Sigma}} = {\bm{V}} \Lambda {\bm{V}}^T$ be the eigendecomposition of the population covariance matrix ${\bm{\Sigma}}$. Let $\hat{{\bm{S}}} = \hat{{\bm{V}}}_k \hat{\Lambda} \hat{{\bm{V}}}_k^T$ be the eigendecomposition of the empirical covariance matrix $\hat{{\bm{S}}}$. Then we have with probability $1-2\exp(-\delta)$,
\begin{flalign}
\mathbb{E}\|{\bm{\xi}} - \Pi_{\hat{{\bm{V}}}_k}({\bm{\xi}})\|_2^2 \le \sum_{k'>k} \lambda_{k'}({\bm{\Sigma}}) + \sqrt{kC/m} + T \sqrt{2\delta/m},
\end{flalign}
where $C= \left[\mathbb{E} \|{\bm{\xi}}\|^4 - \sum_i \lambda_i^2({\bm{\Sigma}})\right] + \left[\frac{1}{m}\sum_{j=1}^m \|\hat{\bm{\xi}}_j\|^4 - \sum_i \lambda_i^2(\hat{\bm{S}})\right]$, $\Pi_{\hat{{\bm{V}}}_k}$ is a projection operator onto the subspace $\hat{{\bm{V}}}_k$ and the $\mathbb{E}$ is taken over the randomness of ${\bm{\xi}}\sim {\mathcal{P}}$.
\end{lemma}
\begin{proof}
The proof is an adaptation of Theorem 3.1 in \cite{blanchard2007statistical}.
\end{proof}
From Lemma~\ref{lma:approx_error}, we can see the larger the number of anchor gradients and the dimension of the anchor subspace $k$, the smaller the residual gradients. We can choose $m,k$ properly such that the upper bound on the expected residual gradient norm is small. This indicates that we may use a smaller clipping threshold and consequently apply smaller noises with achieving the same privacy guarantee.
We next empirically examine the projection error ${\bm{r}}=\sum_{i}{\bm{R}}_{i,:}$ by training a $20$-layer ResNet on CIFAR10 dataset. We try two different types of auxiliary data to compute the anchor gradients: 1) samples from the same source as private data with correct labels, i.e., $2000$ random samples from the test data; 2) samples from different source with random labels, i.e., $2000$ random samples from ImageNet. The relation between the dimension of anchor subspace $k$ and the projection error rate ($\norm{\frac{1}{n}{\bm{r}}}/\norm{{\bm{g}}}$) is presented in Figure~\ref{fig:approx_error}. We can see that the project error is small and decreases with $k$, and the benefit of increasing $k$ diminishes when $k$ is large, which is implied by Lemma \ref{lma:approx_error}. In practice one can only use small or moderate $k$ because of the memory constraint. GEP needs to store at least $k$ individual gradients and each individual gradient consumes the same amount of memory as the model itself. Moreover, we can see that the projection into anchor subspace of random labeled auxiliary data yields comparable projection error, corroborating our argument that unlabeled auxiliary data are sufficient for finding the anchor subspace.
We also verify that the redundancy of residual gradients is small, by plotting the stable rank of residual gradient matrix in Figure~\ref{fig:rgp_stable_rank}. The stable rank of residual gradient matrix is an order of magnitude higher than the stable rank of original gradient matrix. This implies that it could be hard to further approximate ${\bm{R}}$ with low-dimensional embeddings.
\begin{figure}
\centering
\begin{minipage}{0.64\textwidth}
\begin{center}
\centerline{\includegraphics[width=0.8\linewidth]{imgs/arrox_error_two_types.pdf}}
\caption{ Relative projection error ($\norm{\frac{1}{n}{\bm{r}}}/\norm{{\bm{g}}}$) of the second stage in ResNet20. The number of anchor gradients is $2000$. The dimension of anchor subspace is $k$. The learning rate is decayed by $10$ at epoch 30. The left plot uses random samples from ImageNet. The right plot uses random samples from test data. The benefit of increasing $k$ becomes smaller when $k$ is larger. }
\label{fig:approx_error}
\end{center}
\end{minipage}
\hfill
\begin{minipage}{0.33\textwidth}
\begin{center}
\centerline{\includegraphics[width=0.8\linewidth]{imgs/rgp_stable_rank.pdf}}
\caption{Stable rank of the residual gradient matrix versus original gradient matrix. The gradients are computed on full batch data for the first stage in ResNet20. The dimension of anchor subspace is $k=1000$. }
\label{fig:rgp_stable_rank}
\end{center}
\end{minipage}
\end{figure}
We next compare the GEP with a scheme that simply discards the residual gradients and only outputs the perturbed gradient embedding, i.e., the output is $\tilde{{\bm{u}}} := \tilde {\bm{w}}^{T}{\bm{B}}/n$.
\begin{remark}
\label{thm:error_cp}
Let $\tilde {\bm{u}} := \tilde {\bm{w}}^{T}{\bm{B}}/n$ be the reconstructed gradient from noisy gradient embedding and $\tilde{{\bm{v}}}$ be the output of GEP. If ignoring the effect of gradient clipping, we have
\begin{flalign}
\mathbb{E}[{\tilde{\vu}}]= {\bm{g}}-{\bm{r}}/n, \quad \mathbb{E} [{\tilde{\vv}}] = {\bm{g}}.
\end{flalign}
where ${\bm{r}}=\sum_{i}{\bm{R}}_{i,:}$ is the aggregated residual gradients, $\tilde{\bm{w}}, {\bm{B}}$ are given in Algorithm~\ref{alg:ppg} and the expectation is over the added random noises.
\end{remark}
This indicates that ${\tilde{\vu}}$ contains a systematic error that makes ${\tilde{\vu}}$ always deviate from ${\bm{g}}$ by the residual gradient. This systematic error is the projection error, which is plotted in Figure~\ref{fig:approx_error}. The systematic error cannot be mitigated by reducing the noise magnitude (e.g., increasing the privacy budget or collecting more private data). We refer to the algorithm releasing ${\tilde{\vu}}$ directly as \emph{Biased-GEP} or \emph{B-GEP} for short, which can be viewed as an efficient implementation of the algorithm in \citep{zhou2020bypassing}. In our experiments, B-GEP can outperform standard gradient perturbation when $k$ is large but is inferior to GEP. We note that the above remark is made with ignoring the clipping effect (or set a large clipping threshold). In practice, we do apply clipping for the individual gradients at each time step, which makes the expectations in Remark \ref{thm:error_cp} obscure \citep{chen2020understanding}. We note that the claim that ${\tilde{\vv}}$ is an unbiased estimator of ${\bm{g}}$ is not that precise when applying gradient clipping.
\vspace{-1mm}
\subsection{Private learning with gradient embedding perturbation}
\vspace{-1mm}
\label{sec:learning_ppg}
GEP (Algorithm \ref{alg:ppg}) describes how to release one-step gradient with privacy guarantee. In this section, we compose the privacy losses at each step to establish the privacy guarantee for the whole learning process. The differentially private learning process with GEP is given in Algorithm~\ref{alg:dp_gd} and the privacy analysis is presented in Theorem~\ref{thm:privacy_gd}.
\begin{algorithm}
\caption{Differentially private gradient descent with GEP.}
\label{alg:dp_gd}
\begin{algorithmic} [1]
\STATE {\bfseries Input:} private dataset ${\mathbb{D}}$; auxiliary dataset ${\mathbb{D}}^{(a)}$; number of updates $T$; learning rate $\eta$; configuration of GEP ${\mathbb{C}}$; loss function $\ell$;
\STATE {\bfseries Output:} Differentially private model $\boldsymbol\theta_{T}$.
\FOR{$t=0$ {\bfseries to} $T-1$}
\STATE Compute the private gradients ${\bm{G}}_{t}$ and anchor gradients ${\bm{G}}^{(a)}_{t}$ of loss with respect to $\boldsymbol\theta_{t}$.
\STATE Call GEP with ${\bm{G}}_{t},{\bm{G}}^{(a)}_{t}$ and configuration ${\mathbb{C}}$ to get $\tilde {\bm{v}}_{t}$.
\STATE Update model $\boldsymbol\theta_{t+1}=\boldsymbol\theta_{t}-\eta\tilde{\bm{v}}_{t}$.
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{restatable}{thm}{privacygd}
\label{thm:privacy_gd}
For any $\epsilon<2\log(1/\delta)$ and $\delta\in(0,1)$, the output of Algorithm~\ref{alg:dp_gd} satisfies $(\epsilon,\delta)$-DP if we set $\sigma\geq 2\sqrt{2T\log(1/\delta)}/\epsilon$.
\end{restatable}
If the private gradients are randomly sampled from the full batch gradients, the privacy guarantee can be strengthened via the \emph{privacy amplification by subsampling} theorem of DP \citep{balle2018privacy,wang2019subsampled,zhu2019poission, mironov2019renyi}. Theorem~\ref{thm:convergence} gives the expected excess error of Algorithm~\ref{alg:dp_gd}. Expected excess error measures the distance between the algorithm's output and the optimal solution in expectation.
\begin{restatable}{thm}{convergence}
\label{thm:convergence}
Suppose the loss $L(\boldsymbol\theta)=\frac{1}{n}\sum_{({\bm{x}},y)\in{\mathbb{D}}}\ell(f_{\boldsymbol\theta}({\bm{x}}),y)$ is 1-Lipschitz, convex, and $\beta$-smooth. If $\eta=\frac{1}{\beta}$, $T=\frac{n\beta\epsilon}{\sqrt{p}}$, and $\boldsymbol{\bar \theta}=\frac{1}{T}\sum_{t=1}^{T}\boldsymbol\theta_{t}$, then we have
$\mathbb{E}[L(\boldsymbol{\bar\theta})]-L(\boldsymbol{\theta_{*}})\leq \mathcal{O}\left(\frac{\sqrt{k\log(1/\delta)}}{n\epsilon}+\frac{\bar r\sqrt{p\log(1/\delta)}}{n\epsilon}\right)$,
where $\bar r=\frac{1}{T}\sum_{t=0}^{T-1}r^{2}_{t}$ and $r_{t}=\max_{i} \norm{({\bm{R}}_{t})_{i,:}}$ is the sensitivity of residual gradient at step $t$.
\end{restatable}
The $\bar r$ term represents the average projection error over the training process. The previous best expected excess error for gradient perturbation is $\mathcal{O}(\sqrt{p\log(1/\delta)}/(n\epsilon))$ \citep{wang2017differentially}. As shown in Lemma \ref{lma:lowrank_error}, if the gradients locate in a $k$-dimensional subspace over the training process, $\bar{r}=0$ and the excess error is $\mathcal{O}(\sqrt{k\log(1/\delta)}/(n\epsilon))$, independent of the problem ambient dimension $p$. When the gradients are in general position, i.e., gradient matrix is not exact low-rank, Lemma \ref{lma:approx_error} and the empirical result give a hint on how small the residual gradients could be. However, it is hard to get a good bound on $\max_i\|({\bm{R}}_t)_{i,:}\|$ and the bound in Theorem \ref{thm:convergence} does not explicitly improve over previous result. One possible solution is to use a clipping threshold based on the expected residual gradient norm. Then the output gradient becomes biased because of clipping and the utility/privacy guarantees in Theorem \ref{thm:convergence}/\ref{thm:privacy_gd} require new elaborate derivation. We leave this for future work.
|
\section*{Acknowledgements}
This research is supported by National Science Foundation (NSF) Awards CNS-1824357 and CNS-1824470. Support from USRA Cycle 3 Research Oppurtunity Program allowed machine time on a D-Wave Quantum Annealer machine hosted at NASA Ames Research Center.
\section{Design}
\label{sec:design}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/ising.pdf}
\caption{The figure demonstrates the embedding process of Eq.~\ref{eqn:example}. In the figure, \textbf{(a)} shows the direct problem connectivity of Eq.~\ref{eqn:example} and \textbf{(b)} shows its physical connectivity on QA hardware.}
\label{fig:example}
\end{figure}
Quantum Annealing based Vector Perturbation (QAVP) envisions a scenario where a Quantum Annealer (QA) machine is co-located with a centralized data center for computational processing in the C-RAN architecture~\cite{checko2014cloud} which allows for low latency communication between base stations and the QA machine. QAVP converts a VPP problem to a Quadratic Unconstrained Binary Optimization (QUBO) problem, then fine-tunes the QUBO problem coefficients to mitigate the adverse effects of QA hardware noise and increase the probability of finding the correct solution with QA. The fine-tuned QUBO is next embedded onto the physical QA hardware (see Fig.\ref{f:chimera}) for running the problem. We conduct multiple anneals for a given QUBO problem, where each anneal generates a candidate solution bit-string for VPP. The QUBO solutions are converted to perturbation vectors by inverting the transform described in Eq.~\ref{eq:vpqubo}. Among these returned solutions, the solution that minimizes the VPP objective function is chosen by the base station. In scenarios where QAVP fails to find a solution with lower transmit power scaling than Zero-Forcing (ZF) precoding, the base station discards the QAVP solution and uses ZF precoding instead of VPP.
The steps involved in QAVP computation, excluding embedding and annealing, can be executed by the physical layer of base station or can be offloaded to a server in close proximity to QA machine. Fig.~\ref{fig:qavpdep} illustrates a typical deployment scenario for QAVP, where $N_r$ users are receiving downlink data streams from a base station with $N_{t}$ antennas. We next demonstrate QAVP's QUBO formulation for the VPP problem, QAVP's pre-processing, and embedding considerations.
\subsection{QAVP's QUBO Formulation}
\label{sec:vpQubo}
\begin{figure}
\centering
\includegraphics[width=0.8\linewidth]{figures/vpp_qa.png}
\caption{A typical deployment scenario of QAVP. $N_r$ users are receiving downlink data streams from a base station with $N_t$ antennas.}
\label{fig:qavpdep}
\end{figure}
The downlink VPP problem is to find an optimal perturbation vector $\textbf{v}^{\star}$ that minimizes the transmit power at the base station, whose NP-hard objective function is represented in Eq.~\ref{eq:vstar}. As the search objective in Eq.\ref{eq:vstar} is over the vector \textbf{v} $\in$ $\mathbb{G}^{N_r\times 1}$, we construct each entry in \textbf{v} by a linear combination of binary variables as follows.
Let $q_i$ denote the $i^{th}$ QUBO form solution variable, and let $a_k$ + j$b_k$ be the $k^{th}$ entry in \textbf{v}, where $a_k$ and $b_k$ are integers. We design each $a_k$ and $b_k$ $\forall k$ in an identical fashion using $t+1$ distinct solution variables as:
\begin{align}
a_k\big/b_k = \sum_{m=1}^{t} {2^{m-1}\cdot q_m} - 2^{t}*q_{t+1}
\label{eq:vpqubo}
\end{align}
Since all the solution variables ($q_i$) are binary, this formulation allows $a_k$ and $b_k$ to take all integer values in the range $[-2^{t}, 2^{t}-1]$, where each integer corresponds to a unique configuration of solution variables. The value of \textit{t} determines the search range of \textbf{v}. Yuen~\textit{et~al.}~\cite{vpRange} show that perturbation values are most likely to be $\{-1,0,1\}$, and hence $t$ = 1 which allows perturbation values in the range $[-2,1]$, is usually sufficient. We substitute the entries of $\mathbf{v}$ with their corresponding solution variables into Eq.\ref{eq:vstar} to obtain the QAVP's QUBO:
\begin{align}
\arg\min_{\forall q} \big\{\sum_{\forall i} f_i(\textbf{H}, \textbf{u}) q_i + \sum_{\forall i, j} g_{ij}(\textbf{H}, \textbf{u}) q_iq_j\big\}
\label{eq:fqubo}
\end{align}
where $f$ and $g$ are functions of MIMO channel matrix and user data vector. The linear and quadratic coefficients obtained from Eq.~\ref{eq:fqubo} can be programmed into the QA processor. The computational complexities of the proposed QA technique and the optimal case Sphere Encoder are $O(e^{\sqrt{N_u}})$ \cite{mukherjee2015multivariable} and $O(e^{N_u})$ \cite{mohaisen2009fixed} respectively, where $N_u$ is the number of users. We next present our QUBO pre-processing considerations.
\subsection{QAVP's Pre-processing}
Our pre-processing scheme aims to mitigate the adverse effects of QA hardware ICE noise (\S\ref{sec:qa}) by scaling and eliminating minor QUBO coefficients. We note that a generic QUBO form of Eq.\ref{eqn:qubo_bits} can be equivalently expressed as:
\begin{equation}
E = \sum_{i}f_{i}q_{i} + \sum_{i<j}g_{ij}q_{i}q_{j} = \mathbf{q^{T}}\mathbf{Q}\mathbf{q}
\end{equation}
where $\mathbf{q} = [q_{1}, q_{2}, ...]^{T}$, and $\mathbf{Q}$ is an upper triangular matrix with $Q_{ij}$ = $f_{i}$ ($i=j$), $Q_{ij}$ = $g_{ij}$ ($i<j$), $Q_{ij}$ = $0$ ($i>j$). Let $Q_{max}$ = $\max_{i,j} |Q_{ij}|$ be the maximum QUBO coefficient value, $T_{high}$ and $T_{low}$ be our chosen upper and lower bounds for the QUBO coefficient values. We define a \textit{scaleFactor} as:
\begin{equation}
scaleFactor =
\left\{ {\begin{array}{cc}
1 & Q_{max} \leq T_{high} \\
\dfrac{T_{high}}{Q_{max}} & Q_{max} > T_{high} \\
\end{array} } \right.,
\end{equation}\\
Our approach is to first scale each entry in $\mathbf{Q}$ with the scaleFactor, then eliminate coefficients below a chosen threshold of $10^{T_{low}}$. This scaling process is summarized below.
\parahead{Step 1:} Set $\mathbf{Q} \leftarrow \mathbf{Q}*scaleFactor$.\\
\parahead{Step 2:} For every $i\leq j$, if $Q_{ij}< 10^{T_{low}}$ then set $Q_{ij} \leftarrow 0$.\\
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/postProcessingLoss_new.png}
\caption{QAVP's pre-processing loss for different choices of $T_{high}$ and $T_{low}$, showing that $T_{high} = 6$ is near-optimal and $T_{low} = -2$ leads to negligible PPL.}
\label{fig:quboPPL}
\end{figure}
Let $\mathbf{Q}_{pre}$ correspond to the QUBO obtained after this pre-processing step. Let $\mathbf{q}_{pre}^{\star}$ = $\arg\min_\textbf{q} \mathbf{q^{T}}\mathbf{Q}_{pre}\mathbf{q}$, and $\mathbf{q}^{\star} = \arg\min_q \mathbf{q^{T}}\mathbf{Q}\mathbf{q}$. We define a \textit{pre-processing loss} \textit{(PPL)} as:
\begin{equation}
\text{PPL} = \dfrac{|(\mathbf{q}_{pre}^{*^T}\mathbf{Q}\mathbf{q}_{pre}^{*} - \mathbf{q^{*^T}}\mathbf{Q}\mathbf{q^{*}})|}{|(\mathbf{q^{*^T}}\mathbf{Q}\mathbf{q^{*}})|}
\end{equation}
PPL is used to quantify deterioration in the optimal value of the VPP problem due to the modification of the original QUBO problem. We see from Fig.~\ref{fig:quboPPL} that PPL increases with an increase in $T_{low}$ as a higher $T_{low}$ causes more coefficients to be reduced to zero. We further observe that PPL reduces initially with $T_{high}$, and then becomes constant. While having a high negative value for $T_{low}$ leads to low PPL, it makes it more susceptible to QA ICE noise. The optimal choices for $T_{high}$ and $T_{low}$ depend on the distribution of QUBO coefficients, magnitude of QA ICE noise, and problem embedding. Our empirical studies find that a $T_{high} = 6$ and a $T_{low} = -2$ obtains good solutions with the DW2Q QA.
\subsection{QAVP's Embedding and Parallelism.} We use a D-Wave's heuristic algorithm described in Ref.~\cite{cai2014practical} and implemented in D-Wave MinorMiner software libraries for mapping of QAVP's QUBO design onto the QA hardware. It is to note that multiple QUBOs or multiple instances of a QUBO can be parallelly processed by mapping different problems onto distinct physical locations in the QA chip. For instance, QAVP for 7$\times$7 MIMO system requires approximately 200 qubits, implying that 9--10 such problems can be simulataneously solved on the DW2Q QA with 2,048 qubits. The number of qubits in the QA hardware has been steadily doubling each year and this trend is expected to continue \cite{10.1145/3372224.3419207}, with the latest 5,000 qubit chip just released at time of writing. While the evolution of QA technology is currently at an early stage (2011-), future QA processors with higher qubit counts will potentially increase the supported parallelism for QAVP.
\section{Evaluation}
\label{sec:eval}
In this section, we first present microbenchmarks for QA machine parameters. We then investigate QAVP's end-to-end error performance, comparing head-to-head against Zero Forcing (ZF) and Fixed Complexity Sphere Encoder (FSE) algorithms. The DW2Q QA system performance is majorly affected by the choice of \textit{annealing time} ($T_a$), number of anneals ($N_a$), and \textit{chain strengths} ($J_F$). Recall that number of anneals~($N_a$) refers to the number of times annealing process is repeated for each QUBO problem. Our evaluation methodology is as follows: we obtain the VPP solutions from the QA (DW2Q), and use these solutions to simulate an end-to-end downlink data transmission on a trace-driven MIMO simulator (implemented in MATLAB). We simulate data transmission over Rayleigh fading wireless channels and empirically measure the BER corresponding to different users.
\subsection{Effect of Chain Strength ($J_F$)}
\begin{figure}
\centering
\includegraphics[width=0.65\linewidth]{figures/p10.png}
\caption{Evaluation of the overall transmission power ($P_t$) amplitudes over various choices of $|J_F|$ at $T_a$ = $100$~$\mu$s and 10~dB channel SNR. The plot shows that QAVP's performance variance is not significant for $J_F$ values above $0.4J_m$.}
\label{f:poweranalysis}
\end{figure}
We evaluate QAVP's performance over an $10\times 8$ MIMO system with 20 QAVP problem instances. Our channel matrix \textbf{H} is random Rayleigh fading, modulation scheme is BPSK, and wireless channel noise is Gaussian. Let $J_m$ be the maximum coupler strength value of our QUBO problem at hand.
In Fig.~\ref{f:poweranalysis} we investigate the sensitivity of chain strengths $\abs{J_F}$ over the overall transmission power $P_t$. For this evaluation, we set a high annealing time of $100~\mu s$ and a moderate SNR of $10~dB$ to ensure minimal disturbance from the time limit and the channel SNR respectively. We observe that the performance of QAVP varies with $J_F$ and that the sensitivity of QAVP performance beyond $|J_F|$ = $0.4J_m$ is not significant. The average $P_t$ is much higher for $|J_F| < 0.4J_m$, and the variance in performance of QAVP is barely distinguishable for higher values of $|J_F| > 0.4J_m$. While the optimal value of $J_F$ depends on the parameters of the system (\textit{e.g.,} modulation and MIMO size), we empirically determine that $|J_F| = 1.2J_m$ is the optimal setting for $6\times 6$ and $12\times 12$ MIMO systems with 64 QAM modulation scheme.
\begin{figure}
\centering
\includegraphics[height=0.48\linewidth,width=\linewidth]{figures/6c6_final_ta.png}
\caption{Downlink multi-user $6\times 6$ MIMO: BER for various QAVP anneal times $T_{a}$. We compare error performance of QAVP for various values of anneal time ($T_a$) while keeping the number of annneals ($N_a = 10^4$) fixed. We set $|J_F| = 1.2J_m$.}
\label{fig:mimo6x6Ta}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=0.46\linewidth,width=\linewidth]{figures/TTB.png}
\caption{Downlink multi-user $6\times 6$ MIMO: Bit Error Rate vs Total Computation Time ($N_a \times T_a$). We demonstrate the BER performance, at two values of SNR, over different combinations of anneal time ($T_a$) and number of anneals ($N_a$). We set $|J_F| = 1.2J_m$.}
\label{fig:mimo6x6TTb}
\vspace{-10pt}
\end{figure}
\subsection{Effect of Anneal Time ($T_a$)}
Another critical QA parameter that affects the QAVP performance is the anneal time ($T_a$), along with its associated number of anneals ($N_a$). We heretofore quantify the QA performance with the total compute time: $T_a\times N_a$. Fig.~\ref{fig:mimo6x6Ta} compares the BER performance of FSE, ZF and QAVP, and Fig.~\ref{fig:mimo6x6TTb} reports the BER performance of QAVP at different QA compute times, for a $N_{t} = 6$, $N_{r} = 6$ MIMO system with 64 QAM modulation, for various choices of $T_a$ and $N_a$.
We first observe in Fig.~\ref{fig:mimo6x6Ta} that a higher anneal time leads to better BER performance from QA. Although a higher anneal time allows QA to return better quality solutions, it increases the communication latency (due to a higher computation time) as Fig.~\ref{fig:mimo6x6TTb} depicts. The tradeoff between BER performance and anneal time represents a fundamental design problem in QAVP. We observe from Figs.~\ref{fig:mimo6x6Ta} and \ref{fig:mimo6x6TTb} that QAVP achieves an acceptable BER with a low anneal time of $T_a = 10\mu s$, while keeping the computation time lower by an order of magnitude.
\subsection{End-to-end BER performance}
In this section, we consider a large $12\times 12$ MIMO system with
Rayleigh fading channel, White Gaussian noise, and 64 QAM modulation. Fig.~\ref{fig:mimo12x12} reports results (with $|J_F| = 1.2J_m$, $T_a = 299$~$\mu$s, $N_a = 10^4$) for mean BER and normalized throughput (ratio of achieved to maximum possible throughput), with packets size of 1500 bytes, for uncoded data transmission. We observe from Fig.~\ref{fig:mimo12x12} that QAVP outperforms ZF and FSE by achieving a 1--2 orders of magnitude lower BER at $E_b/N_0$ of 27dB. We see from the throughput curves that QAVP achieves a 3~dB gain over FSE and around 6~dB gain over ZF. It is to note that a 12$\times$12 MIMO system with VPP, using theoretically optimal Sphere Encoder, is practically infeasible due to its high computational complexity.
\begin{figure}
\centering
\includegraphics[height = 0.48\linewidth,width=1\linewidth]{figures/12x12_final_big.png}
\caption{Downlink multi-user $12\times 12$ MIMO: BER and throughput performance. We compare the BER performance of QAVP against FSE and ZF. We set $|J_F| = 1.2J_m$, $T_a = 299\mu$s and $N_a = 10^4$.}
\label{fig:mimo12x12}
\vspace{-10pt}
\end{figure}
\section{Conclusion}
\label{sec:concl}
In this paper we propose QAVP, a novel technique that performs Vector Perturbation Precoding using Quantum Annealing. We evaluate QAVP on a real DW2Q QA device over a variety of machine parameters. Our studies show that for large MIMO systems, if the analysis is restricted to computation time, QAVP can outperform practically feasible state-of-the-art techniques like FSE and ZF algorithms for VPP. While prior work investigates QA technology for problems in the uplink \cite{10.1145/3372224.3419207, qaMIMO1}, our studies investigate its potential in the downlink. Our analysis disregards the engineering and system integration overheads of currently available commercial QA systems (e.g. latency, programming time and thermalization times between the runs), since they can be optimized heavily if the system is built for a specific application deployment. Nevertheless, the techniques we propose in this work, in the more distant future, may enable the applicability of VPP to large MIMO systems by exploiting stronger QA devices and parallelism.
\subsection{Future Work}
Our work has several possible improvements along the lines of QUBO pre-processing, problem embedding, and QA machine parameter selection. It is of interest to further refine our pre-processing step to optimize the QUBO representation based on the embedding of the problem. Further, a more sophisticated tuning and selection of QA parameters such as chain strengths and annealing times can potentially improve the QAVP performance. Porting our model to other alternative emerging Ising-model related technologies such as quantum-inspired algorithms, reverse quantum annealing, and quantum-classical hybrid approaches is also a potential work direction.
\section{Introduction}
\label{s:intro}
Modern wireless networks are experiencing tremendous growth
in traffic loads at base stations, and hence to meet the resulting
computational and latency requirements,
designers continue to investigate new architectures
and hardware for today's 5G and
tomorrow's 6G networks. A large component of cellular baseband processing comprises of downlink data traffic due to a significant rise in the popularity and usage of video streaming platforms (\textit{e.g.,} Netflix). To meet the ever-growing user demand, it is critical for the base stations to enhance the quality of downlink data streams in terms of throughput, error rate, and latency.
In a multi-user multiple-input multiple-output (MIMO) downlink data transmission, \textit{precoding} techniques can be used to eliminate the effect of inter-user interference and allow users to detect their respective data non-cooperatively, minimizing error-rate and maximizing throughput. In this work, we focus on \textit{Vector Perturbation Precoding} \textit{(VPP)} \cite{peel2005vector}. VPP is a widely studied non-linear precoding technique that performs transmit-side channel inversion over a perturbed user data vector to reduce the transmit power scaling. Although VPP has been shown to achieve better error performance compared to other precoding techniques (\textit{e.g.,} zero-forcing, Tomlinson-Harashima Precoding~\cite{thp}), finding an optimal perturbation for user data in VPP is known to be NP-hard, making its implementation in massive/large MIMO systems to be infeasible.
A promising and cost-effective architecture to address
the increased computational burden of wireless networks
is a \textit{Centralized Radio Access Network
(C-RAN)} \cite{checko2014cloud} architecture,
which aggregates the computationally\hyp{}demanding
processing at many wireless base stations. Since baseband physical\hyp{}layer processing is highly
time\hyp{}critical and the required stream of baseband signal samples
has a high data\hyp{}rate, this type of C-RAN deployment imposes both
latency and bandwidth requirements on the interconnect between each
base station and the data center. Related studies to this end are investigating quantum computation for solving networking problems in the uplink: such as Channel coding\cite{10.1145/3372224.3419207} and ML detection\cite{qaMIMO1}. However, investigating quantum computation for problems in the downlink remains unexplored to the best of our knowledge.
In this paper, we propose a radically different processing to the NP-hard VPP problem in the downlink: \textit{Quantum Annealing based Vector Perturbation Precoding} \textit{(QAVP)}. Our proposed technique leverages recent advances in quantum computational devices and applies them to the problem of VPP. QAVP's approach is to represent the VPP problem as an optimization problem over a quadratic polynomial with binary variables \textit{i.e.}, \textit{Quadratic Unconstrained Binary Optimization} (QUBO),
which is an optimization form that a \textit{Quantum Annealer (QA)} machine takes
as input, then refine the QUBO
coefficients to mitigate the adverse
effects of QA hardware noise and the process of mapping the polynomial onto the physical QA qubit hardware topology. After these \emph{preprocessing} and \emph{embedding} steps,
QAVP uses a real Quantum Annealing machine to solve the resultant QUBO problem and then constructs the VPP perturbation from the solutions returned by the QA machine. Our results show that, from the standpoint of computation time, QAVP can outperform popular encoding algorithms for large MIMO systems.
\section{Quantum Annealing}
\label{sec:qa}
Quantum Annealing (QA) is a heuristic approach that implements in hardware a quantum computing algorithm inspired by the Adiabatic Theorem of quantum mechanics~\cite{albash2018adiabatic}. The method aims to find the lowest energy \textit{spin configuration} (solution) of the class of \textit{quadratic unconstrained binary optimization} \textit{(QUBO)} problems in their equivalent \textit{Ising} specification. The Ising form is described by:
\begin{equation}
E = \sum_{i}h_{i}s_{i} + \sum_{i<j}J_{ij}s_{i}s_{j}
\label{eqn:qubo}
\end{equation}
where \textit{E} is the \textit{energy}, $s_i$ is a solution variable, $h_{i}$ and $J_{ij}$ are problem parameters called \textit{bias} and \textit{coupler strength} respectively. QUBO form is obtained from Eq.\ref{eqn:qubo} by a simple variable transformation ($s_i \longrightarrow 2q_i - 1$), where $s_i$ and $q_i$ represent Ising and QUBO form variables respectively. Ising form solution variables ($s_i$) take on values in $\{-1, +1\}$, and QUBO form solution variables ($q_i$) take on values in $\{0, 1\}$. The equivalent QUBO form of Eq.~\ref{eqn:qubo} is described by:
\begin{equation}
E = \sum_{i}f_{i}q_{i} + \sum_{i<j}g_{ij}q_{i}q_{j}
\label{eqn:qubo_bits}
\end{equation}
\subsection{Primer: Quantum Annealers.} Quantum Annealers are specialized quantum computers, essentially comprised of two types of resources: \textit{qubits} (quantum bits) and \textit{couplers}, whose regular connectivity structure is organized in \textit{unit cells}. Fig.~\ref{f:chimera} shows the hardware structure of the QA we adapt in this study, the D-Wave 2000Q (DW2Q) quantum annealing machine. The QA device programs the linear (\textit{i.e.,} biases) and quadratic (\textit{i.e.,} coupler strengths) coefficients of Eq.\ref{eqn:qubo} onto qubits and couplers using controllable inductive elements in proximity of superconducting Josephson Junctions present on the chip~\cite{10.1145/3372224.3419207}.
\begin{figure}
\centering
\includegraphics[width=0.65\linewidth]{figures/check.png}
\caption{The figure shows \textit{qubits} (nodes) and \textit{couplers} (edges) in DW2Q QA hardware. Each group of eight qubits is called \emph{unit cell}.}
\label{f:chimera}
\end{figure}
\textit{Embedding.} The process of mapping the problem at hand onto the physical QA hardware topology follows the graph-theory problem of graph minor \textit{embedding}. We demonstrate here the embedding process with an example Ising problem:
\begin{equation}
E = J_{12}s_1s_2 + J_{13}s_1s_3 + J_{23}s_2s_3
\label{eqn:example}
\end{equation}
The direct connectivity of this example problem is shown in Fig.\ref{fig:example}\textbf{(a)}. However a three-node complete connectivity graph structure does not exist in the QA hardware (\textit{cf.} Fig.\ref{f:chimera}). Hence the minor-embedding approach is to map one of the problem variables in Fig.\ref{fig:example}\textbf{(a)} (\textit{e.g.,} $s_3$) onto two physical qubits (\textit{e.g.,} $s_{3a}$ and $s_{3b}$) as Fig.\ref{fig:example}\textbf{(b)} shows, such that the resulting connectivity can be realized on the QA hardware. The couplers between the qubits $s_{3a}$ and $s_{3b}$ are tasked to enforce these physical qubits to be correlated in order to end up with the same value at the end of the annealing process. This is implemented through a strong ferromagnetic interaction energy called \textit{chain strength} $J_F$ (see in Fig.\ref{fig:example}\textbf{(b)}).
\textit{Coefficient considerations.} Today's QA devices provide support for bias values in $[-2, +2]$ and coupler strength values in $[-2, +1]$ with a bit-precision guaranteed to 4--5 bits only. Further, the QA introduces an analog machine noise distinct from communication channel noise called \textit{intrinsic control errors} or \textit{ICE}. ICE noise, a collection of errors caused by qubit flux noise, susceptibility, among others \cite{dwave_tech}, essentially alters the problem coefficients ($h_{i}\rightarrow h_{i}\pm \delta h_{i}$, $J_{ij}\rightarrow J_{ij}\pm \delta J_{ij}$). Although the errors $\delta h_{i}$ and $\delta J_{ij}$ are currently on the order of $10^{-2}$, these may degrade the solution quality of some problems in scenarios where ICE noise erases significant information from the ground state of the input problem.
\subsection{Annealing process}
QA processors simulate systems in the transverse field Ising model described by the time-dependent energy functional (Hamiltonian):
\begin{equation}
H = -A(s)\sum_i\sigma_i^x + B(s)\big\{\sum_i h_{i}\sigma_i^z + \sum_{i<j}J_{ij}\sigma_i^z\sigma_j^z\big\}
\label{eqn: hamiltonian}
\end{equation}
where $\sigma_i^{x,z}$ are spin operators (Pauli matrices) acting on the $i^{th}$ qubit, $h_i$ and $J_{ij}$ are problem parameters, \textit{s} (= $t/t_a$) is called \textit{annealing schedule} where \textit{t} is the time and $t_a$ is the \textit{annealing time}. $A(s)$ and $B(s)$ are two monotonic scaling signals in the annealer such that at time $t = 0$, $A(0) \gg B(0)\approx 0$ and at time $t = t_a$, $B(1) \gg A(1)\approx 0$. The annealing algorithm initializes the system in the ground state of $\sum_i\sigma_i^x$ where each qubit is in a \textit{superposition} state $\frac{1}{\sqrt{2}} \left(\ket{0} + \ket{1}\right)$, then adiabatically evolves this Hamiltonian from time $t = 0$ until $t = t_a$ (\textit{i.e.,} decreasing \textit{A(s)} , increasing \textit{B(s)}). The time-dependent evolution described by the Schroedinger Equation driven by these signals \textit{A} and \textit{B} is essentially the annealing algorithm. While in real processor the dynamics is not ideal but it is dominated by dissipative noise instead~\cite{marshall2019power}, the expectations of the model still hold for the most part. The process of optimizing a problem in the QA is called an \textit{anneal}, while the time taken for an anneal is called \textit{annealing time}.
\section{Vector Perturbation Precoding}
\label{sec:downlink}
We consider a general MIMO downlink scenario where a base station equipped with $N_t$ transmit antennas communicate data streams with $N_r$ ($\leq$~$N_t$) single-antenna non-cooperative users independently and simultaneously.
In VPP, the user data symbol vector \textbf{u} $\in$ $\mathbb{C}^{N_r\times 1}$ is perturbed by an integer vector \textbf{v} $\in$ $\mathbb{G}^{N_r\times 1}$. This maps the data symbols to a wider constellation space, forming a perturbed transmit vector, \textbf{d} = \textbf{u} + $\tau$\textbf{v}. Here, \textbf{v} is a vector of Gaussian integers\footnote{The set of Gaussian integers $\mathbb{G}$ = $\mathbb{Z}$ + j$\mathbb{Z}$ consists of all complex numbers whose real and imaginary parts are integers.} and $\tau$ = 2( $\abs{c_{max}}$ + $\Delta$/2) is a constant chosen to provide symmetric decoding regions around the constellation points. $\abs{c_{max}}$ is the magnitude of the largest constellation symbol and $\Delta$ is the spacing between the constellation symbols \cite{vpPaper}. The perturbed vector \textbf{d} is then precoded with a precoder matrix (\textbf{P} $\in$ $\mathbb{C}^{N_t\times N_r}$), where the choice of \textbf{P} inverts the wireless channel \textbf{H}$\in$ $\mathbb{C}^{N_r\times N_t}$ and reduces the effect of wide range of eigenvalues of the channel coefficients \cite{vpPaper}. The precoder matrix \textbf{P} = $\textbf{H}^H(\textbf{H}\textbf{H}^H)^{-1}$ is used in Zero Forcing (ZF) precoding. The received symbol vector \textbf{y} corresponding to the transmitted symbol vector \textbf{x} = $\textbf{P}\textbf{d}/\sqrt{P_t}$ is given by,
\begin{equation}
\textbf{y} = \frac{1}{\sqrt{P_t}}(\textbf{H}\textbf{P}\textbf{d}) + \textbf{n}
\label{eq:received}
\end{equation}
where the scalar $P_t$ = $\norm{\textbf{P} \textbf{d}}^2$ is the transmission power scaling factor, which is assumed to be known at the receiver, and \textbf{n} is wireless channel noise. The receiver decodes by applying a modulo $\tau$ operation to the received signal \textbf{y}. The transmission power scaling ($P_t$) causes noise amplification during the decoding process. In order to minimize the transmission power scaling, an optimal choice of the perturbation vector ($\textbf{v}^\star$) that leads to the smallest $P_t$ is computed as in \cite{vpPaper}:
\begin{equation}
\textbf{v}^\star = \text{arg} \min_{\textbf{v}} \norm{\textbf{H}^H(\textbf{H}\textbf{H}^H)^{-1}(\textbf{u} + \tau \textbf{v})}^2.
\label{eq:vstar}
\end{equation}
\section{Related Work}
\label{sec:relatedWork}
The Sphere encoder \cite{vpPaper} builds on the Fincke and Pohst algorithm \cite{finPohst} by expressing the precoding matrix $\mathbf{P}$ as $\mathbf{P} = \mathbf{Q}\mathbf{R}$ by QR decomposition, where $\mathbf{Q}$ is unitary and $\mathbf{R}$ is upper triangular. It then performs a tree search, utilizing the upper triangular structure and limiting the search to the points within a hyper-sphere of a suitably chosen radius, and hence avoids exhaustive search over all possibilities. This algorithm is very similar to the Sphere decoder used for the MIMO receiver. While Sphere encoder is much better than exhaustive search, its expected complexity is still exponential. Park~\textit{et~al.}~\cite{decopVP} provide an approximation to the VPP problem by minimizing the real and imaginary parts of the VPP cost function separately. This reduces the complexity of VPP computation at the cost of error performance. The Thresholded Sphere Search algorithm \cite{vppThreshOpt} imposes an additional stopping criteria to Sphere encoder algorithm~\cite{vpPaper} based on an SNR dependent threshold heuristic. Another scheme for reducing the search space of Sphere encoder is discussed in \cite{modLoss}. It restricts the values of each perturbation to four possible values (two for real and two for imaginary) and hence reducing the search space of the Sphere encoder. Despite improvements in computation costs over Sphere encoder, the search space of all these methods is still exponential with respect to the MIMO size.
Several approximations for VPP (with polynomial complexity) exist in literature. The Fixed Complexity Sphere Encoder (FSE), adapts the Sphere encoder to have a lower and fixed polynomial complexity by pruning a large number of branches during the tree search but leads to degradation in error performance \cite{vpFSE}. Degree-2 Sparse Vector Perturbation (D2VP), presented in~\cite{sparseVP}, is a low complexity algorithm for vector perturbation precoding. It reduces the complexity of finding the VPP solution by assuming that only 2 elements of the perturbation vector can be non-zero and then improves the solution over multiple iterations.
|
\section{Introduction}
The number of new heavy-quark exotic-hadron candidates, presumptive
tetraquark and pentaquark states, increases every year. In the past
18 years, over 40 candidates have been observed at multiple
facilities and their hosted experiments. However, no single
theoretical picture to describe the structure of these states has
emerged as an undisputed favorite. Both the broad scope of
experimental results and competing theoretical interpretations have
been reviewed by many in recent years~\cite{Lebed:2016hpi,
Chen:2016qju,Hosaka:2016pey,Esposito:2016noz,Guo:2017jvc,Ali:2017jda,
Olsen:2017bmm,Karliner:2017qhf,Yuan:2018inv,Liu:2019zoy,
Brambilla:2019esw}.
Among these competing physical approaches, the dynamical diquark
picture~\cite{Brodsky:2014xia} was developed to provide a mechanism
through which diquark~($\delta^{\vphantom{1}}$)-antidiquark~(${\bar\delta}$) states could
persist long enough to be identified as such experimentally.
Diquarks are formed through the attractive channels ${3} \otimes
\bf{3} \! \to \! \bar{\bf{3}}$ [$\delta \! \equiv \!
(Q q)_{\bar{\bf 3}}$] and $\bar{\bf 3} \otimes \bar{\bf 3} \! \to \!
{\bf 3}$ [${\bar\delta} \! \equiv \! (\bar Q {\bar q}^\prime)_{\bf 3}$]
between color-triplet quarks. In this physical picture, the heavy
quark $Q$ must first be created in closer spatial proximity to a
light quark $q$ than to a light antiquark $\bar q^\prime$ (and vice
versa for $\bar Q$). This initial configuration provides an
opportunity for the formation of fairly compact $\delta^{\vphantom{1}}$ and ${\bar\delta}$
quasiparticles, in distinction to an initial state in which the
strongly attractive $\bf{3} \otimes \bar{\bf{3}} \! \to \! \bf{1}$
coupling immediately leads to $(Q\bar q^\prime)(\bar Q q)$ meson
pairs. Second, the large energy release of the production process
(from a heavy-hadron decay or in a collider event) drives apart the
$\delta^{\vphantom{1}}$-${\bar\delta}$ pair before immediate recombination into a meson pair
can occur, creating an observable resonance. A similar mechanism
extends the picture to pentaquark formation~\cite{Lebed:2015tna},
by means of using color-triplet ``antitriquarks'' ${\bar\theta} \! \equiv \! [\bar
Q_{\bar{\bf 3}} (q_1 q_2)_{\bar{\bf 3}}]^{\vphantom{(}}_{\bf 3}$.
This physical picture was subsequently developed into the dynamical
diquark {\em model}~\cite{Lebed:2017min}: The separated $\delta^{\vphantom{1}}$-${\bar\delta}$
pair is connected by a color flux tube, whose quantized states are
best described in terms of the potentials computed using the
Born-Oppenheimer (BO) approximation. These are the same potentials
as appear in QCD lattice gauge-theory simulations that predict the
spectrum of heavy-quarkonium hybrid mesons~\cite{Juge:1997nc,
Juge:1999ie,Juge:2002br,Morningstar:2019,Capitani:2018rox}. The BO
potentials are introduced into coupled Schr\"{o}dinger equations that
are solved numerically in order to produce predictions for the
$\delta^{\vphantom{1}}$-${\bar\delta}$ spectrum, as shown in Ref.~\cite{Giron:2019bcs}. As one
of the primary results of that work, all the observed exotic
candidates are shown to be accommodated within the ground-state
BO potential $\Sigma^+_g$, with the specific multiplets in order of
increasing average mass being $1S$, $1P$, $2S$, $1D$, and $2P$. A
full summary of the BO potential notation is presented in
Ref.~\cite{Lebed:2017min}.
The mass spectrum and preferred decay modes (organized by eigenstates
of heavy-quark spin) of the 6 isosinglets and 6 isotriplets
comprising the ${c\bar{c}q\bar{q}^\prime}$ positive-parity $\Sigma^+_g(1S)$ multiplet
(where $q,q^\prime \in \{ u, d \}$) were studied in
Ref.~\cite{Giron:2019cfc}. This was the first work to differentiate
$I \! = \! 0$ and $I \! = \! 1$ states in a diquark model. The
specific model of Ref.~\cite{Giron:2019cfc} naturally produces
scenarios in which $X(3872)$ is the lightest $\Sigma^+_g(1S)$ state,
and moreover predicts that the lighter of the two $I \! = \! 1$,
$J^{PC} \! = \! 1^{+-}$ states in $\Sigma^+_g(1S)$ [$Z_c(3900)$]
naturally decays almost exclusively to $J/\psi$ and the heavier one
[$Z_c(4020)$] to $h_c$, as is observed. The model of
Ref.~\cite{Giron:2019cfc} uses a 3-parameter Hamiltonian consisting
of a common multiplet mass, an internal diquark-spin coupling, and a
long-distance isospin- and spin-dependent coupling (analogous to
$\pi$ exchange) between the light quark $q$ in $\delta^{\vphantom{1}}$ and light
antiquark $\bar q^\prime$ in ${\bar\delta}$. Similar conclusions using QCD
sum rules have been obtained in Ref.~\cite{Ghalenovi:2020zen}.
The dynamical diquark model was developed further through the
corresponding analysis~\cite{Giron:2020fvd} of the negative-parity
${c\bar{c}q\bar{q}^\prime}$ $\Sigma^+_g(1P)$ multiplet and its 28 constituent
isomultiplets (14 isosinglets and 14 isotriplets), which includes
precisely four $Y$ ($I \! = \! 0$, $J^{PC} \! = \! 1^{--}$) states.
In this case, the simplest model has 5 parameters: the 3 listed
above, plus spin-orbit and tensor terms. An earlier diquark analysis
using a similar Hamiltonian, but not including isospin dependence,
appears in Ref.~\cite{Ali:2017wsf}.
The success of Ref.~\cite{Giron:2019bcs} in predicting the correct
mass splittings between the observed bands ($1S,1P,2S$) of exotic
hadrons, and Ref.~\cite{Giron:2019cfc} in effectively representing
the fine structure within the lowest multiplets [especially
$\Sigma^+_g(1S)$] provides strong {\it a posteriori\/} support for
the applicability of the dynamical diquark model. In particular, one
may certainly question whether treating the exotics as quasi-two-body
states within a BO approximation, rather than including full 4- (or
5-) body interactions to represent the internal evolution of the
quasiparticles, is sensible. However, while such effects are
undoubtedly present at some level, the current experimental
evidence appears to support the presence of a scale separation that
allows the quasiparticles to be treated identifiable subunits within
the hadrons. As an example, Ref.~\cite{Giron:2019bcs} showed in
numerical simulations that the diquarks need not be pointlike
particles, but could have substantial spatial extent (characteristic
radii as large as $0.4$~fm) before the full hadron mass spectrum
changes significantly.
An analysis within this model of the 12 isomultiplets comprising the
${b\bar{b}q\bar{q}^\prime}$ $\Sigma_g^+(1S)$ multiplet and the 6 states of the
${c\bar c} s\bar{s}$ $\Sigma_g^+(1S)$ multiplet appears in
Ref.~\cite{Giron:2020qpb}. By using only experimental inputs for the
states $Z_b(10610)$ and $Z_b(10650)$, which includes their masses and
relative probability of decay into $h_b$ versus $\Upsilon$ states, the
entire ${b\bar{b}q\bar{q}^\prime}$ mass spectrum is predicted. In particular, the mass
of the bottom analogue to $X(3872)$ is highly constrained
($\approx \! 10600$~MeV), and the lightest ${b\bar{b}q\bar{q}^\prime}$ state
($I \! = \! 0$, $J^{PC} \! = \! 0^{++}$) lies only a few MeV above
the $B\bar{B}$ threshold. Furthermore, starting with the assumption
that $X(3915)$ is the lowest lying ${c\bar c} s\bar{s}$
state~\cite{Lebed:2016yvr} and $Y(4140)$ is the sole $J^{PC} \! = \!
1^{++}$ ${c\bar c} s\bar{s}$ state in $\Sigma_g^+(1S)$, the remaining 4
masses in the multiplet are predicted. Emerging naturally in the
spectrum is $X(4350)$, a $J/\psi$-$\phi$ resonance seen by
Belle~\cite{Shen:2009vs}, while $Y(4626)$ and $X(4700)$ are found to
fit well within the $\Sigma_g^+(1P)$ and $\Sigma_g^+(2S)$
${c\bar c} s\bar{s}$ multiplets, respectively.
The dynamical diquark model has also recently been extended to the
case in which the light quarks $q$ are replaced with heavy quarks $Q$
to produce fully heavy tetraquark states $Q_1 \overline Q_2 Q_3
\overline Q_4$, where $Q_i \! = \! c$ or $b$. Sparked by the recent
LHCb report of at least one di-$J/\psi$ resonance near $6900$
MeV~\cite{Aaij:2020fnh}, Ref.~\cite{Giron:2020wpx} determined the
spectrum of $c\bar{c}c\bar{c}$ states in the dynamical diquark model.
In this system, the minimal model predicts each $S$-wave multiplet to
consist of 3 degenerate states ($J^{PC} \! = \! 0^{++},\,1^{+-},\,
2^{++}$) and 7 $P$-wave states. $X(6900)$ was found to fit most
naturally as a $\Sigma_g^+(2S)$ state, with other structures in the
measured di-$J/\psi$ spectrum appearing to match $C \! = \! +$
members of the $\Sigma^+_g(1P)$ multiplet.
In this paper we use the dynamical diquark model to predict
radiative transitions between exotic states. So far, very few
theoretical papers have investigated exotic-to-exotic transitions
(and of these papers, only diquark models have been
considered~\cite{Maiani:2014aja,Chen:2015dig}). One of the
distinctive features of the $P$-wave study in
Ref.~\cite{Giron:2020fvd} is the direct calculation of decay
probabilities to eigenstates of heavy-quark spin. Indeed,
Ref.~\cite{Giron:2020fvd} uses the heavy quark-spin content of states
as the main criterion for associating observed resonances with
particular states in the $\Sigma^+_g(1P)$ multiplet, and identifies
using likelihood fits two particularly plausible assignments for the
states. Using the same decay probabilities, we calculate here
the transition amplitudes for $\Sigma^+_g(1P) \! \to \! \gamma
\Sigma^+_g(1S)$. We directly adapt the well-known expression for
electric dipole (E1) radiative transitions used to great effect for
conventional quarkonium. Since the E1 transition formula depends
sensitively upon the initial and final wave functions, a comparison
between our predictions and data provides an important test of the
hypothesis that the purported $\Sigma^+_g(1P)$ and $\Sigma^+_g(1S)$
states, such as in $Y(4220) \! \to \! \gamma X(3872)$, truly share a
common structure. The corresponding magnetic dipole (M1) expression
within this model is also presented, in anticipation of the
observation of relevant transitions such as $\Sigma^+_g(2S) \! \to \!
\gamma \Sigma^+_g(1S)$, or even between two $\Sigma^+_g(1S)$ states
such as $Z_c(4020)^0 \! \to \! \gamma X(3872)$.
This paper is organized as follows: In Sec.~\ref{sec:Expt} we review
the current experimental data on transitions between ${c\bar{c}q\bar{q}^\prime}$ states.
Section~\ref{sec:pwave_review} reprises the relevant phenomenological
aspects of Ref.~\cite{Giron:2020fvd}. In Sec.~\ref{sec:analysis} we
calculate the decay widths and decay probabilities for
exotic-to-exotic radiative transitions and focus upon two of the more
probable $P$-wave state assignments in Ref.~\cite{Giron:2020fvd}.
We conclude in Sec.~\ref{sec:conclusions}.
\section{Experimental Review of Exotic-to-Exotic Transitions}
\label{sec:Expt}
Although the number of exotic-candidate discoveries continues to
increase at a remarkable pace, only a handful of exotic-to-exotic
decays have been observed to date, through
radiative~\cite{Ablikim:2013dyn,Ablikim:2019zio} and
pionic~\cite{Ablikim:2013mio,Liu:2013dau,BESIII:2020pov} transitions.
Considering first the radiative decays that form the topic of this
work, thus far only E1 transitions (as indicated by changing parity
$\Delta P \! = \! -$) have been observed in two states at BESIII, the
$J^{PC} \! = \! 1^{--}$ $Y(4260)$~\cite{Ablikim:2013dyn} and
$Y(4220)$~\cite{Ablikim:2019zio}, both seen to decay to a photon and
the $J^{PC} \! = \! 1^{++}$ $X(3872)$. Indeed, an increasing amount
of evidence from BESIII ({\it e.g.}, in Ref.~\cite{Ablikim:2016qzw})
suggests that the well-known $Y(4260)$ is actually a collection of
resonances, of which $Y(4220)$ is just one component.
Observed exotic-to-conventional radiative transitions are also rather
few in number, due to the large decay widths of exotics that follows
from the dominance of their strong decay modes. To date, only
$X(3872) \! \to \! \gamma J/\psi$ and $\gamma \psi(2S)$, also both E1
transitions, have definitely been seen ({\it e.g.}, in
Ref.~\cite{Aaij:2014ala}). BESIII has also recently announced an
interesting negative result~\cite{Ablikim:2021rba}, an upper limit
for $Z_c(4020)^0 \, (J^{PC} \! = \! 1^{+-}) \! \to \! \gamma
X(3872)$. Indeed, to date no M1 radiative decay
($\Delta P \! = \! +$) of any exotic candidate has yet been seen at
any experiment.
As for pionic transitions, both BESIII~\cite{Ablikim:2013mio} and
Belle~\cite{Liu:2013dau} have observed (indeed, discovered)
$Z_c(3900)^\pm$ through $Y(4260) \! \to \! \pi^+ \pi^- J/\psi$, and
BESIII recently observed $Z_c(3900)^0$ via $Y(4220) \! \to \! \pi^0
\pi^0 J/\psi$~\cite{BESIII:2020pov}. Assuming just a similarity of
hadronic structure between various exotic candidates, one may expect
several more exotic-to-exotic pionic (or other light-meson)
transitions to be observed in the future. An essential criterion for
how such transitions may best be studied relies on the size of the
pion momentum $p_\pi$ in such processes; for example, in the decays
listed above, $p_\pi \! \approx \! 300$~MeV\@. Processes with
smaller $p_\pi$ values may be reliably studied using conventional
chiral perturbation theory, while studies of processes with larger
$p_\pi$ values require modifications to the perturbative calculation
to improve their convergence. Since the methods associated with
radiative transitions (particularly E1 transitions) present fewer
computational ambiguities, we defer a study of exotic-to-exotic
pionic transitions for future work.
The expressions for E1 and M1 transition widths used below
[Eqs.~(\ref{eqn:e1width}) and (\ref{eqn:m1width}), respectively] are
almost identical to the forms derived in standard quantum mechanics
textbooks. As such, they are manifestly nonrelativistic, and
furthermore are developed using the photon long-wavelength
approximation, $\exp ( i \bm{k} \cdot \bm{r} ) \! \to \! 1$.
Nevertheless, the expressions can also be derived directly from the
fundamental Lagrangian $j_\mu A^\mu$ couplings of the electromagnetic
current $j^\mu$ of charged quarks to the photon field $A^\mu$ (see,
{\it e.g.}, Ref.~\cite{Brambilla:2004wf}). In
Sec.~\ref{sec:pwave_review} we discuss the effect of including
certain corrections to the textbook expressions.
In the dynamical diquark model, all states in the multiplet
$\Sigma^+_g(1S) [(1P)]$ have $P \! = \! + [-]$~\cite{Lebed:2017min}.
The current observed properties of the $J^{PC} \! = \! 1^{--}$ ($Y$)
states identified with the multiplet $\Sigma^+_g(1P)$, whose
spectroscopy is analyzed extensively in Ref.~\cite{Giron:2020fvd},
are summarized in Table~\ref{tab:Expt}.
\begin{table*}
\caption{$J^{PC} \! = \! 1^{--}$ charmoniumlike exotic-meson
candidates catalogued by the Particle Data Group
(PDG)~\cite{Zyla:2020zbs}, which are identified with specific
states within the $\Sigma^+_g(1P)$ multiplet of the dynamical
diquark model, as summarized by the cases presented in
Ref.~\cite{Giron:2020fvd} and repeated in
Sec.~\ref{sec:pwave_review}. Both the particle name most commonly
used in the literature and its label as given in the PDG are
listed.}
\label{tab:Expt}
\centering
\setlength{\extrarowheight}{1.2ex}
\begin{tabular}{cccccc}
\hline\hline
Particle
& PDG label
& $I^{G}J^{PC}$
& Mass [MeV]
& Width [MeV]
& Production and decay \\ \hline
$Y(4220)$ & $\psi(4230)$
& $0^-1^{--}$
& $4218^{+5}_{-4}$
& $59^{+12}_{-10}$
& $e^+e^- \to Y$; $Y \to \left\{ \begin{array}{l}
\omega \chi_{c0} \\
\eta J/\psi\\
\pi^+ \pi^- h_c \\
\pi^+ \pi^- \psi(2S) \\
\pi^+ D^0 D^{*-} \\
\pi^0 Z_c^0(3900) \\
\gamma X(3872)\\
\end{array} \right. $ \\ %
$Y(4260)$ & $\psi(4260)$
& $0^-1^{--}$
& $4230 \pm 8$
& $55 \pm 19$
& $e^+e^- \to \gamma Y$ or $Y$; $Y \to \left\{ \begin{array}{l}
\pi^+ \pi^- J/\psi \\
f_0(980) J/\psi \\
\pi^\mp Z_c^\pm (3900) \\
K^+ K^- J/\psi \\
\gamma X(3872) \end{array} \right. $ \\ %
$Y(4360)$ & $\psi(4360)$
& $0^- 1^{--}$
& $4368 \pm 13$
& $96 \pm 7$
& $e^+e^- \to \gamma Y$ or $Y$; $Y \to \bigg\{ \begin{array}{l}
\pi^+ \pi^- \psi(2S) \\
\pi^0 \pi^0 \psi(2S) \end{array}$ \\ %
$Y(4390)$ & $\psi(4390)$
& $0^- 1^{--}$
& $4392 \pm 7$
& $140^{+16}_{-21}$
& $e^+e^- \to Y$; $Y \to \bigg\{ \begin{array}{l}
\eta J/\psi \\
\pi^+ \pi^- h_c \end{array}$ \\ %
$Y(4660)$ & $\psi(4660)$
& $0^- 1^{--}$
& $4643 \pm 9$
& $72 \pm 11$
& $e^+e^- \to \left\{ \begin{array}{l}
\gamma Y; Y \to \pi^+ \pi^- \psi(2S) \\
\ \; Y; Y \to \Lambda^+_c \Lambda^-_c
\end{array} \right.$ \\ \hline\hline
\end{tabular}
\end{table*}
\section{Theoretical Review of $P$-Wave Exotic States}
\label{sec:pwave_review}
The full spectroscopy of diquark-antidiquark ($\delta^{\vphantom{1}}$-${\bar\delta}$)
tet\-raquarks and diquark-antitriquark ($\delta^{\vphantom{1}}$-${\bar\theta}$) pentaquarks
connected by a gluonic field of arbitrary excitation quantum numbers,
and including arbitrary orbital excitations between the $\delta^{\vphantom{1}}$-${\bar\delta}$
or $\delta^{\vphantom{1}}$-${\bar\theta}$ pair, is presented in Ref.~\cite{Lebed:2017min}.
As discussed in that work, the gluonic-field excitations combined
with the quasiparticle sources $\delta^{\vphantom{1}},{\bar\delta},{\bar\theta}$ produce states
analogous to ordinary quarkonium hybrids; therefore, these states
may likewise be classified according to the quantum numbers provided
by BO-approximation static gluonic-field potentials. The numerical
studies of Ref.~\cite{Giron:2019bcs} show that the exotic analogues
to hybrid quarkonium states lie above the exotic states within the
corresponding BO ground-state potential $\Sigma^+_g$ by at least
1~GeV (just as for conventional quarkonium). Since the entire range
of observed hidden-charm exotic candidates [not counting $c\bar c
c\bar c$ candidates such as $X(6900)$] spans only about
800~MeV~\cite{Lebed:2016hpi}, it is very likely that all known
hidden-charm exotic states occupy energy levels within the
$\Sigma^+_g$ BO potential. All known ${c\bar{c}q\bar{q}^\prime}$ candidates can be
accommodated by the lowest $\Sigma^+_g$ levels: $1S$, $1P$, $2S$,
$1D$, and $2P$, in order of increasing mass~\cite{Giron:2019bcs}.
A detailed enumeration of the possible ${Q\bar Q} q\bar q^\prime$ states,
in which the light quarks $q$,$\bar q^\prime$ do not necessarily
carry the same flavor, is straightforward for the $S$ wave. Assuming
zero relative orbital angular momenta between the quarks, any two
naming conventions for the states differ only by the order in which
the 4 quark spins are coupled. In the diquark basis, defined by
coupling in the order $(qQ) \! + \! (\bar q \bar Q)$, the 6 possible
states are denoted by~\cite{Maiani:2014aja}:
\begin{eqnarray}
J^{PC} = 0^{++}: & \ & X_0 \equiv \left| 0_\delta^{\vphantom{1}} , 0_{\bar\delta} \right>_0 \,
, \ \ X_0^\prime \equiv \left| 1_\delta^{\vphantom{1}} , 1_{\bar\delta} \right>_0 \, ,
\nonumber \\
J^{PC} = 1^{++}: & \ & X_1 \equiv \frac{1}{\sqrt 2} \left( \left|
1_\delta^{\vphantom{1}} , 0_{\bar\delta} \right>_1 \! + \left| 0_\delta^{\vphantom{1}} , 1_{\bar\delta} \right>_1 \right)
\, ,
\nonumber \\
J^{PC} = 1^{+-}: & \ & \, Z \ \equiv \frac{1}{\sqrt 2} \left( \left|
1_\delta^{\vphantom{1}} , 0_{\bar\delta} \right>_1 \! - \left| 0_\delta^{\vphantom{1}} , 1_{\bar\delta} \right>_1 \right)
\, ,
\nonumber \\
& \ & \, Z^\prime \equiv \left| 1_\delta^{\vphantom{1}} , 1_{\bar\delta} \right>_1 \, ,
\nonumber \\
J^{PC} = 2^{++}: & \ & X_2 \equiv \left| 1_\delta^{\vphantom{1}} , 1_{\bar\delta} \right>_2 \,
,
\label{eq:Swavediquark}
\end{eqnarray}
where outer subscripts indicate total quark spin $S$. The same
states may be expressed in any other basis by using angular momentum
recoupling coefficients in the form of the relevant $9j$ symbol. For
the purposes of this work, the most useful alternate basis is that of
definite heavy-quark (and light-quark) spin, $({Q\bar Q}) \! + \! ({q\bar q})$:
\begin{eqnarray}
\lefteqn{\left< (s_q \, s_{\bar q}) s_{q\bar q} , (s_Q \, s_{\bar Q}) s_{Q\bar Q}
, S \, \right| \left. (s_q \, s_Q) s_\delta^{\vphantom{1}} , (s_{\bar q} \, s_{\bar Q})
s_{\bar\delta} , S \right> } & & \nonumber \\
& = & \left( [s_{q\bar q}] [s_{Q\bar Q}] [s_\delta^{\vphantom{1}}] [s_{\bar\delta}] \right)^{1/2}
\left\{ \begin{array}{ccc} s_q & s_{\bar q} & s_{q\bar q} \\
s_Q & s_{\bar Q} & s_{Q\bar Q} \\ s_\delta^{\vphantom{1}} & s_{\bar\delta} & S \end{array} \! \right\}
\, , \ \ \label{eq:9jTetra}
\end{eqnarray}
where $[s] \! \equiv \! 2s + 1$ denotes the multiplicity of a
spin-$s$ state. Using Eqs.~(\ref{eq:Swavediquark}) and
(\ref{eq:9jTetra}), one then obtains
\begin{eqnarray}
J^{PC} = 0^{++}: & \ & X_0 = \frac{1}{2} \left| 0_{q\bar q} , 0_{Q\bar Q}
\right>_0 + \frac{\sqrt{3}}{2} \left| 1_{q\bar q} , 1_{Q\bar Q} \right>_0 \, ,
\nonumber \\
& & X_0^\prime = \frac{\sqrt{3}}{2} \left| 0_{q\bar q} , 0_{Q\bar Q}
\right>_0 - \frac{1}{2} \left| 1_{q\bar q} , 1_{Q\bar Q} \right>_0 \, ,
\nonumber \\
J^{PC} = 1^{++}: & \ & X_1 = \left| 1_{q\bar q} , 1_{Q\bar Q} \right>_1 \, ,
\nonumber \\
J^{PC} = 1^{+-}: & \ & \, Z \; = \frac{1}{\sqrt 2} \left( \left|
1_{q\bar q} , 0_{Q\bar Q} \right>_1 \! - \left| 0_{q\bar q} , 1_{Q\bar Q} \right>_1 \right)
\, , \nonumber \\
& \ & \, Z^\prime = \frac{1}{\sqrt 2} \left( \left| 1_{q\bar q} ,
0_{Q\bar Q} \right>_1 \! + \left| 0_{q\bar q} , 1_{Q\bar Q} \right>_1 \right) \, ,
\nonumber \\
J^{PC} = 2^{++}: & \ & X_2 = \left| 1_{q\bar q} , 1_{Q\bar Q} \right>_2 \, .
\label{eq:SwaveQQ}
\end{eqnarray}
Once light-quark flavor is included, one obtains 12 states: 6 each
with $I \! = \! 0$ and $I \! = \! 1$, and spin structures in the form
of Eqs.~(\ref{eq:Swavediquark}) or (\ref{eq:SwaveQQ}).\footnote{If
strange quarks are included, one obtains 6 SU(3)$_{\rm flavor}$
octets and 6 singlets.} Using these states and the most minimal
$3$-parameter Hamiltonian [the $M_0$, $\kappa_{qQ}$, and $V_0$ terms
of Eq.~(\ref{eq:FullHam}) below],
Refs.~\cite{Giron:2019cfc,Giron:2020qpb} calculate the masses of all
$12$ $S$-wave states in the hidden-charm and hidden-bottom sectors
using known masses of $X(3872)$, $Z_c(3900)$, and $Z_c(4020)$ for
the former; and the known masses of $Z_b(10610)$, $Z_b(10650)$, and
their relative $h_b$ to $\Upsilon$ branching fractions for the
latter. These results incorporate isospin dependence (the $V_0$
term), a feature not explicitly integrated into other diquark models.
Reference~\cite{Giron:2020fvd} extends this analysis by examining the
$P$-wave multiplet, whose mass spectrum is dictated by the most
minimal $5$-parameter Hamiltonian:
\begin{eqnarray}
H & = & M_0 + 2 \kappa_{qQ} ({\bf s}_q \! \cdot \! {\bf s}_Q +
{\bf s}_{\bar q} \! \cdot \! {\bf s}_{\bar Q}) + V_{LS} \,
{\bf L} \cdot {\bf S} \nonumber \\ & & + V_0 \, {\bm \tau}_q
\! \cdot \! {\bm \tau}_{\bar q} \; {\bm \sigma}_q \! \cdot \!
{\bm \sigma}_{\bar q} + V_T \, {\bm \tau}_q \! \cdot
\! {\bm \tau}_{\bar q} \; S_{12}^{({q\bar q})} \, ,
\label{eq:FullHam}
\end{eqnarray}
where $M_0$ is the common mass of the multiplet, $\kappa_{qQ}$
represents the strength of the spin-spin coupling within each
diquark, $V_{LS}$ is the spin-orbit coupling strength, $V_0$ is the
isospin-dependent coupling,\footnote{$V_0$ in Eq. (\ref{eq:FullHam})
is analogous to the axial coupling in $NN\pi$ interactions.} $V_T$
represents the tensor coupling, and $S_{12}^{({q\bar q})}$ is the tensor
operator defined as
\begin{equation}
\label{eq:Tensor}
S_{12}^{({q\bar q})} \equiv 3 \, {\bm \sigma}_q \! \cdot {\bm r} \,
{\bm \sigma}_{\bar{q}} \! \cdot {\bm r} / r^2 - {\bm \sigma}_q \!
\cdot {\bm \sigma}_{\bar{q}} \, .
\end{equation}
The well-known tabulated expressions for matrix elements of
$S_{12}^{({q\bar q})}$~({\it e.g.}, in Ref.~\cite{Shalit:1963nuclear})
directly apply neither in the basis of $s_{q\bar q}, s_{Q\bar Q}$ spins nor
$s_\delta^{\vphantom{1}}, s_{\bar\delta}$ spins, but rather refer to the basis of total
light-quark angular momentum $J_{q\bar q}$:
\begin{equation} \label{eq:JqqDef}
{\bf J}_{q\bar q} \equiv {\bf L}_{q\bar q} + {\bf s}_{q\bar q} \, .
\end{equation}
Assuming that $\delta^{\vphantom{1}}$ and ${\bar\delta}$ have no internal orbital excitation
so that $L_{q\bar q} \! = \! L$, the matrix elements of $S_{12}^{({q\bar q})}$
are most easily computed in the $J_{q\bar q}$ basis, with results that are
then related back to the $s_{q\bar q}, s_{Q\bar Q}$ basis by means of recoupling
using $6j$ symbols:
\begin{eqnarray}
\lefteqn{\mathcal{M}_{J_{q\bar q}} \equiv
\left< (L,s_{q\bar q}),J_{q\bar q},s_{Q\bar Q}, J | L,(s_{{q\bar q}},s_{{Q\bar Q}}), S, J \right>
} & & \nonumber\\
= & & (-1)^{L+s_{q\bar q}+s_{Q\bar Q}+J}\sqrt{[J_{q\bar q}][S]}
\left\{ \begin{array}{ccc}
L & s_{q\bar q} & J_{q\bar q}\\
s_{Q\bar Q} & J & S
\end{array} \right\} . \label{eq:JqqCoef}
\end{eqnarray}
Using this expression, $S_{12}^{({q\bar q})}$ matrix elements for all
relevant states are tabulated in Ref.~\cite{Giron:2020fvd}.
The experimental status of the $P$-wave $J^{PC} \! = \! 1^{--}$
exotic candidates remains in flux, with BESIII providing the majority
of the most recent data. With reference to the information presented
in Table~\ref{tab:Expt}, we have already noted that the analysis of
the BESIII Collaboration~\cite{Ablikim:2016qzw} favors the
interpretation of $Y(4260)$ as a superposition of states, the lowest
component of which is $Y(4220)$. They identify the higher component
with $Y(4360)$, although the previous mass measurements of this state
given in Table~\ref{tab:Expt} are rather higher, and one of several
scenarios considered in Ref.~\cite{Giron:2020fvd} proposes that
$Y(4360)$ and $Y(4390)$ are the same state, while the higher-mass
component in Ref.~\cite{Ablikim:2016qzw} can be interpreted as a
distinct ``$Y(4320)$''. Alternately, if the only lower states are
$Y(4220)$, $Y(4360)$, and $Y(4390)$, then $Y(4660)$ becomes the
fourth $I \! = \! 0$, $1^{--}$ candidate state in $\Sigma^+_g(1P)$.
With the mass spectrum of these charmoniumlike states not yet
entirely settled, Ref.~\cite{Giron:2020fvd} also employs information
on their preferred charmonium decay modes as classified by
heavy-quark spin: $\psi$ ($s_{{Q\bar Q}} \! = \! 1$) or $h_c$ ($s_{{Q\bar Q}} \!
= \! 0$). Assuming heavy-quark spin symmetry as expressed by the
conservation of $s_{{Q\bar Q}}$ in the decays, the heavy-quark spin content
$P_{s_{{Q\bar Q}}}$ of each state becomes an invaluable diagnostic in
disentangling the $J^{PC} \! = \! 1^{--}$ spectrum. For example,
from Table~\ref{tab:Expt} one sees that $Y(4220)$ decays to both
$\psi$ states and $h_c$, while if $Y(4360)$ and $Y(4390)$ are in fact
one state, the same can be said for them as well.
Reference~\cite{Giron:2020fvd} also introduces a parameter $\epsilon$
designed to enforce the goodness-of-fit to a particular value $f$ of
$P_{s_{{Q\bar Q}}}$, which in the case of $s_{{Q\bar Q}} \! = \! 0$ reads
\begin{equation}
\Delta \chi^2 = \left( \frac{\ln P_{s_{{Q\bar Q}} = 0} - \ln f}{\epsilon}
\right)^2 .
\end{equation}
In terms of the parameters $P_{s_{{Q\bar Q}}}$, $f$, and $\epsilon$, the 5
cases discussed in Ref.~\cite{Giron:2020fvd} designed to represent a
variety of interpretations of the current data are:
\begin{enumerate}
\item $Y(4220)$, $Y(4260)$, $Y(4360)$, $Y(4390)$ masses are as given in
the PDG (Table~\ref{tab:Expt}). No constraint is placed upon
$P_{s_{{Q\bar Q}} = 0}^{Y(4220)}$ or $P_{s_{{Q\bar Q}} =
0}^{Y(4390)}$. \label{JunkFit19}
\item $Y(4220)$, $Y(4260)$, $Y(4360)$, $Y(4390)$ masses are as given in
the \@. $P_{s_{{Q\bar Q}} = 0}^{Y(4220)}$ is fit to $f \! = \! \frac 1
3$ with $\epsilon \! = \! 0.1$, and $P_{s_{{Q\bar Q}} = 0}^{Y(4390)}$ is
unconstrained. \label{JunkFit7B}
\item $Y(4220)$, $Y(4360)$, and $Y(4390)$ masses are as given in
the PDG, while $m_{Y(4260)} \! = \! 4251 \! \pm \! 6$~MeV, which is
the weighted average of the 3 PDG values not including the low BESIII
value~\cite{Ablikim:2016qzw}. $P_{s_{{Q\bar Q}} = 0}^{Y(4220)}$ is fit to
$f \! = \! \frac 1 3$ with $\epsilon \! = \! 0.2$, and $P_{s_{{Q\bar Q}} =
0}^{Y(4390)}$ is fit to $f \! = \! \frac 2 3$ with $\epsilon \! = \!
0.05$. \label{JunkFit9B}
\item $Y(4360)$, $Y(4390)$, and $Y(4660)$ masses are as
given in the PDG, but $Y(4260)$ is assumed not to exist, and
$m_{Y(4220)} \! = \! 4220.1 \! \pm 2.9$~MeV is the weighted average
of the PDG values combined with the newer BESIII
measurements~\cite{Ablikim:2018vxx,Ablikim:2019apl}. $P_{s_{{Q\bar Q}} =
0}$ values are as given in Case~\ref{JunkFit9B}. \label{JunkFit21B}
\item $m_{Y(4220)}$ is as given in Case~\ref{JunkFit21B};
$m_{Y(4260)}$ is as given in Case~\ref{JunkFit9B};
$m_{\text{``$Y(4320)$''}} \! = \! 4320 \! \pm \! 13$~MeV is the lower
BESIII $Y(4360)$ mass measurement from~\cite{Ablikim:2016qzw};
$m_{Y(4390)} \! = \! 4386 \! \pm \! 4$~MeV is the weighted average of
the PDG value and the upper BESIII $Y(4360)$ mass measurement
from~\cite{Ablikim:2017oaf}. $P_{s_{{Q\bar Q}} = 0}$ values are as given
in Case~\ref{JunkFit9B}.
\label{JunkFit18B}
\end{enumerate}
We previously suggested the importance of heavy-quark spin-symmetry
($s_{Q\bar Q}$) conservation in the decays of exotics, particularly for
$Z_c(3900)$ and $Z_c(4020)$, but also for several other exotic
candidates that to date have only been observed to decay to
charmonium states carrying one specific value of $s_{Q\bar Q}$ ({\it e.g.},
to $\psi$ {\em or\/} to $h_c$). We assume that a state like
$Y(4220)$ is able to decay to channels with either value of $s_{Q\bar Q}$
due to the initial state being a mixture of $s_{Q\bar Q}$ eigenstates,
rather than to the value of $s_{Q\bar Q}$ changing in the decay process
through a heavy-quark spin-symmetry violation. In addition, in this
analysis we take the well-known radiative transition selection rules
to apply to the light degrees of freedom, which carry the total
angular momentum $J_{q\bar q}$ defined in Eq.~(\ref{eq:JqqDef}). As usual,
the operators defining E1 and M1 transitions transform as $J^{P} \! =
\! 1^-$ and $J^{P} \! = \! 1^+$, respectively.
Explicit expressions for radiative transitions between quarkonium
states (themselves transcribed from textbook atomic-physics formulae)
appear in the literature ({\it e.g.}, Ref.~\cite{Barnes:2005pb}), and
may readily be adapted to the present case. In particular, the
quarkonium orbital angular momentum $L$ is replaced with $J_{q\bar q}$, and
the heavy quark mass $m_Q$ is replaced with the diquark mass $m_\delta^{\vphantom{1}}$.
For E1 partial widths, one has
\begin{widetext}
\begin{equation}\label{eqn:e1width}
\Gamma_{\rm E1}\left(n^{\, 2s_{Q\bar Q}+1} \!
\left(J^{\vphantom\prime}_{{q\bar q}}\right)^{\vphantom\dagger}_J
\rightarrow n'^{\; 2s_{Q\bar Q}'+1} \!
\left(J'_{{q\bar q}} \right)^{\vphantom\dagger}_{J'} +
\gamma\right)=\frac{4}{3}C_{fi} \,
\delta_{s^{\vphantom\prime}_{Q\bar Q} s_{Q\bar Q}'} \alpha \, Q_{\delta^{\vphantom{1}}}^2
\left\vert\left\langle \psi_f \middle| r \middle| \psi_i
\right\rangle\right\vert^2E_\gamma^3
\frac{E_f^{({Q\bar Q} q\bar{q}')}}{M_i^{({Q\bar Q} q\bar{q}')}},
\end{equation}
\end{widetext}
where
\begin{equation}
\label{eq:Cfactors}
C_{fi} \equiv \mathrm{max}\left(J_{q\bar q}^{\vphantom\prime},
J^\prime_{q\bar q}\right) \left(2J'+1\right)
\begin{Bmatrix}
J^\prime_{q\bar q} & J' & s_{Q\bar Q}\\
J & J_{q\bar q} & 1
\end{Bmatrix}^2.
\end{equation}
The labels $i$ and $f$ refer to initial and final states,
respectively. The initial exotic state ${Q\bar Q} q\bar q^\prime$, of mass
$M_i^{({Q\bar Q} q\bar q^\prime)}$, decays in its rest frame into a final
exotic state with the same flavor content and energy $E_f^{({Q\bar Q} q\bar
q^\prime)}$, and a photon of energy $E_\gamma$. $\alpha$ is the
fine-structure constant. $\psi$ denotes radial wave functions of
the exotic hadrons, and $r$ is the spatial separation between the
$\delta^{\vphantom{1}}$-${\bar\delta}$ pair centers. $Q_\delta^{\vphantom{1}}$ is the total electric charge (in
units of proton charge) to which the photon couples; in
Ref.~\cite{Chen:2015dig}, the diquarks are treated as pointlike, in
which case one simply takes $Q_\delta^{\vphantom{1}} \! = \! Q_Q \! + \! Q_q$.
Alternately, one may argue that the diquarks $\delta^{\vphantom{1}}$ are of sufficient
spatial extent that the photon couplings to the distinct quarks in $\delta^{\vphantom{1}}$
should add through incoherent diagrams, in which case one takes
$Q_{\delta^{\vphantom{1}}}^2 \! = \! Q_Q^2 \! + \! Q_q^2$. In our calculation we use
the first option, but note in addition that a $Y$ state, being an
isosinglet, contains an equal superposition of $u$ and $d$ quarks.
We thus take
\begin{equation}
\label{eq:charge}
Q_{\delta^{\vphantom{1}}}^2 \to \frac{1}{2} \left[ \left( Q_c + Q_u \right)^2 +
\left( Q_c + Q_d \right)^2 \right] = \frac{17}{18} \, .
\end{equation}
Other schemes give rise to coefficients that differ from this value
only at $O(1)$. Corrections that arise from treating the distinct
quarks within each diquark as separated entities, for example through
electromagnetic form factors of the $\delta^{\vphantom{1}},{\bar\delta}$ composite
quasiparticles, would be incorporated in this model through the
factor $Q_{\delta^{\vphantom{1}}}^2$.
The corresponding expression for M1 partial widths, involving no
change in parity but a flip of the heavy-quark spin $s_{{Q\bar Q}}$ (hence
breaking heavy-quark spin symmetry), reads
\begin{widetext}
\begin{equation}\label{eqn:m1width}
\Gamma_{\rm M1}\left(n^{\, 2s_{Q\bar Q}+1}
\left(J^{\vphantom\prime}_{{q\bar q}}\right)^{\vphantom\dagger}_J
\rightarrow n'^{\; 2s_{Q\bar Q}'+1}
\left(J'_{{q\bar q}} \right)^{\vphantom\prime}_{J'} +
\gamma\right)=\frac{4}{3} \frac{2J' +1}{2J_{q\bar q} +1}
\delta_{J^{\vphantom\prime}_{q\bar q} J'_{q\bar q}}
\delta_{s^{\vphantom\prime}_{Q\bar Q}, \, s_{Q\bar Q}'\pm 1} \, Q_{\delta^{\vphantom{1}}}^2
\frac{\alpha}{m_{\delta^{\vphantom{1}}}^2}\left\vert\left\langle \psi_f \middle| \psi_i
\right\rangle\right\vert^2E_\gamma^3
\frac{E_f^{({Q\bar Q} q\bar{q}')}}{M_i^{({Q\bar Q} q\bar{q}')}} .
\end{equation}
\end{widetext}
This expression is presented here for completeness, in light of the
current lack of experimental evidence for such transitions. However,
in Sec.~\ref{sec:analysis} we use it to calculate the expected
radiative width for the yet-unobserved~\cite{Ablikim:2021rba}
transition $Z_c (4020)^0 \! \to \! \gamma X(3872)$.
As noted above, Eqs.~(\ref{eqn:e1width}) and (\ref{eqn:m1width}) are
almost identical to textbook nonrelativistic results. The only
exception in each case is the inclusion of a factor $E_f/M_i$ to
represent relativistic phase space associated with recoil of the
final-state hadron. In fact, Ref.~\cite{Brambilla:2004wf}
discusses several distinct relativistic corrections that could be
included in a more complete study. Since this work represents the
first attempt to calculate the radiative widths for a spectrum of
states whose experimental interpretation remains ambiguous, we
include only a minimal set of physical effects in the analysis.
Lastly, corrections to the long-wavelength approximation discussed in
Sec.~\ref{sec:pwave_review} that are derived by retaining the full
photon plane-wave factor $\exp ( i \bm{k} \cdot \bm{r} )$ have also
been computed ({\it e.g.}, Ref.~\cite{Brambilla:2004wf}).
Explicitly, Eqs.~(\ref{eqn:e1width}) and (\ref{eqn:m1width}) are
modified through the substitutions
\begin{equation}
\langle \psi_f | r | \psi_i \rangle \to \left< \psi_f \! \left|
\frac{3}{k} \left[ \frac{kr}{2} j_0 \left( \frac{kr}{2} \right) - j_1
\left( \frac{kr}{2} \right) \right] \right| \psi_i \right> \, ,
\end{equation}
and
\begin{equation}
\langle \psi_f | \psi_i \rangle \to \left< \psi_f \! \left| \, j_0
\left( \frac{kr}{2} \right) \right| \psi_i \right> \, ,
\end{equation}
respectively,where $j_0$ and $j_1$ are spherical Bessel
functions. The corresponding series expansions of these functions
read
\begin{equation}
\label{eqn:E1corr}
r - \frac{1}{20} k^2 r^3 + O(k^4 r^5) \, ,
\end{equation}
and
\begin{equation}
\label{eqn:M1corr}
1 - \frac{1}{24} k^2 r^2 + O(k^4 r^4) \, .
\end{equation}
Note especially the small subleading-term numerical coefficient in
each case, suggesting that the long-wavelength approximation holds
relatively well even for substantial values of $kr$. We examine
specific examples in Sec.~\ref{sec:analysis}.
\section{Analysis and Results}\label{sec:analysis}
Possible assignments of observed $Y$ states to members of the
$\Sigma^+_g(1P)$ multiplet in this model are described by the 5 cases
discussed extensively in Ref.~\cite{Giron:2020fvd} and summarized
in Sec.~\ref{sec:pwave_review}. Of these cases, all have excellent
goodness-of-fit values $\chi^2_{\rm min}/{\rm d.o.f.}$ except Case~3;
however, we argue this case and Case~5 to be the most
phenomenologically relevant ones, since they enforce the important
physical constraint that both $Y(4220)$ and $Y(4390)$ are observed
(see Table~\ref{tab:Expt}) to have substantial couplings to $h_c$
($s_{{Q\bar Q}} \! = \! 0$). Since $\Sigma^+_g(1S)$ contains only one
$I \! = \! 0$, $J^{PC} \! = \! 1^{--}$ state with $s_{{Q\bar Q}} \! = \!
0$, the requirement of providing a substantial component of this
state to both of the well-separated $Y(4220)$ and $Y(4390)$ mass
eigenstates is one of the primary obstacles to achieving a good fit.
Case~5 relieves the tension of Case~3 by identifying, as discussed in
Sec.~\ref{sec:Expt}, a new state ``$Y(4320)$'' from the data of
Ref.~\cite{Ablikim:2016qzw}. In addition, Cases~1, 2, 3, and 5 all
predict the sole $I \! = \! 1$, $J^{PC} \! = \! 0^{--}$ state in
$\Sigma^+_g(1P)$ to lie in the range 4220--4235~MeV, which agrees
well with the unconfirmed state $Z_c(4240)$ carrying these quantum
numbers that is observed in the LHCb paper~\cite{Aaij:2014jqa}
confirming the existence of $Z_c(4430)$.
Case~4 also satisfies the $Y(4220)/Y(4390)$ $s_{{Q\bar Q}} \! = \! 0$
criterion, but additionally assigns the rather high-mass $Y(4660)$ to
the $\Sigma^+_g(1P)$ multiplet; the cost is a much higher prediction
($\approx \! 4440$~MeV) for the mass of the $\Sigma^+_g(1P)$ $I \! =
\! 1$, $J^{PC} \! = \! 0^{--}$ state, in conflict with the value of
$m_{Z_c(4240)}$.
We therefore single out the fits of Cases~3 and 5 for the
decomposition of $Y$ states with respect to the total light-quark
angular momentum $J_{{q\bar q}}$ in Tables~\ref{tab:case3} and
\ref{tab:case5}, respectively. For completeness, we also provide the
corresponding information for Cases~1, 2, and 4 in
Table~\ref{tab:cases124}.
\begin{table}
\caption{Decomposition of $Y$ ($I \! = \! 0$, $J^{PC} \! = \!
1^{--}$) charmoniumlike exotic candidates into a basis of good
light-quark spin $s_{{q\bar q}}$, heavy-quark spin $s_{{Q\bar Q}}$, and total
light-quark angular momentum $J_{{q\bar q}}$, performed for the 4
experimentally observed candidate states as described in Case~3 above
and in Ref.~\cite{Giron:2020fvd}. A minus sign on the probability
$(-|P|)$ means that the corresponding amplitude is $-|P|^{1/2}$, the
same convention as is used for Clebsch-Gordan coefficients by the
PDG~\cite{Zyla:2020zbs}.}
\label{tab:case3}
\setlength{\extrarowheight}{0.5ex}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\fill}}c c c c c}
\hline\hline
Particle &$s_{{q\bar q}}$ & $s_{{c\bar c}}$ &\;\;$J_{{q\bar q}}$\;\;& Probability\\
\hline
$Y(4220)$ & $0$ & $0$ & $0$ & $+0.231$\\
$ $& $1$ & $1$ & $0$ & $+0.012$\\
$ $& & $ $ & $1$ & $-0.577$\\
$ $& & $ $ & $2$ & $+0.181$\\
$Y(4260)$ & $0$ & $0$ & $0$ & $+0.061$\\
$ $& $1$ & $1$ & $0$ & $+0.004$\\
$ $& & $ $ & $1$ & $+0.352$\\
$ $& & $ $ & $2$ & $+0.583$\\
$Y(4360)$ & $0$ & $0$ & $0$ & $+0.069$\\
$ $& $1$ & $1$ & $0$ & $+0.835$\\
$ $& & $ $ & $1$ & $+0.020$\\
$ $& & $ $ & $2$ & $-0.075$\\
$Y(4390)$ & $0$ & $0$ & $0$ & $+0.638$\\
$ $& $1$ & $1$ & $0$ & $-0.149$\\
$ $& & $ $ & $1$ & $+0.051$\\
$ $& & $ $ & $2$ & $-0.161$\\
\hline\hline
\end{tabular*}
\end{table}
\begin{table}
\caption{Decomposition of $Y$ ($I \! = \! 0$, $J^{PC} \! = \!
1^{--}$) charmoniumlike exotic candidates as in Table~\ref{tab:case3},
except now performed for the 4 experimentally observed candidate
states as described in Case~5 above and in
Ref.~\cite{Giron:2020fvd}.}
\label{tab:case5}
\setlength{\extrarowheight}{0.5ex}
\begin{tabular*}{\columnwidth}{@{\extracolsep{\fill}}c c c c c}
\hline\hline
Particle & $s_{{q\bar q}}$ & $s_{{c\bar c}}$ &\;\;$J_{{q\bar q}}$\;\; & Probability\\
\hline
$Y(4220)$ & $0$ & $0$ & $0$ & $-0.264$\\
$ $& $1$ & $1$ & $0$ & $-0.007$\\
$ $& & $ $ & $1$ & $+0.543$\\
$ $& & $ $ & $2$ & $-0.186$\\
$Y(4260)$ & $0$ & $0$ & $0$ & $+0.060$\\
$ $& $1$ & $1$ & $0$ & $+0.036$\\
$ $& & $ $ & $1$ & $+0.380$\\
$ $& & $ $ & $2$ & $+0.523$\\
``$Y(4320)$'' & $0$ & $0$ & $0$ & $+0.025$\\
$ $& $1$ & $1$ & $0$ & $+0.870$\\
$ $& & $ $ & $1$ & $+8 \times 10^{-4}$\\
$ $& & $ $ & $2$ & $-0.105$\\
$Y(4390)$ & $0$ & $0$ & $0$ & $-0.651$\\
$ $& $1$ & $1$ & $0$ & $+0.086$\\
$ $& & $ $ & $1$ & $-0.076$\\
$ $& & $ $ & $2$ & $+0.187$\\
\hline
\end{tabular*}
\end{table}
\begin{table*}
\caption{Decomposition of $Y$ ($I \! = \! 0$, $J^{PC} \! = \!
1^{--}$) charmoniumlike exotic candidates as in
Tables~\ref{tab:case3}--\ref{tab:case5}, except now performed for the
4 experimentally observed candidate states as described in Cases~1,
2, and 4 above and in Ref.~\cite{Giron:2020fvd}.}
\label{tab:cases124}
\setlength{\extrarowheight}{0.5ex}
\begin{tabular*}{\textwidth}{@{\extracolsep{\fill}}c c c c c | c c c c c | c c c c c}
\hline\hline
\multicolumn{5}{c|}{Case 1}& \multicolumn{5}{c|}{Case 2}& \multicolumn{5}{c}{Case 4}\\
\hline
Particle & $s_{{q\bar q}}$ & $s_{{c\bar c}}$ &\;\;$J_{{q\bar q}}$\;\; & Probability& Particle & $s_{{q\bar q}}$ & $s_{{c\bar c}}$ &\;\;$J_{{q\bar q}}$\;\; & Probability &Particle & $s_{{q\bar q}}$ & $s_{{c\bar c}}$ &\;\;$J_{{q\bar q}}$\;\; & Probability\\
\hline
$Y(4220)$ & $0$ & $0$ & $0$ & $+0.771$ & $Y(4220)$ & $0$ & $0$ & $0$ & $-0.336$ & $Y(4220)$ & $0$ & $0$ & $0$& $-0.233$\\
$ $& $1$ & $1$ & $0$ & $-0.019$ & $ $ & $1$ & $1$ & $0$ & $+0.032$& $ $& $1$ & $1$ & $0$& $-0.048$\\
$ $& & $ $ & $1$ & $-0.211$ & $ $ & & $ $ & $1$ & $+0.631$ & $ $& & $ $ & $1$& $+0.376$\\
$ $& & $ $ & $2$ & $-4\times 10^{-7}$ & $ $& & $ $ & $2$ & $+8\times 10^{-4}$ & $ $& & $ $ & $2$& $-0.343$\\
$Y(4260)$ & $0$ & $0$ & $0$ & $+0.212$ & $Y(4260)$ & $0$ & $0$ & $0$ & $+0.588$ & $Y(4360)$ & $0$ & $0$ & $0$& $-0.119$\\
$ $& $1$ & $1$ & $0$ & $+0.130$ & $ $ & $1$ & $1$ & $0$ & $+0.056$ & $ $& $1$ & $1$ & $0$& $+0.101$\\
$ $& & $ $ & $1$& $+0.597$ &$ $ & & $ $ & $1$ & $+0.246$ & $ $& & $ $ & $1$& $-0.473$\\
$ $ & & $ $ & $2$ &$+0.062$ & $ $& & $ $ & $2$ & $+0.109$ & $ $ & & $ $ & $2$ & $-0.308$ \\
$Y(4360)$ & $0$ & $0$ & $0$ & $+0.006$ & $Y(4360)$ & $0$ & $0$ & $0$ & $-0.046$ & $Y(4390)$ & $0$ & $0$ & $0$& $+0.647$\\
$ $& $1$ & $1$ & $0$ & $+0.252$ & $ $& $1$ & $1$ & $0$ & $-0.117$ & $ $& $1$ & $1$ & $0$ & $+0.003$\\
$ $& & $ $ & $1$ & $-5\times 10^{-6}$ & $ $& & $ $ & $1$ & $-0.012$ & $ $& & $ $ & $1$& $+0.009$\\
$ $& & $ $ & $2$ & $-0.742$ & $ $& & $ $ & $2$ & $+0.824$ & $ $& & $ $ & $2$& $-0.342$\\
$Y(4390)$ & $0$ & $0$ & $0$ & $-0.012$ & $Y(4390)$ & $0$ & $0$ & $0$ & $-0.029$ &$Y(4660)$ & $0$ & $0$ & $0$ & $-0.002$\\
$ $& $1$ & $1$ & $0$ & $+0.599$ & $ $ & $1$ & $1$ & $0$ & $+0.795$ & $ $& $1$ & $1$ & $0$ & $+0.848$\\
$ $& $ $ & $ $ & $1$ & $-0.193$ & $ $ & & $ $ & $1$ & $-0.111$ & $ $& & $ $ & $1$ & $+0.143$\\
$ $& $ $ & $ $ & $2$ & $+0.196$ & $ $ & & $ $ & $2$ & $+0.066$ & $ $& & $ $ & $2$ & $+0.007$\\
\hline\hline
\end{tabular*}
\end{table*}
Using the mass eigenvalues for the $Y$ states in
Table~\ref{tab:Expt}, the state decompositions according to $J_{{q\bar q}}$
in Tables~\ref{tab:case3}, \ref{tab:case5}, and \ref{tab:cases124},
the coefficient factors in Eq.~(\ref{eq:Cfactors}), and the effective
squared-charge $Q_\delta^2$ from Eq.~(\ref{eq:charge}), one may
calculate the E1 radiative partial decay widths for $\Sigma^+_g(1P)
\! \to \! \gamma \Sigma^+_g(1S)$ transitions from
Eq.~(\ref{eqn:e1width}). The only nontrivial new input to the
calculation is that of the transition matrix element $\langle \psi_f
| r | \psi_i \rangle$. Using the numerical methods for solving
Schr\"{o}dinger equations developed in Ref.~\cite{Giron:2019bcs}, and
particularly the fits performed in Ref.~\cite{Giron:2020qpb} to
obtain the fine structure of the $\Sigma^+_g(1S)$ multiplet, the
optimal diquark mass is found to be
\begin{equation}
\label{eq:diquarkmass}
m_{\delta^{\vphantom{1}}} = m_{{\bar\delta}} = 1.933 \pm 0.005 \, {\rm GeV} \, ,
\end{equation}
as one varies over the static gluonic-field potentials $\Sigma^+_g$
obtained in the lattice calculations of Refs.~\cite{Juge:1997nc,
Juge:1999ie,Juge:2002br,Morningstar:2019,Capitani:2018rox}. We then
compute the relevant matrix element to be
\begin{equation}
\label{eq:TransMatEl}
\left< \psi_f (1S) \left| r \right| \psi_i (1P) \right> = 0.402 \pm
0.001 \, {\rm fm} \, .
\end{equation}
Note in particular that this numerical input appears in all
$\Sigma^+_g(1P) \! \to \! \gamma \Sigma^+_g(1S)$ transitions, not
simply those of $Y \! \to \! \gamma X(3872)$ that are compiled
according to the 5 cases in Table~\ref{tab:partialwidth}. Moreover,
Table~\ref{tab:partialwidth} and additional calculated width values
presented subsequently in this work exhibit only central values for
$\Gamma$; the small uncertainties in Eqs.~(\ref{eq:diquarkmass}) and
(\ref{eq:TransMatEl}) only refer to variation over different lattice
simulations, and do not take into account other much more significant
potential sources of uncertainty, such as effects due to finite
diquark size. Nevertheless, such effects were
shown~\cite{Giron:2020qpb} to change expectation values like $\langle
r \rangle$ no more than 10\%, a value that we adopt as a benchmark
uncertainty for all $\Gamma$ values computed here.
Also noteworthy is the magnitude of $k \left< \psi_f (1S) \! \left| r
\right| \! \psi_i (1P) \right>$ for each case, which provides an
indication of the reliability of the long-wavelength approximation.
Indeed, for $Y(4220)$, $k \! = \! 334$~MeV, and using
Eq.~(\ref{eq:TransMatEl}) gives $kr \! \to \! 0.680$, while the
corresponding value for $Y(4660)$ ($k \! = \! 699$~MeV) is 1.424.
However, the same simulations as in Eq.~(\ref{eq:TransMatEl}) also
produce
\begin{equation}
\left< \psi_f (1S) \left| r^3 \right| \psi_i (1P) \right> = 0.135 \pm
0.001 \, {\rm fm}^3 \, ,
\end{equation}
from which one computes the relative magnitude of the first
correction term in Eq.~(\ref{eqn:E1corr}) to be only 0.033 for
$Y(4220)$ and, surprisingly, only 0.300 for $Y(4660)$.
\begin{table*}
\caption{Radiative E1 partial widths (in keV) to $\gamma X(3872)$
calculated using Eq.~(\ref{eqn:e1width}), for the 5 cases of possible
$Y$ state assignments defined above and in Ref.~\cite{Giron:2020fvd}.
For each case, note that two of the $Y$ states (indicated by dashes)
are assumed either not to exist or not to belong to the
$\Sigma^+_g(1P)$ multiplet.}
\centering
\setlength{\extrarowheight}{1.2ex}
\begin{tabular}{c | r | r | r | r | r | r }
\hline\hline
Case &$Y(4220)$ & $Y(4260)$ & ``$Y(4320)$'' & $Y(4360)$ & $Y(4390)$ & $Y(4660)$\\
\hline
$1$ & $30.4$ & $145.6$ & {--- \ } & $721.3$ & $981.8$ & {--- \ } \\
$2$ & $81.1$ & $80.6$ & {--- \ } & $616.0$ & $1127.0$ & {--- \ } \\
$3$ & $105.1$ & $211.2$ & {--- \ } & $1016.1$ & $319.2$ & {--- \ } \\
$4$ & $136.0$ & {--- \ } & {--- \ } & $432.1$ & $231.6$ & $3363.9$ \\
$5$ & $102.4$ & $216.2$ & $807.6$ & {--- \ } & $253.8$ & {--- \ } \\
\hline
\end{tabular}
\label{tab:partialwidth}
\end{table*}
One observes from Table~\ref{tab:partialwidth} that the widths
$\Gamma_{Y(4220) \to \gamma X(3872)}$ and
$\Gamma_{Y(4260) \to \gamma X(3872)}$ assume almost the same
values in Cases~3 and 5 (102--105~keV and 211--216~keV,
respectively). $\Gamma_{Y(4390) \to \gamma X(3872)}$ also exhibits
fairly modest variation, from 254--319~keV\@. Indeed, some of the
large radiative width values in Table~\ref{tab:partialwidth}, such as
3.4~MeV for $Y(4660) \! \to \! \gamma X(3872)$ in Case 4, can serve
as vital criteria for eliminating possible assignments of $Y$ states
to the $1P$ multiplet: Glancing at the measured total
$\Gamma_{Y(4660)}$ in Table~\ref{tab:Expt}, one sees that were
$Y(4660)$ truly a $1P$ state, then its large phase space for
radiative decay to $X(3872)$ [evident from the $E_\gamma^3$ factor of
Eq.~(\ref{eqn:e1width})] would generate a radiative branching
fraction of at least several percent.
The transition matrix element of Eq.~(\ref{eq:TransMatEl}) has
already been noted to apply to all $\Sigma^+_g(1P) \! \to \! \gamma
\Sigma^+_g(1S)$ transitions. The only observed hidden-charm
tetraquark candidates with $P \! = \! -$ apart from the $Y$ states
are $Z_c(4240)$ and $Y(4626)$; the latter has thus far been observed
to decay only to various $D_s$ meson pairs~\cite{Jia:2019gfe,
Jia:2020epr}, and therefore is very likely a $c\bar c s\bar s$
state~\cite{Giron:2020qpb}. As for $Z_c(4240)$, only its charged
isobar has yet been observed, but assuming the existence of a
degenerate $Z_c(4240)^0$, one may input its quantum numbers
$s_{{Q\bar Q}} \! = \! 1$, $J_{{q\bar q}} \! = \! 1$, $J \! = \!
0$~\cite{Giron:2020fvd} into Eq.~(\ref{eqn:e1width}) to obtain
\begin{equation}
\Gamma \left[ Z_c (4240)^0 \to \gamma X(3872) \right] = 503 \
{\rm keV} \, .
\end{equation}
Lastly, we noted with Eq.~(\ref{eqn:m1width}) that M1 transitions
occur only with a flip of the heavy-quark spin. Such is the case for
the $\Sigma^+_g(1S) \! \to \! \gamma \Sigma^+_g(1S)$ transition
$Z_c(4020)^0 \! \to \! \gamma X(3872)$ ($s_{{Q\bar Q}} \! = \! 0 \! \to
s_{{Q\bar Q}} \! = \! 1$). Using Eq.~(\ref{eq:diquarkmass}), we calculate
\begin{equation}
\Gamma \left[ Z_c (4020)^0 \to \gamma X(3872) \right] = 7.91 \
{\rm keV} \, ,
\end{equation}
noting from Eq.~(\ref{eqn:m1width}) that the underlying matrix
element $\langle \psi_f | \psi_i \rangle \! = \! 1$ since both states
share the same radial wave function. In comparison, the molecular
model, in which $X(3872)$ and $Z_c(4020)$ are $D^0 \bar D^{*0} \! +
\! {\bar D}^0 D^{*0}$ and $D^* \bar D^*$ bound states, respectively,
and the decay $Z_c(4020)^0 \! \to \! \gamma X(3872)$ proceeds via
$D^{*0} \! \to \! \gamma D^0$, produces a rather larger radiative
width: The calculation of Ref.~\cite{Voloshin:2019ivc} predicts a
branching fraction of about $5 \! \times \! 10^{-3}$, which for
$\Gamma_{Z_c(4020)} \! = \! 13 \! \pm \! 5$~MeV~\cite{Zyla:2020zbs}
amounts to at least 40~keV.
In light of our investigations for $\Sigma^+_g (1P) \! \to \!
\Sigma^+_g(1S)$ E1 transitions, the long-wavelength approximation for
M1 transitions within the single multiplet $\Sigma^+_g(1S)$ is
undoubtedly satisfactory [for example, in $Z_c(4020)^0 \! \to \!
\gamma X(3872)$, $k$ is only 150~MeV]. Indeed, one may press the
approximation of Eq.~(\ref{eqn:M1corr}) to consider a transition that
is forbidden in the long-wavelength limit of Eq.~(\ref{eqn:m1width})
due to the orthogonality of wave functions, such as $\Sigma^+_g(2S)
\! \to \! \gamma \Sigma^+_g(1S)$. Assuming that $Z_c^0(4430)$ is the
$2S$ partner to $Z_c^0(4020)$, then $Z_c^0(4430) \! \to \! \gamma
X(3872)$ has $k \! \approx \! 565$~MeV, while we compute
\begin{equation}
\left< \psi_f (1S) \left| r^2 \right| \psi_i (2S) \right> = 0.152 \pm
0.001 \, {\rm fm}^2 \, ,
\end{equation}
and the first nontrivial term of Eq.~(\ref{eqn:M1corr}) evaluates to
$-0.052$. Using this correction in Eq.~(\ref{eqn:m1width}) leads to
a radiative width $\Gamma_\gamma \! \simeq \! 1$~keV, to be compared
with $\Gamma_{Z_c^+(4430)} \! \approx \!
180$~MeV~\cite{Zyla:2020zbs}. The observation a radiative transition
with such a small branching fraction is not impossible, but likely
will not occur in the near future.
\section{Conclusions}\label{sec:conclusions}
In this paper we have calculated exotic-to-exotic had\-ronic
radiative transitions using the dynamical diquark model. The most
phenomenologically relevant final state is $X(3872)$, which is a
member of the model's hidden-charm ground-state multiplet
$\Sigma^+_g(1S)$. We use the results from a recent
study~\cite{Giron:2020fvd} of this model for the lowest $P$-wave
multiplet [$\Sigma^+_g(1P)$] of hidden-charm tetraquark states, in
which the $\Sigma^+_g(1P)$ states are identified with the observed
$I \! = \! 0$, $J^{PC} \! = \! 1^{--}$ ($Y$) states according to a
variety of scenarios, based upon both their mass spectra and
preferred decay modes to eigenstates of heavy-quark spin ({\it e.g.},
$J/\psi$ {\it vs.} $\! h_c$). We calculate E1 and M1 transition
amplitudes for $\Sigma_g^{+}(1P) \! \to \! \gamma\Sigma_g^{+}(1S)$
and $\Sigma_g^{+}(1S) \! \to \! \gamma \Sigma_g^{+}(1S)$ processes,
respectively, and present corresponding values for the radiative
decay widths of a number of particular exclusive channels.
This analysis shows that if $Y(4220)$ and $X(3872)$ have a similar
underlying diquark structure, then one expects $\Gamma_{Y(4220) \to
\gamma X(3872)} \! \approx \! 100$~keV\@. Moreover, similar values
(albeit somewhat larger due to increased $\gamma$ phase space) are
expected for the heavier $Y$ states in $\Sigma^+_g(1P)$. The extreme
possibility of $Y(4660)$ belonging to the $1P$ multiplet would lead
to a $\gamma X(3872)$ branching fraction of several percent, and so
the absence of such a remarkably large signal would appear to
relegate $Y(4660)$ instead to the $\Sigma^+_g(2P)$ multiplet.
Furthermore, we found that the observed but uncon\-firmed
$Z_c(4240)$, a candidate for the sole $I \! = \! 1$, $J^{PC}=0^{--}$
state in $\Sigma^+_g(1P)$, should have a substantial ($\approx \!
500$~keV) radiative decay width to $X(3872)$ through its neutral
isobar, and therefore this decay is a good candidate for future
experimental investigation. Indeed, many of the $\Sigma^+_g(1P)$
states have not yet been observed, offering multiple potential future
tests of the model.
M1 transitions within a single multiplet, such as $Z_c(4020)^0 \! \to
\! \gamma X(3872)$, produce much narrower widths ($< \! 10$~keV in
this model), and can provide sensitive tests of substructure
({\it e.g.}, diquarks {\it vs.} meson molecules).
One may also study exotic-to-exotic radiative transitions in other
heavy-quark sectors ({\it e.g.}, hidden-bottom or $c\bar c s\bar s$
exotics). Indeed, Eqs.~(\ref{eqn:e1width}) and (\ref{eqn:m1width})
are general for any tetraquark state in the diquark-antidiquark
configuration. For example, Ref.~\cite{Giron:2020qpb} calculates the
mass of $X_b$ [the hidden-bottom analogue to $X(3872)$] to lie in a
rather narrow range $m_{X_b}\in[10598,10607]$ MeV, only slightly
below the observed $Z_b(10610)^0$. The M1 transition $Z_b(10610)^0
\! \to \! \gamma X_b$ is thus expected from Eq.~(\ref{eqn:m1width})
to produce a tiny [$O({\rm eV})$ or less] radiative width, owing to
not only the small phase space, but also the larger ($b$-containing)
diquark mass. We conclude that even very coarse experimental results
in other sectors can be decisive in verifying or falsifying
particular models.
\begin{acknowledgments}
This work was supported by the National Science Foundation (NSF) under
Grant No.\ PHY-1803912.
\end{acknowledgments}
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
One of fundamental tools in diophantine geometry is the notion of height functions and this height function measures the geometric and arithmetic complexities of rational points on an algebraic variety. These are crucial to various finiteness results in diophantine geometry such as Mordell-Weil theorem, Siegel's theorem, Mordell-Faltings' theorem, and so on. One of basic properties of height functions is the Northcott property which claims that for a height function associated to an ample divisor, the set of rational points whose height is less than $T$ is finite. Thus one may consider the counting function of rational points of bounded height, and one natural question is the asymptotic formula for such a counting function when $T$ goes to infinity.
Around the late 1980's, Yuri Manin and his collaborators proposed a general framework to understand this asymptotic formula in terms of geometric and arithmetic invariants of the underlying projective variety, and this leads to Manin's conjecture whose formulation is developed in a series of papers \cite{FMT89}, \cite{BM90}, \cite{Peyre}, \cite{BT98}, \cite{Peyre03}, \cite{Pey17}, and \cite{LST18}. One of fertile testing grounds for this conjecture is a class of equivariant compactifications of homogeneous spaces, and there are mainly two methods available, i.e., the method of mixing and the height zeta function method.
Mixing is a concept from ergodic theory, and this idea has been successfully used to prove equidistribution of rational points on homogeneous spaces acted by semi-simple groups (\cite{GMO08} and \cite{GO11}). The height zeta function method can be applied to a variety of equivariant compactifications of connected algebraic groups including, but not limited to, generalized flag varieties (\cite{FMT89}), toric varieties (\cite{BT96} and \cite{BT98}), equivariant compactifications of vector groups (\cite{CLT02}), wonderful compactifications of semi-simple groups of adjoint type (\cite{STBT07}), and biequivariant compactifications of unipotent groups (\cite{ST16}).
The height zeta function method also has its advantage to studying the counting problem of integral points associated to a reduced boundary divisor, and this has been implemented for equivariant compactifications of vector groups (\cite{CLT12}), toric varieties (\cite{CLT10toric}), wonderful compactifications of semi-simple groups of adjoint type (\cite{TBT13} and \cite{Chow19}), and biequivariant compactificaitons of the Heisenberg group (\cite{Xiao20}). These results suggest that there should be an analogous formulation of log Manin's conjecture for integral points, however certain subtleties of geometric and arithmetic nature prevent a general formulation of such a conjecture.
Campana and subsequently Abramovich proposed the notion of Campana points in \cite{Cam05} and \cite{Abr09} and this notion interpolates between rational points and integral points. The counting problem of Campana points has been originally featured in \cite{VVal11}, \cite{BVV12}, and \cite{VVal12}. Recently many mathematicians started to look at this problem and develop a series of results, attested by \cite{BY19}, \cite{PSTVA20}, \cite{PS20}, \cite{Xiao20}, and \cite{Str20}. In \cite{PSTVA20}, Pieropan, Smeets, V\'arilly-Alvarado, and the author initiated a systematic study of the counting problem for Campana points, and formulated a log Manin's conjecture for klt Campana points. Then we confirmed this conjecture for equivariant compactifications of vector groups using the height zeta function method for vector groups which is developed by Chambert-Loir and Tschinkel in \cite{CLT02} and \cite{CLT12}.
In this survey paper, we discuss the formulation of log Manin's conjecture for klt Campana points and applications of the height zeta function method to study this problem for various equivariant compactifications of connected algebraic groups.
\
Here is a plan of this paper:
In Section~\ref{sec:height}, we review the notion of height functions. In Section~\ref{sec:campana}, we introduce two definitions of (weak) Campana points. In Section~\ref{sec:logmanin}, we discuss a formulation of log Manin's conjecture for klt Campana points. Finally in Section~\ref{sec:heightzeta}, we discuss the height zeta function method and its applications to equivariant compactifications of algebraic groups.
\
\noindent
{\bf Acknowledgements:}
The author would like to thank Marta Pieropan, Arne Smeets, and Tony V\'arilly-Alvarado for collaborations helping to shape his perspective on log Manin's conjecture for klt Campana points.
The author would like to thank the organizers of RIMS workshop Problems and Prospects in Analytic Number Theory for an opportunity to give a talk there. This work was supported by the Research Institute for Mathematical Sciences, an International Joint Usage/Research Center located in Kyoto University.
Sho Tanimoto was partially supported by Inamori Foundation, by JSPS KAKENHI Early-Career Scientists Grant number 19K14512, and by MEXT Japan, Leading Initiative for Excellent Young Researchers (LEADER).
\section{Height functions}
\label{sec:height}
In this section we review the notion of height functions and their basic properties. The main references are \cite{HS00} and \cite{CLT10}, and they include different treatments of height functions. In \cite{HS00} height functions are introduced using the machinery of Weil height machine and some basic properties of height functions such as the Northcott property are proved. In \cite{CLT10}, adelic metrizations are used to define height functions, and this definition is frequently used in the literature in Manin's conjecture. It is well-known that two definitions are essentially equivalent. See \cite{HS00} for more details. In this paper, we employ the definition of height functions using adelic metrizations described in \cite{CLT10}.
\
Let us fix our notation: let $F$ be any number field and $\mathcal O_F$ be its ring of integers. We denote the set of places of $F$ by $\Omega_F$, the set of archimedean places by $\Omega_F^\infty$, and the set of non-archimedean places by $\Omega_F^{< \infty}$.
For any finite set $S \subset \Omega_F$ containing $\Omega_F^\infty$, $\mathcal O_{F, S}$ denotes the ring of $S$-integers. For each $v \in \Omega_F$, we denote the completion of $F$ with respect to $v$ by $F_v$. When $v$ is non-archimedean, we denote the ring of integers for $F_v$ by $\mathcal O_v$ with maximal ideal $\mathfrak m_v$ and residue field $k_v$ of size $q_v$. We denote the adele ring of $F$ by $\mathbb A_F$.
For each $v \in \Omega_F$, $F_v$ is a locally compact subgroup and it comes with a self-Haar measure $\mathrm dx_v = \mu_v$ which is normalized in a way Tate did in \cite{Tate}. We define the absolute value $|\cdot|_v$ on $F_v$ by requiring
\[
\mu_v(xB) = |x|_v \cdot \mu_v(B).
\]
This normalization satisfies the product formula, i.e., for any $x \in F^\times$, we have
\[
\prod_{v \in \Omega_F} |x|_v = 1
\]
See \cite{CLT10} for more details.
\
Let $F$ be a number field and $v \in \Omega_F$ be a place of $F$.
Let $U$ be an open set of $F_v^n$ in the analytic topology. A complex valued function on $U$ is smooth if it is $C^\infty$ when $v$ is archimedean and it is locally constant when $v$ is non-archimedean. This notion is local and extends to any $v$-adic analytic manifold.
Let $X$ be a smooth variety defined over $F_v$ and $L$ be a line bundle on $X$.
For each local point $x \in X(F_v)$, we denote the fiber of $L$ at $x$ by $L_x$.
\begin{defi}
A smooth metric on $L$ is a collection of metrics $\|\cdot\| : L_x(F_v) \rightarrow \mathbb R_{\geq 0}$ for all $x \in X(F_v)$ such that
\begin{itemize}
\item for $\ell \in L_x(F_v) \setminus \{0\}$, $\|\ell \| > 0$;
\item for any $a \in F_v$, $x \in X(F_v)$, and $\ell \in L_x(F_v)$, $\|a\ell \| = |a|_v \|\ell\|$, and;
\item for any open subset $U \subset X(F_v)$ and any non-vanishing section $\mathsf f \in \Gamma(U, L)$, the function $x \mapsto \|\mathsf f(x)\|$ is smooth.
\end{itemize}
\end{defi}
An integral model of a projective variety can be used to define a metric on it:
\begin{exam}
Let $X$ be a smooth projective variety defined over $F_v$ and $L$ be a line bundle on $X$ where $v$ is non-archimedean. Suppose that we have a flat projective $\mathcal O_v$-scheme $\mathcal X$ and a line bundle $\mathcal L$ on $\mathcal X$ extending $X$ and $L$. Let $x \in X(F_v) = \mathcal X(\mathcal O_v)$. Then we define a smooth metric on $L$ by insisting that for any $\ell \in L_x(F_v)$
\[
\|\ell\| \leq 1 \iff \ell \in \mathcal L_x(\mathcal O_v).
\]
This metric is called as the induced metric by an integral model $(\mathcal X, \mathcal L)$.
\end{exam}
Next we define adelic metrizations on a smooth projective variety defined over a number field $F$.
\begin{defi}
Let $X$ be a smooth projective variety defined over a number field $F$ and $L$ be a line bundle on $X$. An adelic metrization on $L$ is a collection of $v$-adic smooth metrics $\{\|\cdot\|_v\}_{v \in \Omega_F}$ on $X$ such that there exist a finite set $S$ of places including $\Omega_F^\infty$, a flat $\mathcal O_{F, S}$-projective model $\mathcal X$, and a line bundle $\mathcal L$ on $\mathcal X$ extending $X$ and $L$ such that for any $v \not\in S$, the metric $\|\cdot\|_v$ is induced by $(\mathcal X, \mathcal L)$. Note that two integral models are isomorphic outside of finitely many places so that two adelic metrizations differ only at finitely many places.
\end{defi}
Finally we define the notion of height functions:
\begin{defi}
Let $X$ be a smooth projective variety defined over a number field $F$ and $\mathcal L = (L, \{\|\cdot \|_v\})$ be an adelically metrized line bundle on $X$.
For each rational point $x \in X(F)$, choose $\ell \in L_x(F)$ and we define the height function $\mathsf H_{\mathcal L} : X(F) \to \mathbb R_{\geq 0}$ by
\[
\mathsf H_{\mathcal L}(x) = \prod_{v \in \Omega_F} \|\ell\|_v^{-1}.
\]
This is well-defined due to the product formula mentioned above.
\end{defi}
Here is an example of height functions for the projective space:
\begin{exam}
Let $X = \mathbb P^n$ and $L = \mathcal O_X(1)$. We consider the standard integral model $\mathcal X = \mathbb P^n_{\mathcal O_F}$. For each non-archimedean place $v \in \Omega_F$, we let $\|\cdot\|_v$ be the metric at $v$ induced by $\mathcal X$. For any archimedean place $v$, we define a smooth metric at $v$ by insisting
\[
\|\ell(x) \|_v = \frac{|\ell(x)|_v}{\sqrt{\sum_{i = 0}^n |x_i|_v^2}},
\]
where $x = (x_0: \cdots : x_n) \in X(F)$ and $\ell \in H^0(X, \mathcal O_X(1))$. Then it is an easy exercise to prove that the height function associated to $L$ with this adelic metrization is given by
\[
\mathsf H(x) = \prod_{v \in \Omega_F^{< \infty}} \max \{ |x_0|_v, \cdots, |x_n|_v\} \prod_{v \in \Omega_F^\infty} \sqrt{|x_0|_v^2 + \cdots + |x_n|_v^2}.
\]
When $F = \mathbb Q$, we may assume that $x_i$'s are integers and $\gcd (x_0, \cdots, x_n) = 1$. In this situation, the above formula reduces to
\[
\mathsf H(x) = \sqrt{|x_0|_\infty^2 + \cdots + |x_n|_\infty^2}.
\]
\end{exam}
Let us mention a few basic properties of height functions:
\begin{prop}
Let $X$ be a smooth projective variety defined over a number field $F$ and $\mathcal L = (L, \{\|\cdot\|_v\})$ be an adelically metrized line bundle on $X$.
Then the following statements are true:
\begin{itemize}
\item Let $\mathcal L'$ be another adelically metrized line bundle associated to $L$. Then there exist positive constants $C_1 \leq C_2$ such that for any $x \in X(F)$, we have
\[
C_1 \mathsf H_{\mathcal L'}(x) \leq \mathsf H_{\mathcal L}(x) \leq C_2 \mathsf H_{\mathcal L'}(x);
\]
\item Let $B$ be the base locus of the complete linear series $|L|$. Then there exists a positive constant $C > 0$ such that for any $x \in (X \setminus B)(F)$, we have
\[
\mathsf H_{\mathcal L}(x) \geq C;
\]
\item When $L$ is ample, for any real number $T > 0$ the set
\[
\{x \in X(F) \, \mid \, \mathsf H_{\mathcal L}(x)\ \leq T \}
\]
is a finite set.
\end{itemize}
\end{prop}
The last property is called as the Northcott property which is fundamental in diophantine geometry and it is also foudational for Manin's conjecture. For more details, see \cite{HS00}.
\section{Campana points}
\label{sec:campana}
In this section, we review the notion of Campana points. Campana points were originally considered by Campana for curves in \cite{Cam05}, and its higher dimensional analogue was explored by Abramovich in \cite{Abr09}. One may consider Campana points as integral points on Campana orbifolds developed by again Campana himself:
\begin{defi}
Let $F$ be an arbitrary field and $X$ be a smooth projective variety defined over $F$. Let $D_\epsilon = \sum_{\alpha \in \mathcal A} \epsilon_\alpha D_\alpha$ be an effective $\mathbb Q$-divisor on $X$ with $D_\alpha$'s irreducible and distinct. We say $(X, D_\epsilon)$ is a Campana orbifold if the following statements are true:
\begin{itemize}
\item For any $\alpha \in \mathcal A$, a non-negative rational number $\epsilon_\alpha$ takes the form of
\[
1- \frac{1}{m_\alpha},
\]
where $m_\alpha$ is a positive integer or $+\infty$;
\item the reduced divisor $D = \sum_{\alpha \in \mathcal A} D_\alpha$ is a strict normal crossings divisor.
\end{itemize}
We say a Campana orbifold $(X, D_\epsilon)$ is Fano if $-(K_X + D_\epsilon)$ is ample.
\end{defi}
Let $(X, D_\epsilon)$ be a Campana orbifold. Then $(X, D_\epsilon)$ is a divisorial log terminal (dlt for short) pair in the sense of birational geometry. When $\epsilon_\alpha < 1$ for any $\alpha$, $(X, D_\epsilon)$ is a kawamata log terminal (klt for short) pair. See \cite{KM98} for the definitions and their basic properties. We say a Campana orbifold $(X, D_\epsilon)$ is klt if $\epsilon_\alpha < 1$ for every $\alpha \in \mathcal A$.
\
To define the notion of Campana points, one needs to fix an integral model of a Campana orbifold.
Let $(X, D_\epsilon)$ be a Campana orbifold defined over a number field $F$ with $D_\epsilon = \sum_{\alpha \in \mathcal A} \epsilon_\alpha D_\alpha$. Let $S$ be a finite set of places including all archimedean places. A good integral model of $(X, D_\epsilon)$ away from $S$ is a flat projective $\mathcal O_{F, S}$-scheme $\mathcal X$ such that $\mathcal X$ is extending $X$ and $\mathcal X$ is regular. Let $\mathcal D_\alpha$ be the Zariski closure of $D_\alpha$ in $\mathcal X$ and let $\mathcal D_\epsilon = \sum_{\alpha \in \mathcal A} \epsilon_\alpha \mathcal D_\alpha$.
Let us fix a good integral model of a Campana orbifold $(X, D_\epsilon)$ as above. Let $\mathcal A_\epsilon = \{\alpha \in \mathcal A \, | \, \epsilon_\alpha \neq 0\}$. We set $X^\circ = X\setminus \cup_{\alpha \in \mathcal A_\epsilon} D_\alpha$. Let $P \in X^\circ(F)$ be a rational point and $v \not\in S$ be a non-archimedean place of $F$. Then we may consider $P$ as an $\mathcal O_v$-point $\mathcal P_v \in \mathcal X(\mathcal O_v)$ by valuative criterion for properness. Since $\mathcal P_v \not\subset \mathcal D_\alpha$ for any $\alpha \in \mathcal A_\epsilon$, the pullback of $\mathcal D_\alpha$ via $\mathcal P_v$ defines an ideal in $\mathcal O_v$. We denote its colength by $n_v(\mathcal D_\alpha, P)$. When $P \in D_\alpha$ for some $\alpha \in \mathcal A_\epsilon$, we formally set $n_v(\mathcal D_\alpha, P) = + \infty$. The total intersection number is given by
\[
n_v(\mathcal D_\epsilon, P) = \sum_{\alpha \in \mathcal A_\epsilon} \epsilon_\alpha n_v(\mathcal D_\alpha, P).
\]
Now we are ready to define two notions of Campana points:
\begin{defi}
We say $P \in X(F)$ is a weak Campana $\mathcal O_{F, S}$-point on $(\mathcal X, \mathcal D_\epsilon)$ if the following statements are true:
\begin{itemize}
\item we have $P \in (\mathcal X \setminus \cup_{\epsilon_\alpha = 1}\mathcal D_\alpha)(\mathcal O_{F, S})$, and;
\item for $v \not\in S$, if $n_v(\mathcal D_\epsilon, P) > 0$, then
\[
n_v(\mathcal D_\epsilon, P) \leq \left (\sum_{\alpha \in \mathcal A_\epsilon} n_v(\mathcal D_\alpha, P) \right)-1.
\]
\end{itemize}
\end{defi}
We denote the set of weak Campana $\mathcal O_{F, S}$-points by $(\mathcal X, \mathcal D_\epsilon)_w(\mathcal O_{F, S})$.
\begin{defi}
We say $P \in X(F)$ is a Campana $\mathcal O_{F, S}$-point on $(\mathcal X, \mathcal D_\epsilon)$ if the following statements are true:
\begin{itemize}
\item we have $P \in (\mathcal X \setminus \cup_{\epsilon_\alpha = 1}\mathcal D_\alpha)(\mathcal O_{F, S})$, and;
\item for $v \not\in S$ and for all $\alpha \in \mathcal A_\epsilon$ with $\epsilon_\alpha < 1$ and $n_v(\mathcal D_\alpha, P) > 0$, we have
\[
n_v(\mathcal D_\alpha, P) \geq m_\alpha,
\]
where $\epsilon_\alpha = 1-\frac{1}{m_\alpha}$.
\end{itemize}
A Campana $\mathcal O_{F, S}$-point is klt when the underlying Campana orbifold is a klt pair.
\end{defi}
We denote the set of Campana $\mathcal O_{F, S}$-points by $(\mathcal X, \mathcal D_\epsilon)(\mathcal O_{F, S})$. Then we have the following inclusions:
\[
\mathcal X^\circ(\mathcal O_{F, S}) \subset (\mathcal X, \mathcal D_\epsilon)(\mathcal O_{F, S}) \subset (\mathcal X, \mathcal D_\epsilon)_w(\mathcal O_{F, S}) \subset X(F),
\]
where $\mathcal X^\circ = \mathcal X \setminus (\cup_{\alpha \in \mathcal A_\epsilon}\mathcal D_\alpha)$. When $\epsilon_\alpha = 0$ for all $\alpha \in \mathcal A$, the rightmost two inclusions are equalities. When $\epsilon_\alpha = 1$ for all $\alpha \in \mathcal A_\epsilon$, the leftmost two inclusions are equalities.
Here is an example of klt Campana points:
\begin{exam}
For simplicity, let us assume that $F = \mathbb Q$ and $S = \{\infty\}$. Let $X = \mathbb P^n$ and $H = V(x_0)$ be a hyperplane. Let $m$ be a positive integer and $\epsilon = 1-1/m$. We define
\[
D_\epsilon = \epsilon H.
\]
We consider the standard integral model of $X$. Then a rational point $x = (x_0 : \cdots: x_n) \in X(\mathbb Q)$ with $x_i \in \mathbb Z$ and $\gcd(x_0, \cdots, x_n) = 1$ is a Campana $\mathbb Z$-point if $x_0 = 0$ or $x_0 \neq 0$ and the following statement is true: for any prime number $p$ we have
\[
p \mid x_0 \implies p^m \mid x_0.
\]
Any non-zero integer with this property is said to be $m$-full. When $m = 2$, it is said to be squarefull.
\end{exam}
\section{Log Manin's conjecture}
\label{sec:logmanin}
Let $X$ be a smooth projective variety defined over a number field $F$ and $\mathcal L = (L, \{\|\cdot\|_v\})$ be an adelically metrized line bundle on $X$. We consider the associated height function
\[
\mathsf H_{\mathcal L} : X(F) \to \mathbb R_{>0}.
\]
When $L$ is ample, this height function satisfies the Northcott property so that for any subset $Q \subset X(F)$ and any positive real number $T>0$ one may define the counting function
\[
N(Q, \mathcal L, T) = \# \{P \in Q \, | \, \mathsf H_{\mathcal L}(P) \leq T\}.
\]
Manin's conjecture predicts the asymptotic formula of the above function for an appropriate $Q$, and a natural question is to extend this conjecture to integral points and Campana points. In \cite{PSTVA20}, Pieropan, Smeets, V\'arilly-Alvarado and the author formulated this log version of Manin's conjecture when the underlying Campana orbifold is a klt log Fano pair. In this section, we review a general formulation of this log Manin's conjecture.
\subsection{Two birational invariants}
Let $X$ be a smooth projective variety defined over a field $F$. Let $D_1, D_2$ are $\mathbb Q$-divisors on $X$. We say $D_1$ and $D_2$ are numerically equivalent if for any curve $C \subset X$, we have $D_1.C = D_2.C$. In this case we write $D_1 \equiv D_2$. We define the space of $\mathbb Q$-divisors up to numerical equivalence as
\[
N^1(X)_{\mathbb Q} = \{D : \text{$\mathbb Q$-divisors}\}/\equiv.
\]
We set $N^1(X) := N^1(X)_{\mathbb Q} \otimes_{\mathbb Q}\mathbb R$.
Then we define the cone of pseudo-effective divisors by
\[
\overline{\mathrm{Eff}}^1(X) := \overline{\text{the cone of effective $\mathbb Q$-divisors}} \subset N^1(X).
\]
Now we are ready to introduce two birational invariants which play central roles in Manin's conjecture:
\begin{defi}
Let $(X, D_\epsilon)$ be a klt Campana orbifold defined over a field $F$ and $L$ be an ample $\mathbb Q$-divisor on $X$. We define the Fujita invariant or $a$-invariant by
\[
a(X, D_\epsilon, L) := \inf \{t \in \mathbb R \, |\, tL + K_X + D_\epsilon \in \overline{\mathrm{Eff}}^1(X)\}.
\]
Next assume that $a(X, D_\epsilon, L) >0$. Then we define the $b$-invariant by
\begin{align*}
b(F, X, D_\epsilon, L) := & \text{ codimension of the minimal face of $\overline{\mathrm{Eff}}^1(X)$}\\ & \text{ containing $a(X, D_\epsilon, L)L + K_X + D_\epsilon$.}
\end{align*}
It is explained in \cite[Section 3.6.2]{PSTVA20} that these invariants are birational invariants.
\end{defi}
\begin{exam}
Let $(X, D_\epsilon)$ be a klt Fano orbifold defined over a field $F$ and $L = -(K_X + D_\epsilon)$.
Then we have
\[
a(X, D_\epsilon, L) = 1, \quad b(F, X, D_\epsilon, L) = \rho (X) = \dim N^1(X).
\]
\end{exam}
\subsection{Thin exceptional sets}
The notion of thin sets has been explored by Serre to study Galois inverse problem, and it is also fundamental to Manin's conjecture. Let us give the definition of thin sets for Campana points:
\begin{defi}
Let $(X, D_\epsilon)$ be a klt Campana orbifold defined over a number field $F$. Let $S$ be a finite set of places of $F$ including $\Omega_F^\infty$ and we fix a good integral model away from $S$ $\mathcal X \to \mathrm{Spec} \, \mathcal O_{F, S}$.
A type I thin set is a set of the form
\[
V(F) \cap (\mathcal X, \mathcal D_\epsilon)(\mathcal O_{F, S}),
\]
where $V \subset X$ is a proper closed subset of $X$.
A type II thin set is a set of the form
\[
f(Y(F))\cap (\mathcal X, \mathcal D_\epsilon)(\mathcal O_{F, S}),
\]
where $f : Y \to X$ is a dominant generically finite morphism of degree $\geq 2$ defined over $F$ with $Y$ integral.
A thin set is any subset of a finite union of type I and type II thin sets.
\end{defi}
Here is an example of thin sets:
\begin{exam}
Let $X = \mathbb P^1$ with $D_\epsilon = 0$ defined over a number field $F$.
We consider the morphism
\[
f : \mathbb P^1 \to \mathbb P^1, \, (x_0:x_1) \mapsto (x_0^d:x_1^d)
\]
with $d \geq 2$.
Then $f(X(F)) \subset X(F)$ is a thin set.
\end{exam}
\subsection{Log Manin's conjecture for klt Campana points}
Finally we state log Manin's conjecture for klt Campana points:
\begin{conj}[Log Manin's conjecture for klt Campana points]
\label{conj:main}
Let $(X, D_\epsilon)$ be a klt Fano orbifold defined over a number field $F$ and $\mathcal L = (L, \{\|\cdot\|_v\})$ be an adelically metrized ample line bundle. Assume that $(\mathcal X, \mathcal D_\epsilon)(\mathcal O_{F, S})$ is not thin. Then there exists a thin set $Z \subset (\mathcal X, \mathcal D_\epsilon)(\mathcal O_{F, S})$ such that
\[
N((\mathcal X, \mathcal D_\epsilon)(\mathcal O_{F, S}) \setminus Z, \mathcal L, T) \sim c(F, \mathcal X, \mathcal D_\epsilon, \mathcal L, Z)T^{a(X, D_\epsilon, L)} (\log T)^{b(F, X, D_\epsilon, L)-1},
\]
as $T \to \infty$. Here the leading constant $c(F, \mathcal X, \mathcal D_\epsilon, \mathcal L, Z)$ is analogous to Peyre's constant developed in \cite{Peyre} and \cite{BT98} and its definition is given in \cite[Section 3.3]{PSTVA20}.
\end{conj}
\begin{rema}
For a smooth geometrically rationally connected projective variety $X$ defined over a number field $F$, it is expected that $X(F)$ is not thin as soon as there is a rational point. Indeed, Colliot-Th\'etl\`ene's conjecture predicts that the set of rational points is dense in the Brauer-Manin set, and this implies that $X$ satisfies weak weak approximation. It is known that weak weak approximation property implies non-thinness of the set of rational points. The corresponding statement for klt Campana points, i.e., weak weak approximation for klt Campana sets implies non-thiness of the set of klt Campana points is established in \cite{NS20}. So it is natural to expect that the assumption of Conjecture~\ref{conj:main} is true as long as there is a klt Campana point.
\end{rema}
\begin{rema}
It is well-documented in the case of rational points that in Conjecture~\ref{conj:main} it is important to remove the contribution of a thin set $Z$ from the counting function. There is a series of papers (\cite{LTDuke}, \cite{Sen17}, and \cite{LST18}) studying birational geometry of thin exceptional subsets for rational points. In \cite{LST18}, Lehmann, Sengupta, and the author proposed a conjectural description of thin exceptional subsets and proved that it is indeed a thin set using the minimal model program and the boundedness of singular Fano varieties. It would be interesting to perform a similar study for klt Campana points.
\end{rema}
Conjecture~\ref{conj:main} is known in the following cases:
\begin{itemize}
\item projective space with a boundary being the union of hyperplanes (\cite{VVal11}, \cite{VVal12}, \cite{BVV12}, and \cite{BY19});
\item equivariant compactifications of vector groups (\cite{PSTVA20});
\item toric varieties defined over $\mathbb Q$ (\cite{PS20}) and;
\item biequivaraint compactifications of the Heisenberg group (\cite{Xiao20}).
\end{itemize}
One can also consider a similar counting problem for weak Campana points, however this problem is much harder than Conjecture~\ref{conj:main}. At the moment of writing this paper, we do not know how one should formulate a log Manin's conjecture for weak Campana points, but \cite{Str20} takes the first step towards to this problem.
\section{Height zeta functions}
\label{sec:heightzeta}
Let $F$ be a number field and $G$ be a connected linear algebraic group defined over $F$.
Let $X$ be a smooth projective equivariant compacitification of $G$, i.e., $X$ contains $G$ as a Zariski open subset, and the right action of $G$ extends to $X$. In this situation, the boundary
\[
D = X \setminus G = \bigcup_{\alpha \in \mathcal A} D_\alpha
\]
is a divisor where each $D_\alpha$ is an irreducible component. After applying an equivariant resolution, we may assume that $D = \sum_{\alpha \in \mathcal A}D_\alpha$ is a divisor with strict normal crossings. We also fix an adelic metriazation for $\mathcal O(D_\alpha)$ for every $\alpha \in \mathcal A$.
For each $\alpha \in \mathcal A$, we choose $m_\alpha$ which is a positive integer or $+ \infty$, and set $\epsilon_\alpha = 1-\frac{1}{m_\alpha}$. We consider
\[
D_\epsilon = \sum_{\alpha \in \mathcal A}\epsilon_\alpha D_\alpha,
\]
and $(X, D_\epsilon)$ is a Campana orbifold.
Let us fix a finite set $S$ of places including $\Omega_F^\infty$ and a good integral model $\mathcal X$ away from $S$ extending $X$.
When $-(K_X + D_\epsilon)$ is ample (or more generally big), it is natural to consider the counting problem of $\mathcal O_{F, S}$-Campana points on $(\mathcal X, \mathcal D_\epsilon)$. There is a general approach to this problem which is called as the height zeta function method.
Let $\mathrm{Pic}(X)^G$ be the Picard group of $G$-linearlized line bundles on $X$ up to isomorphisms. (If the reader is not familiar with $G$-linearlizations, she/he may ignore this term for now.) After tensoring by $\mathbb Q$, boundary components $D_\alpha$ form a basis for $\mathrm{Pic}(X)^G_{\mathbb Q}$.
We choose a section
\[
\mathsf f_\alpha \in H^0(X, \mathcal O(D_\alpha)),
\]
corresponding to $D_\alpha$. Then we define a local height pairing: for any place $v\in \Omega_F$,
\[
\mathsf H_v : G(F_v) \times \mathrm{Pic}(X)^G_{\mathbb C} \to \mathbb C^\times, \left(g_v, \sum_{\alpha \in \mathcal A} s_\alpha D_\alpha\right) \mapsto \prod_{\alpha \in \mathcal A} \|\mathsf f_\alpha (g_v)\|_v^{-s_\alpha}.
\]
Using this local height pairing, we define the global height pairing as the Euler product:
\[
\mathsf H := \prod_{v \in \Omega_F} \mathsf H_v : G(\mathbb A_F) \times \mathrm{Pic}(X)^G_{\mathbb C} \to \mathbb C^\times.
\]
Applying the definition of Campana points to local points, for each $v \not\in S$, one can define the Campana set
\[
(\mathcal X, \mathcal D_\epsilon)(\mathcal O_v) \subset X(F_v).
\]
We set
\[
G(F_v)_\epsilon = G(F_v) \cap (\mathcal X, \mathcal D_\epsilon)(\mathcal O_v),
\]
and let $\delta_{\epsilon, v}(g_v)$ be the characteristic function of $G(F_v)_\epsilon$ on $G(F_v)$.
When $v \in S$, we set $\delta_{\epsilon, v}\equiv 1$ and define $\delta_\epsilon$ as the Euler product:
\[
\delta_\epsilon = \prod_{v \in \Omega_F} \delta_{\epsilon, v} : G(\mathbb A_F) \to \mathbb R_{\geq 0}.
\]
For $g \in G(\mathbb A_F)$ and $\bold s \in \mathrm{Pic}(X)^G_{\mathbb C}$, we define the height zeta function by
\[
\mathsf Z(g, \bold s) := \sum_{\gamma \in G(F)} \mathsf H(\gamma g, \bold s)^{-1}\delta_\epsilon(\gamma g)
\]
When $\Re(\bold s)$ is sufficiently large, this function converges to a continuous function in $g \in G(F)\backslash G(\mathbb A_F)$ and a holomorphic function in $\bold s \in \mathrm{Pic}(X)^G_{\mathbb C}$.
A relation of this height zeta function to log Manin's conjecture is given by Tauberian theorem. Indeed, if one can prove that for an ample (or big) line bundle $L$, $\mathsf Z(\mathrm{id}, sL)$ admits a meromorphic continuation to a half plane $\Re(s) \geq a$ with a unique pole at $s = a$ of order $b$ with $a >0$ a positive real number, then one can conclude
\[
N(G(F)_\epsilon, \mathcal L, T) \sim c T^{a} (\log T)^{b-1},
\]
where $c$ is a positive constant related to the leading constant of $\mathsf Z(\mathrm{id}, sL)$ at $s = a$. Thus our goal is reduced to obtain a meromorphic continuation of $\mathsf Z(\mathrm{id}, \bold s)$.
To this end, for $\bold s \gg 0$, one can prove that
\[
\mathsf Z(g, \bold s) \in \mathsf L^2(G(F)\backslash G(\mathbb A_F)),
\]
thus one may apply spectral decomposition of this Hilbert space to $\mathsf Z(g, \bold s)$ and use this spectral decomposition to obtain a meromorphic continuation.
This program has been pioneered mainly by Tschinkel and his collaborators, and has been carried out in the following cases:
\begin{itemize}
\item rational points on toric varieties (\cite{BT96}, \cite{BTtoric});
\item rational points on equivariant compacitifications of vector groups (\cite{CLT02});
\item rational points on wonderful compactifications of semi-simple groups of adjoint type (\cite{STBT07});
\item rational points on biequivariant compactifications of unipotent groups (\cite{ST16});
\item integral points on equivariant compacitificaitons of vector groups (\cite{CLT12});
\item integral points on toric varieties (\cite{CLT10toric});
\item integral points on wonderful compactificaitons of semi-simple groups of adjoint type (\cite{TBT13} and \cite{Chow19});
\item Campana points on equivariant compacitifications of vector groups (\cite{PSTVA20});
\item Campana points on biequivariant compacitifications of the Heisenberg group (\cite{Xiao20}), and;
\item weak Campana points on certain toric varieties (\cite{Str20}).
\end{itemize}
It would be interesting to explore Campana points on other algebraic groups. In particular, the treatment of integral points on toric varieties (\cite{CLT10toric}) is known to be incomplete, and there is some technical issue on this paper. It would be interesting to apply the height zeta function method to klt Campana points on toric varieties and see whether we have a similar issue.
Finally for the readers who are interested in working examples of this program, we recommend them to consult \cite[Interlude I]{PSTVA20}.
\nocite{*}
\bibliographystyle{alpha}
|
\section{Introduction} \label{sec:introduction}
Classical physical theories begin with scientific laws as ans\"atze, which are validated by repeated scientific experiment.
From these laws, one derives a set of equations (usually differential equations) that can be solved, either completely or partially, to deduce various conclusions about the physical system under consideration.
In this paper, we follow a different approach to
learning physical equations: we solve an optimization problem that isolates the most likely physical model (differential equations) that would deliver certain physical measurements (data).
This approach is aligned with a growing trend in data-driven science; see, e.g., \cite{crutchfield1987equations,brunton2016discovering,raissi2018hidden,chen2018neural,long2018pde,han2019uniformly,wu2019numerical,rackauckas2020universal,huang2020learning,pmlr-v107-sun20a,yazdani2020systems,dandekar2020bayesian}.
The focus of this paper is on learning mechanical models for binary black hole (BBH) systems through gravitational wave measurements.
As the black holes orbit one another, the motion of these massive objects generate
gravitational waves that radiate away to the far-field where they
can be observed by an international-network of detectors~\cite{abbott2020gwtc}.
Complicated partial differential equations (PDEs) govern the entire process, and in particular
connect the near-field dynamics to the far-field gravitational radiation. Traditionally,
black hole orbital dynamics and
gravitational waves have been computed by expensive simulation codes~\cite{lehner2014numerical}
or approximations to general relativity such as the post-Newtonian formalism~\cite{blanchet2014gravitational}.
Our principal contribution is to show that two-body relativistic orbital models can be deduced from gravitational wave (GW) measurements
by solving an inverse problem \cite{vogel2002computational,arridge_maass_oktem_schonlieb_2019} where the control variable is the vector of weights and biases in a neural network. Our numerical examples use
gravitational waveform measurements from both a noisy detector and
``clean" measurements from numerical relativity (NR)
simulations. As such, the techniques described here may apply to
both traditional modeling endeavors that require NR data
for calibration and GW astronomy. In the latter case,
where the training data is comprised of GW observations,
our inversion strategy avoids the need to solve Einstein's equation of general relativity
to learn the orbital model.
A key goal of our paper is to
develop the computational framework for learning binary
black hole dynamical models from gravitational waves,
which is a new approach to the modeling problem. As such,
we focus on simple modeling choices and apply them
to illustrative examples. We will show that
simple ansatz models parameterized by feed-forward neural networks
can be used to discover complicated dynamics. Indeed, despite starting
with an essentially Newtonian ansatz model (cf. Eq.~\eqref{eq:UDEModel}),
our trained models accurately capture both the relativistic dynamics
and the waveform (cf. Sec.~\ref{sec:examples}). Potential applications are considered
in the Discussion section.
\section{Methodology}
\subsection{Universal differential equations} \label{sec:universal_differential_equations}
In this work, we rely on the following general class of dynamical system models referred to in \cite{rackauckas2020universal} as (autonomous) universal differential equations (UDEs),
\begin{align}
\label{eq:UDE}
\dot{\mathbf{x}}
=
\mathbf{f}(\mathbf{x},\mathcal{F}(\mathbf{x}))
,
\qquad
\mathbf{x}(0) &= \mathbf{x}_0
\,,
\end{align}
where $\mathbf{x}(t)$ is the solution vector and $\mathbf{x}_0$ specifies the initial conditions.
Here, $\mathcal{F}(\mathbf{x})$ is neural network and the overdot symbol ``$~\dot{}~$'' denotes differentiation with respect to time $t \in [0,T]$.
For example, if $\mathcal{F}(\mathbf{x}) = \mathcal{F}(\mathbf{x};\bm{\xi})$ is a feed-forward deep neural network with two hidden layers, it can be written as
\begin{equation}
\label{eq:TwoHiddenLayers}
\mathcal{F}(\mathbf{x};\bm{\xi})
=
W_3\,\sigma_2(W_2\,\sigma_1(W_1(\mathbf{x}) + b_1)+b_2)+b_3 \,,
\end{equation}
with $\bm{\xi} = (W_1,W_2,W_3,b_1,b_2,b_3)$.
Here, $W_i$ are matrices (weights), $b_i$ are vectors (biases), and $\sigma_i$ are the chosen activation functions.
We note that the UDE paradigm permits a immense variety of different parameterized functions $\mathcal{F}(\mathbf{x})$, not only neural network-based parameterizations.
This flexibility provides numerous advantages, for example, prior scientific knowledge of the solution may be incorporated into the choice of function parametrization.
We choose to use feed-forward neural networks, such as~\cref{eq:TwoHiddenLayers}, because of the ease with which their parametrization can be determined using existing software \cite{rackauckas2020universal}.
\subsection{BBH modeling} \label{sec:physically_motivated_models}
Our task is to define a family of physical models that can be used to describe
relativistic orbital dynamics of two spherical objects of mass $m_1$ and $m_2$. That is, we
ask that our model provides the position of object 1, $\mathbf{r}_1(t)$, and
object 2, $\mathbf{r}_2(t)$, which
are the solutions to the dynamical system model.
In Newtonian physics,
this is the familiar two-body problem of Kepler whose solution has been known since
the earliest days of classical mechanics.
In relativistic physics, the field of
computational relativity is largely devoted to providing numerical solutions to this
problem by solving the equations of general
relativity (a nonlinear, coupled, hyperbolic-elliptic PDE system) on
large supercomputers~\cite{lehner2014numerical}. In order to motivate
a dynamical system model, we consider special cases whereby the
dynamical motion described by PDEs can be
approximately reduced to ordinary differential equations (ODEs).
Such approximations have been well developed over many decades,
and we refer the reader
to~\cite{poisson2014gravity,blanchet2014gravitational,chandrasekhar1998mathematical}
and references therein. Throughout this paper,
we use geometric units where both the
speed of light, $c$, and the gravitational constant, $G$,
are set to unity.
First, for slow moving
objects ($v \ll c$), a powerful formalism
known as the post-Newtonian approximation
provides a systematic framework for
adding relativistic corrections in powers of $v/c$ to the Keplerian equations of motion.
The post-Newtonian framework provides us with justification for treating the two-body problem
as an effective one-body problem. Here, the relevant equations that govern the separation
vector, $\mathbf{r} = \mathbf{r}_1 - \mathbf{r}_2$, can be used to reconstruct the two-body motion through
the relations,
\begin{align}
\label{eq:body_map}
\mathbf{r}_1 = (m_2/M) \mathbf{r} \,, \qquad \mathbf{r}_1 = -(m_1/M) \mathbf{r} \,,
\end{align}
where $M = m_1 + m_2$.
We call this an effective one-body problem as one often views there to be an ``effective" body
located at $\mathbf{r}$ relative to the system's center-of-mass.
A different technique, known as the post-Minkowskian approximation, informs us that the dominant
contribution to the
gravitational radiation field can be computed according to the quadrupole formula, first derived by Einstein.
Collectively, the two-body-to-one-body map and the quadrupole formula neglect a
substantial amount of physics including terms proportional to
$v / c$, as well as many more higher-order terms, some of which have been computed and others that
remain unknown.
Second, in the limit of $m_1 \gg m_2$, the two-body problem reduces to a
simpler setup whereby the larger
object is fixed at the coordinate system's origin. The smaller orbiting object's motion is then
described by a geodesic path in the Schwarzschild geometry set by the
larger black hole. Recent numerical evidence suggests
that the geodesic equations of motion~\eqref{eq:EMRIModel}
with self-force corrections may work unreasonably well
even for near-equal mass systems~\cite{lewis2017fundamental,zimmerman2016redshift,le2011periastron,le2012gravitational,le2014overlap,le2013periastron,rifat2020surrogate,pound2019secondorder,van2020intermediate}.
More importantly for our purposes, we know from
blackhole perturbation theory results that using geodesic equations of motion
to describe the two-body problem neglects a
substantial amount of physics including terms proportional to $m_2 / m_1$.
Having motivated some of the physics behind the problem, we now outline
our strategy to write down a model
inspired by a combination of Newtonian and relativistic physics.
Our orbital model will omit a significant amount of important physics
that will be accounted for by deep neural networks trained on
gravitational waveform data.
We write the two-body problem
as an effective one-body one, and associate the orbital separation vector, $\mathbf{r}$,
with the location of the fictitious effective body orbiting a fixed, spherically-symmetric central object.
Owing to the spherical symmetry of the central object,
we may assume, without loss of generality, that the effective object's trajectory
lies in the equatorial plane, which we take to be
the plane perpendicular to the z-axis and where the
\emph{angle} $\phi$ is between $\mathbf{r}$ and the x-axis.
Orbits are specified by
the orbital parameters
{\em eccentricity} $e(t)$ and {\em semi--latus rectum} $p(t)$; in the Newtonian
case these are constants while in the general relativistic case they are
often interpreted as time-dependent
functions. Finally, we use a well-known parameterization for the Euclidean norm of $\mathbf{r}$,
\begin{align} \label{eq:r_parameterization}
r(t) = p(t) M / (1+e(t) \cos\chi(t)) \,,
\end{align}
and evolve the \emph{anomaly} $\chi(t)$
instead of $r(t)$ because the anomaly increases
monotonically through radial turning points. To summarize, we assume the equations of motion
for the effective object can be described by four time-dependent variables
$\phi$, $\chi$, $p$, and $e$. The effective object's trajectory
is provided $\phi,~\chi$ whereas $p,~e$ parameterize
the orbital configuration.
\begin{figure*}
\centering
\includegraphics[width=0.9\linewidth]{Flowchart2.png}
\caption{\label{fig:Flowchart}
Flowchart of algorithm to solve the inverse problem~\cref{eq:InverseProblem}.
}
\end{figure*}
Upon denoting $\mathbf{x} = (\phi,\chi,p,e)$, we propose the following family of UDEs to describe
the two-body relativistic dynamics:
\begin{subequations}
\label{eq:UDEModel}
\begin{align}
\label{eq:UDEModel_phi}
\dot \phi
&=
\frac{(1+e\cos(\chi))^2}{Mp^{3/2}}
\big(
1 + \mathcal{F}_1(\cos(\chi),p,e)
\big),
\\\label{eq:UDEModel_chi}
\dot \chi
&=
\frac{(1+e\cos(\chi))^2}{Mp^{3/2}}
\big(
1 + \mathcal{F}_2(\cos(\chi),p,e)
\big),
\\\label{eq:UDEModel_p}
\dot p
&=
\mathcal{F}_3(p,e),
\\\label{eq:UDEModel_e}
\dot e
&=
\mathcal{F}_4(p,e),
\end{align}
\end{subequations}
with $\mathbf{x}(0) = (\phi_0,\chi_0,p_0,e_0)$.
Note that the functional form of Eqs.~\eqref{eq:UDEModel} have been inspired by
Eq.~\eqref{eq:EMRIModel}, which are the geodesic equations of motion for an infinitesimally small
``particle" orbiting a super-massive blackhole.
In particular, Eqs.~\cref{eq:UDEModel} are rotationally invariant because the right-hand side omits the $\phi$-variable.
Moreover, when $\mathcal{F}_3 = \mathcal{F}_4 = 0$, orbital energy, $E(p,e)$, and orbital angular momentum, $L(p,e)$, are conserved:
\begin{align}
\dot{E} = \frac{\partial E}{\partial p}\dot{p} + \frac{\partial E}{\partial e}\dot{e}
,
\qquad \dot{L} = \frac{\partial L}{\partial p}\dot{p} + \frac{\partial L}{\partial e}\dot{e}
.
\end{align}
Due to the emission of gravitational waves (so-called radiation-reaction), we have that both $\dot{E},\dot{L}<0$ for all time. When each $\mathcal{F}_j = 0$, we recover Newtonian orbits.
Eqs.~\Crefrange{eq:UDEModel_phi}{eq:UDEModel_e} define a family of
trajectories $\mathbf{x}(t) = \mathbf{x}(t;{\bm{\xi}})$.
Through Eq.~\eqref{eq:body_map}, these trajectories determine the
black hole orbits,
\begin{subequations}
\label{eq:Distance}
\begin{align}
\label{eq:r1}
\mathbf{r}_1(t)
=
\frac{r(t) m_2}{M}\left(\cos(\phi(t)),\sin(\phi(t)), 0 \right) \,,
\\\label{eq:r2}
\mathbf{r}_2(t)
=
-\frac{r(t) m_1}{M}\left(\cos(\phi(t)),\sin(\phi(t)), 0 \right)
\,.
\end{align}
\end{subequations}
Gravitational waves are generated
by orbiting black holes, and so the waves encode detailed information about the
dynamical variables $\mathbf{r}_1(t)$ and $\mathbf{r}_2(t)$.
General relativity tells us that the dynamics and waves are connected through
PDEs, which is a familiar scenario in the modeling of waves.
In the next section
we summarize how to learn $\mathcal{F}_j$ from gravitational wave measurements.
Despite the simplicity of Eqs.~\eqref{eq:UDEModel}, we show that the learned ODEs can describe dynamics beyond the base mechanical model (the base model corresponds to setting the neural network parameters $\bm{\xi}$ to zero and, thus, each $\mathcal{F}_j = 0$). In our first numerical experiment~\ref{sub:extreme_mass_ratio_waveforms}, for example, we show that bound orbits of a test particle following geodesic motion on a Schwarzschild geometry can be accounted for. In our final set of numerical experiments we show that the dissipative dynamics can also be accounted for. Due to the flexible framework of our waveform inversion technique,
one can easily swap out our base model~\eqref{eq:UDEModel} for others; e.g., the EOB model \cite{buonanno1999effective}.
This suggests many possible future applications of gravitational waveform inversion.
\subsection{Quadrupole formula, the loss function, and model discovery} \label{sec:model_discovery}
Very far from a BBH system, where gravitational wave detectors are located, the gravitational radiation
field is an outgoing spherical wavefront. On a sufficiently large sphere we can expand the radiation field
into a complete basis of (tensorial) spherical harmonics labeled by $(\ell, m)$ harmonic indices.
In this paper, we consider only the dominant $(\ell, m)=(2,2)$-mode gravitational waveforms (cf.~\Cref{sub:gravitational_waves_from_an_orbit}), however, our waveform inversion technique could be modified to easily include subdominant modes.
Accordingly, we denote all waveforms by the variable $w = (r/M)\cdot\mathrm{Re}\{h^{22}\}$, where
\begin{align}
\label{eq:gw_22_v2}
h^{22}(t) &=\dfrac{1}{r}\sqrt{\dfrac{4\pi}{5}}\left(\ddot{\mathcal{I}}_{xx}
- 2 i\ddot{\mathcal{I}}_{xy}-\ddot{\mathcal{I}}_{yy}\right)\,
\end{align}
and the trace-free mass quadrupole tensors $\mathcal{I}_{xx}$, $\mathcal{I}_{xy}$, and $\mathcal{I}_{yy}$ are defined in Eq.~\cref{eq:gw_mathcalI} (see also \cite[Eqs.~54--56]{bishop2016extraction}). Eq.~\eqref{eq:gw_22_v2} is the well-known quadrupole formula which
expresses the measurable waveform, $h^{22}$, in terms of the orbits $\mathbf{r}_1 = (x_1,y_1,0)$ and $\mathbf{r}_2 = (x_2,y_2,0)$. The quadrupole formula
is a very simple approximation that will necessarily introduce systematic error when learning $\mathcal{F}_j$, but
is sufficient for our purpose.
We assume that our waveform measurements appear as ordered pairs $(t_k,w_k)$, where $w_k$ denotes the value of the waveform data at time $t_k \in [0,T]$.
In this setting, we define the mean-squared waveform error
\begin{equation}
\label{eq:Loss}
\mathcal{J}(\mathbf{x})
=
\langle J(\mathbf{x},\cdot) \rangle
:=
\frac{1}{T}
\int_{0}^{T}
J(\mathbf{x},t)
\,
\mathrm{d}t
,
\end{equation}
where $J(\mathbf{x},t) = \sum_{k} \big( w_k - w(t) \big)^2\delta(t-t_k)$
and bracket notation, $\langle \cdot \rangle$, denotes
denotes averaging over the time interval.
Accordingly, we choose to solve the inverse problem:
\begin{gather}
\label{eq:InverseProblem}
\min_{{\bm{\xi}}}
\mathcal{J}(\mathbf{x})
\quad
\text{subject to~\Crefrange{eq:UDEModel_phi}{eq:UDEModel_e}}.
\end{gather}
In some situations, convergence to the solution of~\cref{eq:InverseProblem} can be improved by adding well-chosen,
physics-informed penalty and regularization to~\cref{eq:Loss}; cf. \Cref{sub:radiation_reaction_in_numerical_relativity_experiments}.
We note that the exclusive use of gravitational-wave data
in the loss function is motivated by the consideration
that in experimental settings only gravitational-wave
observations will be available and never a direct view
of black hole orbits. Even in computational relativity simulations,
the numerical measurement of black hole trajectories are complicated
by coordinate ambiguities of general relativity that
make it difficult to assign physical significance to their values.
Waveforms computed from computational relativity simulations,
on the other hand, are well-defined and physically meaningful.
The ODE-constrained optimization problem~\cref{eq:InverseProblem} delivers the calibrated dynamical system model $\dot{\mathbf{x}} = \mathbf{f}(\mathbf{x},\mathcal{F}(\mathbf{x};\bm{\xi}^\star))$, where $\bm{\xi}^\star$ denotes the optimizer found by solving~\cref{eq:InverseProblem}.
This inverse problem can be solved with a number of standard methods.
We choose to use a BFGS algorithm with backtracking line search \cite{nocedal2006numerical} and an adjoint-based (implicit differentiation/adjoint sensitivity method) calculation of gradients \cite{boltyanskiy1962mathematical} implemented with the Julia \cite{bezanson2017julia} software package DiffEqFlux \cite{rackauckas2020universal}.
The algorithm is described by the flowchart in~\Cref{fig:Flowchart}.
Our code is available for download at \cite{JuliaCode}.
\section{Results} \label{sec:examples}
In this section, we present results with three different examples.
The first demonstrates the ability of~\cref{eq:InverseProblem} to recover known orbital equations.
The second two showcase the discovery of new equations of motion for equal mass binary black hole mergers.
\begin{figure*}[!htp]
\centering
\begin{minipage}{0.55\linewidth}
\includegraphics[width=\linewidth]{EMRI_trajectories.pdf}
\end{minipage}
~~
\begin{minipage}{0.32\linewidth}
\begin{flushleft}
\includegraphics[width=\textwidth]{EMRI_error_surf_phi.png}\\[2pt]
\includegraphics[width=\textwidth]{EMRI_error_surf_chi.png}
\end{flushleft}
\end{minipage}
\\
\includegraphics[width=\linewidth]{EMRI_waveforms.pdf}
\caption{\label{fig:EMRIWaveform} Summary of our first experiment, where we have used gravitational-wave observations
(black dots; bottom panel) to learn the underlying
two-dimensional dynamical system model governing the relativistic two-body problem in the extreme mass ratio limit
with orbital parameters $p=100,~e=0.5$.
Top left: Learned (dashed red) and exact (solid blue) trajectories extrapolated $4\times$ the training interval.
We also show the portion of the orbit (black) corresponding to the gravitational-wave training window, although no
orbital data was used to learn the dynamics.
Top right: Relative error between the learned model~\cref{eq:UDEModel} and the exact model~\cref{eq:EMRIModel}.
Bottom: Learned (dashed red) and exact (solid blue) waveforms extrapolated $4\times$ the training interval.
}
\end{figure*}
\subsection{Extreme mass ratio systems} \label{sub:extreme_mass_ratio_waveforms}
As our first motivating example, we consider a special case of the relativistic
two-body problem where the exact solution is known. We show that from short-duration gravitational
wave observations we are able to discover
differential equations that are valid over much longer time-scales.
In the regime of $m_1 \gg m_2$, formally the limit $m_1 \rightarrow M$, $m_2$ is a ``test particle"
whose motion obeys~\cite{chandrasekhar1998mathematical,cutler1994gravitational,martel2004gravitational,field2009discontinuous}
\begin{subequations}
\label{eq:EMRIModel}
\begin{align}
\dot{\phi}
&=
\frac{(p-2-2e\cos\chi)(1+e\cos\chi)^2}{Mp^{3/2}\big[(p-2)^2-4e^2\big]^{1/2}}
,
\\
\dot{\chi}
&=
\frac{(p-2-2e\cos\chi)(1+e\cos\chi)^2\big[p-6-2e\cos\chi\big]^{1/2}}{Mp^2 \big[(p-2)^2-4e^2\big]^{1/2}}
,
\end{align}
\end{subequations}
while $\mathbf{r}_1 = (0,0,0)$ and $\dot{e}=\dot{p}=0$.
We shall be interested in the
parameter restriction $0\leq e <1$, for which the radial motion occurs between two
turning points, $pM/(1+e)$ and $pM/(1-e)$ and the orbit is bounded.
When $e = 0$, the orbit is circular.
We let $\mathcal{F}_3 = \mathcal{F}_4 = 0$ and
provide values for the initial conditions $\phi_0 = 0$ and $\chi_0 = \pi$.
In our example, we set $e = 0.5$, $p = 100$, and $m_1 = 1$, although the results we show
remain largely the same for other parameter values we have tested. For simplicity, in this
first example we provide known values for $e_0, p_0$ while
in Sec.~\ref{sub:radiation_reaction_in_numerical_relativity_experiments} we show how
our approach performs when these parameters are also learned.
To prepare our ground-truth data,
we numerically solve Eq.~\eqref{eq:EMRIModel} on a dense time grid, thereby
generating the black hole trajectory $\mathbf{r}_2(t)$.
We then apply the quadropole formula Eq.~\eqref{eq:gw_22_v2} to generate a gravitational waveform
sampled at $250$ equally-spaced points spanning the time interval $[0, 0.6\cdot 10^5]$,
and shown in Fig.~\ref{fig:EMRIWaveform} (bottom panel; black dots).
Note that in the extreme mass ratio limit,
$m_2 \rightarrow 0$, and the waveform $h^{22} \propto m_2 / m_1$
goes to zero.
Therefore, in this example, we use
$w = (m_1/m_2) \cdot \mathrm{Re}\{rh^{22}\}$ as gravitational-wave data;
$w$ is now independent of $m_2$.
Using the procedure summarized in Sec.~\ref{sec:model_discovery},
we recover the governing equations by optimizing for $\mathcal{F}_1$ and $\mathcal{F}_2$.
In this setting, both abstract functions only depend on $\cos\chi$.
We exploit this periodic structure by defining $\mathcal{F}_1$ and $\mathcal{F}_2$ with cosine activation functions, $\sigma_j = \cos$.
We then construct $\mathcal{F}_1$ and $\mathcal{F}_2$ as feed-forward neural networks with two hidden layers each; see, e.g.,~Eq.~\cref{eq:TwoHiddenLayers}.
The exact network architecture and numerical discretization can be found in the file \texttt{EMR.jl} in \cite{JuliaCode}.
Finally, we learn the corresponding neural network weights and biases by optimizing~\cref{eq:InverseProblem}.
This process delivers the red trajectory and waveform presented in~\Cref{fig:EMRIWaveform}.
Not only do both waveforms and trajectories match over the training interval consisting of about about 6 orbits, they continue to agree when the learned dynamics are extrapolated to about 31 orbits, after which the orbit's perihelion precession
has undergone a full cycle.
In \Cref{fig:EMRIWaveform} we compare the true waveforms/trajectories to the learned waveforms/trajectories over the extended time interval $[0, 3\cdot 10^5]$.
To compare the learned model~\cref{eq:UDEModel} to the exact model~\cref{eq:EMRIModel}, we also compute the error in $\dot{\phi}$ and $\dot{\chi}$ over the extended time interval.
Evidently, not only do the waveforms and trajectories match upon visual inspection, the learned model matches the
true mechanical model to about two orders of magnitude. The learned model also recovers important
relativistic effects, notably perihelion precession, from just a few gravitational-wave cycles. Finally, we note
that once the dynamical model is known, it can be used to generate very long orbits and gravitational wave
signals by integrating the ODE~\cref{eq:UDEModel} and post-processing the solution with the quadrupole formula~\cref{eq:gw_22_v2}.
This experiment demonstrates the potential power of waveform inversion in a simple setting with a known
solution.
It also demonstrates how the information content in the original waveform can be used to infer UDEs.
Nevertheless, this system is conservative ($\dot{e}=\dot{p}=0$), the quadrupole formula is prescribed exactly, and the learnable dynamics depend only on the $\chi$-variable. The following examples are more challenging because none of the aforementioned simplifications hold.
\begin{figure*}
\centering
\begin{minipage}{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{NR1_trajectories_overlap.pdf}
\end{minipage}
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[height=4.5cm]{NR1_pe.pdf}\\[4pt]
\includegraphics[height=4.5cm]{NR1_trajectory_error.pdf}
\end{minipage}
\includegraphics[width=0.95\linewidth]{NR1_waveforms.pdf}
\caption{\label{fig:NR1Trajectories} Summary of our second experiment, where have used gravitational-wave observations
(black dots; bottom figure) to learn the underlying dynamical system model governing the relativistic two-body problem for two equal mass black holes in quasi-circular orbit.
Top left: Trajectories of the centers of mass of the black hole system \texttt{SXS:BBH:0217} (black lines), taken from the SXS Gravitational Waveform Database \cite{Boyle:2019kee}. We also show the orbit computed from our learned dynamical system (blue and red lines). In the upper right panel,
we show the evolution of the eccentricity and semi-latus rectum from our learned-dynamical system. The middle right panel shows the disagreement between the NR trajectories and the ones computed from the learned dynamical system. We caution the reader that this figure should not be understood as a relative error because the numerical relativity black hole trajectories and our learned model are expressed in different coordinate systems that are impossible to relate.
Bottom: Learned (red line) and computational relativity (black dots) waveform data corresponding to the real part of the $h^{22}$ mode.
}
\end{figure*}
\subsection{General relativistic orbital dynamics of binary black holes} \label{sub:radiation_reaction_in_numerical_relativity_experiments}
In this pair of examples, we consider numerically generated waveform measurements from equal mass $m_1 = m_2 = 0.5$
binary black hole systems. Unlike the previous experiment, the orbital dynamics for these systems
is much more complicated; the exact equations of motion are unknown and the dynamics include time-dependent
values of the eccentricity and semi-latus rectum.
Although an extensive body of literature exists for deriving these equations
from approximations of general
relativity~\cite{blanchet2014gravitational,buonanno1999effective,poisson2011motion}, we are unaware of any data-driven approaches focused on discovering orbital dynamics from waveform measurements.
Computational relativity codes provide
exact (up to numerical discretization error) solutions to
the general relativistic two-body problem, including both the
corresponding trajectories of the center of the black holes and
gravitational-wave data.
Although the location of the black holes
are
coordinate-dependent they can still be used to compare with the trajectories obtained from our model.
However, such comparisons should no longer be understood as model error since the coordinate system
used for the data and model are necessarily different.
The simulations for this work
were performed using the Spectral Einstein
Code (SpEC)~\cite{SpECwebsite,York:1998hy, Pfeiffer:2002iy, Lindblom:2005qh,Rinne:2008vn,Boyle:2019kee,SXSCatalog}
developed by the Simulating eXterme Spacetimes (SXS)
collaboration~\cite{SpECwebsite} and made publicly available through the
Gravitational Waveform Database \cite{Boyle:2019kee}.
From now on, we augment the loss function $\mathcal{J}$ in~\cref{eq:Loss} with non-negative penalty and regularization terms, motivated below:
\begin{equation}
\label{eq:NewJ}
\mathcal{J}(\mathbf{x},\bm{\xi})
=
\langle J(\mathbf{x},\cdot) \rangle
+
\mathcal{P}_1(\mathbf{x})
+
\mathcal{P}_2(\mathbf{x})
+
\mathcal{R}(\bm{\xi})
.
\end{equation}
In this new expression, we define
\begin{equation}
\label{eq:Penalty}
\mathcal{P}_1(\mathbf{x})
=
\gamma_1\langle (\dot{p})_+^2 \rangle
+
\gamma_2\langle (\ddot{p})_+^2 \rangle
,
\end{equation}
where $(f(t))_+ = \max\{f(t),0\}$,
\begin{equation}
\label{eq:Penalty2}
\mathcal{P}_2(\mathbf{x})
=
\gamma_3\langle (-e)_+^2 \rangle
+
\gamma_4\langle (e-e_0)_+^2 \cdot \bm{1}_{\{p>6+2e_0\}} \rangle
,
\end{equation}
where $\bm{1}_{\Omega}$ denotes the indicator function on the set $\Omega \subset [0,T]$, and finally
\begin{equation}
\label{eq:Regularization}
\mathcal{R}(\bm{\xi})
=
\gamma_5\|\bm{\xi}\|^2
,
\end{equation}
where $\|\bm{\xi}\|$ denotes the $\ell^2$-norm of the expanded parameter vector $\bm{\xi}$.
It is standard practice to use large coefficients for penalty terms and small coefficients for regularization terms.
However, as explained in the paragraphs below, our penalty terms are present to help avoid nonphysical local minima and are not active in the optimized model. For this reason, we do not make a concerted attempt to tune these coefficients.
In both of the coming experiments, we somewhat arbitrarily fix $\gamma_1 = 10^3,~ \gamma_2 = 10^2,~ \gamma_3 = 10^1$, and $\gamma_5 = 10^{-1}$.
In the first experiment, we take $\gamma_4 = 1$, while in the second experiment we use $\gamma_4 = 0$.
The physical motivation for the terms in~\cref{eq:Penalty} relies on Eqs.~\cref{eq:Distance}.
From these equations, we have that the distance between the two black holes $r$ is proportional to $p$.
Due to energy loss from the emitted gravitational waves, $r(t)$ converges to zero at a rate
that increases throughout the system's evolution.
The penalty terms $\langle (\dot{p})_+^2 \rangle$ and $\langle (\ddot{p})_+^2 \rangle$ have been chosen to encourage the selection of solutions with this physical behavior.
The first term in Eq.~\cref{eq:Penalty2} encourages the selection of a
positive eccentricity function $e(t)$ for all time $t$.
On the other hand, the final term in this definition is motivated by the stability condition $p \geq 6+2e$ for bound orbits.
It is widely accepted that $e$ decays in this range \cite{cutler1994gravitational}, and this term helps to direct the solution toward models with this property.
Clearly, if $\dot{p},\ddot{p},-e,(e-e_0)^2\cdot \bm{1}_{\{p>6+2e_0\}} \leq 0$, then $\mathcal{P}_1(\mathbf{x}) = \mathcal{P}_2(\mathbf{x}) = 0$.
Our experiments appear to indicate that optimal solutions $p(t)$ and $e(t)$ satisfy each of these bounds, therefore, the penalty terms only act as guardrails throughout the optimization process.
The Tikhonov--Phillips regularization term~\cref{eq:Regularization} helps convergence by ensuring continuous dependence between the data and the solution \cite{benning_burger_2018}.
The Tikhonov regularization term $\|\bm{\xi}\|^2$ can also help avoid model degeneracies and overfitting in the presence of noisy data.
For example, when the orbit is circular ($e=0$)
the model is degenerate in $\chi$ and, therefore, it is also degenerated in the weights and biases defining $\mathcal{F}_2$.
Other penalty and regularization terms could be considered in future studies.
In the following pair of examples, we construct feed-forward neural network parameterizations of $\mathcal{F}_j$, $j=1,\ldots,4$, with $\tanh$ activation functions.
The exact network architecture we use can be found in files \texttt{SXS1.jl} and \texttt{SXS2.jl} \cite{JuliaCode}.
\subsubsection{Near-circular orbits from clean GW observations} \label{ssub:low_eccentricity_orbit}
For this experiment, we consider a binary black hole system with negligible eccentricity during the initial inspiral.
For inspection, the center of mass-corrected trajectories of the binary black hole system are depicted in the top left-hand corner of~\Cref{fig:NR1Trajectories} (solid black lines), with the associated 1000 equally-spaced waveform data points
in the bottom panel (black dots).
From now on, we let $[0,T]$ denote the time interval between the first ($t=0$) and final ($t=T$) measurement, where the final measurement occurs shortly \emph{before} merger.
As in the previous experiment, we adopt the initial conditions $\phi_0 = 0$ and $\chi_0 = \pi$
and assume that $r_0$ is known.
Using Eq.~\eqref{eq:r_parameterization}, these assumptions provide us with an explicit expression for $p_0$, $M p_0 = r_0\cdot(1+e_0\cos(\chi_0))$.
Due to the nearly-zero eccentricity of the initial trajectories, we opt for the simple initial condition $e_0 = 0$.
In the next and final experiment, we treat the more realistic case where both $e_0$ and $\chi_0$ are unknown.
In order to avoid local minima, we solve~\cref{eq:InverseProblem} on a sequence of increasing time intervals $[0,T_0] \subsetneq [0,T_1] \subsetneq \cdots \subsetneq [0,T]$, using the optimal parameters $\bm{\xi}^\star$ from each preceding optimization problem (plus a small amount of Gaussian noise) as initial data for the subsequent problem.
Using this incremental procedure, we are able to recover the overwhelming majority of the black hole trajectories, as indicated by the visual agreement between the learned (red and blue) and NR (black) trajectories shown in the top left-hand panel of~\Cref{fig:NR1Trajectories}. We also depict the relative disagreement between the NR trajectory of the first black hole $\hat{\mathbf{r}}_1$ and the learned trajectory ${\mathbf{r}}_1$. The model also recovers important general
relativistic effects, notably the learned functions $\mathcal{F}_3$ and $\mathcal{F}_4$ cause a
runaway inspiral process that drives the black holes to merge.
This process is seen most clearly by monitoring the behavior of $p(t)$
in the upper right-hand panel of~\Cref{fig:NR1Trajectories}.
We also observe that our model is able to naturally include both the inspiral ($p > 6$)
and plunge ($p< 6$) orbital regimes. Note the upper right-hand panel of~\Cref{fig:NR1Trajectories} shows
that near this transition region the eccentricity quickly grows, which is at odds with
our physical expectation for stable orbits and indicates a very different dynamical regime.
One complication in validating our learned dynamical system is how to perform meaningful comparisons with other models. Indeed, besides the waveform, the other three sub-panels shown in Fig.~\ref{fig:NR1Trajectories} depict gauge-dependent quantities; that is their value depends on the coordinate system being used. In particular, our trajectories are not expressed in
the same damped harmonic gauge coordinates used by SpEC simulations~\cite{szilagyi2009simulations}. Nevertheless, recent studies have noted surprisingly good agreement between NR trajectories and those computed with post-Newtonian (PN) models~\cite{boyle2014gravitational} and PN-augmented dynamical models~\cite{varma2019binary,blackman2017surrogate,blackman2017numerical}. In particular, Ref.~\cite{varma2019binary} conjectures that the main source of disagreement is due to the PN formula being expressed in the harmonic gauge~\cite{blanchet2014gravitational}. The close agreement between NR and UDE trajectories shown in Fig.~\ref{fig:NR1Trajectories} is another example of surprisingly good agreement~\cite{varma2019binary,boyle2014gravitational}.
To avoid gauge ambiguities, comparisons of BBH dynamics focus on comparing gauge-invariants that are computable within different frameworks. For example, Refs.~\cite{damour2012energy,nagar2016energetics} explore the conservative dynamics by comparing the relationship between the total energy and total angular momentum from NR data to the corresponding analytical predictions from PN and EOB theory. However, the identification of a conserved energy or angular momentum within our setup is not obvious as our equations are not derived from a Hamiltonian; we will return this issue in the Discussion section. Instead, we follow Ref.~\cite{van2020intermediate} and compute the accumulated orbital phase as a function of the orbital frequency. This quantity includes both dissipative and conservative effects, and can be computed within different modeling frameworks.
The NR orbital phase is defined in terms of the waveform data as follows:
\begin{equation}
\label{eq:nr_orbital_phase}
\phi_{\rm NR}^{\rm orb}(t) =\frac{1}{2} \arg h_{22}(t) \,.
\end{equation}
We also compute the orbital phase from a recently developed precessing
EOB model ({\tt SEOBNRv4PHM})~\cite{ossokine2020multipolar}, a numerical relativity surrogate model ({\tt NRSur7dq4})~\cite{varma2019surrogate}, and our UDE model. Both {\tt NRSur7dq4} and {\tt SEOBNRv4PHM} are considered state-of-the-art and have been used by the LIGO-Virgo Collaboration to analyze recent gravitational-wave observations~\cite{abbott2020gw190521,abbott2020gw190412,abbott2020gwtc}. We represent each orbital phase by a degree 3 spline (using a smoothing factor of 0.0002), from which we compute the orbital frequency,
\begin{equation}
\label{eq:orbital_frequency}
\Omega = \frac{d \phi^{\rm orb}}{dt} \,,
\end{equation}
by taking a derivative of the spline, and finally forming the function $\phi^{\rm orb}(\Omega) = \phi^{\rm orb}(t(\Omega))$. We then compare $\phi_{\rm NRSur7dq4}^{\rm orb}$, $\phi_{\rm SEOBNRv4P}^{\rm orb}$, and $\phi_{\rm UDE}^{\rm orb}$ to $\phi_{\rm NR}^{\rm orb}$. For each comparison, we form the difference, $\Delta \phi = \phi_{\rm model}^{\rm orb} - \phi_{\rm NR}^{\rm orb}$ after phase alignment\footnote {We allow for $\phi_{\rm model} \rightarrow \phi_{\rm model} + c$, for some constant $c$, which is equivalent to a rotation in the orbital plane.}. Figure \ref{fig:phase_comparison} shows $\Delta \phi$ for each model. Over the range of orbital frequencies shown, the $L_2$-error in the phase is $2.5\times 10^{-2}$ (UDE model), $2.0\times 10^{-2}$ ({\tt NRSur7dq4}), and $2.2 \times 10^{-2}$ ({\tt SEOBNRv4PHM}). All three models do an excellent job at tracking the NR orbital phase throughout the late inspiral phase to $\Omega \approx 0.16$, which is when a common apparent horizon appears in the NR simulation.
One of the most important practical uses of a dynamical model is as an intermediate step towards generating gravitational waveforms. While a full study is outside the scope of this paper, we provide a preliminary look at this here. We compute an $L_2$-type error measurement (see Eq.~21 from Ref.~\cite{blackman2017surrogate}) between the complexified NR waveform and each model's prediction of the waveform after optimizing for phase and time alignments. We find the errors to be $3.1 \times 10^{-3}$ (UDE model), $1.2 \times 10^{-3}$ ({\tt SEOBNRv4PHM}), and $1.1 \times 10^{-5}$ ({\tt NRSur7dq4}). All three models have been calibrated to $q=1$ NR waveform data, so this comparison is only meant to be suggestive of how well the modeling techniques can perform on the training set and not its generalization error.
We note that all three models used in our comparisons have been built in very different ways. The {\tt SEOBNRv4PHM} model is a highly sophisticated analytical model thats been under investigation for two decades~\cite{buonanno1999effective} while the {\tt NRSur7dq4} model was trained against 1528 NR simulations using numerical techniques that have been in development for nearly a decade~\cite{field2011reduced,field2014fast}. By comparison, our UDE model is new and our modelization choices are simple. Given that our UDE model is able to perform comparably well against state-of-the-art models demonstrates the potential of waveform inversion as new tool for model builders to consider in future work.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{orbital_phase_error.pdf}
\caption{\label{fig:phase_comparison} Comparison of the accumulated orbital phase as a function of orbital frequency for the second experiment~\ref{ssub:low_eccentricity_orbit}, an equal mass quasi-circular inspiral of nonspinning black holes. The orbital phasing obtained from an NR simulation is shown in the inset figure. We show the difference between the NR data and (i) our UDE orbital model (solid blue line), (ii) {\tt NRSur7dq4}, a numerical relativity surrogate model (dashed red line), and (iii) {\tt SEOBNRv4PHM}, a recently developed precessing EOB model (dash-dot black line). All three models show good agreement with the NR orbital phasing. The UDE (learned ODE) model is competitive with these two state-of-the-art models despite being built in a very different way.}
\end{figure}
\subsubsection{Eccentric orbits from noisy GW measurements} \label{ssub:low_eccentricity_orbit}
This experiment proceeds in much the same way as the previous one.
Here, however, we learn the dynamics of an eccentric binary black hole
system whose trajectories are depicted in the top left-hand corner of~\Cref{fig:NR2Trajectories} (solid black lines) with the associated waveform data in the bottom panel (black dots).
Unlike the previous experiments, we do not assume known values for the initial conditions $e_0,p_0,$ or $\chi_0$ but instead
make these part of the learning process. We continue to adopt the initial conditions $\phi_0 = 0$.
As can be seen in \Cref{fig:NR2Trajectories}, we introduce additive Gaussian noise to the waveform data of the form
$w(t_i) + n(t_i)$, where $n(t_i)$ is draw from a normal distribution of mean 0 and standard deviation of $\sigma = 10^{-2}$.
As the typical waveform amplitude is $\sim 0.1$, this corresponds to a coefficient of variation of around $\sigma/0.1 = 0.1$.
\begin{figure*}
\centering
\begin{minipage}{0.45\linewidth}
\centering
\includegraphics[width=\linewidth]{NR2_trajectories_overlap.pdf}
\end{minipage}
\begin{minipage}[t]{0.5\linewidth}
\includegraphics[height=4.5cm]{NR2_pe.pdf}\\[11pt]
\includegraphics[height=4.5cm]{NR2_trajectory_error.pdf}
\end{minipage}
\includegraphics[width=0.95\linewidth]{NR2_waveforms.pdf}
\caption{\label{fig:NR2Trajectories}
Summary of our third experiment, where have used noisy gravitational-wave observations
(black dots; bottom figure) to learn the underlying dynamical system model governing the relativistic two-body problem for two equal mass black holes in an eccentric orbit.
Top left: Trajectories of the centers of mass of a black hole system \texttt{SXS:BBH:1356} (black lines), taken from the SXS Gravitational Waveform Database \cite{Boyle:2019kee}. We also show the orbit computed from our learned dynamical system (blue and red lines). In the upper right panel,
we show the evolution of the eccentricity and semi-latus rectum from our learned-dynamical system. The middle right panel shows the disagreement between the NR trajectories and the ones computed from the learned dynamical system. We caution the reader that this figure should not be understood as a relative error because the numerical relativity black hole trajectories and our learned model are expressed in different coordinate systems that are impossible to relate.
Bottom: Learned (red line) and computational relativity (black dots) waveform data corresponding to the real part of gravitational waveforms.
Here, we also include the learned imaginary part of the waveform (blue line), reconstructed with the quadrupole formula, and compare it with the reference imaginary part taken from the SXS database (black line).
}
\end{figure*}
In spite of these new challenges, we are still able to recover
the original trajectories
as indicated by the visual agreement between the learned (red and blue) and NR (black) trajectories shown in the top left-hand panel of~\Cref{fig:NR2Trajectories}.
This is achieved by simultaneously optimizing for both $e_0$ or $\chi_0$ in~\cref{eq:InverseProblem}, in addition to the neural network parameters $\bm{\xi}$, and deducing the associated value of $p_0$ directly from Eq.~\cref{eq:r_parameterization}.
The model also recovers important general
relativistic effects, notably, as in the previous example,
radiation-reaction effects. In this case,
the tendency for the orbit to circularize, $e(t) \rightarrow 0$, is seen
in the upper right-hand panel of~\Cref{fig:NR2Trajectories}. As before, the
eccentricity has an inflection point around $p \approx 6$ and quickly grows thereafter.
\begin{figure}
\begin{center}
\includegraphics[width=\linewidth]{NR2_SNR_vs_RMSE.pdf}
\end{center}
\caption{\label{fig:vsSNR}
We quantify the method's robustness to increasingly amounts of Gaussian
noise by comparing the true NR waveform, $h_{\rm NR}^{22}$, to the learned waveform, $h^{22}$. The amount of noise is quantified by the signal's SNR. We compute the waveform error according to $\frac{1}{2}\|h_{\rm NR}^{22} - h^{22}\|_2^2\, /\, \|h_{\rm NR}^{22}\|_2^2$, which is related to the overlap error (cf. Ref.~\cite{blackman2017surrogate}) commonly used in gravitational-wave data analysis.
}
\end{figure}
As follow-on to this experiment, we test the stability of the learned solution to different signal-to-noise ratios (SNRs).
Specifically, we use the learned solution parameters $\bm{\xi}$, $e_0$, and $\chi_0$ from~\Cref{fig:NR2Trajectories} as initial guesses in a new set of model discovery problems that fit the original \texttt{SXS:BBH:1356} waveform data, but have white Gaussian noise of different variance superimposed. After solving each of these new model discovery problems, we measure how much the learned waveform differs from the true waveform. The curve in~\Cref{fig:vsSNR} shows the relative error in the learned waveform as a function of the SNR in the waveform measurement. The relative error grows as the SNR decreases, however, the results are surprisingly accurate down to SNR$\approx 32$. We note that while all BBH gravitational-wave detections to-date have had SNRs below 32~\cite{abbott2020gwtc}, future detectors such as LISA~\cite{2017arXiv170200786A}, Cosmic Explorer~\cite{Evans:2016mbw}, or Einstein Telescope~\cite{Hild:2010id} will
routinely detect events with an SNR greater than 32.
\section{Discussion} \label{sec:discussion}
We have presented a new, data-driven
gravitational waveform inversion strategy which generates
mechanical models of binary black hole systems. We start with a structurally very simple set of universal differential equations
and parameterize the space of models with feed-forward neural networks. Our differential equations are trained
by solving a physics-informed constrained optimization problem that seeks to minimize the waveform error. We tested our method
on various BBH systems including extreme and comparable mass ratio systems in eccentric and non-eccentric orbits,
and train on portions of the waveform corresponding to orbital plunge right up to the time of merger.
We find that the resulting differential equations agree remarkably well
with the black hole trajectories computed through purely numerical means.
Our models can be extrapolated in time and recover various known relativistic effects despite these being
previously unknown to the universal differential equations. The main contribution of our work is to show that
two-body relativistic models can be deduced from gravitational wave measurements.
To describe the computational framework, we have focused on simple choices
such as our ansatz neural ODE model Eq.~\eqref{eq:UDEModel} and the quadrupole formula for computing the harmonic modes.
These modeling choices are distinct from the overall computational framework and can be easily swapped out for different choices. For example, future high-accuracy studies should seek better orbit-to-waveform maps as the
quadrupole formula is likely a large source of systematic error.
Our framework for learning the dynamics of binary black holes is quite general, and we expect that it can be applied to a variety of cases we have not considered including unequal masses, aligned-spins systems, and precessing systems.
Our method should be especially useful for discovering equations of motion for systems
where traditional approaches are less well-developed including
eccentric binaries, the highly relativistic late-inspiral and plunge dynamical regimes, and beyond-GR
theories. Given the close agreement with NR trajectories, other possible applications could include setting NR initial data whereby the neural ODEs could be used to predict an NR trajectory before the simulation is performed.
One of the most important applications of our technique may involve calibrating existing orbital dynamics models (and high-accuracy gravitational-wave models) by using a base model different from Eq.~\eqref{eq:UDEModel}. Given that all modern inspiral-merger-ringdown waveform models require calibration of unknown parameters to numerical relativity simulations, the waveform inversion technique described here could benefit these efforts. For example, if implemented within the effective-one-body (EOB) approach, a suitably modified version of our methodology could be used to calibrate for missing terms in the EOB Hamiltonian. Precessing NR surrogate models also require a dynamical model, which is found through a direct fitting for the right-hand-side of the relevant ODEs. This computationally costly step might be simplified and accelerated with our techniques.
Another potential use of our methodology is training dynamical models entirely from gravitational-wave datasets instead of solving or analyzing Einstein's equation of general relativity as is traditionally done. In Fig.~\ref{fig:vsSNR} we explore how well the algorithm performs as the signal's SNR is systematically varied. We find that, at least for the examples considered here, the method continues to work down to an SNR of about 32. Consequently, our method is most applicable for future GW detections, including those made with LISA~\cite{2017arXiv170200786A}, Cosmic Explorer~\cite{Evans:2016mbw}, Einstein Telescope~\cite{Hild:2010id}, or perhaps the upcoming LIGO-Virgo-Kagra science run. Our method will need to be modified to achieve SNR levels of approximately 10 (which would cover most GW detections to date~\cite{abbott2020gwtc}) without compromising accuracy. Some approaches could include: (i) comparing waveforms with the Wasserstein metric which is known to be more robust to phase trappings, which shows up in higher noise, (ii) detecting an ensemble of noisy signals, and training on this observation set, (iii) using Bayesian networks where the network parameters are probability functions, or (iv) to apply our technique to filtered waveforms using a model-agnostic approach such as wavelet methods or denoising methods. We leave such extensions for future work.
\bigskip
\begin{acknowledgments}
We thank Alvin Chua, Sebastian Khan, Gaurav Khanna, Tom O'Leary-Roseberry, and Harald Pfeiffer for helpful discussions.
S.E.F.\ is supported by NSF grants No. PHY-1806665, PHY-2110496, and No. DMS-1912716.
This manuscript was written while the authors were in residence at the Institute for Computational and Experimental Research in Mathematics (ICERM) in Providence, RI, during the Advances in Computational Relativity program, supported by the National Science Foundation under Grant No. DMS-1439786.
This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344, LLNL-JRNL-819108.
\end{acknowledgments}
|
\section{Introduction}
\justify
Acoustically excited bubbles are known to be nonlinear oscillators \cite{A,B} with "infinite complexity" \cite{A}. The addition of a lipid stabilizing shell increases this complexity, thus, even after over a decade of study, the dynamics of ultrasonically excited lipid coated microbubbles (MBs) are not fully understood. Interestingly, lipid coated MBs have been shown to exhibit 1/2 order subharmonic (SH) oscillations even when the excitation amplitude is low ($<$30 kPa \cite{1,2,3}). Despite the increased inherent damping due to the coating, such low threshold values contradict the predictions of the theoretical models as these values are even below the thresholds expected for uncoated free MBs \cite{4,5}.\\
The lipid coating may cause compression dominated oscillations \cite{6} or limit the MB oscillations to only above a certain pressure threshold \cite{7}. It has been shown in \cite{1} that the low pressure threshold for 1/2 order SH emissions is due to the compression only behavior of the MBs due to the buckling of the shell. Overvelde et al.\cite{8} showed that the lipid coating may enhance the nonlinear MB response at acoustic pressures \textit{as low as 10 kPa}. In addition, even a small ($\approx10 kPa$) increase in the acoustic pressure amplitude leads to a significant decrease in the main resonance frequency \cite{8} resulting to a pronounced skewness of the resonance curve. The origin of the “thresholding” \cite{7} behavior has been linked to the shift in resonance \cite{8}. Nonlinear resonance behavior of the lipid coating has also been observed at higher frequencies (5-15 MHz\cite{9}), (8-12 MHz \cite{10}) and (11-25 MHz \cite{11}). Through theoretical analysis of the Marmottant model for lipid coated MBs \cite{12}, Prosperetti \cite{4} attributed the lower SH threshold of the lipid MBs to the variation in the mechanical properties of the coating in the neighborhood of a certain MB radius (e.g. the occurrence of buckling). The coating may also result in expansion dominated behavior in liposome-loaded MBs \cite{13}. Expansion dominated oscillations occur when the initial surface tension of the lipid coated MB is close to that of the water \cite{11,13}. In this regime, and in contrast to compression dominated behavior, the MB expands more than it compresses. Expansion-dominated behavior was used to explain the enhanced non-linearity at higher frequencies (25 MHz) \cite{11}. The Marmottant model effectively captures the behavior of the MB including expansion-dominated behavior \cite{11,12,14}, compression only behavior \cite{6}, thresholding \cite{7} and enhanced non-linear oscillations at low excitation pressures \cite{1,2,8,14,15,16}. Despite the numerous studies which employed the methods of nonlinear dynamics and chaos to investigate the dynamics of acoustically excited MBs \cite{A,B,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44}, the detailed bifurcation structure and nonlinear dynamics of the lipid coated MBs have not been studied. \\ Lipid coated MBs are being routinely used in diagnostic ultrasound \cite{2,45,46,47,48}. Moreover, they have been used in pioneering non-invasive treatments of brain disorders and tumors in humans \cite{49}. Currently there are several investigations on the potential use of lipid coated micro-bubbles (MBs) in high resolution and high contrast imaging procedures \cite{50} as well as non-invasive ultrasonic treatments and localized drug/gene delivery \cite{51}. Despite the promising results of these investigations, the complex dynamics of the system makes it difficult to optimize the behavior of lipid coated MBs. Moreover, from the nonlinear dynamics point of view, the lipid coated MB oscillator is an interesting topic of investigation due to the highly nonlinear nature of the system. The complex behavior of the uncoated MB is interwoven with the nonlinear behavior of the lipid coating which enables unique dynamical properties for this oscillator. \\
In this work we study the bifurcation structure of the lipid coated MBs as a function of size and frequency at different pressure values. Numerical results are accompanied by experimental observations of single MB scattering events at low pressure excitation. We show for the first time, both experimentally and numerically that in addition to 1/2 order SHs, higher order SHs (e.g. 1/3, 1/4) can be generated at low excitation amplitudes. Analysis of the bifurcation structure of the system reveals a unique property of the lipid coated MB which is the generation of chaos at excitation pressures as low as 5kPa.
\section{Methods}
\subsection{Experimental method}
Very dilute solutions of Definity$^{\textregistered}$ MBs were sonicated with 25 MHz continuous pulse trains using the Vevo 770 ultrasound imaging machine (VisualSonics Inc. Toronto, Ontario). The pulse length was held constant at 30 cycles while the applied acoustic pressure was varied over the range of $\approx 70-300 kPa$. The backscattered signals from individual MBs were extracted and different nonlinear modes of oscillations were identified. Acquisition of signals from single MB were similar to the approach in \cite{52}. Fig. 1 shows a schematic of an acquired signal from a single MB event ($\approx250 kPa$ and $f=25 MHz$).
\begin{figure*}
\begin{center}
\includegraphics[scale=3]{770-eps-converted-to.pdf}\\
(a)
\includegraphics[scale=0.5]{exp.png}\\
(b)
\caption{a) Schematic of the Vevo 770 (Visualsonics$^{\textregistered}$) machine which was used in the experiments to detect the signals from single MB events in the region of interest (ROI). b) Left: ROI selected for an ultrasound pulse train at 25 MHz and 250 kPa of pressure (each large subdivision is $\approx$ 130 $\mu m$), and Right: Signal (red) from a single period-3 MB event. The frequency spectrum of the received signal is shown in blue exhibiting 1/3 order SHs at 8.33 and 2/3 order SHs 16.66 MHz.}
\end{center}
\end{figure*}
\subsection{Numerical procedure}
\subsubsection{Marmottant Model}
Dynamics of coated MBs undergoing buckling and rupture can be effectively modeled using the Marmottont equation \cite{12}:
\begin{linenomath*}
\begin{equation}
\begin{gathered}
\rho \left(R\ddot{R}+\frac{3}{2}\dot{R}^2\right)=\\\left[P_0+\frac{2\sigma(R_0)}{R_0}\right](\frac{R}{R_0})^{-3k}\left(1-\frac{3k}{c}\dot{R}\right)- P_0-\frac{2\sigma(R)}{R}-\frac{4\mu\dot{R}}{R}-\frac{4k_s\dot{R}}{R^2}-P_a(t)
\end{gathered}
\end{equation}
\end{linenomath*}
In this equation, R is radius at time t, $R_0$ is the initial MB radius, $\dot{R}$ is the wall velocity of the bubble, $\ddot{R}$ is the wall acceleration, $\rho{}$ is the liquid density (998 $\frac{kg}{m^3}$), c is the sound speed (1481 m/s), $P_0$ is the atmospheric pressure, $\sigma(R)$ is the surface tension at radius R, $\mu{}$ is the liquid viscosity (0.001 Pa.s), $k_s$ is the coating viscosity and $P_a(t)$ is the amplitude of the acoustic excitation ($P_a(t)=P_a sin(2\pi f t)$) where $P_a$ and \textit{f} are the amplitude and frequency of the applied acoustic pressure. The numerical values in the parentheses are for pure water at 293 K. The gas inside the MB is C3F8 and water is the host medium.\\
The surface tension $\sigma(R)$ is a function of radius and is given by:
\begin{linenomath*}
\begin{equation}
\sigma(R)=
\begin{dcases}
0 \hspace{3cm} if \hspace{1cm} R \leq R_b\\
\chi(\frac{R^2}{R_b^2}-1) \hspace{1.2cm} if\hspace{1cm} R_b \leq R\leq R_{r}\\
\sigma_{water} \hspace{2cm} if\hspace{.5cm} \hspace{.5cm} R > R_r
\end{dcases}
\end{equation}
\end{linenomath*}
$\sigma_{water}$ is the water surface tension (0.072 N/m), $R_b=\frac{R_0}{\sqrt{1+\frac{\sigma_0}{\chi}}}$ is the buckling radius, $R_r=R_b\sqrt{1+\frac{\sigma_r}{\chi}}$ is the rupture radius, $\chi$ is the shell elasticity, $\sigma_0$ is the initial surface tension at $R=R_0$, and $\sigma_r$ is the rupture surface tension. Similar to \cite{11} in this paper $\sigma_r=\sigma_{water}$. Shear thinning of the coating is included in the Marmottant model using \cite{53}:
\begin{linenomath*}
\begin{equation}
k_s=\frac{4k_0}{1+\alpha \frac{|\dot{R}|}{R}};
\end{equation}
\end{linenomath*}
where $k_0$ is the shell viscous parameter and $\alpha$ is the characteristic time constant associated with the shear rate. In this work shell parameters of $\chi=0.975 N/m$, $k_0=2.98\times 10^{-10} kg s^{-1}$ and $\alpha=1\times 10^{-6} s$ are used for the $Definity^{\textregistered}$ MBs. These values are adopted from the estimated parameters in \cite{53,54,Carly, Qian}.\\ Due to the negligible thermal damping of C3F8 even at high amplitude oscillations \cite{US2,US3} thermal damping is neglected in this paper.
\subsubsection{Keller-Miksis model}
The dynamics of the uncoated MBs were also visualized alongside the lipid coated MBs to highlight the contributions of the coating to the nonlinear bheavior of the bubble. To model the uncoated MB dynamics the Keller-Miksis model \cite{55} is used:
\begin{linenomath*}
\begin{equation}
\rho[(1-\frac{\dot{R}}{c})R\ddot{R}+\frac{3}{2}\dot{R}(1-\frac{R}{3c})]=(1+\frac{\dot{R}}{c})(G)+\frac{R}{c}\frac{d}{dt}(G)
\label{eq:5}
\end{equation}
\end{linenomath*}
where $G=\left( P_0+\frac{2\sigma_{water}}{R_0} \right)(\frac{R}{R_0})^{-3k}-\frac{4\mu_L\dot{R}}{R}-\frac{2\sigma}{R}-P_0-P_A sin(2 \pi f t)$.\\
\subsubsection{Scattered pressure by MBs}
The pressure scattered (re-radiated) by the oscillating MB can be calculated using \cite{56,57}:
\begin{linenomath*}
\begin{equation}
P_{sc}=\rho\frac{R}{r}(R\ddot{R}+2\dot{R}^2)
\end{equation}
\end{linenomath*}
here $r$ is the distance from the MB center. The scattered pressure ($P_{sc}$) at 15 cm (approximate path length of the MBs in experiments) is calculated for 30 cycle pulses to match experimental conditions. The calculated $P_{sc}$ is then convolved with the one way transducer response accounting for attenuation effects in water ($0.000221 \frac{dB}{mmMHz^2}$ \cite{58}). Moreover, to better compare with experimental observations, the sampling frequency for simulations is chosen to be equal to the sampling frequency in experiments ($460 MHz$).
\subsubsection{Investigation tools}
The results of the numerical simulations were visualized using a comprehensive bifurcation analysis method \cite{43}. In this method the bifurcation structure of the normalized MB oscillations ($\frac{R}{R_0}$) are plotted in tandem versus a control parameter using two different bifurcation methods (Poincaré section at each driving period and method of peaks). The bifurcation diagrams of the normalized MB oscillations ($R/R_0$) are calculated using both methods . When the two results are plotted alongside each other, it is easier to uncover more important details about the superharmonic (SuH) and ultraharmonic (UH) oscillations, as well as the SH and chaotic oscillations. This gives insight into the nonlinear behavior over a wide range of parameters, and enables the detection of SuH and UH oscillations alongside SH and chaotic oscillations \cite{43,POF,US}. This approach reveals the intricate details of the oscillations. In this paper the bifurcation diagrams of the normalized radial oscillations of the $Definity^{\textregistered}$ MBs were plotted versus the MB initial diameter for fixed frequencies of 25 MHz and for the range of the pressure values studied in the experiments.
\section{Results}
\subsection{Bifurcation structure}
In \ref{subsection:AA} we present the bifurcation structure of an uncoated C3F8 MB and a Definity MB with $2\mu m$ initial size as a function of frequency at $P_a=5 kPa$ for different values of initial surface tension $\sigma_0$. We show that the lipid buckling or rupture of the lipid shell enhances the unexpected generation of the nonlinear behavior. Despite the reduced damping, the unocated MB of the same size does not exhibit nonlinear oscillations. In \ref{subsection:BB} we study the bifurcation structure of the uncoated and Definity MBs as a function of the initial size at different $\sigma_0$. The sonication frequency is fixed at 25 MHz (freqeuncy of the experiments) and $P_a=250 kPa$. We show that depending on the bubble size and $\sigma_0$, different nonlinear regimes can be enhanced.
\subsubsection{Bifurcation of $\frac{R}{R_0}$ as a function of frequency for uncoated and coated MBs with different $\sigma(R_0)$} \label{subsection:AA}
\begin{figure}
\begin{tabular}{c c}
\includegraphics[scale=0.43]{uncoatedf3-eps-converted-to.pdf} \hspace{-1cm}
\includegraphics[scale=0.36]{f-0-eps-converted-to.pdf}\\
\hspace{-7cm} (a) & \hspace{-5cm} (b)\\
\includegraphics[scale=0.36]{f-0,01-eps-converted-to.pdf} \hspace{-1cm}
\includegraphics[scale=0.36]{f-0,03-eps-converted-to.pdf}\\
\hspace{-7cm} (c) & \hspace{-5cm} (d)\\
\includegraphics[scale=0.36]{f-0,062-eps-converted-to.pdf} \hspace{-1cm}
\includegraphics[scale=0.36]{f-0,072-eps-converted-to.pdf} \\
\hspace{-7cm} (e) & \hspace{-5cm} (f)\\
\end{tabular}
\caption{The bifurcation structure (blue represent the maxima and the red represents the conventional method) of the $\frac{R}{R_0}$ of a $2 \mu m$ MB as a function of frequency for $P_a=5 kPa$ for the: a) uncoated MB and for the lipid MB with b) $\sigma(R_0)=0 N/m$, c) $\sigma(R_0)=0.01 N/m$, d) $\sigma(R_0)=0.03 N/m$, e) $\sigma(R_0)=0.062 N/m$ and f) $\sigma(R_0)=0.072 N/m$.}
\end{figure}
In order to visualize the dynamics of the MBs at low pressures, the bifurcation structure of a $2 \mu m$ MB is plotted as a function of frequency when $P_a=5 kPa$ (Fig. 2). The blue graph represents the results using the maxima method and the red graph represents the results using the conventional method (Poincaré section at each period). The uncoated MB (Fig. 2a) exhibits a P1 signal over 1-30 MHz with one maximum and resonant oscillations at $\approx 4.7 MHz$. Contrary to Fig. 2a, the lipid MB with $\sigma(R_0)=0$ exhibits significant non-linearity at 5 kPa (Fig. 2b). Higher order SuHs ($5th-2nd$ order) are seen for $f<\approx 2.88 MHz$. For example the 5th order SuH at 1MHz is a P1 signal with 5 maxima. P1 resonance occurs at $f\approxeq 4.63 MHz$ with P2 oscillations over a wide frequency range of $\approx$ $6.9MHz<f<12.22MHz$ with a small window of P4-2 and chaos. We call this P4-2 oscillations as it occurs when P2 oscillations undergo a period doubling (PD) to P4 \cite{23,44}. P3 occurs between $\approx$ $12.85MHz<f<18.2MHz$. P4-1 occurs in the frequency range between $18.2MHz<f<19.5MHz$ (highlighted in subplot within Fig. 2b). We call this P4-1 regime as it occurs when P1 oscillations undergo a saddle node bifurcation to P4 oscillations \cite{26,29,30}.\\ The lipid MB with $\sigma(R_0)=0.01N/m$ (Fig. 2c) exhibits P1 oscillations with one maximum. The pressure dependent resonance frequency ($PDf_r$ \cite{42}) occurs at $f\approx 9.15 MHz$. The MB behavior is of P1 with one maximum for all the studied frequencies.\\ The lipid MB with $\sigma(R_0)=0.03 N/m$ (Fig. 2d) exhibits P1 behavior with 1 maximum and a resonance at $\approx 10.66MHz$. The behavior of the MB with $\sigma=0.062 N/m$ (Fig. 2e) is similar to $\sigma=0.01 N/m$ with $PDf_r\approxeq9.6 MHz$.\\ The MB with $\sigma(R_0)=0.072 N/m$ exhibits a similar behavior to the MB with $\sigma(R_0)=0 N/m$ demonstrating $6th-2nd$ order SuHs between the freqeuncy range of 1-3.7 MHz. Some of the SuH between $1MHz<f<2MHz$ are highlighted in a subplot within Fig. 2f). P2 occurs between $9.3MHz<f<16.5MHz$. P3 occurs between $16.5MHz<f<19 MHz$ (highlighted as a subplot inside Fig. 2f), between $19.7<f<20.7 MHz$ and between $20.3MHz<f<21.3 MHz$. P4-1 regime occurs between $23.4MHz<f<24 MHz$ (highlighted within a subplot in Fig. 2f). The MB also demonstrates P1 resonance frequency at $\approx5.7 MHz$.\\Notably, despite the higher damping due to the coating, the coated MB undergoing shell rupture exhibits a greater oscillation amplitude (Figs. 2a and 2f).
\begin{figure}
\begin{tabular}{c c}
\includegraphics[scale=0.36]{uncoatedbubble-eps-converted-to.pdf} \hspace{-1cm}
\includegraphics[scale=0.36]{r-0-eps-converted-to.pdf}\\
\hspace{-7cm} (a) & \hspace{-5cm} (b)\\
\includegraphics[scale=0.36]{r-0,01-eps-converted-to.pdf} \hspace{-1cm}
\includegraphics[scale=0.36]{r-0,03-eps-converted-to.pdf}\\
\hspace{-7cm} (c) & \hspace{-5cm} (d)\\
\includegraphics[scale=0.36]{r-0,062-eps-converted-to.pdf} \hspace{-1cm}
\includegraphics[scale=0.36]{r-0,72-eps-converted-to.pdf} \\
\hspace{-7cm} (e) & \hspace{-5cm} (f)\\
\end{tabular}
\caption{The bifurcation structure of the $\frac{R}{R_0}$ (blue represent the maxima and the red represents the conventional method) as a function of size (MB diameter) at $P_a=250 kPa$ and $f=25 MHz$ for the: a) uncoated MB and for the lipid MBs with b)$\sigma(R_0)=0 N/m$, c)$\sigma(R_0)=0.01 N/m$, d)$\sigma(R_0)=0.03 N/m$, e) $\sigma(R_0)=0.062 N/m$ and f) $\sigma(R_0)=0.072 N/m$.}
\end{figure}
\subsubsection{Bifurcation $\frac{R}{R_0}$ as a function of size (initial diameter) for uncoated and coated MBs with different $\sigma(R_0)$}
\label{subsection:BB}
In order to investigate the nonlinear behavior of the commercially available $Definity^{\textregistered}$ MBs for the experimental exposure conditions, the bifurcation structure of the $\frac{R}{R_0}$ is studied as a function of MB size when $P_a=250 kPa$ and $f=25MHz$. This is because of the polydisperse nature of the $Definity^{\textregistered}$ MBs \cite{54}, and since in the experiments we are limiting our analysis to the transducer focal zone with small variations in pressure and the fixed sonication frequency. The size distribution in the simulations replicates the distribution of the native $Definity^{\textregistered}$ \cite{54} Thus, the $\frac{R}{R_0}$ plot versus MB size will provide insight relevant to the experimental conditions in this study.\\
Fig. 3a, shows the bifurcation structure of an uncoated MB as a function of size. MB with sizes between $0.27 \mu m$-$0.28 \mu m$ exhibit 2nd order SuH (P1 oscillation with 2 maxima as highlighted in a subplot) and MBs with sizes $0.54 \mu m$ are resonant. Fig. 3a shows that at $f=25 MHz$ and $P_a=250 kPa$ the uncoated MB cannot produce SHs.\\ Fig. 3b, shows the bifurcation structure of the $Definity^@$ MBs with $\sigma(R_0)=0 N/m$. In stark contrast to the uncoated MB (Fig. 3a), an abundance of nonlinear behavior is observed. This includes 4th, 3rd and 2nd order SuHs for MB sizes smaller than $0.345 \mu m$ (some are highlighted in a subplot within Fig. 3b), P1 resonance for $\approx 0.59 \mu m$ MBs , P2, P4-2 and chaotic behavior for MB sizes of $\approx$ $0.74 \mu m<2R_0<1.13\mu m$, P3 oscillations for $\approx$ $1.13 \mu m<2R_0<1.49\mu m$, and intermittent P4-1 oscillations for $\approx$ $1.49 \mu m<2R_0<1.7\mu m$ (P4-1 is highlighted in a subplot within Fig. 3b).\\ $Definity^{\textregistered}$ MBs with $\sigma(R_0)=0.01 N/m$ (Fig. 3c), exhibit enhanced nonlinear behavior including $5th-2nd$ order SuHs (highlighted in a subplot), P2, P4-2, P3 and chaos. Fig. 3d represents the behavior of MBs with $\sigma(R_0)=0.03 N/m$. 2nd order SuH (highlighted in a subplot), $PDf_r$, P2 and P3 oscillations are observed.\\ Fig. 3e-f represent the MBs with initial surface tension close to that of water and thus with a higher tendency for rupture and expansion dominated behavior. For $\sigma(R_0)=0.062 N/m$ and for MB sizes $0.2\mu m<2R_0<0.5 \mu m$, 3rd and 2nd order SuH and 5/2 UH regimes are observed. 5/2 UH is a P2 with 4 maxima and is highlighted in a subplot within Fig. 3e. $PDf_r$, P2, P4-2 and P3 (highlighted in a subplot) oscillations are observed for MB sizes $2R_0>0.5 \mu m$. When $\sigma(R_0)=0.072 N/m$ (Fig. 3f), in addition to the nonlinear behavior we observe in Fig. 3e, we observe a 4th order SuH regime (highlighted in a subplot as a P1 with 4 maxima) and P4-1 and the absence of 5/2 UHs.\\ Results indicate that the nonlinear behavior of the MBs is highly sensitive to the initial surface tension as well as the MB size. The closer the surface tension to 0 or that of water ($\sigma_{water}=0.072 N/m$), the greater is the tendency of the MB to exhibit nonlinear behavior. Notably, P4-1 oscillations \textit{were only observed} when $\sigma(R_0)=0$ and $0.072 N/m$ (Fig. 3).\\ The MBs with $\sigma(R_0)=0.062$ (Fig. 3e) and $\sigma(R_0)=0.072$ (Fig. 3f), exhibit higher oscillation amplitude compared to uncoated MBs of the same size.\\ In order to better visualize the effect of the $\sigma(R_0)$ on the MB behavior, the bifurcation structure of the $\frac{R}{R_0}$ of the MB is plotted as a function of $\sigma(R_0)$ for two different MB sizes in Fig. 4. The bifurcation structure of a MB with an initial diameter of $0.92 \mu m$ is depicted in Fig. 4a. The nonlinear behavior occurs only for the two extreme ends of the $\sigma(R_0)$. P2 occurs for $\sigma(R_0)<0.011 N/m$ and $\sigma(R_0)>0.061 N/m$ with P4-2 happening for $0.0032 N/m<\sigma(R_0)<0.048 N/m$ and $\sigma(R_0)>0.069 N/m$. For a MB with initial diameter of $1.89 \mu m$, the same general behavior is observed. For initial surface tension values between $0.0127<\sigma(R_0)<0.057 N/m$ we observe P1 behavior with 1 maximum. As we approach to the lower and higher $\sigma(R_0)$, nonlinear behavior manifests itself in the bifurcation diagrams. P4-1 oscillations occurs for $0.0035 N/m<\sigma(R_0)<0.0041 N/m$ and $\sigma(R_0)>0.07 N/m$. P3 occurs for $0.009 N/m<\sigma(R_0)<0.012 N/m$ and $0.058 N/m<\sigma(R_0)<0.069 N/m$.
\begin{figure*}
\begin{center}
\includegraphics[scale=0.4]{0,92-sr0-eps-converted-to.pdf}\includegraphics[scale=0.4]{1,89-sr0-eps-converted-to.pdf}\\
\hspace{2cm} (a) \hspace{7cm}(b)
\caption{The bifurcation structure of $\frac{R}{R_0}$ as a function of $\sigma(R_0)$ at $f=25 MHz$ and $P_a=250 kPa$ for a MB size of: a) 0.92 $\mu m$ $\&$ b) 1.89 $\mu m$ }
\end{center}
\end{figure*}
\subsubsection{Experiments}
\begin{figure*}
\begin{center}
\includegraphics[scale=0.3]{P1exp-eps-converted-to.pdf}\includegraphics[scale=0.3]{P1sim-eps-converted-to.pdf} \includegraphics[scale=0.3]{P1comp-eps-converted-to.pdf}\\
(a) \hspace{4cm}(b)\hspace{4cm}(c)\\
\includegraphics[scale=0.3]{P2exp-eps-converted-to.pdf}\includegraphics[scale=0.3]{P2sim-eps-converted-to.pdf} \includegraphics[scale=0.3]{P2comp-eps-converted-to.pdf}\\
(d) \hspace{4cm}(e)\hspace{4cm}(f)\\
\includegraphics[scale=0.3]{P3exp-eps-converted-to.pdf}\includegraphics[scale=0.3]{P3sim-eps-converted-to.pdf} \includegraphics[scale=0.3]{P3comp-eps-converted-to.pdf}\\
(g) \hspace{4cm}(h)\hspace{4cm}(i)\\
\includegraphics[scale=0.3]{P42exp-eps-converted-to.pdf}\includegraphics[scale=0.3]{P42sim-eps-converted-to.pdf} \includegraphics[scale=0.3]{P42comp-eps-converted-to.pdf}\\
(j) \hspace{4cm}(k)\hspace{4cm}(l)\\
\includegraphics[scale=0.3]{P41exp-eps-converted-to.pdf}\includegraphics[scale=0.3]{P41sim-eps-converted-to.pdf} \includegraphics[scale=0.3]{P41comp-eps-converted-to.pdf}\\
(m) \hspace{4cm}(n)\hspace{4cm}(o)\\
\caption{ Demonstration of 5 main oscillation regimes acquired experimentally (blue) and simulated (red) choosing MB sizes based on the feature similarity in Fig. 3f. Representative experimental data and simulations of: 1st row P1, 2nd row P2, 3rd row P3, 4th row P4-2 and 5th row P4-1.}
\end{center}
\end{figure*}
In experiments at $P_a\approx250 kPa$ and $f=25 MHz$ we observed 5 main types of backscattered signals in the data collected from single MB events. A representative of each category is shown in Fig. 5a (P1), Fig. 5d (P2), Fig. 5g (P3) , Fig. 5j (P4-2) and Fig. 5m (P4-1). The results of the numerical simulations are presented in the second column and the frequency spectrum of the experimental signals and the numerical simulations are plotted in the third column (blue:experiments, red:simulations). Numerical simulations are for the $Definity^{\textregistered}$ MBs with $\sigma(R_0)=0.072 N/m$ with the corresponding sizes chosen from the bifurcation diagram (Fig. 3f) to match the observed behavior in the experiments.\\
Fig. 5a displays a typical P1 signal observed in experiments. The calculated $P_{sc}$ for a 2 $\mu m$ $Definity^{\textregistered}$ MB is displayed in Fig. 5b (in red color for distinction) and the power spectrum of the signals in Fig. 5a and 5b are shown in Fig. 5c. The scattered pressure has one maximum and the frequency spectrum has a peak at 25 MHz.\\ A representative signal of the P2 oscillations is displayed in the second row of Fig. 5. Both experimental and simulated (initial size of $0.955 \mu m$) signals have two maxima revealing a P2 oscillation regime. The power spectra in Fig. 5f consist of a SH peak at 12.5MHz and a 3/2 UH peak at 32.5 MHz.\\ A representative of the P3 signal is shown in the third row of Fig. 5. The experimental and simulated (initial size of $1.39 \mu m$) signals have 3 maxima and the order of the maxima are consistent between experiments and simulations. The power spectra in Fig. 5i show a good agreement between experiments and simulations with SHs at (1/3 order) 8.33 MHz, (2/3 order) 16.66 MHz and UHs at (4/3 order) 33.33 MHz and (5/3 order) 41.66 MHz.\\
\begin{figure*}
\begin{center}
\includegraphics[scale=0.4]{P4-2pa-eps-converted-to.pdf}\\
(a)\\
\includegraphics[scale=0.4]{P4-1pa-eps-converted-to.pdf}\\
(b)\\
\caption{The bifurcation structure of the $\frac{R}{R_0}$ of the MB as a function of pressure for excitation with $f=25 MHz$ for a lipid MB with $\sigma(R_0)=0.072 N/m$ and a diameter of : a) 0.92 $\mu m$ $\&$ b) 1.89 $\mu m$}
\end{center}
\end{figure*}
P4-2 oscillations are shown in the 4th row of Fig. 5. There is a good agreement between the experimental and the simulated signals (initial size of $0.92 \mu m$). Both signals have 4 peaks in two envelopes and each envelope repeats itself once every two acoustic cycles. In each envelope there are two peaks and the peaks repeat themselves in an amplitude order of (largest, small, large, smallest). The frequency spectra of the signals are shown in Fig. 5l. There are 3 SHs at (1/4 order) 6.25 MHz, (1/2 order) 12.5 MHz and (3/4) order at 18.75 MHz. The 1/2 order SH is the strongest detected SH and due to the weakness of the 1/4 SH this peak is hardly detectable. This is because the transducer sensitivity drops sharply away from the center frequency and especially below 12.5 MHz (transducer bandwidth is 100$\%$). While the numerically simulated $P_{sc}$ in the absence of convolution with transducer response had a clear peak at 6.25 MHz, however, after the signal is convolved with the transducer response, the signal drops below the noise level of -70 dB in our experiments.\\ The last row of Fig. 5 depicts the case of the P4-1 oscillations. Simulations are for a MB with initial size of $1.89 \mu m$. The signals have one envelope with 4 maxima that repeats itself once every 4 acoustic cycles. Amplitudes repeat themselves in the order of smallest, largest, large and small. Both experimental and simulated signals demonstrate the same pattern of peaks and their orders. The power spectra in Fig. 5o shows a good agreement between the orders of the SHs and their locations. There are 3 SHs at (1/4 order) 6.25 MHz, (1/2 order) 12.5 MHz and (3/4 order) at 18.75 MHz. The 3/4 order SH is the strongest detected SH. It should be noted that 1/4 order SH is the strongest peak in the calculated $P_{sc}$ in the absence of convolution with transducer response (see Fig. 7f). Due to the reduced sensitivity of the transducer at 6.25 MHz, the detected strength of the 1/4 order SH diminishes strongly and it drops below all the other SHs. \\
\begin{figure*}
\begin{center}
\includegraphics[scale=0.3]{P4-2radial-eps-converted-to.pdf}\includegraphics[scale=0.3]{P4-2phase-eps-converted-to.pdf} \includegraphics[scale=0.3]{P4-2power-eps-converted-to.pdf}\\
(a) \hspace{4cm}(b)\hspace{4cm}(c)\\
\includegraphics[scale=0.3]{P4-1radial-eps-converted-to.pdf}\includegraphics[scale=0.3]{P4-1phase-eps-converted-to.pdf} \includegraphics[scale=0.3]{P4-1power-eps-converted-to.pdf}\\
(d) \hspace{4cm}(e)\hspace{4cm}(f)\\
\caption{ Characteristics of the two P4 oscillations identified: a) P4-2 radial oscillations, b) P4-2 phase portrait, c) power spectrum of the P4-2 $P_{sc}$, d)P4-1 radial oscillations, e)P4-2 phase portrait and f) power spectrum of P4-2 $P_{sc}$. Here $P_{sc}$ is not convolved with the transducer response. (red circles shows the location of the R every 4 acoustic cycles)}
\end{center}
\end{figure*}
\subsubsection{Difference between P4-1 and P4-2}
Fig. 6a shows the bifurcation structure of a 0.92 $\mu m$ $Definity^{\textregistered}$ MB as a function of pressure when $f=25 MHz$. At low pressures there are linear oscillations with period doubling (Pd) at $\approx 70 kPa$. P2 oscillations undergo further Pd to P4-2 oscillations at $\approx 210 kPa$. The process of P4-2 generation and disappearance is through a bubbling bifurcation. In case of the $1.89 \mu m$ $Definity^@$ (Fig. 6b), P4 oscillations are generated through a direct period quadrupling via a saddle node bifurcation similar to \cite{26}. This is the reason why we named this a P4-1 oscillations. Models for uncoated MBs or coated MBs with pure viscoelastic behavior predict very high pressures for the generation of P4-1 oscillations; however, here we show, for the first time, that the dynamic variation of the shell elasticity including buckling and rupture enhances the generation of the P4-1 oscillations at very low acoustic pressures ( $P_a\approx16.6 kPa$ in Fig. 6b).\\
Fig. 7 compares the radial oscillations, phase portraits and the power spectra of the $P_{sc}$ for both P4 oscillations at $f=25 MHz$ and $P_a=250 kPa$. P4-2 radial oscillations consist of two envelops, with each envelope having 2 maxima or one with 2 maxima and the other with a maxima and a critical point. These envelopes repeat themselves once every two acoustic cycles in Fig. 7a. The phase portrait of the P4-2 oscillations consists of a loop undergoing two internal loops with the largest loop undergoing another internal loop. The power spectrum depicts SHs with strength order of $1/2>3/4>1/4$. P4-1 oscillations in Fig. 7d have one envelope with 4 maxima which repeats itself once every 4 acoustic cycles. The phase portrait consist of a main loop that has undergone 3 bends to create 3 internal loops. The frequency spectrum of $P_{sc}$ depicts SHs in the strength order of $1/4>1/2>3/4$. It should be noted that due to the lower sensitivity of the transducer as we move away from central frequency, the strength order of the SHs that are detected in experiments were different. After, convolving the simulations results with the one way transducer response, experiments and simulations were in good agreement. To our best knowledge, this is the first time that the two types of P4 oscillations are detected experimentally and characterized numerically for a MB oscillator.
\section{Discussion}
A MB oscillator is an extremely complex system that has beneficial applications in a wide range of fields including material science and sonochemistry \cite{59,60,61}, food technology \cite{62} underwater acoustics \cite{63,64} and medical ultrasound (ranging from imaging blood vessels \cite{65}, drug delivery \cite{51} to thrombolysis \cite{66} and the treatment of brain through intact skull \cite{49,67}). In addition to these important applications, the complex dynamical properties of the MB system make it a very interesting subject in the field of nonlinear dynamics. It is well known that an ultrasonically excited MB is a highly nonlinear oscillator. Due to the importance of the understanding of the MB behavior in several applications, numerous studies have employed the methods of nonlinear dynamics and chaos to study the complex behavior of the system. Pioneering works of \cite{20,61,68} have revealed several nonlinear and chaotic properties of the MB oscillations (both numerically and experimentally). Recent extensive studies on the nonlinear behavior of MBs in water \cite{23,24,25,26,27,43,44}, coated MBs \cite{26,27,42}, MBs in highly viscous media \cite{28,29,30,31,32,33}, MBs sonicated with asymmetrical driving acoustic forces \cite{33,34,35,36,37,38,39,40,41} and MBs in non-Newtonian fluids \cite{37} have revealed many nonlinear features in the MB behavior. Occurrence of P2, P3, P4-2, P4-1 and higher periods, as well as chaotic oscillations, has been demonstrated in these works. Moreover, the effect of nonlinear dynamics of MBs on the propagation of sound waves in a bubbly medium is under recent investigation\cite{26,69,70}.\\ Despite these studies that employed the methods of chaos physics to investigate the nonlinear dynamics of the uncoated and coated MBs with viscoelastic behavior, the effect of the lipid coating on the dynamics of the MB especially in the realm of nonlinear dynamics and chaos has not been systematically investigated.\\
In this study we investigated the bifurcation structure of the lipid coated MBs and used the numerical results to help interpret unique signals that we observed experimentally. In stark contradiction to the results of classical theory of uncoated MBs, and despite the increased damping of the coated MBs, lipid coated MBs exhibited higher order nonlinear behavior at low excitation amplitudes (shown here both experimentally and numerically). The numerical and experimental findings can be summarized as follows:\\
a- Theoretically, we have shown that even at pressures as low as $5 kPa$, $6th-2nd$ order SuHs, P4-2, P2, P3 P4-1 and chaotic regimes manifest themselves in the MB behavior. To our best knowledge the existence of higher order SHs and chaotic behavior at such low excitation amplitudes is first reported here.\\
b- The initial surface tension of the MB plays a critical role in the enhanced nonlinear behavior. We have shown that the closer $\sigma(R_0)$ is to 0 (leading to buckling and compression only behavior) or to $\sigma_{water}$ (leading to shell rupture and expansion dominated behavior), the lower the excitation threshold for nonlinear behavior and the higher the order of non-linearity.\\
c- Despite the increased damping of the lipid coated MBs we show that, the MBs with surface tension $\geq0.062 N/m$ may have higher radial oscillation amplitude compared to the uncoated bubble.\\
d- We have experimentally shown that single $Definity^{\textregistered}$ MBs, can exhibit, P2, P3, P4-2 and P4-1 oscillations at high frequencies (25 MHz) and low pressures (250 kPa). These results can not be predicted using conventional coated MB models (with pure viscoelastic behavior) and they even contradict predictions of uncoated MB models with less damping effects.\\
e- Through numerical simulations of Marmottant model \cite{12} and visualization of the results using bifurcation diagrams we showed that $Definity^{\textregistered}$ MBs can exhibit enhanced nonlinear behavior. Using this model and assuming MBs with initial surface tension close to 0 N/m or $\sigma_{water}$ could be used to explain experimental observations of higher order nonlinear oscillations\\
f- The 5 main regimes of oscillations were identified as P1, P2, P3, P4-2 and P4-1. Simulation results of the scattered pressure were in good agreement with experimental observations both in terms of the shape of the amplitude versus time signal and also its frequency content.\\
g- For the first time, the two different P4 oscillations of the MB system were identified and characterized experimentally and numerically. P4-2 oscillations are the result of two consecutive well known period doublings while P4-1 oscillations occur through a single period quadrupling via a saddle node bifurcation. The distinct features of the signal shapes and their unique frequency spectrum were identified both experimentally and numerically. P4-1 oscillations require larger MBs compared to P4-2 oscillations.\\
Previous studies have shown that lipid coated MBs can exhibit 1/2 order subharmonic oscillations even when the excitation amplitude is low ($<$30 kPa \cite{1,2,3}) where such low pressure thresholds are below the thresholds expected even for uncoated free MBs \cite{4,5}. The low pressure threshold for SH emissions has been attributed to the buckling of the coating and compression only behavior \cite{1}. Compression dominated oscillations \cite{6} occur when the coating buckles and the effective surface tension on the MB drops to values close to zero. In such an instance, the MB compresses more than it expands. In addition to compression only behavior, lipid coating may also result in expansion dominated behavior where the MB expands more than it compresses \cite{11,13}. Expansion-dominated behavior occurs when the shell ruptures. This effect was used to explain the enhanced non-linearity at at higher frequencies (25 MHz) \cite{11,72}. Theoretical analysis of the Marmottant model for lipid coated MBs \cite{12} by Prosperetti \cite{4} attributed the lower SH threshold of the lipid MBs to the variation in the mechanical properties of the coating in the neighborhood of a certain MB radius (e.g. occurrence of buckling). In this work we show that there is a symmetry for enhanced non-linearity in the bifurcation structure of the $\frac{R}{R_0}$ of the MB as a function of $\sigma(R_0)$. Both buckling and rupture can be responsible for enhanced non-linearity, where the closer the $\sigma(R_0)$ to the buckling state (0 N/m) or rupture threshold (0.072 N/m), the lower the excitation threshold required for the generation of nonlinear oscillations. Moreover, the closer the $\sigma(R_0)$ to these two limit values, the higher the order of the nonlinearity.\\
Using the estimated parameters for the $Definity^{\textregistered}$ MB in \cite{54} and considering the shear thinning \cite{53}, the observed experimental behavior was only replicated for MBs with initial surface tension close to the two limit values of 0 and 0.072 N/m. However, it should be noted that during the sonication of a polydisperse solution of lipid MBs different values in initial surface tension and coating properties (coating elasticity and viscosity) are expected. It is been reported that even for MBs of the same size, the lipid coating can be different from MB to MB and are shown to be heterogeneous for MBs smaller than 10 $\mu m$\cite{73,74}. Despite the better homogenity of lipid distribution in lipid coated MBs similar to $Definity^{\textregistered}$ \cite{73,74}, the small differences in the lipid distribution in the coating influences the effective coating properties, thus changing the MB response \cite{75,76,77}. Moreover, its shown that the coating elasticity and coating viscosity changes with the MB size \cite{78,79,80}. Despite assuming the same coating properties for all MB sizes in this work, we were still be able to replicate the peculiar higher order nonlinearities in experiments. Moreover, we used the simplest model for lipid coated MBs and we neglected the possible stiffness softening \cite{81} or higher viscoelastic effects. Implementation of these effects are outside of the focus of this study but can be used to better characterizing the coating. In addition, simulation results only implemented a monofreqeuncy ultrasound source, and the effects of nonlinear propagation of sound waves in the medium are neglected. The generation of the SHs and UHs were not due to the nonlinear propagation of waves as nonlinear propagation manifests itself through generation of only harmonics.\\
Effects of the shape oscillations on the MBs response were also neglected in this paper. Holt and Crum observed significant effects of shape oscillations on the nonlinear behavior of the larger MBs with initial radii between $20 \mu m<R_0< 100 \mu m$ \cite{82}. Versluis et al. \cite{83} using high speed optical observations identified time-resolved shape oscillations of mode n= 2 to 6 in the behavior of single air bubbles with radii between $10 \mu m$ and 45 $\mu m$. \cite{83} concludes that close to resonance, bubbles were found to be most vulnerable toward shape instabilities. The effect of non-spherical bubble oscillations on nonlinear bubble behavior is studied in \cite{84} through GPU accelerated large parameter investigations. The active cavitation threshold has been shown to depend on the shape instability of the bubble \cite{84}. \cite{84} also shows that shape instability can affect the subharmonic threshold and nonlinear behavior of bubbles. Nonspherical oscillations of ultrasound contrast agent coated MBs are investigated in \cite{85} through high speed optical observations. They
showed that non-spherical bubble oscillations are significantly present in medically relevant ranges of bubble radii and applied acoustic pressures. Non-spherical oscillations develop preferentially at the resonance radius and may be present during SH oscillations \cite{85}. Thus, for a more accurate modeling of the MB oscillations, more sophisticated theoretical modeling of bubble coating, accounting for membrane shear and bending is required \cite{85}. Recently Guédra et al. \cite{86} showed that at sufficiently large pressure amplitudes, energy transfer from surface to volume oscillations may trigger subharmonic spherical oscillations at smaller pressure amplitudes than predicted by models of spherical only bubble oscillations. Guédra and Inserra studied the nonlinear interactions between spherical and nonspherical modes in \cite{87}. They showed that bubble
shape oscillations could be sustained for excitation amplitudes below the classical parametric threshold.
Experimentally, nonlinear interactions between the spherical, translational, and shape oscillations of micrometer-size bubbles have been reported in detail in \cite{88}. Liu et al \cite{POF2} numerically studied the shape oscillations of encapsulated bubbles. They showed that, in case of very small encapsulated microbubbles, the shape oscillation is less likely to occur since the surface tension suppresses the developments of nonspherical shape modes. Their model however, does not take into account the dynamic variations in the effective surface tension that occurs in case of lipid coated bubbles. Liu and Wang \cite{POF3} showed that shape modes of an encapsulated microbubbles in a viscous Newtonian liquid are stable. In case of encapsulated bubbles with stiffness hardening or softening behavior, Tisiglifis and Plekasis \cite{POF4} numerically showed that parametric instability is possible and result in shape oscillations as a result of subharmonic or harmonic resonance. Generation of the shape modes for encapsulated bubbles exhibitng breakup and buckling have been nurmeically investigated for higher bubble oscillation amplitude \cite{POF5}. Implementation of the nonspeherical bubble oscillations is beyond the scope of current paper, however, it may help to achieve a better agreement between numerical simulations and experimental observations in case of larger bubbles (e.g. Figure 5m-o).\\
Since experimental study was done in very dilute solutions of MBs to record single scattering events, bubble-bubble interaction was not included in the model. Bubble-bubble interaction have been shown to lower the threshold of subharmonic emissions and chaos \cite{44,89,90,91}. However even at very high concentrations \cite{44} the changes to the pressure threshold of subharmonics is minimal compared to the signification decrease in the onset of higher order nonlinearities when the shell undergoes buckling and rupture.\\
Generation of higher order SHs at low pressures may have potential in high resolution SH imaging due to their higher frequencies, higher contrast to tissue and signal to noise ratio. A SH of order 2/3 or 3/4 can be detected more effectively by the transducer as they are closer to the transducer center frequency when compared to 1/2 order SHs. Moreover, the higher scattered pressures, faster oscillations and the lower frequency contents of the oscillations of the higher order SHs may enhance the nondestructive shear stress on cells for enhanced drug delivery or in cleaning applications. Mixing applications are another category of applications that can take advantage of higher order SHs at high frequencies.
\section{Conclusion}
We have shown experimentally and for the first time that higher order SHs (e.g. 1/3,1/4,..) can be generated in the oscillations of lipid coated MBs when insonated at high frequencies and low excitation amplitudes. The bifurcation structure of a simple model of lipid coated MBs were studied as function of frequency and size to explain the experimental observations. We showed that compression only behavior or expansion dominated oscillations due to buckling and rupture of the coating and dynamic variation of the effective surface tension can explain the observed enhanced non-linearity in MBs oscillations.\\
\textbf{Acknowledgments}\\
The work is supported by the Natural Sciences and Engineering Research Council of Canada (Discovery Grant RGPIN-2017-06496), NSERC and the Canadian Institutes of Health Research ( Collaborative Health Research Projects ) and the Terry Fox New Frontiers Program Project Grant in Ultrasound and MRI for Cancer Therapy (project $\#$1034). A. J. Sojahrood is supported by a CIHR Vanier Scholarship and Qian Li is supported by NSF CBET grant $\#$1134420.
|
\section{Introduction}
Let $E/\mathbb{Q}$ be an optimal elliptic curve and let $-D$ be a negative fundamental discriminant, i.e., $-D$ is negative and is the discriminant of a quadratic number field. Assume that $D$ is coprime to the conductor $N$ of $E/\mathbb{Q}$ and let $E^{-D}/\mathbb{Q}$ denote the twist of $E/\mathbb{Q}$ by $-D$. Let $L(E^{-D},s)$ be the $L$-function of $E^{-D}/\mathbb{Q}$ and assume that $L(E^{-D},1)\neq 0$, i.e., that $E^{-D}/\mathbb{Q}$
has analytic rank $0$. In \cite{aga}, under some additional mild hypotheses, Agashe proved that $\frac{L(E^{-D},1)}{\Omega(E^{-D})} \in \mathbb{Z}[\frac{1}{2}]$. The second part of the Birch and Swinnerton-Dyer conjecture for elliptic curves of rank $0$ predicts that $$\frac{L(E^{-D},1)}{\Omega(E^{-D})}= \frac{|\Sha(E^{-D}/\mathbb{Q})| \cdot \prod_{p}c_p(E^{-D})}{|E^{-D}(\mathbb{Q})|^2},$$ see Conjecture $1.2$ of \cite{aga}. Here $\Sha(E^{-D}/\mathbb{Q})$ denotes the Shafarevich-Tate group of $E^{-D}/\mathbb{Q}$ and $c_p(E^{-D})$ is the order of the arithmetic component group of $E^{-D}/\mathbb{Q}$ at $p$, also called the Tamagawa number of $E^{-D}/\mathbb{Q}$ at $p$. This, along with results of \cite{pal}, led Agashe, in Section $5.3$ of the appendix of \cite{pal}, to propose the following conjecture.
\begin{conjecture}\label{agasheconjecture}
Let $E/\mathbb{Q}$ be an optimal elliptic curve of conductor $N$ and let $-D$ be a negative fundamental discriminant such that $D$ is coprime to $N$. Let $E^{-D}/\mathbb{Q}$ denote the twist of $E/\mathbb{Q}$ by $-D$. Suppose that $L(E^{-D},1) \neq 0$. Then
\vspace{-.1cm}
\begin{align*}
|E^{-D}(\mathbb{Q})|^2 \quad \text{divides} \quad |\Sha(E^{-D}/\mathbb{Q})| \cdot \prod_{p|N}c_p(E^{-D}), \quad \text{up to a power of 2.}
\end{align*}
\end{conjecture}
Theorem \ref{maintheorem} below, which is the main theorem of this paper, implies Conjecture \ref{agasheconjecture} (see Corollary \ref{corollaryagasheconjecture} for the proof of this implication). We note that Conjecture \ref{agasheconjecture} and, hence, our theorem can be seen as further evidence for the second part of the Birch and Swinnerton-Dyer conjecture in the analytic rank zero case.
\begin{theorem}\label{maintheorem} {\bf (see Corollary \ref{corollarytorsion} and Theorem \ref{prop:dpowerof3}) }
Let $E/\mathbb{Q}$ be an elliptic curve of conductor $N$ and let $d \neq \pm 1$ be a square-free integer such that $d$ is coprime to $N$.
\begin{enumerate}
\item If $d\neq \pm3$, then the torsion subgroup of $E^d(\mathbb{Q})$ contains only points of order dividing $8$.
\end{enumerate}
\noindent Assume also that $L(E^d,1) \neq 0$. Then the groups $E^d(\mathbb{Q})$ and $|\Sha(E^d/\mathbb{Q})|$ are finite and, up
to a power of $2$, the following conditions hold.
\begin{enumerate}[label=(\roman*),resume]
\item If $d \neq 3$, then $
|E^d(\mathbb{Q})|^2 \quad \text{divides} \quad |\Sha(E^d/\mathbb{Q})| \cdot \displaystyle\prod_{p|N}c_p(E^d).$
\item If $d=3$, then $
|E^d(\mathbb{Q})|^2 \quad \text{divides} \quad |\Sha(E^d/\mathbb{Q})| \cdot \displaystyle\prod_{p|2N}c_p(E^d).$
\end{enumerate}
\end{theorem}
\begin{remark}
If moreover $d \neq -3$, then Part $(i)$ of Theorem \ref{maintheorem} implies Part $(ii)$. We also note that the case $d = \pm 3$ is by far
the most difficult case to treat in the proof.
\end{remark}
\begin{remark}
Clearly, the analogue of Part (i) of Theorem \ref{maintheorem} with $d=1$ does not hold, since there exist elliptic curves $E/\mathbb{Q}$ of rank $0$ which have $\mathbb{Q}$-rational torsion points of order 3, 5, or 7.
Consider now the elliptic curve with LMFDB \cite{lmfdb} label 176.a2. Then $E^{-1}/\mathbb{Q}$, the twist of $E/\mathbb{Q}$ by $-1$, is the elliptic curve with LMFDB label 44.a2 and $E^{-1}(\mathbb{Q})=\mathbb{Z}/3\mathbb{Z}$. This proves that Part $(i)$ of Theorem \ref{maintheorem} is false if $d=-1$.
Moreover, $|\Sha(E^{-1}/\mathbb{Q})|=1$ and $\prod_{p} c_p(E^{-1})=c_2(E^{-1})=3$. Therefore, $|E^{-1}(\mathbb{Q})|^2=9$ while $|\Sha(E^{-1}/\mathbb{Q})|\prod_{p} c_p(E^{-1})=3$. This proves that the analogues of Theorem \ref{maintheorem}, Parts $(ii), (iii)$ are false when $d= -1$.
\end{remark}
The proof of Part $(i)$ of Theorem \ref{maintheorem} is postponed to Corollary \ref{corollarytorsion} and the proof of Parts $(ii)$ and $(iii)$ is completed in Theorem \ref{prop:dpowerof3}. To prove Parts $(ii)$ and $(iii)$ we prove Theorem \ref{3torsionshaandtamagawa}, which has applications to another consequence of the Birch and Swinnerton-Dyer conjecture, discovered by Agashe and Stein in \cite{agashestein} (see Remark \ref{remarkagashestein}).
This paper is organized as follows. First, we prove that if $d \neq \pm 1, \pm 3$, then $E^d(\mathbb{Q})$ only contains points of order a power of $2$. This explains some observations made by Agashe, in the case where $d$ is a negative fundamental discriminant, which can be found on page $8$ of \cite{aga}. Next, we prove that for $d=-3$ or $3$, $E^d(\mathbb{Q})$ cannot contain points of order $5$ or $7$. All this is done in section \ref{section2}. Finally, in section \ref{section3} we consider in detail the case where $d=\pm 3$ and $E^d(\mathbb{Q})$ contains a point of order $3$.
\begin{acknowledgement}
This work is part of the author's doctoral dissertation at the University of Georgia. The author would like to thank his advisor Dino Lorenzini for valuable help during the preparation of this work as well as for many useful suggestions on improving the exposition of this manuscript. The author thanks the anonymous referee for many insightful comments.
\end{acknowledgement}
\section{Restrictions on the torsion subgroup}\label{section2}
Let $E/\mathbb{Q}$ be an elliptic curve and let $d$ be any non-zero integer. Recall that the twist of $E/\mathbb{Q}$ by $d$, which is denoted by $E^d/\mathbb{Q}$, is an elliptic curve over $\mathbb{Q}$ that becomes isomorphic over $\mathbb{Q}(\sqrt{d})$ to $E_{\mathbb{Q}(\sqrt{d})}/\mathbb{Q}(\sqrt{d})$ but which is not in general isomorphic to $E/\mathbb{Q}$ over $\mathbb{Q}$. More explicitly, if $E/\mathbb{Q}$ is given by a Weierstrass equation of the form $y^2=x^3+ax+b$, then $E^d/\mathbb{Q}$ is given by $dy^2=x^3+ax+b$ or, after making a change of variables, by $y^2=x^3+d^2ax+d^3b$.
In this section, we obtain results that put restrictions on the rational torsion subgroup of $E^d/\mathbb{Q}$ provided that $d$ and the conductor of $E/\mathbb{Q}$ are coprime.
\begin{proposition}\label{no5or7torsion}
Let $E/\mathbb{Q}$ be an elliptic curve defined over $\mathbb{Q}$ of conductor $N$, and let $d \neq \pm 1$ be any square-free integer that is coprime to $N$. Then $E^d(\mathbb{Q})$ cannot contain a rational point of order $5$ or $7$.
\end{proposition}
\begin{proof}
We prove the proposition by contradiction. Let $\ell$ be equal to $5$ or $7$, assume that $E^d(\mathbb{Q})$ contains a point of order $\ell$, and that $d \neq \pm 1$. Since $E^d/\mathbb{Q}$ has a $\mathbb{Q}$-rational point of order $\ell$, Proposition $8.4$ in Chapter XV of \cite{mfl} implies that $E^d/\mathbb{Q}$ is semi-stable away from $\ell$.
Assume that there exists an odd prime $p$ with $p \mid d$. Then $p \nmid N$ because $d$ and $N$ are coprime. Therefore, Proposition $1$ of \cite{com} implies that $E^d/\mathbb{Q}$ has reduction type I$_0^*$ modulo $p$. As a result, since $E^d/\mathbb{Q}$ is semi-stable away from $\ell$, we obtain that $p$ can only be equal to $\ell$. However, Proposition \ref{mentzelospaper5,7} below implies that $E^d/\mathbb{Q}$ cannot have reduction of type I$_0^*$ modulo $\ell$. Therefore, $d= \pm 2$ since $d$ is square-free.
Assume now that $d=-2$ or $2$. We will arrive at a contradiction. Since we assume that $d$ and $N$ are coprime, we get that $2 \nmid N$. Since $E/\mathbb{Q}$ has good reduction modulo $2$, by Table II of \cite{com} (see also the hypotheses for this table at the bottom of page 58 of \cite{com}) we obtain that $E^d/\mathbb{Q}$ has reduction of type I$_8^*$ or II modulo $2$. In \cite{com}, good reduction is denoted as usual with the symbol I$_0$, and our twist $E^d$ is denoted by $E^{\chi}$. Note that Table II is independent of the existence of a torsion point. We now use our assumption that $E^d$ has a $\mathbb{Q}$-rational point of order $\ell$. Since $E^d$ has a $\mathbb{Q}$-rational point of order $\ell$, it has semi-stable reduction modulo $2$, which is a contradiction. This proves that $E^d(\mathbb{Q})$ cannot contain a point of order $\ell$.
\end{proof}
A proof of the following proposition can be found in the forthcoming paper \cite{mentzeloskodairandtorsion}.
\begin{proposition}\label{mentzelospaper5,7}
Let $E/\mathbb{Q}$ be an elliptic curve with a $\mathbb{Q}$-rational point of order $p$, where $p$ is a prime number. Assume that $E/\mathbb{Q}$ has additive reduction modulo $p$. Then $p \leq 7$, and
\begin{enumerate}[label=(\roman*),topsep=2pt]\label{5and7torsionreduction}
\itemsep0em
\item If $p=5$, then $E/\mathbb{Q}$ can only have reduction of type II or III modulo $5$.
\item If $p=7$, then $E/\mathbb{Q}$ can only have reduction type II modulo $7$.
\end{enumerate}
\end{proposition}
We now investigate whether $E^d(\mathbb{Q})$ can contain a point of order $3$. Let $E/\mathbb{Q}$ be an elliptic curve with a $\mathbb{Q}$-rational point of order $3$. Then, by translating that point to $(0,0)$ and performing a change of variables if necessary (see Remark $2.2$ in Section $4.2$ of \cite{hus}) we obtain a Weierstrass equation for $E/\mathbb{Q}$ of the form $$y^2+cxy+dy=x^3,$$ with $c,d \in \mathbb{Q}$. If $u \in \mathbb{Z}$, then the transformation $(x,y) \rightarrow{(\frac{x}{u^2}, \frac{y}{u^3})}$ gives a new Weierstrass equation of the same form with $c$ replaced by $uc$ and with $d$ replaced by $u^3d$ (see page 185 of \cite{aec}). Therefore, by picking $u$ to be a large power of the product of the denominators of $c,d$ (if any), we can arrange that $c,d \in \mathbb{Z}$. Moreover, by applying the transformation $(x,y) \rightarrow{(x, -y)}$ if necessary, we can arrange that $d > 0$. We now show that we can find a new Weierstrass equation of the above form with coefficients, which we will call $a$ and $b$ below, that also have the property that for every prime $q$ either $q \nmid a$ or $q^3 \nmid b$. If there is no prime $q$ such that $q \mid c$ and $q^3 \mid d$, then set $a=c$ and $b=d$. Otherwise, let $q_1,...,q_s$ be the set of primes such that $q_i \mid c$ and ${q_i}^3 \mid d$, and let $n_i= \mathrm{min}\{ \mathrm{ord}_{q_i}(c), \lfloor \frac{\mathrm{ord}_{q_i}(d)}{3} \rfloor \} $, where $\lfloor - \rfloor$ is the floor function. Let $u=\prod_{i=1}^{s} q_i^{n_i}$. By applying the transformation $(x,y) \rightarrow{(u^2x, u^3y)}$, we obtain a new Weierstrass equation of the form $$y^2+\frac{c}{u}xy+\frac{d}{u^3}y=x^3.$$ Setting $a=\frac{c}{u}$ and $b=\frac{d}{u^3}$ we see that $a,b \in \mathbb{Z}$, $b >0$, and for every prime $q$ either $q \nmid a$ or $q^3 \nmid b$. Therefore, we have proved that we can choose a Weierstrass equation for $E/\mathbb{Q}$ of the form
\begin{align*}\label{eq:3torsion}
y^2+axy+by=x^3, \tag{$2.3$}
\end{align*}
where $a,b$ are integers, $b>0$, and for every prime $q$ either $q \nmid a$ or $q^3 \nmid b$. Also, we must have $a^3-27b \neq 0$, since the discriminant of Equation $($\ref{eq:3torsion}$)$ is $$\Delta=b^3(a^3-27b) \quad \text{and we also have} \quad c_4=a(a^3-24b).$$
\stepcounter{theorem}
\begin{proposition}\label{prop:3torsionreduction} (See Proposition $3.5$ and Lemma $3.6$ of \cite{koz})
Let $E/\mathbb{Q}$ be an elliptic curve given by Equation $($\ref{eq:3torsion}$)$. Write $D:=a^3-27b$ and let $p$ be any prime (note that either $p \nmid a$ or $p^3 \nmid b$). Then the reduction of $E/\mathbb{Q}$ modulo $p$ is determined as follows:
\begin{enumerate}
\item $\text{If} \; 3ord_p(a) \leq ord_p(b)$ : $\begin{cases}
3ord_p(a)< ord_p(b) \rightarrow \text{split I}_{3ord_p(b)}, \quad c_p(E)=3ord_p(b) \\
3ord_p(a)= ord_p(b) \rightarrow \begin{cases}
ord_p(D)>0 \rightarrow \text{I}_{ord_p(D)}\\
ord_p(D)=0 \rightarrow \text{Good reduction I}_0
\end{cases}
\end{cases}$
\item $\text{If} \; 3ord_p(a) > ord_p(b)$ : $\begin{cases}
ord_p(b) = 0 \rightarrow \begin{cases}
p=3 \rightarrow \text{Go to} \; (iii) \\
p \neq 3 \rightarrow \text{Good reduction I}_0
\end{cases} \\
ord_p(b) = 1 \rightarrow \text{IV, } c_p(E)=3\\
ord_p(b) = 2 \rightarrow \text{IV}^*,\text{ } c_p(E)=3.
\end{cases}$
\item If $p=3$ and $ord_p(a)>0=ord_p(b)$: $ord_p(D)=$ $\begin{cases}
3 \rightarrow \text{II or III}\\
4 \rightarrow \text{II}\\
5 \rightarrow \text{IV}\\
n \rightarrow \text{I}_{n-6}^* \text{ , for } n \geq 6.
\end{cases}$
\end{enumerate}
\end{proposition}
\begin{proposition}\label{no3torsiondnotequaltopm3}
Let $E/\mathbb{Q}$ be an elliptic curve over $\mathbb{Q}$ of conductor $N$ and let $d \neq \pm1, \pm 3$ be a square-free integer coprime to $N$. Then $E^d(\mathbb{Q})$ cannot contain a point of order $3$.
\end{proposition}
\begin{proof}
The proof is by contradiction. Assume that $E^d(\mathbb{Q})$ contains a point of order $3$ and that $d$ has a prime divisor $p \neq 3$. Assume first that $p \neq 2$. Since $p \nmid N$, Proposition $1$ of \cite{com} implies that $E^d/\mathbb{Q}$ has reduction of type I$_0^*$ modulo $p$. However, this is impossible because by Proposition \ref{prop:3torsionreduction}, $E^d(\mathbb{Q})$ cannot have reduction of type I$_0^*$ modulo $p$ for $p \neq 3$. Assume now that $p=2$. Since $2 \nmid N$, we obtain that $E/\mathbb{Q}$ has good reduction modulo $2$ and, hence, by Table II of \cite{com} (see also the hypotheses for this table at the bottom of page 58 of \cite{com}), we get that $E^d/\mathbb{Q}$ has reduction of type I$_8^*$ or II modulo $2$. However, $E^d/\mathbb{Q}$ cannot have reduction of type I$_8^*$ or II modulo $2$ by Proposition \ref{prop:3torsionreduction}. This implies that $2 \nmid d$. Therefore, $d$ can only be divisible by $3$ so $d= \pm 1$ or $\pm3$.
\end{proof}
\begin{corollary}\label{corollarytorsion}
Let $E/\mathbb{Q}$ be an elliptic curve over $\mathbb{Q}$ of conductor $N$ and let $d \neq \pm1$ be a square-free integer such that $d$ is coprime to $N$. Then
\begin{enumerate}[label=(\roman*),topsep=2pt]
\itemsep0em
\item $E^d(\mathbb{Q})_{tors}$ contains only points of order $2^{\alpha}3^{\beta}$ where $\alpha, \beta \geq 0$.
\item If $d \neq \pm3$, then $E^d(\mathbb{Q})_{tors}$ contains only points of order dividing $8$.
\end{enumerate}
\end{corollary}
\begin{proof}
By a Theorem of Mazur (see \cite{maz} Theorem $(8)$) the only primes that can divide $|E^d(\mathbb{Q})_{\mathrm{tors}}|$ are $2,3,5$ and $7$. Therefore, Proposition \ref{no5or7torsion} implies Part $(i)$. Moreover, Propositions \ref{no5or7torsion} and \ref{no3torsiondnotequaltopm3} imply that if $d \neq \pm 1, \pm 3$, then $E^d(\mathbb{Q})_{tors}$ contains only points of order dividing $2^{\alpha}$, where $\alpha \geq 0$. Finally, Part $(ii)$ follows by applying Mazur's Theorem (see \cite{maz} Theorem $(8)$).
\end{proof}
\section{Elliptic curves over $\mathbb{Q}$ with a rational point of order 3}\label{section3}
In this section, we prove Parts $(ii)$ and $(iii)$ of Theorem \ref{maintheorem} in Theorem \ref{prop:dpowerof3}. To achieve this goal, we will use Theorem \ref{3torsionshaandtamagawa} below, which might be of independent interest (see Remark \ref{remarkagashestein}).
\begin{theorem}\label{3torsionshaandtamagawa}
Let $E/\mathbb{Q}$ be an elliptic curve with a $\mathbb{Q}$-rational point of order $3$. Assume that the analytic rank of $E/\mathbb{Q}$ is $0$ and that $E/\mathbb{Q}$ has reduction of type I$_n^*$ modulo $3$, for some $n \geq 0$.
\begin{enumerate}[label=(\alph*)]
\item If $E/\mathbb{Q}$ has semi-stable reduction away from $3$, then $9 \mid |\Sha(E/\mathbb{Q})| \cdot \prod_{p}c_p(E)$.
\item If $E/\mathbb{Q}$ has more than two places of additive reduction, then $9 \mid \prod_{p}c_p(E)$.
\item If $E/\mathbb{Q}$ has exactly two places of additive reduction, then $9 \mid |\Sha(E/\mathbb{Q})| \cdot \prod_{p}c_p(E)$.
\end{enumerate}
\end{theorem}
\begin{proof}[\it Proof of Theorem \ref{3torsionshaandtamagawa}]
Let $j_E$ be the $j$-invariant of $E/\mathbb{Q}$. Since some of our arguments below do not work for $j_E=0$ or $1728$, we first handle these cases with the following claim.
\begin{claim}\label{claimjinvariants}
Let $E/\mathbb{Q}$ be an elliptic curve with $j$-invariant $0$ or $1728$ that has a $\mathbb{Q}$-rational point of order $3$. Then $E/\mathbb{Q}$ cannot have reduction of type I$_n^*$ modulo $3$, for any $n \geq 0$.
\end{claim}
\begin{proof}
Let $E/\mathbb{Q}$ be an elliptic curve with $j$-invariant $j_E=0$. Since, $j_E=\frac{{c_4}^3}{\Delta}$ we obtain that $c_4=0$. Tableau II of \cite{pap} implies that $E/\mathbb{Q}$ can only have reduction type II, II$^*$, III, III$^*$, IV, or IV$^*$ modulo $3$. Therefore, $E/\mathbb{Q}$ cannot have reduction type I$_n^*$ modulo $3$, for any $n \geq 0$, as needed.
Let $E/\mathbb{Q}$ be an elliptic curve with $j$-invariant $j_E=1728$ and with a $\mathbb{Q}$-rational point of order $3$. Since $E/\mathbb{Q}$ has a $\mathbb{Q}$-rational point of order $3$, it has a Weierstrass equation of the form $($\ref{eq:3torsion}$)$. Assume that $E/\mathbb{Q}$ has reduction type I$_n^*$ modulo $3$, for some $n \geq 0$, and we will find a contradiction. Proposition \ref{prop:3torsionreduction}, Parts $(i)$ and $(ii)$ imply that $3 \nmid b$ and $3 \mid a$. Since $j_E=\frac{c_4^3}{\Delta}$ and $j_E=1728$, we obtain that $c_4^3=1728\Delta$. Writing $c_4$ and $\Delta$ in terms of $a$ and $b$, we get that $(a(a^3-24b))^3=1728b^3(a^3-27b)$ and, hence, $$3{\rm ord}_3(a)+3{\rm ord}_3(a^3-24b)={\rm ord}_3(1728)+3{\rm ord}_3(b)+{\rm ord}_3(a^3-27b).$$ Since $3 \mid a$ and $3 \nmid b$, we get that ord$_3(a^3-24b)=1$. Therefore, $$3{\rm ord}_3(a)+3=3+0+{\rm ord}_3(a^3-27b),$$ which yields
$$3{\rm ord}_3(a)={\rm ord}_3(a^3-27b).$$
If ${\rm ord}_3(a)=1$, then ${\rm ord}_3(a^3-27b)=3$, which is a contradiction because Proposition \ref{prop:3torsionreduction}, Part $(iii)$ implies that $E/\mathbb{Q}$ has reduction type II or III. If ${\rm ord}_3(a) \geq 2$, then ${\rm ord}_3(a^3-27b)=3$ because ${\rm ord}_3(b)=0$, and this is again a contradiction. This proves our claim.
\end{proof}
Claim \ref{claimjinvariants} implies that to prove our theorem, it is enough to assume that $j_E \neq 0, 1728$. We assume from now on that $j_E \neq 0, 1728$.
Since $E/\mathbb{Q}$ has a $\mathbb{Q}$-rational point of order $3$, it can be given by an equation of the form $($\ref{eq:3torsion}$)$ where $(0,0)$ is a $\mathbb{Q}$-rational point of order $3$. Let $\widehat{E}:=E/<(0,0)>$ and let $\phi: E \rightarrow{} \widehat{E}$ be the associated $3$-isogeny. We denote by $\hat{\phi}$ the dual isogeny.
We recall now some important facts in preparation for the proof of Lemma \ref{tamagawagreaterorequalto2} below. The following formula is due to Cassels (see Theorem $1$ of \cite{ks} or Theorem $2.1$ of \cite{koz})
\begin{align}\label{eq:selmerandtamagawa}
\frac{|\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})|}{|\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})|}=\frac{|E[\phi](\mathbb{Q})|\Omega(\widehat{E})\prod_{p}c_p(\widehat{E})}{|\widehat{E}[\hat{\phi}](\mathbb{Q})|\Omega(E)\prod_{p}c_p(E)}. \tag{3.3}
\end{align}
Here $\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})$ and $\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})$ are the $\phi$- and $\hat{\phi}$-Selmer groups, respectively. Moreover, $\Omega(E):=\displaystyle\int_{E(\mathbb{R})}|\omega_{min}|$ and $\Omega(\widehat{E}):=\displaystyle\int_{\widehat{E}(\mathbb{R})}|\widehat{\omega}_{min}|$, where $\omega_{min}$ and $\widehat{\omega}_{min}$ are two minimal invariant differentials on $E/\mathbb{Q}$ and $\widehat{E}/\mathbb{Q}$ respectively (see Section III.1 of \cite{aec} and page 451 of \cite{aec}). The interested reader can consult section X.$4$ of \cite{aec} concerning the definitions of Selmer groups of elliptic curves as well as relevant background.
Since $E[\phi](\mathbb{Q}) \cong \mathbb{Z}/3\mathbb{Z}$, we get that $\widehat{E}[\hat{\phi}](\mathbb{Q})$ is trivial by the Weil pairing (see \cite{aec} exercise III.3.15 on page $108$). Hence,
\stepcounter{theorem}
\begin{align}\label{kernelofisogeny}
\frac{|E[\phi](\mathbb{Q})|}{|\widehat{E}[\hat{\phi}](\mathbb{Q})|}=3. \tag{3.4}
\end{align}
We know that $\phi^*\widehat{\omega}_{min}=\lambda \omega_{min}$, for some $\lambda \in \mathbb{Z}$ (see the second paragraph of page $284$ of \cite{vatsal}). Moreover, the fact that $\hat{\phi}\circ \phi=[3]$ implies that $|\lambda|=1$ or $3$.
Therefore, $|\frac{\omega_{min}}{\phi^*\widehat{\omega}_{min}}|$ is equal to $1$ or $3^{-1}$. By Lemma $7.4$ of \cite{dd},
\begin{align*}
\frac{\Omega(E)}{\Omega(\widehat{E})}=\frac{|{\rm ker} (\phi :E(\mathbb{R}) \longrightarrow \widehat{E}(\mathbb{R}))|}{|{\rm coker} (\phi :E(\mathbb{R}) \longrightarrow \widehat{E}(\mathbb{R}) )|}\cdot|\frac{\omega_{min}}{\phi^*\widehat{\omega}_{min}}|
\end{align*}
Since ${\rm ker} (\phi) \subset E(\mathbb{R})$ and $\phi$ has degree $3$, Proposition $7.6$ of \cite{dd} implies that $\frac{|{\rm ker} \phi :E(\mathbb{R}) \longrightarrow \widehat{E}(\mathbb{R})|}{|{\rm coker} (\phi :E(\mathbb{R}) \longrightarrow \widehat{E}(\mathbb{R}) )|}=3$. Therefore, since $|\frac{\omega_{min}}{\phi^*\widehat{\omega}_{min}}|$ is equal to $1$ or $3^{-1}$,
\stepcounter{theorem}
\begin{align}\label{Omega}
\frac{\Omega(E)}{\Omega(\widehat{E})}=1 \; \text{or} \; 3. \tag{3.5}
\end{align}
\stepcounter{theorem}
The following lemma will be used repeatedly in what follows. In pursuing this idea, we were inspired by work of Byeon, Kim, and Yhee in \cite{bky}.
\begin{lemma}\label{tamagawagreaterorequalto2}
Let $E/\mathbb{Q}$ be an elliptic curve given by a Weierstrass equation of the form $($\ref{eq:3torsion}$)$ and assume that $E/\mathbb{Q}$ has analytic rank $0$. Let $\widehat{E}:=E/<(0,0)>$ and let $\phi: E \rightarrow{} \widehat{E}$ be the associated $3$-isogeny.
\begin{enumerate}
\item If $\mathrm{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})\geq 2$ or $\rm{dim}_{\mathbb{F}_3}(\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})) \geq 2$, then $9$ divides $|\Sha(E/\mathbb{Q})|$.
\item If $\frac{|\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})|}{|\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})|}=1$, then $9$ divides $|\Sha(E/\mathbb{Q})|$.
\end{enumerate}
\end{lemma}
\begin{proof}\text{}
Since the analytic rank of $E/\mathbb{Q}$ is zero, work of Gross and Zagier, on heights of Heegner points \cite{grosszagierpaper}, as well as work of Kolyvagin, on Euler systems \cite{kolyvagineulersystems}, imply that $E/\mathbb{Q}$ has (algebraic) rank $0$ and that $\Sha(E/\mathbb{Q})$ is finite (see Theorem $3.22$ of \cite{darmonmodularellipticcurves} for a sketch of the proof).
Thus the statements of the lemma make sense.
{\it Proof of $(i)$:} First we prove that if dim$_{\mathbb{F}_3}(\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})) \geq 2$, then $9$ divides $|\Sha(E/\mathbb{Q})|$. There is a short exact sequence $$0 \longrightarrow{} E(\mathbb{Q})/3E(\mathbb{Q}) \longrightarrow{} \mathrm{Sel}^{3}(E/\mathbb{Q}) \longrightarrow{} \Sha(E/\mathbb{Q})[3] \longrightarrow{} 0$$ (see Theorem X.$4.2$ of \cite{aec}).
Therefore, using the fact that $E(\mathbb{Q})/3E(\mathbb{Q}) \cong \mathbb{Z}/3\mathbb{Z}$, since $E/\mathbb{Q}$ has rank $0$ and $E[3](\mathbb{Q}) \cong \mathbb{Z}/3\mathbb{Z}$, we see that dim$_{\mathbb{F}_3}(\mathrm{Sel}^{3}(E/\mathbb{Q})) \geq 2$ implies that $\Sha(E/\mathbb{Q})[3]$ has a nontrivial element. Since the order of $\Sha(E/\mathbb{Q})$ is a square (see \cite{aec} Corollary $17.2.1$) if $\Sha(E/\mathbb{Q})[3]$ has a nontrivial element, then $9$ divides $|\Sha(E/\mathbb{Q})|$. By Corollary 1, section 2 of \cite{ks} and the fact that $\widehat{E}[\hat{\phi}](\mathbb{Q})$ is trivial, we obtain that the natural map $\mathrm{Sel}^{(\phi)}(E/\mathbb{Q}) \longrightarrow \mathrm{Sel}^{3}(E/\mathbb{Q})$ is injective. Therefore, if dim$_{\mathbb{F}_3}(\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})) \geq 2$, then $9$ divides $|\Sha(E/\mathbb{Q})|$.
Finally, if $\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})\geq 2$, then Equation $($\ref{eq:selmerandtamagawa}$)$, combined with $($\ref{kernelofisogeny}$)$ and $($\ref{Omega}$)$, implies that dim$_{\mathbb{F}_3}(\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})) \geq 2$ so by the first part of the proof, we find that $9$ divides $ |\Sha(E/\mathbb{Q})|$. This concludes the proof of Part $(i)$.
{\it Proof of $(ii)$:}
By Corollary 1 of section $2$ of \cite{ks} and the fact that $\widehat{E}[\hat{\phi}](\mathbb{Q})$ is trivial there is an exact sequence:
\stepcounter{theorem}
\begin{align*}\label{exactsequencephiselmergroup}
0 \rightarrow \mathrm{Sel}^{(\phi)}(E/\mathbb{Q}) \rightarrow \mathrm{Sel}^3 (E/\mathbb{Q}) \rightarrow \mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q}) \rightarrow \Sha(\widehat{E}/\mathbb{Q})[\hat{\phi}]/\phi(\Sha(E/\mathbb{Q})[3]) \rightarrow 0. \tag{3.7}
\end{align*}
Moreover, by Theorem $4.3$ of \cite{shnidmanquadraticrm}, the Cassels-Tate pairing on $\Sha(\widehat{E}/\mathbb{Q})$ restricts to a non-degenerate alternating pairing on $\Sha(\widehat{E}/\mathbb{Q})[\hat{\phi}]/\phi(\Sha(E/\mathbb{Q})[3])$ (see also Theorem $3$ of \cite{fisherctp}). Therefore, it follows that $\mathrm{dim}_{\mathbb{F}_3}\Sha(\widehat{E}/\mathbb{Q})[\hat{\phi}]/\phi(\Sha(E/\mathbb{Q})[3])$ is even.
Let $$2\alpha:= \mathrm{dim}_{\mathbb{F}_3}\Sha(\widehat{E}/\mathbb{Q})[\hat{\phi}]/\phi(\Sha(E/\mathbb{Q})[3]).$$ Since $\frac{|\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})|}{|\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})|}=1$ by hypothesis, we get that
$$\beta:=\mathrm{dim}_{\mathbb{F}_3}\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})=\mathrm{dim}_{\mathbb{F}_3}\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q}) .$$ If $\gamma:=\mathrm{dim}_{\mathbb{F}_3}\mathrm{Sel}^3 (E/\mathbb{Q}) $, then by the exact sequence $($\ref{exactsequencephiselmergroup}$)$ we obtain that $$\beta-\gamma+\beta-2\alpha=0$$ and, therefore, $\gamma$ must be even. There is a short exact sequence $$0 \longrightarrow{} E(\mathbb{Q})/3E(\mathbb{Q}) \longrightarrow{} \mathrm{Sel}^{3}(E/\mathbb{Q}) \longrightarrow{} \Sha(E/\mathbb{Q})[3] \longrightarrow{} 0$$ (see Theorem X.$4.2$ of \cite{aec}). Since $E/\mathbb{Q}$ has a $\mathbb{Q}$-rational point of order three and rank $0$, we obtain that $E(\mathbb{Q})/3E(\mathbb{Q}) \cong \mathbb{Z}/3\mathbb{Z}$ and, hence, $\gamma \geq 1$. Since $\gamma$ is even, we obtain that $\gamma \geq 2$. Since $\gamma=\mathrm{dim}_{\mathbb{F}_3}\mathrm{Sel}^3 (E/\mathbb{Q}) \geq 2$ and $E(\mathbb{Q})/3E(\mathbb{Q}) \cong \mathbb{Z}/3\mathbb{Z}$, $\Sha(E/\mathbb{Q})[3]$ is not trivial. Finally, since the order of $\Sha(E/\mathbb{Q})$ is a square (see \cite{aec} Corollary $17.2.1$) if $\Sha(E/\mathbb{Q})[3]$ has a nontrivial element, then $9$ divides $|\Sha(E/\mathbb{Q})|$. This concludes the proof of Part $(ii)$.
\end{proof}
\begin{emptyremark}\label{rootnumbers}
We now collect some facts concerning root numbers. Let $w(E):=\displaystyle \prod_{p \in M_{\mathbb{Q}} } w_p(E)$ where $w_p(E)$ is the local root number of $E/\mathbb{Q}$ at $p$ and $M_{\mathbb{Q}}$ is the set of places of $\mathbb{Q}$. The number $w(E)$ is called the global root number of $E/\mathbb{Q}$.
By Theorem $3.22$ of \cite{darmonmodularellipticcurves}, $\Sha(E/\mathbb{Q})$ is finite and the rank of $E/\mathbb{Q}$ is $0$. Consequently, by Theorem $1.4$ of \cite{dd1} we get that $1=(-1)^{rk(E/\mathbb{Q})}=w(E)$.
Recall that we assume that $j_E \neq 0, 1728$. The local root number of $E/\mathbb{Q}$ at $p$ is as follows (see \cite{connel} page $95$ and \cite{rohrlich} page $132$ for $j_E \neq 0, 1728$ and $p \geq 5$, and \cite{halberstadt} page 1051 for $p=2$ or $3$)
\[ \begin{cases}
w_{\infty}(E)=-1 \\
w_p(E)=1 & \mathrm{if} \ E/\mathbb{Q} \ \mathrm{has} \ \mathrm{modulo} \ p \ \mathrm{good} \ \mathrm{or} \ \mathrm{nonsplit} \ \mathrm{multiplicative} \ \mathrm{reduction}.\\
w_p(E)=-1 & \mathrm{if} \ E/\mathbb{Q} \ \mathrm{has} \ \mathrm{modulo} \ p \ \mathrm{split} \ \mathrm{multiplicative} \ \mathrm{reduction}.\\
w_3(E)=-1 & \mathrm{if} \ E/\mathbb{Q} \ \mathrm{has} \ \mathrm{modulo} \ 3 \ \mathrm{reduction} \ \mathrm{of} \ \mathrm{type} \ \mathrm{I}_n^*\ ,n\geq 0.
\end{cases}
\]
\end{emptyremark}
\begin{lemma}\label{lemmatamagawanumber}
Let $E/\mathbb{Q}$ be an elliptic curve given by a Weierstrass equation of the form $($\ref{eq:3torsion}$)$ such that $E/\mathbb{Q}$ has modulo $3$ reduction of type I$_n^*$, for some $n \geq 0$. If $9 \nmid \prod_{p}c_p(E)$, then either $b=1$ or $b=r^m$ for some prime $r \neq 3$.
\end{lemma}
\begin{proof}
Since $E/\mathbb{Q}$ has modulo $3$ reduction of type I$_n^*$, for some $n \geq 0$, Proposition \ref{prop:3torsionreduction} implies that $3 \nmid b$, as we now explain. First, if $3 \mid b$ and $3 \nmid a$, then Proposition \ref{prop:3torsionreduction}, Part $(i)$, implies that $E/\mathbb{Q}$ has modulo $3$ multiplicative reduction, which is a contradiction. Moreover, if $3 \mid b$ and $3 \mid a$, then, $3^3 \nmid b$ so $3\textrm{ord}_3(a)>\textrm{ord}_3(b)$. Therefore, using Proposition \ref{prop:3torsionreduction}, Part $(ii)$, we obtain that $E/\mathbb{Q}$ has modulo $3$ reduction of type IV or IV$^*$, which is again a contradiction.
If a prime $r$ divides $b$, then either $r^3 \nmid b$ or $r \nmid a$. Therefore, since $r \neq 3$, by Proposition \ref{prop:3torsionreduction}, Parts $(i)$ and $(ii)$, $E/\mathbb{Q}$ has reduction of type IV, IV$^*$ or (split) I$_{3\text{ord}_{r}(b)}$, and in any of these cases $3 \mid c_r(E)$. Hence, if $9 \nmid \prod_{p}c_p(E)$, then $b$ can have at most one prime divisor.
\end{proof}
\begin{proof}[\it Proof of Theorem \ref{3torsionshaandtamagawa} Part $(a)$.]
Assume that $E/\mathbb{Q}$ is semi-stable away from $3$.
If $9$ divides $\prod_{p}c_p(E)$, then Theorem \ref{3torsionshaandtamagawa}, Part $(a)$ is true. Therefore, we can assume from now on that $9 \nmid \prod_{p}c_p(E)$ and, hence, by Lemma \ref{lemmatamagawanumber} that either $b=1$ or $b=r^m$, for some prime $r$ with $r \neq 3$. We now split the proof into two cases depending on whether $b=1$ or $b=r^m$.
\medskip \underline{Case 1:} $b=1$. By a proposition of Hadano (see Theorem $1.1$ of \cite{had}) since $b=1$, $\widehat{E}/\mathbb{Q}$ is given by $$y^2+(a+6)xy+(a^2+3a+9)y=x^3.$$ The discriminant of this equation is $\widehat{\Delta}=(a^3 - 27)^3$ and $\widehat{c_4}=a(a^3+216)$.
We first show that $\widehat{E}/\mathbb{Q}$ has a prime $q$ of split multiplicative reduction. By Proposition \ref{prop:3torsionreduction}, Part $(iii)$, we obtain that $3 \mid a$ because $E/\mathbb{Q}$ has reduction of type I$_n^*$ modulo $3$. Therefore, we can write $a=3a'$ for some integer $a'$. Then $$a^2+3a+9=9((a')^2+a'+1)$$ and $(a')^2+a'+1$ cannot be a power of $3$ unless $a'=1,-2$ because it is always non zero modulo $9$. If $(a')^2+a'+1=3$, then $a'=1$ or $-2$. Moreover, $a'=1$ gives $a=3$, and together with $b=1$ gives an equation of the form $($\ref{eq:3torsion}$)$ which has discriminant $0$ and, hence, is not an elliptic curve. Therefore, $a'=1$ is not allowed. If $a'=-2$, which implies $a=-6$, we obtain, combined with $b=1$, an elliptic curve $E/\mathbb{Q}$ of the form $($\ref{eq:3torsion}$)$ which does not have reduction of type I$_n^*$ at 3 because $a^3-27=(-6)^3-27=243$ is not divisible by $3^6$, see Proposition \ref{prop:3torsionreduction}, Part $(iii)$. Moreover, if $(a')^2+a'+1= \pm 1$, then $a'=-1$. If $a'=-1$, which implies $a=-3$, we obtain, combined with $b=1$, an elliptic curve $E/\mathbb{Q}$ of the form $($\ref{eq:3torsion}$)$ which does not have reduction of type I$_n^*$ at 3 because $a^3-27=(-3)^3-27=-54$ is not divisible by $3^6$, see Proposition \ref{prop:3torsionreduction}, Part $(iii)$. What this proves is that $a^2+3a+9$ has a prime divisor not equal to $3$, say $q$. By looking at the Weierstrass equation for $\widehat{E}/\mathbb{Q}$ we get that $\widehat{E}/\mathbb{Q}$ has split multiplicative reduction modulo $q$. Indeed, if $q \mid a^2+3a+9$ and $q \mid a+6$, then $$q \mid a^2+3a+9-(a+6)^2=a^2+3a+9-a^2-12a-36=-9(a+3)$$ and, hence, $q \mid a+3$. Since $q \mid a+6$, we obtain that $q=3$, which is a contradiction. This proves that $q \nmid a+6$ and, thus, $\widehat{E}/\mathbb{Q}$ has split multiplicative reduction modulo $q$.
Since $\widehat{E}/\mathbb{Q}$ has split multiplicative modulo $q$, we get that $c_q(\widehat{E})=\mathrm{ord}_q((a^3-27)^3)=3\mathrm{ord}_q(a^3-27)$. Since $E/\mathbb{Q}$ and $\widehat{E}/\mathbb{Q}$ are isogenous over $\mathbb{Q}$, they have
the same $L$-function (see Korollar 1 of \cite{faltingsisogeny}) and, hence, $a_q(E)=a_q(\widehat{E})$ (see page $366$ of \cite{diamondshurman}). Moreover, $a_q(\widehat{E})=1$ because $\widehat{E}/\mathbb{Q}$ has split multiplicative reduction modulo $q$ (see page $329$ of \cite{diamondshurman}), which implies that $a_q(E)=1$. Therefore, $\widehat{E}/\mathbb{Q}$ has split multiplicative reduction modulo $q$ and $c_q(E)=\mathrm{ord}_q(\Delta)$. However, $\Delta=a^3-27$ so $c_q(E) = \mathrm{ord}_q(a^3-27)$, which implies that $\textrm{ord}_3 (\frac{c_q(\widehat{E})}{c_q(E)})>0$.
We will now show that $9$ divides $|\Sha(E/\mathbb{Q})$|. By Part $(i)$ of Lemma \ref{tamagawagreaterorequalto2}, it is enough to show that $\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})\geq 2.$ We achieve this in the following claim.
\begin{claim}
We have $$\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)}) \geq 2.$$
\end{claim}
\begin{proof}
First, we claim that because the analytic rank of $E/\mathbb{Q}$ is zero, $E/\mathbb{Q}$ has an even number of places of split multiplicative reduction. Indeed, by \ref{rootnumbers} $w(E)=1$, $w_3(E)=-1$, and $w_{\infty}(E)=-1$. Moreover, $E/\mathbb{Q}$ is semi-stable away from $3$ and by \ref{rootnumbers} for $p\neq 3$ we obtain that $w_p(E)=-1$ if and only if $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$. This proves that $E/\mathbb{Q}$ has an even number of places of split multiplicative reduction.
Let $p \neq 3$ be any prime such that $E/\mathbb{Q}$ has multiplicative reduction modulo $p$. If $E/\mathbb{Q}$ has nonsplit multiplicative reduction modulo $p$, then by lines $5,6$, and $7$ of Theorem $6.1$ of \cite{dd} we obtain that $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=0$. Note that in Theorem $6.1$ of \cite{dd}, $\widehat{E}$ is denoted by $E'$ and $\delta,\delta'$ are the valuations of the two discriminants. Assume now that $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$. Recall that we assume that $b=1$ so $\Delta=a^3-27$ and $\widehat{\Delta}=(a^3-27)^3$. If $\textrm{ord}_p(\Delta)=\gamma$, then $\textrm{ord}_p(\widehat{\Delta})=3\gamma$. Therefore, by line $3$ of Theorem $6.1$ of \cite{dd} we obtain that $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=1$. Since $E/\mathbb{Q}$ is semi-stable away from $3$, the above arguments prove that if $p \neq 3$ is a prime, then $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$ if and only if $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=1$, and $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=0$ otherwise.
We now claim that ord$_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})$ is even. By line $13$ of Theorem $6.1$ of \cite{dd}, since $E/\mathbb{Q}$ has modulo 3 reduction of type I$_n^*$ for some $n\geq 0$, we obtain that $\textrm{ord}_3 (\frac{c_3(\widehat{E})}{c_3(E)})=0$. Combining all the above with the fact that $E/\mathbb{Q}$ has an even number of places of split multiplicative reduction, we obtain that ord$_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})$ is even.
Finally, since $\textrm{ord}_3 (\frac{c_q(\widehat{E})}{c_q(E)})>0$ and ord$_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})$ is even, we get that $\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)}) \geq 2$. This proves our claim.
\end{proof}
\underline{Case 2:} $b=r^m$ for some prime $r$, and $m>0$. Recall that we assume that $E/\mathbb{Q}$ has semi-stable reduction outside of $3$ and reduction of type I$_n^*$ modulo $3$, for some $n \geq 0$. Lemma \ref{lemmatamagawanumber} implies that $r \neq 3$.
We claim that $E/\mathbb{Q}$ has split multiplicative reduction modulo $r$. Indeed, by our assumption $E/\mathbb{Q}$ is semi-stable away from $3$ and $r \neq 3$. Since $\textrm{ord}_r(b)>0$ and we know that either $r \nmid a$ or $r^3 \nmid b$, we obtain that either $3\textrm{ord}_r(a)>\textrm{ord}_r(b)$ or $\textrm{ord}_r(a)=0$. The case $3\textrm{ord}_r(a)>\textrm{ord}_r(b)$ gives that $E/\mathbb{Q}$ has reduction of type IV or IV$^*$ modulo $r$, by Part $(ii)$ of Proposition \ref{prop:3torsionreduction}, which contradicts our assumption that $E/\mathbb{Q}$ has semi-stable reduction outside of $3$. If $\textrm{ord}_r(a)=0$, then we obtain that $3\textrm{ord}_r(a)<\textrm{ord}_r(b)$ and, hence, by Proposition \ref{prop:3torsionreduction}, Part $(i)$ we get that $E/\mathbb{Q}$ has split multiplicative reduction modulo $r$.
If $3 \mid m$, then by Proposition \ref{prop:3torsionreduction}, Part $(i)$ we obtain that $3m \mid \prod_{p}c_p(E)$ and, hence, $9 \mid \prod_{p}c_p(E)$. Thus, if $3 \mid m$, then Part $(a)$ of Theorem \ref{3torsionshaandtamagawa} is satisfied. Therefore, we can assume that $3 \nmid m$ in what follows. Finally, since $3\textrm{ord}_r(a)<\textrm{ord}_r(b)$, by Theorem $4.1$ of \cite{koz} we obtain $\textrm{ord}_3(\frac{c_r(\widehat{E})}{c_r(E)})=-1$.
\begin{claim}\label{claimevenorder1}
The number ord$_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})$ is even and non-negative.
\end{claim}
\begin{proof}
There exists at least one prime $q$ different from $r$ such that $E/\mathbb{Q}$ has split multiplicative reduction modulo $q$. Indeed, by \ref{rootnumbers} $w(E)=1$, $w_r(E)=-1$, $w_3(E)=-1$, and $w_\infty(E)=-1$. Moreover, since $E/\mathbb{Q}$ is semi-stable away from $3$, by \ref{rootnumbers} for $p\neq 3$, we get that $w_p(E)=-1$ if and only if $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$. This proves that $E/\mathbb{Q}$ has an even number of places of split multiplicative reduction. Therefore, since $E/\mathbb{Q}$ has split multiplicative reduction modulo $r$, we obtain that there is a prime $q \neq r$, such that $E/\mathbb{Q}$ has split multiplicative reduction modulo $q$.
Let $p \neq 3$ be any prime such that $E/\mathbb{Q}$ has multiplicative reduction modulo $p$. If $E/\mathbb{Q}$ has nonsplit multiplicative reduction modulo $p$, then by lines $5,6$, or $7$ of Theorem $6.1$ of \cite{dd} we obtain that $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=0$.
The elliptic curve $\widehat{E}/\mathbb{Q}$ is given by a Weierstrass equation of the form
\begin{align*}
y^2+axy-9by=x^3-(a^3+27b)b,
\end{align*}
where $a,b$ are as in $($\ref{eq:3torsion}$)$ (see \cite{koz} equation $(3.2)$). The discriminant of this Weierstrass equation is $$\widehat{\Delta}=(a^3-27b)^3b.$$
We now prove that for any prime $p \neq 3,r,$ if $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$, then $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=1$. Assume that $p \neq r$ is a prime such that $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$. Note that since $b=r^m$ and $p\neq r$, $p\nmid b$. Recall that $\Delta=b^3(a^3-27b)$ and $\widehat{\Delta}=b(a^3-27b)^3$. If $\textrm{ord}_p(\Delta)=\gamma$, then $\textrm{ord}_p(\widehat{\Delta})=3\gamma$. Therefore, by line $3$ of Theorem $6.1$ of \cite{dd} we obtain that $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=1$.
Recall that there exists a prime $q$ different from $r$ such that $E/\mathbb{Q}$ has split multiplicative reduction modulo $q$. We have proved so far that $\textrm{ord}_3(\frac{c_r(\widehat{E})}{c_r(E)})=-1$, $\textrm{ord}_3(\frac{c_q(\widehat{E})}{c_q(E)})=1$, and that for any prime $p \neq 3,r$, if $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$, then $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=1$. Moreover, we have proved above that if $E/\mathbb{Q}$ has nonsplit multiplicative reduction modulo $p$, then $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=0$. Therefore, since $E/\mathbb{Q}$ is semistable away from $3$ and has an even number of places of split multiplicative reduction, we obtain that ord$_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})$ is even and non-negative, as claimed.
\end{proof}
If ord$_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)}) \geq 2$, then Part $(i)$ of Lemma \ref{tamagawagreaterorequalto2} implies that $9$ divides $ |\Sha(E/\mathbb{Q})|$.
Otherwise, ord$_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)}) = 0$. Since the degree of $\phi$ is $3$, by Proposition $2$ of \cite{ks} we obtain that ord$_s(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)}) = 0$ for every prime $s \neq 3$. Therefore, if $\frac{\Omega(E)}{\Omega(\widehat{E})}=3$, then Equation $($\ref{eq:selmerandtamagawa}$)$ implies that $\frac{|\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})|}{|\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})|}=1$ and, hence, by Part $(ii)$ of Lemma \ref{tamagawagreaterorequalto2} we obtain that $9$ divides $|\Sha(E/\mathbb{Q})|$. If $\frac{\Omega(E)}{\Omega(\widehat{E})}=1$, then, since $b=r^m$ with $3 \nmid m$, the following lemma shows that $9$ divides $|\Sha(E/\mathbb{Q})|$. This concludes the proof of Part $(a)$ of Theorem \ref{3torsionshaandtamagawa}.
\end{proof}
\begin{lemma}\label{lemmadescent}
Let $E/\mathbb{Q}$ be as in Theorem \ref{3torsionshaandtamagawa}, and given by a Weierstrass equation as in $($\ref{eq:3torsion}$)$. Assume moreover that there exists a prime $r \neq 3$ such that $b=r^m$, for some integer $m$ coprime to $3$. If $\mathrm{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})=0$ and $\frac{\Omega(E)}{\Omega(\widehat{E})}=1$, then $9$ divides $|\Sha(E/\mathbb{Q})|$.
\end{lemma}
\begin{proof}
Since $\frac{\Omega(E)}{\Omega(\widehat{E})}=1$, by Equation $($\ref{eq:selmerandtamagawa}$)$, combined with $($\ref{kernelofisogeny}$)$, we get that $$|\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})|=\frac{|\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})|\cdot 3\cdot \prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)}.$$
Since $\mathrm{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})=0$, we obtain that $\mathrm{dim}_{\mathbb{F}_3}\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})=\mathrm{dim}_{\mathbb{F}_3}\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})+1$. Therefore, if we can show that $\mathrm{dim}_{\mathbb{F}_3}\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q}) \geq 1$, then we get that $\mathrm{dim}_{\mathbb{F}_3}\mathrm{Sel}^{(\phi)}(E/\mathbb{Q}) \geq 2$ and, hence, by Part $(i)$ of Lemma \ref{tamagawagreaterorequalto2} we obtain that $9$ divides $|\Sha(E/\mathbb{Q})|$.
We now show that $\mathrm{dim}_{\mathbb{F}_3}\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q}) \geq 1$. There is a short exact sequence $$0 \longrightarrow{} E(\mathbb{Q})/\hat{\phi}(\widehat{E}(\mathbb{Q})) \longrightarrow{} \mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q}) \longrightarrow{} \Sha(\widehat{E}/\mathbb{Q})[\hat{\phi}] \longrightarrow{} 0$$ (see Theorem X.$4.2$ of \cite{aec} applied to $\hat{\phi}: \widehat{E} \rightarrow{} E$). Recall that $E/\mathbb{Q}$ has rank $0$ and that $E(\mathbb{Q})$ contains a point of order $3$. The rank of $\widehat{E}/\mathbb{Q}$ is $0$ because it is isogenous to $E/\mathbb{Q}$. Moreover, since $b$ is not a cube, Theorem $1.1$ of \cite{had} implies that $\widehat{E}(\mathbb{Q})$ does not contain a point of order $3$. Therefore, $E(\mathbb{Q})/\hat{\phi}(\widehat{E}(\mathbb{Q}))$ contains a point of order $3$ which injects into $\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})$. This proves that $\text{dim}_{\mathbb{F}_3}(\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})) \geq 1$.
\end{proof}
\begin{proof}[\it Proof of Theorem \ref{3torsionshaandtamagawa} Part $(b)$.]
We know that $E/\mathbb{Q}$ has a Weierstrass equation as in $($\ref{eq:3torsion}$)$. By Proposition \ref{prop:3torsionreduction} if $p \neq 3$ is any prime such that $E/\mathbb{Q}$ has additive reduction, then $p \mid b$, the reduction type modulo $p$ is IV or IV$^*$, and $3 \mid c_p(E)$. Thus, if $E/\mathbb{Q}$ has more than two places of additive reduction, then $9\mid \prod_{p \mid N}c_p(E)$.
\end{proof}
\begin{proof}[\it Proof of Theorem \ref{3torsionshaandtamagawa} Part $(c)$.]
Our strategy in this proof is to show that if $9$ does not divide $\prod_{p|N}c_p(E)$,
then 9 divides $|\Sha(E/\mathbb{Q})|$.
We know that $E/\mathbb{Q}$ has a Weierstrass equation as in $($\ref{eq:3torsion}$)$. By assumption $E/\mathbb{Q}$ has exactly two places of additive reduction, say at $3$ and $r$.
Assume that $9 \nmid \prod_{p|N}c_p(E)$ otherwise the theorem is proved. Our assumptions force $b=r$ or $b=r^2$, and $a=3rm$ where $m$ is an integer, as we now explain. Indeed, if $b=1$, then Parts $(i)$ and $(ii)$ of Proposition \ref{prop:3torsionreduction} imply that $E/\mathbb{Q}$ is semi-stable away from $3$ and that $E/\mathbb{Q}$ has only one place of additive reduction. If $b$ had two or more prime divisors, Parts $(i)$ and $(ii)$ of Proposition \ref{prop:3torsionreduction} imply that $9$ divides $\prod_{p|N}c_p(E)$. If $r \nmid a$, then by Proposition \ref{prop:3torsionreduction}, Part $(i)$, we obtain that $E/\mathbb{Q}$ has multiplicative reduction modulo $r$, contrary to our assumption. This proves that $r \mid a$. Since $r\mid a$ and $E/\mathbb{Q}$ has additive reduction modulo $r$, Proposition \ref{prop:3torsionreduction} implies that $r \mid b$. Moreover, since we must have that $r \nmid a$ or $r^3 \nmid b$ in Equation $($\ref{eq:3torsion}$)$, we obtain that $b$ is equal to either $r$ or $r^2$. Finally, since $E/\mathbb{Q}$ has reduction I$_n^*$ modulo $3$, Proposition \ref{prop:3torsionreduction}, Part $(iii)$ implies that $3 \mid a$.
By Proposition \ref{prop:3torsionreduction}, Part $(ii)$, $E/\mathbb{Q}$ has reduction of type IV or IV$^*$ modulo $r$. By the table on page $46$ of \cite{tatealgorithm} for $r \neq 2$ and by Tableau IV of \cite{pap} for $r=2$, we obtain then that ord$_r(\Delta)=4$ or $8$. Therefore, Proposition $2$ of \cite{rohrlich} implies that $w_r(E)=\big( \frac{-3}{r} \big)$. If $\big( \frac{-3}{r} \big)=1$, then $r \equiv 1 \; (\text{mod} \; 3)$ and if $\big( \frac{-3}{r} \big)=-1$, then $r \equiv 2 \; (\text{mod} \; 3)$. We split the proof into two cases, when $w_r(E)=1$ and $r \equiv 1 \; (\text{mod} \; 3)$, and when $w_r(E)=-1$ and $r \equiv 2 \; (\text{mod} \; 3)$.
Let us first prove the following claim.
\begin{claim}\label{claimsplitmultiplicative}
If $p$ is any prime such that $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$, then $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=1$.
\end{claim}
\begin{proof}
Note that $p \nmid 3b$ because $E/\mathbb{Q}$ has additive reduction modulo $3$ and $r$. Recall that $\widehat{E}/\mathbb{Q}$ is given by a Weierstrass equation of the form
\begin{align*}
y^2+axy-9by=x^3-(a^3+27b)b,
\end{align*}
where $a,b$ are as in $($\ref{eq:3torsion}$)$ (see \cite{koz} equation $(3.2)$). The discriminant of this Weierstrass equation is $$\widehat{\Delta}=(a^3-27b)^3b.$$ Since $\Delta=b^3(a^3-27b)$ and $\widehat{\Delta}=b(a^3-27b)^3$, if $\textrm{ord}_p(\Delta)=\gamma$, then $\textrm{ord}_p(\widehat{\Delta})=3\gamma$. Therefore, by line $3$ of Theorem $6.1$ of \cite{dd} we obtain that $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=1$. This proves that if $p$ be any prime such that $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$, then $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=1$.
\end{proof}
\underline{Case 1:} $w_r(E)=1$ and $r \equiv 1 \; (\text{mod} \; 3)$.
\begin{claim}
The number
$\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})$ is even and non-negative.
\end{claim}
\begin{proof}
First, \ref{rootnumbers} implies that $w_\infty(E)=-1$ and that $w_3(E)=-1$. Recall that $E/\mathbb{Q}$ is semi-stable away from $3$ and $r$. For $p\neq 3$ we have that $w_p(E)=-1$ if and only if $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$. Since $w(E)=1$, we obtain that $E/\mathbb{Q}$ has an even number of places of split multiplicative reduction.
Since $r \equiv 1 \; (\text{mod} \; 3)$, we obtain that $\zeta_3 \in \mathbb{Q}_r$ and since we also have that $E/\mathbb{Q}$ has reduction of type IV or IV$^*$ modulo $r$, by line $10$ of Theorem $6.1$ of \cite{dd} we obtain that $\textrm{ord}_3(\frac{c_r(\widehat{E})}{c_r(E)})=0$. If $E/\mathbb{Q}$ has nonsplit multiplicative reduction modulo $p$, then $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=0$ by line $4$ of Theorem $6.1$ of \cite{dd}. Also, If $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$, then $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=1$ by Claim \ref{claimsplitmultiplicative}. Finally, line $13$ of Theorem $6.1$ of \cite{dd} implies that $\textrm{ord}_3(\frac{c_3(\widehat{E})}{c_3(E)})=0$. Putting all those together we obtain that $\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})$ is even and greater than or equal to zero. This concludes the proof of the claim.
\end{proof}
We are now ready to conclude the proof of Case $1$. Assume first that $\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})=0.$ Since the degree of $\phi$ is $3$, we can use Proposition $2$ of \cite{ks} and obtain that ord$_p(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)}) = 0$ for every prime $p \neq 3$. Hence, if $\frac{\Omega(E)}{\Omega(\widehat{E})}=3$, then Equation $($\ref{eq:selmerandtamagawa}$)$ implies that $\frac{|\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})|}{|\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})|}=1.$ Therefore, Part $(ii)$ of Lemma \ref{tamagawagreaterorequalto2} implies that $9$ divides $|\Sha(E/\mathbb{Q})$|. If $\frac{\Omega(E)}{\Omega(\widehat{E})}=1$, then, since $b$ is equal to either $r$ or $r^2$, and we assume that $\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})=0$, Lemma \ref{lemmadescent} implies that $9$ divides $|\Sha(E/\mathbb{Q})|$. Assume now that $\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)}) \geq 2.$
Then $9$ divides $|\Sha(E/\mathbb{Q})|$ by Part $(i)$ of Lemma \ref{tamagawagreaterorequalto2}. This proves Theorem \ref{3torsionshaandtamagawa}, Part $(c)$, in the case where $w_r(E)=1$.
\medskip \underline{Case 2:} $w_r(E)=-1$ and $r \equiv 2 \; (\text{mod} \; 3)$.
\begin{claim}
The number ord$_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})$ is even and non-negative.
\end{claim}
\noindent {\it Proof.}
By \ref{rootnumbers} $w(E)=1$, $w_3(E)=-1$, and $w_{\infty}(E)=-1$. If $p \neq 3,r$ is a prime such that $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$, then $w_p(E)=-1$. If $p \neq 3,r$ is a prime such that $E/\mathbb{Q}$ has nonsplit multiplicative or good reduction modulo $p$, then $w_p(E)=1$. Therefore, since $E/\mathbb{Q}$ is semi-stable away from $3$ and $r$, we obtain that $E/\mathbb{Q}$ has an odd number of primes of split multiplicative reduction. Note that in particular $E/\mathbb{Q}$ has at least one prime of split multiplicative reduction.
If $p$ is any prime such that $E/\mathbb{Q}$ has split multiplicative reduction modulo $p$, then Claim \ref{claimsplitmultiplicative} implies that $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=1$. By line $4$ of Theorem $6.1$ of \cite{dd}, if $E/\mathbb{Q}$ has nonsplit multiplicative reduction modulo $p$, then $\textrm{ord}_3(\frac{c_p(\widehat{E})}{c_p(E)})=0$. Moreover, by line $10$ of Theorem $6.1$ of \cite{dd} we obtain that $\textrm{ord}_3(\frac{c_r(\widehat{E})}{c_r(E)})=-1$ because $r \equiv 2 \; (\text{mod} \; 3)$ implies $\zeta_3 \notin \mathbb{Q}_r$ and $E/\mathbb{Q}$ has reduction of type IV or IV$^*$ modulo $r$. Finally, line $13$ Theorem $6.1$ of \cite{dd} implies that $\textrm{ord}_3(\frac{c_3(\widehat{E})}{c_3(E)})=0$. Since there is an odd number of primes of split multiplicative reduction, and in particular at least one prime of split multiplicative reduction, we obtain that ord$_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})$ is even and non-negative. This proves our claim.
\end{proof}
We are now ready to conclude the proof of Case $2$. Assume first that $\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})=0.$ Since the degree of $\phi$ is $3$, by line $5$ Theorem $6.1$ of \cite{dd} (or Proposition $2$ of \cite{ks}) we obtain that ord$_q(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)}) = 0$ for every prime $q \neq 3$. Therefore, if $\frac{\Omega(E)}{\Omega(\widehat{E})}=3$, Equation $($\ref{eq:selmerandtamagawa}$)$ implies that $\frac{|\mathrm{Sel}^{(\phi)}(E/\mathbb{Q})|}{|\mathrm{Sel}^{(\hat{\phi})}(\widehat{E}/\mathbb{Q})|}=1$ and, hence, by Part $(ii)$ of Lemma \ref{tamagawagreaterorequalto2} we obtain that $9$ divides $|\Sha(E/\mathbb{Q})|$. If $\frac{\Omega(E)}{\Omega(\widehat{E})}=1$, since $b$ is equal to $r$ or $r^2$, and we assume that $\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)})=0$, Lemma \ref{lemmadescent} implies that $9$ divides $|\Sha(E/\mathbb{Q})|$. Assume now that $\text{ord}_3(\frac{\prod_{p}c_p(\widehat{E})}{\prod_{p}c_p(E)}) \geq 2.$
Then $9$ divides $|\Sha(E/\mathbb{Q})|$ by Part $(i)$ of Lemma \ref{tamagawagreaterorequalto2}. This proves Theorem \ref{3torsionshaandtamagawa}, Part $(c)$, in the case $w_r(E)=-1$. This completes the proof of Theorem \ref{3torsionshaandtamagawa}.
\end{proof}
\begin{remark}
The condition that $E/\mathbb{Q}$ has modulo $3$ reduction I$_n^*$ for some $n\geq0$ in Theorem \ref{3torsionshaandtamagawa} is necessary. Indeed, the conclusion of Theorem \ref{3torsionshaandtamagawa} for the elliptic curve $E/\mathbb{Q}$ with Cremona \cite{cremonabook} label 324d1 (or LMFDB \cite{lmfdb} label 324.b1) is not true. This curve has reduction II modulo $3$, reduction IV$^*$ modulo $2$, $\prod_{p}c_p(E)=3$, $\Sha(E/\mathbb{Q})$ is trivial, and $E(\mathbb{Q})\cong \mathbb{Z}/3\mathbb{Z}$.
Moreover, the condition that $E/\mathbb{Q}$ has analytic rank $0$ in Theorem \ref{3torsionshaandtamagawa} is necessary. Indeed, the conclusion of Theorem \ref{3torsionshaandtamagawa} for the elliptic curve $E/\mathbb{Q}$ with Cremona \cite{cremonabook} label 171b2 (or LMFDB \cite{lmfdb} label 171.b2) is not true. This curve has reduction I$_0^*$ modulo $3$, rank $1$, $\prod_{p}c_p(E)=6$, $\Sha(E/\mathbb{Q})$ is trivial, and $E(\mathbb{Q}) \cong \mathbb{Z} \times \mathbb{Z}/3\mathbb{Z}$.
\end{remark}
\begin{remark}\label{remarkagashestein}
Let $E/\mathbb{Q}$ be an optimal elliptic curve of analytic rank $0$. It follows from work of Agashe and Stein that the Birch and Swinnerton-Dyer conjecture, combined with the conjecture that the Manin constant is $1$, imply that {\it the odd part of $|E(\mathbb{Q})|$ divides $|\Sha(E/\mathbb{Q})| \cdot \prod_p c_p(E)$} (see the end of Section 4.3 of \cite{agashestein}). Lorenzini has proved the above statement up to a power of $3$ (see Proposition 4.2 of \cite{lor}).
Without the assumption that $E/\mathbb{Q}$ is optimal, if $E/\mathbb{Q}$ has reduction of type I$_n^*$ modulo $3$, then Theorem \ref{3torsionshaandtamagawa} proves that the odd part of $|E(\mathbb{Q})|$ divides $|\Sha(E/\mathbb{Q})| \cdot \prod_p c_p(E)$. Note however that the curve $E/\mathbb{Q}$ with Cremona \cite{cremonabook} label 14a4 has a $\mathbb{Q}$-rational point of order $3$, $\prod_p c_p(E)=2$, and $|\Sha(E/\mathbb{Q})|=1$. Thus the assumption that the curve $E/\mathbb{Q}$ is optimal is necessary for the statement that the odd part of $|E(\mathbb{Q})|$ divides $|\Sha(E/\mathbb{Q})| \cdot \prod_p c_p(E)$ to be true in general.
\end{remark}
\begin{theorem}\label{prop:dpowerof3}
Let $E/\mathbb{Q}$ be an elliptic curve of conductor $N$ with $3 \nmid N$. Let $d = - 3$ or $3$ and assume that $L(E^d,1) \neq 0$. Then
\begin{enumerate}
\item If $d=-3$, then $
|E^d(\mathbb{Q})|^2 \quad \text{divides} \quad |\Sha(E^d/\mathbb{Q})| \cdot \displaystyle\prod_{p|N}c_p(E^d), \quad \text{up to a power of } 2.$
\item If $d=3$, then $
|E^d(\mathbb{Q})|^2 \quad \text{divides} \quad |\Sha(E^d/\mathbb{Q})| \cdot \displaystyle\prod_{p|2N}c_p(E^d), \quad \text{up to a power of } 2.$
\end{enumerate}
\end{theorem}
\begin{proof}
We first claim that
$|E^d(\mathbb{Q})|^2$ divides $|\Sha(E^d/\mathbb{Q})| \cdot \prod_{p}c_p(E^d)$, up to a power of 2. By Corollary \ref{corollarytorsion}, the only primes that can divide $|E^d(\mathbb{Q})|$ are $2$ and $3$. Therefore, if $E^d/\mathbb{Q}$ does not have a point of order $3$, our claim is proved since $E^d(\mathbb{Q})$ has order a power of $2$. So assume that $E^d(\mathbb{Q})$ contains a $\mathbb{Q}$-rational point of order $3$. A theorem of Lorenzini (see \cite{lor} Proposition $1.1$) implies that if $E/\mathbb{Q}$ is an elliptic curve with a $\mathbb{Q}$-rational point of order $9$, then $9^2$ divides $\prod_{p} c_p(E)$ except for the curve that has Cremona label \cite{cremonabook} 54b3 with $\prod_{p} c_p(E)=27$. If $E/\mathbb{Q}$ is an elliptic curve with $3 \nmid N$, then $E/\mathbb{Q}$ has good reduction I$_0$ modulo $3$. Therefore, if $d=-3$ or $3$, we obtain, using Proposition $1$ of \cite{com}, that $E^d/\mathbb{Q}$ has reduction of type I$_0^*$ modulo $3$. However, the curve with Cremona label \cite{cremonabook} 54b3 does not have reduction of type I$_0^*$ modulo $3$ and, hence, it cannot be a twist by $d$ of an elliptic curve with good reduction modulo $3$. Therefore, we can assume that $E^d(\mathbb{Q})$ does not contain a point of order $9$.
Since we assume that $E^d(\mathbb{Q})$ does not contain a point of order $9$ and the only primes that can divide $|E^d(\mathbb{Q})|$ are $2$ and $3$, to prove that
$|E^d(\mathbb{Q})|^2$ divides $|\Sha(E^d/\mathbb{Q})| \cdot \prod_{p}c_p(E^d)$, up to a power of 2, it is enough to show that $9$ divides $|\Sha(E^d/\mathbb{Q})| \cdot \prod_{p}c_p(E^d)$. Since $3 \nmid N$ and, hence, $E/\mathbb{Q}$ has good reduction I$_0$, by Proposition $1$ of \cite{com}, we obtain that $E^d/\mathbb{Q}$ has reduction of type I$_0^*$ modulo $3$. If $E^d/\mathbb{Q}$ is semi-stable away from $3$, then applying Part $(a)$ of Theorem \ref{3torsionshaandtamagawa} to $E^d/\mathbb{Q}$ proves that $9$ divides $|\Sha(E^d/\mathbb{Q})| \cdot \prod_{p}c_p(E^d)$. If $E^d/\mathbb{Q}$ has more than two places of additive reduction, then by Part $(b)$ of Theorem \ref{3torsionshaandtamagawa} we obtain that $9 \mid \prod_{p}c_p(E^d)$. Finally, if $E^d/\mathbb{Q}$ has exactly two places of additive reduction, then Part $(c)$ of Theorem \ref{3torsionshaandtamagawa} implies that $9$ divides $|\Sha(E^d/\mathbb{Q})| \cdot \prod_{p}c_p(E^d)$.
\textit{Proof of $(i)$:} We just need to show that the product can be taken over all the primes of bad reduction for $E/\mathbb{Q}$. If $d=-3$, then the only prime that ramifies in $\mathbb{Q}(\sqrt{d})$ is $3$. Therefore the primes of bad reduction of $E^d/\mathbb{Q}$ consist of the primes of bad reduction of $E/\mathbb{Q}$ as well as 3. Hence, it is enough to show that $3 \nmid c_3(E^d)$. By Proposition $1$ of \cite{com} we see that $E^d/\mathbb{Q}$ has modulo $3$ reduction of type I$_0^*$ and $c_3(E^d)=1, 2$ or $4$ by page $367$ of \cite{silverman2}.
\textit{Proof of $(ii)$:} If $d=3$, then $2$ and $3$ are the only primes that ramify in $\mathbb{Q}(\sqrt{d})$. Therefore the primes of bad reduction of $E^d/\mathbb{Q}$ consist of all the odd primes of bad reduction of $E/\mathbb{Q}$ as well as $3$ and possibly $2$. By Proposition $1$ of \cite{com} we see that $E^d/\mathbb{Q}$ has modulo $3$ reduction of type I$_0^*$ and $c_3(E^d)=1, 2$ or $4$ by page $367$ of \cite{silverman2}. This completes the proof.
\end{proof}
We are now ready to prove a slightly stronger form of Agashe's Conjecture \ref{agasheconjecture}.
\begin{corollary}\label{corollaryagasheconjecture}
Let $E/\mathbb{Q}$ be an elliptic curve of conductor $N$ and let $-D$ be a negative fundamental discriminant such that $D$ is coprime to $N$. Suppose that $L(E^{-D},1) \neq 0$. Then
\vspace{-.1cm}
\begin{align*}
|E^{-D}(\mathbb{Q})|^2 \quad \text{divides} \quad |\Sha(E^{-D}/\mathbb{Q})| \cdot \prod_{p|N}c_p(E^{-D}), \quad \text{up to a power of 2.}
\end{align*}
\end{corollary}
\begin{proof}
Recall that an integer is called a fundamental discriminant if it is the discriminant of a quadratic number field. If $n$ is a square-free integer, then the discriminant of $\mathbb{Q}(\sqrt{n})$ is $n$ in the case where $n \equiv 1 \; (\text{mod}\; 4)$ and $4n$ otherwise. If $-D$ is square-free, which happens in the case where $-D \equiv 1 \; (\text{mod} \; 4)$, then Part $(ii)$ of Corollary \ref{corollarytorsion} implies that the only prime that can divide $|E^{-D}(\mathbb{Q})_{tors}|$ is $2$, except possibly for the case where $D=3$. We note that since $-1$ is not a fundamental discriminant, we must have that $D \neq 1$. If $D=3$, then Theorem \ref{prop:dpowerof3} implies the desired result.
Assume now that $-D$ is not square-free. Then we know that $-D=4n$, where $n$ is a square-free integer. Since $E^{-D}=E^n$ and $n$ is square-free, by applying Part $(i)$ of Corollary \ref{corollarytorsion} we obtain that the only prime that can divide $|E^{-D}(\mathbb{Q})_{tors}|$ is $2$, except possibly for the cases where $n=-1$ or $-3$. If $n=-3$, then since $E^{-12}=E^{-3}$, Part $(i)$ of Theorem \ref{prop:dpowerof3} implies the desired result. Therefore, in order to prove our corollary it remains to provide a proof for the case where $n=-1$, i.e., where $-D=-4$.
Assume that $D=4$. By our assumption $2 \nmid N$. We will show that $E^{-4}(\mathbb{Q})=E^{-1}(\mathbb{Q})$ can only contain points of order a power of $2$. Recall that since $L(E^{-1},1) \neq 0$, Theorem $3.22$ of \cite{darmonmodularellipticcurves} implies that the rank of $E^{-1}/\mathbb{Q}$ is zero. Since $2 \nmid N$, by Tables I and II of \cite{com} we obtain that $E^{-1}/\mathbb{Q}$ has reduction of type I$_4^*$, I$_8^*$, II, or II$^*$ modulo $2$. If $E^{-1}/\mathbb{Q}$ had a point of order $5$ or $7$, then $E^{-1}/\mathbb{Q}$ would have had multiplicative reduction modulo $2$ which is impossible. Therefore, the only primes that can divide $|E^{-1}(\mathbb{Q})|$ are $2$ and $3$. Assume that $E^{-1}/\mathbb{Q}$ has a point of order $3$, and we will arrive at a contradiction. Since $E^{-1}/\mathbb{Q}$ has a point of order $3$, $E^{-1}/\mathbb{Q}$ has an equation of the form $($\ref{eq:3torsion}$)$. By Proposition \ref{prop:3torsionreduction}, $E^{-1}/\mathbb{Q}$ has either semi-stable reduction or reduction of type IV or IV$^*$ modulo $2$. However, this is a contradiction because $E^{-1}/\mathbb{Q}$ has reduction of type I$_4^*$, I$_8^*$, II or II$^*$ modulo $2$. This proves that $E^{-1}(\mathbb{Q})$ only contains points of order a power of $2$. This concludes our proof.
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}
The random cluster model is of fundamental importance in statistical mechanics, with its special cases including percolation and the Ising and Potts models \cite{Potts52,FK1972,Wu82}, and touches a wide range of subject areas. Percolation, for example, has proven to be fertile ground for many fascinating rigorous \cite{Kesten82,Smirnov,Schramm2001} and non-rigorous \cite{Cardy92} results in mathematics, while also serving as a simple model of forest fires \cite{vonNiessen1986,Loehle1996} and epidemics \cite{Ziff2021}. The Potts model, aside from being a general model of ferromagnetism, likewise has found a wide range of surprising applications. For only one example, the three-dimensional 3-state model is used to study the heavy quark limit in lattice quantum chromodynamics \cite{Svetitsky1982,Alford2001}. Thus, the search for efficient numerical methods for the random cluster model is important, and new techniques can potentially impact many different fields.
Critical slowing down has long troubled Monte Carlo simulations of the Potts model. The Swendsen-Wang (SW) algorithm \cite{SwendsenWang}, which greatly mitigates this problem, is now the basis for most modern Potts model simulation techniques \cite{Wolff1989,Machta,Ferrenberg1988}. However, it is restricted to integer $q$ and is therefore not completely general, as there are many applications for which non-integer $q$ is interesting. For example, the range $0 \le q \le 1$ was shown to be in the same universality class as the gelation transition in branched polymers \cite{Lubensky1978}, $q=1/2$ is related to the dilute spin glass problem \cite{Wu82}, and in three dimensions the transition from a second- to first-order critical point occurs for $2 <q <3$ \cite{Grollau2001}. In addition, conformal field theory \cite{DiFrancesco} makes many predictions conjectured to be exact in the continuous range of $q$. While methods exist that handle real $q$ \cite{Sweeny1983,Elci2013,Chayes1998,Gliozzi2002}, there are often trade-offs involved either in efficiency or ease of use compared to SW.
Here, we present a Monte Carlo scheme, based not on any of these previous techniques but on the Newman-Ziff (NZ) approach to percolation. Advantages of our method are that a) it works for real $q>0$ and that b) results for an arbitrary number of different $q$ can be obtained from a single simulation. One price is that the memory requirement is somewhat steeper than other methods, scaling as $\mathrm{O}(L^3)$ for a lattice with $L$ vertices on an edge. Another, more serious, disadvantage is that as $L$ is increased, configurations that are important when $q>1$ are increasingly poorly sampled. So, although this method does not exactly have the critical slowdown problem, we are nevertheless limited in the size of the systems we can study. On the other hand, because the critical polynomial \cite{Scullard2012,ScullardJacobsen2012,JacobsenScullard2013} provides very accurate estimates of transition points, even on relatively small graphs, it can be used very effectively with the present Monte Carlo scheme. In the standard polynomial method, the polynomial, or more often just its root \cite{Jacobsen15}, is computed via an analytic technique \cite{Jacobsen2014} and, for two-dimensional planar lattices, provides accuracy far surpassing what is currently possible with Monte Carlo methods \cite{ScullardJacobsen2020}. However, relying as it does on a transfer matrix calculation, it is somewhat challenging to generalize to non-planar, not to mention higher-dimensional, lattices. Additionally, even in the planar case, it is rather more complex to implement than a typical Monte Carlo algorithm and one may prefer to trade some accuracy for simplicity.
\section{Algorithm}
In a configuration of the random cluster model, each edge of a lattice graph is chosen to be open or closed. The partition function with edge weight $p$ is given by \cite{FK1972}
\begin{equation}
Z=\sum_{\{\omega\}} p^n (1-p)^{N-n} q^{C(\omega)}
\end{equation}
where $n$ is the number of edges present in the configuration $\omega$, $C(\omega)$ is the number of connected components (including isolated vertices) in $\omega$, and the sum is over all configurations. It is convenient to write $Z$ in the form
\begin{equation}
Z=\sum_{n=0}^N \sum_{C=1}^{L^2} p^n (1-p)^{N-n} q^C \Gamma_n(C), \label{eq:Z}
\end{equation}
for a two-dimensional lattice with $L^2$ vertices and $N$ edges, and where $\Gamma_n(C)$ is the number of configurations consisting of $n$ edges and $C$ connected components. This function is obviously not known analytically in general, but it is a simple matter to estimate it during the simulation. The probability of some event $A$ is then
\begin{equation}
P(A)=\frac{1}{Z} \sum_{n=0}^N \sum_{C=1}^{L^2} p^n (1-p)^{N-n} q^C \Gamma_n(C) P(A|n,C) \label{eq:PA}
\end{equation}
where $P(A|n,C)$ is the probability of $A$ for fixed $n$ and $C$. Therefore, to compute the probability of an event $A$, the quantities $P(A|n,C)$ and $\Gamma_n(C)$ can be estimated for all $n$ and $C$ and the results combined in Equation (\ref{eq:PA}). All configurations with a given $n$ and $C$ have the same probability, and thus they might be said to constitute a microcanonical ensemble. More concretely, we have
\begin{equation}
P(A|n,C)=\frac{\Omega_n(A,C)}{\Gamma_n(C)}
\end{equation}
where $\Omega_n(A,C)$ is the number of configurations consisting of $n$ edges and containing $C$ clusters for which the event $A$ occurs. Random sampling is therefore used only to count configurations. The numerical approach encapsulated in Equation (\ref{eq:PA}) is the generalization to arbitrary $q$ of that of Newman and Ziff \cite{NewmanZiff2000,NewmanZiff} (NZ), to which this method reduces when $q=1$. Their sampling procedure will be used here as it is straightforward to implement and hardly needs any generalization for the present purposes. Although we describe it below, we recommend the reader consult their paper for more details.
The sampling algorithm for a single run works as follows. First, the lattice is empty, with all edges closed. We then add edges one by one in a completely random order until they are all open. At step $n$, we find the number of clusters, $C$, which is simple to do; adding an edge can only reduce $C$ by one or leave it unchanged, and when $n=0$ we have $C=L^2$. Performing multiple runs we then find $f_n(A,C)$, the fraction of runs for which the event $A$ occurs at a given $n$ and $C$. The estimate of the number of configurations is then
\begin{equation}
\Omega_n(A,C)={N \choose n} f_n(A,C) .
\end{equation}
The only difference between runs is the order in which the edges are added to the system. Permutations of the edge order are handled in exactly the same way as in NZ, and in fact they provide a function written in C called \verb!permutation()! which will work unchanged here. Adding all $N$ edges while tracking clusters takes a time that is essentially $\mathrm{O}(N)$ \cite{NewmanZiff}. Although we do not have independence between step $n$ and step $n+1$ within a single run, all that matters is that there be independence between runs and that all configurations with a fixed $n$ and $C$ should appear with equal probability, which is plainly the case here. We can now see how the critical slowdown and the restriction to integer $q$ are evaded by this algorithm. The configuration sampling knows nothing about the temperature (the probability $p$ in our case) or $q$, but rather these appear as parameters in a convolution, Eq. (\ref{eq:PA}), that occurs after the sampling has been done. For the same reason, we get results for every value of $q$ in a single simulation.
To identify and merge clusters we use the standard \cite{Galler1964,Sedgewick} tree-based union/find described in section IIB of NZ. Although we define configurations by their edge states, clusters are tracked only on the vertices. The key idea is that every cluster has a unique root site and that any given site can be quickly traced to its root. To achieve this, we define an array of integer pointers, \verb!ptr[i]!, which for vertex $i$ either gives its parent site in the cluster or, if $i$ is a root, its value is the negative of the size of the cluster. Thus, every site is connected by a path of parents to the root. So if an edge joins the two sites $i$ and $j$, we take the following steps
\begin{enumerate}
\item Find the roots of $i$ and $j$ by traversing their trees.
\item If $i$ and $j$ have the same root, they are in the same cluster. This situation will be important below, but as far as the cluster configuration goes we need do no more.
\item If $i$ and $j$ are different, determine which cluster is smaller. If, say, $j$ is in the smaller cluster, we make the root of $j$ point to the root of $i$ and add the size of the smaller cluster to the larger one by adjusting \verb!ptr[i]!. Obviously, if both clusters are the same size we can choose arbitrarily which cluster to add to which.
\end{enumerate}
The algorithm is most efficient if all sites point to their roots. Thus, after we have found the roots with step 1, it is worthwhile to backtrack over the path and set all pointers to point directly to the root we have just found. See NZ for a fuller discussion of this path compression.
\section{Critical polynomials}
To locate the critical point, we estimate the critical polynomial. This is defined (up to a sign) on a doubly periodic finite lattice, or basis, $B$ by \cite{JacobsenScullard2013}
\begin{equation}
P_B(p,q) \equiv P(\mathrm{2D};p,q)-q P(\mathrm{0D};p,q) \label{eq:PB}
\end{equation}
where $P(\mathrm{2D})$ is the probability that there is a cluster that wraps both dimensions and $P(\mathrm{0D})$ is the probability that no cluster wraps either. A third possibility, which we will encounter later, is referred to as 1D and is the event that a wrapping spans only one dimension. See Figure \ref{fig:wrapping} for examples of these events. That the root in $[0,1]$ of $P_B(p)$ provides an estimate for the transition point is well established \cite{ScullardZiff08,ScullardJacobsen2012,Mertens2016} and is a consequence of universality.
The advantage of using the critical polynomial, and the reason for its natural compatibility with our Monte Carlo scheme, is that it can be used with small lattices. For lattices for which the critical points can be determined analytically using a duality argument, such as the square lattice, equation (\ref{eq:PB}) will give the exact answer for any $L$. For an unsolved problem, the critical polynomial does not give the exact answer for any size of basis. However, the estimates it provides are generally very accurate even when the basis is of a size normally considered small for Monte Carlo. For example, for the kagome lattice, a basis of size $L=4$ already produces estimates of the percolation threshold accurate to six digits \cite{ScullardJacobsen2012}. By the time $L \sim 20$ a polynomial prediction for a given system (provided it does not possess some long-range non-planar features) is generally accurate enough that it cannot be checked with Monte Carlo. Thus, using these relatively small lattices is perfectly acceptable in the present approach.
Estimates are obtained by computing averages of $P(\mathrm{2D}|n,C)$, $P(\mathrm{0D}|n,C)$ and $\Gamma_n(C)$ and combining them with Eq. (\ref{eq:PA}). As a function of $p$, $P_B(p,q)$ is monotonic in $p$, taking the value $-q$ at $p=0$ and 1 at $p=1$. A simple root-finding algorithm suffices to find the estimate for $p_c$, and here we use the bisection method. By making good initial guesses, perhaps derived from smaller-sized lattices, one can minimize the number of evaluations needed by this root-finding step.
\begin{figure}
\begin{center}
\includegraphics{wrappings.eps}
\caption{The three possible wrapping events on a periodic lattice, or basis.}
\label{fig:wrapping}
\end{center}
\end{figure}
The next piece of the algorithm is the identification of the different wrapping events. This is done by using a variation of an approach first suggested by Machta et al. \cite{Machta}. For each site $i$, in addition to the pointers \verb!ptr[i]!, we also associate \verb!x[i]! and \verb!y[i]!, the $x$ and $y$ lattice displacements to the root site of the vertex $i$. Upon joining two vertices that are already in the same cluster, if at least one vertex' displacement vector points through a boundary then this operation might produce a wrapping cluster. That is, the root of the cluster can now be connected to itself via a path that winds through one or both of the periodic directions. The exact manner of this winding can be encoded in a connection vector, as shown in Figure \ref{fig:convec}. It is a simple matter to compute this vector for each join we perform. If we are joining vertices, say 1 and 2, for which the vector connecting 1 to 2 is ${\bf u}_{12}$, and their root displacement vectors are ${\bf d}_1$ and ${\bf d}_2$ then the connection vector is given by
\begin{equation}
{\bf c}={\bf d}_2-{\bf d}_1+{\bf u}_{12} \mathrm{\ mod\ } L
\end{equation}
(note that the overall sign is irrelevant). If the connection vector is zero, then we have created no new wrapping. Upon forming a 1D wrapping, we must store the connection vector associated with that cluster \footnote{As a practical matter it can be stored in the same array as the displacement vectors, as we know the displacement vector for the root is zero and its entry can be used for something else}. Call this vector ${\bf c}_0$, and if upon later joining two further vertices of this cluster we form a new connection vector, ${\bf c}_1$, then we must determine the span of the two vectors. One way is to evaluate the cross product, and if
\begin{equation}
{\bf c}_0 \times {\bf c}_1 \neq {\bf 0} \label{eq:crossprod}
\end{equation}
then we know we have just formed a 2D wrapping. In higher dimensions, one might want to determine the span in some other way, such as forming a matrix out of all the wrapping vectors associated with a cluster and transforming it into row-echelon form.
Note that if we were to restrict ourselves only to planar lattices, the state of the wrapping would be a global property. For example, if we formed a wrapping with ${\bf c}=(1,0)$ after already having encountered a cluster with ${\bf c}=(0,1)$, then these clusters are necessarily attached and we have formed a 2D wrapping. However, in a non-planar lattice (including higher-dimensional lattices) it is possible to have disconnected non-colinear wrappings. We must therefore check the cross products of connection vectors every time we join clusters, with a 2D wrapping formed whenever (\ref{eq:crossprod}) is satisfied. Similarly, whether the lattice is planar or not we must be sure that the connection vector is inherited correctly when joining a 1D wrapping to a non-wrapping cluster.
\begin{figure}
\begin{center}
\includegraphics{connect_vector.eps}
\caption{The 1D wrapping event of Figure \ref{fig:wrapping} tiled to demonstrate the meaning of the connection vector, indicated by the red arrow. The circle denotes the root of the cluster. The connection vector is $(2,-1)$ in this case.}
\label{fig:convec}
\end{center}
\end{figure}
\section{Convolution}
Finally, we need some means of numerically computing the convolution in Eq.~(\ref{eq:PA}). In the summation, the maximum $C$ is $L^2$ and $N$ is $\mathrm{O}(L^2)$, so this sum in principle contains $\mathrm{O}(L^4)$ terms for a two-dimensional lattice. In reality, not all these terms will be encountered by the sampling algorithm, as some states are either impossible or very unlikely. Rather, for each $n$ we will have data only between $C_{\mathrm{min}}(n)$ and $C_{\mathrm{max}}(n)$. We can consider either the theoretically possible $C_{\mathrm{min}}(n)$ and $C_{\mathrm{max}}(n)$ or the actually sampled $C_{\mathrm{min}}(n)$ and $C_{\mathrm{max}}(n)$, which will be a smaller range. Eliminating the impossible states from the summation divides the number of $(n,C)$ states only by a constant factor. Eliminating unsampled states (theoretically possible but very unlikely) gives a much larger improvement. To realize this improvement in memory and run time, sampled data is stored in a linked list instead of an array, so that we are only allocating memory for terms that have actually appeared. We then computationally find that the memory requirement is $\mathrm{O}(L^3)$. This memory requirement also follows theoretically from McDiarmid's bounded difference inequality, which applied here gives that the tail distribution on the probability of $C$ drops off like $e^{-O(C^2)}$ for $p\in(0,1)$ and $q=1$ \cite{McDiarmid}.
We also need to calculate binomial-type terms of the form
\begin{equation}
B(N,n,C;p,q) \equiv {N \choose n} p^n (1-p)^{N-n}q^C . \label{eq:binom}
\end{equation}We can not evaluate (\ref{eq:binom}) directly, as individual terms may be very large. We follow NZ here by setting the largest term, at a point we denote $(n',C')$, to 1 and calculating all other terms relative to that with recursion. When $q=1$, the largest term occurs at $n'=N p$. However, for general $q$, we must take into account how $q^C$ skews the distribution of likely $n$. To find $(n',C')$, we can iterate through all sampled $(n,C)$ values, renormalize the maximum found so far to 1, and calculate further $B$ values by the recursive formulas that will be introduced below. In practice, we only need to iterate over one edge of the list, as for $q\le 1$ ($q\geq1$), the largest value of $B$ will come at the lowest (highest) sampled $C$ value for a given $n$, and thus this iteration will be $\mathrm{O}(L^2)$.
We assign this point $(n',C')$ to have a $B$ value of 1. We then use recursion to fill in the other terms:
\begin{equation}
B(N,n+1,C;p,q) = \frac{p(N-n)}{(1-p)(n+1)}B(N,n,C;p,q)
\end{equation}
\begin{equation}
B(N,n,C+1;p,q)=qB(N,n,C;p,q)
\end{equation}
In this way, we can compute all the relevant binomial coefficients subject to our arbitrary normalization. As long as we use these same coefficients to compute $Z$ in Eq.~(\ref{eq:Z}), Eq.~(\ref{eq:PA}) will be normalized correctly.
\section{Run-Time and Memory}
Our algorithm has three phases. The first is the sampling phase, which does not depend on $p$ or $q$. In this, we can run $t$ Newman-Ziff trial runs, each of which adds $\mathrm{O}(L^2)$ edges in averaged constant time each. Thus, this step takes approximately $\mathrm{O}(L^2t)$ time, though in practice it is more like $\mathrm{O}(L^{2.5}t)$ due to memory requirements. The linked list in which we store our data will have $\mathrm{O}(L^3)$ nodes, and thus we have an $\mathrm{O}(L^3)$ memory requirement, which the bounded difference inequality again shows will have a $t$-dependence of $\mathrm{O}(L^3\sqrt{\log(t)})$ (reflecting the fact that the size of the list must increase, although only slowly, as more rare configurations appear).
The next step is that, for a given $q$, we can do a pre-summation to reduce our run time and memory. Since the inner sum of Eq.~(\ref{eq:PA}) only depends on $q$, we can compute it before we know the $p$ value. In other words, we collapse the $(n,C)$ linked list to two arrays of length $N+1$, which store the sum and normalization for that $n$. To avoid overflow or underflow, we normalize to the highest sampled $C(n)$ (if $q\ge 1$, or lowest if $q\le 1$), which we also store in a third array for correct normalization in the post-summation step. This step requires $\mathrm{O}(L^3)$ time and reduces our memory from $\mathrm{O}(L^3)$ to $\mathrm{O}(L^2)$.
Then, we do a root-finding algorithm for this $q$ to find the critical $p$. The post-summation step for a given $(p,q)$ pair takes only $\mathrm{O}(L^2)$ time to sum over the arrays from the pre-summation step.
If the list of $q$ we are interested in is specified in advance, then it is possible to reduce the memory of the entire to $\mathrm{O}(L^2)$ by combining the sampling and pre-summation steps. This does require renormalizing as we sample based off the sampled $C(n)$, but takes approximately the same amount of time as the normal sampling algorithm.
\section{Results}
In Table \ref{tab:square}, we present results on the square lattice for various $q$, where the critical points are known to be \cite{Baxter1973}:
\begin{equation}
p_c(q)=\frac{\sqrt{q}}{1+\sqrt{q}}.
\end{equation}
As mentioned earlier, if the critical point of a lattice is known exactly, the critical polynomial will find the exact solution for any size of basis. To illustrate this, calculations are presented for both $L=3$ using $10^{10}$ runs and $L=16$ using $10^9$. Although both are in agreement with the exact solution it is clear that for $L=16$ and $q>1$, accuracy suffers somewhat because of estimates that rely on some portion of $(n,C)$ space that is relatively difficult to resolve. We would expect this situation to improve eventually as $q$ becomes large because the probabilities become more weighted toward the least-occupied configurations.
As an example of an unsolved problem, we calculate critical points for the matching graph of the square lattice, the $L=3$ basis of which is shown in Figure \ref{fig:sq_match}. This is a non-planar lattice, with the diagonals in each face crossing but not directly connected to each other. Its site percolation threshold is $1-p_c(\mathrm{square\ site}) \approx 0.4073$, unsolved but known numerically to high precision \cite{Jacobsen15}. Thresholds are reported in Table \ref{tab:sq_match} for various $q$ including that for $q=1$, the bond percolation threshold, which has previously been found numerically by several authors \cite{FengDengBlote08, Ouyang2018} and very recently by Xu et al. \cite{Xu2021} using a Monte Carlo-critical polynomial technique. Their value is $0.250\,368\,40(4)$. Because none of these cases is exact, we expect the estimates to depend on $L$, and we used a few different sizes around $L=25$ just to confirm that the same predictions were made. These results were found with $10^8$ total runs, far short of what one might normally use but our goal is only to demonstrate that the algorithm works, not to present values of definitive accuracy. Note that one may instead want to do computations on smaller lattices and then extrapolate to infinite $L$, as this may be the best way to maximize the accuracy of the estimates. There is, however, not yet a good theory governing this extrapolation, and scaling exponents, which are known to be lattice-dependent \cite{Jacobsen15,ScullardJacobsen2020}, must be calculated empirically, so we do not explore this here.
\begin{figure}
\begin{center}
\includegraphics{sq_match.eps}
\caption{The $L=3$ basis for matching graph of the square lattice.}
\label{fig:sq_match}
\end{center}
\end{figure}
\begin{table}
\begin{center}
\begin{tabular}{l|l|l|l}
$q$ & $L=3$ & $L=16$ & exact \\ \hline
1.0 & 0.5000002(7) & 0.500000(2) & 1/2 \\
1.5 & 0.5505103(6) & 0.550511(3) & 0.5505102572...\\
2.5 & 0.6125740(5) & 0.61252(5) & 0.6125741133...\\
3.5 & 0.6516683(4) & 0.6515(3) & 0.6516685226...\\
9.5 & 0.7550342(2) & 0.7552(2) & 0.7550344704...\\
10.0 & 0.7597467(2) & 0.7599(2) &0.7597469266...\\
\end{tabular}
\caption{Potts critical points, $p_c$, on the square lattice computed with the Monte Carlo algorithm for bases of size 3 and 16. These both agree with the known exact solution, as is standard for critical polynomials. These estimates are found by calculating averages of $P(\mathrm{2D}|n,C)$, $P(\mathrm{0D}|n,C)$, and $\Gamma_n(C)$, combining these with Eq. (\ref{eq:PA}) to get $P(\mathrm{2D})$ and $P(\mathrm{0D})$ and numerically finding the root of equation (\ref{eq:PB}). Results for all $q$ are found in a single simulation.}
\label{tab:square}
\end{center}
\end{table}
\begin{table}
\begin{center}
\begin{tabular}{l|l|l|l}
$q$ & $p_c$ & $q$ & $p_c$ \\ \hline
1.0 & 0.250368(2) & 7.0 & 0.4473(2) \\
1.5 & 0.28812(3) & 7.5 & 0.4546(2) \\
2.0 & 0.3164(4) & 8.0 & 0.46141(9) \\
2.5 & 0.3388(5) & 8.5 & 0.46778(9) \\
3.0 & 0.3577(5) & 9.0 & 0.47376(9) \\
3.5 & 0.3739(4) & 9.5 & 0.47940(9) \\
4.0 & 0.3881(3) & 10.0 & 0.48472(9) \\
4.5 & 0.4005(3) & 10.5 & 0.48976(9) \\
5.0 & 0.4117(3) & 11.0 & 0.49455(8) \\
5.5 & 0.4217(3) & 11.5 & 0.49910(8) \\
6.0 & 0.4310(3) & 12.0 & 0.50345(8) \\
6.5 & 0.4395(2) & & \\
\end{tabular}
\caption{Critical points, $p_c$, for the random cluster on the square matching lattice.}
\label{tab:sq_match}
\end{center}
\end{table}
\section{Discussion}
Although we have only studied the square and square matching lattices here, our algorithm works for all planar and non-planar periodic lattices and easily generalizes to higher dimensions. Its adaptability also gives it the potential to be used for related models, such as the Ashkin-Teller model \cite{AshkinTeller}, though for some generalizations such as multisite occupancy \cite{multisite}, the issue of a large number of disallowed edge configurations may hinder the algorithm's effectiveness. Our algorithm is also fast, as it can sample in $\mathrm{O}(L^{2.5}t)$ time for $t$ trials on an $L \times L$ lattice, and then find the critical point for each $q$ in $\mathrm{O}(L^3)$ time. The $\mathrm{O}(L^3\sqrt{\log{t}})$ memory requirement is steeper than some other algorithms, but this is only for the most general calculation; if a list of $q$ is specified in advance, the memory can be reduced to $\mathrm{O}(L^2)$. However, we also note that, because there is not much need to use lattices greater than $L \sim 30$ in 2D, memory might not be much of a consideration. A straightforward approach using arrays instead of a linked list carries an $\mathrm{O}(L^4)$ memory cost, and some might find this acceptable in exchange for a simpler implementation.
The main weakness of our algorithm is that, because we have separated the factor $q^C$ from the sampling and made $q$ a convolution parameter, some of the unsampled, ``unlikely'' configurations might prove to be important when their true weight is taken into account. Because our sampled region is of size $\mathrm{O}(L^3\sqrt{\log{t}})$ while the total possible region is of size $\mathrm{O}(L^4)$, this problem becomes more serious for large $L$. Increasing $t$ mitigates this problem, but only with the small effect of $\sqrt{\log{t}}$. How serious this is depends on the quantity being studied. For example, we have found that it is very difficult to accurately calculate the partition function, or free energy, for $q>1$ and large $L$, but the root of the critical polynomial still gives reasonable approximations to the critical point for $L=300$ (although such large calculations are hardly necessary). Thus, we mainly recommend the algorithm for finding critical points of the random cluster model, as it appears to be very useful for that purpose particularly if one wants to study multiple values of $q$ simultaneously.
The aforementioned problem is akin to the one seen in Hu's algorithm \cite{Hu1992}. There, sampling was done for $q=1$ and another value of $q$ could be obtained by a mapping which was subsequently \cite{Heringa1993} shown to behave poorly for large lattices. Although our method is very different, our sampling gives equal weight to all configurations with the same $n$, which is best suited to $q=1$. The ideal solution would be to find a sampling method that gives equal weight to all configurations with a given $(n,C)$, but where the weights of a state with $(n,C)$ and $(n,C+1)$ differ by a factor $1/q$. Unfortunately, it seems difficult to do this in an efficient manner. Wang, Kozan, and Swendsen \cite{Wang2002,Wang2003,WangConf} give a simple survival process that satisfies these criteria but they point out its extreme inefficiency when $q>1$. To address this, they devise a clever technique they call ``binary tree summation'', which is able to fill in a larger range of $(n,C)$ than our algorithm but is also limited to small graphs and is otherwise not very general; for example, it would require significant modification to be able to measure wrappings to compute the critical polynomial. It remains to be seen whether a simple alternative exists that can fully extend the advantages of the Newman-Ziff algorithm to the random cluster model.
\section{Conclusion}
We have presented a Monte Carlo algorithm for the random cluster that, in the spirit of the Newman-Ziff algorithm for percolation, separates the edge and cluster weights $p$ and $q$ from the configuration sampling. The result is a method that can be used to calculate results corresponding to multiple values of real $q>0$ in a single simulation. Although it is generally limited to small graphs, using it to calculate the critical polynomial results in good estimates for the critical point, which for non-integer $q$ might be difficult to calculate with other methods.
\section{Acknowledgments}
Part of this research was performed while the authors were at the Institute for Pure and Applied Mathematics under funding from the National Science Foundation grant DMS-1925919. We thank IPAM and UCLA for hosting, and Susana Serna for her program leadership. We also thank Sina Zareian (Claremont Graduate University) for helpful conversations, advice, and resources. CRS thanks Robert Ziff for a stimulating and helpful discussion of this problem, and Martin Weigel for useful comments. THB is supported by NSF Graduate Research Fellowship Grant Nos. DGE-1745016 and DGE-2140739. This work was partially performed under the auspices of the U.S. Department of Energy at the Lawrence Livermore National Laboratory under Contract No.\ DE-AC52-07NA27344 and was supported by the LLNL-LDRD Program under Project No.\ 19-DR-013.
\section*{References}
\bibliographystyle{iopart-num}
|
\section{INTRODUCTION} \label{sec:intro}
Real-time detection and tracking of circular and elliptic shapes using an onboard vision system are essential for many real-world (mainly robotics) applications. For example, many aerial robot' landing zones consist of an elliptical shape~\cite{cargozone}, and autonomous cars need to detect the circular road signs~\cite{roadsign}.
Detecting ellipses in images has been a topic of interest for researchers for a long time \cite{tsuji1977detection, tsuji1978detection}. In general, it is possible to classify the available ellipse detection approaches into four classes.
The first class contains voting-based algorithms, including Hough Transform (HT) \cite{cite:hough} and the methods based on it. The HT algorithm uses a 5-dimensional parametric space for the ellipse detection task and is too slow for real-time applications. Other methods try to enhance HT by reducing the dimensionality in parametric space \cite{cite:houghdim1, cite:houghdim2, cite:zhang}, performing piecewise-linear approximation for the curved segments \cite{cite:houghapprox1} or randomizing the method \cite{cite:houghrand1, cite:houghrand2, cite:houghrand3}. Some of these modified HT-based methods are fast but generally less accurate and unsuitable for many robotics applications.
The second class contains optimization-based methods. Most popular methods convert the ellipse fitting problem into a least-squares optimization problem and solve the problem to find the best fit \cite{cite:fitls, fitzgibbon1995buyer}. These methods are generally not robust and tend to produce many false positives. However, their high processing speed is useful for quickly estimating the ellipse as the first step in other methods. Another group of optimization-based methods try to solve the non-linear optimization problem of fitting an ellipse using a genetic algorithm \cite{cite:genetic1, cite:genetic2}. These algorithms perform well in noisy images with multiple ellipses, but their processing time makes them impractical for real-time applications.
The third class consists of methods based on edge linking and curve grouping \cite{cite:mai}. These methods can detect ellipses from complex and noisy images but are generally computationally expensive and cannot be used in real-time applications.
The last class contains the methods that use an ad-hoc approach or combine the methods from the first three classes to achieve a faster and more accurate ellipse detection. A real-time method proposed by Nguyen et al. \cite{cite:nguyen} detects arc-segments from the image and groups them into elliptical arcs to estimate the ellipse parameters using a least-square optimization. A method proposed by Fornaciari et al.~\cite{fornaciari2014ellipsedetect} combines arc grouping with Hough Transform in a decomposed parameter space to estimate the ellipse parameters. In this way, it achieves a real-time performance comparable to slower, more robust algorithms.
Currently, in many robotics applications, the detection of the previously-known elliptic objects is done by general learning-based object detection methods~\cite{bhattacharya2021}. However, this approach requires collection and enhancement of large amount of data for each individual object and access to powerful computing systems for training. Recently, there has been an effort to implement a learning-based method specifically for ellipse detection~\cite{cnnellipse}. The method has been shown to work well for occlusions and partial views. However, it requires a powerful processor and is not yet suitable for real-time execution on onboard computers.
While many ellipse detection algorithms are proposed for computer vision tasks (\cite{WANG2021107741, Liu2020, 8972900}), the performance of these methods drops when used in real-world robotics tasks. Some of the challenges in these applications include:
\begin{itemize}[leftmargin=*]
\item The algorithm should work online (with a frequency greater than 10 Hz), generally on the robot's resource-limited onboard computer.
\item The elliptical object should be detected when it is far from or close to the robot.
\item The shape of the ellipse is transformed by a projective distortion, which occurs when the shape is seen from different views.
\item There is a wide range of possible illumination conditions (e.g., cloudy, sunny, morning, evening, indoor lighting).
\item Due to reflections (e.g., from sources like sun, bulbs), the pattern may not always be seen in all the frames, even when the camera is close.
\item In some frames, there may be shadows on the pattern (e.g., the shadow of the robot or trees around).
\item In some frames, the pattern may be seen only partially (due to occlusion or being outside of the camera view).
\item In some frames, the pattern may not be present.
\end{itemize}
Different approaches have been devised in the literature to detect and track an elliptic pattern in real applications to address the challenges above. For example, \cite{beul2018} uses the circular Hough-transform algorithm for the initial detection, which is a slow algorithm, and then uses other features in their pattern for the tracking. The authors of~\cite{jin2018} have developed a Convolutional Neural Network trained with over 5,000 images collected from the pattern moving at a maximum speed of $15$\,km/h at various heights to select (validate) the desired ellipse from all the ellipses detected using the method in~\cite{jia2017}. The authors of~\cite{li2018} use the method proposed in~\cite{fornaciari2014ellipsedetect} for ellipse detection only when their robot is far from the pattern and switch to other features in their pattern for the closer frames.
It is possible to use a predictor filter along with the detector to narrow down the region of interest (ROI) and improve the detection speed of ellipses across the frames. The authors of~\cite{ellipsekalman1} and \cite{ellipsekalman2} use Kalman filter to predict the direction of the ellipse movement in the frames, increase the detection speed and reduce the false positives. In many robotics applications, the combined simultaneous motions of the elliptic pattern, the robot, and the camera complicate the prediction provided by such predictors, resulting in the loss of tracking on many frames, triggering a detection on the entire frame and slowing the tracking speed.
This paper proposes a novel ellipse detection method that performs on resource-limited onboard computers in real-time. Our proposed ad-hoc method first extracts all contours from the input image and then uses a least-square method to fit an ellipse to each contour. It tests how well the estimated ellipse fits the contour and starts rejecting the contours using several criteria. The remaining contours are accepted as resulted ellipses, and if there are no contours left (e.g., due to the absence of ellipses in the frame), the algorithm returns an empty set of ellipses. The detection method is combined with a simple tracking algorithm that changes detection parameters as necessary and significantly increases the elliptical pattern detection's precision and performance. This simple tracker has been shown to be faster than using predictor filters due to the reduced tracking losses. The resulting detector and tracking method can deal with lighting variations and detect the pattern even when the view is partial. By comparing the results of our experiments on a collected dataset to the other methods tested on similar datasets we show that our method outperforms all the other real-time ellipse detection methods proposed so far.
Section~\ref{sec:approach} explains the developed method for detection and tracking of the elliptical targets; Section~\ref{sec:tests} shows an example real-time application that uses the proposed algorithm and compares the performance with some other available well-known methods. Finally, Section~\ref{sec:conclusion} discusses how the proposed method can be further improved in the future.
\section{ELLIPSE DETECTION AND TRACKING} \label{sec:approach}
The idea of the proposed ellipse detection algorithm is to fit ellipses to all the contours in a frame or the region of interest and then decide if the ellipse is a good fit. With real-time ellipse fitting algorithms and fast criteria for checking the fit, the result is a real-time detection algorithm that can detect ellipses as long as the elliptic pattern's contours are (at least partially) extracted during the contour extraction. With a suitable contour extraction method, the whole detection becomes robust to the lighting and environmental changes.
The pseudocode for the proposed ellipse detector is shown in Algorithm~\ref{alg:detection}.
\begin{algorithm}[!t]
\caption{Proposed approach for ellipse detection}
\label{alg:detection}
\begin{algorithmic}[1]
\LineComment {This function detects all the ellipses found using the input criteria (thresholds)}
\Function{DetectEllipses}{\textit{frame}, \textit{thresholds}}
\LineComment {Initialize an empty set for the detected ellipses}
\State $\textit{Detections} \gets \text{$\emptyset$}$
\State
\LineComment {Detect all edges in the frame}
\State $\textit{edges} \gets \text{\textsc{DetectEdges}(\textit{frame})}$
\State
\LineComment {Extract contours from the detected edges}
\State $\textit{contours} \gets \text{\textsc{ExtractContours}(\textit{edges})}$
\State
\ForAll {$\textit{c} \in \textit{contours}$}
\LineComment {Reject if the contour is too small}
\IIf {$|c.pixels| < minContourSize$} \Continue \EndIIf
\State
\LineComment {Fit an ellipse to each contour}
\State $\textit{ellipse} \gets \text{\textsc{FitEllipse}(\textit{c})}$
\State
\LineComment {Reject ellipses with unreasonable dimensions}
\IIf {$ellipse.largeAxis > mxAxSize$} \Continue \EndIIf
\IIf {$ellipse.smallAxis < mnAxSize$} \Continue \EndIIf
\State $\textit{axisRatio} \gets \textit{ellipse.largeAxis / ellipse.smallAxis}$
\IIf {$axisRatio > maxAxisRatio$} \Continue \EndIIf
\State
\LineComment {Reject contours with small overlap with ellipse}
\State $\textit{overlap} \gets \textit{$c.pixels \ \cap \ ellipse.pixels$}$
\IIf {$|\textit{overlap}| \textit{ / } |\textit{c.pixels}| \text{ is small}$} \Continue \EndIIf
\State
\LineComment {Reject ellipses with small overlap with the edges}
\State $\textit{overlap} \gets \textit{$ellipse.pixels \ \cap \ edges.pixels$}$
\IIf {$|\textit{overlap}| \textit{ / } |\textit{ellipse.pixels}| \text{ is small}$} \Continue \EndIIf
\State
\LineComment {Add the detected ellipse to the set of detections}
\State $\textit{Detections} \text{.Insert(\textit{ellipse})}$
\EndFor
\State
\LineComment {Return the detections after all contours are processed}
\State \Return $\textit{Detections}$
\EndFunction
\end{algorithmic}
\end{algorithm}
The detection function receives a frame and a set of threshold values used in the function and returns a set of detected ellipses. A step-by-step explanation of the algorithm is as follows:
\begin{enumerate}[leftmargin=*]
\item The first step of the algorithm is to extract edges from the input frame and create the edge image. We used the Canny edge detector \cite{canny1986computational}, considering that the thresholds should be selected carefully to extract suitable edges in a large variety of conditions (e.g., illumination) while preventing the generation of too many edges. Usually, it is beneficial for smaller targets to produce more edges; this action will increase the processing time but reduces the probability of not having an edge for the elliptic target.
\item The resulted edges are utilized to extract contours using the algorithm proposed by Suzuki and Abe~\cite{suzuki1985topological}. This step helps make connections between relevant edges and enables the extraction of shapes in a frame. Ideally, each contour is a collection of connected points constructing a shape's border in the edge image.
\item Each contour is processed individually to determine if it is a part of an ellipse or not. For robotics tasks, an ellipse can have one of the following contour types:
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{figures/circletypes.png}
\caption{Contour types of ellipses in different conditions: (a) A single contour containing a full ellipse. (b) A single contour containing a full ellipse with additional connected contours from the rest of the pattern. (c) Multiple contours, together constructing a full ellipse. (d) A single contour containing an ellipse that is partially occluded by another object. (e) A single contour containing a partially-seen ellipse. (f) A single contour containing a full ellipse with additional connected contour branches that is partially seen in the frame.}
\label{fig:circletypes}
\end{figure}
\begin{itemize}[leftmargin=*]
\item A single contour containing a full ellipse without any occlusions, broken segments or additional contours (Fig.~\ref{fig:circletypes}(a)).
\item A single contour containing an ellipse with additional connected contours from the rest of the pattern (Fig.~\ref{fig:circletypes}(b)).
\item Multiple contours, together constructing a full ellipse (Fig.~\ref{fig:circletypes}(c)).
\item A single contour containing an ellipse that is partially occluded by other objects (Fig.~\ref{fig:circletypes}(d)).
\item A single contour containing an ellipse that is partially seen in the frame (Fig.~\ref{fig:circletypes}(e)).
\item A combination of the above contour types (Fig.~\ref{fig:circletypes}(f)).
\end{itemize}
In order to correctly detect the above contour types, the following process is performed on each contour:
\begin{enumerate}[leftmargin=*]
\item If a contour has a very small number of pixels, it is ignored since it is most probably just noise.
\item An ellipse is fit to the contour using the least-square approximation method described by Fitzgibbon and Fisher \cite{fitzgibbon1995buyer}. The method fits an ellipse to any input contour; therefore, the contour should be further processed to determine the actual ellipses.
\item The resulting ellipse will be ignored if any of the axes are too small or if the ellipse's eccentricity is high (close to 1). In high eccentricity, the resulting ellipse is similar to a line and is not really an ellipse.
\item The current contour is intersected with the perimeter of the resulted ellipse, described by its center point, minor and major axes, and rotation angle. Then, the ratio of the number of pixels in the intersection to the number of all pixels in the contour is calculated:
%
\begin{equation}
ContourOverlap = \frac{|Contour \ \cap \ Ellipse|}{|Contour|},
\label{eq:01}
\end{equation}
where $| \cdot |$ is used for the number of pixels. A low result means that the contour and the resulted ellipse do not fit well, and a significant portion of the contour is not lying on the fitted ellipse. In this case, the contour is ignored and not further processed.
\item Finally, the ellipse is intersected with the edges, and the ratio of the number of pixels in the intersection to the number of all the pixels in the ellipse is calculated:
\begin{equation}
EllipseOverlap = \frac{|Ellipse \ \cap \ Edges|}{|Ellipse|},
\label{eq:02}
\end{equation}
where $| \cdot |$ is used for the number of pixels. A low result means that a significant portion of the ellipse does not correspond to any contours in the image. The reason that the ellipse is intersected with the edge image instead of only its constructing contour is that due to noise or imperfect contour detection step, sometimes an ellipse is broken into two or more contours (e.g., the cases like Fig. \ref{fig:circletypes}(c)). In these cases, checking the ellipse against a single contour will give a low ratio and results in false negatives. To take care of the cases similar to Figure~\ref{fig:circletypes}(e), it is essential to count only the ellipse pixels that are actually lying in the image; otherwise, the result will be too low, and a partially viewed ellipse may get rejected. Additionally, to make the algorithm more robust to slightly imperfect ellipses, increasing the edges' thickness before intersecting them with the ellipse is beneficial.
\item If an ellipse is not rejected in the previous steps, it represents a real ellipse in the image and is added to the set of detected ellipses.
\end{enumerate}
\item Due to the target ellipse's thickness, there is a chance of detecting two or more concentric ellipses. Therefore, the returned set of detected ellipses in the frame is further processed to find the concentric ellipses. All the non-concentric sets of ellipses are ignored when this happens.
\end{enumerate}
The proposed algorithm can also detect the ellipses with partial occlusion or the ellipses exceeding the image boundaries. The rejection criteria can be chosen in a way to accept ellipses in such cases.
Choosing higher rejection criteria for ellipse detection generally helps to eliminate potential false positives. Therefore, it is beneficial to have higher rejection thresholds when there is no prior information about the pattern location and size in the image. However, after the first detection of the elliptical pattern, it is possible to change the initial parameters and conditions for the detection to enhance the performance and increase the detection rate.
Decreasing the detection threshold values reduces the probability of losing the target ellipse in the following frames due to illumination variation, noise, occlusion, or other changes in the conditions. Additionally, if the approximate movement of the target is known, setting the region of interest (ROI) to the area where the pattern is expected to be seen will decrease both the processing time and the probability of falsely detecting other similar shapes that were rejected in the initial frame due to the higher thresholds.
Furthermore, whenever the detected target is large enough to be detected in the input frame with a smaller scale, the frame can be scaled down to increase the processing speed. Performing ellipse detection on the smaller frame takes less CPU time and is much faster.
We propose these steps that can be combined with the ellipse detection algorithm to track the detected elliptic pattern in the next frames more efficiently:
\begin{enumerate}[leftmargin=*]
\item Significantly decrease the $ContourOverlap$ and $EllipseOverlap$ threshold values. This threshold change increases the detection rate and helps the algorithm to keep tracking the target.
\item Determine the region of interest, which includes the current detected target and expands in all directions based on the distance from the target, the robot's relative speed and the pattern, and other available information. For example, if the distance is far and the relative linear and angular speeds are low, the target is expected to be found close to the current detection coordinates in the next frame.
\item Decreasing the scale of the frame by order of two (up until a set threshold) every time that the detected target is larger than a specified size and scaling the frame up again by order of two (up to the actual frame scale) every time the detected target is smaller than a chosen threshold. To make the approach more robust, ellipse detection with initial higher parameters is performed once again on a higher scale if no candidate targets are found on a lower scale. The scale change is performed to reduce the execution time, as the algorithm needs to process only a quarter number of pixels every time it scales the frame down.
\end{enumerate}
\begin{algorithm}
\caption{Proposed approach for elliptic target tracking}
\label{alg:tracking}
\begin{algorithmic}[1]
\LineComment {This function reads the frames from a video stream and performs tracking of a target}
\Function{TrackTarget}{\textit{videoStream}}
\LineComment {Set frame scale, ROI and tracking status}
\State \textit{isTracking} $\gets$ \textbf{false}
\State \textit{scale} $\gets$ 1
\State \textit{roi} $\gets$ $\emptyset$
\State
\While {videoStream $\neq$ $\emptyset$}
\State \textit{frame} $\gets$ \textsc{ReadFrame}(\textit{videoStream})
\LineComment {Detect the ellipse in the frame or ROI}
\If {\textit{isTracking} $=$ \textbf{false}}
\State \textit{offsetTarget} $\gets$ \textsc{DetectTarget}(\textit{frame}, \textit{scale})
\Else
\State \textit{offsetTarget} $\gets$ \textsc{DetectTarget}(\textit{roi}, \textit{scale})
\EndIf
\State
\If {\textit{offsetTarget} $=$ $\emptyset$ \textbf{and} \textit{scale} $>$ 1}
\LineComment {If target not found, try on higher frame scale}
\State \textit{scale} $\gets$ \textit{scale} $\div$ 2
\If {\textit{isTracking} $=$ \textbf{false}}
\State \textit{offsetTarget} $\gets$ \textsc{DetectTarget}(\textit{frame}, \textit{scale})
\Else
\State \textit{offsetTarget} $\gets$ \textsc{DetectTarget}(\textit{roi}, \textit{scale})
\EndIf
\Else
\LineComment {If found, set the proper scale for next frame}
\State \textit{isTracking} $\gets$ \textbf{true}
\If {\textit{offsetTarget.size} $>$ \textit{maxTargetSize}}
\If {\textit{scale} $<$ \textit{maxScale}}
\State \textit{scale} $\gets$ \textit{scale} $\times$ 2
\EndIf
\ElsIf {\textit{offsetTarget.size} $<$ \textit{minTargetSize}}
\If {\textit{scale} $>$ 1}
\State \textit{scale} $\gets$ \textit{scale} $\div$ 2
\EndIf
\EndIf
\EndIf
\State
\LineComment {Compensate the target offset resulted from detection in ROI. Do nothing if ROI is null}
\State \textit{target} $\gets$ \textsc{CompensateOffset}(\textit{offsetTarget}, \textit{roi})
\State
\LineComment {Set the tracking status and ROI for the next frame}
\If {\textit{target} $=$ $\emptyset$}
\State \textit{isTracking} $\gets$ \textbf{false}
\State \textit{roi} $\gets$ $\emptyset$
\Else
\State \textit{isTracking} $\gets$ \textbf{true}
\LineComment {Expand the area around the detected target}
\State \textit{roi} $\gets$ \textsc{CalculateROI}(\textit{target})
\EndIf
\State
\State \Return $\textit{target}$
\EndWhile
\EndFunction
\end{algorithmic}
\end{algorithm}
\begin{algorithm}
\caption{Proposed approach for elliptic target detection in a given frame scale}
\label{alg:targetdetection}
\begin{algorithmic}[1]
\LineComment {This function detects the target in the input frame at the specified frame scale}
\Function{DetectTarget}{\textit{frame}, \textit{scale}, \textit{isTracking}}
\State
\LineComment {Resize the frame to make the processing faster}
\State \textit{scaledFrame} $\gets$ \textsc{ScaleImage}(\textit{frame}, 1 / \textit{scale})
\State
\LineComment {Set detection parameters based on inputs}
\State \textit{thresholds} $\gets$ \textsc{DetermineParams}(\textit{isTracking}, \textit{scale})
\State
\LineComment {Detect ellipses using the specified thresholds}
\State \textit{detections} $\gets$ \textsc{DetectEllipses}(\textit{scaledFrame}, \textit{thresholds})
\State
\LineComment {Select the actual target from the detected ellipses}
\State \textit{scaledTarget} $\gets$ \textsc{SelectTarget}(\textit{detections})
\State
\LineComment {Rescale the detected target to original scale}
\State \textit{target} $\gets$ \textsc{ScaleSize}(\textit{scaledTarget})
\State
\LineComment {Return the detected target (or $\emptyset$ if not detected)}
\State \Return $\textit{target}$
\EndFunction
\end{algorithmic}
\end{algorithm}
Algorithms~\ref{alg:tracking} and~\ref{alg:targetdetection} show the described method in more detail.
\section{EXPERIMENTS AND RESULTS} \label{sec:tests}
\subsection{Elliptic Target Dataset and Parameter Selection} \label{sec:dataset}
We created a dataset with sequences recorded using a UAV from a stationary and moving vehicle carrying an elliptical platform at various distances, angles, and illumination conditions. The dataset contains 1,511 frames (1,378 positive and 133 negative frames) and 456 frames of a thinner version of the same pattern. The size of the frames is $640 \times 360$, and the ground truth for the detections is provided. The dataset can be accessed from \url{http://theairlab.org/landing-on-vehicle}.
The thresholds selection of the proposed ellipse detection algorithm depends on the tolerance for false positives vs. false negatives in the application. However, in practice, the detection is not too sensitive to the parameters in most cases, and they can be selected from a broad range. For our tests on the AirLab Elliptic Target Detection Dataset, we empirically chose the values shown in Table~\ref{tbl:ellipse-detection-parameters}. The GUI tool provided with the code helps with the calibration process letting the user see the parameters' effects in real-time. The ellipse detection parameters are independent of the lighting conditions. Therefore after a one-time calibration, the algorithm should detect the pattern in a wide range of weather conditions (e.g., sunny, cloudy, snowy) as long as the light is enough for the camera to capture the pattern.
\begin{table}[!t]
\centering
\begin{threeparttable}
\caption{Ellipse Detection parameters chosen for the tests on the AirLab Elliptic Target Detection Dataset.}
\label{tbl:ellipse-detection-parameters}
\begin{tabular}{|l|c|l|}
\hline
\rowcolor[HTML]{EFEFEF}
Parameter & Value & Justification \\ \hline
\textit{ContourOverlap} for detection & 0.95 & To prevent False Positives. \\ \hline
\textit{ContourOverlap} for tracking & 0.7 & To enhance target tracking. \\ \hline
\textit{EllipseOverlap} for detection & 0.95 & To prevent False Positives. \\ \hline
\textit{EllipseOverlap} for tracking & 0.3 & To enhance target tracking. \\ \hline
\end{tabular}
\end{threeparttable}
\end{table}
In order to assess the sensitivity of the algorithm against the thresholds, Figure~\ref{fig:contour-parameter-selection} shows the performance of the algorithm for different values of \textit{ContourOverlap} with the value of \textit{EllipseOverlap} fixed to $0.7$. Additionally, Figure~\ref{fig:ellipse-parameter-selection} shows the performance of the algorithm for different values of \textit{EllipseOverlap} with the value of \textit{ContourOverlap} fixed to $0.5$.
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.238\textwidth}
\includegraphics[width=\textwidth]{figures/cot_measures.png}
\caption{~}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.238\textwidth}
\includegraphics[width=\textwidth]{figures/cot_tptn.png}
\caption{~}
\end{subfigure}
\medskip
\begin{subfigure}[b]{0.238\textwidth}
\includegraphics[width=\textwidth]{figures/cot_fpfn.png}
\caption{~}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.238\textwidth}
\includegraphics[width=\textwidth]{figures/cot_times.png}
\caption{~}
\end{subfigure}
\caption{Performance and execution times of our algorithm vs. contour overlap threshold (\textit{ContourOverlap} parameter). The value of \textit{EllipseOverlap} is fixed to $0.7$ for all the experiments. Wrong Positives are the wrong detections when the target was present in the frame.
(a) Accuracy, Recall, and F1 Score of the algorithm increase with the increase of \textit{ContourOverlap} up to a point. (b) Increasing \textit{ContourOverlap} increases the number of True Negatives while may result in fewer True Positives after some point. (c) Increasing \textit{ContourOverlap} results in a lower number of False Positives and Wrong Detections, while it may result in an increase of the number of False Negatives after some point. (d) The execution time decreases by increasing the \textit{ContourOverlap} parameter.}
\label{fig:contour-parameter-selection}
\end{figure}
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.238\textwidth}
\includegraphics[width=\textwidth]{figures/eot_measures.png}
\caption{~}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.238\textwidth}
\includegraphics[width=\textwidth]{figures/eot_tptn.png}
\caption{~}
\end{subfigure}
\medskip
\begin{subfigure}[b]{0.238\textwidth}
\includegraphics[width=\textwidth]{figures/eot_fpfn.png}
\caption{~}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.238\textwidth}
\includegraphics[width=\textwidth]{figures/eot_times.png}
\caption{~}
\end{subfigure}
\caption{Performance of our algorithm vs. ellipse overlap threshold (\textit{EllipseOverlap} parameter). The value of \textit{ContourOverlap} is fixed to $0.5$ for all the experiments.
(a) Accuracy, Recall, and F1 Score of the algorithm very slowly increase with the increase of \textit{EllipseOverlap} up to a breaking point, where they suddenly drop. (b) Increasing \textit{EllipseOverlap} increases the number of True Negatives while reducing the number of True Positives after a certain point. (c) Increasing \textit{EllipseOverlap} results in a lower number of False Positives, while it may result in an increase of the number of False Negatives after some point. (d) The execution time generally decreases by increasing the \textit{EllipseOverlap} value.}
\label{fig:ellipse-parameter-selection}
\end{figure}
Let us define $N_{TP}$, $N_{FP}$, $N_{TN}$, $N_{FN}$, $N_{All}$, and $N_{WD}$ as the number of True Positive detections, False Positives, True Negatives, False Negatives, the total number of frames and the number of frames with visible targets where the detection was wrong. The measures in the plots are defined as follows using the five outcome types above.
\begin{itemize}[leftmargin=*]
\item Accuracy measures the ratio of all the frames that the algorithm gives the correct result (either the target is detected correctly or no target is detected in a frame without a target). It is defined as $(N_{TP} + N_{TN}) / N_{All}$.
\item Precision measures what ratio of all the target detections is actually correct. It is defined as $N_{TP} / (N_{TP} + N_{FP} + N_{WD})$.
\item Recall or sensitivity measures the ratio of all the frames containing targets that are correctly detected. It is defined as $N_{TP} / (N_{TP} + N_{FN} + N_{WD})$.
\item F1 Score is the weighted average of precision and recall and takes both false positives and false negatives into account. F1 Score is defined as $2 \times (Recall \times Precision) / (Recall + Precision)$.
\end{itemize}
It can be seen that increasing the value of \textit{ContourOverlap} up to some point generally decreases the number of false results (both false positives and false negatives) and therefore increases the number of correct results and statistical measures. Although, after some point, the number of detections (true or false positives) starts dropping, and the performance slowly decreases. Additionally, the higher \textit{ContourOverlap} results in more candidate contours being eliminated before further processing, reducing the algorithm's execution time.
On the other hand, the value of \textit{EllipseOverlap} has a smaller effect on performance and execution time, slightly improving the performance up to a point. If it is set too high, the algorithm starts rejecting more ellipses, causing a drop in the number of positive results, significantly decreasing the algorithm's performance.
The choice of other parameters in the algorithm depends on the application. For example, when the pattern is not expected to be far, setting a higher number for $minContourSize$ will improve the speed by eliminating the contours that cannot belong to the elliptic pattern. On the other hand, setting this parameter too high can result in false negatives when the pattern is seen partially or consists of several partial contours. In practice, however, the method has shown to be rather insensitive to these parameters, and generally, there is no need to change the default values. In all our tests, we have used the following values: $mnAxSize = 5$, $mxAxSize = 700$, $maxAxisRatio = 5$, $minContourSize = 50$, $maxScale = 100$.
\subsection{Comparison With Other Methods} \label{sec:comparison}
To compare the performance of our algorithm with other methods, we ran four other methods on the dataset introduced in Section~\ref{sec:dataset}:
the MATLAB implementation of a Hough Transform-inspired approach proposed by Xie \& Ji~\cite{xie2002ellipsedetect} with a random sub-sampling inspired by the work in~\cite{basca2005htsampling},
and the C++ implementations of the methods proposed by Fornaciari et al.~\cite{fornaciari2014ellipsedetect}, Prasad et al.~\cite{prasad2012}, and Jia et al.~\cite{jia2017}.
Table~\ref{tbl:ellipse-methods-performance} shows the results and performance of the algorithms on our dataset.
\begin{table}[!t]
\centering
\begin{threeparttable}
\caption{Performance of the ellipse detector methods on the AirLab Elliptic Target Detection Dataset.}
\label{tbl:ellipse-methods-performance}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|}
\hline
\rowcolor[HTML]{EFEFEF}
Method & Accuracy$^{\star}$ & Precision$^{\star}$ & Recall$^{\star}$ & F1 Score$^{\star}$ \\ \hline
Ours & \textbf{96.56\%} & \textbf{99.77\%} & \textbf{96.44\%} & \textbf{0.981} \\ \hline
\cite{xie2002ellipsedetect} & 3.64\% & 3.64\% & 3.99\% & 0.038 \\ \hline
\cite{fornaciari2014ellipsedetect} & 88.75\% & 99.67\% & 87.66\% & 0.933 \\ \hline
\cite{prasad2012} & 89.81\% & 99.45\% & 90.22\% & 0.965 \\ \hline
\cite{jia2017} & 78.36\% & 95.50\% & 80.04\% & 0.871 \\ \hline
\end{tabular}
\begin{tablenotes}
\small
\item $^{\star}$ As defined in Section~\ref{sec:dataset}.
\end{tablenotes}
\end{threeparttable}
\end{table}
The results show that our implemented algorithm outperforms the other methods in all the criteria. The method by Xie \& Ji's~\cite{xie2002ellipsedetect} was unable to perform well on the real frames of our test environment due to the relatively low resolution of our frames and the small size of the target in the frames; the few cases it could detect the elliptic target were when the target was covering a large portion of the frame. The main problem with the method by Fornaciari~\cite{fornaciari2014ellipsedetect} was that it was unable to detect the elliptic targets when more than 25\% of the target was outside of the frame (case (e) in Figure~\ref{fig:circletypes}). The method by Jia et al.~\cite{jia2017} is comparatively fast but has a high false positive rate and is unable to detect the ellipses when they are far away (small in the frame). Prasad's method~\cite{prasad2012} has comparable performance to Fornaciari's but is significantly slower. At the same time, our proposed algorithm was fast and still able to detect the target's elliptic pattern in partial views or when it was small. Additionally, we should note that all the false positive cases of our algorithm on the dataset, detected elliptical drawings on the ground, which would have been rejected if the UAV's altitude information was used.
Figure~\ref{fig:deckresults} shows results for the detection of the elliptical pattern in some sample frames from the dataset. The method by Fornaciari et al. is unable to detect the ellipses in Figure~\ref{fig:deckresults}(a) and Figure~\ref{fig:deckresults}(d).
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/decktrack1.png}
\caption{~}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/decktrack2.png}
\caption{~}
\end{subfigure}
\medskip
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/decktrack3.png}
\caption{~}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/decktrack4.png}
\caption{~}
\end{subfigure}
\caption{Result of the proposed algorithm on sample frames. The red ellipse indicates the detected pattern on the deck of the moving vehicle.}
\label{fig:deckresults}
\end{figure}
\begin{table*}[t]
\centering
\begin{threeparttable}
\caption{Execution times of ellipse detector methods on the AirLab Elliptic Target Detection Dataset.$^\star$}
\label{tbl:ellipse-methods-times}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{\textbf{Method}} & \multicolumn{3}{c|}{\textbf{True Positive}} & \multicolumn{3}{c|}{\textbf{True Negative}} & \multicolumn{3}{c|}{\textbf{False Positive}} & \multicolumn{3}{c|}{\textbf{False Negative}} \\
& \textit{Avg.} & \textit{Max.} & \textit{Min.} & \textit{Avg.} & \textit{Max.} & \textit{Min.} & \textit{Avg.} & \textit{Max.} & \textit{Min.} & \textit{Avg.} & \textit{Max.} & \textit{Min.} \\ \hline
Our Method & 32.2 & \textbf{64.1} & \textbf{1.5} & 48.9 & 68.1 & \textbf{6.6} & 48.4 & 49.6 & 46.4 & 39.1 & 64.0 & \textbf{2.5} \\ \hline
Fornaciari, et al \cite{fornaciari2014ellipsedetect
& 26.5 & 87.3 & 15.5 & 25.4 & 54.2 & 15.3 & - & - & - & \textbf{23.8} & \textbf{60.3} & 14.7 \\ \hline
Prasad, et al \cite{prasad2012
& 36.2 & 70.0 & 25.5 & 34.4 & 64.9 & 20.0 & - & - & - & 38.8 & 72.1 & 26.6 \\ \hline
Jia, et al \cite{jia2017
& \textbf{24.4} & 67.3 & 13.1 & \textbf{21.2} & \textbf{52.7} & 11.6 & - & - & - & \textbf{23.8} & 63.3 & 12.8 \\ \hline
\end{tabular}
\begin{tablenotes}
\small
\item $^\star$ All times are in milliseconds.
\end{tablenotes}
\end{threeparttable}
\end{table*}
Table~\ref{tbl:ellipse-methods-times} compares the execution times of the ellipse detection methods on the same dataset (using Intel Core i5-4460 CPU @ 3.20 GHz). The implementation of Xie \& Ji's method is done in MATLAB, which gives much higher execution times than C++ implementations. Therefore we excluded it from Table~\ref{tbl:ellipse-methods-times}. Fornaciari et al. method has similar average execution times to our algorithm. However, their approach provides slightly more consistent execution times, which can be convenient for control systems using the detection output for robot control. On the other hand, our method's speed significantly increases (with frame processing time going down to just a few milliseconds) when the robot gets closer to the target pattern. This increase in speed especially helps the system to have a much higher detection rate when a higher processing speed is needed for the robot to approach the moving vehicle for landing.
\subsection{Example Application} \label{sec:application}
To test the proposed ellipse detection and tracking method's performance, it was used with a visual servoing method for an autonomous UAV landing on a circular pattern painted on top of moving platforms in indoor, outdoor, and simulated environments~\cite{keipour2021visualservoing}. Figure~\ref{fig:landing-tests} shows screenshots of the method in these different lighting conditions. Videos for these experiments and the project details can be accessed at \url{http://theairlab.org/landing-on-vehicle}.
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/resindoor2.png}
\caption{~}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth]{figures/resfield4.png}
\caption{~}
\end{subfigure}
\medskip
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth, height=2.4cm]{figures/simulation1.jpg}
\caption{~}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.23\textwidth}
\includegraphics[width=\textwidth, height=2.4cm]{figures/fpv.jpg}
\caption{~}
\end{subfigure}
\caption{Screenshots from video sequences showing our autonomous UAV landing on a pattern moving at up to $4.2$~m/s speed in various lighting and environmental conditions.}
\label{fig:landing-tests}
\end{figure}
\section{FUTURE WORK AND DISCUSSION} \label{sec:conclusion}
The proposed ellipse detection and tracking algorithm has shown its performance in an example application and has outperformed other standard methods in our tests. However, the proposed method has mainly been tested on a set of elliptical patterns used in our current and prior research. Utilizing the method for other applications may require further testing and improvements. The following suggestions can further enhance the performance and increase the method's usability in real-world robotics applications.
The underlying algorithms used in the ellipse detection steps are the most common methods already available in the OpenCV library. The choice has been made to allow fast implementation by the potential reader and convenience. If better performance is required, the whole method's execution speed and performance can be improved by replacing steps such as ellipse fitting with faster and better algorithms.
Finally, if the robot's camera is not perfectly rectified, it may lose track of the elliptic target at close distances where only a small portion of the pattern is visible at such a skewed angle that it causes the circle to be seen as non-elliptic in the camera. The problem exists for any ellipse detection algorithm but can be improved using a robust tracker (such as Kernelized or Discriminative correlation filters) instead of a detector to track the target ellipse when the robot's camera is too close to the elliptical pattern.
\section*{ACKNOWLEDGMENT}
Authors want to thank Near Earth Autonomy for their support in this project. The project was sponsored by Carnegie Mellon University Robotics Institute and Mohamed Bin Zayed International Robotics Challenge. During the realization of this work, G. Pereira was supported by UFMG and CNPq/Brazil.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The Lie-Trotter product formula~\cite{trotter1959product,suzuki1976generalized} was originally used by Lloyd~\cite{lloyd1996universal} to establish the first method for efficiently approximating the dynamics $U(t) = e^{-i t H}$ generated by a local Hamiltonian $H$ with a universal quantum computer. After many refinements~\cite{childs2019theory, childs2019nearly, campbell2019random} this approach (often called ``Trotterization'') continues to be an appealing method for Hamiltonian simulation from both experimental and mathematical perspectives.
The method is based on dividing $U(t)$ into $L$ short-time evolutions $U(t) = U^{L}(\delta t), t = L \delta t$, and replacing each $U(\delta t)$ with an approximation $T(\delta t)$. The parameter $L$ is the number of Trotter steps and $\delta t > 0$ is the Trotter step size. Given a decomposition of the Hamiltonian into a sum of layers $H = \sum_{n=1}^\Gamma H_n$~\footnote{As usual in trotterization, the guideline for the decomposition is that one has a way to implement $e^{i t H_j}$ efficiently for each $H_j$. A sufficient condition for this is for each $H_j$ to be a sum of pairwise commuting local terms.}, the first order product formula approximation is
\begin{equation}
T(\delta t) \equiv \prod_{n=1}^{\Gamma}e^{-iH_{n}\delta t},
\label{equ:1st}
\end{equation}
where $\delta t$ and $L$ are chosen to depend on the tolerable level of error, as we subsequently discuss.
Most prior works quantify the Trotter error in terms of the operator norm $\|U(\delta t) - T(\delta t)\|$, but in this work we directly compare states evolved under the exact and approximate time evolution to produce a tighter error estimate for the specific class of initial states we consider. We quantify the Trotter error in terms of the phase error $\theta$ and the fidelity error $f$,
\begin{equation}
T(\delta t)^{L}|\psi\rangle = \sqrt{1-f}e^{i\theta}U(t)|\psi\rangle + \sqrt{f}U(t)|\psi^{\perp}\rangle.
\label{equ:setup}
\end{equation}
where $|\psi\rangle$ is the initial state. While the operator norm can always be used to upper bound $f$ and $\theta$ for an arbitrary initial state $|\psi\rangle$, we identify situations in which analyzing $f$ and $\theta$ for specific initial states produces tighter error bounds than the general case.
Our results are based on a spectral analysis of the effective Hamiltonian $\widetilde{H}$ that generates the Trotterized time evolution,
\begin{equation}
\widetilde{H} \equiv i\log(T(\delta t))/\delta t,\quad T(\delta t) = e^{-i\widetilde{H} \delta t}
\label{equ:eff_H}
\end{equation}
Regarding $\delta t$ as a small parameter, perturbative methods can be used to compare the spectrum $\{E_k ,|\psi_k\rangle\}$ of the original Hamiltonian $H$ to the spectrum $\{\widetilde{E}_k ,|\widetilde{\psi}_k\rangle\}$ of the effective Hamiltonian $\tilde{H}$. These methods lead us to consider applications in which the initial state $|\psi\rangle$ is (or is close to) an eigenstate of $H$, enabling an improved upper bound on the Trotter step size in Quantum Phase Estimation (QPE) and Digital Adiabatic Simulation (DAS).
QPE is one of the most important quantum simulation algorithms~\cite{kitaev1995quantum}, which relates Hamiltonian time-evolution $U(t) = e^{-i t H}$ to the measurement of energy eigenvalues~\cite{reiher2017elucidating,kivlichan2020improved, nielsen2002quantum}. In the ideal version of the algorithm, measuring the output of the phase estimation circuit collapses the system into an energy eigenstate of $H$. If we replace the time evolution with the product formula approximation $T(\delta t)^{L}$, then (in the ideal case) we will instead measure an energy eigenvalue of $\widetilde{H}$. Therefore the Trotter error is directly related to the difference in spectrum between $H$ and $\widetilde{H}$. Under conditions which are satisfied in many cases of interest, we show that the first order perturbative correction vanishes and use this to rigorously show that the Trotter error in phase can be reduced from $\mathcal{O}(L\delta t^{2})$ to $\mathcal{O}(L\delta t^{3})$. In terms of the target precision $\epsilon$ of the QPE, this means the Trotter step size can be enlarged from $\delta t = \mathcal{O}(\epsilon)$ to $\delta t = \mathcal{O}(\epsilon^{1/2})$.
DAS can be used to implement adiabatic quantum computation \cite{farhi2001quantum,aharonov2008adiabatic,albash2018adiabatic} and adiabatic state preparation~\cite{aharonov2003adiabatic, wan2020fast} on a digital quantum computer. The scaling of Trotter error in DAS has previously been analyzed in terms of the operator norm~\cite{barends2016digitized, sun2018adiabatic}. We find this measure of error is dominated by the accumulation of a global phase, whereas in adiabatic algorithms it is generally only the fidelity error $f$ that matters. We regard the Trotterized time evolution as a discretized adiabatic evolution under an effective Hamiltonian, and obtain tighter bounds on the fidelity error by applying an adiabatic theorem to the effective Hamiltonian.
This paper is organized as follows: we first illustrate the setting of the problems and our main results, together with the techniques and lemmas used in the proof. The primary mathematical tools are a rigorous perturbation method~\cite{ambainis2004elementary} and the Magnus expansion~\cite{blanes2009magnus}. Then we apply the main results to analyze Trotter error in QPE and DAS, in both cases we achieve improvements in circuit complexity.
\section{Main Results}
\subsection{Set up and Notations}
Usually, the Trotter error is quantified by the norm distance between operators:
\begin{equation}
\Delta \equiv \|\hat{\Delta}\|,\quad\hat{\Delta} \equiv T(\delta t)^{L} - U(t)
\end{equation}
The notation $\|\cdot\|$ refers to the operator norm : $\|M\| = \max_{\|x\|_{2} = 1}\|M|x\rangle\|_{2}$, where $\|\cdot\|_{2}$ is the Euclidean norm of vector $\|v\|_{2} = \sqrt{vv^{\dagger}}$. To quantify $\Delta$, it's enough to quantify the norm distance error of a single Trotter step $\delta \equiv \|T(\delta t) - U(\delta t)\|$ as $\Delta \le L\delta$. For a given error tolerance $\epsilon$, the restriction of $\Delta \le \epsilon$ determines the gate complexity of the algorithm.
In this paper, we separate the digital error into phase error $\theta$ and fidelity error $f$ defined by
\begin{gather}
f \equiv 1 - |\langle\psi|U^{\dagger}(t)T(\delta t)^{L}|\psi\rangle|^{2}\\
\theta \equiv \text{Arg}\left(\langle\psi|U^{\dagger}(t)T(\delta t)^{L}|\psi\rangle\right)
\end{gather}
where $|\psi\rangle$ is an initial state. For any $L,\delta t$ that satisfy $\Delta \le 1/\sqrt{2}$, the Euclidean distance error $\mathcal{E} \equiv \|\hat{\Delta}|\psi\rangle\|_{2}$ satisfies (see Appendix A)
\begin{equation}
f + \frac{\theta^{2}}{4} \le \mathcal{E}^{2} \le 2f + \theta^{2}
\label{equ:fandtheta}
\end{equation}
Without further assumptions about $H$ and $|\psi\rangle$, the two parameters are bounded by $f = \mathcal{O}(\Delta^{2}), |\theta| = \mathcal{O}(\Delta)$ as $\mathcal{E} \le \Delta$. However, in some special cases, we find that both $f$ and $\theta^{2}$ have a different parameteric scaling with $\Delta$. This means that $\Delta$ does not always reflect the true Trotter error that we are interested in, which leads us to a different approach.
\subsection{Improved $f$ : Spectral Analysis}
The Trotterized evolution operator $T(\delta t)^{L}$ can be viewed as an exact evolution under an effective Hamiltonian $\widetilde{H} \equiv i\log(T(\delta t))/\delta t$. Owing to the tiny size of $\delta t$, the spectrum of $\widetilde{H}$ is $\{\widetilde{E}_{k},|\widetilde{\psi}_{k}\rangle\}$ is close to that of $H$. Based on this observation,
suppose the initial state is one of the eigenstate $|\psi_{k}\rangle$ of $H$, $f$ and $\theta$ can be quantified by the difference between spectrum:
\begin{gather}
f = \mathcal{O}(1-|\langle \psi_{k}|\widetilde{\psi}_{k}\rangle|^{2})\\
\theta = \mathcal{O}(|\widetilde{E}_{k} - E_{k}|t)
\label{equ:relation}
\end{gather}
To ensure there exists a one-to-one correspondence between $|\psi_k\rangle$ and $|\tilde{\psi}_k\rangle$, we assume the spectrum is nondegenerate. Therefore there is some spectral gap $\lambda_k = \min(E_k - E_{k-1}, E_{k+1}-E_k)$ around this eigenstate. We use $\lambda$ as a general lower bound for the spectral gap around an initial eigenstate.
The upper bound for the difference between energy eigenvalues is based on an upper bound for $\|\widetilde{H}\|$. Using the Baker-Campbell-Hausdorff formula,
\begin{equation}
\widetilde{H}(\delta t) = H + \frac{i\delta t}{2}\sum_{l>m}[H_{l},H_{m}] + \mathcal{O}(\delta t^{2}).
\label{equ:linear}
\end{equation}
The first few terms of the standard (Rayleigh-Schrodinger) perturbation theory can be used to estimate $f$ and $\theta$, but to avoid convergence issues and derive rigorous results we use other methods ~\cite{ambainis2004elementary, jansen2007bounds} that are widely used in proofs of adiabatic theorems. By Weyl's inequality, the perturbation in the eigenvalues satisfies
\begin{equation}
|\widetilde{E}_{k} - E_{k}| \le \|\widetilde{H} - H\| \\
\end{equation}
The perturbation of the eigenvectors is derived from the following lemma.
\begin{lemma} [Rigorous perturbation method~\cite{jansen2007bounds}] $H(s)$ is a parameterized Hamiltonian with spectrum :$\{E_{j}(s), P_{j}(s)\}$, define
\begin{equation*}
P(s) = \sum_{j=1}^{m}P_{j}(s)
\end{equation*}
as the projector into a subspace $\mathcal{A}$ spanned by $m$ eigenstates. Its derivative has norm upper bound:
\begin{equation}
\|P^{\prime}(s)\| \le \sqrt{m}\|H^{\prime}(s)\|/\lambda
\end{equation}
where $\lambda$ is the lower bound of energy gap between the eigenstates in and outside region $\mathcal{A}$.
\label{lemma : rigorp}
\end{lemma}
For a single eigenvector, lemma \ref{lemma : rigorp} implies
\begin{equation}
\begin{aligned}
\sqrt{1 - |\langle \widetilde{\psi}_{k}|\psi_{k}\rangle|^{2}} &= \|\widetilde{P}_{k} - P_{k}\|\\
&= \|P(\delta t) - P(0)\| \le \delta t\max_{s\in[0,\delta t]}\|P^{\prime}(s)\|\\
&\le \max_{s\in[0,\delta t]}\|\widetilde{H}^{\prime}(s)\|\delta t/\lambda
\label{equ:norm_to_fid}
\end{aligned}
\end{equation}
The term $\|\widetilde{H}'(s)\|$ quantifies the size of the perturbation. The necessary bound on $\|\widetilde{H}^{\prime}\|$ has already been obtained using the Magnus expansion in previous work, see Appendix B for more details.
\begin{lemma}[Magnus Expansion\cite{tran2020faster}] Given $\widetilde{H}$ defined in Eq : (\ref{equ:eff_H}), when $\delta t = \mathcal{O}(N^{-1})$, for all $s\in [0,\delta t]$:
\begin{gather}
\|\widetilde{H}(s) - H\| = \mathcal{O}(h\delta t),\quad \|\widetilde{H}^{\prime}(s)\| = \mathcal{O}(h)\\
h = \frac{\alpha}{2} + \frac{4}{3}(\beta + 128\alpha\|H\|)\delta t\\
\alpha \equiv \sum_{n>m}||[H_{n},H_{m}]||\nonumber \\
\beta \equiv \sum_{l\ge n>m}||[H_{l},[H_{n},H_{m}]]|| \nonumber
\end{gather}
\label{lemma:magnus}
\end{lemma}
For example, if $H = \sum_{j = 1}^N h_j$ is a local Hamiltonian on $N$ qubits that satisfies $[h_{j},h_{k}] = 0, \forall |j-k|>1$, then $h = \mathcal{O}(N) + \mathcal{O}(N^{2}\delta t) = \mathcal{O}(N)$. In general, the parameter dependence of $h$ is complicated but will always be $\textrm{poly}(N)$ for any $k$-local Hamiltonian. To simplify the notation we retain the form of $h$ in the following paragraph.
The following result follows from Lemma \ref{lemma : rigorp} with Lemma \ref{lemma:magnus}, see Appendix C.
\begin{corollary}[Eigenstate as initial state]
For any $H$ and any eigenstate $|\psi_{k}\rangle$ of $H$ separated from the rest of the spectrum by a spectral gap $\lambda$, the time evolution under the 1st-order product formula satisfies :
\begin{equation*}
T(\delta t)^{L}|\psi_{k}\rangle = \sqrt{1-f}e^{i\theta} U(t)|\psi_{k}\rangle + \sqrt{f}|\psi_{k}^{\perp}\rangle
\end{equation*}
where the fidelity error and phase error satisfy
\begin{gather}
|\theta| = \mathcal{O}(Lh\delta t^{2})\\
f = \min\left\{\mathcal{O}\left(\frac{h^{2}\delta t^{2}}{\lambda^{2}}\right), \mathcal{O}(L^{2}h^{2}\delta t^{4})\right\}
\end{gather}
with $h$ defined in Lemma \ref{lemma:magnus}.
\label{coro:eigen}
\end{corollary}
As a comparison, $\Delta = \mathcal{O}(Lh\delta t^{2})$ in general. Corollary \ref{coro:eigen} implies that the fidelity error $f$ eventually stops growing with with the total number of Trotter steps $L$, which is an extreme example of $f \ll \Delta^{2}$. After a short initial period, the Trotter error only accumulates in the global phase. This fact can be related to the leakage rate property~\cite{csahinouglu2020hamiltonian} of the Trotterized evolution operator. Suppose the initial state is $|\psi\rangle = \sum_{k}c_{k}|\psi_{k}\rangle$, where $|\psi_{k}\rangle$ all belong to a special region $\mathcal{A}$, like low-energy states. The leakage rate is the percentage for $|\psi\rangle$ to go outside that region after $T(\delta t)^{L}$: $1 - \text{Tr}(P_{\mathcal{A}}T(\delta t)^{L}|\psi\rangle\langle\psi|T^{\dagger}(\delta t)^{L})$. Using the argument about $\widetilde{H}$ we prove:
\begin{theorem} [Leakage rate]
$T(\delta t)^{L}$ is Trotterized evolution operator, if the initial state $\rho$ belongs to subregion $\mathcal{A}$ spanned by $m$ eigenstates:
\begin{equation*}
P = \sum_{j=1}^{m}P_{j},\quad \text{Tr}(\rho P) = 1
\end{equation*}
Then the leakage rate can be bounded by the norm distance between $P$ and corresponding effective projector $\widetilde{P}$ induced by $\widetilde{H}$ if $\|P - \widetilde{P}\| < 1$:
\begin{equation*}
\begin{aligned}
1 - \text{Tr}(T(\delta t)^{L}\rho T^{\dagger}(\delta t)^{L}P) &= \mathcal{O}\left(\|P - \widetilde{P}\|^{2}\right)\\
&= \mathcal{O}\left(\frac{mh^{2}\delta t^{2}}{\lambda^{2}}\right)
\end{aligned}
\end{equation*}
where $\lambda$ is the lower bound of energy gap between the eigenstates in and outside region $\mathcal{A}$.
\label{coro:multi}
\end{theorem}
\subsection{Improved $\theta$ : Special Perturbation}
The bound $\theta = \mathcal{O}(\Delta)$ can be tight, even when the initial state is an eigenstate. However, we show that it can be improved by a factor of $\delta t$ under assumptions that are satisfied for many Hamiltonians of interest, and this improves the scaling of the Trotter step size needed for QPE. This can be illustrated by the simple case $H = H_A + H_B$ in which the Hamiltonian is decomposed into two commuting layers,
\begin{gather*}
T(\delta t) = e^{-iH_{A}\delta t}e^{-iH_{B}\delta t} = e^{-i\widetilde{H}\delta t},\\
\widetilde{H} = H - i\frac{\delta t}{2}[H_{A},H_{B}] + \mathcal{O}(\delta t^{2}).
\end{gather*}
The leading perturbation term is $V \equiv -i\delta t[H_{A},H_{B}]/2$. In standard perturbation theory, the 1st order correction in energy is $E^{(1)} = \langle \psi_{k}|V|\psi_{k}\rangle$. However,
\begin{equation*}
\langle \psi_{k}|[H_{A},H_{B}]|\psi_{k}\rangle = \langle \psi_{k}
|[H,H_{B}]|\psi_{k}\rangle = 0.
\end{equation*}
In previous section we prove an upper bound of Trotter error in energy of order $\mathcal{O}(h\delta t)$. While under this special situation, because $E^{(1)} = 0$, the shift in energy at most has order $\mathcal{O}(\delta t^{2})$. This improvement can be applied to a general decomposition $H = \sum_{n=1}^{\Gamma}H_{n}$. Whenever the leading order correction,
\begin{equation*}
V = \frac{i\delta t}{2}\sum_{l>m}[H_{l},H_{m}],
\end{equation*}
is off-diagonal in the eigenbasis of $H$, we can reduce the Trotter error in energy from $\mathcal{O}(\delta t)$ to $\mathcal{O}(\delta t^{2})$. In the setting of Corollary \ref{coro:eigen}, the following result is proven by rigorous perturbative methods in Appendix E.
\begin{lemma}$H$ is a normalized local Hamiltonian on $N$ sites with spectrum $\{E_{k},|\psi_{k}\rangle\}$, $\widetilde{H}$ is its corresponding effective Hamiltonian induced from 1st order product formula. The new spectrum is $\{\widetilde{E}_{k},|\widetilde{\psi}_{k}\rangle\}$. The first perturbation of $\widetilde{H}$ is off-diagonal in the eigenbasis of $H$:
\begin{equation}
\forall |\psi_{k}\rangle,\quad \langle\psi_{k}|\widetilde{H} - H|\psi_{k}\rangle = \mathcal{O}(\delta t^{2})
\label{equ : criteria}
\end{equation}
Further if $\delta t = \mathcal{O}(\lambda/N)$, where $\lambda$ is the lower bound of spectral gap between $|\psi_{k}\rangle$ and neighboring eigenstates, then the shift in energy satisfies:
\begin{equation}
|\widetilde{E}_{k} - E_{k}| = \mathcal{O}\left(N^{2}\delta t^{2}\max\left\{1,\frac{1}{\lambda^{2}}\right\}\right)
\end{equation}
\label{lemma:spe}
\end{lemma}
In addition to Hamiltonians that can be decomposed into two commuting layers (of which a prominent class of examples are Hamiltonians $H = H_X + H_Z$ that have local terms which involve only Pauli $X$ or Pauli $Z$ operators), we list other different conditions where Eq : (\ref{equ : criteria}) is satisfied.
\begin{itemize}
\item Real Hamiltonians. Assume all of the local terms of $H$ have real matrix elements in some basis. The components of an eigenstate $|\psi_{k}\rangle$ of any real symmetric matrices can all be taken to be real. Consider an arbitrary commutator in $V$, $\langle\psi_{k}|H_{l}H_{m}|\psi_{k}\rangle$ is conjugate to $\langle\psi_{k}|H_{m}H_{l}|\psi_{k}\rangle$, and both are real numbers. So they are equal and appear with opposite signs in the commutator. Therefore, $\forall k,l,m, \langle\psi_{k}|[H_{l},H_{m}]|\psi_{k}\rangle = 0$.
\item Any Hamiltonian whose layers (or local terms) can be totally ordered to satisfy $[H_{l},H_{m}] = 0, \forall |l-m|>1$. This condition is satisfied by 1D Hamiltonians with nearest-neighbor interactions, as well as general lattice Hamiltonians regarded as 1D chains of super-sites (since our results do not depend on the local dimension). This condition leads to a recursive relation: $[H_{l},H_{l+2}] = [H_{l}, H - H_{l-1} - H_{l+1}] = 0$. If $[H_{l-1},H_{l}]$ is off-diagonal, $[H_{l},H_{l+1}]$ is also off-diagonal as any operator in the form of $[O,H]$ is off-diagonal in the eigenbasis of $H$. The case of $l=1$ is special as we don't have $H_{0}$. Thus $[H_{1}, H_{2}] = [H_{1},H]$ is off-diagonal. As a result, $\forall l, [H_{l},H_{l+1}]$ is off-diagonal. These assumptions can be satisfied for any Hamiltonian with geometrically local terms in 1D.
\item Frustration-free Hamiltonians~\cite{bravyi2010complexity}. This type of Hamiltonian satisfies $H = \sum_{j}\Pi_{j}, H|\psi_{0}\rangle = E|\psi_{0}\rangle, \Pi_{j}|\psi_{0}\rangle = E_{j}|\psi_{0}\rangle$ where $|\psi_{0}\rangle$ is ground state. With this property, when the initial state is the ground state, there will be no Trotter error no matter how big $\delta t$ is. In Lemma \ref{lemma:spe}, the upper bound on the Trotter error is inversely proportional to the spectral gap $\lambda$. However, it's possible for frustration-free Hamiltonian to be gapless~\cite{bravyi2012criticality}. This example shows that our methods can still overestimate the Trotter error for gapless Hamiltonians.
\end{itemize}
The second part of Eq : (\ref{equ:fandtheta}) indicates that $f$ and $\theta$ can be used to bound the Euclidean distance error as well. We have just proved that $f$ and $\theta^{2}$ can both be much smaller than $\Delta$, which means under these conditions the norm distance can't reflect Trotter error in Euclidean distance either.
Finally, we provide an example $H = H_1 + H_2 + H_3$ in which $V$ is not off-diagonal to show the result in Lemma \ref{lemma:spe} is not fully general. Let $H$ be a diagonal matrix $\Lambda$ in the eigenbasis of itself. In this basis choose:
\begin{equation*}
H_{1} = X\otimes I,\quad H_{2} = Y\otimes I,\quad H_{3} = \Lambda - H_{1} - H_{2}
\end{equation*}
Thus:
\begin{equation*}
V/\delta t = i[(X+Y)\otimes I,\Lambda]/2 + Z\otimes I
\end{equation*}
The first term is off-diagonal, the second term is not. Thus $V$ is not off-diagonal.
\subsection{Improved $f$ : Adiabatic Theorem}
DAS is a special type of time-dependent evolution simulation task~\cite{poulin2011quantum,berry2020time} that leverages the quantum adiabatic theorem~\cite{cheung2011improved,jansen2007bounds}. Physically, when the Hamiltonian evolves with time slowly enough, an initial state in some eigenspace will stay close to that eigenspace of the time-dependent Hamiltonian at all times. The evolution operator under $\hat{H}(t)$ has expression:
\begin{gather*}
i\frac{d}{dt}|\psi(t)\rangle = \hat{H}(t)|\psi(t)\rangle,\quad |\psi(T)\rangle = \hat{A}(T)|\psi(0)\rangle\\
\hat{A}(T) = \exp_{\mathcal{T}}\left(-i\int_{0}^{T}\hat{H}(t)dt\right)
\end{gather*}
We restrict out attention to the linear adiabatic path $\hat{H}(t) = (1-t/T)H_{i} + t/T H_{f}$. In terms of the dimensionless parameter $s = t/T$,
\begin{gather*}
H(s) \equiv \hat{H}(Ts) = (1-s)H_{i} + sH_{f}\\
A(T) \equiv \exp_{\mathcal{T}}\left(-iT\int_{0}^{1}H(s)ds\right) = \hat{A}(T)
\end{gather*}
$T\to\infty$ corresponds to the case where evolution is performed adiabatically:
\begin{gather*}
\lim_{T\to \infty}A(T)|\psi_{i}\rangle = |\psi_{f}\rangle
\end{gather*}
$|\psi_{i}\rangle$ is one eigenstate of initial Hamiltonian $H_{i}$ and $|\psi_{f}\rangle$ is the corresponding one of $H_{f}$. $T$ quantifies how slowly the evolution is, it also reflects how close the evolved state is to $|\psi_{f}\rangle$. We apply the following rigorous adiabatic theorem.
\begin{lemma}[Adiabatic theorem~\cite{jansen2007bounds}] The error of adiabatic evolution is quantified by:
\begin{equation}
\epsilon_{adb} \equiv \|P_{f} - A(T)P_{i}A^{\dagger}(T)\|
\end{equation}
where $P_{f} = |\psi_{f}\rangle\langle\psi_{f}|$, $P_{i} = |\psi_{i}\rangle\langle\psi_{i}|$. $A(T)$ is the adiabatic evolution operator. Then:
\begin{equation}
\epsilon_{\text{adb}} \le \mathcal{G}(T,H)
\end{equation}
with
\begin{equation}
\begin{aligned}
\mathcal{G}(T,H) &\equiv \frac{1}{T}\left(\frac{\|H^{\prime}(0)\|}{\lambda(0)^{2}} + \frac{\|H^{\prime}(1)\|}{\lambda(1)^{2}}\right)\\
& + \frac{1}{T}\int_{0}^{1}\frac{\|H^{\prime\prime}(s)\|}{\lambda^{2}(s)} + 7\frac{\|H^{\prime}(s)\|^{2}}{\lambda^{3}(s)}ds
\end{aligned}
\end{equation}
$\lambda(s)$ is the lower bound of energy gap between $|\psi(s)\rangle$ and other eigenstates during evolution. $|\psi(s)\rangle$ is the eigenstate in $H(s)$ associated with $|\psi_{i}\rangle$.
\label{lemma:continu}
\end{lemma}
\noindent Notice the adiabatic theorem refers to fidelity error,
\begin{equation*}
\epsilon_{\text{adb}} = \sqrt{1 - |\langle\psi_{f}|A(T)|\psi_{i}\rangle|^{2}}.
\end{equation*}
To simulate this process with Trotterization, we first approximate $A$ by a product of short time evolutions,
\begin{gather}
A_{d} \equiv \prod_{a=1}^{M}U_{a},\quad U_{a} \equiv e^{-iH_{a}\bar{\delta t}}\\
\bar{\delta t} \equiv \frac{T}{M},\quad H_{a} \equiv H\left(\frac{a}{M}\right)
\end{gather}
Here $M$ is the discretization number. Note that going from $A$ to $A_d$ already incurs some discretization error. Each short time evolution operator is further Trotterized,
\begin{gather}
A_{t} \equiv \prod_{a=1}^{M}U^{t}_{a},\quad U^{t}_{a} \equiv U^{t}\left(\frac{a}{M}\right)\\
U^{t}(s) \equiv e^{-i\bar{\delta t}(1-s)H_{i}}e^{-i\bar{\delta t} s H_{f}}
\label{equ:dig_adb}
\end{gather}
Therefore the parameter $M$ determines both the discretization error and the additional Trotter error. The expression of $U^{t}(s)$ actually depends on how the Trotterization is performed, while our argument applies to general product formulas, thus we use the above definition as an instance.
\iffalse
\begin{center}
\begin{tabular}{ |p{2cm} | p{6cm}| }
$\epsilon_{\text{adb}}$ & Total error of continuous adiabatic evolution with $H$. \\ \hline
$\epsilon_{\text{adb}}^{\prime}$ & Total error discretized adiabatic evolution with $H$. \\ \hline
\end{tabular}
\end{center}
\fi
The total error of DAS, which is the fidelity distance between $A_{t}|\psi_{i}\rangle$ and $|\psi_{f}\rangle$, is divided into three parts. The first part comes from adiabatic evolution itself, it can be quantified by Lemma \ref{lemma:continu}. The other two stem from discretization and Trotterization steps. To simplify the question, we only study the case where the error caused by discretization is negligible comparing to the error from adiabatic theorem, which is to say:
\begin{gather}
\epsilon^{\prime}_{\text{adb}} \equiv \|P_{f} - A_{d}P_{i}A^{\dagger}_{d}\|\\
\epsilon_{\text{dis}} \equiv |\epsilon^{\prime}_{\text{adb}} - \epsilon_{\text{adb}}| \ll \epsilon_{\text{adb}}
\end{gather}
\begin{figure}
\includegraphics[scale=0.5]{err_total.png}
\includegraphics[scale=0.5]{err_adb_tro.png}
\caption{Error scaling of DAS. The example is $H_{i} = -\sum_{i}X_{i}, H_{f} = -\sum_{i}(Z_{i} + Z_{i}Z_{i+1})$ on $N = 8$ sites without periodic boundary conditions. The initial state is the ground state of $H_{i}$. The discretization number is fixed at $M = 2000$. The figures illustrate how $\epsilon^{\prime}_{\text{tot}}, \epsilon^{\prime}_{\text{adb}}$ and $\epsilon_{\text{tro}}$ scale with parameter $T$ ranging from $[M/50, M]$ evenly. Each point is a complete run of DAS and there are 50 of them in one line. The vertical axis is shown on a $\log_{10}$ scale. The first figure is only about $\epsilon^{\prime}_{\text{tot}}$, the second figure is a combination of $\epsilon^{\prime}_{\text{adb}}$ and $\epsilon_{\text{tro}}$. As to $\epsilon^{\prime}_{\text{adb}}$, the overall scaling of $\mathcal{O}(T^{-1})$ indicates the correctness of prediction by adiabatic theorem. Note that the Trotter error remains small when $T,M$ are comparable to each other. $\epsilon^{\prime}_{\text{tot}}$ is close to the summation of $\epsilon^{\prime}_{\text{adb}}$ and $\epsilon_{\text{tro}}$. The overall scaling is similar to that of $\epsilon^{\prime}_{\text{adb}}$, while after a turning point the error grows rapidly, and the curve matches with that of $\epsilon_{\text{tro}}$. The turning point is the place where the minimal error is reached. Approximately it's $0.75M$. }
\label{fig:err}
\end{figure}
We focus on the Trotter error in DAS, \begin{equation}
\epsilon_{\text{tro}} \equiv \|A_{d}P_{i}A^{\dagger}_{d} - A_{t}P_{i}A^{\dagger}_{t}\| \le \|A_{d} - A_{t}\|
\end{equation}
which has previously been bounded in terms of operator norm : $\|A_{d} - A_{t}\| = \mathcal{O}(T^{2}/M)$~\cite{barends2016digitized}. However, as indicated by the adiabatic theorem, all we need to control is the fidelity error, and by building on the techniques in Corollary \ref{coro:eigen} we will demonstrate that $\epsilon_{\text{tro}}$ has a tighter upper bound that differs from $\|A_{d} - A_{t}\|$. We associate a time-evolving effective Hamiltonian $\widetilde{H}(s)$ with $U^{t}(s)$,
\begin{equation}
\widetilde{H}(s) \equiv i\log(U^{t}(s))/\bar{\delta t}
\end{equation}
Notice that at the beginning point $s=0$ and end point $s=1$, $\widetilde{H}(s)$ is the same as $H_{i}$ and $H_{f}$. So the adiabatic evolution under $\widetilde{H}(s)$ can also transform $|\psi_{i}\rangle$ to $|\psi_{f}\rangle$. Naturally the total error should also be quantified with adiabatic theorem.
\begin{gather}
\epsilon^{\prime}_{\text{tot}} \equiv \|P_{f} - A_{t}P_{i}A^{\dagger}_{t}\|
\end{gather}
Again, we require that the error caused by discretization is negligible:
\begin{gather*}
\epsilon_{\text{tot}} \equiv \|P_{f} - \widetilde{A}(T)P_{i}\widetilde{A}^{\dagger}(T)\|\\
\widetilde{A}(T) \equiv \exp_{\mathcal{T}}\left(-iT\int_{0}^{1}\widetilde{H}(s)ds\right)\\
\widetilde{\epsilon}_{\text{dis}} \equiv |\epsilon^{\prime}_{\text{tot}} - \epsilon_{\text{tot}}| \ll \epsilon_{\text{tot}}
\end{gather*}
Here $\epsilon_{\text{tot}}$ is the adiabatic error of evolution under $\widetilde{H}(s)$, which is also the continuous limit of $\epsilon^{\prime}_{\text{tot}}$. Using the adiabatic theorem we have:
\begin{equation}
\epsilon_{\text{tot}} \le \mathcal{G}(T,\widetilde{H})
\end{equation}
Combining with Lemma \ref{lemma:magnus}, Lemma \ref{lemma:continu} and the triangle inequality $\epsilon_{\text{tro}} \le \epsilon^{\prime}_{\text{adb}} + \epsilon^{\prime}_{\text{tot}}$, we prove the following proposition in Appendix F.
\begin{prop}
Consider a digital adiabatic evolution with initial Hamiltonian $H_{i}$, final Hamiltonian $H_{f}$, total evolution time $T$, and discretization number $M$. If the discretization error is negligible comparing to the continuous limit then
\begin{equation}
\epsilon_{dis} \ll \epsilon_{adb},\quad \widetilde{\epsilon}_{dis} \ll \epsilon_{tot}
\label{equ:c2}
\end{equation}
then the Trotter error has upper bound:
\begin{equation}
\epsilon_{\text{tro}} = \mathcal{O}(\mathcal{G}(T,H) + \mathcal{G}(T,\widetilde{H}))
\end{equation}
with $\mathcal{G}$ introduced in Lemma \ref{lemma:continu}. Define:
\begin{gather*}
C_{0}\equiv\|H_{i}\|,\quad C_{1}\equiv\|[H_{i},H_{f}]\|\\
D\equiv\|H_{i}-H_{f}\|
\end{gather*}
when $D/\lambda \gg 1$, where $\lambda$ is the lower bound of spectral gap during adiabatic evolution under $H(s)$ and $\widetilde{H}(s)$, and $(C_{0} + 3D/2)T<M/4$, we obtain:
\begin{equation}
\epsilon_{\text{tro}} = \mathcal{O}\left(\frac{D^{2}}{T\lambda^{3}}\right) + \mathcal{O}\left(\frac{C_{1}^{2}T}{M^{2}\lambda^{3}}\right)
\end{equation}
\label{prop:1}
\end{prop}
From Proposition \ref{prop:1} we already derive a result different from $\|A_{d} - A_{t}\|$. Numerical results (see Fig. \ref{fig:err}) seem to indicate that $\epsilon_{\text{tro}}$ is close to $\epsilon^{\prime}_{\text{tot}} - \epsilon^{\prime}_{\text{adb}}$, which means this result can be further improved. Also this result only works for $T/M = \mathcal{O}(N^{-1})$.
Though taken for granted in most works, the criterion for Eq : (\ref{equ:c2}) to hold is critical to the error analysis. This question has its independent interest. The discretized adiabatic evolution has been the subject of some analysis~\cite{ambainis2004elementary}, but we believe the result can be intrinsically improved and thus leave it for future work.
\section{Applications}
\subsection{Quantum Phase Estimation}
The QPE algorithm constructs a quantum circuit to detect the phase $\theta$ of a unitary operator: $U|\psi\rangle = e^{i\theta}|\psi\rangle$.
For an exact QPE algorithm, the measurement outcome is the integer $a$ closest to $2^{l}\theta$, where $l$ is the size of the first register. Since it's unlikely for $2^{l}\theta$ to be an integer, there is an inherent error $\xi = \mathcal{O}(2^{-l})$ in this algorithm. The probability of measuring the value closest to the true $\theta$ is at least $4/\pi^{2}$~\cite{nielsen2002quantum}.
The influence of the Trotter error comes from two aspects. Again we regard the Trotterized evolution operator as an exact evolution operator under the effective Hamiltonian $\widetilde{H}$. This effective Hamiltonian has an eigenstate $|\widetilde{\psi}\rangle$, which is very close to initial eigenstate $|\psi\rangle$ : $|\psi\rangle = \sqrt{1-p}|\widetilde{\psi}\rangle + \sqrt{p}|\widetilde{\psi}^{\perp}\rangle$. As a result, the final phase detected should be $\widetilde{\theta}$ and the success rate should be decreased by a factor of $(1-p)$. However, since usually $p$ is much smaller comparing to 1, this change in success rate is almost negligible.
More importantly, the Trotter error in phase $\delta\theta = |\widetilde{\theta} - \theta|$ should satisfy $\delta\theta < \xi$, otherwise the phase error caused by Trotterization will be detected. This relation gives us a constraint on the Trotter error:
\begin{equation}
|\widetilde{E} - E|t \le \xi
\end{equation}
In QPE, $\theta$ should be set to be close to 1 to avoid wasting the accuracy provided by the first register, thus $t = \mathcal{O}(1/
|E|)$. However, we can only guess about $E$ before the algorithm. Here we use $t_{0}$ to denote an appropriate choice of time scale in $U$. Thus
\begin{equation}
L = \mathcal{O}\left(\frac{t_{0}}{\delta t}\right)
\end{equation}
where $L$ is the Trotterization number.
The following result follows from Lemma \ref{lemma:spe}.
\begin{corollary}
Suppose there's a quantum circuit performing QPE, the size of the first register is $l$ thus the inherent error is $\xi = \mathcal{O}(2^{-l})$. The unitary operator is $U = e^{-iHt_{0}}$ where $H$satisfies one of the conditions from Lemma \ref{lemma:spe} in the previous section. To guarantee that the Trotter error in phase is smaller than the inherent error $\xi$, we have:
\begin{gather}
\delta t = \mathcal{O}\left(\frac{1}{N}\sqrt{\frac{\xi}{t_{0}}}\min\{1,\lambda\}\right)\\
L = \mathcal{O}\left(N\sqrt{\frac{t_{0}^{3}}{\xi}}\max\left\{1,\frac{1}{\lambda}\right\}\right)\\
\text{Circuit Depth} = \mathcal{O}\left(\sqrt{\frac{N^{3}t_{0}^{3}}{\xi^{3}}}\max\left\{1,\frac{1}{\lambda}\right\}\right)
\end{gather}
$\lambda$ is the lower bound of spectral gap between initial state $|\psi\rangle$ and its neighboring eigenstates.
\end{corollary}
As a comparison, in general case with $|\widetilde{E} - E| = \mathcal{O}(N\delta t)$, the final circuit depth is $\mathcal{O}(N^{2}t^{2}_{0}/\xi^{2})$.
\subsection{Optimizing Digital Adiabatic Simulation}
Although a better upper bound of $\epsilon_{\text{tro}}$ has been derived in Proposition $\ref{prop:1}$, it is really $\epsilon^{\prime}_{\text{tot}}$ rather than $\epsilon_{\text{tro}}$ that is the most relevant quantity, for it doesn't matter whether the error originates from numerical procedure or the finite size of $T$. Motivated by this, here we elaborate on a different question: given a DAS task with $H_{i}$, $H_{f}$ and $|\psi_{i}\rangle$ specified, and a quantum computer with fixed circuit depth $M$, find the optimal $T$ to minimize the estimated total error $\epsilon^{\prime}_{\text{tot}}$.
There exists a trade-off between Trotter error and adiabatic error: on one hand, $T$ can't be too small as the adiabatic error is inversely proportional to $T$; on the other hand, the error caused by Trotterization increases with $T$ (for a fixed depth $M$). The trade-off is balanced when the two errors are of the same magnitude. We use $T_{c}$ to denote the balanced point, which is exactly the turning point in Fig. \ref{fig:err}, if our estimation of $\epsilon_{\text{tot}}$ is accurate enough. The value of $T_{c}$ depends on the estimation of $\epsilon^{\prime}_{\text{tot}}$. In Proposition \ref{prop:1} we have derived an upper bound for $\epsilon^{\prime}_{\text{tot}}$ and denote it with the function $\mathcal{G}(T,\widetilde{H})$. The critical value of $T_c$ and the optimal error are defined by
\begin{equation}
\epsilon_{\text{opt}} \equiv \inf_{T}\mathcal{G}(T,\widetilde{H}),\quad T_{c}\equiv \arg\min \mathcal{G}(T,\widetilde{H}).
\end{equation}
The next Corollary indicates that $T_{c}$ is proportional to $M$ and their optimal ratio determines the gate complexity of DAS. The result follows from Proposition \ref{prop:1}.
\begin{corollary}[Optimal choice of $T$]
In the setting of Proposition \ref{prop:1}, we intend to perform a digital adiabatic evolution on a quantum device with circuit depth $M$. If $(8C_{0} + 12D)D \le 3C_{1}$, then the upper bound of $\epsilon^{\prime}_{\text{tot}}$ reaches its minimal value at:
\begin{equation}
T_{c} = \frac{2MD}{3C_{1}}
\end{equation}
Accordingly,
\begin{equation}
\epsilon_{\text{opt}} = \mathcal{O}\left(\frac{DC_{1}}{M\lambda^{3}}\right)
\end{equation}
Alternatively, to achieve an error $\epsilon$ we can keep the ratio $T_{c}/M$ fixed and take the depth of quantum circuit $M$ to be:\begin{equation}
\text{Circuit Depth} = \mathcal{O}\left(\frac{NDC_{1}}{\lambda^{3}\epsilon}\right)
\end{equation}
$N$ is the width of quantum circuit and $\lambda$ is the lower bound of spectral gap.
\label{coro:aqc}
\end{corollary}
However, the $T_{c}$ derived still deviates a lot from the turning point in Fig. \ref{fig:err}, because our estimation of $\epsilon^{\prime}_{\text{tot}}$ is much larger than the actual value. A tighter bound for $\epsilon^{\prime}_{\text{tot}}$ would lead to a more accurate estimate of $T_{c}$. Additional factors may contribute to the location of the turning point; for example, if the spectral gap of $\widetilde{H}(s)$ closes as $\delta t$ gets larger, the application of the adiabatic theorem will eventually fail.
\section{Conclusion $\&$ Outlook}
Our main contribution is the observation that refined estimation of Trotter error can be established from effective Hamiltonian $\widetilde{H}$. When the initial state is an eigenstate, we first relate the fidelity error and phase error to the spectrum analysis of $\widetilde{H}$, and find that during evolution, most error accumulates in the phase. Further, if the leading perturbation term of $\widetilde{H}$ vanishes in the eigenbasis of $H$, the Trotter error in energy is reduced from $\mathcal{O}(\delta t)$ to $\mathcal{O}(\delta t^{2})$, which results in improvement of QPE. We remark that the improvement in Trotter step size with phase estimation error $\epsilon$ from $\epsilon$ to $\epsilon^{1/2}$ can be compared to the step size of $\epsilon^{1/2}$ that is obtained from the use of a second-order product formula approximation. Similar results apply to other QPE methods such as robust phase estimation~\cite{russo2020evaluating} (see Appendix G), as long as the Trotterized unitary operator is used. Finally, we show that the spectral analysis method is particularly suitable to analyzing Trotter error in DAS, and we demonstrate how consideration of the various types of error leads to an optimal time parameter $T_{c}$ for DAS when the circuit depth of the quantum circuit is fixed.
There are many targets to pursuit in future work. Here we have only studied Trotter error for the 1st order product formula, and it will be interesting to see whether similar properties exist for higher-order product formulas. One may also seek examples of $f \ll \Delta^{2}$ in time-dependent Hamiltonian situation, in which the initial state is the eigenstate of the initial Hamiltonian, generalizing our results for DAS. Third, numerically the Trotter error in DAS is more close to $\epsilon^{\prime}_{\text{tot}} - \epsilon^{\prime}_{\text{adb}}$ rather than the summation of them, we believe with detail analysis this improvement can be achieved. The forth point is the criterion for the time step scale $T/M$ in DAS that keeps the error caused by discretization negligible. Finally, we would like to try the effective Hamiltonian idea in other quantum simulation algorithms~\cite{campbell2019random,low2019hamiltonian,low2017optimal,berry2015simulating}.
\section{Acknowledgement}
We thank Rolando Somma and Burak \c{S}ahino\u{g}lu for helpful discussions. This material is based upon work supported by the U.S. Department of Energy, Office of Science, National Quantum Information Science Research Centers, Quantum Systems Accelerator (QSA).
|
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigconf]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigconf]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
ACM's consolidated article template, introduced in 2017, provides a
consistent \LaTeX\ style for use across ACM publications, and
incorporates accessibility and metadata-extraction functionality
necessary for future Digital Library endeavors. Numerous ACM and
SIG-specific \LaTeX\ templates have been examined, and their unique
features incorporated into this single new template.
If you are new to publishing with ACM, this document is a valuable
guide to the process of preparing your work for publication. If you
have published with ACM before, this document provides insight and
instruction into more recent changes to the article template.
The ``\verb|acmart|'' document class can be used to prepare articles
for any ACM publication --- conference or journal, and for any stage
of publication, from review to final ``camera-ready'' copy, to the
author's own version, with {\itshape very} few changes to the source.
\section{Template Overview}
As noted in the introduction, the ``\verb|acmart|'' document class can
be used to prepare many different kinds of documentation --- a
double-blind initial submission of a full-length technical paper, a
two-page SIGGRAPH Emerging Technologies abstract, a ``camera-ready''
journal article, a SIGCHI Extended Abstract, and more --- all by
selecting the appropriate {\itshape template style} and {\itshape
template parameters}.
This document will explain the major features of the document
class. For further information, the {\itshape \LaTeX\ User's Guide} is
available from
\url{https://www.acm.org/publications/proceedings-template}.
\subsection{Template Styles}
The primary parameter given to the ``\verb|acmart|'' document class is
the {\itshape template style} which corresponds to the kind of publication
or SIG publishing the work. This parameter is enclosed in square
brackets and is a part of the {\verb|documentclass|} command:
\begin{verbatim}
\documentclass[STYLE]{acmart}
\end{verbatim}
Journals use one of three template styles. All but three ACM journals
use the {\verb|acmsmall|} template style:
\begin{itemize}
\item {\verb|acmsmall|}: The default journal template style.
\item {\verb|acmlarge|}: Used by JOCCH and TAP.
\item {\verb|acmtog|}: Used by TOG.
\end{itemize}
The majority of conference proceedings documentation will use the {\verb|acmconf|} template style.
\begin{itemize}
\item {\verb|acmconf|}: The default proceedings template style.
\item{\verb|sigchi|}: Used for SIGCHI conference articles.
\item{\verb|sigchi-a|}: Used for SIGCHI ``Extended Abstract'' articles.
\item{\verb|sigplan|}: Used for SIGPLAN conference articles.
\end{itemize}
\subsection{Template Parameters}
In addition to specifying the {\itshape template style} to be used in
formatting your work, there are a number of {\itshape template parameters}
which modify some part of the applied template style. A complete list
of these parameters can be found in the {\itshape \LaTeX\ User's Guide.}
Frequently-used parameters, or combinations of parameters, include:
\begin{itemize}
\item {\verb|anonymous,review|}: Suitable for a ``double-blind''
conference submission. Anonymizes the work and includes line
numbers. Use with the \verb|\acmSubmissionID| command to print the
submission's unique ID on each page of the work.
\item{\verb|authorversion|}: Produces a version of the work suitable
for posting by the author.
\item{\verb|screen|}: Produces colored hyperlinks.
\end{itemize}
This document uses the following string as the first command in the
source file:
\begin{verbatim}
\documentclass[sigconf]{acmart}
\end{verbatim}
\section{Modifications}
Modifying the template --- including but not limited to: adjusting
margins, typeface sizes, line spacing, paragraph and list definitions,
and the use of the \verb|\vspace| command to manually adjust the
vertical spacing between elements of your work --- is not allowed.
{\bfseries Your document will be returned to you for revision if
modifications are discovered.}
\section{Typefaces}
The ``\verb|acmart|'' document class requires the use of the
``Libertine'' typeface family. Your \TeX\ installation should include
this set of packages. Please do not substitute other typefaces. The
``\verb|lmodern|'' and ``\verb|ltimes|'' packages should not be used,
as they will override the built-in typeface families.
\section{Title Information}
The title of your work should use capital letters appropriately -
\url{https://capitalizemytitle.com/} has useful rules for
capitalization. Use the {\verb|title|} command to define the title of
your work. If your work has a subtitle, define it with the
{\verb|subtitle|} command. Do not insert line breaks in your title.
If your title is lengthy, you must define a short version to be used
in the page headers, to prevent overlapping text. The \verb|title|
command has a ``short title'' parameter:
\begin{verbatim}
\title[short title]{full title}
\end{verbatim}
\section{Authors and Affiliations}
Each author must be defined separately for accurate metadata
identification. Multiple authors may share one affiliation. Authors'
names should not be abbreviated; use full first names wherever
possible. Include authors' e-mail addresses whenever possible.
Grouping authors' names or e-mail addresses, or providing an ``e-mail
alias,'' as shown below, is not acceptable:
\begin{verbatim}
\author{Brooke Aster, David Mehldau}
\email{dave,judy,<EMAIL>}
\email{<EMAIL>}
\end{verbatim}
The \verb|authornote| and \verb|authornotemark| commands allow a note
to apply to multiple authors --- for example, if the first two authors
of an article contributed equally to the work.
If your author list is lengthy, you must define a shortened version of
the list of authors to be used in the page headers, to prevent
overlapping text. The following command should be placed just after
the last \verb|\author{}| definition:
\begin{verbatim}
\renewcommand{\shortauthors}{McCartney, et al.}
\end{verbatim}
Omitting this command will force the use of a concatenated list of all
of the authors' names, which may result in overlapping text in the
page headers.
The article template's documentation, available at
\url{https://www.acm.org/publications/proceedings-template}, has a
complete explanation of these commands and tips for their effective
use.
Note that authors' addresses are mandatory for journal articles.
\section{Rights Information}
Authors of any work published by ACM will need to complete a rights
form. Depending on the kind of work, and the rights management choice
made by the author, this may be copyright transfer, permission,
license, or an OA (open access) agreement.
Regardless of the rights management choice, the author will receive a
copy of the completed rights form once it has been submitted. This
form contains \LaTeX\ commands that must be copied into the source
document. When the document source is compiled, these commands and
their parameters add formatted text to several areas of the final
document:
\begin{itemize}
\item the ``ACM Reference Format'' text on the first page.
\item the ``rights management'' text on the first page.
\item the conference information in the page header(s).
\end{itemize}
Rights information is unique to the work; if you are preparing several
works for an event, make sure to use the correct set of commands with
each of the works.
The ACM Reference Format text is required for all articles over one
page in length, and is optional for one-page articles (abstracts).
\section{CCS Concepts and User-Defined Keywords}
Two elements of the ``acmart'' document class provide powerful
taxonomic tools for you to help readers find your work in an online
search.
The ACM Computing Classification System ---
\url{https://www.acm.org/publications/class-2012} --- is a set of
classifiers and concepts that describe the computing
discipline. Authors can select entries from this classification
system, via \url{https://dl.acm.org/ccs/ccs.cfm}, and generate the
commands to be included in the \LaTeX\ source.
User-defined keywords are a comma-separated list of words and phrases
of the authors' choosing, providing a more flexible way of describing
the research being presented.
CCS concepts and user-defined keywords are required for for all
articles over two pages in length, and are optional for one- and
two-page articles (or abstracts).
\section{Sectioning Commands}
Your work should use standard \LaTeX\ sectioning commands:
\verb|section|, \verb|subsection|, \verb|subsubsection|, and
\verb|paragraph|. They should be numbered; do not remove the numbering
from the commands.
Simulating a sectioning command by setting the first word or words of
a paragraph in boldface or italicized text is {\bfseries not allowed.}
\section{Tables}
The ``\verb|acmart|'' document class includes the ``\verb|booktabs|''
package --- \url{https://ctan.org/pkg/booktabs} --- for preparing
high-quality tables.
Table captions are placed {\itshape above} the table.
Because tables cannot be split across pages, the best placement for
them is typically the top of the page nearest their initial cite. To
ensure this proper ``floating'' placement of tables, use the
environment \textbf{table} to enclose the table's contents and the
table caption. The contents of the table itself must go in the
\textbf{tabular} environment, to be aligned properly in rows and
columns, with the desired horizontal and vertical rules. Again,
detailed instructions on \textbf{tabular} material are found in the
\textit{\LaTeX\ User's Guide}.
Immediately following this sentence is the point at which
Table~\ref{tab:freq} is included in the input file; compare the
placement of the table here with the table in the printed output of
this document.
\begin{table}
\caption{Frequency of Special Characters}
\label{tab:freq}
\begin{tabular}{ccl}
\toprule
Non-English or Math&Frequency&Comments\\
\midrule
\O & 1 in 1,000& For Swedish names\\
$\pi$ & 1 in 5& Common in math\\
\$ & 4 in 5 & Used in business\\
$\Psi^2_1$ & 1 in 40,000& Unexplained usage\\
\bottomrule
\end{tabular}
\end{table}
To set a wider table, which takes up the whole width of the page's
live area, use the environment \textbf{table*} to enclose the table's
contents and the table caption. As with a single-column table, this
wide table will ``float'' to a location deemed more
desirable. Immediately following this sentence is the point at which
Table~\ref{tab:commands} is included in the input file; again, it is
instructive to compare the placement of the table here with the table
in the printed output of this document.
\begin{table*}
\caption{Some Typical Commands}
\label{tab:commands}
\begin{tabular}{ccl}
\toprule
Command &A Number & Comments\\
\midrule
\texttt{{\char'134}author} & 100& Author \\
\texttt{{\char'134}table}& 300 & For tables\\
\texttt{{\char'134}table*}& 400& For wider tables\\
\bottomrule
\end{tabular}
\end{table*}
Always use midrule to separate table header rows from data rows, and
use it only for this purpose. This enables assistive technologies to
recognise table headers and support their users in navigating tables
more easily.
\section{Math Equations}
You may want to display math equations in three distinct styles:
inline, numbered or non-numbered display. Each of the three are
discussed in the next sections.
\subsection{Inline (In-text) Equations}
A formula that appears in the running text is called an inline or
in-text formula. It is produced by the \textbf{math} environment,
which can be invoked with the usual
\texttt{{\char'134}begin\,\ldots{\char'134}end} construction or with
the short form \texttt{\$\,\ldots\$}. You can use any of the symbols
and structures, from $\alpha$ to $\omega$, available in
\LaTeX~\cite{Lamport:LaTeX}; this section will simply show a few
examples of in-text equations in context. Notice how this equation:
\begin{math}
\lim_{n\rightarrow \infty}x=0
\end{math},
set here in in-line math style, looks slightly different when
set in display style. (See next section).
\subsection{Display Equations}
A numbered display equation---one set off by vertical space from the
text and centered horizontally---is produced by the \textbf{equation}
environment. An unnumbered display equation is produced by the
\textbf{displaymath} environment.
Again, in either environment, you can use any of the symbols and
structures available in \LaTeX\@; this section will just give a couple
of examples of display equations in context. First, consider the
equation, shown as an inline equation above:
\begin{equation}
\lim_{n\rightarrow \infty}x=0
\end{equation}
Notice how it is formatted somewhat differently in
the \textbf{displaymath}
environment. Now, we'll enter an unnumbered equation:
\begin{displaymath}
\sum_{i=0}^{\infty} x + 1
\end{displaymath}
and follow it with another numbered equation:
\begin{equation}
\sum_{i=0}^{\infty}x_i=\int_{0}^{\pi+2} f
\end{equation}
just to demonstrate \LaTeX's able handling of numbering.
\section{Figures}
The ``\verb|figure|'' environment should be used for figures. One or
more images can be placed within a figure. If your figure contains
third-party material, you must clearly identify it as such, as shown
in the example below.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{sample-franklin}
\caption{1907 Franklin Model D roadster. Photograph by Harris \&
Ewing, Inc. [Public domain], via Wikimedia
Commons. (\url{https://goo.gl/VLCRBB}).}
\Description{A woman and a girl in white dresses sit in an open car.}
\end{figure}
Your figures should contain a caption which describes the figure to
the reader.
Figure captions are placed {\itshape below} the figure.
Every figure should also have a figure description unless it is purely
decorative. These descriptions convey what’s in the image to someone
who cannot see it. They are also used by search engine crawlers for
indexing images, and when images cannot be loaded.
A figure description must be unformatted plain text less than 2000
characters long (including spaces). {\bfseries Figure descriptions
should not repeat the figure caption – their purpose is to capture
important information that is not already provided in the caption or
the main text of the paper.} For figures that convey important and
complex new information, a short text description may not be
adequate. More complex alternative descriptions can be placed in an
appendix and referenced in a short figure description. For example,
provide a data table capturing the information in a bar chart, or a
structured list representing a graph. For additional information
regarding how best to write figure descriptions and why doing this is
so important, please see
\url{https://www.acm.org/publications/taps/describing-figures/}.
\subsection{The ``Teaser Figure''}
A ``teaser figure'' is an image, or set of images in one figure, that
are placed after all author and affiliation information, and before
the body of the article, spanning the page. If you wish to have such a
figure in your article, place the command immediately before the
\verb|\maketitle| command:
\begin{verbatim}
\begin{teaserfigure}
\includegraphics[width=\textwidth]{sampleteaser}
\caption{figure caption}
\Description{figure description}
\end{teaserfigure}
\end{verbatim}
\section{Citations and Bibliographies}
The use of \BibTeX\ for the preparation and formatting of one's
references is strongly recommended. Authors' names should be complete
--- use full first names (``Donald E. Knuth'') not initials
(``D. E. Knuth'') --- and the salient identifying features of a
reference should be included: title, year, volume, number, pages,
article DOI, etc.
The bibliography is included in your source document with these two
commands, placed just before the \verb|\end{document}| command:
\begin{verbatim}
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
To date, many social media platforms observed an explosive number of user-created content posted everyday from Twitter to YouTube to Instagram and more.
Following the acceleration of online contents which becomes even faster partly due to COVID-19, it has also become easier for people to access sensitive content that may not be appropriate for the general purpose.
Owing to the scale of these content and users' abilities to share and repost them in a flash, it becomes extremely costly to detect the sensitive content solely by manual work.
Current social media platforms have adopted various (semi)automated content moderation methods including a deep learning-based classification (e.g., Microsoft Azure Content Moderator \cite{MSAzure}, DeepAI's Nudity Detection API \cite{DeeapAINudity}, Amazon Rekognition Content Moderation \cite{rekognition_moderation}).
Meanwhile, since desktop 3D printers have been flooded into the consumer market, 3D printing specific social platforms such as Thingiverse \cite{thingiverse} have also gained popularity, contributing to the proliferation of shared 3D contents that are easily downloadable and replicable among community users.
Despite a massive number of 3D contents shared for free to date---As of 2020 2Q, there are near 1.8 million 3D models available for download, excluding empty entries due to post deletion---, there has been relatively little attention to sensitive 3D contents.
This might result in not only a lack of a dataset to be used as a bench mark, but also
a lack of discussion on fair rationales to be utilized in building a algorithmic 3D content moderation that integrates everyone's perspectives with a different background.
Along with significant advances in technology of machine mechanisms and materials (e.g., 3D printing in metals), the 3D printing community may present an even greater impact from the spread of content due to its limitless potential for replication and reproduction.
In view of various stakeholders who have different perspectives in consuming and interpreting content
---from K-12 teachers who may seek 3D files online to design curricula to artists who depict their creativity in digitized 3D sculptures---, moderating 3D content with fairness becomes more challenging.
3D contents online often consist of images and text that are possibly useful to adopt existing moderation schemes including text (e.g., \cite{prabowo2009sentiment, baccianella2010sentiwordnet, saha2018hateminers, Ahluwalia2018hateSpeech}) or image based (e.g., \cite{minaee2019machine, kumar2020image, llanso2020artificial}) approaches.
However, there exist 3D printing specific features (e.g., print support to avoid overhangs, uni-colored outcome, segmented in parts, etc.) that may prevent direct adoption of those schemes, requiring further consideration about implementing advanced 3D content moderation techniques.
In this work, we first study the existing content moderation efforts that has potential to be used in 3D content moderatio
and discuss shared concerns in examining transparency and fairness issues in algorithmic content moderation.
As a potential solution, we propose a semi-automated human-in-the-loop validation pipeline using augmented learning that incrementally trains the model with the input from the human workforce.
We highlight potential biases that are likely to be propagated from different perspectives of human moderators who provide final decisions and labeling for re-training a classification model.
To mitigate those biases, we propose an image annotation interface to develop an explainable dataset and the system that reflects various stakeholders' perspectives in understanding the 3D content.
We conclude with initial recommendations for metadata design to (1) require consent and (2) inform previously unaware users of consent for publicizing the content which might invade copyright or privacy
\begin{comment}
This lack of attention is mostly due to the common misunderstanding that 3D printing communities are not as popular as other social media platforms. On the contrary, Thingiverse, which is the largest 3D printing community, has over 2 million active monthly users, 1.7 million downloads per month \cite{thingiverse2015downloads}, and over 3 million 3D printable \textit{Things} \cite{thingiverse2018celebrates}. By the virtue of openness of such communities and proliferation of consumer-grade 3D printers, the number of shared contents is rapidly growing, as well as sensitive 3D designs.
\end{comment}
\section{Algorithmic Content Moderation}
Manual moderation relying on a few trusted human workforce and voluntary reports has been common solutions to review shared contents.
Unfortunately, it becomes increasingly difficult to meet the demands of growing volumes of users and user-created content \cite{crowdmoderation}.
Algorithmic content moderation has taken an important place in popular social media platforms to prevent various sensitive content in real-time, including graphic violence, sexual abuse, harassment, and more.
As with other media post
, 3D contents available online appear as web documents that consist of images and text.
For example, to attract audiences and help others understand the design project, creators in Thingiverse voluntarily include various information such as written descriptions of the model, tags, as well as photos of a 3D printed design; thus, 3D content can provide us an ample opportunity to employ the existing text and image based moderation schemes.
Among various text-based solutions, sentiment analysis is one traditionally popular approach that categorizes input text into either two or more categories: positive and negative, or more detailed \textit{n}-point scales (e.g., highly positive, positive, neutral, negative, highly negative) \cite{prabowo2009sentiment, baccianella2010sentiwordnet}.
Moderators can consider categorization results in deciding whether the content is offensive or discriminatory \cite{SentimentAnalysis}.
Various classifiers, such as Logistic Regression Model, Support Vector Machine, and random forest, are actively used in detecting misogynistic posts on Twitter (e.g., \cite{saha2018hateminers, Ahluwalia2018hateSpeech}).
Jigsaw and Google's Counter Abuse Technology suggested Perspective API \cite{perspective_api} provide a score on how \textit{toxic} (i.e., rude, disrespectful, or unreasonable) the text comment is, using a machine learning (ML) model that was trained by people's rating of internet comments.
With the rapid improvement of Computer Vision (CV) technologies with machine learning, several image datasets (e.g., NudeNet Classifier dataset\cite{nudenet_dataset}) and moderation APIs enable developers to apply these ready-to-use mechanisms to their applications.
For example, Microsoft Azure Content Moderator \cite{MSAzure} classifies adult images into several categories, such as explicitly sexual in nature, sexually suggestive, or gory.
DeepAI's Nudity Detection API \cite{DeeapAINudity} enables automatic detection of adult images and adult videos.
Amazon Rekognition content moderation \cite{rekognition_moderation} detects inappropriate or offensive features in images and provides detected labels and prediction probabilities.
However, many off-the-shelf services and APIs are often obscured, because it is hard for users to expect that the models are trained with fair ground-truths that can offer reliable results to various stakeholders with different cultural or social backgrounds without any biases, which we will discuss more in a detailed way in the following section.
\section{Transparency and Fairness Issues in Content Moderation}
\subsection{Transparency: Black Box that Lacks Explanation}
Content moderation has long been controversial due to its non-transparent and secretive process \cite{gorwa2020algorithmic}, resulting from lacking explanations for community members about how the algorithm works.
To meet the growing demands for transparent and accountable moderation practice as well as to elevate public trust, recently, popular social media platforms have begun to dedicate their efforts to make their moderation process more obvious and candid \cite{gorwa2020algorithmic, granados2013transparency, transparency_democracy, maccarthy2020transparency}.
As a reasonable starting point, those services provided detailed terms and policies (e.g., Facebook's Community Standards \cite{facebook_community}) describing the bounds of acceptable behaviors on the platform \cite{gorwa2020algorithmic}.
In 2018, as a collective effort, researchers and practitioners proposed the Santa Clara Principles on Transparency and Accountability in Content Moderation (SCP) \cite{scp}.
SCP suggests one requirement that
social media platforms should provide detailed guidance to the members about which content and behaviors are discouraged, including examples of permissible and impermissible content,
as well as an explanation of how automated tools are used across each category of content.
It also recommends for content moderators to give users a \textit{rationale} for content removal to assure about what happens behind the content moderation.
Making the moderation process transparent and explainable is crucial to the success of the community \cite{juneja2020through}, in order not only to maintain its current scale but also to invite new users, because it may affect users' subsequent behaviors.
For example, given no explanation about the content removal, users are less likely to upload new posts in the future or leave the community, because they may believe that their content was treated unfairly thus get frustrated owing to an absence of communication \cite{jhaver2019did}.
Reddit \cite{Reddit}, which is one of the most popular social media, has equipped volunteer-based moderation schemes resulting in the removal of almost one fifth of all posts every day \cite{jhaver2018did} due to violation of their community policy \cite{reddit_policy} (e.g., Rule 4: \textit{Do not post or encourage the posting of sexual or suggestive content involving minors.}) or individual rules of the subreddits (i.e., subcommunity of Reddit that has a specific individual topic) according to their own objectives (e.g., One of the rules in 3D printing subreddit: ``Any device/design/instructions which are \textit{intended} injure people or damage property will be removed.'').
Users being aware of community guidelines or receiving explanations for content removal are more likely to perceive that the removal was fair \cite{jhaver2019did} and showcase more positive behaviors in the future.
As many social platforms including 3D open communities such as Thingiverse highly rely on voluntary posting of the user-created content \cite{ccomlekcci2019custodians}, the role of a transparent system in content moderation becomes more significant in maintaining the communities themselves.
Even if many existing social media platforms have their full gears to implement artificial intelligence (AI) in content moderation, it has long been in the black box
\cite{juneja2020through}, thus not understandable for users due to the complexity of the ML model.
To address the issue of the uninterpretable model that hinders the users from understanding how it works, researchers shed lights on the blind spot by studying various techniques to make the model \textit{explainable} (e.g., \cite{letham2015interpretable, wang2017bayesian, freitas2014comprehensible}).
Explainability has been on the rise to be an effective way of enhancing transparency of ML models \cite{lepri2018fair}.
In order to secure explainability, the system must enable stakeholders to understand the high-level concepts of the model, the reasoning used by the model, and the model's resulting behavior \cite{bhatt2020explainable}.
For example, as shown in the Fairness, Accountability, and Transparency (FAT) model, supporting users to know which variables are important in the prediction and how they will be combined is one powerful way to enable them to understand and finally trust the decision made by the model \cite{lakkaraju2016interpretable}.
\subsection{Fairness: Implicit Bias and Inclusivity Issues}
People often overlook fairness of the moderation algorithm and tend to believe that the systems automatically make unbiased decisions \cite{garfinkel2017toward}.
In fact, the human adjudication of user-generated content has been occurred in secret and for relatively low wages by unidentified moderators \cite{roberts2019behind}.
In some platforms, users are even unable to know the presence of moderators or who they are \cite{roberts2016commercial}, and thus it is hard for them to know what potential bias, owing to different reasoning processes, has been injected into the moderation procedure.
For example, there have been worldwide actions that strongly criticize the sexualization of women's bodies without inclusive inference (e.g., `My Breasts Are Not Obscene' protest by the global feminist group Femen \cite{femen} to denounce a museum's censorship of nudity.).
Similarly, Facebook's automatic turning down of postings and selfies that include women's topless photo by tagging them as \emph{Sexual/Porn} ignited `My Body is not Porn' movement \cite{MyBodyNotPorn, facebook_porn}.
The different points of view in perceiving and reasoning towards the same piece of work makes it yet hard to decide the absolute sensitivity.
It is nearly impossible that the sole group of users represent all, therefore, it is difficult for users to expect a \textit{ground-truth} in the decision-making process, and trust the result while believing experts made the final decisions based on thoughtful consideration with an unbiased rationale.
Subsequently, many studies (e.g., \cite{blodgett2016demographic, binns2017like}) have explored potential risks of algorithmic decision-making that are potentially biased and discriminatory to a certain group of people such as underrepresented groups of gender, race, disability.
Classifier has been one common approach in content moderation, but developing a perfectly fair set of classifiers in content moderation is complex compared to those in common recommendation or ranking systems, as classifiers tend to inevitably embed a \textit{preference} to the certain group over others to decide whether the content is offensive or not \cite{gorwa2020algorithmic}.
\subsection{Transparency \& Fairness Issue in 3D Content Moderation}
Through a text feature based classification, we identified there are three main categories of sensitive 3D content: (1) sexual/suggestive, (2) dangerous weaponry, and (3) drug/smoke.
Due to the capability of unlimited replication and reproduction in 3D printing, unawareness of these 3D contents could be crucial.
We noticed that Thingiverse limits access to \textit{some} of sensitive things that are currently labeled as NSFW (Not Safe for Work) by replacing their thumbnail images with the black warning images.
It is a secretive process because there are no clear rationale or explanations offered to users behind this process.
Therefore, users cannot expect whether Thingiverse operates based on an unbiased and fair set of rules.
While the steep acceleration of increments of 3D models \cite{thingiverse2018celebrates} is making automatic detection of sensitive 3D content imperative,
moderating 3D content also faces fairness issues and users are suffering from lacking explanations.
We need to take our account into various stakeholders' points of view that affect their decision on potentially sensitive 3D content, as well as further discussions to mitigate bias and discrimination of the algorithmic decision-making system.
Here we propose an explainable human-in-the-loop 3D content moderation system to enable various users who have distinct rules to participate in calibrating algorithmic decisions to decrease bias or discrimination of the algorithm itself.
Although we focus on specific issues in shared 3D content online, our proposed pipeline generally applies to advancing a semi-automatic process toward an explainable and fair content moderation for all.
\section{Challenges of Moderating 3D Content}
\subsection{Challenges in Moderating 3D Content}
As we noted earlier, 3D contents appear as web documents that consist of text descriptions, auto-generated preview images, and user-uploaded images to help others comprehend the content at a glance.
Although it is technically possible to utilize existing text and image based moderation schemes, 3D models have unique features that make it hard to directly adopt the existing CV techniques to their rendered images or photos.
\subsubsection{3D specific features that hamper the use of existing CV techniques}
We identified four characteristics that make sensitive elements undetectable by the existing algorithms.
\begin{figure*}%
\centering
\subfloat[\centering Rotated model]{{\includegraphics[width=.5\columnwidth]{figures/2_align.png} }}%
\subfloat[\centering Support structure]{{\includegraphics[width=.505\columnwidth]{figures/2_support.png} }}%
\subfloat[\centering Texture on surface]{{\includegraphics[width=.494\columnwidth]{figures/2_texture.png} }}%
\subfloat[\centering Divided into parts]{{\includegraphics[width=.498\columnwidth]{figures/2_part.png} }}%
\caption{Example images for the mainly 4 characteristics that make it hard to use the existing CV techniques; each thing is reachable using its unique ID through the url of https://thingiverse.com/thing:ID}%
\label{fig:challenges}%
\end{figure*}
\noindent
\textbf{Challenge 1. Difficulties in Locating Features from Images of the Current Placement.}
Thingiverse automatically generates rendered images of the 3D model when a 3D file is uploaded, and this is used as a representative image if the designer does not provide any photos of real 3D prints.
In many cases, these files are placed in the best orientation that guarantees print-success in FDM (Fused Deposition Modeling) printers, aligning the design to minimize overhangs.
As the preview is taken in a fixed angle, so it might not be in a \textit{perfect} angle that shows the main part of the model thoroughly (e.g., Fig~\ref{fig:challenges}(a)).
It hinders
the existing image-based APIs from accurate detection of sensitivity in the preview images, because sensitive parts might not be visible.
\vspace{2.0ex}
\noindent\textbf{Challenge 2. Support Structure that Occludes the Features.}
Following the model alignment strategy of FDM printing, designers often include a custom support structure to prevent overhangs and to avoid printing failures and deterring surface textures with auto-generated supports from slicers (i.e., 3D model compiler) such as Cura \cite{cura}.
These special structures easily occlude the design's significant features (e.g., Fig~\ref{fig:challenges}(b)).
Since the model is partly or completely occluded, the existing CV techniques barely detect sensitivity of the design.
\vspace{2.0ex}
\noindent\textbf{Challenge 3. Texture and Colors.}
Current 3D printing technologies enable users to use various print settings and other postprocessing techniques.
Accordingly, the printed model may present unique appearances compared to general real-world entities.
Often the model is single-colored and can have a unique texture such as linear lines on the surface (e.g., Fig~\ref{fig:challenges}(c)) due to the nature of 3D printing mechanisms of accumulating materials layer-by-layer, which might let the existing CV algorithms overlook the features.
\vspace{2.0ex}
\noindent\textbf{Challenge 4. Models Separated into Parts for Printing.}
As one common 3D printing strategy to minimize printing failures from a complex 3D designs such as a human body, many designers divide their models into several parts to ease the printing process, and let users post-assemble as shown in Fig~\ref{fig:challenges}(d).
In this case, it is hard for the existing CV techniques to get the whole assembled model, resulting in a failure to recognize its sensitivity.
\section{Towards Explainable 3D Moderation System}
A potential solution to examine 3D contents' sensitivity with fairness is employing the human workforce with ample experiences in observing and perceiving with various perspectives.
We suggest a human-in-the-loop pipeline, based on the idea of incremental learning \cite{RN15} that the human workforce can collaborate with an intelligent system, concurrently classifying data input and annotate features with the explanation for the decision.
\subsection{Building an Inclusive Moderation Process}
Making decisions on the sensitivity of a 3D model can be subjective due to various factors such as cultural differences, the nature of the community, and the purpose of navigating 3D models.
To reflect different angles in discerning the nature and intention of contents, we need to deliberate various interpretations taken from various groups of people.
For example, there are lots of 3D printable replicas of artistic statues or Greek sculptures that are reconstructed by 3D scanning of the original in the museums \cite{openculture}.
Speculative K-12 teachers designing their STEAM education curriculum using 3D models are not likely to want any NSFW designs revealed to their search results.
On the other hand, there are many activists and artists who may want to investigate the limitless potential of the technology, sharing a 3D scanned copy of the naked body of herself \cite{vice} or digitizing nude sculptures available in the museum to make the intellectual assets accessible to everyone, etc.
The nude sculpture has been one popular form of artistic creation in history, and it is not simple to stigmatize these works as `sensitive'.
Everyone has their own right to `leave the memory of self' in a digital form.
Forcing to adapt a preset threshold of sensitivity and filter these wide array of user-created contents could unfairly treat one's creative freedom.
As the extent that various stakeholders perceive the sensitivity could be distinct,
our objective is to design an inclusive process in accepting and adopting the sensitivity.
\subsection{Solution 1: Human-in-the-loop with Augmented Learning}
Automated content moderation could help review of a vast amount of data and provide filtered cases for humans to support a decision-making process \cite{jhaver2019did}, if we well-echo diverse perspectives in understanding contents.
\begin{figure*}%
\centering
\subfloat[\centering Human-in-the-loop pipeline]{{\includegraphics[width=\columnwidth]{figures/pipeline.png} }}%
\subfloat[\centering User interface mockup]{{\includegraphics[width=\columnwidth]{figures/interface.png} }}%
\quad
\caption{(a) Overview of the human-in-the-loop pipeline powered by human moderators to acknowledge various perceptions of sensitivity and (b) an user interface mockup for the moderators to validate prediction results and provide annotations regarding their rationale, thus to augment the model.}%
\label{fig:pipeline_interface}%
\end{figure*}
In our proposal of the human-in-the-loop pipeline (Fig~\ref{fig:pipeline_interface}(a)), an input image dataset of 3D models will be used for the initial model training, then the result will be reviewed by multiple human moderators step by step.
We trained the model with 1,077 things that are already labeled as NSFW by Thingiverse and 1,077 randomly selected non-NSFW things.
All input images are simply categorized as NSFW or not, with no annotation for specific image features to provide the reasoning.
Human moderators recruited from various groups of people now review the classification results whether they agree.
They are asked to annotate image segments using a bounding box where they referred to make the final decision with the category.
At the same time, they provide the rough level of how much the part affected the entire sensitivity and a written rationale for the decision.
These features will enhance the data quality so to be used to fine-tune the model with the weighted score, thus the model becomes able to recognize previously unknown sensitive models based on the similarity and now can \textit{explain} sensitive features.
When two different groups of people with different standards do not agree on the same model's classification results,
the model uses their decision, annotated features, and levels of sensitivity to differentiate the extent of perceived sensitivity and reflect to the different threshold.
For example, one moderator thinks that the model is sensitive while the other does not, the model will have a higher threshold in categorizing the content.
Different decisions on the same model finally could be brought to the table for further discussion if needed, for example, to regulate policy guidelines, or used as search criteria for other community users who have similar goals in viewing and unlocking analogous 3D contents.
To summarize, one iteration contains the following steps:
\begin{enumerate}
\item The pre-trained model presents prediction results.
\item The human moderator can enter disagreement/agreement with the results and annotate sensitive parts with a sensitivity level and a decision rationale.
\item The annotated image is used to fine-tune the model.
\item If the decision for the image is different from other moderators, annotations and sensitivity levels are used to set the different threshold.
\end{enumerate}
We elaborate more on feedback from the moderators by showing three possible scenarios: (1) the moderator's agreement with the prediction results, (2) sensitive parts not detected, and (3) false-classification of insensitive features sensitive.
\noindent\textbf{Case 1. Agreement with the Prediction Result}
In case that the moderators agree with the decision, they can either finalize it or reject the classification,
by selecting provided top-level categories (e.g., sexual/suggestive, weaponry, drug/smoke) and second-level categories (e.g., under sexual/suggestive, explicit nudity, adult toys, sexual activity, etc.).
We currently refer to a two-level hierarchical taxonomy of Amazon Rekognition to label categories of inappropriate or offensive content.
\noindent\textbf{Case 2. Sensitive Parts Ignored by the Algorithm}
Another possible case is that the specific feature in the image that the moderator perceives as sensitive is missing in the detection results.
In this case, human moderators can label that part and provide rationales using \textit{enter the level of sensitivity} field from 1 (slightly sensitive) to 5 (highly sensitive), how each specific part affects the entire sensitivity of the model.
\noindent\textbf{Case 3. False Negative}
It is also possible that some parts detected by the model are not sensitive for the moderator due to the higher tolerance to sensitivity.
The moderator can either submit the disagreement or provide more detailed feedback by excluding specific results.
Different \textit{degrees} of sensitivity perception from various stakeholders can reflect distinct points of view, which may manifest fairness in algorithmic moderation through multiple iterations of this process.
In our interface for the end-users that assists searching 3D designs, we let users set their desired threshold.
For those who might find it difficult to decide a threshold that perfectly fits their need, we show several random example images that have detected sensitive labels with the corresponding threshold.
This pipeline also helps obtain the explainable moderation algorithm.
Our model can help users understand the rationales of the model by locating detected features/prediction probabilities in the image and providing written descriptions that the moderators entered for data classification.
\subsection{Solution 2: New Metadata Design to Avoid Auto-Filtering}
Another potential problem in open 3D communities is copyright or privacy-invasive contents that are immediately marked as NSFW by Thingiverse indicating they are \textit{inappropriate}.
Currently, Thingiverse lacks notification and explanation for content removal, while a majority of them might invade copyrights.
Its obscurity results in a negative impact on the user's future behaviors.
For example, creators are frustrated at the un-notified removal of their content thus decided to quit their membership (e.g., \cite{quit_thingiverse}), which might not happen if they saw an informative alert when they post the content.
Along with advanced 3D scanning technologies
\cite{all3dp}, many creators are actively sharing 3D scanned models (e.g., As of December 2020, Thingiverse has 1150 things that tagged with `3D\_scan' and 308 things with the tag `3D\_scanning').
With arising concerns over possible privacy invasion in sensitive 3D designs, what caught our attention is 3D scanned replicas of human bodies.
Many of them do not include an explicit description of whether the creator received the consent from the subject (e.g., \cite{scan_amber, scan_mel}).
Some designers quoted the subject's permission,
for example, one creator describes that the subject, Nova, has agreed to share her scanned body on Thingiverse \cite{scan_nova}.
Still, this process relies on the users' voluntary action given no official guidelines, resulting in a lack of awareness that the users must be granted the consent to upload possibly privacy-invasive contents at the time of posting those content in public spaces regardless of the commercial purpose.
Without explicit consent, the content is very likely to be auto-filtered by Thingiverse, which decreases fairness by hampering artistic/creative freedom.
To iron out a better content-sharing environment in the these open communities, redesigning of metadata must be considered and adapted by system admins that invoke responsible actions.
For example, providing a checkbox that asks \textit{``If the design is made of 3D scanned human subject, I got an agreement from the subject''} can inform previously unaware users about the need for permission to post potentially privacy-breaching contents.
Including the subject's consent can also protect creative freedom from auto-filtering, by adding that the content is not breaching copyright or privacy and can be shared in the public spaces.
In addition, it can enable users to understand that an absence of consent could be the reason for filtering.
\section{Conclusion}
As an inclusive process to develop transparent and fair moderation procedure in 3D printing communities, our study proposes to build an explainable human-in-the-loop pipeline.
We aim to employ diverse group of human moderators to collect their rationales, which can be used to enhance the model's incremental learning.
Our objective is not to censor 3D content but to build a pleasant 3D printing community for all, by safeguarding search as well as guaranteeing creative freedom, through the pipeline and new metadata design that has potential to minimize issues related with privacy or copyright.
|
\section{Introduction to tumor forecasting}\label{sec:intro}
Cancers are highly \emph{heterogeneous} diseases supported by diverse biological mechanisms occurring, interacting, and evolving at multiple spatial and temporal scales \cite{Marusyk2010}. These phenomena span from the phenotypic and genotypic cellular diversity within the tumor to the regional variations of the tumor microenvironment (e.g., vasculature and extracellular matrix), which can result in epigenetic changes in cancer cells or gradients in nutrient availability. Hence, the heterogeneous nature of cancer makes each patient's case unique.
However, established, standard-of-care methods determine diagnosis, stage, treatment regimen, and response to treatment according to historical population averages.
This paradigm only enables the observation of cancer evolution and the outcome of treatment at fixed time points, offers a limited individualization of disease management, and largely ignores the intrinsic heterogeneity of cancers, which may result in treatment failure \cite{Jain1998,Withers1985}.
Thus, a new clinical paradigm that effectively integrates the spatiotemporal dynamics of tumor growth and treatment response to identify effective clinical strategies for each patient is desperately needed. We posit that mathematical modeling informed by clinically-relevant data can provide the framework to address this challenge \cite{Yankeelov2013,Yankeelov2015,Rockne2019}.
\emph{Computational oncology} is a rapidly growing field that attempts to leverage mathematical models of the key biological mechanisms that characterize cancer to predict how a patient's tumor will grow and respond to treatment \cite{Rockne2019,Yankeelov2013,Yankeelov2015}.
Computer simulations of these models provide personalized \emph{tumor forecasts}, designed to ultimately assist oncologists in clinical decision-making.
For example, tumor forecasts may predict disease progression, thereby providing much needed guidance on the optimal intervention strategy early in the course of therapy.
Indeed, we hypothesize that treatment optimization can be achieved through the development and rigorous validation of practical mathematical models and efficient computational methods that can provide accurate personalized predictions of cancer development and treatment response.
A fundamental challenge in computational oncology is accomplishing the \emph{patient-specific parameterization} of the biological mechanisms involved in cancer models (e.g., tumor cell mobility, proliferation and death rates, or therapy efficacy). In general, these parameters are extremely difficult to measure \emph{in vivo} in human tumors.
However, medical imaging may provide a viable source of data for this purpose.
Clinical oncology currently focuses on \emph{anatomical imaging} for the diagnosis, treatment, monitoring, and assessment of therapeutic response of solid tumors \cite{Therasse2000} (e.g., measuring tumor size, identifying invasion into adjacent structures, and detecting metastasis).
Unfortunately, anatomical imaging frequently fails to capture the heterogeneous underlying biology within tumors.
Alternatively, \emph{quantitative imaging} techniques enable the measurement of clinically-relevant biological features of tumors, such as tumor cell density, blood volume fraction, and perfusion \cite{HormuthII2019}.
Thus, these quantitative imaging techniques can be used to assess the spatiotemporal evolution of a cancer's heterogeneous architecture, morphology, growth dynamics, and response to therapy, thereby providing the necessary data to parameterize predictive models of tumor growth and treatment \cite{Yankeelov2013,HormuthII2019a}.
In this chapter, we will discuss how quantitative imaging can be used to enable tumor forecasting and optimization of therapeutic response. We will begin by identifying relevant quantitative imaging data types and how they are incorporated into existing image-based models of cancer growth and treatment. We will also provide insights into the technical aspects of numerical implementation, model calibration, and model selection. Then, we will introduce a promising framework to optimize patient treatment plans. We will conclude with a discussion of the barriers to successfully translating image-based computational tumor forecasting into patient care.
\section{Relevant data types from medical imaging}\label{sec:img}
While measuring tumor size throughout therapy is central in oncological response assessments \cite{Therasse2000}, the dynamics of tumor size changes are often temporally downstream of intratumoral biological and physiological responses to therapy.
Magnetic resonance imaging (MRI) and positron emission tomography (PET) provide non-destructive and non-invasive 3D quantitative measurements of biological properties within and around the tumor. Hence, the acquisition of these imaging data at several timepoints is well-suited to initialize and parameterize mathematical models of tumor growth and treatment response.
In this section, we will briefly introduce the relevant MRI and PET measurements that have been commonly used in computational oncology (see Fig.~\ref{fig:img} for representative images of these techniques). For a detailed review of advanced MRI and PET techniques in oncology, the reader is respectively referred to \cite{HormuthII2019} and \cite{Gambhir2002}.
\subsection{Diffusion weighted magnetic resonance imaging}\label{sec:dwi}
Diffusion weighted (DW-) MRI is an established technique that has been applied in oncology as a noninvasive assessment of cellularity changes during treatment \cite{Padhani2009}. DW-MRI is sensitive to the diffusion of water molecules within tissue.
In a DW-MRI experiment, water molecules are first tagged based on their spatial location. Then, after a short delay of typically 20–-60 ms, a second spatial-encoded tag is applied. During this delay, water molecules move throughout the tissue due to diffusion. If the water molecules do not travel far, the first spatial-encoded tag can be largely removed by the second spatial-encoded tag and there is no loss (or gain) in signal intensity. However, if the water molecules move throughout the domain, there is a net-difference between the two spatial-encoded tags resulting in a decrease in signal intensity. Thus, the signal intensities within each voxel in the resulting image are ``weighted'' based on water diffusion. In practice, several diffusion weighted experiments are performed with different settings (e.g., varied diffusion-sensitizing gradient amplitudes of the magnetic field) to spatially quantify the apparent diffusion coefficient ($ADC$) of water. However, water diffusion in tissue is heavily restricted by cells, macromolecules, and extracellular structures. Hence, these physical barriers reduce the measured $ADC$. This phenomenon has been observed in several studies showing an inverse correlation between $ADC$ and cellularity \cite{Anderson2000,Jiang2016,Barnes2015}. Following these reports' results, $ADC$ can be used to estimate cellularity using:
\begin{equation}\label{eq:NADC}
N(\vec{x},t) = \theta\left( \frac{ADC_w - ADC(\vec{x},t)}{ADC_w - ADC_{\min}}\right)
\end{equation}
where $\theta$ represents the maximum tumor cell carrying capacity for an imaging voxel (determined by the voxel dimensions and assumptions in cell geometry and packing density), $ADC_w$ is the $ADC$ of free water at 37$^{\circ}$C (i.e., $2.5\cdot 10^{-3}$ mm$^2$/s; \cite{Whisenant2014}), $ADC(\vec{x},t)$ is the $ADC$ value at a given 3D position $\vec{x}$ and time $t$, and $ADC_{\min}$ is the minimum $ADC$ value observed within the tumor. Fig.~\ref{fig:img} displays a representative $ADC$ map from breast and brain cancer.
While there are significant correlations between cellularity and the measured $ADC$, cellularity is not the sole factor in dynamic changes in $ADC$. Changes in cell size, cell permeability, and tissue tortuosity may alter the measured $ADC$ \cite{Padhani2009}. Other diffusion-based imaging approaches can also report on cell size \cite{Jiang2016} and diffusion anisotropy \cite{Sundgren2004}. The reader is referred to \cite{Koh2007} for a technical review of DW-MRI and its applications in oncology.
\subsection{Dynamic contrast-enhanced magnetic resonance imaging}\label{sec:dce}
Dynamic contrast-enhanced (DCE-) MRI consists of the rapid acquisition of a series of heavily $T_1$-weighted images before, during, and after the injection of a $T_1$ altering-contrast agent (typically a Gadolinium chelate) to probe vascular properties in tissue \cite{Yankeelov2009}. Using a pre-contrast $T_1$ map, any post-contrast $T_1$ changes can be related to the concentration of the contrast agent. Thus, each image voxel yields a signal intensity time course that can be related to the concentration of the contrast agent within that voxel. The subtraction images obtained from pre- and post-contrast enhanced images are often used to identify tumor regions, which usually show areas of rapid and intense enhancement due to their higher and more permeable vascularity than the neighboring healthy tissue. The dynamics of signal intensity are commonly analyzed with a two-compartment pharmacokinetic model describing the extravasation of the contrast agent from the plasma space to the tissue space \cite{Yankeelov2009}. The solution to this model is given by
\begin{equation}\label{eq:dce}
C_t(\vec{x},t) = K^{trans}(\vec{x})\int_0^t C_p(u)e^{-\frac{K^{trans}(\vec{x})}{v_e(\vec{x})}(t-u)}du + v_p(\vec{x})C_p(t),
\end{equation}
\noindent where $C_t(\vec{x},t)$ is the concentration of the contrast agent in tissue at position $\vec{x}$ and time $t$, $C_p(t)$ is the concentration of the contrast agent in the plasma space at time $t$, $ K^{trans}(\vec{x})$ is the volume transfer constant from the plasma to tissue space, $v_e(\vec{x})$ is the extravascular-extracellular volume fraction, and $v_p(\vec{x})$ is the plasma volume fraction. Importantly, $C_t$, $K^{trans}$, $v_e,$ and $v_p$ are all voxel-specific and are related to structural (cell density) and physiological (vessel permeability and perfusion) properties. $C_p(t)$ can be measured directly for individual subjects from a large artery within the image field of view or can be replaced with a population-based estimate \cite{Li2011}. Fig.~\ref{fig:img} shows $K^{trans}$, $v_e,$ and $v_p$ maps from a preclinical and a clinical study.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{figure_1_v2}
\caption{Representative quantitative imaging measurements from clinical and preclinical settings. (a) Parameter maps extracted from DW-MRI, DCE-MRI, and $^{18}$FDG-PET through the central slice of a breast tumor (all MR data are from the same patient, PET data are from a different individual).
DW-MRI provides estimates of $ADC$, while DCE-MRI yields estimates of $K^{trans}$ and $v_e$. The $^{18}$FDG-PET $SUV$ map shows increased glucose uptake within the breast tumor relative to surrounding tissue. (b) Parameter maps acquired in a preclinical murine model of glioblastoma from DW-MRI ($ADC$) and DCE-MRI ($K^{trans}$, $v_e$, and $v_p$). }
\label{fig:img}
\end{figure}
\subsection{Molecular imaging with positron emission tomography}\label{sec:pet}
PET relies on the injection of a radiopharmaceutical (or PET tracer) to generate image contrast. As there is no endogenous signal, PET has excellent sensitivity to detect and localize the distribution of radiopharmaceuticals throughout the body. Several radiopharmaceuticals have been developed to probe tumor properties, such as glucose metabolism (\emph{via} $^{18}$F-fludeoxyglucose or $^{18}$FDG), hypoxia (\emph{via} $^{18}$F-fluoromisonidazole or $^{18}$F-MISO), cellular proliferation (\emph{via} $^{18}$F-Flurodeoxythymidine \cite{Woolf2014}), and receptor status (e.g., $^{64}$Cu-diethylenetriaminepen-taacetic acid Trastuzumab for HER2+ positive cancers \cite{Mortimer2018}). We will primarily focus on $^{18}$FDG and $^{18}$F-MISO as they are well-established in oncology, but the analysis techniques are similar for other PET tracers.
In an $^{18}$FDG-PET study, a single image is acquired following the injection of the glucose analogue $^{18}$FDG, which cells uptake in a similar fashion to glucose.
However, once internalized, $^{18}$FDG is phosphorylated and trapped intracellularly. The resulting image intensities are proportional to the concentration of $^{18}$FDG within each voxel. PET tracer uptake can be quantified using the standardized uptake value ($SUV$), which is the ratio of the concentration of $^{18}$FDG radioactivity in tumor tissue to the total injected dose and divided by the patient's body weight. In oncology studies, contrast between tissues is typically generated due to variations in glucose uptake due to an overexpression of glucose transporters and hexokinase activity in tumor cells relative to healthy cells \cite{Castell2008}. This difference in $^{18}$FDG uptake is also shown in Fig.~\ref{fig:img}.
Likewise, in $^{18}$F-MISO PET a single image is also acquired following the injection of $^{18}$F-MISO, which is a radiopharmaceutical that produces images sensitive to oxygen concentration in tissue \cite{Rajendran2015}. After $^{18}$F-MISO is internalized by cells, it is reduced to produce a radical anion. In normoxic or oxygen-rich environments, oxygen accepts the electron from the radical anion enabling $^{18}$F-MISO to leave the cell. Conversely, in hypoxic or oxygen-poor environments, the radical anion of $^{18}$F-MISO binds to other intracellular macromolecules trapping it within the cell. Thus, the concentration of $^{18}$F-MISO and the produced PET signal within a voxel are inversely proportional to the oxygen concentration. $^{18}$F-MISO uptake is quantified using the standardized uptake value ($SUV$) or the oxygen enhancement ratio ($OER$), which is the ratio of signal intensity in tumor relative to blood.
\section{Image-based mathematical models of cancer}\label{sec:cmodels}
Medical imaging provides an excellent way to develop, calibrate, and validate personalized mathematical models of cancer evolution and treatment response \cite{Yankeelov2013} for three main reasons.
First, medical imaging enables the \emph{in vivo} measurement of relevant biological properties in tumor and healthy tissues, which would otherwise be impractical or impossible to measure in individual patients.
Second, medical imaging data can be obtained frequently throughout the clinical management of the patient's tumor, which enables model calibration.
Third, medical imaging data are acquired on a regular voxel grid, which facilitates their computational processing.
In this section, we discuss common image-based models of tumor growth and treatment response that leverage the quantitative imaging data types introduced in Section~\ref{sec:img}. Fig.~\ref{fig:models} shows simulation outputs of many of the models discussed in this section.
\subsection{Baseline tumor growth models}\label{sec:base}
In mathematical oncology \cite{Rockne2019}, the \emph{logistic growth} model is one of the simplest and most common approaches to describe changes in tumor volume \cite{Benzekry2014} or cell number \cite{Atuegwu2011} over time. It is a flexible model that can be adapted to \textsl{in vitro} and \textsl{in vivo} data alike. The formulation of the logistic growth model over a certain tissue region of interest follows the partial differential equation (PDE):
\begin{equation}\label{eq:log}
\frac{\partial N(\vec{x},t)}{\partial t}=k(\vec{x})N(\vec{x},t)\left(1-\frac{N(\vec{x},t)}{\theta}\right),
\end{equation}
where $N(\vec{x},t)$ is the \emph{tumor cell density} at position $\vec{x}$ and time $t$, $k(\vec{x})$ is a spatially-varying net proliferation rate, and $\theta$ is the carrying capacity.
The image-informed applications of this model have typically been posed voxelwise, such that $N(\vec{x},t)$ is redefined as the number of cancer cells within the voxel in position $\vec{x}$ at time $t$.
Atuegwu \textsl{et al.} \cite{Atuegwu2011} used this approach to predict tumor growth in breast cancer patients receiving neoadjuvant chemotherapy.
First, they used Eq.~\eqref{eq:NADC} to estimate $N(\vec{x},t)$ from $ADC$ maps obtained \emph{via} DW-MRI. The estimates of $N$ at baseline (pre-treatment) and after one cycle of therapy were used to determine $k(\vec{x})$.
Then, Atuegwu \textsl{et al.} used their model equipped with the resulting $k(\vec{x})$ to predict $N$ at the conclusion of therapy. They observed a strong correlation between the predictions and data estimates of $N$ over the entire tumor (Pearson correlation coefficient, PCC, of 0.95) and for individual voxels (PCC=0.70). Fig.~\ref{fig:models} shows an example of this approach for a clinical breast cancer model and a preclinical glioblastoma model.
However, the logistic growth model fails to capture the potential movement of cells that may occur over time. To overcome this limitation, the logistic growth model can be extended to a \emph{reaction-diffusion} model given by the PDE
\begin{equation}\label{eq:rd}
\frac{\partial N(\vec{x},t)}{\partial t} = \nabla\cdot\left(D \nabla N(\vec{x},t)\right) + k(\vec{x})N(\vec{x},t)\left(1-\frac{N(\vec{x},t)}{\theta}\right),
\end{equation}
where the first term on the right is a diffusion term describing the movement of tumor cells with a constant diffusion coefficient $D$, while the second term on the right is a reaction term represented by the logistic growth of cancer cells.
Eq.~\eqref{eq:rd} is well established in computational oncology \cite{Hormuth2019,Weis2013,Jarrett2018a,Rockne2015,Swan2018,Hogea:2008b,Wong:2017a,Roque2018}.
The work of Swanson \textsl{et al.} \cite{Baldock2013} in high grade gliomas showed one of its first image-informed applications, using anatomical MRI data to provide segmentation of tumor boundaries and fixed cell-density counts in enhancing and non-enhancing disease. The spatiotemporal changes in tumor boundaries were used to estimate a constant tumor-specific proliferation rate $k$ and tissue-specific $D$, which were then used to predict patient survival. This approach has had promising results in relating growth kinetics to patient outcomes \cite{Baldock2013,Neal2013}. However, it does not describe the intratumoral heterogeneity of cell density.
Hormuth \textsl{et al.} \cite{Hormuth:2015a} addressed this limitation by estimating $N$ from $ADC$ maps obtained \emph{via} DW-MRI using Eq.~\eqref{eq:NADC} in a murine model of glioma. In this study, animals were imaged up to seven times over ten days. The first three imaging datasets were used to initialize $N$ as well as to calibrate $k(\vec{x})$ and $D$. In a separate calibration, a spatially-constant $k$ (i.e., tumor specific) was also calibrated along with $D$. The calibrated model was then used to predict $N(\vec{x},t)$ at the remaining imaging visits. While both calibration scenarios overestimated future tumor growth, the predictions with a spatially-varying $k(\vec{x})$ rendered lower tumor volume errors, higher Dice correlation coefficients, and higher concordance correlation coefficients (CCC; all $p<0.05$). These results highlight the importance of accounting for the intratumoral heterogeneous dynamics to obtain accurate tumor forecasts and the promising potential of quantitative imaging to provide the required data for this purpose. Other studies have used alternative quantitative imaging measures to inform the reaction-diffusion model in Eq.~\eqref{eq:rd}; examples include incorporating anisotropic diffusion \emph{via} diffusion tensor imaging (DTI, a form of DW-MRI \cite{Sundgren2004}) \cite{Swan2018,Jbabdi2005}, using cell density measurements \emph{via} contrast-enhanced computed tomography \cite{Wong:2017a}, and estimating cell phenotypes \emph{via} DCE-MRI \cite{Roque2018}.
Alternatively, \emph{phase-field models} are another common paradigm to describe tumor growth \cite{Lima2016,Agosti2018,Lorenzo2016,Lorenzo2019,Xu2020}. The \emph{phase field} $\phi(\vec{x},t)$ identifies healthy tissue (e.g., $\phi=0$) from tumor tissue (e.g., $\phi=1$), showing a smooth and thin transition between either region. Phase-field models usually focus on the dynamics of tumor morphology through the evolution of the healthy-tumor interface, which is implicitly defined by a phase-field isosurface.
These models rely on more complex physics than those described above, usually requiring the definition of an energy functional that drives tumor growth \cite{Gomez2017,Lima2016,lima17}.
Phase-field modeling has been notably applied in brain tumors \cite{Lima2016,Agosti2018}, prostate cancer \cite{Lorenzo2016,Lorenzo2019}, and tumor angiogenesis \cite{Xu2020,Vilanova2017}.
However, there is a paucity of studies using quantitative imaging data to inform phase-field models. This is partly due to their more complex dynamics, which usually requires a higher number of parameters, larger and richer patient-specific datasets, more advanced numerical methods, and more computational resources.
We identified two illustrative works that use imaging measurements to initialize, calibrate, and/or validate phase-field models of cancer.
Lima \textsl{et al.} \cite{Lima2016,lima17} have been investigating model selection to identify the best formulation of murine glioma growth according to longitudinal anatomical MRI tumor measurements, including an array of phase-field models also accounting for the local tumor-induced mechanical stress field (see Section~\ref{sec:mech}). Their work shows that phase-field models are plausible formulations of tumor growth and in \cite{Lima2016} they emerge as the best models indeed.
Additionally, Agosti \textsl{et al.} \cite{Agosti2018} developed a phase-field model of glioblastoma multiforme that uses quantitative DTI data to define anisotropic tumor cell motility and nutrient diffusion. Their work focuses on the prediction of tumor recurrence after surgical resection and subsequent radiotherapy. By accounting for post-surgery changes in tissue architecture, they obtained a Jaccard index of 0.71 post-radiotherapy.
\subsection{Mechanically-coupled models}\label{sec:mech}
Local mechanical tissue properties and tumor-induced mechanical stresses are known to affect cancer growth dynamics \cite{Helmlinger1997,Nagelkerke2015,Jain2014}.
For example, Helmlinger \textsl{et al.} \cite{Helmlinger1997} observed that tumor spheroid growth \textsl{in vitro} was increasingly inhibited as the substrate matrix stiffness was augmented. Uncontrolled tumor growth can also severely deform healthy tissue structures, thereby adversely impacting patient health and quality of life. Therefore, several mathematical models of cancer couple tumor growth dynamics with local mechanical equilibrium \cite{Lorenzo2019,Jarrett2018a,Weis2013,Hogea:2008b,HormuthII2017,Liu2014,Clatz2005}.
\begin{figure}[!t]
\centering
\includegraphics[width=\textwidth]{figure_2_v2}
\caption{Differences between image-based models in clinical and preclinical settings. (a) Example from a clinical breast cancer model, where measured initial, measured final, and model forecasts of the final distributions of tumor cell density are shown. The logistic model fails to capture the expansion of the tumor into nearby tissue, while the remaining models incorporating a diffusion term perform better. (b) Example from a preclinical model of glioblastoma, where measured initial, measured final, and model predicted final distributions of tumor cell density are shown. The logistic model also fails to predict the expansion, but does predict an intratumoral low cell density area.}
\label{fig:models}
\end{figure}
A common approach \cite{Weis2013} is to dampen the diffusion coefficient in Eq.~\eqref{eq:rd} with a function of local tissue stress:
\begin{equation}\label{eq:vm}
D(\vec{x},t)=D_0e^{-\gamma_v\sigma_{vm}(\vec{x},t)},
\end{equation}
where $D(\vec{x},t)$ is now a spatially and temporally varying diffusion coefficient, $D_0$ is the tumor cell diffusion coefficient in the absence of stress, $\gamma_v$ is an empirical coupling constant, and $\sigma_{vm}(\vec{x},t)$ is the von Mises stress. Here, $\sigma_{vm}(\vec{x},t)$ is used to summarize the local mechanical stress field, which is calculated assuming quasistatic linear elastic equilibrium with tissue-specific mechanical properties:
\begin{equation}\label{eq:lineq}
\nabla\cdot\left(\lambda\left(\nabla\cdot\vec{u}\right)\vec{I} + \mu\left(\nabla\vec{u}+\nabla\vec{u}^T\right)\right) - \gamma_N\nabla N = \vec{0},
\end{equation}
where $\lambda$ and $\mu$ are the Lam\'{e} coefficients (related to the tissue's Young modulus $E$ and Poisson's ratio $\nu$), $\vec{u}$ is the displacement field due to tumor cell growth, and $\gamma_N$ is another empirical coupling constant. In Eq.~\eqref{eq:vm}, the first term on the left-hand side represents the linear elastic tissue response to the local tumor-induced forces described by the second term on the left.
Weis \textsl{et al.} \cite{Weis2013} used tumor cell number estimates from DW-MRI data using Eq.~\eqref{eq:NADC} to initialize and calibrate a mechanically-coupled reaction-diffusion model of breast cancer growth during neoadjuvant chemotherapy consisting of Eqs.~\eqref{eq:rd}--\eqref{eq:lineq}. Their work shows that the mechanically-coupled model rendered more accurate predictions of $N$ (PCC=0.85) than the baseline reaction-diffusion model (PCC=-0.29). Several subsequent studies have also used Eqs.~\eqref{eq:vm}--\eqref{eq:lineq} to couple mechanics to breast and brain tumor dynamics \cite{Jarrett2018a,HormuthII2017}, as shown in Fig.~\ref{fig:models}. Lima \emph{et al.} further considered a mechanical inhibition of tumor proliferation following a similar formulation to Eq.~\eqref{eq:vm} \cite{Lima2016,lima17}. Moreover, the prostate cancer model of Lorenzo \emph{et al.} \cite{Lorenzo2019} extended Eq.~\eqref{eq:vm} to combine the measure of mechanical tissue distortion \emph{via} $\sigma_{vm}$ with hydrostatic stress, which is not captured by $\sigma_{vm}$ and contributes to a more precise description of intratumoral stress.
Other mathematical models couple local mechanics to tumor growth dynamics through a reaction-advection-diffusion equation \cite{Hogea:2008b,Liu2014} in which the tumor cell drift velocity is related to the displacement field, thereby explicitly simulating the displacement of cells due to mechanical deformation.
Additionally, Wong \textsl{et al.} \cite{Wong:2017a} leveraged a hyperelastic biomechanical model. Interestingly, in this study tumor dynamics was described using a reaction-diffusion model in which the proliferation rate $k(\vec{x})$ was calibrated with $^{18}$FDG-PET $SUV$ data, as follows:
\begin{equation}\label{eq:ksuv}
k(\vec{x}) = \frac{\alpha SUV(\vec{x}) - \beta c(\vec{x},t)}{c(\vec{x},t)\left(1-c(\vec{x},t)\right)},
\end{equation}
where $SUV(\vec{x})$ is the standardized uptake value at position $\vec{x}$, $\alpha$ and $\beta$ are unknown constants to be calibrated, and $c(\vec{x},t)$ is the cell volume fraction estimated from computed tomography at position $\vec{x}$ and time $t$.
\subsection{Vasculature-coupled models}\label{sec:vasc}
Co-opting of local vasculature and recruitment of new blood vessels \emph{via} angiogenesis is a critical component of cancer development that is needed to support growth past 2--3 mm$^3$ in size \cite{Gillies1999,Jain2007}.
Thus, understanding the evolving distribution and function of the tumor-supporting vasculature is crucial to accurately model tumor growth and treatment response.
There is an extensive literature on mathematical models of tumor angiogenesis \cite{Vilanova2017}. However, very few describe this phenomenon at the imaging/tissue scale \cite{Hormuth2019,Swanson2011} or personalize it for individual tumors.
Hormuth \textsl{et al.} \cite{Hormuth2019} developed a murine model of glioma growth coupled with angiogenesis that was initialized and calibrated using tumor cell number estimates obtained from DW-MRI \emph{via} Eq.~\eqref{eq:NADC} and estimates of the blood volume fraction extracted from DCE-MRI.
The spatiotemporal evolution of tumor cells and vasculature was described using two coupled reaction-diffusion equations.
In this model, vasculature influenced the direction of tumor growth and was coupled to the carrying capacity. Similarly, tumor cells also influenced the direction of vasculature evolution. The animals were imaged up to seven times over a period of ten days. Model parameters were calibrated using the first three imaging datasets, and then used in a forward evaluation of the model to predict tumor growth at the remaining imaging time points. The authors observed that their model resulted in less than 10.3$\%$ error in tumor volume predictions and less than 9.4$\%$ error at the voxel-level for all prediction time points. Fig.~\ref{fig:models} shows an example of this approach in a pre-clinical model of glioblastoma.
Roque \textsl{et al.} \cite{Roque2018} developed a vasculature-informed preclinical reaction-diffusion model of breast cancer accounting for normoxic, hypoxic, and necrotic cancer subpopulations along with nutrient dynamics, which regulates normoxic cell proliferation as well as the normoxic-hypoxic and hypoxic-necrotic transfer rates. While not explicitly evolving the tumor-supporting vascular network, the authors used vasculature-derived parameters obtained from DCE-MRI (e.g., blood flow, mean transit time, and maximum enhancement) to initialize all model variables and calibrate key parameters. While the study results suggested that further model development is needed to capture individual differences in tumor growth, this work is a unique effort to identify tumor subpopulations using quantitative imaging data.
\subsection{Radiotherapy}\label{sec:rt}
Radiotherapy is a common treatment for many cancers \cite{Miller2016}. However, intratumoral heterogeneity may result in significant variations in treatment response, which may ultimately lead to poor therapeutic outcomes \cite{Gillies1999,Baumann2016}.
Image-based modeling could prove valuable to predict the response to radiotherapy and hence optimize treatment protocols for individual patients.
To this end, several studies have investigated incorporating imaging measures from PET \cite{Rockne2015} and MRI \cite{Hormuth2020,Hormuth2018,Hathout2016,lima17} into reaction-diffusion based models to characterize patient response to radiotherapy. The usual approach to model radiotherapy effects is by instantaneously killing a fraction of tumor cells at treatment times \cite{Rockne2015,Hormuth2020,Hormuth2018,Hathout2016,lima17}.
This strategy may be further combined with a transient or permanent reduction in tumor cell proliferation \cite{Hormuth2020,Hormuth2018,lima17}. These radiotherapy effects are usually modeled as a function of the prescribed dose, which may also account for local tumor cell and vascular densities.
Rockne \textsl{et al.} \cite{Rockne2015} adapted the glioblastoma model by Swanson \textsl{et al.} \cite{Baldock2013} to explicitly incorporate cell death due to radiotherapy based on $^{18}$F-MISO PET data.
Oxygen concentration and the degree of hypoxia in tumors are known to significantly impact response to radiotherapy \cite{Vaupel2007}.
Thus, Rockne \textsl{et al.} used $^{18}$F-MISO PET to assess the level of hypoxia by calculating the $OER$, which is then used along with the usual linear quadratic model of radiotherapy response \cite{Douglas1976} to calculate cell survival, $S$, as follows:
\begin{equation}\label{eq:oer}
S=\exp\left( -\alpha\left(OER(\vec{x})\right)\left( d + \frac{d^2}{\alpha/\beta\left(OER(\vec{x})\right)} \right)\right),
\end{equation}
where $d$ is the prescribed radiation dose while $\alpha\left(OER(\vec{x})\right)$ and $\alpha/\beta\left(OER(\vec{x})\right)$ are radiosensitivity parameters as a function of the $OER$ at position $\vec{x}$.
Rockne \textsl{et al.} observed that predictions by a model featuring Eq.~\eqref{eq:oer} outperformed those obtained with a model with uniform radiosensitivity (1.1$\%$ \emph{vs} 14.6$\%$ error in tumor volume, respectively).
\subsection{Chemotherapy}\label{sec:ct}
Chemotherapy is another common treatment for most cancers \cite{Miller2016}.
Unlike the localized nature of radiation therapy, chemotherapy relies on drugs that are administered systemically throughout the body.
While chemotherapy has traditionally leveraged cytotoxic drugs (i.e., promoting cell death), recent approaches also use drugs targeting specific cancer cell markers to decrease proliferation or triggering particular immune responses.
Similar to radiation therapy, challenges for modeling chemotherapies stem from quantifying how much drug is distributed in the tumor (and healthy tissues) and patient-specific treatment efficacy. Multiple mathematical models have been proposed to describe the effect of chemotherapy on tumor growth \cite{Swan1990,Yin2019,Swierniak2009}, but only a few are informed by quantitative imaging data.
In particular, the contribution by Jarrett \emph{et al.} \cite{Jarrett2018a} extended the mechanically-coupled reaction diffusion model consisting of Eqs.~\eqref{eq:rd}--\eqref{eq:lineq} by including the dynamic effect of chemotherapy in the tumor growth equation:
\begin{align}\label{eq:ct}
\frac{\partial N(\vec{x},t)}{\partial t} = \nabla\cdot\left(D(\vec{x},t) \nabla N(\vec{x},t)\right) + k(\vec{x})N(\vec{x},t)\left(1-\frac{N(\vec{x},t)}{\theta}\right) - \alpha C_d(\vec{x},t) N(\vec{x},t)
\end{align}
where $\alpha$ is the patient-specific drug efficacy and $C_d(\vec{x},t)$ is the drug concentration in the tissue. $C_d(\vec{x},t)$ was approximated patient-wise by means of the two-compartment model commonly used to analyze the contrast agent pharmacokinetics in DCE-MRI data (see Section~\ref{sec:dce}). This approach has two central limitations: it assumes that the drug and the contrast agent have similar dynamics, and that all chemotherapies explicitly induce tumor cell death. However, Jarrett \emph{et al.} showed that their drug-informed model predictions outperformed those of the mechanically-coupled reaction-diffusion model without the drug term when compared to patient-specific estimates of tumor cell density extracted from DW-MRI \emph{via} Eq.~\eqref{eq:NADC} at the end of chemotherapy; in particular, the CCC improved from 0.85 to 0.99 ($p<0.01$).
\section{Computational methods to solve image-based cancer models}\label{sec:comp}
Mechanistic models of cancer usually consist of coupled, nonlinear PDEs. Using the appropriate numerical strategies, these cancer models can be solved and rendered as a computer simulation of the spatiotemporal development of a patient's tumor; i.e., a tumor growth forecast.
In this section we will provide an elementary description of the Finite Difference Method (FDM) \cite{Leveque2007}, Finite Element Analysis (FEA) \cite{Hughes2000} and Isogeometric Analysis (IGA) \cite{Cottrell2009}. All these numerical methods have been widely used to solve PDEs in science and engineering.
\subsection{The Finite Difference Method}\label{sec:fdm}
The FDM relies on a direct approximation of the derivatives involved in the PDEs of the model by means of Taylor series expansions \cite{Leveque2007}.
To apply the FDM, we define a global time interval for the simulation $\left[0,T\right]$ and a geometric domain $\Omega$ consisting of a 3D box that includes the tumor-harboring organ.
Let us discretize $\left[0,T\right]$ with a constant time step $\Delta t$, leading to a partition in time subintervals $[t_n, t_{n+1}]$, such that $t_{n+1}-t_n=\Delta t$, $t_0=0$, $t_{n_t}=T$, and $n=0,\ldots,n_t-1$.
We discretize $\Omega$ with a uniform 3D cartesian grid composed of $n_p=n_xn_yn_z$ nodes numbered $A=1,\ldots,n_p$, where $n_x$, $n_y$ and $n_z$ are the number of nodes in each spatial direction. Let $\vec{g}=(i,j,k)$ further denote the grid coordinates of each node, such that $i=0,\ldots,n_x-1$, $j=0,\ldots,n_y-1$, and $k=0,\ldots,n_z-1$. Then, the spatial coordinates of each node $A$ can be written as $\vec{x}_A=\vec{x}_0+\vec{h}\cdot\vec{g}$, where $\vec{h}=(h_x,h_y,h_z)$ is a vector holding the grid spacing in each spatial direction. Fig.~\ref{fig:fdm} illustrates an FDM grid in 2D.
The standard FDM uses first-order approximations of the time and spatial derivatives in the PDE at the grid nodes and at a certain time instant $\tilde{t}\in[t_n, t_{n+1}]$. For example, the time derivative in Eq.~\eqref{eq:rd} would be approximated by
\begin{equation}
\frac{\partial N(\vec{x},t)}{\partial t} \approx \frac{ N(\vec{x}_A,t_{n+1}) - N(\vec{x}_A,t_{n})}{\Delta t} = \frac{ N^{n+1}_{i,j,k} - N^{n}_{i,j,k}}{\Delta t}
\end{equation}
\noindent on each node at $t=t_{n+1}$. Higher-order derivatives are recursively approximated with first-order approximations of the subsequent lower-order derivatives.
Ultimately, the FDM method reduces the PDE on every node to an algebraic equation involving a combination of values of the PDE solution on the current and adjacent nodes in each spatial direction (see Fig.~\ref{fig:fdm}) at instants $t_n$ and $t_{n+1}$.
Then, the general strategy is to recursively use the known nodal values of the PDE solution at $t_n$, $\{N(\vec{x}_A,t_{n})\}_{A=1,\ldots,n_p}$, to calculate the nodal values at $t_{n+1}$, $\{N(\vec{x}_A,t_{n+1})\}_{A=1,\ldots,n_p}$.
\begin{figure}[t]
\includegraphics[width=\linewidth]{fdm
\caption{The Finite Difference Method (FDM).
(a) In FDM, the model equations are solved on a rectangular grid of physical nodes, which approximates the physical space $\Omega$ representing the problem's geometry. Ancillary fictitious nodes may also be required to construct the FDM grid. Spatial derivatives are approximated with linear combinations of the model solution at times $t_n$ and $t_{n+1}$ on each node and adjacent neighbors in each grid direction.
(b) The FDM can be formulated using imaging voxel data by defining one node per voxel. The organ segmentation can be used to build a map $g(\mathbf{x})$ to define $\Omega$ and hence identify physical ($g(\mathbf{x})=1$) and fictitious nodes ($g(\mathbf{x})=0$).}
\label{fig:fdm}
\end{figure}
Depending on the choice of $\tilde{t}$, there are three common FDM approaches in practice: the explicit Euler method ($\tilde{t}=t_n$), the implicit Euler method ($\tilde{t}=t_{n+1}$), and the Crank-Nicolson method ($\tilde{t}=t_n+\Delta t/2$) \cite{Leveque2007}. The explicit method enables the direct calculation of the PDE solution on the grid nodes at $t_{n+1}$, $\{N(\vec{x}_A,t_{n+1})\}_{A=1,\ldots,n_p}$, from the nodal values of the solution at $t_n$. This approach involves a minimal computational cost, which has been exploited for the recursive model resolutions involved in the patient-specific calibration of image-based models of brain and breast cancer \cite{Jarrett2018a,Hormuth2019}.
For example, the application of the explicit method to Eq.~\eqref{eq:rd} yields
\begin{align}\label{eq:explicit}
\frac{N^{n+1}_{i,j,k} - N^{n}_{i,j,k}}{\Delta t} & = D\left( \frac{ N^{n}_{i-1,j,k} - 2 N^{n}_{i,j,k} + N^{n}_{i+1,j,k}}{h^2_x} + \frac{ N^{n}_{i,j-1,k} - 2 N^{n}_{i,j,k} + N^{n}_{i,j+1,k}}{h^2_y}\right. + \notag \\ & + \left.\frac{ N^{n}_{i,j,k-1} - 2 N^{n}_{i,j,k} + N^{n}_{i,j,k+1}}{h^2_z}\right) + k_{i,j,k}N^{n}_{i,j,k}\left(1-\frac{N^{n}_{i,j,k}}{\theta}\right)
\end{align}
\noindent at every grid node, where we have denoted $k(\vec{x}_A)=k_{i,j,k}$. Note that in Eq.~\eqref{eq:explicit}, we can directly compute $N^{n+1}_{i,j,k}$ from a linear combination of nodal values at $t_n$. However, the explicit method usually requires small time steps to ensure numerical stability. The implicit and Crank-Nicolson methods lead to a system of $n_p\times n_p$ algebraic equations whose resolution provides $\{N(\vec{x}_A,t_{n+1})\}_{A=1,\ldots,n_p}$. These FDM schemes are computationally more intensive, but show better numerical stability and enable the use of larger time steps. Application of these methods to nonlinear PDEs like Eq.~\eqref{eq:rd} results in a nonlinear algebraic system, which can be solved with Newton's method by iteratively solving the corresponding linearized system \cite{Leveque2007}. Alternatively, an implicit-explicit method can use an implicit approach for the diffusion operator and an explicit scheme for the nonlinear logistic term \cite{Leveque2007,Ruuth1995}, which leads to a linear algebraic system. Currently, multiple sparse-matrix algorithms enable a computationally efficient resolution of most linear systems emanating from the application of the implicit, Crank-Nicolson, and implicit-explicit methods \cite{Leveque2007}.
Boundary conditions (BCs) in FDM are applied to the grid nodes lying on the boundary of $\Omega$ (i.e., $\partial\Omega$). The usual approach is to fix the value of the PDE solution on them (Dirichlet BCs) or to approximate a differential boundary condition with an FDM scheme (Neumman and Robin BCs) \cite{Leveque2007}. However, organ borders have complex geometries that rarely coincide with the FDM cartesian grid. This leads to the partition of $\Omega$ into the physical domain, corresponding to the tumor-harboring organ, and a fictitious domain, as shown in Fig.~\ref{fig:fdm}. FDM codes label grid nodes as physical or fictitious and only solve the PDE on the former. Additionally, FDM codes need to identify the physical nodes closer to organ borders to apply BCs.
FDM may be appealing for image-based cancer models because of its simplicity, rapid implementation, and that the cartesian grid can naturally fit the voxel datasets obtained with the imaging technologies described in Section~\ref{sec:img}, as shown in Fig.~\ref{fig:fdm}.
However, the FDM neglects the approximation of the organ geometry and simply relies on placing sufficient grid nodes to capture the organ's border. This impedes an accurate implementation of BCs and may also compromise the resolution of geometry-sensitive problems (e.g., mechanics). FEA and IGA overcome these limitations, also providing superior numerical results that are supported by a strong and rigorously demonstrated mathematical basis \cite{Hughes2000,Cottrell2009}.
\subsection{Finite Element Analysis and Isogeometric Analysis}\label{sec:feaiga}
\subsubsection{General framework}
The central constituents of FEA and IGA are (i) the weak or variational formulation of the strong form of the model, and (ii) a robust approximation of this variational formulation using finite-dimensional function spaces with powerful approximation properties \cite{Hughes2000,Cottrell2009}. To define and illustrate these ideas, let us start by considering the stationary heat equation over a certain physical domain $\Omega$:
\begin{equation}\label{eq:sheat}
\nabla\cdot\left(\kappa\nabla u(\vec{x})\right) + f(\vec{x}) = 0,
\end{equation}
where $\kappa$ is the constant heat conductivity, $u(\vec{x})$ is the spatial map of temperatures over $\Omega$, and $f(\vec{x})$ is a heat source. We further consider homogeneous Dirichlet BCs (i.e., $u(\vec{x})=0$ on $\partial\Omega$), which together with Eq.~\eqref{eq:sheat} constitute the \emph{strong form} of the problem. To derive the \emph{weak form} or variational formulation of this PDE model, we define the \emph{trial function space} $\mathcal{U}$, where the PDE solution resides, and the \emph{weighting function space} $\mathcal{V}$. To this end, we choose $\mathcal{U},\mathcal{V} \subset \mathcal{H}^1$, which is the Sobolev space of square-integrable functions with square-integrable first derivatives. Standard FEA and IGA follow a Bubnov-Galerkin approach. For our heat problem, this translates in $\mathcal{U}=\mathcal{V}$ with functions $u\in\mathcal{U}$ and $w\in\mathcal{V}$ verifying $u(\vec{x})=w(\vec{x})=0$ on $\partial\Omega$. The interested reader is referred to \cite{Hughes2000,Cottrell2009} for a rigorous construction of $\mathcal{U}$ and $\mathcal{V}$. We obtain the weak form of our heat problem as follows: we multiply all terms in Eq.~\eqref{eq:sheat} by an arbitrary $w(\vec{x})\in\mathcal{V}$, integrate the PDE in space over $\Omega$, and integrate the diffusive term by parts using the divergence theorem recalling that $w(\vec{x})=0$ on $\partial\Omega$, which cancels the boundary integral. As a result, the weak form is
\begin{equation}\label{eq:wheat}
\int_\Omega \nabla w(\vec{x}) \cdot\left(\kappa\nabla u(\vec{x})\right) d\vec{x} - \int_\Omega w(\vec{x})f(\vec{x})d\vec{x} = 0,
\end{equation}
\noindent which accounts for both the PDE and the BCs \cite{Hughes2000,Cottrell2009}. Let us define the finite-dimensional subspaces $\mathcal{U}^h\subset\mathcal{U}$ and $\mathcal{V}^h\subset\mathcal{V}$ to approximate the infinite-dimensional spaces $\mathcal{U}$ and $\mathcal{V}$, respectively. We choose a set of basis functions $\{B_A(\vec{x})\}_{A=1,\ldots,n_p}$ spanning $\mathcal{U}^h$ and $\mathcal{V}^h$, where $n_p=\mbox{dim}(\mathcal{U}^h)=\mbox{dim}(\mathcal{V}^h)$. This enables us to \emph{discretize} the weak form in space.
Now, our aim is to find $u^h(\vec{x})=\sum_{B=1}^{n_p}u_B B_B(\vec{x})$ in $\mathcal{U}^h$ satisfying Eq.~\eqref{eq:wheat} for any $w^h(\vec{x})=\sum_{A=1}^{n_p}w_AB_A(\vec{x})$ in $\mathcal{V}^h=\mathcal{U}^h$. In these expressions, the coefficients $u_B$ and $w_A$ are real constants. Using the definition of $w^h$ and recalling that all $w_A$ are arbitrary, we can simplify Eq.~\eqref{eq:wheat} to the Galerkin form
\begin{equation}\label{eq:gheat}
\int_\Omega \nabla B_A(\vec{x}) \cdot\left(\kappa\nabla u^h(\vec{x})\right) d\vec{x} - \int_\Omega B_A(\vec{x})f^h(\vec{x})d\vec{x} = 0,
\end{equation}
for all $A=1,\ldots,n_p$ and where $f^h(\vec{x})=\sum_{A=1}^{n_p}f_A B_A(\vec{x})$. Then, by introducing $u^h(\vec{x})=\sum_{B=1}^{n_p}u_B B_B(\vec{x})$ in Eq.~\eqref{eq:gheat} and rearranging terms, we obtain
\begin{equation}\label{eq:ssheat}
\sum_{B=1}^{n_p} u_B \int_\Omega \nabla B_A(\vec{x}) \cdot\left(\kappa\nabla B_B(\vec{x})\right) d\vec{x} = \int_\Omega B_A(\vec{x})f^h(\vec{x})d\vec{x} ,
\end{equation}
\noindent for all $A=1,\ldots,n_p$. Eq.~\eqref{eq:ssheat} corresponds to a linear algebraic system $\vec{K}\vec{U}=\vec{F}$, where $\vec{F}=\{F_A\}$ and $\vec{K}=\{K_{AB}\}$ are given by
\begin{equation}\label{eq:mheat}
F_A = \int_\Omega B_A(\vec{x})f^h(\vec{x})d\vec{x} \quad\mbox{and} \quad
K_{AB} = \int_\Omega \nabla B_A(\vec{x}) \cdot\left(\kappa\nabla B_B(\vec{x})\right) d\vec{x}.
\end{equation}
\noindent The solution $\vec{U}=\{u_B\}$ provides the coefficients to determine the FEA or IGA approximation $u^h(\vec{x})=\sum_{B=1}^{n_p}u_B B_B(\vec{x})$ to our original model in Eq.\eqref{eq:sheat}. In this process, the construction of the finite spaces $\mathcal{U}^h$ and $\mathcal{V}^h$ along with the basis $\{B_A(\vec{x})\}_{A=1,\ldots,n_p}$ are key steps that ultimately control the convergence and accuracy of the numerical scheme, and that exhibit methodological differences between FEA and IGA.
\begin{figure}[!t]
\includegraphics[width=\linewidth]{femiga
\caption{Finite Element Analysis (FEA) and Isogeometric Analysis (IGA).
(a) In FEA, we approximate the physical domain by repeatedly mapping a common parent element $\hat{\Omega}$ over $\Omega$ using a geometric map $\vec{x}^e$ to generate each element in the mesh. The parent element supports a local function basis $\{b_1,b_2,b_3\}$, which, once mapped to $\Omega$ from each element, contributes to the definition of a global function basis. This is used to approximate the model equations in variational form and is usually integrated using quadrature rules also defined on the parent element (red crosses).
(b) Unstructured FEA meshes can be built to match the segmentation of an organ extracted from medical images (left). Alternatively, immersed-boundary approaches define a FEA mesh matching the voxel grid and a map $g(\mathbf{x})$ to identify the physical domain in which the model will be solved (right).
(c) In IGA, the physical domain is approximated with a topologically equivalent parametric space $\hat{\Omega}$, that is globally mapped onto $\Omega$. The parametric space results from the tensor product of univariate piecewise spline basis $\{b_i\}_{i=1,\ldots,n_i}$ and $\{b_j\}_{j=1,\ldots,n_j}$. The resulting multivariate spline basis is used to approximate the model equations in variational form. These can be integrated using quadrature rules (red crosses) defined over the quadrature space $\tilde{\Omega}$, which is mapped to $\Omega$ \textit{via} composition of the geometric maps $\phi^e$ and $\vec{x}^e$ for each element.
(d) IGA meshes can be built to match the segmentation of an organ extracted form medical images (left). Alternatively, immersed-boundary strategies define an IGA mesh aligning with the voxel grid and a map $g(\mathbf{x})$ to identify the physical domain in which the model will be solved (right).}
\label{fig:femiga}
\end{figure}
Standard FEA uses piecewise Lagrangian polynomial bases to approximate $u^h(\vec{x})$ \cite{Hughes2000}.
The piecewise architecture of FEA bases enables the partition of $\Omega$ in a mesh of $n_e$ subregions termed \emph{elements}, as shown in Fig.~\ref{fig:femiga}.
FEA bases are also \emph{isoparametric} \cite{Hughes2000}, which is a crucial property enabling the use of the same basis functions to describe the geometry $\Omega$ of our problem by means of a function $G(\vec{x})=\sum_{A=1}^{n_p} \vec{x}_A B_A(\vec{x})$, where $\vec{x}_A$ are the physical coordinates of a known set of points over the elements termed \emph{global nodes}.
FEA bases $\{B_A(\vec{x})\}_{A=1,...,n_p}$ are built from a canonical \emph{local basis} defined on a \emph{parent element} $\hat{\Omega}$, which is common for all the elements in the mesh.
The local basis $\{b_a(\bm{\xi})\}_{a=1,...,n_b}$ is composed of $n_b$ functions constructed on the local coordinate system $\bm{\xi}$ of the parent element (see Fig.~\ref{fig:femiga}).
Each local basis function $b_a$ is associated to a unique point in the parent element termed \emph{local node}, with local coordinates $\bm{\xi}_a$ (see Fig.~\ref{fig:femiga}).
For each element $e$ in the mesh, we can build a \emph{geometric mapping} from the parent element given by $\vec{x}^e(\bm{\xi})=\sum_{a=1}^{n_b}\vec{x}^e_a b^e_a(\bm{\xi})$, where $\vec{x}^e_a$ are the physical coordinates of the local nodes of element $e$ in $\Omega$. Hence, we can repeatedly map the parent element and the local basis to each of the elements in the mesh (see Fig.~\ref{fig:femiga}), thereby obtaining the definition of the local basis over each element $e$, i.e., $\{b^e_a(\vec{x}^e)\}_{a=1,\ldots,n_b}$.
In this process, we also build a \emph{connectivity array} of the form $A=c(e,a)$ to identify the global nodes $\vec{x}_A$ and global basis functions $B_A(\vec{x})$ associated to each element's local nodes $\vec{x}^e_a$ and local functions $b^e_a(\vec{x}^e)$ \cite{Hughes2000}. Note that the combination of all element geometric mappings ultimately renders the geometric function $G(\vec{x})$ representing all $\Omega$.
The geometric map $\vec{x}^e$ is \emph{invertible}, such that we can use its inverse to map each element $e$ in the mesh back to the parent element.
This \emph{pull-back} enables us to integrate any basis function over the common parent element and always use the same quadrature rule.
This is a key idea to efficiently calculate vector $\vec{F}$ and matrix $\vec{K}$, using processes called \emph{formation} and \emph{assembly}.
This strategy consists of looping over the elements of the mesh, such that for each element $e$ we (i) use the connectivity array to identify the local nodes and basis functions, (ii) pull them back to the parent element, (iii) calculate the integrals participating in Eq.~\eqref{eq:mheat} using Gaussian quadrature, and (iv) assemble the resulting local values $f_a^e$ and $K_{ab}^e$ by summing them into their corresponding global counterparts $F_A$ and $K_{AB}$ as indicated by the connectivity array. Note that $F_A$ and $K_{AB}$ will receive a contribution from each of the elements sharing node $A$.
In step (ii) we can also pull back spatially varying functions over each element, such as $f(\vec{x})$ or even $u^h(\vec{x})$ (e.g., in nonlinear problems).
Thus, the processes of formation and assembly capitalize on the piecewise definition of FEA bases over the elements to efficiently calculate the integrals in Eq.~\eqref{eq:mheat} specifically wherever the basis functions are defined, instead of performing an inefficient integration over the whole physical domain $\Omega$\cite{Hughes2000}.
IGA is considered a generalization of FEA because it relies on the same core ideas. However, IGA employs more sophisticated polynomial functions coming from computer graphics because its root idea is to use the functions \emph{exactly} describing a computer-generated geometric model of $\Omega$ (e.g., an engineering design, an organ segmentation) to numerically solve the PDE problem posed on such geometry \cite{Cottrell2009}. Conversely, in FEA we first choose the basis to approximate the solution $u^h(\vec{x})$ and then we use it to describe $\Omega$, which usually results in an approximation of $\Omega$ as well. Thus, IGA bases are \emph{geometrically-exact} and isoparametric. IGA bases also show higher global continuity, which ultimately yields superior accuracy \cite{Cottrell2009}.
Non-Uniform Rational B-splines (NURBS) define the most usual function space in IGA \cite{Cottrell2009,Farin1999}.
Univariate NURBS bases are globally defined by a \emph{knot vector}, which is a set of non-decreasing coordinates termed \emph{knots} enabling the definition of all NURBS basis functions over a segment \emph{parametric space} $\hat{\Omega}$ (see Fig.~\ref{fig:femiga}). Multivariate NURBS bases are defined by the tensor product of univariate NURBS bases. Likewise, the tensor product of the corresponding knot vectors results in the definition of the complete IGA mesh formed by $n_e$ elements in a multivariate \emph{parametric space} $\hat{\Omega}$, which is topologically equivalent to the physical space $\Omega$. The isogeometric elements are formed by the knot lines in each parametric direction (see Fig.~\ref{fig:femiga}).
Each of the resulting $n_p$ basis functions $B_A(\bm{\xi})$ defined in $\hat{\Omega}$ is associated to a \emph{control point}, with parametric coordinates $\bm{\xi}_A$ and physical coordinates $\vec{x}_A$.
This results in the definition of an invertible \emph{global geometric mapping} $\vec{x}^g(\bm{\xi})=\sum_{A=1}^{n_p} \vec{x}_A B_A(\bm{\xi})$, bringing the whole IGA mesh from the parametric space $\hat{\Omega}$ into the physical space $\Omega$ and providing an explicit definition of the problem geometry (i.e., $G(\vec{x})$).
Contrary to global nodes in FEA, the control points in IGA do not necessarily align with the mesh and may even be placed out of $\Omega$ \cite{Farin1999,Cottrell2009}.
The restriction of $\vec{x}^g$ to each of the elements also enables the construction of the invertible \emph{element geometric mapping} $\vec{x}^e$, which relies on the identification of the $n_b$ local basis functions and associated local control points defined over the element $e$ by means of a \emph{connectivity array} as in FEA. Additionally, the connectivity array in IGA further accounts for the univariate basis functions that gave rise to the multivariate basis functions \cite{Cottrell2009}.
As the knot vectors are arbitrary, the elements in parametric space may have varying sizes. Thus, we further define a unique \emph{quadrature space} $\tilde{\Omega}$, which is common to all elements, and an ancillary invertible element mapping $\phi^e(\tilde{\bm{\xi}})$ from $\tilde{\Omega}$ to each element in $\hat{\Omega}$. During the process of \emph{formation} and \emph{assembly}, we combine $\vec{x}^e$ and $\phi^e$ to perform the \emph{pull-back} from $\Omega$ to $\tilde{\Omega}$ and calculate the integrals participating in matrices and vectors of the final system (see Eq.~\eqref{eq:mheat}). The rest of the assembly steps in IGA are essentially the same as in FEA (see \cite{Cottrell2009}).
\subsubsection{FEA and IGA for image-based cancer models}
Let us now consider the reaction-diffusion cancer growth model in Eq.~\eqref{eq:rd}. Standard FEA and IGA approaches are \emph{boundary-fitted}, i.e., the physical space $\Omega$ represents the patient's organ (see Fig.~\ref{fig:femiga}). Here, we will derive the weak form of Eq.~\eqref{eq:rd} with the usual no-flux BC $\nabla N\cdot\vec{n}=0$. We choose $\mathcal{U},\mathcal{V} \subset \mathcal{H}^1$ and, using a Bubnov-Galerkin approach, we obtain $\mathcal{U}=\mathcal{V}$ \cite{Hughes2000,Cottrell2009}.
We define $\mathcal{U}^h\subset\mathcal{U}$ and $\mathcal{V}^h\subset\mathcal{V}$ and we choose a basis $\{B_A(\vec{x})\}_{A=1,\ldots,n_p}$ to span $\mathcal{U}^h$ and $\mathcal{V}^h$. Standard FEA and IGA approximate time-dependent functions $f(\vec{x},t)$ as $f^h(\vec{x},t)=\sum_{A=1}^{n_p}f_A(t)B_A(\vec{x})$. Then, the Galerkin discretization of the reaction-diffusion model's weak form is
\begin{equation}\label{eq:galerkin}
\int_{\Omega} B_A\frac{\partial N^h}{\partial t} d\Omega + \int_{\Omega} \nabla B_A \cdot(D\nabla N^h) d\Omega - \int_{\Omega} B_A k^h N^h \left(1-\frac{N^h}{\theta} \right) d\Omega = 0,
\end{equation}
for all $A=1,\ldots,n_p$ and where $k^h(\vec{x})=\sum_{A=1}^{n_p}k_A B_A(\vec{x})$. To discretize and integrate in time, we propose the \emph{generalized-$\alpha$ method} \cite{Chung1993,Jansen2000,Cottrell2009}. This approach relies on a partition of $\left[0,T\right]$ in subintervals $[t_n, t_{n+1}]$ with an arbitrary time step $\Delta t_n$, the discretization of $\partial N^h(\vec{x},t)/\partial t = \sum_{A=1}^{n_p}\dot{N}_A(t)B_A(\vec{x})$, and the definition of vectors $\vec{N}^n=\vec{N}(t_n)=\{ N_A(t_n)\}$ and $\dot{\vec{N}}^n=\dot{\vec{N}}(t_n)=\{ \dot{N}_A(t_n)\}$. The generalized-$\alpha$ method is a predictor-multicorrector algorithm that provides $\vec{N}^{n+1}$ and $\dot{\vec{N}}^{n+1}$ given $\vec{N}^n$ and $\dot{\vec{N}}^n$. This method can be proven to be second-order accurate and $A$-stable by adequately choosing its defining parameters (see \cite{Cottrell2009,Colli2020}). The resulting system of nonlinear algebraic equations can be solved iteratively with Newton's method (see \cite{Cottrell2009,Colli2020}).
Finally, there are multiple approaches to generate \emph{boundary-fitted} meshes in FEA and IGA \cite{Zhang2018}. \emph{Unstructured meshes} are the usual strategy in FEA and consist of populating a geometric model of the patient's tumor-hosting organ generated from its segmentation with tetrahedral or hexahedral elements \cite{Agosti2018,Weis2013,lima17}. In IGA, a common approach is \emph{parametric mapping}, whereby a known geometric model that is topologically equivalent to the organ's geometry is deformed to match the organ's segmentation \cite{Lorenzo2016,Lorenzo2019}. However, image-based cancer models may be more amenable to \emph{immersed-boundary} methods \cite{Mittal2005,Burman2015,Schillinger2015,Parvizian2007,Duester2008}. These strategies rely on constructing the FEA/IGA mesh to align with the voxel grid and defining a function $g(\vec{x})$ accurately representing the organ's boundary. Hence, the function $g(\vec{x})$ enables the definition of the physical space $\Omega$ and a fictitious space over the rest of the mesh (see Fig.~\ref{fig:femiga}). The model PDEs are solved over the whole mesh but their parameters are weighted with $g(\vec{x})$, such that they take their usual value in the physical domain and a negligible value in the fictitious domain. Immersed-boundary methods may also rely on local refinement to improve the discretization of $g(\vec{x})$ \cite{Schillinger2015}.
\section{Calibrating image-based mathematical oncology models}\label{sec:invprob}
Mathematical models of cancer define a \emph{forward problem}, whose solution provides \emph{state variables} (e.g., tumor cell density). In general, these models are parameterized by \emph{unknown} biophysical parameters (and possibly initial conditions) that typically manifest substantial variability across subjects~\cite{Swanson:2020a,Lipkova:2019a,Scheufele:2020}.
The estimation of these unknown variables (also called \emph{inversion variables}) should be patient-specific and can be mathematically posed as an \emph{inverse problem}, which aims at optimizing an objective function constrained by the model.
Since image-based cancer models are usually represented by PDEs, the resulting inverse problem is formally a PDE-constrained optimization problem.
In this section, we outline the general formulation of the inverse parameter estimation problem and discuss the standard methods to compute its solution, keeping in mind the ultimate goal of patient-specific tumor characterization and model prediction. Fig.~\ref{fig:model-calibration} illustrates the typical image-based inverse problem workflow in the context of brain tumors.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{radiomics-overview.pdf}\\ (a) Imaging workflow in brain tumor radiomics \vspace{2em}\\
\includegraphics[width=\textwidth]{model-calibration.pdf}\\ (b) Inverse problem \\
\caption{Workflow for an image-driven biophysical inverse problem. (a) The goal is to extract clinically relevant biomarkers from input multi-parametric MRI data that can predict and guide intervention so as to improve clinical outcome. Purely imaging-based approaches (feature extraction) can be integrated with biophysical priors (biophysical model inversion) to develop tools that can assist in treatment and prognosis. (b) The input data for patients with brain tumors (multi-parametric MRI scans) is at a single time point (treatment typically follows immediately after diagnosis) and is translated into model observables for tumor concentration $c_\text{obs}$.
The inverse problem seeks to estimate unknown biophysical parameters $\vec{p}$ of a given mathematical model. The basic idea is to perform a number of model simulations with different parameters so that the model-predicted tumor matches the observed data. The panels in this figure are adapted from~\cite{Mang:2020}.}\label{fig:model-calibration}
\end{figure}
\subsection{Inverse problems for oncology models}
Let $\mathcal{F}(\vec{p},c) = 0$ denote a cancer model consisting of a PDE system, where $\vec{p}$ is a vector of unknown parameters and $c(\vec{x},t)$ represents the state variables.
The inverse problem seeks to estimate $\vec{p}$ such that the model state variables $c(\vec{x},t)$ match given observational patient data (see Fig.~\ref{fig:model-calibration}).
In oncology, the input data is usually a series of medical images (e.g., MRI, PET) at specific time instances $\{t_j\}_{j = 1, \cdots, n_t}$ within a finite time horizon $(0,T]$.
These define our observations $\hat{c}_j(\vec{x})$ of the state variables in the forward model $c(\vec{x},t)$, e.g., tumor cell density (see Section~\ref{sec:dwi}).
Then, the inverse problem can be mathematically formulated as
\begin{equation} \label{eq:pdeopt}
\min_{\vec{p}}
\mathcal{J}(\vec{p},c) = \frac{1}{2} \sum_{j=1}^{n_t} \int_{\Omega}\!\! \big(c(\vec{x},t_j) - \hat{c}_j(\vec{x})\big)^2 d \vec{x} + \mathcal{R}(\vec{p}) \;\; \text{s.t.} \;\; \mathcal{F}(\vec{p},c) = 0,
\end{equation}
where $\Omega$ is the spatial domain representing the patient's tumor-harboring organ. The objective function $\mathcal{J}(\vec{p},c)$ minimizes the mismatch between the predicted tumor and observed data at times $\{t_j\}_{j = 1, \cdots, n_t}$ using an $L^2$ distance measure. Additionally, it balances this data fidelity/mismatch term with a regularization operator $\mathcal{R}(\vec{p})$.
Designing algorithms for the efficient and effective numerical solution of such PDE-constrained optimization problems is a challenging task~\cite{Leugering:2014a,Akcelik:2006a}. While derivative-free optimization strategies are popular due to ease of implementation, they are typically associated with slow convergence and can become prohibitively expensive (especially if the dimensionality of $\vec{p}$ is large). Hence, optimization algorithms that utilize gradient information are preferable. In addition to improving convergence and computation time, these methods can reveal important characteristics of the objective function landscape, which can be exploited to design better algorithms and help us understand the sensitivities and ill-posedness inherent to PDE-constrained optimization problems.
There exist multiple options for evaluating the gradient (and higher order derivatives) of the objective function, such as automatic differentiation, numerical approximation through finite differences, and adjoint-based methods.
For oncology models, several groups have employed adjoints for inversion~\cite{Hogea:2008b,Gholami:2016a,Subramanian20IP,Colin:2014a,Feng:2018a,Knopoff:2013a}. Some efforts also employ Hessian information to accelerate convergence~\cite{Gholami:2016a,Scheufele:2018a}. Other strategies such as derivative-free optimization~\cite{Chen:2012b,Wong:2017a,Mi:2014a,Konukoglu:2010b} or finite difference approximations~\cite{Hormuth:2015a} have also been considered in literature, but are usually less effective than adjoint-based methods. For large-scale 3D inversion, parallel (distributed memory) algorithms have been considered in~\cite{Gholami:2016a,Scheufele:2018a,Subramanian20IP}.
While the deterministic approaches just described are successful in estimating the optimal parameters to the minimization problem in Eq.~\eqref{eq:pdeopt}, their utility could be limited due to uncertainties arising from modeling errors and noise in measurements/data. A probabilistic (Bayesian) formulation can mitigate this drawback by characterizing our confidence in the inversion variables $\vec{p}$ using probability density functions. This approach will be described in Section~\ref{sec:probcal}, as part of a comprehensive Bayesian framework for model selection. In the following section, we briefly describe adjoint-based inversion methods.
\subsection{Adjoint methods for inverse problems}
The standard technique for solving the inverse problem posed in Eq.~\eqref{eq:pdeopt} is to introduce Lagrange multipliers $\lambda(\vec{x},t)$, which are termed \emph{adjoint variables} or simply \emph{adjoints}, and construct the Lagrangian functional as
\begin{equation}\label{eq:lagrangian}
\mathcal{L}(c, \vec{p}, \lambda) = \mathcal{J}(\vec{p},c) + \langle \mathcal{F}(\vec{p},c), \lambda \rangle,
\end{equation}
where $\langle . \rangle$ denotes an appropriate inner product (typically $L^2$) and $\mathcal{J}(\vec{p},c)$ denotes the objective function defined in Eq.~\eqref{eq:pdeopt}. By requiring stationarity with respect to the state, adjoint, and inversion variables, we arrive at the \emph{first order optimality conditions} by taking the following variations:
\begin{subequations}\label{eq:opt}
\begin{align}
\delta_\lambda \mathcal{L} &= 0 \quad \quad\quad \text{\textit{(forward equations)}},\label{eq:forward}\\
\delta_c \mathcal{L} &= 0 \quad \quad\quad \text{\textit{(adjoint equations)}},\label{eq:adjoint}\\
\delta_{\vec{p}} \mathcal{L} &= 0 \quad \quad\quad \text{\textit{(inversion equations)}},\label{eq:inversion}
\end{align}
\end{subequations}
where $\delta_z \mathcal{L}$ denotes the variation of $ \mathcal{L}$ with respect to $z$. The forward equations are simply the tumor growth PDE model. The adjoint equations are linear PDEs in the adjoint variables backward in time. The inversion equations denote the PDE-constrained gradient of $\mathcal{J}(\vec{p},c)$, which is set to zero at the local minimum.
For example, consider the reaction-diffusion cancer model in Eq.~\eqref{eq:rd} with the usual no-flux BC $\nabla N\cdot\vec{n}=0$ and two tumor cell density observations $\hat{N}_0(\vec{x})$ and $\hat{N}_1(\vec{x})$ derived from DW-MRI at $t=t_0=0$ and $t=t_1$ (see Section~\ref{sec:dwi}). We assume that $\hat{N}_0(\vec{x})$ are known initial conditions for Eq.~\eqref{eq:rd}, and use $\hat{N}_1(\vec{x})$ to calibrate $\vec{p}=\{D,k(\vec{x})\}$, where $D$ is a scalar constant and $k(\vec{x})$ is a spatial field, representing the tumor cell diffusion and proliferation rate, respectively (see Section~\ref{sec:dwi}). Then, we define the objective function $\mathcal{J}(\vec{p},N)$ and the Lagrangian functional $\mathcal{L}(N, \vec{p}, \lambda)$ as
\begin{align}\label{eq:glL}
\mathcal{J}(\vec{p},N) &= \frac{1}{2} \int_{\Omega}\!\! \big(N(\vec{x},t_1) - \hat{N}_1(\vec{x})\big)^2 d \vec{x} + \frac{a}{2}\left(D^2 + \int_\Omega\!\! k^2(\vec{x})d\vec{x}\right), \\
\mathcal{L}(N, \vec{p}, \lambda) &= \mathcal{J}(\vec{p},N) + \int_0^{t_1}\int_\Omega\!\!\lambda(\vec{x},t)\left( \frac{\partial N(\vec{x},t)}{\partial t} - \nabla\cdot\left(D \nabla N(\vec{x},t)\right) \right. \notag \\ & \left. - k(\vec{x})N(\vec{x},t)\left(1-\frac{N(\vec{x},t)}{\theta}\right) \right)d\vec{x}dt,
\end{align}
where $a$ is a regularization parameter. Following Eq.~\eqref{eq:adjoint}, the adjoint equation is
\begin{equation}
-\frac{\partial\lambda(\vec{x},t)}{\partial t} = \nabla\cdot\left(D \nabla \lambda(\vec{x},t)\right) + k(\vec{x})\lambda(\vec{x},t)\left(1-\frac{2N(\vec{x},t)}{\theta}\right),
\end{equation}
which is a linear backward problem in time subject to the BC $\nabla \lambda\cdot\vec{n}=0$ on $\partial\Omega$ and the terminal condition $\lambda(\vec{x},t_1)=\hat{N}_1(\vec{x})-N(\vec{x},t_1) $. Following Eq.~\eqref{eq:inversion}, the inversion equations are given by
\begin{align}
aD + \int_0^{t_1}\int_\Omega\!\! \nabla\lambda(\vec{x},t) \cdot \nabla N(\vec{x},t) d\vec{x}dt &= 0, \\
ak(\vec{x}) - \int_0^{t_1}\!\! \lambda(\vec{x},t) N(\vec{x},t)\left(1-\frac{N(\vec{x},t)}{\theta}\right)dt &= 0.
\end{align}
In general, Eq.~\eqref{eq:opt} represents a large, non-linear, coupled system of PDEs, which can be significantly challenging to solve simultaneously. Instead, a standard approach involves a \emph{reduced space} algorithm, which is an iterative strategy that reduces Eq.~\eqref{eq:opt} to a system involving only the inversion variables. In each iteration, we use the current approximation to $\vec{p}$ to solve the state and adjoint equations to respectively get the current approximation to the state and adjoint variables. Then, we update our inversion variables using the inversion equations. This process is repeated until convergence, which is typically set by a user-defined threshold on the parameter update (i.e., the parameter gradient). In contrast to \emph{full space} methods, reduced space methods present more tractable systems that can exploit existing PDE solvers for the state and adjoint equations and are better conditioned. We refer the reader to~\cite{Akcelik:2006a} for more details on adjoint methods in PDE-constrained optimization.
\section{Model selection and identification of relevant parameters}
Given the vast array of cancer growth models in the literature, it is not trivial to choose which is the best to represent the available data and to predict key quantities of interest (e.g., the tumor size, treatment efficacy, or percentage of necrosis) for a certain tumor type.
In this section, we describe the Occam Plausibility Algorithm (OPAL), which has been proposed in \cite{Farrell2015} as an adaptive process for model selection and validation in the presence of uncertainties. The strategy relies on three key steps: sensitivity analysis to identify the relevant model parameters, model calibration of the relevant parameters, and calculation of model selection criteria. The OPAL can be referred to as \emph{model agnostic} in that no single model is advocated; rather, the \emph{best} model is selected based upon our model selection criteria.
Details regarding each step of the OPAL are given in the next subsections.
\subsection{Variance-based sensitivity analysis}\label{sec:sobol}
Sensitivity analysis quantifies how changes in parameter values affect the uncertainty in model output \cite{saltelli2019so}. We can distinguish between \emph{local} and \emph{global} methods. Local methods compute the variation of the model output changing one parameter at a time (i.e., first-order effects) usually \emph{via} derivation, but neglect the interactions between the parameters.
In global methods, the contribution of each parameter along with its interactions with other parameters (i.e., higher-order effects) are taken into account, as all parameters are varied simultaneously over the entire parameter space. In this work, we present the variance-based global sensitivity analysis method, also known as the Sobol method \cite{sobol2001global,saltelli2008global}.
Details regarding local and other global methods can be found in \cite{saltelli2008global}.
Let $\vec{M}(\bm{\theta})$ be a model parameterized by $k$ parameters $\bm{\theta}$, which belong to a parameter space $\bm{\Theta}\subset\mathbb{R}^k$. The computational cost of the sensitivity analysis of model $\vec{M}(\bm{\theta})$ depends on the number of parameters $k$ and the sample size $N$, with the total number of model evaluations given by $N_T=N(k+1)$. There are several approaches to estimate the \emph{total sensitivity index} for each parameter, which quantifies all effects of the parameter on the model output. Here, we present the strategy in \cite{saltelli2010variance}, as it is known to demand a reduced sample size to converge.
First, we randomly generate two sampling matrices, $\vec{A}$ and $\vec{B}$, with size $N\times k$. Each row of these matrices represents a sampled value for the vector of parameter $\bm{\theta}$. Additionally, we create $k$ matrices $\vec{A}_{\vec{B}}^{(k)}$, where we copy the values from the matrix $\vec{A}$ and replace the values from column $k$ with the values from $\vec{B}$. For the case where $N=1$, these matrices are given as: $\vec{A}=[ \theta^a_{1,1} \quad \theta^a_{1,2} \; \ldots \; \theta^a_{1,k}]$, $\vec{B}=[ \theta^b_{1,1} \quad \theta^b_{1,2} \; \ldots \; \theta^b_{1,k}]$, $\vec{A}_{\vec{B}}^{(1)}=[ \theta^b_{1,1} \quad \theta^a_{1,2} \; \ldots \; \theta^a_{1,k}]$, $\vec{A}_{\vec{B}}^{(2)}=[ \theta^a_{1,1} \quad \theta^b_{1,2} \; \ldots \; \theta^a_{1,k}]$, $\ldots$, and $\vec{A}_{\vec{B}}^{(k)}=[ \theta^a_{1,1} \quad \theta^a_{1,2} \; \ldots \; \theta^b_{1,k}]$.
Then, we run the forward model for each row in matrix $\vec{A}$ and all matrices $\vec{A}_{\vec{B}}^{(k)}$.
The outputs of the model are stored in corresponding solution vectors for each matrix; i.e., $\vec{Y}_A$, $\vec{Y}_{AB}^1$, \ldots, $\vec{Y}_{AB}^k$. Finally, we compute the total sensitivity index $S_{T_i}$ for each parameter, which can be approximated \cite{jansen1999analysis,saltelli2008global,saltelli2010variance} by
\begin{align}
S_{T_i}=\frac{1}{2N\mathrm{Var}(\vec{Y}_{A})}\sum_{j=1}^{N}\left(\left(\vec{Y}_{A}\right)_j-\left(\vec{Y}_{AB}^{(i)}\right)_j\right)^2,
\label{eq_tei}
\end{align}
where $\mathrm{Var}(\vec{Y}_{A})$ is the variance of vector $\vec{Y}_{A}$. According to \cite{saltelli2008global}, a parameter $i$ can be considered non-influential if $S_{T_i}=0$. In practice, we define a threshold $\epsilon$ and we identify a parameter $i$ as non-influential if $S_{T_i}<\epsilon$. The choice of $\epsilon$ is relative to other $S_{T_j}$ and problem-dependent.
If $S_{T_i}$ is sufficiently small, then the parameter does not affect the quantities of interest, and the complexity of the model can be reduced by removing or fixing the parameter to any value within the uncertainty range \cite{saltelli2008global}.
\subsection{Model calibration}\label{sec:probcal}
To characterize uncertainties in the observable data and the stochastic behavior of tumor growth, we follow the Bayesian statistical calibration procedure. This method captures these uncertainties by delivering a probabilistic distribution of the model parameters, instead of a single value for each of them \cite{oden2016,lima17,Lima2016,oden2017df}. The basic ideas behind the Bayesian parameter estimation involve the following steps:
\begin{enumerate}
\item Select the observational data $\vec{D}$ to be used (e.g., baseline and follow-up MRI).
\item Establish the prior distribution of the model parameters $\pi_{prior}(\bm{\theta})$. In the cases where we do not have knowledge regarding the distribution of the parameters, and we can only estimate the range of these parameters, the usual approach is to assume a uniform prior distribution.
\item Construct the likelihood function, which, given the values assigned to the parameters $\bm{\theta}$, yields the probability of $\vec{D}$ being observed \cite{oden2017df}. Assuming both the experimental error and the model inadequacy to be Gaussian, and the experimental data to be independent and identically distributed, the likelihood is given as
\begin{align}
\pi_{like}(\vec{D}|\bm{\theta})=\prod_{i=1}^{N_t}\frac{1}{\sqrt{2\pi \sigma^2}}\exp\left(-\frac{(D_i-Y_i(\bm{\theta}))^2}{2\sigma^2}\right),
\label{eq:like}
\end{align}
where $N_t$ is the number of data points, $\sigma$ is the standard deviation of the experimental error and model inadequacy, and $Y(\bm{\theta})$ is the model output.
\item Compute the posterior distribution of the parameters $\pi_{post}(\bm{\theta}|\vec{D})$ as
\begin{align}
\pi_{post}(\bm{\theta}|\vec{D})=\frac{\pi_{like}(\vec{D}|\bm{\theta})\pi_{prior}(\bm{\theta})}{\pi_{evid}(\vec{D})},
\label{eq:post}
\end{align}
where $\pi_{evid}(\vec{D})=\int_\Theta\left(\pi_{like}(\vec{D}|\bm{\theta})\pi_{prior}(\bm{\theta})\right)\,d\bm{\theta}$ is the model evidence. The resulting posterior distribution of the parameters allows the prediction of the quantities of interest taking into account the uncertainties in the parameters.
\end{enumerate}
The posterior probability density function is, in general, non-Gaussian. Sampling schemes such as Markov Chain Monte Carlo (MCMC) methods can be used to evaluate posterior expectations \cite{lima17,Lima2016}. These stochastic methods can incur in a large computational cost, so efficient sampling strategies exploiting the problem structure are actively being investigated \cite{prudencio12,Drzisga2017,Wang2018,Petra2014}.
\subsection{Model selection criteria}
Following the Bayesian framework used for calibration, we approach model selection by computing the \emph{model plausibility} \cite{jeffreys98,chow81,beck10,prudencio12,oden2017df}.
Given a set of $m$ models $\vec{M}=\left\{M_i(\bm{\theta}_i)\right\}_{i=1}^m$, the Bayes' rule in Eq. (\ref{eq:post}) can be rewritten assuming that probabilities are conditional on the model $M_i$ and the set ${M}$:
\begin{align}
\pi_{post}(\bm{\theta}_i|\vec{D},M_i,\vec{M})=\frac{\pi_{like}(\vec{D}|\bm{\theta}_i,M_i,\vec{M})\pi_{prior}(\bm{\theta}_i|M_i,\vec{M})}{\pi_{evid}(\vec{D}|M_i,\vec{M})}.
\label{eq:pospm}
\end{align}
The evidence of each model can be viewed as a likelihood for a discrete Bayesian calculation, yielding a new posterior called the model plausibility, $\pi_{plaus}$, given as
\begin{align}
\pi_{plaus}(M_i|\vec{M},\vec{D})=\frac{\pi_{evid}(\vec{D}|,M_i,\vec{M})\pi_{prior}(M_i|\vec{M})}{\pi_{evid}(\vec{D}|\vec{M})},\qquad 1\leq i \leq m.
\label{eg:plaupm}
\end{align}
If we assume that all models are equally probable, $\pi_{prior}(M_i|\vec{M})=\frac{1}{m}$. The sum of all model plausibilities is equal to one (i.e., $\sum_{i=1}^m\pi_{plaus}(M_i|\vec{M},\vec{D})=1$). The model with highest plausibility is selected as the best model in $\vec{M}$ to capture the data.
Another popular method of model selection is the Akaike Information Criterion (AIC) \cite{konishi08}. In this method, the likelihood of the maximum likelihood estimator, $\hat{\bm{\theta}}$, is penalized according to the number of model parameters, i.e.,
\begin{align}
AIC_i=-2\log \pi_{like}(\vec{D}|\hat{\bm{\theta}})+2k_i, \qquad 1\leq i \leq m,
\end{align}
where $k_i$ the number of parameters in model $M_i$. In this case, the model with the lowest AIC is the best model in $\vec{M}$.
\subsection{The Occam Plausibility Algorithm }
\begin{figure}[t]
\centering
\includegraphics[width=1.0\hsize]{opal}
\caption{The Occam Plausibility Algorithm (OPAL). This framework brings together experimental data for model calibration and validation, sensitivity analysis, selection, and model prediction.}
\label{fig:opalf}
\end{figure}
OPAL has been proposed as a comprehensive framework that combines parameter sensitivity analysis, and model calibration, validation, selection, and prediction \cite{Farrell2015,oden2017df}. Fig. \ref{fig:opalf} shows a schematic representation of the main OPAL steps, which we outline in the following \cite{Farrell2015,oden2016}:
\begin{enumerate}
\item Identify a set $\vec{M}=\{M_1(\bm{\theta}_1),M_2(\bm{\theta}_2),\ldots,M_m(\bm{\theta}_m)\}$ of parametric models, each with parameters $\bm{\theta}_i$ belonging to an appropriate parameter space $\bm{\Theta}_i$, $1\leq i \leq m$.
\item Perform a sensitivity analysis to identify non-influential parameters. Based on this analysis, the values of these parameters are fixed to the mean value used in the sensitivity analysis. If there is a model in $\vec{M}$, whose only difference to other model is given by the non-influential parameters, this model can be eliminated, yielding a reduced set $\bar{\vec{M}}=\{\bar{M}_1(\bm{\theta}_1),\bar{M}_2(\bm{\theta}_2),\ldots,\bar{M}_l(\bm{\theta}_l)\}$ of models, with $l\leq m$.
\item Divide the models in $\bar{\vec{M}}$ into ``Occam Categories'' according to their complexity (e.g., number of parameters). These categories are sorted in ascending order according to their complexity.
\item Calibrate the models in Category 1 using the calibration data $\vec{D}^c$.
\item Select the best model to represent the data in this category (e.g., the model with the highest plausibility).
\item Test the best model identified in the current category in a validation scenario, where the posterior from the calibration step is used as a prior and the distribution of the parameters are updated against the validation data $\vec{D}^v$. If the model is able to represent the data within a preset tolerance, the model is considered ``valid''. If not, we return to step 3 and move to the next Occam category. If we are not able to find a valid model, we need to return to step 1 and include new models.
\item After finding the ``simplest'' valid model, solve the forward model in the prediction scenario and compute the quantities of interest.
\end{enumerate}
All of these steps are designed to consider uncertainties in the choice of model, the model parameters, the observational data, and the target quantities of interest. All uncertainties are generally characterized by probability densities.
\section{Towards the optimization of personalized treatment plans}\label{sec:optcon}
Several image-based mathematical models of cancer growth have shown promise in predicting treatment outcomes in a patient-specific manner, as discussed in Sections~\ref{sec:rt} and \ref{sec:ct}. Those models could provide a means to determine optimal therapeutic regimens to treat a certain type of cancer \emph{in silico}, which could then be investigated within a clinical trial \emph{in vivo}. Hence, this computational approach seeks to help clinicians navigate the vast array of radiotherapy and drug combinations, dosing options, and treatment schedules and select optimal strategies, which are virtually impossible to assess in clinical trials. Ultimately, cancer models could also serve as a \emph{digital twin} for the patient's tumor, thereby enabling the pathological assessment, monitoring, and design of optimal therapeutic regimens for the individual patient \emph{in silico}.
In this section, we discuss the use of image-based predictive tumor growth models accounting for the therapeutic regimen and associated tumor response for the discovery of optimal therapeutic regimens and the design of patient-specific optimal treatment strategies.
\subsection{Potential to select treatment plans for individual patients}\label{sec:8.2}
Selecting a treatment regimen for a patient is a complex process. Oncologists use decision tree algorithms to select therapeutics for each patient considering, for example, tumor grade and cell markers \cite{Bevers2009}. However, the determination of the optimal dosing regimens for these therapies is vastly underinvestigated. This limitation follows from the impossibility to test all the potential dosing strategies within a clinical trial. Additionally, regimens may be altered by the treating oncologist due to considerations like side effects and quality of life for the patient, where doses may be skipped, dosages decreased, and/or supportive medications prescribed. However, these changes are made only with a limited knowledge of their effects on the treatment outcome for any given patient.
As patients present with varying physiologies and sensitivities, the one-size-fits all approach is clearly not optimal for all patients.
Mathematical models of tumor growth and treatment response can help us predict therapeutic efficacy accounting for each patient's specific tumor dynamics and, potentially, select the best therapeutic regimen for each individual patient.
For instance, Jarrett \emph{et al.} \cite{Jarrett2020} used the model in Eq.~\eqref{eq:ct} to investigate alternative dosing regimens of cytotoxic therapies for breast cancer \emph{in silico}.
First, the model was parameterized patient-wise using MRI data collected prior to the start of therapy and after one drug cycle as indicated in Section~\ref{sec:ct}. The resulting personalized model was then simulated to the time of completion of the prescribed therapeutic regimen, and the simulated tumor growth was compared to the actual tumor response measured by MRI for each patient. The model predictions were found to be highly correlated with actual tumor response ($N = 13$, CCC$ > 0.90$, $p < 0.01$ for total cellularity, total volume, and longest axis), so the model was considered valid to reproduce the effects of chemotherapy in breast cancer.
Their validated model was then used to explore alternative therapeutic regimens, which were defined patient-wise by fixing the total dose prescribed in standard regimen and varying frequency and dosage.
The authors indicated that an additional 0-46\% reduction (median=17\%) in total cellularity may have been achievable across the patient cohort ($N = 13$) compared to the standard chemotherapeutic regimens that the patients were prescribed. The dosing regimens that the model predicted to reduce/control each tumor were also found to significantly outperform standard regimens for tumor control ($p < 0.001$), thereby supporting the claim that standard regimens may not be the most effective for every patient.
\subsection{Optimal control theory for personalized treatment planning} \label{sec:8.3}
Consider a dynamical system involving a set of variables $u(t)$ and controls $z(t)$, which are functions describing external forces that can alter the system dynamics.
\emph{Optimal control theory} (OCT) was developed to determine the solution of the system that achieves a particular outcome by adequately adjusting the controls.
The mathematical formulation of the optimal control problem consists of minimizing or maximizing an objective functional $J(u,z)$. Thus, given a particular dynamical system over a certain time interval $[0,T]$, applying OCT largely consists of determining the objective functional, problem-specific constraints, and a method for solving the OCT problem.
The general form of the objective functional for OCT is
\begin{equation}
J(u,z) = \Phi[u(t_j),z(t_j),t_j]+\int_{t_0}^{t_f}L(u(t),z(t),t)dt,
\label{eqn:ObjFunc}
\end{equation}
where $\Phi[u(t_j),z(t_j),t_j]$ includes target values of the variables and the controls at specific times $\{t_k\}_{k=1,\ldots,n_t}$, while $L(u(t),z(t),t)$ accounts for the target dynamics of the variables and the controls over $[0,T]$. $\Phi[u(t_j),$ $z(t_j),t_j]$ has several names in OCT literature, including \emph{endpoint cost} in minimization problems and \emph{terminal payoff} in maximization problems.
The formulation of the objective function can also be divided into three canonical types: \emph{endpoint control}, which only includes $\Phi[u(t_j),z(t_j),t_j]$; \emph{bang-bang control}, which only features a linear $L(u(t),z(t),t)$; and \emph{continuous control}, which only has a quadratic $L(u(t),z(t),t)$. The objective function can also be constructed by combining these canonical types.
The term \emph{bang-bang} refers to the usual dynamics of optimal control $z(t)$ for this type of functional, which switches between the maximum admissible value and $z(t)=0$ (i.e., no effect).
Additionally, the quadratic term in the continuous control is not usually motivated by problem-dependent phenomena, but it ensures that the optimization problem is \emph{convex}. Hence, the optimal control problem has key mathematical features, including the existence of a global minimum.
In the context of cancer, we can apply OCT to obtain optimal treatment strategies by using a mathematical model to simulate tumor growth and therapeutic response as a dynamical system, setting the treatment as a control, and selecting clinically-relevant treatment outcomes in the objective functional.
Here we will briefly discuss the formulation of the OCT problem using a simplification of the PDE model of breast cancer chemotherapy in Eq.~\eqref{eq:ct}, whereby we directly model the total number of tumor cells $n(t)$ using the ordinary differential equation (ODE) \cite{Yin2019,Johnson2020}:
\begin{equation}\label{ode_nat}
\frac{dn(t)}{dt}=kn(t)\left(1-\frac{n(t)}{\theta}\right)-\alpha z(t)n(t)
\end{equation}
where $k$ describes global tumor cell proliferation, $\theta$ is the tissue carrying capacity, $\alpha$ models chemotherapy efficacy, and $z(t)$ represents the dynamics of the concentration of drug(s) in the plasma, which is used to derive the initial concentration of drug in the tissue $C_{drug}(\vec{x},t)$ in Eq.~\eqref{eq:ct} \cite{Jarrett2018a}. We chose this ODE model for the sake of simplicity and because ODE models have been commonly used in the cancer OCT literature \cite{Jarrett2020a, Lenhart2007,Schaettler2015, Anita2011, Swan1990}. However, the following ideas could also be applied to the PDE model in Eq.~\eqref{eq:ct} by defining $n(t)=\int_{\Omega}N(\vec{x},t)d\vec{x}$ (see \cite{Almeida2019}).
A primary goal of OCT problems for cancer treatment is to minimize the tumor burden only at the completion of the therapy, which can be formulated by the endpoint control functional
\begin{equation}
J_1(n)=n(t_f).
\label{eq:endpointcost}
\end{equation}
This optimal control problem requires additional constraints for the therapeutic regimen $z(t)$. For example, these can limit the maximum dose by setting $z(t)<z_{max}$ and/or the total maximum dose by imposing $\int_{0}^{T}z(t)dt < z_{tot}$, which is termed an \emph{isoperimetric constraint} (see Fig.~\ref{fig:optcon}).
The limits $z_{max}$ and $z_{tot}$ are drug-dependent and may be patient-specific (e.g., quality of life, comorbidities). However, the optimization of $J_1(n)$ is insensitive to the dynamics of the drug concentration; i.e., it does not formally adapt the drug regimen to the tumor burden. Another limitation of endpoint control formulations is that they do not consider the potential growth of the tumor during therapy, which may be relevant to select actionable therapeutic regimens depending on the type of cancer.
To address this limitation, we can extend the objective functional in Eq.~\eqref{eq:endpointcost} with a bang-bang term accounting for $z(t)$:
\begin{equation}
J_2(u,z)=w_1 n(t_f)+w_2\int_{0}^{T}z(t)dt.
\label{eq:bangbang}
\end{equation}
where $w_1$ and $w_2$ are problem specific weights that can be included in the objective function to give greater or lesser importance to the different terms in the objective functional during optimization. For example, if a particular treatment has significant adverse side effects, $w_2 > w_1$ may be enforced to focus the problem on minimizing the total drug dose. Similarly, an objective function that includes the continuous control of the therapy can be written as
\begin{equation}
J_3(u,z)=w_1 n(t_f)+\frac{w_2}{2}\int_{0}^{T}(z(t))^2dt.
\label{eq:continuousdrug}
\end{equation}
Furthermore, the objective function can also account for the tumor growth over the entire treatment period:
\begin{equation}
J_4(u,z)=\frac{1}{2}\int_{0}^{T}w_1(n(t))^2+w_2(z(t))^2dt.
\label{eq:continuousboth}
\end{equation}
\begin{figure}[!t]
\includegraphics[width=\textwidth]{figure_oct}
\caption{Optimal control problem solutions for the ODE model of cancer chemotherapy in Eq.~\eqref{ode_nat}.
(a) Without an isoperimetric constraint, the optimal control problem using $J_1$ yields $z(t)=z_{max}$, which would probably induce acute toxicity in patients. The optimal $z(t)$ for the bang-bang control using $J_2$ produces the characteristic stepwise solution, while the optimal $z(t)$ obtained with the continuous controls in $J_3$ and $J_4$ show a continuous decay. The drug regimens calculated with $J_1$, $J_3$ and $J_4$ successfully control tumor growth, while for $J_2$ we observe regrowth once $z(t)$ drops to zero.
(b) If we impose the isoperimetric constraint, the optimal $z(t)$ for $J_1$ and $J_2$ coincide and produce a similar regimen to the bang-bang solution without isoperimetric constraint. The optimal $z(t)$ obtained with the continuous controls in $J_3$ and $J_4$ now take lower concentration values, which translate into a limited tumor control. Balancing the maximum drug delivered during treatment, the maximum reduction in tumor burden, and the relative weights in the formulations of $J_2$, $J_3$, and $J_4$ is a challenge in OCT applications in cancer, which may render very different results. }
\label{fig:optcon}
\end{figure}
The main difference between the optimal treatment solution found for objective function $J_2$ versus $J_3$ and $J_4$ is that $J_2$ assumes that a constant dose is given over a certain interval, while $J_3$ and $J_4$ assume that the drug concentration $z(t)$ can change over time. These differences are shown in Fig.~\ref{fig:optcon}.
In principle, an on-off (bang-bang) control would be more clinically relevant than the continuous control, as patients are not usually treated over time in a continuous manner. However, this type of control would result in a sudden drop in $z(t)$ after the conclusion of each drug cycle (see Fig.~\ref{fig:optcon}), whereas in reality $z(t)$ decays with an approximately exponential trend. With the introduction of take-home infusion pumps for chemotherapy \cite{Zahnd1999}, evaluating continuous control may be a more plausible avenue of investigation.
However, the optimal $z(t)$ obtained under continuous control may also be unachievable with the current drugs for cancer treatment (e.g., due to incompatible pharmacokinetics).
Ultimately, the optimal drug concentrations $z(t)$ obtained with these functionals can render valuable information to guide the design of clinically-feasible optimal therapeutic strategies. Another future goal could even be the synthesis of new drugs or the adaptation of current drug compounds to match target dynamics emanating from the combination of OCT, cancer modeling, and pharmacodynamics \cite{Iyengar2012,Shi2017,Luepfert2005}.
Finally, optimal control problems using $J_2$, $J_3$, and $J_4$ may include clinically-relevant constraints to further focus the solution, including the limitations to $z(t)$ discussed for $J_1$. For instance, we can limit the tumor burden at any time by imposing $n(t)<n_r$, where $n_r$ is an arbitrary threshold. Additional constraints can also limit the frequency $f$ of doses, for example, by setting $f \leq T/n_d$, where $n_d$ is the maximum number of doses per treatment period.
Larger systems of equations that may account for the healthy and immune cell populations may also require additional constraints and/or incorporate these other variables into the optimal control problem itself. Beyond biological concerns are the logistical, monetary, and psychological costs that may also be considered for an optimal regimen.
The reader is referred to \cite{Jarrett2020a, Lenhart2007,Schaettler2015, Anita2011, Swan1990} for further details on applying OCT to cancer models.
\section{Barriers to success}\label{sec:barriers}
One of the principal issues preventing the successful translation of image-based computational modeling technologies to routine clinical practice is access to proper data.
Calibration and validation of image-based cancer models require individual quantitative imaging data at multiple time points during the course of surveillance or treatment.
The acquisition of frequent and rich imaging datasets might be feasible in preclinical settings, thereby enabling the realization of controlled studies to thoroughly assess model validity.
However, such controlled studies are extremely challenging in a clinical scenario for two key reasons. First, the timely and effective treatment of the patient and their quality of life are the utmost priorities. Thus, a controlled validation study would require the acquisition of minimal datasets to ensure a reliable calibration of the model, while ensuring that data collection is not a burden for patients and that those in the model-informed arm will be closely monitored in case immediate action is warranted during the course of the controlled study.
Second, quantitative imaging data are not routinely collected in the current clinical protocols.
Standard-of-care imaging primarily focuses on delineating tumor boundaries for staging and planning interventions (e.g., biopsy, surgery, and radiotherapy). However, anatomical imaging fundamentally limits modeling approaches as it does not quantitatively characterize the unique heterogeneous nature of each patient's tumor.
In addition, standard-of-care images are often acquired only before and following the completion of therapy. This is partly due to financial constraints, hospital workflow, scanner availability, and patient burden.
Thus, carefully designed studies are needed to determine a clinically-feasible strategy to collect sufficient quantitative imaging data enabling an accurate parameterization of predictive models of cancer.
For instance, abbreviated imaging protocols may fit quantitative, research-focused scans into routine clinical visits alleviating the need for separate research scans.
Additionally, many quantitative imaging data types can now be acquired in the community setting (i.e., away from major hospitals or oncology centers), which may be more convenient for patients \cite{Sorace2018}.
From a modeling viewpoint, a central challenge is balancing model complexity, a rational use of computational resources, data requirements to ensure an accurate calibration, and predictive accuracy.
The dynamics of cancer growth and treatment response is extremely complex, involving a multitude of biophysical processes interacting at various spatial and temporal scales \cite{Marusyk2010}.
Mathematical models of cancer are built upon a series of relevant biophysical phenomena, whose selection, formulation, and calibration ultimately determines the predictive power of the model \cite{Yankeelov2013,Yankeelov2015,Rockne2019}.
The accurate modeling of some of these phenomena (e.g., angiogenesis \cite{Vilanova2017}) would require complex equations at multiple scales, a large number of parameters requiring extensive spatiotemporal data, and advanced numerical methods that may incur in a large computational cost.
However, such models would be incompatible with the constraints on quantitative imaging data availability and patient care during the controlled validation studies discussed above.
Thus, clinically-oriented image-based models usually require conservative modeling assumptions, which enables the description of cancer growth and treatment response at organ scale using simple modeling paradigms and involving a minimal set of parameters whose calibration is feasible with existing quantitative imaging techniques \cite{Yankeelov2013}.
However, these model assumptions may incur substantial errors that ultimately limit the predictive power of the model.
Moreover, some biophysical model fields (e.g., tumor cell density) are not directly observed. Instead, the observed imaging data must be preprocessed to create a proxy to the biophysical observables. Such \emph{pseudo-correspondences} introduce additional uncertainties and can significantly affect the reconstruction results~\cite{Gholami:2016a,Swanson:2008a,Le:2017a,Harpold:2007a}.
Likewise, the phenomenological nature of macroscopic cancer models introduces parameters without direct biological counterparts (e.g., tumor cell diffusivity, drug efficacy), which further complicates model assessment and validation.
Nevertheless, the future development of image-based cancer models requires the initial validation of simpler models because this enables the identification of key improvements in model formulation to refine predictive accuracy. In addition, the success of simpler models also provides justification to collect further data to calibrate more complex biophysical mechanisms in subsequent model extensions.
For example, this has been the process behind the extension of the logistic model in Eq.~\eqref{eq:log} to the reaction-diffusion model in Eq.~\eqref{eq:rd}, then to a mechanically-coupled model in Section~\ref{sec:mech}, and later to a vasculature-coupled or drug-informed model in Sections~\ref{sec:vasc} and \ref{sec:ct}, respectively.
Biophysical inversion is a promising strategy to calibrate predictive models of cancer, but also presents several challenges. Complex, typically nonlinear and time-dependent, PDE-based models often result in ill-conditioned and non-convex optimization problems, which require sophisticated numerical algorithms to stabilize the inversion, such as multiresolution continuation, parameter continuation, and regularization schemes \cite{Scheufele:2020,Scheufele:2019a,Gholami:2016a,Subramanian20IP}.
Data scarcity can exacerbate the ill-posedness.
Mitigating this issue can entail imposing additional modeling assumptions and regularization strategies \cite{Subramanian20IP}.
The noise arising from various sources in imaging data also complicates model inversion. Further modeling priors and structure-exploiting algorithms can help mitigate some of these issues. Other mathematical considerations such as the choice of mismatch function (e.g., $L^2$ loss, cross-entropy loss) and regularization models can further complicate the inversion.
Inversion methods may also require several forward model evaluations, so specialized solvers are needed to prevent prohibitive computational costs.
OPAL is an attractive methodology to decide potential model extensions by comprehensively assessing the improvement in model predictions against the increase in model complexity.
In particular, OPAL can guide the modeler to select the \emph{best} valid model representing a certain experimental or clinical setting while accounting for uncertainties in both data and model parameters.
However, OPAL may also face certain challenges.
For example, computationally expensive models might require a more efficient approach for sensitivity analysis than the Sobol method, such as the elementary effects method or metamodels \cite{saltelli2008global}. Another major difficulty in model selection is that every model in the initial set of models might be invalid (i.e., they do not satisfy the validation criteria), which would require to extend such set with further models.
Finally, application of OCT to mathematical models of tumor growth and treatment response is a promising strategy for the optimization of therapeutic regimens \textsl{in silico} \cite{Jarrett2020}. However, the reliability and plausibility of solutions generated by OCT methods depend on several factors including the validity of the model, the accurate definition of the objective functional, the uncertainty in the parameters and data, the application of clinically-relevant constraints, and the accuracy in solving the OCT problem itself.
Additionally, implementation of OCT approaches within the clinical trial system is even more complicated than model validation. Beyond the challenges on data availability, model assumptions, and biophysical inversion described above, a controlled clinical study to validate OCT-generated therapies would involve the test of novel computationally-derived regimen protocols in patients. This requires a close monitoring of treatment response, toxicity, and patient well-being, which can be extremely complex to balance and maintain during such clinical study. Thus, robust preclinical evidence showing the advantages of OCT for the design of therapeutic regimens is required before advancing to clinical scenarios.
\section{Conclusion}\label{sec:conc}
Integrating quantitative data obtained from biomedical imaging with mechanism-based mathematical modeling represents a significant departure from current para-digms in cancer biology and oncology. More specifically, this approach is fundamentally different from the current trend in modeling which emphasizes applying the methods of artificial intelligence to extremely large data sets. However, statistical inference - though enormously powerful - relies on properties of large populations that can frequently obscure important characteristics (or conditions) that are specific to individual patients and may drive the development of their disease or their response to therapy. The high-consequence decisions present in clinical oncology simply must be based on more than data analytics. These decisions must incorporate biophysical processes within a rigorous, mathematical framework that can be calibrated with patient-specific data to make patient-specific predictions. The transformation from population-based care to patient-based care is inevitable, and the intimate integration of quantitative imaging, mechanism-based mathematical modeling, and efficient computational methods enabling precise \emph{in silico} tumor forecasts is a very promising avenue to achieve this important goal.
\section*{Acknowledgements}
GL was partially supported by a Peter O'Donnell Jr. Postdoctoral Fellowship from the Oden Institute for Computational Engineering and Sciences at The University of Texas at Austin and acknowledges funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sk\l{}odowska-Curie grant agreement No. 838786.
EABFL is supported by the Oncological Data and Computational Sciences collaboration between the Oden Institute for Computational Engineering and Sciences, The University of Texas MD Anderson Cancer Center, and Texas Advanced Computing Center.
TEY is a CPRIT Scholar in Cancer Research and acknowledges funding from CPRIT RR160005, NCI U24CA226110, NCI R01CA235800, and NCI U01CA253540.
\begin{small}
\input{CCC_arxiv_v3.bbl}
\end{small}
\end{document}
|
\section{Introduction}
The transport of cosmic rays (CRs) in the Galaxy is a crucial piece in the puzzle that leads to understanding the origin of these high energy particles. The description of CR transport is based on a macroscopic description of the particle motion (advection and diffusion) that is based on microscopic processes (particle wave resonances and CR self-generation) and particle physics (radiative losses and nuclear fragmentation). As such, it represents an endeavour of remarkable proportions, that at best can be pursued by using effective theories. The advection-diffusion equation engulfs such difficulties in a way that led to numerous successes and that requires continuous adjustments to reflect pieces of observations that are being acquired.
The precision measurements made by AMS-02 in the last decade or so have affected our understanding of the origin of CRs in a profound way, by confirming some pillar ideas and by revealing anomalies that might be signs of a need for a deep revision of some aspects of the problem: 1) the observation of spectral breaks in both primary and secondary nuclei ~\cite{PAMELA.2011.phe,AMS02.2015.protons,AMS02.2015.he,AMS02.2018.libeb} has driven us towards developing theories that could accommodate such features. The fact that the breaks appear also in the ratio of secondary-to-primary nuclei is a clear indication that they are most probably due to some new phenomenon occurring in particle transport rather than in the accelerators or in the spatial distribution of sources~\cite{Genolini2017prl}; 2) the AMS-02 measurements showed beyond doubt that protons, helium and heavier nuclei need to be injected at the sources with slightly different spectra, a finding that is at odds with the basic idea of diffusive shock acceleration. 3) The spectrum of positrons in CRs~\cite{AMS02.2013.posfraction,AMS02.2019.positrons} is very different from what one would expect as secondary products of hadronic interactions, thereby requiring additional sources of these antimatter particles~\cite{Hooper2009jcap,Serpico2012aph,Bykov2017ssrv,Manconi2020prd,Evoli2020arxiv}. This might also be the case, although less clearly, for antiprotons~\cite{Boudaud2020prr,Heisig2020prr,AMS02.2016.pbar}. 4) The measurement of beryllium~\cite{AMS02.2018.libeb} provided a firm confirmation that the CR residence time in the Galaxy requires extended magnetized halos, with size $\gtrsim 5$~kpc~\cite{Evoli2020prd,Weinrich2020aaa}.
In most of these lines of investigation, the main limitation to our capability to extract physical information from the high precision data provided by AMS-02 comes from uncertainties in the cross sections for fragmentation of nuclei in spallation reactions~\cite{Webber1998apj,Silberberg1998apj,Mashnik2004adspr,Moskalenko2001icrc,Moskalenko2003icrc,Moskalenko2003apj,Tomassetti2017prd,Genolini2018prc}. It is probably worth stressing that this is also one of the main sources of systematic error in the measurement of the flux of heavy nuclei, due to the fragmentation of those nuclei inside the experiment, an effect that needs to be corrected for.
The recent publication of the results of the measurement by AMS-02 of the spectra of intermediate mass nuclei in CRs~\cite{AMS02.2019.nemgsi} up to rigidities of order $\sim$TV, provides us with the most complete set of information about CR transport in that such nuclei are the parent particles for lighter nuclei through processes of spallation (for stable nuclei) and decay (for unstable isotopes). Such measurements extend now to iron nuclei, which have a predominantly primary nature~\cite{AMS02.2021.iron}.
Provided particles are transported either diffusively or advectively over distances much larger than the thickness of the disc ($h\sim 150$ pc), the transport of CRs in the Galaxy is properly described by a weighted slab model~\cite{Jones2001apj,Aloisio2013jcap}, recently used to derive the basic properties of the CR transport (diffusion coefficient and advection velocity)~\cite{Evoli2019prd} and the size of the confining halo $H\gtrsim 5$ kpc, from the measurement of the Be/B ratio~\cite{Evoli2020prd}.
This last goal could be achieved because of the fact that the $^{10}$Be unstable isotope decays to $^{10}$B, so that both numerator and denominator of the Be/B ratio are affected. The idea that this decay (and those of similarly unstable isotopes of other nuclei) could be used to infer the halo size dates back to~\cite{Ptuskin1998aap,Webber1998apj,Putze2010aap,Strong1998apj,Trotta2011apj} although such calculations were typically carried out using the $^{10}$Be/$^{9}$Be ratio (or similar ratios) at low energies. Notice also that some early papers used the Leaky Box model to derive the confinement time in the halo, a procedure that is known to lead to incorrect estimates.
AMS-02 cannot separately measure the flux of these isotopes, but on the other hand it can measure the Be/B ratio up to high energies, $\sim$~TeV/n, where the size of the halo $H$ can be inferred without the huge uncertainties that derive from lack of information on CR transport in the disc, which is crucial for unstable nuclei at $\lesssim 100$ MeV/n energies.
The set of calculations discussed above showed that a consistent picture of CR transport can be obtained by analyzing the spectra of primary and secondary elements lighter than oxygen, while the secondary/primary ratios allowed us to determine the grammage traversed by CRs inside the Galaxy, thereby confirming that a break in the energy dependence of the diffusion coefficient is needed to describe AMS-02 data.
In fact the measurement of the beryllium flux and of the Be/B ratio led to a slight readjustment of the parameters due to small production of boron from the decay of the unstable $^{10}$Be isotope~\cite{Evoli2020prd} (this effect on the boron spectrum was noticed early on by~\cite{GarciaMunoz1987apjs}). As stressed above, the main limitation to the strength of these conclusions derives from the experimentally poor knowledge of the spallation cross sections, especially the partial cross sections for specific channels of fragmentation. A discussion of these uncertainties was provided in~\cite{Tomassetti2012apj,Genolini2018prc,Evoli2018jcap}.
Another source of uncertainty, though to a smaller extent, was the fact that the abundances of heavier elements, acting as primaries for lighter nuclei, were poorly known and had to be taken from older experiments. This gap has been filled by the recent measurements of the fluxes of intermediate mass and heavy elements by AMS-02.
In this article we apply the same approach, already mastered in~\cite{Evoli2019prd,Evoli2020prd}, to heavy nuclei, up to iron. We show that the spectra of all nuclei heavier than oxygen are described within uncertainties with an injection spectrum at the sources that is the same as for lighter intermediate mass nuclei (CNO), provided the contribution of spallation from even heavier nuclei is taken into account. This is in some tension with the statement by the AMS-02 collaboration that Ne, Mg and Si are three primaries that behave peculiarly, in that they have a different spectral index than He, C and O. We maintain that this behaviour is fully accounted for by properly calculating the secondary contribution to these nuclei from their heavier parent nuclei, a situation similar to that of nitrogen.
The only exception to this general conclusion is the spectrum of iron as measured by AMS-02, which appears to be at odds with the predictions of the standard model. The main anomaly is a hard shape of the Fe spectrum for $R\lesssim 100$ GV, which is not compatible with the grammage inferred from secondary/primary ratios for lighter nuclei. Moreover, the ratio Fe/O measured by AMS-02 appears to be in tension with previous results from HEAO~\cite{Engelmann1990aa} and Voyager~\cite{Cummings2016apj}. We present here the results of several tests proposed to address these problems.
The article is structured as follows: in \S~\ref{sec:model} we provide a brief summary of the theoretical approach adopted here, based on the weighted slab technique applied to the whole chain of nuclei, from H to Fe. In \S~\ref{sec:method} we describe the methodology used to achieve a description of the nuclear spectra, as well as the secondary/primary ratios. In \S~\ref{sec:results} we describe in detail our findings, starting from nuclei of intermediate mass (Ne, Mg, Si) and going up to iron. We summarize our conclusions in \S~\ref{sec:conclusions}.
\section{Propagation model}
\label{sec:model}
The details of the transport equation and the approach we follow to solve it are detailed in~\cite{Jones2001apj,Aloisio2013jcap,Evoli2019prd,Evoli2020prd}. Here, we simply summarize the most important features of the model.
We assume that the CR density of stable and unstable nuclear species obeys a steady-state diffusion-advection equation~\cite{TheBible}:
\begin{multline}\label{eq:slab}
-\frac{\partial}{\partial z} \left[ D_{a} \frac{\partial f_a }{\partial z} \right]
+ v_A \frac{\partial f_a}{\partial z}
- \frac{dv_A}{dz} \frac{p}{3} \frac{\partial f_a}{\partial p}
\\
+ \frac{1}{p^{2}} \frac{\partial}{\partial p} \left[ p^{2} \left(\frac{dp}{dt}\right)_{a,\rm ion} f_a \right]
+ \frac{\mu v(p) \sigma_a}{m} \delta(z) f_a
+ \frac{f_a}{\hat\tau_{d,a}}
\\
= 2 h_d q_{0,a}(p) \delta(z)
+ \sum_{a' > a} \frac{\mu\, v(p) \sigma_{a' \to a}}{m}\delta(z) f_{a'}
+ \sum_{a' > a} \frac{f_{a'}}{\hat\tau_{d,a'}},
\end{multline}
where $f_a (p,z)$ is the distribution function of species $a$ in phase space, $v(p)=\beta(p) c$ is the particles' velocity, and $\mu = 2.3\,$mg/cm$^2$ is the surface density of the disk.
The transport equation in Eq.~\ref{eq:slab} is obtained assuming spatially homogeneous diffusion and that the sources of CRs and the target gas are confined in an infinitely thin disc (half-thickness $h \simeq 100$~pc), in which all the interactions (spallation and ionization energy losses) occur.
We assume that CRs are confined in a low-density infinite slab of half-thickness $H$, extending well beyond the gaseous disc ($H \gg h$). Outside the magnetic halo, the particles can escape freely into intergalactic space so as to reduce the CR density to $f_a(p, z = \pm H) \simeq 0$.
For the diffusion coefficient we adopt the parametrisation proposed in~\cite{Evoli2020prd}:
\begin{equation}\label{eq:diffusion}
D(R) = 2 v_A H + \beta D_0 \frac{(R/{\rm GV})^\delta}{[1+(R/R_b)^{\Delta \delta /s}]^s},
\end{equation}
where $D_0$ and $\delta$ are fitted to secondary over primary ratios, while the other parameters $s$, $\Delta \delta$ and $R_{b}$ are fixed from observations of primary nuclei.
The change of slope above the rigidity $R_b$ in Eq.~\ref{eq:diffusion} has been implemented to account for the presence of similar breaks in the spectra of primaries~\cite{AMS02.2017.heco} and in the secondary/primary ratios~\cite{AMS02.2018.libeb}. As shown in~\cite{Genolini2017prl}, the spectral breaks seen in CR data at a rigidity of $\sim 300$~GV must be attributed to a change of properties in the Galactic transport rather than in the injection.
While a break in the diffusion coefficient was hypothesized by several authors, even before AMS-02 data \cite{Vladimirov2012apj,Evoli2012prd}, theoretical investigations have been carried out immediately after, to motivate this functional form in terms of the transition from self-generated turbulence to preexisting turbulence~\cite{Blasi2012prl,Aloisio2013jcap}, or of a non-separable spatially dependent diffusion coefficient in the Galactic halo~\cite{Tomassetti2012apj}.
The plateau at low energies, where advection dominates transport, was found as a consequence of self-generated turbulence as described in~\cite{Recchia2016mnras}. Advection of CRs with self-generated waves is an unavoidable phenomenon peculiar of these models and virtually incompatible with the existence of second order Fermi acceleration on Galactic scales.
The advection velocity $v_A$ is assumed to be constant above and below the disk, hence adiabatic energy losses (third term in the LHS of Eq.~\ref{eq:slab}) are determined through the expression $dv_A/dz = 2v_{A} \delta(z)$.
The injection of CR nuclei of type $a$ is given as a power law in momentum, having in mind that they are predominantly accelerated at SN shocks~\cite{Blasi2013aarv,Morlino2017book,Blasi2019ncimr}, hence the source term can be written as $q_{0,a} \propto p^{-\gamma_{\rm inj}}$.
The slope $\gamma_{\rm inj}$ is assumed to be the same for all nuclei heavier than He, while the normalizations are chosen for each individual species to match the data.
Notice that the assumption of universal injection for all nuclei, although strongly supported by theoretical grounds, does not apply to light CRs, as it has been shown that the injection of protons(helium) requires a softer(harder) slope than nuclei~\cite{Evoli2019prd,Weinrich2020aab}.
The other terms on the LHS of Eq.~\ref{eq:slab} describe the disappearance of the species $a$ because of spallation ($\sigma_a$) or of radioactive decay ($\hat\tau_{d,a}$). In the RHS, one can find the corresponding terms describing the production of the species $a$ by spallation and decay of heavier ($a' > a$) elements.
The inelastic spallation cross sections are computed using the parametric fits provided by~\cite{Tripathi1996nimpb,Tripathi1997nimpb}, while for the secondary production cross-sections we adopted the parametric fits of~\cite{Evoli2019prd} for the major primary channels of Li, Be and B production, and the approach described in~\cite{Evoli2018jcap} to compute all other cross-sections (including the cumulative contribution of ghost nuclei).
We remind that this approach is based on cross-section measurements (when available) to set the normalization of each spallation channel.
To this aim, we make use of the GALPROP cross-section measurement database~\cite{Strong2001adspr,Moskalenko2001icrc,Moskalenko2003icrc} and a collection of more recent measurements~\cite{VillagrasaCanton2007prc}, in particular for the production of heavy fragments (Ne, Mg, Si, S) by iron spallation~\footnote{In order to facilitate the reproducibility of our results we release the cumulative cross-sections adopted to compute the secondary production via a dedicated repository~\cite{EvoliXsecs}}.
The quantities $\hat \tau_{d, a}=\gamma \tau_{d,a}$ define the Lorentz boosted decay lifetimes of unstable elements.
In~\cite{Evoli2020prd}, some of us addressed the effect of including the decay of $^{10}$Be, and the corresponding production of $^{10}$B, on the calculation of the Boron-over-Carbon ratio (B/C) and on the determination of the transport parameters.
Since the CR propagation in the environment of the Galactic disc might take place under different conditions, hostile to diffusion, we studied the minimum rigidity above which the $^{10}$Be decay takes place predominantly outside the disc. We found that this condition is well satisfied for rigidities $\gtrsim$ few GV's even for halos as small as $H \sim 1$~kpc \cite{Evoli2020prd}.
Above this rigidity, the Beryllium-over-Boron ratio as measured by AMS-02 was found to be best reproduced assuming $H\sim 7$ kpc, which we assume to be fixed here.
As in this work we focus on the propagation of heavy elements, additional unstable isotopes heavier than oxygen must be considered.
The unstable isotopes having lifetimes comparable with the propagation timescales at these energies are the $\beta^-$ unstable isotopes $^{26}$Al, $^{36}$Cl and $^{54}$Mn (see also~\cite{Donato2002aa}).
The shortest lifetime is the one of $^{36}$Cl, which is a factor $\sim 5$ smaller than the one of Be.
From Fig.~1 in~\cite{Evoli2020prd} we conclude that for halo sizes $H\gtrsim 6\,$kpc the condition that the nucleus decays outside the disc of the Galaxy is fulfilled above a few GV.
All other unstable isotopes have lifetimes at rest shorter than few kyrs and are treated as ghost nuclei, which means that we assume their instantaneous decay.
Finally, we comment on the fact that we do not model the decay modes involving electron capture, since it is important only at energies much smaller than $\sim$~GeV/n~\cite{Letaw1985apss}.
We account for the effect of solar modulation by using the force field approximation~\cite{Gleeson1968apj} with a Fisk potential $\phi$.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{BCO.pdf}
\caption{B/C and B/O ratios as derived by using our calculations (solid line) compared to recent AMS-02 data~\cite{AMS02.2018.libeb}.}
\label{fig:light_ratios}
\end{figure}
\section{Methodology}\label{sec:method}
In order to get predictions for the fluxes of heavy elements, like Ne, Mg, Si, S and Fe, we include in our calculation all the stable isotopes from B ($Z=5$) to Ni ($Z=28$). Additionally, we propagate the unstable isotopes $^{10}$Be, $^{26}$Al, $^{36}$Cl and $^{54}$Mn as described in \S~\ref{sec:model}. This procedure requires fitting of numerous parameters, that we summarize here.
Spatial transport, including diffusion and advection, comprises 7 free-parameters: $D_0$, $\delta$, $v_A$, $H$, $R_b$, $\Delta \delta$, $s$. To speed up the analyses, as mentioned before, we fix the halo size to be consistent with the value derived from the analysis of the Be/B ratio, $H = 7$~kpc~\cite{Evoli2020prd}, having in mind that secondary-to-primary ratios are degenerate for $D_0/H$, to an excellent degree of approximation.
Moreover, we infer the three high-rigidity break parameters ($R_b$, $\Delta \delta$, $s$) with a fit to the H and He fluxes. Noticing that they are practically the same as the best-fit values of the recent AMS-02 publication \cite{AMS02.2021.review} , i.e. $s=0.09$ and $\Delta \delta=0.22$, obtained by fitting the proton spectrum.
However we require a somewhat lower $R_b$ of $290\,$GV to simultaneously fit He which is still compatible with the fit provided in \cite{AMS02.2021.review} within their error bars.
These assumptions reduce the number of free parameters which describe the transport to 3.
The injection efficiencies $\epsilon_a$ of the species C, N, O, Ne, Mg, Si, S and Fe are set in such a way that the total flux of each species match the AMS-02 data. That amounts to additional 8 parameters. For other primary species, in particular Na, Al, P, Ar, Ca, Cr, Mn and Ni, the TRACER03~\cite{Ave2008apj}, CRISIS~\cite{Young1981apj}, and HEAO3-C2~\cite{Engelmann1990aa} measurements have been used to fix their abundances and we do not include these normalizations as free parameters in the best-fit search. Notice that the efficiency $\epsilon_a$ pertains to the injection of a given charge, including all isotopes. We distribute the efficiency $\epsilon_a$ among the isotopes in order to reproduce the isotopic abundance in the interstellar medium (ISM) as measured in~\cite{LoddersPalme2009}.
The solar modulation potential, $\phi$, and the injection slope, $\gamma_{\rm inj}$ are two additional free parameters, bringing the total number of free parameters to 13.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{heavyelements.pdf}
\caption{Flux of Ne, Mg, Si and S (solid orange line) compared to recent AMS-02 data. The shaded area shows the effect of cross section uncertainties on the predictions. The dashed lines reflect the fluxes obtained by only including the primary contribution (no secondary production). The red line shows the expected flux for the given element, if the cross sections of the main production channels of Mg are increased by 30\% and the primary component of Mg is adjusted accordingly.}
\label{fig:fluxes}
\end{figure}
To fit the propagation parameters against different datasets we use the MINUIT package~\cite{JamesRoos1975}. To ensure that true minima are found, $O(50)$ minimisations from different starting points are carried out for all our analyses. The quantity we minimise is the $\chi^2$ computed for the AMS-02 measurements of different datasets.
Specifically, the total $\chi^2$ is computed by summing the reduced $\chi^2$ computed over the ratios Be/C, B/C, Be/O, Ne/Mg, Si/Mg, Ne/O, Mg/O, Si/O, and the absolute fluxes of B, C, N, O, Ne, Mg, Si, and S~\cite{AMS02.2017.heco,AMS02.2018.libeb,AMS02.2019.nemgsi}.
To normalize the S primary injection we make use of the AMS-02 preliminary data on the S absolute flux as reported in seminar presentations~\footnote{See, e.g., \url{https://indico.gssi.it/event/80/}}.
The absolute fluxes are fitted only for rigidities larger than $\sim 10$~GV in order to minimize the effect of solar modulation (see also~\cite{Evoli2019prd}).
Finally, for each dataset the total uncertainty is computed by adding systematic and statistic errors in quadrature~(see also~\cite{Weinrich2020aaa} for an attempt to take into account the correlation among systematics).
\section{Results}
\label{sec:results}
\subsection{Intermediate mass nuclei}
\label{sec:heavy}
\begin{figure}
\centering
\includegraphics[width=0.95\columnwidth]{heavy_over_O.pdf}
\caption{Ratios of Ne, Si and Mg over O compared to recent AMS-02 data. The shaded region indicates the effect of cross section uncertainties on the model. The red line shows the case with higher Mg production cross sections.}
\label{fig:Oratios}
\end{figure}
The propagation parameters that best fit the data, using the method described in \S \ref{sec:method} are: $\delta = 0.56$, $\gamma_{\text{inj}} = 4.33$, $D_0/H = 0.35$ (in units of $10^{28}\,$cm$^2$ s$^{-1}$ kpc$^{-1}$, $v_A = 4.4\,$km/s and $\Phi = 0.49\,$GV. We notice that the best-fit values are in perfect agreement with the results in~\cite{Evoli2020prd} where the fit was performed using only the AMS-02 data on nuclei lighter than oxygen.
Following the procedure first introduced in~\cite{Evoli2019prd}, in order to illustrate the impact of cross section uncertainties on our results we have repeated our calculations 500 times for the best-fit scenario.
In each repetition the spallation cross sections of elements heavier than oxygen are each rescaled by an individual factor. For each cross section the factor is sampled from a Gaussian distribution with mean $1$ and a width of $0.1$ for the total cross sections of a given element and $0.3$ for the partial cross sections of different channels respectively.
In all the following figures, the gray-shaded areas represent the 1-$\sigma$ variance of our calculations associated with the uncertainty in the cross-section.
In fact, the uncertainty in the cross sections might also affect the best-fit parameters, in particular the injection efficiencies, hence the error bands shown in our figures only capture part, though most likely the bigger part, of the effect of uncertainties in spallation cross sections.
In Fig.~\ref{fig:light_ratios} we show how the B/C and B/O ratios are correctly reproduced by our best-fit. The variance due to the uncertainty on the heavy-element cross sections is of the order of $\sim 1$\% at $\sim 10$~GV (and smaller at higher energies). This is due to the fact that the secondary contribution of elements heavier than oxygen to boron and carbon is sub-dominant compared to that of carbon, nitrogen, and oxygen~\cite{Evoli2018jcap}. This result justifies our previous attempts to extract the diffusion parameters by fitting the B/C and B/O even with a poor knowledge of the normalization of the flux of heavier elements.
The reduced $\chi^2$ for all the fits shown throughout this article is $\lesssim 1$, hence the overall description of data is excellent. On the other hand one might argue that a few data points in the B/C and B/O ratios at $R\gtrsim 400$ GV appear to be slightly displaced from the theoretical curve. This gives us the opportunity to point out that in this energy range there are at least three guaranteed effects that are expected to produce small deviations from the standard predictions (curves in Fig.~\ref{fig:light_ratios}): 1) CRs must experience some grammage inside the accelerators, which clearly depends on what the sources are.
For instance for the case of supernova remnants (SNRs) as sources of CRs, it was estimated in Ref. ~\cite{Aloisio2015aap} that the roughly energy-independent source grammage should be $\sim 0.1-0.2$~g/cm$^2$.
At $\sim 400$ GV this would cause a $\sim 5-10\%$ correction with respect to the standard estimate. 2) Some of the secondary nuclei produced around the acceleration region should be re-energized at the shock, which should somewhat change the estimate and energy dependence of the source grammage discussed above \cite{Berezhko2003aa,Blasi2009prl,BlasiSerpico2009prl}. 3) Secondary nuclei produced during Galactic transport occasionally encounter SNR shocks and get accelerated there. This effect, discussed in detail by \cite{Blasi2017mnras,Bresci2019mnras}, leads to small deviations from the standard expectation at high rigidities.
Although all these effects are guaranteed to play a role starting at $\sim$ few hundred GV rigidities, we do not have the capability to calculate them at the level of accuracy required to compare predictions with AMS-02 data. However, it is reassuring that even the standard calculations in the same rigidity range are uncertain at the $\sim 10\%$ level (Fig.~\ref{fig:light_ratios})).
Adopting the procedure discussed above, we obtain a best-fit that correctly reproduces the Ne, Mg, Si and the preliminary S data using the same injection slope as the intermediate nuclei, C and O. This can be appreciated in Fig.~\ref{fig:fluxes}. The case of Fe nuclei will be discussed later.
In each panel we show with a dashed line the result of our calculations obtained if the contribution to the flux of that nucleus from spallation of heavier elements were neglected. This allows us to immediately identify the species that are mainly primaries from those that receive a substantial secondary contribution: for instance, the Si flux is predominantly of primary origin (like C and O), while the Ne, Mg and S elements have mixed origin, resembling the case of Nitrogen. In all these cases the secondary production fills the low energy part of the spectrum, as one might easily have expected.
In Fig.~\ref{fig:fluxes}, a marginally worse agreement with data is seen in the case of Mg. However, given the uncertainties in the partial cross section for dominant channels of $\rm ^{26}Mg$ production from $\rm ^{28}Si$, we also studied the effect of increasing such cross sections by 30\% and reducing the source normalization correspondingly.
The resulting total Mg flux is shown as a solid red line in Fig.~\ref{fig:fluxes}, and agrees with AMS-02 data points at high energy visibly better than for the nominal value of the cross section for the branch $\rm^{28}Si\to {^{26}Mg}$.
In Fig.~\ref{fig:Oratios} we show the ratios of Ne, Mg and Si over O, as functions of rigidity. The choice of plotting these ratios with respect to oxygen is due to the fact that oxygen can be considered (together with protons and iron) a pure primary species, meaning that the secondary contribution from spallation of heavier nuclei is negligible at all energies.
All of these ratios are in good agreement with measurements for $R\gtrsim 10\,$GV.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{heavyratios.pdf}
\caption{Ratios of Ne and Si over Mg compared to recent AMS-02 data. The shadowed region indicates the effect of cross section uncertainties on the model. The red line shows the case with higher Mg production cross sections.}
\label{fig:ratios}
\end{figure}
Although Mg/O and Si/O slightly overshoot the data at very high rigidity, the experimental uncertainties of these data points are quite large, and we do not see this as evidence of a problem at this point. Moreover, for the case of the Mg/O ratio, we explored the effect of a possible increase by 30\% in the production cross section of Mg (solid red line, as in Fig.~\ref{fig:fluxes}): the increase of the cross section allows us to slightly reduce the normalization of the primary contribution to the flux of Mg, and hence to get a somewhat better agreement with data points. Similar considerations can be made for other heavy nuclei, for which the partial cross sections for spallation are known with comparably large uncertainty. In turn, this might be seen as a possible indication that our face values for at least some of these cross sections, might be underestimated.
The case of Mg is a good illustration of the point already made earlier in this section: the shaded areas only catch part of the effect due to the uncertainties in the spallation partial cross sections. Changing these values leads to a change in the normalization of the primary contribution that is not accounted for in the shaded areas.
In Fig.~\ref{fig:ratios} we also show the ratios of the Ne and Si fluxes over the Mg flux. Again, the agreement with data is remarkable. The only possible discrepancy with data could be identified in the Ne/Mg ratio at high rigidities. However, even this ratio is better described in terms of a mild increase in the production cross section of Mg, as discussed above (solid red line).
In all these cases, it remains true that the main limitation to uncover the physics of CR transport and acceleration is in the poorly known cross sections for production of elements through spallation reactions.
\subsection{Iron}
\label{sec:Fe}
Iron is as close to a primary nucleus as H and O are, hence the spectrum of iron can be considered as a reliable test of the injection spectra derived from the analysis illustrated above, and of the grammage traversed by CRs. The latter enters the calculation of the equilibrium spectrum of Fe through the severe spallation losses, which are expected to affect the flux of iron for $R\lesssim100$ GV, making it substantially harder than at higher rigidities.
\begin{figure*}
\centering
\includegraphics[width=\columnwidth]{O_exp.pdf}
\includegraphics[width=\columnwidth]{Fe_exp.pdf}
\caption{Spectrum of oxygen (left) and iron nuclei (right) as predicted in our calculations (solid line), with an estimate of the uncertainties induced by the poor knowledge of cross sections (shaded area). The data points from AMS-02 and other experiments are also shown.}
\label{fig:OandFe}
\end{figure*}
The spectra of oxygen (left panel) and iron nuclei (right panel) that we derive using our calculations are shown in Fig.~\ref{fig:OandFe}, together with data points from AMS-02 and other experiments (see labels on the figure). A few comments are due: the spectrum of oxygen is very well reproduced at all energies, and the spectral break at $\sim 300$ GV is well visible, because the transport is dominated by diffusion in that rigidity range. For iron nuclei, the situation is different in that no clear evidence for a break is visible either in the data or in the prediction. This is expected since the transport of Fe nuclei is dominated by spallation even at $\sim 100$ GV. The comparison between predictions of our calculations and AMS-02 data clearly shows that for $R\lesssim 30$~GV, there is a strong disagreement, not reconcilable with error bars quoted by the ASM-02 collaboration.
Following AMS-02~\cite{AMS02.2021.iron}, in Fig.~\ref{fig:FeO} we also show the Fe/O ratio as a function of kinetic energy per nucleon, which was also provided by the AMS-02 collaboration. In the same plot we show the same quantity as measured by ACE-CRIS~\cite{Lave2013apj}, HEAO3~\cite{Engelmann1990aa} and Voyager~\cite{Cummings2016apj}. The predicted ratio of modulated fluxes as derived using our calculations is shown as a solid (red) curve. Since the plot also contains a data point from Voyager, that measures unmodulated quantities, we also plot the results of our calculation of the Fe/O ratio of unmodulated fluxes (dashed line).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{FeO_kin.pdf}
\caption{Fe/O ratio of the modulated fluxes (solid line) as a function of kinetic energy per nucleon, compared with data from AMS-02, ACE-CRIS and HEAO. The dashed line shows the ratio of unmodulated fluxes, compared with the Voyager data.}
\label{fig:FeO}
\end{figure}
The problem mentioned above at low energies is confirmed in the Fe/O ratio, which is predicted to be appreciably higher than AMS-02 measurements for $E_{kin}/n\lesssim 20$ GeV/n. The discrepancy is $\sim 30-40\%$ at $\sim 1$ GeV/n, much larger than the quoted AMS-02 systematic uncertainty in the same energy region.
On the other hand it is worth noticing how the same ratio, measured by HEAO3 is also at odds with the AMS-02 result, while agreeing with our predicted trend and normalization. The Fe/O ratio as measured by ACE-CRIS at lower energies is also quite higher than an extrapolation of the AMS-02 data to lower energies. The case of Voyager data, at even lower energies confirms that the Fe/O ratio seems lower, although it is worth keeping in mind that Voyager measured unmodulated fluxes. On the other hand, our dashed curve (ratio of unmodulated fluxes) is perfectly compatible with the Voyager data point, despite all our fitting procedures were made only using data above $10$ GeV/n. The small bump in the dashed line reflects the energy dependence of spallation cross sections at low energies. A warning is needed here: our calculations are increasingly less trustworthy when moving to lower and lower energies, when ionization energy losses force the whole iron propagation to take place in the disc, that in our calculations is assumed to be infinitely thin. In fact, all calculations of CR transport should be taken with much caution when this happens, because the diffusion coefficient of CRs in the disc is very likely different from the one assumed on Galactic scales, due to the severe ion-neutral damping that suppresses CR scattering in the disc~\cite{Kulsrud1971apjl}.
Given the oddity of the Fe and Fe/O situation, especially when compared with the impressive agreement of our predictions with all other data presented by AMS-02 on nuclei, we decided to explore some possible caveats of our approach that might help unveiling the source of the discrepancy:
{\it 1)} the weighted slab model used in our work to describe particles' transport neglects the grammage accumulated by CRs in the halo. This is a good approximation as long as the density in the halo, $n_H$, is much smaller than $n_d(h/H)\sim 2\times 10^{-2}\rm cm^{-3}$. This appears to be a good approximation since the expected density in the halo is thought to satisfy this condition~\cite{Ferriere2001rvmp}. We actually calculated the solution of the transport equation of Fe nuclei in the presence of target gas in the halo and found that the low energy solution changes only at the percent level for gas density in the halo of order $10^{-3}\rm cm^{-3}$. Hence, it is unlikely that the discrepancy between predictions and AMS-02 data may be attributed to oversimplified CR transport in the halo. These effects might become of some importance only for halo gas density much larger than those suggested by observations.
{\it 2)} The uncertainties in the spallation cross sections are definitely a limiting factor in all these calculations. However most uncertain are the partial cross sections for fragmentation of a nucleus $A$ into a nucleus $A'$, while the total cross sections are somewhat better known. Nevertheless, for the sake of completeness, we tried to increase the spallation cross section of Fe by $40\%$, but even such drastic change turned out into a bad fit to the spectrum of Fe nuclei. Hence, the disagreement between theory and AMS-02 data below 100 GV cannot be due to an underestimation of the Fe spallation cross section.
{\it 3)} Despite the lack of physical support to the idea that iron nuclei may be injected with a different injection spectrum compared with other nuclei, we adopted an agnostic attitude and tried to find a best fit injection spectrum that could improve the fit (after all this is what is required for H and He). We find that adopting a harder injection spectrum (as required by the low energy Fe AMS-02 data), leads to only a slightly better fit at lower energies, while making the high energy fit worse.
{\it 4)} Even more aesthetically unappealing is the possibility that the solar modulation may take place in a different way for iron and for oxygen. Imposing this unnatural solution, one finds that a good fit would require a Fe modulation about 70\% stronger than for oxygen.
In a recent article~\cite{Boschini2021arxiv}, the authors have noticed the same problem with Fe nuclei when comparing AMS-02 data with GALPROP predictions. It was suggested that in order to fit the AMS-02 iron spectrum it is necessary to introduce multiple breaks in the injection (source) spectrum. In fact the need for these multiple breaks is common to other nuclei as well, as found by~\cite{Boschini2020apjs}, quite at odds with the results illustrated in the present article, where the source spectra are pure power laws.
In~\cite{Boschini2021arxiv}, even at high energies, where all nuclei are expected to behave in the same way under the effect of diffusion, iron is assumed to be injected with a spectrum different from that of other nuclei. Moreover, in order to accommodate the obvious disagreement of the predicted flux with Voyager data, they suggest that the contribution from few local sources could become dominant. Clearly it is difficult to disprove such a model, especially because no modeling of such contribution was discussed. We notice however that the contribution of local sources at energies of few GeV/n seems unlikely. For an iron nucleus at that energy the distance traveled under the effect of losses is $d_l=(D \tau_l)^{1/2}\sim 1$ kpc (where $\tau_l$ is the energy loss timescale including ioinization and spallation, see also~\cite{Morlino2020prd}), too large to expect that something special may be happening. If the authors of~\cite{Boschini2021arxiv} were referring to a fluctuation due to accidentally local and recent sources, then this would in general lead to even larger fluctuations at higher energies where the global fits illustrated above seem to provide an excellent description of the data, with no need for local sources.
Finally, we find the assumption of multiple breaks in the source spectrum in~\cite{Boschini2021arxiv} not only lacking a solid justification on physical grounds, but also rather puzzling in that the breaks are different for different elements, thereby suggesting no clear trend in the yet to be found underlying physics.
We stress once more that the AMS-02 data on the Fe/O flux are not only hard to reconcile with the results of existing calculations of CR transport on Galactic scales, but also and perhaps more important, with the results of previous experimental endeavours.
\subsection{H and He}
\label{sec:HeH}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{HeH.pdf}
\caption{Fluxes of H and He. The model parameters are the same of the last section but the injection slope of H is 4.37 and of He is 4.31. As in the other plots, the dashed line shows the predicted flux if the secondary production were neglected.}
\label{fig:HeH}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{crslopes.pdf}
\caption{Slope of the spectra of nuclei with different mass as a function of rigidity.}
\label{fig:slopes}
\end{figure}
As discussed in past literature, the situation of H and He is very peculiar and certainly puzzling. Here we confirm the conclusion of previous investigations ~\cite{Evoli2019prd} that AMS-02 data on H and He require different injection spectra. A similar conclusion was previously reached based on PAMELA data~\cite{PAMELA.2011.phe}. It turns out that H is best reproduced with a source spectrum with a slope which is $\approx 0.05$ softer than other nuclei, similar to what was found in~\cite{Evoli2019prd}, while He requires a source spectrum with a slope which is $\approx 0.02$ harder than that of nuclei.
At present no real explanation of this trend exists and certainly this result was not expected. Even models which suggest a different injection slope for the two cannot explain the difference in the injection spectra between He and intermediate mass nuclei, such as C and O~\cite{Malkov2012prl,Hanusch2019He}.
Notice that the remarkable similarity among the measured spectra of He, C and O nuclei (see Fig.~78 of~\cite{AMS02.2021.review}) is probably the most effective way to motivate the need for a different source spectrum, since the spallation of these three elements is quite different and yet the measured spectra are very similar.
Our best-fit to the H and He spectra is shown in Fig.~\ref{fig:HeH}, where, as usual, the shaded regions provide an estimate of the effect of uncertainties in the cross sections. The dashed line shows the primary contribution alone to the He flux, and its distance from the solid line shows the secondary contribution (take into account that the flux plotted in the figure is the sum of $^4\rm He$ and $^3\rm He$, and the latter is also produced in spallation reactions of $^4\rm He$).
\section{Conclusions}
\label{sec:conclusions}
The AMS-02 data provide us with the tools for what could be the most substantial step forward in our understanding of the origin of CRs. In the present article we discussed in detail the implications of these data in terms of CR transport in the Galaxy.
The general framework that emerges from our calculations is one in which three slopes are required to describe the source spectra for elements of different masses: a steep injection for protons, one sightly harder spectrum for nuclei heavier than He and one even harder for He itself. The difference between the slopes is at the level of few percents. This might seem a small difference, but the high precision data now available force us to provide a description of physical processes at this level. Taken at face value, this difference, however small, imposes a big strain on our models of the origin of CRs: all the acceleration processes that we have devised through the years are rigidity dependent, namely nuclei of the same rigidity should experience the same acceleration and manifest the same spectrum. The situation is somewhat more complex, in that the spectrum of particles released into the ISM is not the same as at the acceleration site (see for instance~\cite{Caprioli2009mnras,Caprioli2010aph,Cristofari2020aph}). But it remains true that nuclei of different mass but the same rigidity should escape the acceleration region in the same way.
It has also been speculated that the integration over time during a SNR evolution might give rise to a difference in the spectra of H and He~\cite{Malkov2012prl,Hanusch2019He}. However, aside from the details of the model, it predicts that He and heavier nuclei should share the same spectrum.
It may also be speculated that different types of sources may release particles with slightly different spectra and different relative abundances (for instance some sources may be He-rich), but such solutions all appear very fuzzy at the present time, and it should be admitted that no clear understanding of the origin of the different source spectra for H, He and heavier nuclei is yet available.
Once it is accepted that three injection slopes are necessary to describe observations, we have been able to show that the spectra of nuclei are all properly accounted for, with the possible exception of iron, that will be discussed later. In order to reach such a conclusion it is of the utmost importance that all chains of spallation reactions and radioactive decays are taken into account~\cite{Webber1998apj,Silberberg1998apj,Mashnik2004adspr,Moskalenko2001icrc,Moskalenko2003icrc,Moskalenko2003apj,Tomassetti2017prd,Genolini2018prc}. The cross sections for these processes, or rather the poorness with which such cross sections are currently known, represent the chief limitation to our capability to extract physical information from available data. For some nuclei, for instance Mg, the comparison between observed and predicted spectra would benefit from slight increases in the production cross sections, compatible with known uncertainties.
The spectra that result from propagation also have a trend in terms of observed spectral shape, as illustrated in Fig.~\ref{fig:slopes}: 1) for increasingly more massive nuclei, there is a trend towards globally harder spectra, as a result of the enhanced role of spallation. This is particularly evident for nuclei which are mainly of primary origin (O, Si and Fe), while deviations from this trend can be noticed for nuclei which receive a substantial secondary contribution. 2) At low energies, where spallation reactions play a more important role the slope decreases for heavier nuclei. 3) For rigidities $R\gg 10^3$ GV all nuclei heavier than He tend to have the same spectral slope, since at such rigidities CR transport is dominated by diffusion, which behaves in the same way for all nuclei at given rigidity. Small differences around these trends can be easily identified (and have been discussed above) for nuclei that are substantially contaminated by the contribution of spallation from heavier nuclei (for instance this is the case of N~\cite{AMS02.2018.n}). For such nuclei, deviations are visible at low rigidities where the secondary contribution becomes important.
The global picture of CR transport that emerges from our calculations is that, aside from energy losses, the phenomenon is well described by diffusion and advection. A few comments on both ingredients may be useful: the character of particle diffusion is fully reconstructed using the proton spectrum and secondary/primary ratios, in addition to the Be/B ratio that provides an estimate of the halo size~\cite{Evoli2019prd,Weinrich2020aab}.
Both these pieces of observations suggest that the spectral breaks observed in spectra at $\sim 300$~GV are due to a corresponding break in the diffusion coefficient. Such a break was proposed to originate from a transition from self-generated to pre-existing turbulence in the Galaxy~\cite{Blasi2012prl,Aloisio2013jcap,Aloisio2015aap,Evoli2018prl} (see also \cite{Blasi2019galax,Amato2018adspr} for recent reviews) or from a non trivial spatial dependence of the diffusion coefficient \cite{Tomassetti2012apj}. At rigidity $R\lesssim 10$ GV, advection with Alfv\'en waves becomes important. Although in this article we do not include self-generation explicitly, it was previously shown~\cite{Blasi2012prl} that such terms cannot be eliminated in theories where CRs are allowed to generate their own scattering centers. Such models are also incompatible with having second order Fermi acceleration at low energies, since waves move preferentially away from the galactic disc.
As mentioned above, the only exception to this otherwise satisfactory picture of CR transport is represented by the spectrum of iron nuclei measured by AMS-02~\cite{AMS02.2021.iron,Boschini2021arxiv} at $R\lesssim 30$ GV. In this region the discrepancy between our predictions and the measured spectra are at the level of $\sim 20-40\%$, much larger than the quoted systematic error in the same energy region. The discrepancy is emphasized even better in the Fe/O ratio, where AMS-02 data also seem to be at odds with HEAO data at the same kinetic energy and with ACE-CRIS, if extrapolated at lower energies. We explored several possible avenues to reduce the discrepancy between our predictions and AMS-02 data but none of them came close to be either a resolution of the problem or exemplary in plausibility.
To add to this puzzling situation, the predictions based on our calculations smoothly connect the high energy Fe/O ratio to the ACE-CRIS data. Moreover, when the ratio Fe/O is calculated on unmodulated fluxes, as it would be appropriate for Voyager data, our predicted ratio is in impressive agreement with Voyager data points.
At this point it is hard to draw any firm conclusion on the origin of this discrepancy, but there is a serious possibility that the problem with the iron spectrum may be of experimental nature and that only a detailed comparison between different techniques and different analyses can unveil the source of the disagreement. It is probably useful to stress that other discrepancies would deserve the same attention, such as the marked difference between the C and O spectra of AMS-02 compared with those of PAMELA~\cite{PAMELA.2014.bc} and CALET~\cite{CALET.2020.co}. At low energies the effects of nuclear fragmentation inside the experiment itself are rather large and for heavy elements, such as iron, this effect requires a careful correction. In this sense, it would be interesting and instructive to see a detailed description of the analysis of these data for different experiments, perhaps with different models for the relevant cross sections.
Several CR detectors, based on various detection techniques, are currently in operation and are expected to release the spectrum of O and Fe nuclei in the near future. These data will serve as additional tests of the nature of the discrepancy.
\section*{Acknowledgements}
We are grateful to the anonymous referee for several very interesting comments that helped us improve the quality of the manuscript. We are grateful to R. Aloisio, M. Boezio, I. De Mitri and R. Munini for useful conversations. The research activity of PB was partially funded through Grant ASI/INAF n. 2017-14-H.O. We acknowledge the use of the CRDB~\cite{Maurin2020crdb} and ASI~\cite{DiFelice2017asi} databases of CR measurements.
\input{output.bbl}
\end{document}
|
\section{Introduction}
The human population gets along with different microorganisms. Some of them lead to transmitted diseases that result in epidemics, even pandemics such as SARS-Cov-2. It is very important to define reliable measures to characterize the spread of those pathogens, both at the beginning and in the course of epidemics.
The reproduction number $\mathcal{R}(t)$ indicates the number of new infections that result from a single infected individual at any time $t$. When it is greater than 1, each infected individual tends to generate more than one infected individuals resulting in the occurrence of an epidemic outbreak. The reproduction number is usually known from its initial value, when all of the population is susceptible, known as the basic reproduction number $\mathcal{R}_0$. Due to the recovery of infected individuals, a portion of the population becomes immune to the disease, thus, herd immunity begins to emerge in the population. This acts as a barrier to the transmission of the disease, so $\mathcal{R}(t)$ takes into account the evolution of the number of susceptible individuals in the population and is usually referred as the effective reproduction number. It provides us with information on how disease transmission occurs during the outbreak it is an important metric to make healthcare managers aware of the current level of the disease propagation.
By considering heterogeneities in the homogeneous compartmental models, the reproduction number is affected. Indeed different classes of infected individuals can lead to different contributions for the generation of new infections. A lot of studies have taken the heterogeneity into account when dealing with the basic reproduction number $\mathcal{R}_0$ \cite{Hyman,Brauer}, however, this is not the same for the effective reproduction number, $\mathcal{R}(t)$.
To study reproduction number, the renewal equation, a Lotka-Euler type equation \cite{Wallinga}, is usually used. This equation emerges from the mathematical models dynamics and can be derived using infection-age models. This type of modelling comes from the study of age-structured population growth and was first used within the scope of mathematical demography with the aim of modeling the birth dynamics due to the offspring produced by an individual over its lifespan. Analogously, the same methodology can be applied to the infectious diseases scenario by modelling the generation of new infected individuals during the infection period of those previously infected \cite{Martcheva}. Infection-age models are the origins of the widely known SIR and SEIR models developed by Kermack-Mckendrick \cite{Kermack}. Most known compartment ODE models for infectious diseases have their infection-age identical counterparts. The advantage with respect to the infection-age models is that they allow us to have access to the distribution of the infected individuals during the infectious period.
The estimation of the reproduction number in the course of an outbreak usually rely on the daily count of new cases along with the renewal equation. This approach uses the generation interval distribution, also called the serial interval or generation time distribution, and was popularized by \cite{Wallinga}. Several works indicate how to obtain this distribution through empirical data\cite{Park,Huber,Lessler,Nishiura 2020} or from other known distributions and models \cite{Nishiura,Wallinga,Champredon,Akkouchi,Champredon2}. However, these distributions rely on epidemiological and empirical studies and several systems are very difficult to analyse or even do not have the required data for the envisaged evaluation. Moreover, highly heterogeneous systems do not have a general methodology for estimating the reproduction number and its generation interval distribution \cite{Nishiura 2009,Park 2019, Wesley, Fraser}.
In this study, we use mathematical models to bridge the gap between the reported cases and the reproduction number. Therefore, we present a methodology that derives reproduction numbers and the generation interval distribution for an arbitrary compartment mathematical model. With such methodology, we can investigate highly heterogeneous systems using appropriate models and evaluate their reproduction numbers. With that in mind, apply the method to a meta-population model to analyse the role of spatial heterogeneity in the spreading of COVID-19 in the Brazilian territory.
Our work is separated as follow: in section 2, we introduce heterogeneity by allowing the existence of different groups of individuals in a population that can be sorted into compartments \cite{Hethcote}, considering then a very general heterogeneous model that can be reduced to most models in literature. We use this general formulation to develop our methodology to obtain the reproduction numbers and expressions that correlate it to actual data. Then, in section 3, we apply it to a specific scenario related to the role of the inter-municipal commuter flow of people and extract its reproduction number and the generation interval distribution. In section 4 and 5 we use actual data of the emerging SARS-Cov-2 coronavirus pandemic in the municipalities of the state of Rio de Janeiro, Brazil, to estimate the reproduction numbers that emerge from the system. Additionally, we reconstruct the time series of cases from the contributions of each municipality in the propagation of the disease.
\section{Reproduction number in a heterogeneous population}
\subsection{A general infection-age model}
In a heterogeneous population, individuals can be discernible by age-groups, spatial locations, behaviour, different susceptibility to diseases or any other factor that may distinguish them for one another.
In this section, we consider a general heterogeneous model that separates the individuals in $m$ homogeneous compartments. For the purpose of evaluating the reproduction number, we first consider a sub-set of variables encompassing the infected compartments, i.e., those compartments containing individuals that carry the etiologic agent in their organisms. These compartments are denoted by $\bm{x}(t,\tau)=(x_1(t,\tau) , ...,x_n(t,\tau))$, where $t$ and $\tau$ indicate, respectively, the calendar time and the infection-age, the time elapsed since they got the infection. Of course the condition $\bm{x}(t,\tau)=0$ for $\tau >t$ must be satisfied. If $m=m_i+m_0$, where $m_i$ and $m_0$ indicate the number of infected and non-infected compartments in the homogeneous model, $n=m_i$. Therefore, $\bm{x}$ represents the number of individuals in each compartment at a specific calendar time ``$t$'' and at infection age ``$\tau$''. $\bm{x}$ entries can be called the infection-age distributions. The total number of individuals in each compartment, denoted as $X_i(t)$, can be obtained by integrating $x_i(t,\tau)$ with respect of $\tau$ from zero to infinity, where $\bm{X}(t)=(X_1(t) , ...,X_n(t))$. For simplicity, we will be using matrix notation in some equations. The matrices and vectors will be identified by using bold font, whereby $\bm{M}(y,z)= \Big [M_{ij}(y,z) \Big]$. Also, as usual, integrals and derivatives with respect of scalar variables operate componentwise.
Drawing a parallel with Van den Driessche's next generation method \cite{van den Driessche}, we distinguish the new infections from the other compartment flows, defining $\bm{\mathcal{F}}(t)=(\mathcal{F}(t) , ...,\mathcal{F}_n(t))$ as the rate of appearance of new infections and $\bm{\mathcal{V}}(t,\tau)=(\mathcal{V}_1(t,\tau) , ...,\mathcal{V}_n(t,\tau))$ as the rate of transfer into compartments. Therefore, $\bm{\mathcal{F}}(t)$ describes the flow from non-infected compartments into infected ones and depends on $\bm{X}(t)$. On the other hand,$\bm{\mathcal{V}}(t,\tau)$ is related to the flow between infected compartments or from infected to non-infected ones and must depend on $\bm{x}(t,\tau)$. Thus, an usual infection-age model can be written as:
\begin{align}
\Big( \frac{\partial}{\partial t} +\frac{\partial}{\partial \tau}\Big)\bm{x}(t,\tau)=& -\bm{\mathcal{V}}(t,\tau),\label{Infec_V}\\
\bm{x}(t,\tau=0)=& \bm{\mathcal{F}}(t). \label{Infec_F}
\end{align}
Infection-age models are partial differential equations (PDE) that takes into account the time elapsed since the infection $\tau$. The equations \eqref{Infec_V} and \eqref{Infec_F} can usually be linked to their respective ODE models, and in fact the Kermack-Mckendrick SIR and SEIR models \cite{Kermack} are the special cases of their infection-age correspondents. This modeling allows us to have access to the distribution of the infected during the infectious period, which enable us to estimate the number of active infected individuals and the reproduction number along the outbreak using the reported data. Integrating \eqref{Infec_V} from zero to infinity of with respect of $\tau$ we obtain:
\begin{equation}\label{correspondece}
\frac{d}{dt} \bm{X}(t) = \bm{\mathcal{F}}(t) - \int_0 ^{\infty} \bm{\mathcal{V}}(t,\tau) d\tau,
\end{equation}
\noindent which is the correspondent to the sub-set of equations for the infected variables of the ODE model. Therefore, even though we are dealing with an infection-age model, one can simply build $\bm{\mathcal{F}}$ and $\bm{\mathcal{V}}$ from an ODE system and apply the methodology in this work. We demonstrate this for usual ODE models in the Supplementary Material 1.
Since $\mathcal{F}_i(t)$ and $\mathcal{V}_i(t,\tau)$ are well defined, we proceed to solving equation \eqref{Infec_V} using the method of integration along the characteristic line. Thus, the left side of the equation \eqref{Infec_V} indicates that the characteristics of those PDE's are lines of slope 1, which implies that $t=\tau+c$ with $c$ being an arbitrary constant. We fix a point ($t_0$,$\tau_0$) and introduce a variable $\omega$ such that $u_i(s)=x_i(t_0+\omega,\tau_0+\omega)$ are functions that provide the values of the compartment densities along the characteristic. After straightforward calculations \cite{Martcheva}, we obtain
\begin{equation}\label{solve_V}
\frac{d}{d\omega} \bm{u}(\omega) = -\bm{\overline{\mathcal{V}}}(\omega),
\end{equation}
\noindent where $\overline{\mathcal{V}}_i(\omega)= \mathcal{V}_i(t_0 +\omega, \tau_0 +\omega)$. In epidemic models, $\Bar{\mathcal{V}}_i(\omega)$ are usually linear equations, making this system easy to solve. In that case, the system can be represented as:
\begin{equation}\label{solve_V_x}
\frac{d}{d\omega} \bm{u} = -\frac{\partial \; \bm{\overline{\mathcal{V}}}}{\partial \bm{u}} \; \bm{u},
\end{equation}
\noindent where $-\frac{\partial\bm{\overline{\mathcal{V}}}}{\partial \bm{u}}= \Big[ -\frac{\partial}{\partial u_j} \overline{\mathcal{V}}_i(\omega) \Big]$ is the matrix of the linear system. Assuming that there are no infected individuals prior to $t=0$, we only need to take into account the solution where $t>\tau$, leading to $\omega=\tau$, $t=\tau +t_0$ and $\tau_0 = 0$. The solution for a linear system can be written as
\begin{equation}\label{Solution_u}
\bm{u}(\omega) = \bm{\overline{\Gamma}}(\omega) \; \bm{u}(0) ,
\end{equation}
\noindent where $\bm{\overline{\Gamma}}(\omega)=\bm{\Gamma}(t,\tau)$, is the fundamental matrix obtained by solving \eqref{solve_V}. Therefore, we identify $\bm{\mathcal{F}}(t_0)= \bm{u}(0)$ in \eqref{Infec_F}. So, \eqref{Solution_u} becomes
\begin{equation}\label{Solution_V}
\bm{x}(t,\tau) = \bm{\Gamma}(t,\tau) \; \bm{\mathcal{F}}(t-\tau).
\end{equation}
For a linear $\bm{\mathcal{V}}(t,\tau)$, $\bm{\Gamma}(\tau)$ components are exponential functions. Equation \eqref{Solution_V} can also be used to express the solution of a general non-linear $\bm{\mathcal{V}}(t,\tau)$, but if it is not the case, such system will not be considered in this work. On the other hand, we can also assume that $\mathcal{F}_i(t)$ can be written as
\begin{equation}\label{F_het}
\mathcal{F}_i(t)= \sum_{j}^{n} \int_{0}^{\infty} \Omega_{ij}(t,\tau) \; x_j(t,\tau) d\tau,
\end{equation}
\noindent in which $\bm{\Omega}(t,\tau)$ is related to the generation of infected individuals in "$i$" due to "$j$". If $\bm{\Omega}$ does not depend on $\tau$, as in ODE models, it assumes the form of
\begin{equation}\label{Omega_deriv}
\bm{\Omega}(t)= \Big[ \frac{\partial}{\partial x_j} \mathcal{F}_i(t) \Big]
\end{equation}
Most system in literature satisfies \eqref{Solution_V} and \eqref{F_het}, thus the method in this work is very general and can be applied to a wide range of models.
\subsection{Obtaining the reproduction numbers}
To estimate the reproduction number using the reported data of new infections, we need to link them with the equations of the model. This can be done by substituting \eqref{Solution_V} in \eqref{F_het}. Arranging the equation we get
\begin{equation}\label{renewal_het}
\bm{\mathcal{F}}(t)=\int_{0}^{\infty} \bm{A}(t,\tau) \bm{\mathcal{F}}(t-\tau) d\tau,
\end{equation}
\noindent for
\begin{equation}\label{A_het}
\bm{A}(t,\tau)= \bm{\Omega}(t,\tau) \bm{\Gamma}(t,\tau).
\end{equation}
The functions $A_{ij}(t,\tau)$, analogously to \cite{Nishiura}, are referred as the rate of new infections in $X_i$ due to $X_j$ previous infections at a calendar time $t$ and infection-age $\tau$, whereby $A_{ij}(t,\tau>t)\equiv 0$. So, we can describe the new cases in "$X_i$" from the cases that occurred previously in all the groups. In fact, \eqref{renewal_het} is a general form for the widely known renewal equation, \cite{Nishiura, Wallinga}, actually a sum of renewal equations. Thus, by defining the number of new infections in the "$i$" compartment due to "$j$" previews infections as
\begin{equation}\label{renewal_het_ij}
\mathcal{J}_{ij}(t)= \int_{0}^{\infty} A_{ij}(t,\tau) \mathcal{F}_j(t-\tau) d\tau,
\end{equation}
\noindent it becomes clear that $\mathcal{F}_i(t)=\sum_j^{n} \mathcal{J}_{ij}(t)$. Thus, it is possible to quantify the influence of the infections occurred in one compartment new ones in another compartment. To obtain the expected number of new infections in $X_i$ a newly infected individual form $X_j$ is expected to generate, thus the reproduction number, we integrate $A_{ij}(t,\tau)$ from zero to infinity with respect to $\tau$, as in \cite{Nishiura}.
\begin{equation}\label{Rt_ij}
\mathcal{R}_{ij}(t)= \int_0^\infty A_{ij}(t,\tau) d\tau.
\end{equation}
$\bm{\mathcal{R}}$ is usually called the next-generation matrix, where its entries corresponds to the reproduction numbers of the system \cite{Nishiura 2009}. It is remarkable that, for a ODE model where $\bm{\mathcal{V}}$ is linear, the next-generation matrix evaluated at the free-disease fixed point $\bm{\mathcal{R}|}_{x_0}$, is equivalent to the one developed in \cite{van den Driessche}. Naturally from \eqref{Rt_ij}, we define
\begin{equation}\label{g_ij}
g_{ij}(t,\tau) = \frac{A_{ij}(t,\tau)}{\int_0^\infty A_{ij}(t,\tau) d\tau},
\end{equation}
\noindent where $g_{ij}(t,\tau)$ is normalized and denoted as the generation interval distribution \cite{Nishiura,Wallinga}. Thus, it is related to the flow of individuals between infected compartments and their recover process. We assume a generation interval distribution depending on $t$ and $\tau$, which is general enough to be applied to models whose dynamics can change over time. Therefore, using \eqref{Rt_ij} and \eqref{g_ij} in \eqref{renewal_het_ij} we obtain:
\begin{equation}\label{J_ij_R}
\mathcal{J}_{ij}(t)= \mathcal{R}_{ij}(t) \int_0^\infty g_{ij}(t,\tau) \;\mathcal{F}_j(t-\tau) d\tau.
\end{equation}
It is important to highlight that $ \mathcal {R} _ {ij} (t) $ is not necessarily the number of new infected in $ X_i $ generated by $ X_j $. Instead, the meaning of $ \mathcal {R} _ {ij} (t) $ is linked to the generation of new infected ones in ``$i$'', $ \mathcal {F} _i (t) $, due to infected individuals previously generated in ``$ j $'', $ \mathcal {F} _j (t- \tau) $, regardless of what stage of the disease these infected individuals who entered ``$ j $'' are at instant $ t $. That is the case because a new infected individual may generate new infections through out many stages of the disease. In the SEIR model, for example, all of the new infections are in the $E$ compartment, but the individual only becomes infective when it goes to the $I$. To access the reproduction number a newly infected in $X_j$ is expected to generate to all of the other compartments, we can just sum the $\mathcal{R}_{ij}$ over $i$, defining
\begin{equation}\label{R_bar_j}
\overline{\mathcal{R}}_j(t) = \sum_i ^n\mathcal{R}_{ij}(t).
\end{equation}
\noindent Through out this work, the over-line represents the merge of the first index , in this case, in the form of a sum. This lead us to analogously define $\overline{A}_j= \sum_i ^n A_{ij}$, whereby its integral from zero to infinity with respect to $\tau$ is $\overline{\mathcal{R}}_j$. Thus, we are able to write
\begin{equation}\label{J_bar}
\overline{\mathcal{J}}_{j}(t)= \overline{\mathcal{R}}_j(t) \int_0^\infty \overline{g}_{j}(t,\tau) \;\mathcal{F}_j(t-\tau) d\tau,
\end{equation}
\noindent where $\overline{\mathcal{J}}_{j}(t) = \sum_i ^{n} \mathcal{J}_{ij}(t)$ and
\begin{equation}\label{g_bar}
\overline{g}_{j}(t,\tau) = \frac{\overline{A}_{j}(t,\tau)}{\int_0^\infty \overline{A}_{j}(t,\tau) d\tau}= \frac{\sum_i ^n \mathcal{R}_{ij} (t) g_{ij}(t,\tau)}{\sum_i ^n \mathcal{R}_{ij} (t) }
\end{equation}
Therefore, $\overline{\mathcal{J}}_j(t)$ is the total number of new infections generated by previews infections in $X_j$. It is interesting that the generation interval distribution $\overline{g}_j(t,\tau)$ takes the form of a
weighted average in which the $g_{ij}(t,\tau)$ are the weights.
Thus, the implementation of the proposed method in this work amounts to: identifying the terms $\bm{\mathcal{F}}$ and $\bm{\mathcal{V}}$ from a model; using them to find $\bm{\Omega}$ and $\bm{\Gamma}$; obtaining $\bm{A}$ and integrating to get $\bm{\mathcal{R}}$ and $\bm{g}$. Further in this work we present applications of the method and estimations using actual data. Examples of the method in different types of models can be found on the Supplementary Material 1.
\subsection{The total reproduction number}
After obtaining the reproduction numbers of the constituents of the system, we now intend to define a reproduction number for the whole system. The number of new infections in a group, $\mathcal{F}_i(t)$, can be described as a fraction of the total number of cases from all groups, $\mathcal{F}^T(t)=\sum_i ^n \mathcal{F}_i(t)$ such that
\begin{equation}\label{F_prop}
\mathcal{F}_i(t)= \alpha_i(t) \mathcal{F}^T(t).
\end{equation}
\noindent Since $\alpha_i(t)$ is the proportion of the total number of cases that $\mathcal{F}_i(t)$ represents, the condition $1=\sum_i \alpha_i(t)$ must be satisfied. Thus, using \eqref{renewal_het} with \eqref{F_prop}, we obtain:
\begin{equation}\label{renewal_het_prop}
\mathcal{F}^T(t)= \mathcal{R}^T(t) \int_{0}^{\infty} g^T(t,\tau)\mathcal{F}^T(t-\tau) d\tau,
\end{equation}
\noindent where $\mathcal{R}^T(t)= \bm{\alpha \cdot \overline{\mathcal{R}}}$ is the reproduction number of the system and
\begin{equation}\label{g_prop}
g^T(t,\tau) = \frac{\sum_i \alpha_i(t) \overline{\mathcal{R}}_i(t) \; \overline{g}_i(\tau)}{\sum_i \alpha_i(t) \overline{\mathcal{R}}_i(t)} .
\end{equation}
Since $ \mathcal{R}^T (t) $ is the scalar product between $ \bm {\alpha} $ and $ \bm{\overline{\mathcal{R}}} $, we can interpret it as the projection of the $ \bm{\overline{\mathcal{R}}} $ over the fractions $ \bm {\alpha} $ forming a weighted average. Thus, it is interesting to note that we can describe the system as an average of its heterogeneities. Since the definition of the total number of reproductions $ \mathcal{R}^T $ is very general, it is not always meaningful. For example, if we form a system of two independent dynamics, that is, ($ \mathcal{R}_{ij} = 0 \text{ for } i \neq j $) it is still possible to obtain a $ \mathcal{R}^T $, even if there is no meaning to it. The $\alpha_i(t)$ functions are the key for analysing the feasibility of a total reproduction number that has a dynamical meaning.
We focus our attention on a case where the $\alpha_i(t)$ appears naturally in the equations. If $\Omega_{ij}(t,\tau)$ can be separated in a function of $t$ depending on the "$i$" index and another function of $t$ and $\tau$ depending on the "$j$" index, then it can be written as
\begin{equation}\label{Omega_prop}
\bm{\Omega} = \bm{\alpha} \otimes \overline{\bm{\Omega}}= \Big [\alpha_i(t) \overline{\Omega}_j(t,\tau) \Big]
\end{equation}
\noindent where $ \otimes $ represents a tensor product. We define $ \overline {\Omega}_j = \sum_i \Omega_ {ij} $, and noticed that $ \overline {A} _j = \sum_ {k} \overline {\Omega} _k \Gamma_ {kj} $ e $ \bm {A} = \bm {\alpha} \otimes \overline {\bm { A}} $. Thus, $ \bm {\Omega} $ and $ \bm {A} $ can be separated in proportions, $ \bm {\alpha} $ and $ \overline {\bm {\Omega}} $. In fact, the above equation also impacts \eqref{R_bar_j} and \eqref{J_bar} which can be factorized in terms of $ \bm {\mathcal {R}} = \bm {\alpha} \otimes \overline {\bm {\mathcal {R} }} $ e $ \bm {\mathcal {J}} = \bm {\alpha} \otimes \overline {\bm {\mathcal {J}}} $. Using \eqref{g_bar} we can also get $ g_ {ij} (t, \tau) = \overline {g} _j (t, \tau) $. Furthermore, because the next generation matrix is obtained from a tensor product of vectors, the largest eigenvalue of $ \bm {\mathcal {R}} $ corresponds to the scalar product of $ \bm {\overline {\mathcal {R }}} $ e $ \bm {\alpha} $, that is $ \mathcal {R} ^ T $. Thus, in these systems the total reproduction number assessed at the disease-free equilibrium point, $t=0$, corresponds to the basic reproduction number, $\mathcal {R} ^ T (0)= \mathcal{R}_0$. Also, in that case, the $\mathcal{R}^T$ is the spectral radius of $\bm{\mathcal{R}}$, as in \cite{van den Driessche}. This is very common in disease transmission models, in fact the SIR and SEIR model are examples of this case.
\subsection{Estimations with real data }
So far, we've developed a general framework to estimate the reproduction numbers from the rate of new infections $\mathcal{F}$. However, when we deal with real data what we have is the collection of all of the new infections in an $\Delta t$ period of time, which leads to the definition of $\bm{\mathcal{B}}$ and $\bm{\mathcal{T}}$.
\begin{equation}\label{reported_cases}
\mathcal{B}_i(t) = \rho_i \int_{t} ^{t+\Delta t} \mathcal{F}_i(t') dt', \qquad \mathcal{T}_{ij}(t) = \rho_i \int_{t} ^{t+\Delta t} \mathcal{J}_{ij}(t') dt'.
\end{equation}
Therefore $\mathcal{B}_i(t)$ are the reported cases in $X_i$, where $\rho_i$ is a constant related to a higher or lower notification, due to sub/super-notification. Analogously we define $\mathcal{T}_{ij}(t)$ as the number of reported cases in $X_i$ due to previous cases in $X_j$. By assuming that $\mathcal{R}_{ij}$, $g_{ij}(t,\tau)$ and $\mathcal{F}_i $ are approximately constants during a $\Delta t$ interval, we use \eqref{J_ij_R} and \eqref{F_het} to derive
\begin{equation}\label{T}
\mathcal{T}_{ij}(t) = \mathcal{R}_{ij} (t) \sum_{\tau=0}^{t} \frac{\rho_i}{\rho_j} \; g_{ij} (t,\tau) \mathcal{B}_j(t-\tau) \Delta t
\end{equation}
\noindent for $\mathcal{B}_i(t)=\sum_j^n\mathcal{T}_{ij}$. \eqref{T} is a general form for the discrete version of the renewal equation. In fact, by using \eqref{renewal_het_prop} we are able to recover a well known result in literature \cite{Fraser}
\begin{equation}
\mathcal{R}^T(t)= \frac{\mathcal{B}^T(t)}{\sum_{\tau=0}^{t} \frac{\rho_i}{\rho_j} \; g^T (t,\tau) \mathcal{B}^T(t-\tau)\Delta t },
\end{equation}
\noindent where $\mathcal{B}^T(t) = \sum_i ^n \mathcal{B}_i(t)$.
\section{Explicit expression of $\mathcal{R}(t)$ for two meta-population models}
In this section we analyse two meta-population models detailed in Supplementary Material 2. The methodology developed on this study is applied to both models to obtain the correspondent reproduction numbers and generation interval distributions, see Supplementary Material 1 for detailed calculations.
\subsection{SIR-type meta-population model}
In this section we will consider a meta-population model takes into account groups of spatially separated "island" populations with some level of interaction. Such models are widely used for systems in which the movement of individuals between meta-populations is considered \cite{van den Driessche_spatial,Bichara,Lloyd,Lajmanovich,Arino,Keeling,Miranda}. Since each population is connected with the others, the system can be interpreted as a network for which the nodes represent the meta-populations and the weight of their edges represent the intensity of the movement between them. This type of model does not describe the daily movement of individuals explicitly, but as an interaction of the meta-populations. It is suitable when the population sizes are not permanently affected by the flow of individuals, as in the case of commuter movement between locations of residence, work and study. This type of movement is obligatory cyclical, predictable and recurring regularly, most of the time on a daily basis. In this model, we assumed that each meta-population $i$, with $N_i$ individuals, has its own transmission rate $\beta_i(t)$. Also, the movement of individuals between meta-populations is taken into account, where we introduce $\Phi_{ij}(t)$ as the density of flow between the populations $i$ and $j$. That is, the amount of $i$ resident individuals commuting from $i$ to $j$ divided by the total population of $i$, $N_i$. The parameters $\beta_i(t)$ and $\Phi_{ij}(t)$ are time dependent, so we can incorporate the changes in the behavior of the populations on those variables.In Supplementary Material 2 we derive this model, inspired by \cite{Miranda}, and in Supplementary Material 1 we obtain the corresponding reproduction numbers and generation interval distribution for it, which reads:
\begin{equation}
R_{ij}(t) = S_i(t) \frac{\lambda_{ij}(t)}{\gamma}, \qquad g_{ij}(\tau)= g(\tau)= \gamma \; e^{-\gamma \, \tau}.
\end{equation}
\noindent Here $\lambda_{ij}$ is related to the transmission of the disease from a meta-population ``$i$'' to other meta-population ``$j$'' and is derived based on simple assumptions about the commuter movement of individuals in the network, see Supplementary Material 2. Notably, if we isolate the meta-populations from the network, $\Phi_{ij}(t)=0$, for all $i$ and $j$, then the reproduction numbers and the generation interval distributions becomes identical to the classical SIR model \cite{Nishiura}.
\subsection{A meta-population model for Covid-19 (SEIIR)}
The classical SIR model is a very simple and qualitative approach to disease transmission dynamics, but it does not provide the best description for most diseases. For now on, we will be focusing on a model for a specific disease, the SARS-Cov-2 coronavirus. In this case, the transmission can be facilitated by the existence of individuals whose symptoms are very weak or even nonexistent \cite{Li_R}, therefore, this heterogeneity can change the dynamics. In order to have consistence description of this aspect, it is wise to assume that the existence of two classes of infected individuals, the symptomatic and the asymptomatic/undetected ones, as considered in a general model for the SARS-Cov-2 coronavirus \cite{Oliveira}. Therefore, we now take into account infected individuals who do not need to be hospitalized and are not recorded in official registered data, thus becoming undetectable. For the sake of simplicity, we will refer to those individuals only as asymptomatic ones, for simplicity. In Supplementary Material 2 we proceed to derive this model, based in the meta-population SIR-type approach described in the previous section. In Supplementary Material 1 the expressions for the reproduction number and generation interval distribution are detailed derived. There, we show that we only need $\bm{\mathcal{R}}$ for $i,j\leq n$ to describe the dynamics. Thus, in this main framework, whenever we refer to $\bm{\mathcal{R}}$ or $\bm{g}$ we will be alluding to their $i,j\leq n$ elements. Thus, we obtain:
\begin{equation}
R_{ij}(t) = S_i(t)\lambda_{ij}(t) \Big[ \frac{p}{\gamma_s} + \frac{\delta (1-p)}{\gamma_a}\Big], \qquad g_{ij}(\tau)= g(\tau)= \frac{\frac{p}{\gamma_s} g^s(\tau) + \frac{\delta (1-p)}{\gamma_a}g^a(\tau)}{\frac{p}{\gamma_s} + \frac{\delta (1-p)}{\gamma_a}}.
\end{equation}
\noindent the $\lambda_{ij}$'s expressions are the same presented for the SIR-type model in Supplementary Material 2. $\delta$ is a factor that reduces or enhances the asymptomatic infectivity, $p$ is the proportion of individuals that becomes symptomatic when infected, $\gamma_a$ and $\gamma_s$ are the recover rates of the asymptomatic and symptomatic individuals, respectively. $g^a(\tau)$ and $g^s(\tau)$ are expressed in terms of exponential functions described according to \eqref{eqgas}.
\begin{equation}\label{eqgas}
g^a(\tau) = \frac{\kappa \gamma_a}{\gamma_a - \kappa} (e^{-\kappa \tau} - e^{-\gamma_a \tau} ), \qquad \qquad \qquad g^s(\tau) = \frac{\kappa \gamma_s}{\gamma_s - \kappa} (e^{-\kappa \tau} - e^{-\gamma_s \tau} ) .
\end{equation}
\noindent Whereby, if we isolate the meta-populations from the network, $\Phi_{ij}(t)=0$, for all $i$ and $j$, the reproduction numbers and generation interval distributions returns to the expression obtained in \cite{Oliveira}.
\section{Applications for the meta-population models using actual data}
In this section, we present results for the methodology applied to the meta-population model developed in the previews section. We use actual data of the first six months of the COVID-19 pandemic in Brazilian cities, such as: reported cases in each municipality, daily commuter movement due to work between municipalities and daily mobility tends towards workplaces. In Supplementary Material 3 we derive and present the expressions and parameters needed to estimate the reproduction numbers for both meta-population models. Thus, we obtain a daily time series of the reproduction numbers for each model.
\subsection{Database}
In this work, we use daily notifications of new cases due to COVID-19 in Brazil, obtained from public websites: \url{https://covid.saude.gov.br/} and \url{https://brasil.io/datasets/} , which provide data from the Health Ministry. We obtained intermunicipal commuter movement of workers and students data from a study on population arrangements and urban concentrations in Brazil conducted by IBGE (Brasilian Institute of Geography and Statistics) in 2015 that can be found at \cite{ibge}. In addition, we obtain daily mobility data for each Brazilian state from a public report by Google, accessed at: \url{https://google.com/covid19/mobility/}.
We used the data observed until September 14th, 2020, and performed a 10-day moving average, in order to attenuate noise and better express the data trend. To take the social distancing restrictions into account, we considered only the commuter movement data related to work, since, with the mitigation measures of COVID-19, the flow due to education were significantly reduced. In addition, because the movement towards work also dropped due to social isolation policies, we used the mobility data obtained from the community mobility report provided by Google to estimate this reduction. This database compares for each state the daily mobility to workplaces with the past trends, therefore, we can access the percentage of reduction in commuting to work. A moving average is performed in the mobility time series and we reduce the intermunicipal work flow according to the percentage indicated on the data. Thus, we obtain the flow of individuals due to commuter work that leads to $\Phi_{ij}(t)$. The parameters used to feed the model were obtained in \cite {Jorge} and are displayed in Supplementary Material 3.
The data from the state of Rio de Janeiro was selected for this analysis. Ten cities with the highest commuter flow with the capital Rio de Janeiro (RJ), were chosen, namely: Duque de Caxias (DdC), Nova Iguaçu (NI), São João de Meriti (SJdM), Niterói (Nt), São Gonçalo (SG), Belford Roxo (BR), Nilópolis (Ns), Mesquita (Mq), Queimados (Q), Magé (Ma). All of the chosen cities for this work are part of the Rio de Janeiro's metropolitan region. Additional information about the municipalities is presented in Supplementary Material 4.
\subsection{Analyses of the results}
In our first results, shown in Figure \ref{fig:1}, we present a comparison between the SIR and SEIIR outputs. Using the daily time series of the reproduction numbers, see Supplementary Material 3, we obtain the series of $\overline{\bm{\mathcal{R}}}(t)$ and $\bm{\mathcal{T}}(t)$ from \eqref{R_bar_j} and \eqref{T}, respectively. We observe that the $\overline{\bm{\mathcal{R}}}$ of SEIIR model is, in the average, $33\%$ higher then the corresponding results for the SIR model. On the other hand, the estimations of the total number of exported cases that are reported, $ \sum _{t}\sum_{i} ^n \mathcal{T}_{ij}(t)$ for $i \neq j$, is very similar for both SIR and SEIIR models Also, it seems that the total commuter movement, which is the sum of all the inflow and outflow happening in a municipality, is not the only main factor that determine the number of exported cases of a municipality. This non-linearity can be observed when comparing São João de Meriti (SJdM) and Niterói (Nt) or Duque de Caxias (DdC) and Nova Iguaçu (NI) (see Figure \ref{fig:1}b). Those municipalities have a similar amount of total flow but very different results for the exported cases. Interestingly, even not having the highest $\overline{\mathcal{R}}_j$, the capital, Rio de Janeiro (RJ) presents the largest amount of exported cases, which also showcases the non-linear dynamics of the phenomenon.
\begin{figure}
\centering
\includegraphics[width={1.\linewidth}]{Figures/Compare.pdf}
\caption{\textbf{Comparison between SIR and SEIIR outputs.} SIR results in blue and SEIIR ones in orange. The bar graph (a) compares the SIR and SEIIR $\overline{\mathcal{R}}_i$ averages in time for all municipalities. In (b), for each municipality, estimations for the total number of exported cases that are reported for both models are displaced with the total commuter movement in that city. The names of the municipalities are abbreviated using acronyms: Rio de Janeiro (RJ), Duque de Caxias (DdC), Nova Iguaçu (NI), São João de Meriti (SJdM), Niterói (Nt), São Gonçalo (SG), Belford Roxo (BR), Nilópolis (Ns), Mesquita (Mq), Queimados (Q), Magé (Ma).}
\label{fig:1}
\end{figure}
From now on, we will only look at the results of the SEIIR model. With $\mathcal{T}_{ij}(t)$ we are able to access the contribution of each municipality on the outbreaks happening in the state. Thus, by dividing $\mathcal{T}_{ij}(t)$ by $\mathcal{B}_i(t)$ in every time step, we obtain a time series for the proportion of the total cases in ``$i$'' generated by ``$j$''. Therefore, we proceed into evaluating the mean value through time of $\mathcal{T}_{ij}/\mathcal{B}_i$ , where it is displaced in Figure \ref{fig:2}. It is observed a high autochthonous behavior on the disease transmission, whereby the highest influence of a city is on itself. Therefore, most of the cases generated in a municipality are caused by its own individuals. However, we also identify cities where the cases generated by other municipalities on it are very important. The $\bm{\mathcal{R}}(t)$ matrix also corroborates the presence of an important autochthonous behavior as its diagonal elements correspond to the highest values of the reproduction numbers. We also observed lots of very smalloff-diagonal elements low values through out the matrix.
\begin{figure}
\centering
\includegraphics[width={0.75\linewidth}]{Figures/Matrix_SEIIR.pdf}
\caption{\textbf{Influence on number of cases of a municipality in another.} The heatmap captures the average, through time, influence on number of cases of a municipality in another as a proportion of the total number of daily cases, $\mathcal{T}_{ij}/\mathcal{B}_i$. $i$ corresponds to the rows and $j$ to the columns. }
\label{fig:2}
\end{figure}
In Figure \ref{fig:3} we illustrate the results of Figure \ref{fig:2}, whereby only non-autochthonous influences above 5\% are considered. Thus, we must highlight the capital, Rio de Janeiro (RJ), as the most important agent on the disease transmission to the municipalities on the network. However, cities like Nova Iguaçu (NI) also presents itself as a relevant disseminator of the pathogen. As shown in Figure \ref{fig:1}, Nova Iguaçu NI is the highest city in case exportation, besides the capital. In Figure \ref{fig:3} we identify cities like: São João de Meriti (SJdM), Belford Roxo (BR), Nilópolis (Ns), Mesquita (Mq) and Queimados(Q) as the main receptors of those cases. Niterói (Nt), even having a NI-like number of exported cases, did not presented a high influence on a lot of cities. On the other hand, Niterói (Nt) generates a significant amount of cases in São Gonçalo (SG), highlighting the importance of the connection of those two municipalities.
\begin{figure}
\centering
\includegraphics[width={0.63\linewidth}]{Figures/Rplot.pdf}
\caption{\textbf{Visualization of the influence between municipalities.} Here, we present a visualization of the results from Figure \ref{fig:2}. Only non-autochthonous influences above 5\% were considered. The thickness of the lines connecting municipalities is proportional to the number of cases that one generates on the other. The color of each line represents the municipality that is generating the cases.}
\label{fig:3}
\end{figure}
\begin{figure}
\centering
\includegraphics[width={1.\linewidth}]{Figures/Time_Series.pdf}
\caption{\textbf{Reconstruction of the time series.} The black dots represents the daily reported cases in (a) São Gonçalo (SG), (b) São João de Meriti (SJdM). The blue dots are the amount of cases generated in (a) São Gonçalo (SG), (b) São João de Meriti (SJdM) due to the capital, Rio de Janeiro (RJ). In red we have the number of cases reported in (a) São Gonçalo (SG) due to Niterói (Nt); (b) São João de Meriti (SJdM) due to Nova Iguaçu (NI).}
\label{fig:4}
\end{figure}
We illustrate the time series reconstruction in Figure \ref{fig:4} where two scenarios are displaced. In the first one, we focus on the cases reported in São Gonçalo (SG) and compare total amount, $\mathcal{B}_i(t)$, with the number of daily cases in SG generated by Rio de Janeiro (RJ) and Niterói (Nt). We observe that Nt has a predominance over the capital, presenting a higher number of cases generated in all times. The second scenario is related to the total number of cases in São João de Meriti (SJdM) and the contributions due to Rio de Janeiro (RJ) and Nova Iguaçu (NI). In this case, the capital presents the largest number of cases generated in that city, besides the city itself. It is also interesting to observe in both scenarios how the cities contributions merge into a part of the total cases notified on a municipality.
\section{Discussion}
Understanding and dealing with infectious diseases is an on going challenge to humankind. The methodology presented in this work is very general and can be applied to multiple disease transmission systems. By merging the model dynamics with the available epidemic data, this method is able to estimate key epidemiological factors, like the reproduction numbers and the generation interval distributions. These theoretical results are the basis for a lot of possible data analyses, specially for highly heterogeneous systems that have been demanding for a suitable methodology for estimating the reproduction number through actual data. The method is robust and reproduces known results in literature, as shown in Supplementary Material 1 for the SIR, SEIR and SEIIR models. This methodology opens room for the analyses of more sophisticated models, that leads to a better understanding and control of infectious diseases.
With the emerging of the COVID-19 pandemic, due to SARS-Cov-2 coronavirus, in December 2019, scientists allover the world have joined forces to formulate control strategies \cite{Manica}. Factors like the presence of asymptomatic individuals, risk groups and spatial distribution of the disease brings out a system with multiples layers of heterogeneity \cite{Li_R, Eikenberry, Miranda, Gomes}.
Although the application of some vaccines starts recently, mitigating COVID-19 with non-pharmacological strategies is essential. Therefore, the second half of this work we focused on the application of the developed methodology for two models with spatial heterogeneity, one focusing on the COVID-19 specific dynamics. Each municipal demography, captured by $\beta$ parameters, and the commuter movement were combined by the model lead to some interesting results. The reproduction numbers are obtained for both models indicate that they only differ by multiplying factors, as $1/\gamma$ and $ \frac{p}{\gamma_s} + \frac{\delta (1-p)}{\gamma_a}$, as in \cite{Oliveira, Jorge}. The measured results shows that the asymptomatic presence is related to an increase on the reproduction number, because the model predicts more infected individuals than what is reported. Another substantial result is obtained when comparing the exported cases with the total commuter movement in a municipality. In this case, the relationship is not direct, since the intricate topology of the network fosters a non-linear dynamic in the phenomenon. This result points up the importance of a model to provide epidemiological meaning of the available data. For this specific case, it became evident that analysing the system only by the commuter movement data wouldn't be enough to point out the key cities on the disease transmission dynamic, as we did using the models and methodology.
The results of the model display the role of each municipality on the epidemic on the network. Cities like Nova Iguaçu, Niterói and São Gonçalo pops out as important agents on the spread of the pathogen through out the metropolitan region of Rio de Janeiro, Brazil.
The capital is highlighted as the main hub of spreading, which is related to its high incidence of the disease combined with its central role on the movement behavior of individuals on the network, as also observed in \cite{Jorge, Miranda, Liu}. However, cases like Niterói and São Gonçalo, where the interaction of both cities is higher than the capital's influence, cannot be neglected. This call's the attention for the relevance of analysis, like in this work, that provides the epidemiological interpretation of data. In this work, we choose to present an analysis with actual data in which the reproduction number is not the main result. We proceeded with this approach to portrait the reproduction number not as a dead end result, but as a tool for obtaining deeper analyses, such as the reconstruction of the time series and the number of exported cases.
Our results for the Rio de Janeiro metropolitan area, however, have limitations. The model developed in this work provides a very simple approximation of intercity commuter flow, while more sophisticated models available in the literature are required to provide more precise description of the movement behavior. Another limitation of this work is that it does not consider the interstate flow, since the state of Rio de Janeiro was portrayed as a closed system Also, we did not take into account further heterogeneous features, which could be accounted as well by the general theoretical formalism, besides space and asymptomatic presence, leaving a gap for other key COVID-19 dynamics characteristics, like age-groups. In addition, the examples presented in this work considered actual reported data of confirmed cases and mobility, which may present problems of underestimation and report delay. Taking those limitations into account, the results displaced here are still able to give substantial understanding of the system studied, which is a common feature in mathematical modelling. Finally, we reinforce the importance of the method proposed on this work and highlight its broad application on infectious disease models.
\enlargethispage{20pt}
\ethics{Since all data handled in this study is publicly available, an approval by an ethics committee is not required, according to Resolutions 466/2012 and 510/2016 (article 1, sections III and V) from the National Health Council (CNS), Brazil.}
\dataccess{All data and codes are gathered and presented in our public GitHub repository at \url{https://github.com/danielcpj/Rt-heterogeneous-models}}
\aucontribute{DCPJ, STRP and RFSA conceived of and designed the methodology presented in this study. DCPJ, STRP, JGVM and JFO formulated and interpreted the meta-population models and applications. DCPJ performed the data analysis and drafted the manuscript. All authors read, reviewed and approved the manuscript.}
\competing{The authors declare that they have no competing interests.}
\funding{DCPJ was funded by a Scientfiic Initiation scholarship from CNPq (process
number 117568/2019-8). JFO was supported by the Fiocruz Program of Promotion of Innovation - innovative ideas and products - COVID-19, orders and strategies - Inova Fiocruz (Processo VPPIS-005-FIO-20-2-40), and the Center of Data and Knowledge Integration for Health (CIDACS) through the Zika Platform - a long-term surveillance platform for Zika virus and microcephaly (Unified Health System (SUS), Brazilian Ministry of Health). STRP was supported by an International Cooperation grant (process number INT0002/2016) from Bahia
Research Foundation (FAPESB). RFSA was supported by Brazilian agency CNPq through Grants No. 422561/2018-5 and
304257/2019-2. STRP and RFSA were supported by the National Institute of Science and Technology - Complex Systems from CNPq, Brazil. JGVM acknowledges the support of the National Council of Technological and Scientific Development, CNPq, Brazil (Grant number: 307828/2018-2).}
\ack{The authors acknowledge the discussions and suggestions from members of the CoVida Network (\url{http://www.redecovida.org}).}
\beginsupplement
\section*{Supplementary Materials}
\paragraph*{Supplementary Material 1-}{ Methodology applied for epidemiological compartment models}
\vspace{-10pt}
\paragraph*{Supplementary Material 2-}{ Meta-population models formulation.}
\vspace{-10pt}
\paragraph*{Supplementary Material 3-}{Expressions and parameter values to estimate the reproduction numbers of the meta-population models}
\vspace{-10pt}
\paragraph*{Supplementary Material 4-}{Additional information about the municipalities.}
\section*{Applying the method to epidemiological models}
We proceed to illustrate the application of the method developed in this work to compartmental epidemic models. We chose two simple models that are known in literature and two variations of a meta-population model that is used in the main framework and developed in Supplementary Material 2. We seek to show the method step by step, presenting its detailed calculations. All models presented are composed of ordinary differential equations. In the transitions between infected compartments described in these models, $ \mathbfcal{V} (t, \tau) $, the the infected compartments $\bm {x}(t,\tau)$ appear with linear dependence. Therefore we can simplify:
\begin{equation}\label{cte}
\dfrac{\partial\overline{\mathcal{V}_i}}{\partial u_j} = \dfrac{\partial \mathcal{V}_i}{\partial x_j}=cte \qquad \text{and} \qquad \frac{d}{d\omega} \bm{u}(\omega) = -\frac{\partial {\mathbfcal{V}}}{\partial \bm{x}} \bm{u}(\omega).
\end{equation}
\noindent In addition, all the parameters of the models in this Supplementary Material are constant, which leads to
\begin{equation}
\bm{\Omega}(t)= \frac{\partial}{\partial \bm{X}} {\mathbfcal{F}}(t) \qquad \text{and} \qquad \overline{\bm{\Gamma}}(\omega) = \bm{\Gamma}(\tau)
\end{equation}
\subsection*{SEIR model}
The SEIR model is designed by introducing a new exposed $ E $ stage of the disease into the SIR model \cite{SIRfirst}. We can assume that when an individual becomes infected, it must go through a latency period before showing its first symptoms and starting to infect other individuals. This is accomplished by introducing the exposed compartment $ E $ and its removal rate $ \kappa $. Thus, all individuals who are infected start in the exposed state and, on average, after a $ 1 / \kappa $ latency time are introduced into the $ I $ compartment. It is also possible to introduce a factor related to a $ \epsilon $ pre-symptomatic infection. This way, individuals can start to infect in the exposed compartment, before presenting symptoms. Thus, this model, considering pre-symptomatic infection, can be written as
\begin{align}
\frac{dS}{dt}& =- \frac{\beta S}{N}\Big [ I + \epsilon E \Big ],\\
\frac{dE}{dt}& = \frac{\beta S}{N}\Big [ I + \epsilon E \Big ] - \kappa E,\\
\frac{dI}{dt}& = \kappa E - \gamma I,\\
\frac{dR}{dt}&= \gamma I.
\end{align}
Thus, we can sort the infected compartments as $\bm{X}(t)=[E(t), \;I(t)]$. In this way the distributions of the infectious phase are defined as $\bm{x}(t,\tau)=[i_e(t,\tau) ,i_i(t,\tau)]$. We get
${\mathbfcal{F}}(t)$ e ${\mathbfcal{V}}(t,\tau)$:
\begin{equation}
{\mathbfcal{F}}(t) =
\begin{pmatrix}
\frac{\beta S}{N}\big [ I + \epsilon E \big ]\\[1ex]
0
\end{pmatrix}, \qquad
{\mathbfcal{V}}(t,\tau) = \begin{pmatrix}
\kappa i_e(t,\tau)\\[1ex]
\gamma i_i(t,\tau) - \kappa i_e(t,\tau)
\end{pmatrix},
\end{equation}
\noindent as we recover the sub-set of equations for the infected compartments with:
\begin{equation}
\frac{d}{dt} \bm{X}(t) = \mathbfcal{F}(t) - \int_0 ^{\infty} \mathbfcal{V}(t,\tau) d\tau.
\end{equation}
\noindent The change from $ E $ to $ I $ is not considered a new infection, but the progression of the disease stage. Therefore, new infections only occur in the exposed compartment $ E $, causing $\mathcal{F}_2(t)=0$. Thus, from ${\mathbfcal{F}}(t)$, we obtain $\bm{\Omega}(t)$:
\begin{equation}
\bm{\Omega}(t) = \Big[ \frac{\partial}{\partial \bm{X}} {\mathbfcal{F}}(t) \Big] =
\begin{pmatrix}
\epsilon \frac{\beta S}{N} & \frac{\beta S}{N} \\
0 & 0
\end{pmatrix}.
\end{equation}
\noindent We proceed to obtain $\Gamma(\tau)$ by solving
\begin{equation}
\frac{d}{d\omega} \bm{u}(\omega) = - \frac{\partial {\mathbfcal{V}}}{\partial \bm{x}} \bm{u}(\omega) =
\begin{bmatrix}
-\kappa & 0 \\
\kappa & -\gamma
\end{bmatrix} \bm{u}(\omega).
\end{equation}
\noindent In order to obtain
\begin{equation}
\bm{u}(\omega) =
\begin{bmatrix}
e^{-\kappa\omega} & 0 \\
\frac{\kappa}{\gamma -\kappa} [e^{-\kappa\omega} - e^{-\gamma\omega} ]& e^{-\gamma\omega}
\end{bmatrix} \bm{u}(0).
\end{equation}
\noindent Therefore:
\begin{equation}
\bm{\Gamma}(\tau) = \begin{bmatrix}
e^{-\kappa\omega} & 0 \\
\frac{\kappa}{\gamma -\kappa} [e^{-\kappa\tau} - e^{-\gamma\tau} ]& e^{-\gamma\tau}
\end{bmatrix}
\end{equation}
With $\bm{\Omega}(t)$ and $\bm{\Gamma}(\tau)$
we perform the multiplication between the matrices, in order to obtain
\begin{equation}
\bm{A}(t,\tau) =
\begin{bmatrix}
\epsilon \frac{\beta S}{N} & \frac{\beta S}{N} \\
0 & 0
\end{bmatrix}
\begin{bmatrix}
e^{-\kappa\tau} & 0 \\
\frac{\kappa}{\gamma -\kappa} [e^{-\kappa\tau} - e^{-\gamma\tau} ]& e^{-\gamma\tau}
\end{bmatrix}.
\end{equation}
\noindent All terms in the second line are null, leaving only
\begin{align}
A_{11}(t,\tau)&= \epsilon \frac{\beta S}{N}e^{-\kappa\tau} + \frac{\beta S}{N}\frac{\kappa}{\gamma -\kappa} [e^{-\kappa\tau} - e^{-\gamma\tau} ],\\
A_{12}(t,\tau)&= \frac{\beta S}{N} e^{-\gamma\tau}.
\end{align}
\noindent Performing the integral from zero to infinity with respect to $ \tau $ we obtain the reproduction numbers of the system
${\mathbfcal{R}}(t)$:
\begin{equation}
{\mathbfcal{R}}(t)=\beta\frac{S}{N}
\begin{pmatrix}
\dfrac{\epsilon}{\kappa}+ \dfrac{1}{\gamma} & \dfrac{1}{\gamma} \\[2ex]
0 & 0
\end{pmatrix}.
\end{equation}
\noindent Where it is clear that
\begin{equation}
\bm{\overline{\mathbfcal{R}}}= \beta\frac{S}{N} \begin{pmatrix}
\dfrac{\epsilon}{\kappa}+ \dfrac{1}{\gamma} \\[2ex] \dfrac{1}{\gamma}
\end{pmatrix}.
\end{equation}
Since there is only generation of new infected in the exposed compartment, we have to $ \mathcal {F} _1 (t) = \mathcal {F} ^ T (t) $. Thus, the vector $ \bm {\alpha} (t) $ can be written as $ \bm {\alpha} = \big (1 , 0 \big) $, thus
\begin{equation}
{\mathbfcal{R}} = \bm{\alpha} \otimes \bm{\overline{\mathbfcal{R}}}=
\begin{pmatrix}
1 \\
0
\end{pmatrix} \bigotimes
\beta\frac{S}{N}
\begin{pmatrix}
\dfrac{\epsilon}{\kappa}+ \dfrac{1}{\gamma} \\[2ex] \dfrac{1}{\gamma}
\end{pmatrix}.
\end{equation}
\noindent Thus, to obtain the total reproduction number of the system, it is enough to make the scalar product between $\bm{\overline{\mathbfcal{R}}}$ and $\bm{\alpha}$:
\begin{equation}
\mathcal{R}^T(t)= \bm{\alpha} \bm{\cdot} \bm{\overline{\mathbfcal{R}}} = \frac{\beta S}{N} \Big [ \frac{\epsilon}{\kappa}+ \frac{1}{\gamma} \Big ],
\end{equation}
\noindent which leads to
\begin{equation}
g^T(\tau)= \frac{\epsilon \; e^{-\kappa\tau} + \frac{\kappa}{\gamma -\kappa} [e^{-\kappa\tau} - e^{-\gamma\tau} ]}{\epsilon/\kappa+ 1/\gamma} .
\end{equation}
When we make $ \epsilon \to 0 $, we retrieve the result of the reproduction number of the classic SEIR model (without pre-symptomatic infection),$ \mathcal {R} _0 = \beta / \gamma $. Similarly, the generation interval distribution of the SEIR model is also reduced to the well-known form in the literature \cite{champredon2018equivalence}, demonstrating robustness in the method. It is trivial to apply the method to the SIR model, whose analysis corresponds to tanking a limit of $\kappa \to \infty$. Therefore, both the reproduction number and the generation interval distribution, $g(\tau)= \gamma e^{-\gamma \tau}$, return to the known results from literature \cite{nishiura2009effective}.
\subsection*{SIIR model}
This is an adaptation of the SIR model, where we include two different manifestations of the disease, $ I_1 $ and $ I_2 $. In this model, there is only one susceptible population whose individuals can evolve into two compartments that carry the infectious agent. Both types of the disease carry the same pathogen, so that individuals infected with either type can go for $ I_1 $ and $ I_2 $. When infected, an individual has the probability $ p $ of manifesting the type $ I_1 $ of the disease and $ q = (1-p) $ of manifesting the type $ I_2 $. We assume two independent infection rates $ \beta_1 $ and $ \beta_2 $ for $ I_1 $ and $ I_2 $. Likewise, each slot has its own $ \gamma_1 $ or $ \gamma_2 $ removal rate. So, we write the model equations:
\begin{align}
\frac{dS}{dt}& =- \frac{(\beta_1 I_1 +\beta_2 I_2)S}{N},\\
\frac{dI_1}{dt}& = p\frac{(\beta_1 I_1 +\beta_2 I_2)S}{N} - \gamma_1 I_1,\\
\frac{dI_2}{dt}& = q\frac{(\beta_1 I_1 +\beta_2 I_2)S}{N} - \gamma_2 I_2,\\
\frac{dR}{dt}&= \gamma_1 I_1 + \gamma_2 I_2.
\end{align}
We sort the infected compartments as $\bm{X}(t)=[I_1(t), \;I_2(t)]$ and $\bm{x}(t,\tau)=[i_1(t,\tau) ,i_2(t,\tau)]$, where it is clear that $\bm{X}(t)= \int_{0}^{\infty}\bm{x}(t,\tau) d\tau$. We obtain ${\mathbfcal{F}}(t)$ and ${\mathbfcal{V}}(t,\tau)$ as:
\begin{equation}
{\mathbfcal{F}}(t) =\dfrac{S}{N}
\begin{pmatrix}
p (\beta_1 I_1 +\beta_2 I_2)\\[1ex]
q (\beta_1 I_1 +\beta_2 I_2)
\end{pmatrix}, \qquad
{\mathbfcal{V}}(t,\tau) = \begin{pmatrix}
\gamma_1 i_1(t,\tau)\\
\gamma_2 i_2(t,\tau)
\end{pmatrix}.
\end{equation}
\noindent Thus
\begin{equation}
\bm{\Omega}(t) = \dfrac{S}{N}
\begin{pmatrix}
p\beta_1 & p\beta_2 \\[1ex]
q\beta_1 & q\beta_2
\end{pmatrix}, \qquad -\frac{\partial {\mathbfcal{V}}}{\partial \bm{x}} =
\begin{pmatrix}
-\gamma_1 & 0 \\
0 & -\gamma_2
\end{pmatrix}.
\end{equation}
The linear O.D.E system described by the matrix $-\partial {\mathbfcal{V}}/\partial \bm{x}$ it is simple to solve, since this is a diagonal matrix. So, we get
\begin{equation}
\bm{\Gamma}(\tau) =
\begin{pmatrix}
e^{-\gamma_1 \tau} & 0 \\
0 & e^{-\gamma_2 \tau}
\end{pmatrix},
\end{equation}
\noindent which, when multiplied by the matrix $ \bm {\Omega} (t) $ results in
\begin{equation}
\bm{A}(t,\tau) = \dfrac{S}{N}
\begin{pmatrix}
p\beta_1 \; e^{-\gamma_1 \tau} & p\beta_2\; e^{-\gamma_2\tau} \\[2ex]
q\beta_1 \; e^{-\gamma_1 \tau} & q\beta_2 \; e^{-\gamma_2 \tau}
\end{pmatrix}
\end{equation}
\noindent It is interesting to realize that this is one of the cases where $A_{ij}(t,\tau)= \alpha_{i} \overline{A}_j(t,\tau)$, recalling that $\overline{A}_j(t,\tau)= \sum_{i} A_{ij}(t,\tau)$. Therefore $\bm{\alpha} =\big(p,q\big)$ and we can factorize the matrix into $\overline{\bm{A}}$ and $\bm{\alpha}$ as:
\begin{equation}
\bm{A} = \bm{\alpha} \otimes \overline{\bm{A}} =
\begin{pmatrix}
p \\
q
\end{pmatrix} \bigotimes
\begin{pmatrix}
\dfrac{\beta_1 S}{N} \; e^{-\gamma_1 \tau} \\[2ex]
\dfrac{\beta_2 S}{N}\; e^{-\gamma_2\tau} \\
\end{pmatrix}.
\end{equation}
\noindent Where $\otimes$ represents the tensorial product, $\bm{\alpha} \otimes \bm{\overline{\bm A}}= \Big [ \alpha_i \overline{A}_j \Big ]$. Integrating $\overline{\bm{A}}(t,\tau)$ in relation to $\tau$ we get:
\begin{equation}
\bm{\overline{\mathbfcal{R}}}(t) = \int_0 ^{\infty}\overline{\bm{A}}(t,\tau) d\tau =\dfrac{ S}{N}
\begin{pmatrix}
\beta_1/\gamma_1\\[1ex]
\beta_2/\gamma_2 \\
\end{pmatrix}.
\end{equation}
\noindent Therefore, we proceed to the scalar product between $\bm \alpha$ and $\overline{\mathcal{\bm{R}}}$:
\begin{equation}
\mathcal{R}^T (t) = \bm{ \alpha \cdot \overline{\mathbfcal{R}}}=\frac{S(t)}{N} \bigg [ p \frac{\beta_1}{\gamma_1} + q\frac{\beta_2}{\gamma_2} \bigg ].
\end{equation}
\noindent When $t \to 0$, $\mathcal{R}_0 = p \beta_1/\gamma_1 + q\beta_2/\gamma_2 $. We realize that the total reproduction number is the sum of the reproduction numbers of the two types of infection times the percentage of occurrence of each. Thus, we proceed to obtain the distribution of the generation interval, which takes the form:
\begin{equation}
g^T(\tau) = \dfrac{ p \beta_1 \; e^{-\gamma_1 \tau} + q \beta_2 \; e^{-\gamma_2 \tau} }{ \beta_1/\gamma_1 + \beta_2/\gamma_2 }.
\end{equation}
\subsection*{SIR-type meta-population model}
In this section we consider a meta-population model that will be used in the main framework and is detailed at Supplementary Material 2. Here we summarize the model. We consider the existence of ``$ n $'' meta-populations with coupled SIR-type dynamics. Where, due to the movement of individuals between meta-populations, an infected compartment in one meta-population can influence the disease transmission process of all the others. The coupling of the equations happens by the transmission rates $ \lambda_ {ik} $ related to the contamination process that emerges from the flow of individuals. The SIR-type model for $n$ meta-populations can be written as
\begin{align}
\frac{dS_i}{dt} =& - \sum_{j}^{n} \lambda_{ij}(t)\;I_j(t) \; S_i(t),\label{eqS}\\
\frac{dI_i}{dt}=& \sum_{j}^{n} \lambda_{ij}(t)\;I_j(t)\; S_i(t) -\gamma \; I_i(t),\label{eqI}\\
\frac{dR_i}{dt}=& \gamma I_i(t),\label{eqR}
\end{align}
\noindent in which $S_i(t)$, $I_i(t)$ and $R_i(t)$ correspond to the susceptible, infected and removed individuals that belong to the meta-population ``$i$''. The $\lambda_{ij}$ parameter is related to the transmission between the meta-populations ``$i$'' and ``$j$'', see Supplementary Material 2. We assume that the recover rate $\gamma$ is uniform for all meta-populations.The infected compartments are sorted as $\bm{X}(t)=[I_1(t), \;I_2(t), \hdots , I_n(t)]$ and $\bm{x}(t,\tau)=[i_1(t,\tau), \;i_2(t,\tau), \hdots , i_n(t,\tau)]$, where it is clear that $\bm{X}(t)= \int_{0}^{\infty}\bm{x}(t,\tau) d\tau$. We proceed to identify $\mathcal{F}_i(t)$ and $\mathcal{V}_i(t,\tau)$ as
\begin{equation}
\bm{\mathbfcal{F}}(t) = \Bigg [ \sum_j ^n \lambda_{ij}(t) I_j(t) S_i(t) \Bigg ], \qquad \bm{\mathbfcal{V}}(t,\tau) = \Bigg [ -\gamma \; i_i(t,\tau) \Bigg ]
\end{equation}
Therefore we get:
\begin{equation}
\bm{\Omega}(t) =
\begin{pmatrix}
\lambda_{11} S_1 & \lambda_{12} S_1 & \ldots & \lambda_{1n} S_1 \\
\lambda_{21} S_2 & \lambda_{22} S_2 & \ldots & \lambda_{2n} S_2 \\
\vdots & \vdots & \ddots & \vdots \\
\lambda_{n1} S_n & \lambda_{n2} S_n & \ldots & \lambda_{nn} S_n
\end{pmatrix},\qquad
-\frac{\partial \bm{\mathbfcal{V}}}{\partial \bm{x}} =
\begin{pmatrix}
\gamma & 0 & \ldots & 0 \\
0 & \gamma & \ldots & 0 \\
\vdots & \vdots & \ddots & \vdots \\
0 & 0 & \ldots & \gamma
\end{pmatrix}.
\end{equation}
It is clear to see that $-\partial \bm{\mathbfcal{V}}/\partial \bm{x}$ is a diagonal matrix, such that $\dfrac{d}{d\omega}\bm{u}(\omega) = -\dfrac{\partial \bm{\mathbfcal{V}}}{\partial \bm{x}} \bm{u}(\omega)$ are of trivial solution
\begin{equation}
\bm{\Gamma}(\tau) = e^{-\gamma \tau} \mathds{I},
\end{equation}
\noindent where $\mathds{I}$ represents the identity matrix of dimension ``$n$''. Thus, when multiplying the matrices $ \bm {\Omega} $ and $ \bm {\Gamma} $ we arrive at:
\begin{equation}
\bm{A}(t,\tau) = \Big [ \lambda_{ik} S_i \; e^{-\gamma \tau}\Big ]
\end{equation}
\noindent Integramos $\bm{A}$ de forma a obter a matriz de próxima geração do sistema $\bm{\mathbfcal{R}}$, dada por:
\begin{equation}
\bm{\mathbfcal{R}}(t) = \Bigg [ \dfrac{\lambda_{ik} S_i}{\gamma} \Bigg ].
\end{equation}
\noindent That leads to:
\begin{equation}
g_{ij}(\tau)= g(\tau)= \gamma \; e^{-\gamma \, \tau}.
\end{equation}
\noindent Whereby, if only one meta-population is considered, the SIR model results are recovered \cite{nishiura2009effective}.
\subsection*{SEIIR-type meta-population model}
Finally, we present a meta-population model related to the transmission dynamics of SARS-Cov-2 coronavirus. As in the previous section, the meta-populations are connected by the commuter movement of individuals. In this model, individuals that are infected have to pass through a latency period to become infectious, during that time, we consider that those are in the exposed compartment $E$. After a mean latency period of $1/k$, those infected can become symptomatic or asymptomatic, $I^s$ and $I^a$ respectively. While in these compartments, the individuals of ``$j$'' are able to generate new infected ones in ``$i$'' based on the transmission rate $\lambda_{ij}$. However, asymptomatic ones are considered to have a lower transmissibility, thus, the transmission rate must be multiplied by a factor $\delta$ that lowers it's infectivity. The SEIIR model was developed by Oliveira in \cite{Oliveira2020mathematical} and in Supplementary Material 2 we derive an SEIIR-type meta-population model that can be written as:
\begin{align}
\frac{dS_i}{dt} =& -\sum_{j}^{n} \lambda_{ij}(t)\; S_i(t) \Big [I^s_j(t)+\delta I^a_j(t)\Big ] \; S_i(t),\label{eqS-met2}\\
\frac{dE_i}{dt}=& \sum_{j}^{n} \lambda_{ij}(t)\; S_i(t) \Big [I^s_j(t)+\delta I^a_j(t)\Big ] -\kappa \; E_i(t),\\
\frac{dI^s_i}{dt}=& p\kappa \; E_i(t) -\gamma_s I^s _i(t),\\
\frac{dI^a_i}{dt}=& (1-p)\kappa \; E_i(t) -\gamma_a I^a _i(t),\\
\frac{dR_i}{dt}=& \gamma_s I^s_i(t) + \gamma_a I^a_i(t).\label{eqR2}
\end{align}
To proceed with the methodology, we sort the compartments as $X=[ E_1, ..., E_n, {I_1}^{a}, ..., {I_n} ^{a}, {I_1} ^{s}, ..., {I_n} ^{s}]$ and $x=[ {i^e} _{1}, ..., {i^e} _{n}, {i^s} _{1}, ..., {i _n} ^{s}, {i_1} ^{a}, ..., {i_n} ^{a}]$ in a way that if there are $n$ meta-populations we have $3n$ infected compartments. Therefore, we obtain $\mathbfcal{F}(t)$ and $\mathbfcal{V}(t,\tau)$ as:
\begin{align}\label{F_SEIIR}
\mathcal{F}_i(t)= \begin{cases}
\sum_{j}^{n} \lambda_{ij}(t)\; S_i(t) \Big [I^s_j(t)+\delta I^a_j(t)\Big ] , & \text{for $i\leq n$} \\
0, & \text{for $i > n$}
\end{cases}
\end{align}
\begin{align}\label{V_SEIIR}
\mathcal{V}_i(t,\tau)= \begin{cases}
\kappa\; x_i(t,\tau) , & \text{for $i\leq n$} \\
\gamma_s \; x_i(t,\tau) - p\kappa\; x_{i-n}(t,\tau), & \text{for $ n < i \leq 2n$} \\
\gamma_a \; x_i(t,\tau) - (1-p)\kappa\;
x_{i-2n}(t,\tau), & \text{for $ 2n < i$}
\end{cases}
\end{align}
\noindent where $I_j^s=X_{j+n}$ and $I_j^a=X_{j+2n}$. Thus, from \eqref{F_SEIIR}, we obtain:
\begin{align}\label{Omega_SEIIR}
\bm{\Omega}(t)= \begin{pmatrix}
\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} &\bm{\Omega^s} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} & \bm{\Omega^a} \\ \hline
\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} &\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} & \bm{0} \\ \hline
\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} &\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} & \bm{0}
\end{pmatrix}, \qquad \qquad -\frac{\partial \bm{\mathbfcal{V}}}{\partial \bm{x}}= \begin{pmatrix}
-\kappa & 0 & 0 \\
p \kappa & -\gamma_s & 0 \\
(1-p) \kappa & 0 & -\gamma_a
\end{pmatrix} \mbox{\Large{$\otimes$}} \; \mbox{\Large{$\mathds{I}_n$}}.
\end{align}
\noindent Whereby ${\mathds{I}_n}$ is an $n \times n$ identity matrix and $\otimes$ is a tensor product. Also, $\bm{\Omega}$ is divided into nine $n \times n$ submatrices, where $\bm{0}$ is an all-zeroes $n \times n$ matrix, $\bm{\Omega}^s\equiv \Big [\lambda_{ij} S_i \Big ]$ and $\bm{\Omega}^a\equiv \delta \bm{\Omega}^s$. Solving the system of differential equations on the characteristic line, we obtain:
\begin{align}\label{Gamma_SEIIR}
\bm{\Gamma}(\tau)= \begin{pmatrix}
e^{-\kappa \tau} & 0 & 0 \\[2ex]
p\frac{\kappa}{\gamma_s -\kappa} \Big(e^{-\kappa\tau} - e^{-\gamma_s\tau} \Big ) & e^{-\gamma_s \tau} & 0 \\[2ex]
(1-p)\frac{\kappa}{\gamma_a -\kappa} \Big(e^{-\kappa\tau} - e^{-\gamma_a\tau} \Big ) & 0 & e^{-\gamma_a \tau}
\end{pmatrix} \mbox{\Large{$\otimes$}} \; \mbox{\Large{${\mathds{I}_n}$}}.
\end{align}
We proceed into the multiplication of the matrices $\bm{\Omega}$ and $\bm{\Gamma}$ in order to obtain:
\begin{align}\label{Omega_SEIIR}
\bm{A}(t,\tau)= \begin{pmatrix}
\bm{A^e} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} &\bm{A^s} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} & \bm{A^a} \\ \hline
\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} &\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} & \bm{0} \\ \hline
\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} &\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} & \bm{0}
\end{pmatrix}
\end{align}
\noindent whereby the submatrices are defined as:
\begin{equation}
\bm{A^e}(t) \equiv \Bigg [ \lambda_{ij}(t) S_i(t) \bigg( p\frac{\kappa}{\gamma_s -\kappa} \Big(e^{-\kappa\tau} - e^{-\gamma_s\tau} \Big ) + \delta (1-p)\frac{\kappa}{\gamma_a -\kappa} \Big(e^{-\kappa\tau} - e^{-\gamma_a\tau} \Big ) \bigg) \Bigg ],
\end{equation}
\noindent $\bm{A^s}\equiv e^{-\gamma_s \tau}\bm{\Omega^s}$ and $\bm{A^a}\equiv e^{-\gamma_a \tau} \bm{\Omega^a}$. Therefore, the next generation matrix is
\begin{align}\label{Omega_SEIIR}
\bm{\mathbfcal{R}}(t)= \begin{pmatrix}
\bm{\mathbfcal{R}^e} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} &\bm{\mathbfcal{R}^s} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} & \bm{\mathcal{R}^a} \\ \hline
\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} &\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} & \bm{0} \\ \hline
\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} &\bm{0} & \hspace*{-\arraycolsep}\vline\hspace*{-\arraycolsep} & \bm{0}
\end{pmatrix}
\end{align}
\noindent where:
\begin{equation}
\bm{\mathbfcal{R}^e}(t) \equiv \Bigg [ \lambda_{ij} (t) S_i(t) \bigg( \frac{p}{\gamma_s} + \delta \frac{(1-p)}{\gamma_a} \bigg) \Bigg ],
\end{equation}
\noindent Because there is no generation of infected individuals on the symptomatic and asymptomatic compartments, the expressions for $\bm{\mathbfcal{R}^s}$ and $\bm{\mathbfcal{R}^a}$ are not needed. This is so because $\mathcal{F}_i(t)$ for $i>n$ is null for any time. Therefore, the renewal equations of $\mathcal{F}_i(t)$ for $i>n$ have the tautological result that zero equals zero, regardless of the values of $\bm{\mathbfcal{R}^s}$ or $\bm{\mathbfcal{R}^a}$. Thus, all we need to describe the dynamics is $\bm{\mathbfcal{R}}$ for $i,j\leq n$, that is $\bm{\mathbfcal{R}^e}$. Lastly, we obtain the generation interval distribution matrix for $i,j\leq n$:
\begin{equation}
g_{ij}(\tau) \equiv g(\tau)= \frac{\frac{p}{\gamma_s} g^s(\tau) + \frac{\delta (1-p)}{\gamma_a}g^a(\tau)}{\frac{p}{\gamma_s} + \frac{\delta (1-p)}{\gamma_a}},
\end{equation}
\noindent for
\begin{equation}
g^a(\tau) = \frac{\kappa \gamma_a}{\gamma_a - \kappa} (e^{-\kappa \tau} - e^{-\gamma_a \tau} ), \qquad \qquad \qquad g^s(\tau) = \frac{\kappa \gamma_s}{\gamma_s - \kappa} (e^{-\kappa \tau} - e^{-\gamma_s \tau} ) .
\end{equation}
\noindent Whereby, if only one meta-population is considered, we return to the results in literature \cite{Oliveira2020mathematical}.
\newpage
\topskip0pt
\vspace*{\fill}
\begin{center}
\begin{minipage}{.6\textwidth}
\Large{\textbf{Supplementary Material 2}}\\
\large{\textcolor{gray}{\textbf{Meta-population models formulation}}}
\end{minipage}
\end{center}
\vspace*{\fill}
\newpage
\section*{Meta-population models formulation}
In this Supplementary Material we will be, inspired by \cite{miranda2021scaling}, developing a meta-population model that takes into account the movement of individuals between meta-populations to describe the propagation of a disease through out multiple municipalities . We will be interpreting the inter-municipal flow as a complex network where its nodes represent municipalities and the weight of its edges represents the intensity of the flow between the connected municipalities. We also consider that each municipality can be interpreted as a meta-population, with its own compartments and parameters, which can be represented by a vector $\bm{y}_i(t)$, where the index “$i$” indicates the meta-population portrayed.Each entry of $\bm{y}_i(t)$ represents a compartment of this meta-population, such as $\bm{y}_i(t)=\big (S_i(t), I_i(t), R_i(t) \big)$ in the case of a SIR model. In which $S_i(t)$, $I_i(t)$ and $R_i(t)$ correspond to the susceptible, infected and removed individuals that are residents of the meta population "$i$", respectively. The sum of the elements of
$\bm{y}_i$ corresponds to the number of individuals of the meta-population "$i$", in the SIR model we have $S_i(t) + I_i(t) + R_i(t) =N_i(t)$.
We can represent the amount of individuals that goes from the meta-population "$i$" to another meta-population "$j$" each day as $\varphi_{ij}$, so it represents the flow of individuals between those meta-populations. Since each meta-population is described from it's compartments, we can describe the flow of those using $\bm{y}_i(t)$ in
\begin{equation}
\text{Flow from $i$ to $j$} = \varphi_{ij}(t) \frac{\bm{y}_{i}}{N_i}.\centering
\end{equation}
We define $\Phi_{ij}(t) \equiv \frac{\varphi_{ij}(t)}{N_i}$ as the density of flow. In that way, $\Phi_{ij}\; \bm{y_i}$ is the number of individuals from each compartment class that are flowing from "$i$" to "$j$". It's natural to see that sum of $\Phi_{ij}\; \bm{y}_i$ for each compartment class of $\bm{y}_i$ is equal to $\varphi_{ij}$, in the SIR model for example: $\Phi_{ij}\; S_i +\Phi_{ij}\;I_i + \Phi_{ij}\; R_i = \varphi_{ij}$.
Given that the meta-populations are connected through the flow , it is necessary to identify how they. For this purpose, graphs are used which, in general, represent complex networks, given the large number of connections. To represent networks, matrices are used, as in the case of adjacency matrices whose elements represent the density of flow ($\Phi_{ij}$). Thus, the values of each $\Phi_{ij}(t)$ are the weight of the edges with a null value representing the absence of an edge. Since there are no self interactions in this network, we do not consider the flow of a meta-population to itself, thus $\varphi_{ii}\equiv 0$ .
Due to the circulation of individuals through the network, the characteristics of the meta-populations will be changed. Thus, we define the effective population of ``$i$'' described by the vector $\bm{y}_{e,\;i} = \big(S_{e\; ,i}(t),I_{e\; ,i}(t),R_{e\; ,i}(t) \big)$ which corresponds to individuals located, at time t, in the “i” meta-population regardless of their origin meta-population.The effective population represents the new characteristics of a population due to the flow; for example, a meta-population that does not have infected resident individuals may have carriers of the pathogen in their effective population due to the flow from some other meta-population that presents infected individuals. We can write $\bm{y}_{e,\; i}$ as:
\begin{equation}
\bm{y}_{e,\; i}(t) = \overbrace{\bm{y}_i(t)}^\text{ Resident Pop. } -\quad\overbrace{ \sum_{j \neq i}^{n} \, \Phi_{ij} \; \bm{y}_{i}(t) }^\text{Outflow} \quad+ \quad \overbrace{\sum_{j \neq i}^{n} \Phi_{ji} \;\bm{y}_{j}(t)}^\text{Inflow}
\end{equation}
For, $N$ being the number of meta populations. Rearranging the equation:
\begin{equation}\label{Effective_Pop}
\bm{y}_{e,\;i}(t) = \bm{y}_{i}(t)\Big( 1 - \sum_{j}^{n} \, \Phi_{ij} \Big) + \sum_{j}^{n} \Phi_{ji}\; \bm{y}_{j}(t)
\end{equation}
It's clear then that the effective population of "$i$" is the sum of the individuals from "$i$" that are in "$i$" , $\bm{y}_{i}\Big( 1 - \sum_{j \neq i}^{n} \, \Phi_{ij} \Big)$, with the individuals from other meta populations that are in "$i$", $\sum_{j \neq i}^{n} \Phi_{ji}\; \bm{y}_{j}$. If we sum each element of $\bm{y}_{e,\; i}$ we have the effective population number: $N_{e,\; i}= N_i - \sum_{j \neq i}^{n}\varphi_{ij} + \sum_{j \neq i}^{n}\varphi_{ji}$. It is important to highlight that the resident population is not changed, therefore, we do not consider migration but only commuter periodic movement, whereby the individuals of one meta-population go to another to work or study.
We now proceed in order to describe how the transmission of a disease occurs in a meta-population, taking into account the flow of individuals in the network. We must then describe the occurrence of new infections in the meta-population $\bm{y}_i(t)$, which can occur within the “i” meta-population
\begin{equation}\label{Infection_In_city}
\frac{\beta _i}{N_i} \, \times
\overbrace{S_{i} \Big( 1 - \sum_{j}^{n} \, \Phi_{ij} \Big)}^\text{Susceptible Indv. from ``$i$'' in ``$i$''} \times \overbrace{I_{e,\;i}}^\text{Infected Indv. in ``$i$''},
\end{equation}
\noindent or on another meta-population "$j$"
\begin{equation}\label{Infection_Out_city}
\frac{\beta _j}{N_j} \times \overbrace{\Phi_{ij} \, S_{i}}^\text{Susceptible Indv. from ``$i$'' in ``$j$''} \times \overbrace{I_{e,\; j}}^\text{Infected Indv. in ``$j$''}.
\end{equation}
\noindent $\beta_i(t)$ and $\beta_j(t)$ being the transmission rates within the “i” and “j” meta-populations, respectively. The parameters $\beta_i(t)$ and $\Phi_{ij}(t)$ are time dependent, so we can incorporate the changes in the behavior of the populations on those variables. Thus, the number of new infections of individuals that belong to “$i$”, $\mathcal{F}_i(t)$, will be equal to the sum of \eqref{Infection_In_city} and \eqref{Infection_Out_city}. We can substitute equation \eqref{Effective_Pop} in expressions \eqref{Infection_In_city} and \eqref{Infection_Out_city}, in order to expand the term of the effective population of infected individuals($I_{e\; ,i}(t)$ ,$I_{e\; ,i}(t)$) and be able to express them according to the infected residents in each meta-population. Carrying out these operations, we get to:
\begin{equation}\label{Lambda}
\mathcal{F}_i(t) = \sum_{j}^{n} \lambda_{ij}(t)\; S_i(t)I_j(t)
\end{equation}
For:
\begin{align}
\lambda_{ii}&=\frac{\beta _i}{N_i} \; \Big( 1 - \sum_{j}^{n} \, \Phi_{ij} \Big)^2 + \sum_{j}^{n}\; \frac{\beta _j}{N_j} \; \Phi_{ij}^2 \label{lambdaI} \\
\lambda_{ij}&= \frac{\beta _i}{N_i} \; \Phi_{ji} \Big( 1 - \sum_{k}^{n} \, \Phi_{ik} \Big) + \frac{\beta _j}{N_j} \; \Phi_{ij} \; \Big( 1 - \sum_{k}^{n} \, \Phi_{jk} \Big) + \sum_{k}^{n} \frac{\beta _k}{N_k} \Phi_{ik} \; \Phi_{jk} \label{lambdaJ}
\end{align}
The $\lambda_{ij}(t)$ are related to the transmission between individuals of the same meta-population and from individuals of the “j” meta-population to individuals of the “i” meta-population.
\subsection*{SIR-Type model}
In this section, we construct the set of equations for the SIR-type model based in the contamination process modeled in this Supplementary Material. We will be modeling the resident populations of each municipality and will not be considering migration, therefore, resident individuals of one meta-population will always remain residents of that same meta-population. It is assumed that the susceptible population can only decrease due to infection process, therefore, no life and death dynamic. In the same way, the removed individuals are only generated by a removing rate $\gamma$, uniform for all meta-populations. Gathering those considerations, we write the system of equations:
\begin{align}
\frac{dS_i}{dt} =& - \sum_{j}^{n} \lambda_{ij}(t)\;I_j(t) \; S_i(t),\label{eqS-met}\\
\frac{dI_i}{dt}=& \sum_{j}^{n} \lambda_{ij}(t)\;I_j(t)\; S_i(t) -\gamma \; I_i(t),\label{eqI}\\
\frac{dR_i}{dt}=& \gamma I_i(t).\label{eqR}
\end{align}
\noindent Where it is clear that when $\varphi_{ij}=0$, for all $i$'s and $j$'s, each meta-population will be described by the classical SIR homogeneous model. Of course, this SIR approach is not taking into account the other heterogeneities of a disease besides space. On the following section we present a more sophisticated approach focused on the Covid-19 transmission dynamics.
\newpage
\subsection*{A meta-population model for Covid-19 (SEIIR)}
In this section, we establish a more precise description for the dynamics of SARS-Cov-2 coronavirus on the municipalities network. To do so, we consider, as in \cite{Oliveira2020mathematical}, that the infected individuals can be separated in three classes: the exposed $E$, individuals
infected which are in the latency period and do not transmit the disease; the symptomatic individuals $I^s$, that are infectious, present a substantial amount of symptoms and are registered in the official data; the
asymptomatic/undetected ones $I^a$, that are infectious but present mild/non-existing symptoms and are not registered in the official data. The infected individuals always start in the exposed compartment, a portion $p$ of them eventually becomes symptomatic and the other $(1-p)$ portion becomes asymptomatic. Since we have two types of infectious individuals, both of them must be taken in consideration in the generation of new infected. Therefore, assuming that the asymptomatic individuals transmission rate is a fraction $\delta$ of the symptomatic transmission rate, it is simple to derive, similarly to \eqref{Infection_In_city} and \eqref{Infection_Out_city}, the number of infected individuals that are generated on the exposed compartment of each meta-population:
\begin{equation}\label{Lambda2}
\mathcal{F}^e_i(t) = \sum_{j}^{n} \lambda_{ij}(t)\; S_i(t) \Big [I^s_j(t)+\delta I^a_j(t)\Big ].
\end{equation}
\noindent Therefore, by the same assumptions presented on the previews section, the following model is obtained:
\begin{align}
\frac{dS_i}{dt} =& -\sum_{j}^{n} \lambda_{ij}(t)\; S_i(t) \Big [I^s_j(t)+\delta I^a_j(t)\Big ] \; S_i(t),\label{eqS-met2}\\
\frac{dE_i}{dt}=& \sum_{j}^{n} \lambda_{ij}(t)\; S_i(t) \Big [I^s_j(t)+\delta I^a_j(t)\Big ] -\kappa \; E_i(t),\\
\frac{dI^s_i}{dt}=& p\kappa \; E_i(t) -\gamma_s I^s _i(t),\\
\frac{dI^a_i}{dt}=& (1-p)\kappa \; E_i(t) -\gamma_a I^a _i(t),\\
\frac{dR_i}{dt}=& \gamma_s I^s_i(t) + \gamma_a I^a_i(t).\label{eqR2}
\end{align}
\noindent Whereby $\kappa$, $\gamma_s$ and $\gamma_a$ are the removing rate of the exposed, symptomatic and asymptomatic compartments, respectively, and are uniform in all meta-populations.
\newpage
\topskip0pt
\vspace*{\fill}
\begin{center}
\begin{minipage}{.6\textwidth}
\Large{\textbf{Supplementary Material 3}}\\
\large{\textcolor{gray}{\textbf{Expressions and parameter values for evaluating $\mathbfcal{R}(t)$ for the meta-population models}}}
\end{minipage}
\end{center}
\vspace*{\fill}
\newpage
\section*{Expressions for evaluating $\mathbfcal{R}(t)$ using incidence data}
Here, we derive the expressions and parameter values needed to estimate the reproduction numbers for the meta-population models. Firstly, we start by substituting the explicit expressions of the reproduction numbers of each model in the equation 2.24 of the main framework. After some straightforward calculations, we obtain:
\begin{equation}\label{Est}
\mathbfcal{Q}_i (t) = \sum_j ^n \lambda_{ij}(t) a_j(t).
\end{equation}
\noindent whereby, $\mathbfcal{Q}_i (t) = \mathcal{B}_i(t)/S_i(t)$ and
\begin{equation}
a_j= \sum_{\tau=0} ^t \frac{1}{\gamma} g(\tau) \Delta t
\end{equation}
\noindent for the SIR model and
\begin{equation}
a_j= \sum_{\tau=0} ^t \Big[ \frac{p}{\gamma_s} + \frac{\delta (1-p)}{\gamma_a}\Big] g(\tau) \Delta t
\end{equation}
\noindent for the SEIIR model. We consider that the $\mathcal{B}_i(t)$ is the collection of all the new infections during a $\Delta t$ time interval of 1 day. With appropriate units of measure, we have $\Delta t=1$. In the SIR model we consider that every infection is reported, $\rho_i=1$ and in the SEIIR case only the symptomatic individuals report their infections, $\rho_i=p$. We proceed into estimating the susceptible population. If we integrate \eqref{eqS-met} and substitute \eqref{Lambda} and equation 2.23 of the main framework, we get:
\begin{equation}
S_i(t)= N_i - \sum_{t'=0} ^t\frac{\mathcal{B}_i(t')}{\rho_i}.
\end{equation}
Finally, substituting \eqref{lambdaI} and \eqref{lambdaJ} in \eqref{Est} results in:
\begin{equation}\label{Est2}
\mathbfcal{Q}_i (t) = \sum_j ^n \Theta_{ij}(t) \beta_j(t),
\end{equation}
\noindent for:
\begin{align}
\Theta_{ii}&= \frac{1}{N_i} \Bigg[ a_i \Big( 1- \sum_k ^n \varphi_{ik} \Big )^2 + \sum_j ^n a_j \varphi_{ji} \Big(1-\sum_k ^n \varphi_{ik} \Big ) \Bigg ],\label{ThetaI} \\
\Theta_{ij}&= \frac{1}{N_j} \Bigg[ a_j \varphi_{ij} \Big(1-\sum_k ^n \varphi_{jk} \Big ) + \varphi_{ij} \sum_{k} ^{n} a_k \varphi_{kj}\Bigg ].
\label{ThetaJ}
\end{align}
\noindent Therefore value of $\mathcal{Q}_i(t)$ and $\theta_{ij}$ for each day can be obtained using the reported data and parameters. Thus, for each day, this leaves us with algebraic system of ``$n$'' variables, $\beta_j(t)$ and ``$n$'' equations. We can analytically solve this system for every meta-population, with the help of a computer algorithm, and obtain the daily values of $\beta_j(t)$ of every meta population. Substituting the values of every $\beta_j(t)$ in \eqref{lambdaI} and \eqref{lambdaJ}, we can compute the values of the $\lambda_ij(t)$'s which leads us to the values of the reproduction number, equations 3.1 and 3.2 of the main framework. Since the available data is of daily number of cases, the $\bm{\mathcal{R}}(t)$ is evaluated for each calendar day.
\subsection*{Parameters}
To estimate the $\beta_i(t)$ parameters, and consequently estimate the reproduction numbers, the parameters of both models must be obtained. $\delta$, $\gamma_a$, $\gamma_s$, $p$, and $\kappa$ are estimated for the state of Rio de Janeiro in \cite{jorge2020assessing} and can be found on Table \ref{tabPara}. Additionally, in the SIR type model we assume $\gamma=\gamma_s$. The intermunicipal commuter movement of workers and students for the cities of Brazil, $\varphi_{ij}$, can be found in a study conducted by IBGE (Brasilian Institute of Geography and Statistics) \cite{ibge2016arranjos}.
\begin{table}[H]
\caption{Key epidemiological parameters of the SEIIR model obtained in \cite{jorge2020assessing}. }
\label{tabPara}
\centering
\begin{tabular}{lllllll}
\hline
\textbf{Parameter} & \textbf{Description} & \textbf{Value}
\\ \hline
$\delta$ & Asymptomatic/non-detected infectivity factor & $0.258$
\\
$p$ & Proportion of latent (E) that proceed to symptomatic infective & $0.273$
\\
$\kappa^{-1}$ & Mean exposed period (days$^{-1}$) & $1/0.25$
\\
$\gamma_{a}^{-1}$ & Mean asymptomatic period (days$^{-1}$) & $1/0.288$
\\
$\gamma_{s}^{-1}$ & Mean symptomatic period (days$^{-1}$) & $1/0.25$
\\ \hline
\end{tabular}
\end{table}
\newpage
\topskip0pt
\vspace*{\fill}
\begin{center}
\begin{minipage}{.6\textwidth}
\Large{\textbf{Supplementary Material 4}}\\
\large{\textcolor{gray}{\textbf{Additional information about the municipalities}}}
\end{minipage}
\end{center}
\vspace*{\fill}
\newpage
\section*{Additional information about the municipalities}
\begin{table}[H]
\centering
\resizebox{0.9\textwidth}{!}{%
\begin{tabular}{llll}
\hline
\textbf{Municipality} & \textbf{Acronyms} & \textbf{Populational size} & \textbf{Total reported cases} \\ \hline
Belford Roxo & BR & 485,687 & 8,578 \\
Duque de Caxias & DdC & 905,129 & 8,736 \\
Magé & Ma & 242,113 & 3,539 \\
Mesquita & Mq & 800,835 & 1,351 \\
Nilópolis & Ns & 154,749 & 1,245 \\
Niterói & Nt & 497,883 & 12,165 \\
Nova Iguaçu & NI & 167,287 & 5,908 \\
Queimados & Q & 150,333 & 2,376 \\
\textbf{Rio de Janeiro} & RJ & 6,592,227 & 95,444 \\
São Gonçalo & SG & 1,075,372 & 11,601 \\
São João de Meriti & SJdM & 448,340 & 3,167 \\ \hline
\end{tabular}%
}
\caption{}
\caption{ The selected cities of the state of Rio de Janeiro in the Southeast of Brazil with their acronyms, number of inhabitants, and the reported COVID-19 cases until September 14th (2020). In bold, the capital (Rio de Janeiro).}
\label{tab:my-table}
\end{table}
\begin{figure}[h]
\centering
\includegraphics[width={\linewidth}]{Figures/Mapa_Rio.png}
\caption{Distribution of cases by the chosen cities of this study.}
\label{map}
\end{figure}
\begin{figure}[H]
\centering
\includegraphics[width={\linewidth}]{Figures/Rplot12.pdf}
\caption{Average daily commuter movement between municipalities, due to workplaces. The thickness of each line is proportional to the amount of movement from one municipality to another. Colors are used only for the purpose of better visualization and have no specific meaning.}
\label{map}
\end{figure}
\newpage
|
\section{Introduction}
\label{sec:introduction}
If artificial agents are to be effective in the real world, they will need to thrive in environments populated by other agents.
Agents are typically goal-directed, sometimes by definition \citep{franklin1996agent}. While their goals can be different, they often depend on shared salient features of the environment, and may be able to interact with and affect the environment in similar ways.
Humans and other animals make ready use of these similarities to other agents while learning \citep{henrich2017secret,laland2018darwin}.
We can observe the goal-directed behaviours of other humans, and combine these observations with our own experiences, to quickly learn how to achieve our own goals.
If reinforcement learning~\citep[RL,][]{sutton2018reinforcement} agents could similarly interpret the behaviour of others, they could learn more efficiently, relying less on solitary trial and error.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{assets/misc/implementation.png}
\caption{
\textbf{The $\Psi \Phi$-learning algorithm for RL with no-reward demonstrations.}
Demonstrations $\mathcal{D}$ contain data from {\color{matlab-blue}other agents} for \emph{unknown} tasks. We employ \emph{inverse temporal difference learning} (ITD, cf. Section~\ref{subsec:inverse-temporal-difference-learning}) to recover other agents' successor features (SFs) $\Psi^k$ and preferences $\mathbf{w}^k$.
The {\color{matlab-green}ego-agent} combines the estimated SFs of others along with its own preferences $\mathbf{w}^{\text{ego}}$ and SFs $\Psi^{\text{ego}}$ with generalised policy improvement (GPI, cf. Section~\ref{subsec:successor-features-and-cumulants}), generating experience.
Both the demonstrations and the ego-experience are used to learn the {\color{matlab-red}shared cumulants} $\Phi$.
}
\label{fig:implementation}
\end{figure}
To this end, we formalise and address a problem setting in which an agent (the `ego-agent’) is given access to observations and actions drawn from the experiences of other goal-directed agents interacting with the same environment, but pursuing distinct goals.
These observed trajectories are unlabelled in the sense that they lack the goals or rewards of the other agents.
This type of data is readily available in many real-world settings, either from (i) observing other agents acting simultaneously with the ego-agent in the same (multi-agent) environment, or (ii) multi-task demonstrations collected independently from the ego-agent's experiences.
Consider autonomous driving as a motivating example: the car can observe the decisions of many nearby human drivers with various preferences and destinations, or may have access to a large offline dataset of such demonstrations.
Because the other agents are pursuing their own varied goals, it can be difficult to directly use this information with conventional imitation learning methods \citep{widrow1964pattern} or inverse RL (IRL) \citep{ng2000algorithms,ziebart2008maximum}.
While the ego-agent should not copy other agents directly, it is likely that the behaviour of all agents depends on shared features of the environment.
To disentangle such shared features from agent-specific goals, we turn to the framework of successor features \citep{dayan1993improving,kulkarni2016deep,barreto2017successor}.
Successor features are a representation that captures the sum of state features an agent's policy will reach in the future.
An agent's goal is represented separately as a preference vector.
In this paper, we demonstrate how a reinforcement learner can benefit from multi-task demonstrations using the framework of successor features.
The key contributions are:
\begin{enumerate}
\item \textbf{Offline multi-task IRL:}
We propose an inverse RL algorithm, called \emph{inverse temporal difference} (ITD) \emph{learning}.
Using only demonstrations, we learn shared state features, alongside per-agent successor features and inferred preferences.
The reward functions can be trivially computed from these learned quantities.
We show empirically that ITD achieves superior or comparable performance to prior methods.
\item \textbf{RL with no-reward demonstrations:}
By combining ITD with learning from environment interactions, we arrive at a novel algorithm for RL with unlabelled demonstrations, called $\Psi \Phi$-learning (pronounced `Sci-Fi').
$\Psi \Phi$-learning is compatible with sub-optimal demonstrations.
It treats the demonstrated trajectories as being soft-optimal under \emph{some} task and employs ITD to recover successor features for the demonstrators' policies.
$\Psi \Phi$-learning inherits the unbiased, asymptotic performance of RL methods while leveraging the provided demonstrations with ITD.
When the goals of any of the demonstrators are even partially aligned with the $\Psi \Phi$-learner, this enables much faster learning than solitary RL.
Otherwise, when the demonstrations are not useful or even misleading, it gracefully falls back to standard RL, unlike na{\"i}ve behaviour cloning or IRL.
\item \textbf{Few-shot adaptation with task inference:}
Taking full-advantage of the successor features framework, our $\Psi \Phi$-learner can even adapt zero-shot to new goals it has never seen or experienced during training, but which are partially aligned with the demonstrated goals.
This is possible due to the disentanglement of representations into task-specific features (i.e., preferences) and shared state features.
We can efficiently update the task-specific preferences and rely on generalised policy improvement for safe policy updates.
We derive worst-case bounds for the performance of $\Psi \Phi$-learning in zero-shot transfer to new tasks.
\end{enumerate}
We evaluate $\Psi \Phi$-learning in a set of grid-world environments, a traffic-flow simulator~\citep{highway-env}, and a task from the ProcGen suite~\cite{cobbe2020leveraging}, observing advantages over vanilla RL, imitation learning \cite{reddy2019sqil,ho2016generative}, and auxiliary-task baselines \cite{hernandez2017survey}.
Thanks to the shared state features between the ITD and RL components, we find empirically that the $\Psi \Phi$-learner not only improves its ego-learning with demonstrations, but also enhances its ability to model others agents using its own experience.
\section{Background and Problem Setting}
\label{sec:background-and-problem-setting}
We consider a world that can be represented as an infinite horizon controlled Markov process (CMP) given by the tuple: $\mathcal{C} \triangleq \langle \mathcal{S}, \mathcal{A}, P, \gamma \rangle$.
$\mathcal{S}$ and $\mathcal{A}$ represent the continuous state and discrete action spaces, respectively, $\mathbf{s}' \sim P(\cdot \vert \mathbf{s}, \mathbf{a})$ describes the transition dynamics and $\gamma$ is the discount factor.
A \textbf{task} is formulated as a Markov decision process~\citep[MDP,][]{puterman2014markov}, characterised by a reward function, $R: \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$, i.e., $M \triangleq \langle \mathcal{C}, R \rangle$.
The goal of an agent is to find a policy which maps from states to a probability distribution over actions, $\pi: \mathcal{S} \rightarrow \Delta(\mathcal{A})$, maximising the expected discounted sum of rewards $G^{R} \triangleq \sum_{t=0}^{\infty} \gamma^{t} R(\mathbf{s}_t, \mathbf{a}_t)$.
The action-value function of the policy $\pi$ is given by $Q^{\pi, R}(\mathbf{s}, \mathbf{a}) \triangleq \mathbb{E}^{\mathcal{C}, \pi} \left[ \left. G^{R} \right| \mathbf{s}_0 = \mathbf{s}, \mathbf{a}_0 = \mathbf{a} \right]$, where $\mathbb{E}^{\mathcal{C}, \pi} \left[ \cdot \right]$ denotes expected value when following policy $\pi$ in environment $\mathcal{C}$.
\subsection{RL with No-Reward Demonstrations}
\label{subsec:rl-with-no-reward-demonstrations}
We are interested in settings in which, in addition to an environment $\mathcal{C}$, the agent has also access to \textbf{demonstrations without rewards}, i.e., behavioural data of mixed and unknown quality.
The demonstrations are generated by other agents, whose goals and levels of expertise are unknown, and who have no incentive to educate the controlled agent.
We will refer to the controlled agent, i.e., reinforcement learner, as the `ego-agent' and to the agents that generated the demonstrations as `other-agents'.
We denote the demonstrations with $\mathcal{D} = \{ \tau_1, \tau_2, \ldots, \tau_N \}$, where the trajectory $\tau \triangleq \left( \mathbf{s}_0, \mathbf{a}_0, \ldots, \mathbf{s}_T, \mathbf{a}_T; k \right)$ is generated by the $k$-th agent.
Note that each trajectory does include an identifier of the agent that generated it.
The ego-agent also gathers its own experience by interacting with the environment, collecting data $\mathcal{B} = \{ \left( \mathbf{s}, \mathbf{a}, \mathbf{s}', r^{\text{ego}} \right) \}$.
Due to the lack of reward annotations in $\mathcal{D}$, and the fact that the data may be irrelevant to the ego-agent's task, it is not trivial to combine demonstrations from $\mathcal{D}$ with the ego-agent's experience $\mathcal{B}$.
\subsection{Successor Features and Cumulants}
\label{subsec:successor-features-and-cumulants}
To make use of the demonstrations $\mathcal{D}$, we wish to capture the notion that while the agents' rewards may differ, they share the same environment.
To do so we turn to the framework of successor features (SFs) ~\citep{barreto2017successor}, in which rewards are decomposed into cumulants and preferences:
\begin{definition}[Cumulants and Preferences]
The (one-step) rewards are decomposed into task-agnostic \emph{cumulants} $\Phi(\mathbf{s}, \mathbf{a}) \in \mathbb{R}^{d}$, and task-specific \emph{preferences} $\mathbf{w} \in \mathbb{R}^{d}$:
%
\begin{equation}
R^{\mathbf{w}}(\mathbf{s}, \mathbf{a}) \triangleq \Phi(\mathbf{s}, \mathbf{a})^{\top} \mathbf{w} \,.
\label{eq:cumulants-times-preferences}
\end{equation}
\end{definition}
The preferences $\mathbf{w}$ are a representation of a possible goal in the world $\mathcal{C}$, in the sense that each $\mathbf{w}$ gives rise to a task $M^{\mathbf{w}} = \langle \mathcal{C}, R^{\mathbf{w}} \rangle$.
We use `task', `goal', and `preferences' interchangeably when context makes it clear whether we are referring to $\mathbf{w}$ itself, or the corresponding $M^{\mathbf{w}}$ or $R^{\mathbf{w}}$.
The action-value function for a policy $\pi$ in $M^{\mathbf{w}}$ is then a function of the preferences $\mathbf{w}$ and the $\pi$'s successor features.
\begin{definition}[Successor Features]
For a given discount factor $\gamma \in [0, 1)$, policy $\pi$ and cumulants $\Phi(\mathbf{s}, \mathbf{a}) \in \mathbb{R}^{d}$, the successor features (SFs) for a state $\mathbf{s}$ and action $\mathbf{a}$ are:
%
\begin{equation}
\Psi^{\pi}(\mathbf{s}, \mathbf{a}) \triangleq \mathbb{E}^{\mathcal{C}, \pi} \left[ \left. \sum_{t=0}^{\infty} \gamma^{t} \Phi(\mathbf{s}_t, \mathbf{a}_t) \right| \mathbf{s}_0 = \mathbf{s}, \mathbf{a}_0 = \mathbf{a} \right] \,.
\label{eq:sf-definition}
\end{equation}
\end{definition}
The $i$-th component of $\Psi^{\pi}(\mathbf{s}, \mathbf{a})$ gives the expected discounted sum of $\Phi(\mathbf{s}, \mathbf{a})$'s $i$-th component, when starting from state $\mathbf{s}$, taking action $\mathbf{a}$ and then following policy $\pi$.
Intuitively, cumulants $\Phi$ can be seen as a vector-valued reward function and SFs $\Psi^{\pi}$ the corresponding vector-valued state-action value function for policy $\pi$.
An action-value function is then given by the dot product of the preferences $\mathbf{w}$ and $\pi$'s SFs:
\begin{equation}
Q^{\pi, \mathbf{w}}(\mathbf{s}, \mathbf{a}) = \Psi^{\pi}(\mathbf{s}, \mathbf{a})^{\top} \mathbf{w} \,.
\label{eq:Q-Psi-times-w}
\end{equation}
\begin{proof}
See~\citep{barreto2017successor}.
\end{proof}
Note that if we have $\Psi^{\pi}$, the value of $\pi$ for a new preference $\mathbf{w}'$ can be easily computed.
This property allows the successor features of a set of policies to be repurposed for accelerating policy updates, as follows.
\begin{definition}[Generalised Policy Improvement]
Given a set of policies $\Pi = \{ \pi_{1}, \ldots, \pi_{K} \}$ and a task with reward function $R$, generalised policy improvement (GPI) is the definition of a policy $\pi'$ s.t.
%
\begin{equation}
Q^{\pi', R}(\mathbf{s}, \mathbf{a}) \geq \sup_{\pi \in \Pi} Q^{\pi, R}(\mathbf{s}, \mathbf{a}) \,, \forall \mathbf{s} \in \S, \mathbf{a} \in \mathcal{A} \,.
\label{eq:gpi}
\end{equation}
\end{definition}
Provided the SFs of a set of policies, i.e., $\{\Psi^{\pi^{k}}\}_{k=1}^{K}$, we can apply GPI to derive a new policy $\pi'$ whose performance on a task $\mathbf{w}$ is no worse that the performance of any of $\pi \in \Pi$ on the same task, given by
\begin{equation}
\pi'(\mathbf{s}) = \argmax_{a} \max_{\pi \in \Pi} \Psi^{\pi}(\mathbf{s}, a)^{\top} \mathbf{w} \,.
\label{eq:gpi-with-sf}
\end{equation}
Eqn.~(\ref{eq:gpi-with-sf}) suggests that if we could estimate the SFs of other agents, we could utilise them for improving the ego-agent's policy with GPI.
However, to do so with conventional methods we would require access to their rewards, cumulants and/or preferences.
In our setting, we can only observe their sequence of states and actions (Section~\ref{subsec:rl-with-no-reward-demonstrations}).
Next, we introduce our method that only requires no-reward demonstrations to estimate SFs and can be integrated seamlessly with GPI for accelerating reinforcement learning.
\section{Accelerating RL with Demonstrations}
\label{sec:method}
We now present a novel method, $\Psi \Phi$-learning, that leverages reward-free demonstrations to accelerate RL, shown in Figure~\ref{fig:implementation}.
Our approach consists of two components: (i) a novel inverse reinforcement learning algorithm, called \emph{inverse temporal difference} (ITD) \emph{learning}, for learning cumulants, per-agent successor features and corresponding agent preferences from demonstration without reward labels, and (ii) a novel RL
algorithm that combines ITD with generalised policy improvement (GPI, Section~\ref{subsec:successor-features-and-cumulants}).
\subsection{Inverse Temporal Difference Learning}
\label{subsec:inverse-temporal-difference-learning}
Given demonstrations without rewards, $\mathcal{D}$, we model the agents that generated the data (i.e., blue nodes in Figure~\ref{fig:implementation}) as soft-optimal for an \emph{unknown} task.
In particular, the $k$-th agent's policy is soft-optimal under task $\mathbf{w}^{k}$ and is given by
\begin{equation}
\pi^{k}(\mathbf{a} | \mathbf{s}) = \frac{\exp(\Psi^{\pi^{k}}(\mathbf{s}, \mathbf{a})^{\top} \mathbf{w}^{k})}{\sum_{a} \exp(\Psi^{\pi^{k}}(\mathbf{s}, a)^{\top} \mathbf{w}^{k})} \,, \forall \mathbf{s} \in \S, \mathbf{a} \in \mathcal{A} \,.
\label{eq:pi-k}
\end{equation}
We choose to represent the action-value functions of the other agents with their SFs and preferences to enable GPI, and to expose task- and policy-agnostic structure in the form of shared cumulants $\Phi$.
The $k$-th agent's successor features are temporally consistent with these cumulants $\Phi$
\begin{equation}
\Psi^{\pi^{k}}(\mathbf{s}, \mathbf{a}) = \Phi(\mathbf{s}, \mathbf{a}) + \gamma\mathbb{E}^{\mathcal{C}, \pi^{k}} \left[ \Psi^{\pi^{k}}(\mathbf{s}', \pi^{k}(\mathbf{s})) \right] \,.
\label{eq:psi-k}
\end{equation}
To learn these quantities from $K$ demonstrators, we parameterise the SFs with $\boldsymbol{\theta}^{\Psi^{k}}$, preferences with $\mathbf{w}^{k}$, and shared cumulants with $\boldsymbol{\theta}_{\Phi}$.
A schematic of the model architecture and further details are provided in Appendix~\ref{app:implementation-details}.
The parameters are learned by minimising a behavioural cloning and SFs TD loss based on equations \eqref{eq:pi-k} and \eqref{eq:psi-k}.
\textbf{Behavioural cloning loss.}
~Given demonstrations generated only by the $k$-th agent, i.e., $\mathcal{D}^{k} \subset \mathcal{D}$, we train its successor features $\boldsymbol{\theta}_{\Psi^{k}}$ and the preferences $\mathbf{w}^{k}$ by minimising the negative log-likelihood of the demonstrations
\begin{equation}
\begin{tiny}
\! \! \! \! \! \L_{\text{BC-}Q}(\boldsymbol{\theta}_{\Psi^{k}}, \mathbf{w}^{k}) \triangleq - \! \! \! \! \! \mathbb{E}_{\tau \sim \mathcal{D}^{k}} \! \! \! \log \frac{\exp(\Psi(\mathbf{s}_{t}, \mathbf{a}_{t}; \boldsymbol{\theta}_{\Psi^{k}})^{\top} \mathbf{w}^{k})}{\sum_{a} \exp(\Psi(\mathbf{s}_t, a; \boldsymbol{\theta}_{\Psi^{k}})^{\top} \mathbf{w}^{k})} \,.
\label{eq:bcq-loss}
\end{tiny}
\end{equation}
Importantly, Eqn.~(\ref{eq:bcq-loss}) reflects the fact that along a trajectory $\tau$, the successor features are a function of the state and action at each time-step, $\mathbf{s}_t$ and $\mathbf{a}_t$, while the preferences $\mathbf{w}^{k}$ are learnable but consistent across time and trajectories.
The direction of the preference $\mathbf{w}^k$ indicates the goal of the agent $k$ by showing how relatively rewarding it finds the different dimensions of the cumulant features $\Phi$.
The learned magnitude of $\mathbf{w}^k$ can further capture how greedily the agent $k$ pursues this goal.
A sparsity prior, i.e., $\L_{1}$ loss, on preferences $\mathbf{w}^{k}$ is also used to promote disentangled cumulant dimensions (see Figure~\ref{fig:cumulants}).
We found the $\L_{1}$ loss made the algorithm more robust to the choice of dimension of $\Phi$ (see Figure \ref{fig:sensitivity-to-size-of-cumulants} in the Appendix), but did not substantially affect overall performance.
\textbf{Inverse temporal difference loss.}
~The cumulant parameters $\boldsymbol{\theta}_{\Phi}$ are trained to be TD-consistent with all agents' successor features.
This procedure inverts\footnote{Hence the name \emph{inverse TD learning}.} the standard TD-learning framework for SFs~\citep{dayan1993improving,barreto2017successor} where they are trained to be consistent with a fixed cumulant $\Phi$.
Instead, we first train the SFs and preference vectors to `explain' the other agents' behaviour with the behavioural cloning loss Eqn.~(\ref{eq:bcq-loss}), and then train $\boldsymbol{\theta}_{\Phi}$, $\boldsymbol{\theta}_{\Psi^{k}}$ to be (self-)consistent with these SFs by minimising
\begin{align}
\L_{\text{ITD}}(\boldsymbol{\theta}_{\Phi}, &\boldsymbol{\theta}_{\Psi^{k}}) \triangleq \mathbb{E}_{(\mathbf{s}_{t}, \mathbf{a}_{t}, \mathbf{s}_{t+1}, \mathbf{a}_{t+1}, k) \sim \mathcal{D}} \| \Psi(\mathbf{s}_{t}, \mathbf{a}_{t}; \boldsymbol{\theta}_{\Psi^{k}}) \hookleftarrow \nonumber \\ & - \Phi(\mathbf{s}_{t}, \mathbf{a}_{t}; \boldsymbol{\theta}_{\Phi}) - \underdescribe{\gamma \Psi(\mathbf{s}_{t+1}, \mathbf{a}_{t+1}; \tilde{\boldsymbol{\theta}}_{\Psi^{k}})}{\texttt{stop\_gradient}} \| \,.
\label{eq:itd-loss}
\end{align}
In practice, our ITD-learning algorithm alternates between minimising $\L_{\text{BC-}Q}$ for training only the successor features and preferences, and $\L_{\text{ITD}}$ for training both the shared cumulants and successor features, provided only with no-reward demonstrations, as illustrated in blue in Figure~\ref{fig:implementation}.
From the definition of cumulants and preferences, we can recover the $k$-th demonstrator's reward, by applying Eqn.~(\ref{eq:cumulants-times-preferences}), i.e., $R^{k}(\mathbf{s}, \mathbf{a}) \approx \Phi(\mathbf{s}, \mathbf{a}; \boldsymbol{\theta}_{\Phi})^{\top} \mathbf{w}^{k}$.
Our ITD algorithm returns both Q-functions that can be used for imitating a demonstrator and an explicit reward function for each agent, requiring only access to demonstrations without any online interaction with the simulator.
Hence ITD-learning is an offline multi-task IRL algorithm.
ITD is summarised in Algorithm~\ref{algo:itd}.
\begin{theorem}[Validity of the ITD Minimiser]
The minimisers of $\L_{\text{BC-}Q}$ and $\L_{\text{ITD}}$ are potentially-shaped cumulants that explain the observed reward-free demonstrations.
\label{thm:itd}
\end{theorem}
\begin{proof}
See Appendix~\ref{app:proofs}.
\end{proof}
\textbf{Single-task setting.}
~To gain more intuition about the ITD algorithm, consider the simpler case of performing IRL with demonstrations from a single policy.
This obviates the need for a representation of preferences, so we can use $\mathbf{w}=1$.
In this case $\Psi$ is the action-value function $Q$ and $\Phi$ is simply the reward $R$.
Minimising \eqref{eq:bcq-loss} reduces to finding a $Q$-function whose softmax gives the observed policy, and minimising \eqref{eq:itd-loss} finds a scalar reward that explains the $Q$-function.
Our more general formulation, with cumulants $\Phi$ in place of a scalar reward, allows us to perform ITD-learning on demonstrations from many policies, and to efficiently transfer to new tasks, as we show next.
\subsection{$\Psi \Phi$-Learning with No-Reward Demonstrations}
\label{subsec:psi-phi-learning-with-no-reward-demonstrations}
Now we present our main contribution, $\Psi \Phi$-learning,
which combines our ITD inverse RL algorithm with RL and GPI, using no-reward demonstrations from other agents to accelerating the ego-agent's learning.
$\Psi \Phi$-learning is depicted in Figure~\ref{fig:implementation} and summarised in Algorithm~\ref{algo:itd}.
$\Psi \Phi$-learning is an off-policy algorithm based on Q-learning~\citep{watkins1992q,mnih2013playing}. The action-value function is represented with successor features, $\Psi^{\text{ego}}$, and preferences, $\mathbf{w}^{\text{ego}}$, as in Eqn.~(\ref{eq:Q-Psi-times-w}).
The ego-agent interacts with the environment, storing its experience in a replay buffer, $\mathcal{B} \leftarrow \mathcal{B} \cup \{ \left( \mathbf{s}, \mathbf{a}, \mathbf{s}', r^{\text{ego}} \right) \}$.
The $\Psi \Phi$-learner also has estimates for the cumulants, per-agent SFs and preferences obtained with ITD from the demonstrations $\mathcal{D}$.
\textbf{Reward loss.}
~The ego-rewards, $r^{\text{ego}}$, are used to ground the cumulants $\Phi$ and the preferences $\mathbf{w}^{\text{ego}}$, via the loss
\begin{align}
\L_{\text{R}}(\boldsymbol{\theta}_{\Phi}, \mathbf{w}^{\text{ego}}) \triangleq \mathbb{E}_{(\mathbf{s}, \mathbf{a}, r^{\text{ego}}) \sim \mathcal{B}} &\| \Phi(\mathbf{s}, \mathbf{a}; \boldsymbol{\theta}_{\Phi})^{\top} \mathbf{w}^{\text{ego}} - r^{\text{ego}} \| \,.
\label{eq:r-loss}
\end{align}
Importantly, we share the \emph{same} cumulants between the ITD-learning from other agents and the ego-learning, so that they span the joint space of reward functions.
This can be also seen as a representation learning method, where by enforcing all agents, including the ego-agent, to share the same $\Phi$, we transfer information about salient features of the environment from learning about one agent to benefit learning about all agents.
\textbf{Temporal difference learning.}
~The ego-agent's successor features are learned using two losses. First, we train the SFs to fit the Q-values using the Bellman error
\begin{align}
\mathcal{L}_{Q}(\boldsymbol{\theta}_{\Psi^{\text{ego}}}) \triangleq & \mathbb{E}_{(\mathbf{s}, \mathbf{a}, \mathbf{s}', r^{\text{ego}}) \sim \mathcal{B}} \| \Psi(\mathbf{s}, \mathbf{a}; \boldsymbol{\theta}_{\Psi^{\text{ego}}})^{\top} \mathbf{w}^{\text{ego}} \hookleftarrow \nonumber \\ & - r^{\text{ego}}
-\underdescribe{ \gamma \max_{a'} \Psi(\mathbf{s}', \mathbf{a}'; \tilde{\boldsymbol{\theta}}_{\Psi^{\text{ego}}})^\top \mathbf{w}^{\text{ego}}}{\texttt{stop-gradient}} \| \,.
\label{eq:q-loss}
\end{align}
\vspace{4em}
We additionally train the successor features to be self-consistent (i.e. to satisfy Equation~\ref{eq:sf-definition}) using a TD loss $\mathcal{L}_{\text{TD-}\Psi}$.
\begin{align}
\mathcal{L}_{\text{TD-}\Psi}(\boldsymbol{\theta}_{\Psi^{\text{ego}}}) \triangleq & \mathbb{E}_{(\mathbf{s}, \mathbf{a}, \mathbf{s}', \mathbf{a}') \sim \mathcal{B}} \| \Psi(\mathbf{s}, \mathbf{a}; \boldsymbol{\theta}_{\Psi^{\text{ego}}}) \hookleftarrow \nonumber \\ & - \Phi(\mathbf{s}, \mathbf{a}; \tilde{\boldsymbol{\theta}}_{\Phi})
- \gamma \underdescribe{\Psi(\mathbf{s}', \mathbf{a}'; \boldsymbol{\theta}_{\Psi^{\text{ego}}})}{\texttt{stop-gradient}}\| \,.
\label{eq:psi-loss}
\end{align}
\textbf{GPI behavioural policy.}
~Provided SFs estimates for the other agents, $\{\boldsymbol{\theta}_{\Psi^{k}}\}_{k=1}^{K}$, and the ego-agent $\boldsymbol{\theta}_{\Psi^{\text{ego}}}$, and inferred ego preferences, $\mathbf{w}^{\text{ego}}$, we adopt an action selection mechanism according to the GPI rule in Eqn.~(\ref{eq:gpi-with-sf})
\begin{align}
\pi^{\text{ego}}(\mathbf{s}) = \argmax_{a} \max_{\boldsymbol{\theta}_{\Psi}} \Psi(\mathbf{s}, a; \boldsymbol{\theta}_{\Psi})^{\top} \mathbf{w}^{\text{ego}} \,.
\label{eq:pi-ego}
\end{align}
The GPI step lets the agent estimate the value of the demonstration policies on its current task, and then copy the policy that it predicts will be most useful.
We combat model overestimation by acting pessimistically with regard to an ensemble of two successor features approximators.
If the agent's estimated values are accurate and the demonstration policies are useful for the ego task, the GPI policy can obtain good performance faster than policy iteration with only the ego value function.
The next section quantifies this claim.
\begin{figure*}
\centering
\begin{subfigure}[b]{0.26\linewidth}
\centering
\includegraphics[width=\linewidth]{assets/misc/highway/highway-v0.png}
\caption{Highway}
\end{subfigure}
~
\begin{subfigure}[b]{0.238\linewidth}
\centering
\includegraphics[width=\linewidth]{assets/misc/highway/roundabout-v0.png}
\caption{Roundabout}
\label{fig:envs_roundabout}
\end{subfigure}
~
\begin{subfigure}[b]{0.19\linewidth}
\centering
\includegraphics[width=\linewidth]{assets/misc/coingrid/coingrid.pdf}
\caption{CoinGrid}
\end{subfigure}
~
\begin{subfigure}[b]{0.189\linewidth}
\centering
\includegraphics[width=\linewidth]{assets/misc/fruitbot/fruitbot_5.pdf}
\caption{Fruit Bot}
\end{subfigure}
\caption{
Environments studied in this paper. Environments (a-b) are multi-agent environments in which the ego-agent must learn online from other agents, and learn to navigate around other agents in the environment (see Section~\ref{sec:results_accelerate_rl}).
Environments (c-d) are single-agent.
In environment (c) we test whether the ego-agent can learn offline from a set of demonstrations previously collected by other agents (see Section~\ref{sec:results_transfer}. Environment (d) is used to test whether our method can scale to more complex, high-dimensional tasks (see Section~\ref{sec:results_accelerate_rl}).
}
\label{fig:envs}
\vspace{-0.5em}
\end{figure*}
\textbf{Performance Bound.}
~Given a set of demonstration task vectors $\{w_k\}_{k=1}^{K}$ and successor features for the corresponding optimal policies $\{ \Psi^{\pi^{k}} \}_{k=1}^{K}$, \citet{barreto2017successor} show that it is possible to bound the performance of the GPI policy on the ego-agent task $w'$ as a function of the distance of $w'$ from the closest demonstration task $w_j$, and the value approximation error of the predicted value functions $\widetilde{Q}^{\pi_i} = \Psi^{\pi_i} w'$.
We extend this result to explicitly account for the reward approximation error $\delta_r$ obtained by the the learned cumulants and the SF approximation error $\delta_\Psi$.
\begin{theorem}[Generalisation Bound of $\Psi \Phi$-Learning]
Let $\pi^*$ be the optimal policy for the ego task $w'$ and let $\pi$ be the GPI policy obtained from $\{\tilde{Q}^{\pi_i}\}$, with $\delta_r, \delta_\Psi$ the reward and successor feature approximation errors. Then $\forall s,a$
\begin{align}
Q^{*}(s,a) - Q^\pi(s,a) \leq &\frac{2}{1-\gamma} \bigg [ \phi_{\max} \min_{j} \| w' - w_j\| \hookleftarrow \nonumber \\ & + 2 \delta_r + \|w'\| \delta_{\Psi} + \frac{\delta_r}{1-\gamma} \bigg] \; .
\end{align}
\label{thm:gpi}
\end{theorem}
\begin{proof}
See Appendix~\ref{app:proofs} for a formal statement.
\end{proof}
In settings where the agent has a good reward and SFs approximation, and the ego task vector $w'$ is close to the demonstration tasks, Theorem~\ref{thm:gpi} says that the ego-agent will attain near-optimal performance from the start of training.
\section{Experiments}
\label{sec:experiments}
We conduct a series of experiments to determine how well $\Psi \Phi$-learning functions as an RL, IRL, imitation learning, and transfer learning algorithm.
\textbf{Baselines.}
~We benchmark against the following methods: (i) \textbf{DQN:} Deep Q-learning \cite{mnih2013playing}, (ii) \textbf{BC:} Behaviour Cloning, a simple imitation learning method in which we learn $p(\mathbf{a} | \mathbf{s})$ via supervised learning on the demonstration data, (iii) \textbf{DQN+BC-AUX:} DQN with an additional behavior-cloning auxiliary loss \cite{hernandez2019agent}, (iv) \textbf{GAIL:} Generative Adversarial Imitation Learning \cite{ho2016generative}, which uses a GAN-like approach to approximate the expert policy, and (v)
\textbf{SQILv2}: Soft Q Imitation Learning \cite{reddy2019sqil}, a recently proposed imitation technique that combines imitation and RL, and works in the absence of rewards.
For high-dimensional environments, we replace DQN with \textbf{PPO}, Proximal Policy Optimization \cite{schulman2017proximal}. Both DQN and PPO are trained to optimize environment reward through experience, and do not have access to other agents' experiences.
\subsection{Environments}
Experiments are conducted using four environments, shown in Figure \ref{fig:envs}. We cover a broad range of problem setting, including both multi-agent and single-agent environments, as well as learning online during RL training, or offline from previously collected demonstrations.
\textbf{Highway}~\citep{highway-env} is a multi-agent autonomous driving environment in which the ego-agent must safely navigate around other cars and reach its goal.
The other agents follow near-optimal scripted policies for various goals, depending on the scenario.
In the \textbf{single-task} scenario, other agents have the same objective as the ego-agent, so their experience is directly relevant.
In the \textbf{adversarial task}, the other agents do not move, and the ego-agent has to accelerate and go to a particular lane while avoiding other vehicles.
Finally, in the \textbf{multi-task} scenario, the other agents and ego-agent have different preferences over target speed, preferred lane, and following distance. We consider the multi-task scenario to be the most realistic and representative of real highway driving with human drivers. In addition to highway driving, we also study the more complex \textbf{Roundabout} task.
Roundabout is inherently multi-task, in that other agents randomly exit either the first or second exit, while the ego-agent must learn to take the third exit.
\textbf{CoinGrid} is a single-agent grid-world, environment containing goals of different colours. We collect offline trajectories of pre-trained agents with preferences for different goals. red.
During training, the ego-agent is only rewarded for collecting a subset of the possible goals. We can then test how well the ego-agent is able to transfer to a goal that was never experienced during training (Section \ref{sec:results_zeroshot}). This environment also enables learning easily interpretable preference vectors, allowing us to visualize how well our method works as an IRL method for inferring rewards (Section \ref{sec:results_irl}).
\textbf{FruitBot} is a high-dimensional, procedurally generated, single-agent environment from the OpenAI ProcGen~\citep{cobbe2020leveraging} suite.
We use FruitBot to test whether $\Psi \Phi$-learning can scale up to more complex RL environments, requiring larger deep neural network architectures that learn directly from pixels.
The agent must navigate around randomly generated obstacles while collecting fruit, and avoiding other objects and walls.
To create a multi-task version of FruitBot, we define additional tasks which vary agents' preferences over collecting objects in the environment, and train PPO baselines on these task variants.
The ego-agent observes the states and actions of these trained agents playing the game in parallel with its own interactions.
\begin{figure*}
\centering
\includegraphics[width=0.6\linewidth]{assets/misc/fruitbot/legend.pdf}
\begin{subfigure}[b]{0.24\linewidth}
\centering
\includegraphics[width=\linewidth]{assets/misc/highway/single-task.pdf}
\caption{Highway: Single-task}
\label{fig:results_curves_singletask}
\end{subfigure}
\begin{subfigure}[b]{0.24\linewidth}
\centering
\includegraphics[width=\linewidth]{assets/misc/highway/adversarial-task.pdf}
\caption{Highway: Adversarial}
\label{fig:results_curves_adversarial}
\end{subfigure}
\begin{subfigure}[b]{0.24\linewidth}
\centering
\includegraphics[width=\linewidth]{assets/misc/highway/multi-task.pdf}
\caption{Highway: Multi-task}
\label{fig:results_curves_multitask}
\end{subfigure}
\begin{subfigure}[b]{0.24\linewidth}
\centering
\includegraphics[width=\linewidth]{assets/misc/fruitbot/training_curves.pdf}
\caption{FruitBot}
\label{fig:results_curves_fruitbot}
\end{subfigure}
\caption{
Learning curves for $\Psi \Phi$-learning and baselines in three tasks in the multi-agent Highway environment (a-c), and in single-agent FruitBot (d).
Tasks (a) and (b) represent extreme cases where either RL or imitation learning is irrelevant.
In (a) other agents have the same task as the ego-agent, so imitation learning excels.
In the adversarial task (b), other agents exhibit degenerative behaviour, so imitation learning performs extremely poorly and traditional RL (DQN) excels.
In both of these extreme cases, $\Psi \Phi$-learning achieves good performance, showing it can flexibly reap the benefits of either imitation or RL as appropriate.
Task (c) is most realistic; here, other agents have varied preferences and goals that may or may not relate to the ego-agent's task.
$\Psi \Phi$-learning clearly outperforms baseline techniques. Similar results are shown in Fruitbot (d), showing that $\Psi \Phi$-learning scales well to high-dimensional environments, consistently outperforming baselines like PPO and SQIL.
We plot mean performance over 3 runs and individual runs with alpha.
}
\label{fig:social_learning}
\vspace{-1.5em}
\end{figure*}
\subsection{Accelerating RL with No-Reward Demonstrations}
\label{sec:results_accelerate_rl}
This section addresses two hypotheses:
\textbf{H1:} When the unlabelled demonstrations are relevant, $\Psi \Phi$-learning can accelerate or improve performance of the ego-agent when learning with online RL; and
\textbf{H2:} If the demonstrations are irrelevant, biased, or are generated by sub-optimal demonstrators, $\Psi \Phi$-learning can perform at least as well as standard RL.
Figure \ref{fig:social_learning} shows the results of $\Psi \Phi$-learning and the baselines in the Highway and FruitBot environments. In the single-task scenario (\ref{fig:results_curves_singletask}), when other agents' experience is entirely relevant to the ego-agent's task, imitation learning methods like BC and SQILv2 learn fastest. DQN learns slowly because it does not use the other agents' experience. However, $\Psi \Phi$-learning achieves competitive results, out-performing DQN+BC-Aux \cite{hernandez2019agent,ndousse2020multi}.
In the adversarial task (\ref{fig:results_curves_adversarial}), the other agents' behaviors are irrelevant for the ego-agent's task, so imitation learning (BC and SQILv2) performs poorly, while traditional RL techniques (DQN and DQN+BC-Aux) perform best.
The performance of $\Psi \Phi$-learning does not suffer like other imitation learning methods; instead, it retains the performance of standard RL (\textbf{H2}).
$\Psi \Phi$-learning can flexibly reap the benefits of either imitation learning or RL, depending on what is most beneficial for the task.
The multi-task scenario (\ref{fig:results_curves_multitask}) is the most realistic autonomous driving task, in which other agents navigate the highway with varying driving styles.
Here, $\Psi \Phi$-learning clearly out-performs all other methods, suggesting it can leverage information about other agents' preferences in order to learn the underlying task structure of the environment, acclerating performance on the ego-agent's RL task (\textbf{H1)}. FruitBot (\ref{fig:results_curves_fruitbot}) gives consistent results, showing that $\Psi \Phi$-learning scales well to high-dimensional, single-agent tasks while still outperforming BC, SQIL, PPO, and PPO+BC-Aux.
\subsection{Inverse Reinforcement Learning}
\label{sec:results_irl}
We now test hypothesis \textbf{H3:} ITD is an effective IRL method, and can accurately infer other agents' rewards.
We present a quantitative and qualitative study of the rewards for other agents that are inferred by ITD, as well as the learned cumulants and preferences.
Here, we focus solely on offline IRL and use only ITD to learn from offline reward-free demonstrations, without any ego-agent experience.
\begin{table}
\centering
\caption{
We evaluate how well $\Psi \Phi$-learning is able to infer the correct reward function by training an RL agent on the inferred rewards, and comparing this to alternative imitation learning methods in three environments.
All methods are trained on expert demonstrations.
A ``$\diamondsuit$'' indicates methods that infer an \emph{explicit} reward function and then use one of DQN or PPO to train an RL agent, depending on the environment.
A ``$\clubsuit$'' indicates methods that directly learn a policy from demonstrations.
A ``$\dagger$'' indicates methods that use privileged task id information for handling multi-task demonstrations.
We report mean and standard error of \emph{normalised returns} over 3 runs, where higher-is-better and the performance is upper bounded by $1.0$, reached by the same RL agent, trained with the ground truth reward function.
}
\label{tab:irl}
\resizebox{\linewidth}{!}{
\begin{tabular}{lc|c|c}
\toprule
\textbf{Methods} & \texttt{Roundabout}$^{\texttt{DQN}}$ & CoinGrid$^{\texttt{DQN}}$ & \texttt{FruitBot}$^{\texttt{PPO}}$ \\
\midrule
BC$^{\dagger \clubsuit}$~\citep{pomerleau1989alvinn} &
$0.81{\color{black!50}\pm0.02}$ &
$0.69{\color{black!50}\pm0.06}$ &
\fontseries{b}\selectfont 0.37${\color{black!50}\pm0.02}$ \\
SQIL$^{\dagger \clubsuit}$~\citep{reddy2019sqil} &
$0.85{\color{black!50}\pm0.02}$ &
$0.64{\color{black!50}\pm0.05}$ &
\fontseries{b}\selectfont 0.35${\color{black!50}\pm0.03}$ \\
\midrule
GAIL$^{\dagger \diamondsuit}$~\citep{ho2016generative} &
$0.77{\color{black!50}\pm0.07}$ &
$0.73{\color{black!50}\pm0.02}$ &
$0.31{\color{black!50}\pm0.02}$ \\
\rowcolor{ourmethod} ITD$^{\diamondsuit}$ (ours, cf. Section~\ref{subsec:inverse-temporal-difference-learning}) &
\fontseries{b}\selectfont 0.92${\color{black!50}\pm0.01}$ &
\fontseries{b}\selectfont 0.77${\color{black!50}\pm0.03}$ &
\fontseries{b}\selectfont 0.35${\color{black!50}\pm0.04}$\\
\bottomrule
\end{tabular}
}
\vspace{-0.5em}
\end{table}
To quantitatively evaluate how well ITD can infer rewards, we train an RL agent on the inferred reward function, and compare the performance to other imitation learning and IRL methods. Table \ref{tab:irl} gives the performance in terms of normalised returns on all three environments.
Using ITD to infer rewards results in significantly higher performance than BC and SQIL, in two environments, and competitive performance in FruitBot. We note that unlike $\Psi \Phi$-learning, BC and SQIL directly learn a policy from demonstrations, and do not actually infer an explicit reward function. In contrast, GAIL does infer an explicit reward function, and ITD gives consistently higher performance than GAIL in all three environments. These results demonstrate that ITD is an effective IRL technique
(\textbf{H3}).
Qualitatively, we can evaluate how well the cumulants inferred by ITD in the CoinGrid environment span the space of possible goals.
We compute the learned cumulants $\hat{\phi}(s)$ for each square $s$ in the grid.
Figure \ref{fig:cumulants_coingrid} shows the original CoinGrid game, and Figure \ref{fig:cumulants1}-\ref{fig:cumulants3}
shows the first three dimensions of the learned cumulant vector, $\hat{\phi}_1$-$\hat{\phi}_3$ (the rest are given in the Appendix). We find that $\hat{\phi}_1$ is most active for red coins, $\hat{\phi}_2$ for green, and $\hat{\phi}_3$ for yellow.
Clearly, ITD has learned cumulant features that span the space of goals for this game.
See Appendix~\ref{app:visualisations} for more details and visualisations of the learned rewards and preferences.
\begin{figure}[!h]
\centering
\begin{subfigure}[b]{0.24\linewidth}
\centering
\includegraphics[width=\textwidth]{assets/misc/coingrid/board.pdf}
\caption{CoinGrid}
\label{fig:cumulants_coingrid}
\end{subfigure}%
\begin{subfigure}[b]{0.24\linewidth}
\centering
\includegraphics[width=\textwidth]{assets/misc/coingrid/phi_1.pdf}
\caption{$\phi_{1}$}
\label{fig:cumulants1}
\end{subfigure}
\begin{subfigure}[b]{0.24\linewidth}
\centering
\includegraphics[width=\textwidth]{assets/misc/coingrid/phi_2.pdf}
\caption{$\phi_{2}$}
\end{subfigure}
\begin{subfigure}[b]{0.24\linewidth}
\centering
\includegraphics[width=\textwidth]{assets/misc/coingrid/phi_3.pdf}
\caption{$\phi_{3}$}
\label{fig:cumulants3}
\end{subfigure}
\label{fig:irl_fourrooms}
\caption{
Qualitative evaluation of the learned cumulants in the CoinGrid task.
Cumulants $\phi_{1}$, $\phi_{2}$, and $\phi_{3}$ seem to capture the red, green, and yellow blocks, respectively.
Therefore, linear combinations of the learned cumulants can represent arbitrary rewards in the environment, which involve stepping on the coloured blocks.
}
\label{fig:cumulants}
\end{figure}
\subsection{Imitation learning}
\label{sec:results_imitate}
Here, we investigate hypothesis \textbf{H4:} $\Psi \Phi$-learning works as an effective imitation learning method, allowing for accurate prediction of other agents' actions. To test this hypothesis, we train other agents in the Roundabout environment, then split no-reward demonstrations from these agents into a train dataset (80\%) and a held-out test dataset. We use the train dataset to run ITD, which means that we use the data to learn both $\Phi$ and the $\Psi$ and $\mathbf{w}$ for other agents. Because it is specifically designed to accurately predict other agents' actions, we use BC as the baseline. We compare this to using only ITD, and using the full $\Psi \Phi$-learning algorithm including ITD \textit{and} learning from RL and experience to update the shared cumulants $\Phi$.
Accuracy in predicting other agents' actions on the held-out test set is used to measure imitation learning performance.
Figure \ref{fig:predict_others_actions} shows accuracy over the course of training.
At each phase change marked in the figure, the ego-agent is given a new task, to test how the representation learning benefits from diverse ego-experience.
We see that although BC obtains accurate train performance, it generalises poorly to the test set, reaching little over 80\% accuracy.
Without RL, $\Psi \Phi$-learning achieves similar performance. However, when using RL to improve imitation, $\Psi \Phi$-learning performs well on both the train and test set, achieving markedly higher accuracy ($\approx 95$\%) in predicting other agents' behaviour. This suggests that when $\Psi \Phi$-learning uses RL and interaction with the world to improve the estimation of the shared cumulants $\Phi$, this in turn improves its ability to model the $Q$ function of other agents and predict their behaviour. Further, $\Psi \Phi$-learning adapts well when the agent's goal changes, since it uses SFs to disentangle the representation of an agent's goal from environment dynamics. Taken together, these results demonstrate that $\Psi \Phi$-learning also works as a competitive imitation learning method (\textbf{H4}).
\begin{table*}[!h]
\centering
\caption{
We evaluate how well $\Psi \Phi$-learning is able to transfer to new tasks in a few-shot fashion.
We construct a multi-task variant of the CoinGrid environment: The ego-agent is provided demonstrations for either capturing only red coins \texttt{R} or only green coins \texttt{G}.
Then it is evaluated on 4 different tasks: collecting (i) both red and green coins \texttt{R+G}, (ii) collecting red and avoiding green coins \texttt{R-G}, (iii) avoiding red and collecting green coins \texttt{-R+G} and (iv) avoiding both red and green coins \texttt{-R-G}.
A ``$\diamondsuit$'' indicates methods that use a single model for all tasks, while ``$\clubsuit$'' indicates methods that require one model per task, i.e., they comprise of 4 models. Because it disentangles preferences from task representation, $\Psi \Phi$-learning is able to adapt to reach optimal performance on the new tasks after a single episode or improve intra-episode from the first episode after experiencing the first rewards. In contrast, SQIL takes $100$ episodes to adapt.
}
\label{tab:few-shot}
\resizebox{\linewidth}{!}{
\begin{tabular}{lrrrr|rrrr|rrrr}
\toprule
&
\multicolumn{4}{c}{0-shot} &
\multicolumn{4}{c}{1-shot} &
\multicolumn{4}{c}{100-shot} \\
\cline{2-5} \cline{6-9} \cline{10-13} \\
%
\textbf{Methods} &
\texttt{R+G} &
\texttt{R-G} &
\texttt{-R+G} &
\texttt{-R-G} &
\texttt{R+G} &
\texttt{R-G} &
\texttt{-R+G} &
\texttt{-R-G} &
\texttt{R+G} &
\texttt{R-G} &
\texttt{-R+G} &
\texttt{-R-G} \\
%
\midrule
SQILv2$^{\clubsuit}$~\citep{reddy2019sqil} &
$1.0{\color{black!50}\pm0.0}$ &
$0.0{\color{black!50}\pm0.0}$ &
$0.0{\color{black!50}\pm0.0}$ &
$-1.0{\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
$0.0{\color{black!50}\pm0.0}$ &
$0.0{\color{black!50}\pm0.0}$ &
$-1.0{\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ \\
\rowcolor{ourmethod} $\Psi \Phi$-learning$^{\diamondsuit}$ (ours, cf. Section~\ref{subsec:psi-phi-learning-with-no-reward-demonstrations}) &
$1.0{\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 0.2${\color{black!50}\pm0.1}$ &
\fontseries{b}\selectfont 0.2${\color{black!50}\pm0.1}$ &
\fontseries{b}\selectfont $-$0.4${\color{black!50}\pm0.2}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ &
\fontseries{b}\selectfont 1.0${\color{black!50}\pm0.0}$ \\
\bottomrule
\end{tabular}
}
\end{table*}
\begin{figure}[h]
\centering
\begin{subfigure}[c]{0.63\linewidth}
\centering
\includegraphics[width=\textwidth]{assets/misc/highway/roundabout-predict_others_actions.pdf}
\end{subfigure}
\begin{subfigure}[c]{0.35\linewidth}
\centering
\includegraphics[width=\textwidth]{assets/misc/highway/roundabout-predict_others_actions-legend.pdf}
\end{subfigure}
\caption{
Test accuracy in predicting other agents' actions.
The shared cumulants $\Phi$ for modelling others- and ego- reward functions allow our $\Psi \Phi$-learner to improve its ability to predict others' actions by experiencing new ego-tasks.
Pure imitation learning and our ITD inverse RL methods achieve high train accuracy but they do not have a mechanism for utilising RL experience to improve their generalisation to the test set as new tasks are provided.
}
\label{fig:predict_others_actions}
\end{figure}
\subsection{Transfer and Few-Shot Generalisation}
\label{sec:results_transfer}
\label{sec:results_zeroshot}
Since SFs have been shown to improve generalization and transfer in RL, here we test hypothesis \textbf{H5:} $\Psi \Phi$-learning will be able to generalize effectively to new tasks in a few-shot transfer setting.
Using the CoinGrid environment, it is possible to precisely test whether the ego-agent can generalise to a task it has never experienced during training. Specifically, we would like to determine whether: (i) the ego-agent can generalise to tasks it was never rewarded for during training (but which it may have seen other agents demonstrate), and (ii) the ego-agent can generalise to tasks not experienced by \emph{any} agents during training.
Table \ref{tab:few-shot} shows the results of transfer experiments in which agents are given 0, 1, or 100 additional training episodes to adapt to a new task. Unlike SQIL, $\Psi \Phi$-learning is able to adapt 0-shot to obtain some reward on the new tasks, and fully adapt after a single episode to achieve the maximum reward on all transfer tasks. This is because $\Psi \Phi$-learning uses SFs to disentangle preferences (goals) in the task representation, and learn about the space of possible preferences from observing other agents.
To adapt to a new task, it need only infer the correct preference vector.
Task inference is trivially implemented as a least squares regression problem, see Eqn.~(\ref{eq:r-loss}): Having experienced $\mathcal{B}^{\text{new}}$ in the new task, the $\Psi \Phi$-learner identifies the preference vector for the new task by solving $\min_{\mathbf{w}} \sum_{\mathcal{B}^{\text{new}}} \L_{\text{R}}(\boldsymbol{\theta}_{\Phi}, \mathbf{w})$.
In contrast, SQIL requires $100$ episodes to reach the same performance.
\section{Related Work}
\label{sec:related-work}
\textbf{Learning from demonstrations.}
~Learning from demonstrations, also referred to as imitation learning~\citep[IL,][]{widrow1964pattern,pomerleau1989alvinn,atkeson1997robot}, is an attractive framework for sequential decision making when reliable, expert demonstrations are available.
Early work on IL assumed access to high-quality demonstrations and aimed to match the expert policy~\citep{pomerleau1991efficient,heskes1998solving,ng2000algorithms,abbeel2004apprenticeship,billard2008survey,argall2009survey,ziebart2008maximum}.
Building on this assumption, many recent works have studied various aspects of both \emph{single-task}~\citep{ratliff2006maximum,wulfmeier2015maximum,choi2011inverse,finn2016guided,ho2016generative,finn2016connection,fu2017learning,zhang2018deep,rahmatizadeh2018vision} and \emph{multi-task}~\citep{,dimitrakakis2011bayesian,mulling2013learning,stulp2013learning,deisenroth2014multi,sharma2018multiple,codevilla2018end,fu2019language,rhinehart2020deep,filos2020can} IL.
However, these methods are all limited by the performance of the demonstrator that they try to imitate.
Learning from suboptimal demonstrations has been studied by~\citet{coates2008learning,grollman2011donut,zheng2014robust,choi2019robust,shiarlis2016inverse,brown2019extrapolating}, enabling, under certain assumptions, imitation learners to surpass their demonstrators' performance.
In contrast, our method integrates demonstrations into an online reinforcement learning pipeline and can use the demonstrations to improve learning on a new task.
Our inverse TD (ITD) learning offline multi-task inverse reinforcement learning algorithm is similar to the Cascaded Supervised IRL (CSI) approach \citep{klein2013cascaded}.
However, CSI assumes a single-task, deterministic expert while ITD does not.
\textbf{Reinforcement learning with demonstrations.}
~Demonstration trajectories have been used to accelerate the learning of RL agents~\citep{taylor2011integrating,vecerik2017leveraging,rajeswaran2017learning,hester2018deep,gao2018reinforcement,nair2018overcoming,paine2018one,paine2019making}, as well as demonstrations where actions and/or rewards are unknown~\citep{borsa2017observational,torabi2018behavioral,sermanet2018time,liu2018imitation,aytar2018playing,brown2019extrapolating}.
In contrast to the standard imitation learning setup, these methods allow improving over the expert performance as the policy can be further fine-tuned via reinforcement learning.
Offline reinforcement learning with online fine-tuning~\citep{kalashnikov2018scalable,levine2020offline} can be framed under this settings too.
Our method builds on the same principles, however, unlike these works, we do not assume that the demonstration data either come with reward annotations, or that they relate to the same task the RL agent is learning (i.e., we learn from multi-task demonstrations which may include irrelevant tasks).
\textbf{Successor features.}
~Successor features (SFs) are a generalisation of the successor representation~\citep{dayan1993improving} for continuous state and action spaces~\citep{barreto2017successor}. Prior work has used SFs for (i) zero-shot transfer~\citep{barreto2017successor,borsa2018universal,barreto2020fast}; (ii) exploration~\citep{janz2019successor,machado2020count}; (iii) skills discovery~\citep{machado2017eigenoption,hansen2019fast}; (iv) hierarchical RL~\citep{barreto2019option} and theory of mind~\citep{rabinowitz2018machine}.
Nonetheless, in all the aforementioned settings, direct access to the rewards or cumulants was provided.
Our method, instead, uses demonstrations without reward labels for inferring the cumulants and learning the corresponding SFs.
More closely to this work,~\citet{lee2019truly} propose learning cumulants and successor features for a \emph{single-task} IRL setting.
Their approach differs from ours in two key respects:
first, they use a learned dynamics model to learn the cumulants. Second, the learned SFs are used for representing the action-value function, not to inform the behaviour policy with GPI.
\textbf{Model of others in multi-agent learning.}
~Our method draws inspiration and builds on the multi-agent learning setting, where multiple agents participate in the same environment and the states, actions of others are observed~\citep{davidson1999using,lockett2007evolving,he2016opponent,jaques2019social}.
However, we do not explore \emph{strategic} settings, where recursive reasoning~\citep{stahl1993evolution,yoshida2008game} is necessary for optimal behaviour.
\section{Discussion}
\label{sec:discussion}
We have presented two major algorithmic contributions.
The first, ITD, is a novel and flexible offline IRL algorithm that discovers salient task-agnostic environment features in the form of cumulants, as well as learning successor features and preference vectors for each agent which provides demonstrations.
The second, $\Psi \Phi$-learning, combines ITD with RL from online experience. This makes efficient use of unlabelled demonstrations to accelerate RL, and comes with theoretical worst-case performance guarantees.
We showed empirically the advantages of these algorithms over various baselines: how imitation with ITD can improve RL and enable zero-shot transfer to new tasks, and how experience from online RL can help to improve imitation in turn.
\textbf{Future Work.}
~We want to explore ways to: (i) adapt $\Psi \Phi$-learning to multi-agent \emph{strategic} settings, where coordination and opponent modelling~\citep{albrecht2018autonomous} are essential and (ii) use a universal successor features approximator~\citep{borsa2018universal} for ITD, overcoming its current, linear scaling with the number of distinct demonstrators.
\textbf{Acknowledgements.}
~We thank Pablo Samuel Castro, Anna Harutyunyan, RAIL, OATML and IRIS lab members for their helpful feedback.
We also thank the anonymous reviewers for useful comments during the review process.
A.F. is funded by a J.P.Morgan PhD Fellowship and C.L. is funded by an Open Phil AI Fellowship.
|
\section{Introduction.}
\begin{figure*}[t!]
\centering
\includegraphics[width=7cm]{FIG1.eps}
\caption{Schematics of the anharmonicity in materials using a diatomic pair potential model, where the lattice vibrational potential energy as a function of interatomic distance is displayed, whereas dashed line indicates harmonic oscillation, cyan and violet curves are for anharmonic vibrations. The strength of the anharmonicity is quantified by the deviation of the potential energy from the harmonic one. In general, when atom is off equilibrium, the slower increase of the potential energy, the larger of the anharmonicity is in the system.}
\label{Model}
\end{figure*}
Copper-based materials often exhibit some unusual physical properties such as superconductivity\cite{Giustino2008,Bednorz1986}, superionicity\cite{Niedziela2019}, semimetallicity\cite{Tang2016}, antiferromagnetism\cite{Wadley2016}, transparent conductivity\cite{Nie2002}, and outstanding thermoelectricity\cite{Shi2012}, which are of particular importance in the fields of either condensed matter or energy engineering. For instance, thermoelectric devices that transform waste heat into useful electricity require building blocks with ultralow thermal conductivity ($\kappa$)\cite{Zhao2014,He2017,Mukhopadhyay2017} as well as high electrical conductivity. Liu $\it et~al.$ has reported experimentally that copper chalcogenides possess very low thermal conductivity with $\kappa<$1~$\rm W m^{-1} K^{-1}$ at room temperature\cite{Shi2012}. The lattice thermal conductivities of, e.g., $\rm Cu_2 S$\cite{He2014}, $\rm Cu_2 Se$\cite{Shi2012,Byeon2019}, CuCl\cite{Mukhopadhyay2018}, $\rm CuBiS_2$ and $\rm CuMX_2$ (M=Sb, Cr, and X=S, Se)\cite{Niedziela2019,Du2017,Zhang2016,Feng2017} are approximately two-orders of magnitude smaller than their adjacent compounds like GaP, GaAs, ZnS, ZnSe and NaCl in the periodic table\cite{Spitzer1970} and, interestingly, one order smaller than the leading thermoelectric materials such as PbTe\cite{Delaire2011}. The unusual intrinsic low thermal conductivity endows copper compounds as promising candidates for thermoelectric devices.
Traditionally, lower thermal conductivities are observed in materials with complex structures, heavy atoms or disordered arrangements in the unit cell\cite{Slack1973,Chiritescu2007,Snyder2008,Christensen2008,Poudel2008,Bera2010,Biswas2012,Voneshen2013,Katre2017,BLi2018,Morelli2008} in order to achieve small phonon group velocity and short mean free path. These traditional criteria clearly are unable to judge the thermal conductivity of copper-based compounds because they are all crystalline semiconductors with simple periodic structure and composed by relatively light chemical elements such as CuCl or $\rm Cu_2 S$. Liu $\it et~al.$ ascribed the low $\kappa$ of copper-based materials to the copper ion's liquid-like vibrations, which result in strong phonon scattering and some of the vibrational modes are completely suppressed\cite{Shi2012,He2014}. However, understanding the physical origin of the liquid-like vibrations remains ambiguous up to now, which is crucial for thermal management applications. Regarding the lattice dynamic theory\cite{Ziman1960,Mahan2000}, atomic vibrations in solids suffer from intrinsic scattering processes mainly by geometric boundary or scattering by other phonons. If the forces between atoms were purely harmonic, there would be no mechanism for collisions between different phonons. Therefore, the phonon scattering effects that are responsible for the low thermal conductivity here are mainly due to the large anharmonicity of the interatomic potential energy\cite{Slack1982,Mukhopadhyay2017,Shi2012,He2014}. In general, the anharmonicity is quantified by the deviation of the potential energy from the harmonic one, and the slower increase of the potential energy, the larger of the anharmonicity is in the system (see Fig.~1). Thereby, uncovering the nature of the potential energy in heated materials is the key. As for semiconductors, in addition to the ion-involved Coulomb interactions, the electronic coupling contributes significantly to the potential energy. For example, in PbTe, it has been demonstrated that the electronic configuration coined as resonant bonding is the fundamental mechanism of the low thermal conductivity\cite{Delaire2011,Lee2014,Li2015,Shportko2008}. However, for copper-based semiconductors, the correlation between electronic states and liquid-like copper sublattice vibrations (thus anharmonic phonon scattering) has not been established.
In this work, we reveal that the symmetry controlled strong {\it s-d} coupling is fundamentally responsible for the giant phonon anharmonicity in copper-based semiconductors. Cu has a highest occupied {\it 3d} orbital relative to other elements, yielding an extremely strong coupling between the high-lying copper occupied {\it 3d} state and the lowest conduction band state (an admixture of anion's {\it s} and copper's {\it 4s} orbitals), when atoms vibrate away from equilibrium. The remarkable red shift of phonon vibrational frequency with increasing temperature observed in the Raman spectra measurements experimentally confirms our theoretical expectation. This study reshapes the understanding of heat conduction in copper-based compounds and sheds light on the materials design that depends on engineering electron-phonon coupling.
\begin{figure*}[t!]
\centering
\includegraphics[width=12cm]{FIG2.eps}
\caption{Evaluation of lattice thermal conductivity and anharmonic phonon scattering. $\textbf{a}$, First-principles calculated thermal conductivity $\kappa$ (solid curves) as a function of temperature for $\rm Cu_2 Se$, $\rm Cu_2 S$, CuCl, PbTe, NaCl, ZnSe, GaAs and Si, while all experimental data (sparse dots) are taken from Refs.~[\citeonline{Slack1982,Steigmeier1966,Slack1972,Glassbrenner1964,Lu2018,Hakansson1986}], when available. For each material, we choose same color but different symbol to distinguish theoretical and experimental results. $\textbf{b}$, Phonon scattering rate $\gamma$ as a function of wave number for all investigated materials, where we have taken their averaged value of acoustic and optic branches as defined in Eq.~(\ref{meanrate}).}
\label{Conductivity}
\end{figure*}
Fig.~2a shows the comparison of the intrinsic lattice thermal conductivity of several prototypical crystalline semiconductors as a function of temperature. We calculate the temperature-dependent lattice thermal conductivity using the first-principles density functional theory (DFT) combined with the Boltzmann transport equation approach as implemented in VASP and Phono3py packages\cite{Dreizler1990,Kohn1965,Togo2015,Paulatto2015} (computational details are given in Methods and Supplementary Materials). We take $\rm Cu_2 S$, $\rm Cu_2 Se$ and CuCl as examples of copper-based compounds, and show their $\kappa(T)$ (solid line) in comparison with the results of Si, GaAs, ZnSe, NaCl and PbTe. We also show their corresponding experimental data (dots)\cite{Slack1982,Steigmeier1966,Slack1972,Glassbrenner1964,Lu2018,Hakansson1986,Spitzer1970}, when available. It exhibits that theoretical result is in good agreement with experimental data, which is particular true for groups IV, III-V and IIB-VI zinc-blende semiconductors. As temperature rises, the thermal conductivity decreases as expected due to enhanced phonon scattering effect. Moving from Si passing through GaAs, ZnSe, and NaCl, to PbTe, the thermal conductivity decreases monotonically at investigated temperature range. Interestingly, we find that the copper-based compounds exhibit very low thermal conductivity compared to traditional semiconductor materials, no matter how light their chemical component is. The lattice thermal conductivity of $\rm Cu_2 S$ and $\rm Cu_2 Se$ is nearly two orders magnitude smaller than that of adjacent GaAs and ZnSe in periodic table and one order smaller than that of PbTe.
In thermoelectric materials, Lee $\it et~al.$ ascribed the low $\kappa$ of PbTe to the large softening of transverse optic (TO) phonon modes which lead to strong scattering with acoustic phonons\cite{Delaire2011,Lee2014,Li2015,Shportko2008}. In Fig.~2b we show the phonon scattering rates as a function of wave number for all investigated materials. We find that copper-based semiconductors of $\rm Cu_2 S$, $\rm Cu_2 Se$ and CuCl exhibit the largest anharmonic scattering rates. Therefore, using phonon softening model, it is difficult to explain why CuCl has lower thermal conductivity (larger scattering) than PbTe, because the softening of TO modes in CuCl is weaker than that of PbTe (Fig.~S1). Although understanding the temperature dependence of thermal conductivity also requires taking into account changes both in the heat capacity of solids and in the propagation velocities of quasiparticles (Eq.~(\ref{LTC})), the phonon scattering effect dominates the behavior of $\kappa$ of copper compounds between 100 and 500~K compared to other materials (Fig.~S2).
\begin{figure*}[t!]
\centering
\includegraphics[width=12cm]{FIG3.eps}
\caption{Origin of the giant phonon anharmonicity in copper compounds. $\textbf{a}$, Chemical trend of the fully occupied valence {\it d} orbital energy level of elements in the Periodic Table, where all are referred to the same zero energy level at infinity. Obviously, Cu owns the highest valence {\it d} orbital energy level among elements with fully occupied valence {\it d} shell. $\textbf{b}$, First-principles calculated vibrational potential energy as a function of atomic displacement, where the potential energies at equilibrium are set as zero for comparison. We move cations along bond-stretching direction, i.e., Cu, Zn, Ga, or Si,in the copper compounds, ZnSe, GaAs, and Si, respectively (Fig.~S4). Clearly, copper-based compounds exhibit the lowest vibrational potential energy compared to traditional semiconductor materials at a given interatomic spacing. In the inset, we show schematically the band coupling between copper's {\it 3d} and {\it 4s} states. $\textbf{c}$, $\textbf{d}$ and $\textbf{e}$, The vibrational frequency shifts for $\rm Cu_2 S$, CuCl and GaAs, respectively, where $k_z=0$. We defined $\overline{\Delta\omega}=\frac{1}{N_{\nu}} \sum_{\nu}|\omega_{\mathbf{q}\nu}({\rm 300~K})-\omega_{\mathbf{q}\nu}({\rm 0})|$ to characterize the energy changes of phonon branches due to anharmonicity.}
\label{Anharmonicity}
\end{figure*}
To uncover the underlying mechanism why copper-based compounds have the low thermal conductivity, we analyzed the electronic structures and phonon vibrational properties. Compared to traditional semiconductors, the most prominent feature of copper compounds is the high-lying occupied {\it d} orbitals in energy (Fig.~S3). In zinc-blende structure with $T_d$ symmetry point group, the occupied {\it d} orbitals transform to irreducible representations of two-fold degenerate $E$ and three-fold degenerate $T_2$, respectively. Whereas, the high-lying unoccupied {\it s} orbital transforms as $A_1$. Therefore, there is no common symmetry between occupied {\it d} orbitals and unoccupied {\it s} orbital and their coupling is forbidden by symmetry. However, thermal vibrations reduce the crystal symmetry to $C_1$, in which the coupling between occupied {\it d} orbitals and unoccupied {\it s} orbital is allowed since {\it d} orbitals transform to $A$ and {\it s} to $A$. Such {\it s-d} coupling pushes the occupied {\it d} orbital down in energy and makes total energy smaller, which lowers the potential barrier of atomic vibrations. The strength of this allowed {\it s-d} coupling is proportional to the overlap of their wave functions but inversely proportional to the energy separation between occupied {\it d} orbitals and unoccupied {\it s} orbital. Thus, the anharmonicity is expected to enhance as the occupied {\it d} orbital energy increases.
Fig.~3a shows that copper owns the highest occupied {\it 3d} orbital in energy among all elements with fully occupied valence {\it d} states. The atomic energy of {\it 3d} electrons in copper is even higher than that of the outermost valence {\it p} states in chalcogen and halogen elements, but close to its own {\it 4s} state (Table~S1). In copper-based semiconductors such as copper halides, the copper's {\it 3d} orbital dominates the valence band maximum (VBM) state and is below the {\it s}-like conduction band minimum (CBM) by a band gap (about 3-4 eV) (Fig.~S3)\cite{Wei1993,Deng2016,Yang2019}. The small energy difference between copper's {\it 3d} and {\it 4s} states enhances their coupling which lowers the lattice vibrational potential energy. This is consistent with our first-principles calculations as shown in Fig.~3b, where we investigated the lattice vibrational potential energy of solids as a function of displacement of the atoms away from their equilibrium. For considered materials, we displace cations (i.e., Cu, Zn, Ga or Si) in the corresponding compounds along bond-stretching direction (see Fig.~S4). In the inset of Fig.~3b, we show how {\it s-d} coupling effect reduces the system potential energy and thus enhances the phonon anharmonicity. As moving from Cu to Zn and Ga, their occupied {\it 3d} orbital is getting deeper, i.e., away from the VBM (Table~S1), so that contributing less and less to the valence band edges and their {\it s-d} coupling strength is getting weak. It is found that GaAs and ZnSe have vibrational potential energy much higher than that of copper compounds such as $\rm Cu_2 S$ and CuCl, whereas Si exhibits the highest potential energy among them due to absence of the {\it s-d} coupling.
Although coupling between copper's unoccupied {\it 4s} orbital and anion occupied {\it p} state occurs in the copper-based compounds due to symmetry reduction, this effect is weak because (i) the unoccupied {\it 4s} state from cation and occupied {\it p} state from anion are separated in space, different from the {\it s-d} coupling in which both are from same copper ions; (ii) the energy difference between occupied {\it p} state of anion and unoccupied {\it s} state of copper is larger than that of copper's {\it 3d} and {\it 4s} orbitals, leading to the weak coupling strength of the former than the latter.
\begin{figure*}[t!]
\centering
\includegraphics[width=12cm]{FIG4.eps}
\caption{Artificial modulation of s-d coupling effect in the thermal transport of copper-based semiconductors. $\textbf{a}$, Lattice vibrational potential energy of CuCl as a function of copper's displacement along the bond-stretching direction (Fig.~S4) under different Coulomb repulsive energy U, which is applied on Cu {\it 3d} orbitals. $\textbf{b}$ Scattering rate, and $\textbf{c}$ thermal conductivity of CuCl under different U. $\textbf{d}$, Anharmonicity caused frequency shift $\Delta\omega_{\mathbf{q}\nu}$ of optic phonon modes in CuCl between T=300 and 0 K at the center of the Brillouin zone.}
\label{DFTUcalculations}
\end{figure*}
In Figs.~3c and 3d, we investigate the lattice vibrational frequency shifts of $\rm Cu_2 S$ and CuCl due to anharmonic phonon interactions, respectively. To do so, we are employing the density functional perturbation theory as implemented in the Quantum-Espresso package to calculate the temperature dependent phonon frequency, $\Delta\omega_{\mathbf{q}\nu}(T)$, ($\Delta\omega_{\mathbf{q}\nu}(T)=\omega_{\mathbf{q}\nu}(T)-\omega_{\mathbf{q}\nu}(0)$). To reveal the rate of change of phonon energy in different materials, we defined $\overline{\Delta\omega}=\frac{1}{N_{\nu}} \sum_{\nu}|\omega_{\mathbf{q}\nu}(\rm 300 K)-\omega_{\mathbf{q}\nu}(0)|$ with $N_{\nu}$ being the number of phonon branches. As for $\rm Cu_2 S$, we find notable value of $\overline{\Delta\omega}$ near the center of Brillouin zone (Fig.~3c and Fig.~S5), whereas for CuCl, the value is reduced in magnitude (Fig.~3d and Fig.~S6). For comparison, in Fig.~3e, we have also presented $\overline{\Delta\omega}$ of GaAs. Interestingly, we find that the frequency shift of both $\rm Cu_2 S$ and CuCl is remarkable larger than that of GaAs, which has almost zero frequency changes until room temperature due to weak anharmonicity. The large temperature-dependent frequency shift of phonon modes confirms the strong anharmomic lattice dynamics in copper compounds.
Thus far, we have illustrated clearly the origin of the giant phonon anharmonicity in the copper-based semiconductor materials. To further verify our proposed model, in Fig.~4 we performed an artificial modulation by applying an advanced functional, i.e., DFT combined with the Hubbard U model\cite{Hubbard1963,Gao2018} calculations to adjust the {\it s-d} coupling strength to check the changes in the lattice thermal conductivity of copper compounds. According to the above discussion, it is expected that the lattice thermal conductivity will increase if we reduce the s-d coupling effect. Hence, we added a positive Coulomb repulsive energy, U, on Cu {\it 3d} orbital, which can lower the energy level of copper's occupied {\it 3d} orbitals. In this way, we can control the strength of {\it s-d} coupling in the copper-based compounds. As displayed in Fig.~4a, we find that the lattice vibrational potential energy of CuCl increases as U rises, due to the reduced {\it s-d} coupling strength resulting from the lowering of Cu {\it 3d} energy level and accompanied wave function localization\cite{Wei1993,Deng2016,Yang2019} (Fig.~S7). Accordingly, both the scattering rate shown in Fig.~4b and the thermal conductivity shown in Fig.~4c increases due to the reduced anharmonic phonon interactions. Moreover, in Fig.~4d, the calculated vibrational frequency shift of optic phonons at the center of Brillouin zone for CuCl at room temperature compared to low temperature decreases as the copper's {\it 3d} orbital moves away from the VBM. This result further supports that the {\it s-d} coupling is the root of the low thermal conductivity in copper-based materials.
\begin{figure*}[t!]
\centering
\includegraphics[width=16cm]{FIG5.eps}
\caption{Experimental evidence of phonon anharmonicity in copper-based materials. $\textbf{a}$, Wide angle X-ray diffraction patterns were obtained with an ultima IV diffractometer equipped with Cu $\rm K_\alpha$ radiation (40kV, 40mA) over the angle range 20 to 100$\rm ^o$. The XRD spectrum confirms that the CuCl crystal is single phase zinc-blende structure. The inset is the optical microscopy image of CuCl single crystal under a 100$\times$ objective lens. $\textbf{b}$, Raman spectra of CuCl single crystal in temperature range from 4~K to 300~K. From low to high frequency, as Potts $\it et~al.$ pointed out earlier\cite{Potts1974,Fukumoto1973}, $\alpha$ peak is contributed by the combination of longitudinal and transverse acoustic modes, $\beta$ and $\gamma$ bands are mainly from the transverse optic phonon modes, whereas $\delta$ band is due to the vibrations of longitudinal optic modes, respectively. $\textbf{c}$, Raman frequency shift of the four peaks labeled in ($\textbf{b}$). To compare with experimental results, we carried out first-principles calculations of the lattice vibrational frequency at finite temperature, which taking into account the anharmonic phonon interactions (solid lines).}
\label{Ramanmeasurements}
\end{figure*}
To verify the large anharmonicity of copper-based compounds, we also perform temperature-dependent Raman spectra measurements of copper chloride as an example. Fig.~5a shows the X-ray diffraction (XRD) pattern of single crystal CuCl, where sample synthesis and characterization are illustrated in detail in the Methods. We find that XRD shows clearly all the characteristic peaks of zinc-blende structure with space group $\rm F\bar{4}3m$, which confirms the sample quality of CuCl single crystal. In Fig.~5b we present Raman measured vibrational spectra in a wide temperature range from 4~K to 300~K. Interestingly, four pronounced lattice vibrational peaks labeled by $\alpha$, $\beta$, $\gamma$ and $\delta$ corresponding to the phonon energy of $\omega\approx$ ${\rm 65~cm^{-1}}$, ${\rm 155~cm^{-1}}$, ${\rm 166~cm^{-1}}$ and ${\rm 209~cm^{-1}}$ are identified, respectively. Earlier, Potts $\it et~al.$ pointed out that the symmetric $\delta$ peak is from the zone-center vibrations of longitudinal optic modes, whereas $\alpha$ band is a combination of the longitudinal and transverse acoustic modes\cite{Potts1974,Fukumoto1973}; as to the $\beta$ and $\gamma$ bands, it is found that they are dominated by the transverse optic phonon modes vibrations, respectively.
In Fig.~5c we show the Raman frequency shift of the four vibrational peaks indicated in 4b. As temperature rises, all peaks shift to the lower frequency regime, and a remarkable energy reduction for $\alpha$ and $\beta$ bands occurs. This suggests the significant softening of the phonon modes and confirms the giant phonon anharmonicity in copper-based compounds. Our first-principles simulations (solid lines) are consistent with the experimental observations.
Besides the binary copper compounds, recently, it was found that few ternary copper-based materials also exhibit very low thermal conductivity (Table~S2), such as $\rm CuBiS_2$\cite{Feng2017}, $\rm CuSbX_2$\cite{Du2017,Zhang2016} and $\rm CuCrX_2$ (X=S or Se)\cite{Niedziela2019,Bhattacharya2013,Xia2020}. Particularly, Niedziela $\it et~al.$ demonstrated that $\rm CuCrSe_2$ has liquid-like thermal diffusive behavior at a relative high temperature, causing a large lattice anharmonicity due to copper ions dominated phonon modes breaking down in the low-energy regime\cite{Niedziela2019}. Applying above theory, it is easy to understand this behavior, which is caused mainly by the strong band coupling between copper ion's unoccupied {\it 4s} and occupied {\it 3d} states. This study indicates the prominent role of copper ions in the heat conduction in this class of materials.
Using combined first-principles calculations and experimental measurements, we investigated the fundamental origin of the ultralow thermal conductivity observed in the copper-based semiconductor materials. We revealed that the symmetry controlled coupling between copper's occupied {\it 3d} and unoccupied {\it 4s} orbitals lowers the lattice vibrational potential and promotes interactions of the collective oscillations. Therefore, the anharmonic scattering between phonons is enhanced which suppresses the crystalline thermal conductivity in copper-based compounds. This study, connecting the electronic structures and phonon vibrational modes, offers a new insight to understand the behavior of thermal transport in semiconductor materials.
\paragraph{References}
|
\section{0pt}{12pt plus 4pt minus 2pt}{6pt plus 2pt minus 2pt}
\titlespacing\subsection{0pt}{12pt plus 4pt minus 2pt}{3pt plus 2pt minus 3pt}
\titlespacing\subsubsection{0pt}{12pt plus 4pt minus 2pt}{0pt plus 2pt minus 3pt}
\singlespacing
\def\boxit#1{\vbox{\hrule\hbox{\vrule\kern6pt
\vbox{\kern6pt#1\kern6pt}\kern6pt\vrule}\hrule}}
\def\bfred#1{{\color{red}\bf#1}}
\def\bfblue#1{{\color{blue}\bf#1}}
\def\red#1{{\color{red}#1}}
\def\blue#1{{\color{blue}#1}}
\definecolor{orange}{rgb}{1,0.5,0}
\definecolor{MyDarkBlue}{rgb}{0,0.08,0.45}
\def\orange#1{{\color{orange}#1}}
\def\fredcomment#1{\vskip 2mm\boxit{\vskip 2mm{\color{MyDarkBlue}\bf#1}
{\color{MyDarkBlue}\bf -- Fred\vskip 2mm}}\vskip 2mm}
\def\alexcomment#1{\vskip 2mm\boxit{\vskip 2mm{\color{red}\bf#1}
{\color{red}\bf -- Alex\vskip 2mm}}\vskip 2mm}
\newtheorem{corollary}{Corollary}[section]
\newtheorem{definition}{Definition}
\newtheorem{example}{Example}[section]
\newtheorem{lemma}{Lemma}[section]
\newtheorem{proposition}{Proposition}[section]
\newtheorem{remark}{Remark}[section]
\newtheorem{theorem}{Theorem}[section]
\newtheorem{problem}{Problem}[section]
\newtheorem{assumption}{Assumption}[section]
\newtheorem{Def}{Definition}[section]
\renewcommand{\baselinestretch}{1.5}
\interfootnotelinepenalty=10000
\begin{document}
\title{\Large \bfseries Deep Equal Risk Pricing of Financial Derivatives with Multiple Hedging Instruments
}
\author[a] {Alexandre Carbonneau\thanks{Corresponding author\vspace{0.2em}. \newline
{\it Email addresses:} \href{mailto:<EMAIL>@mail.concordia.ca} (Alexandre Carbonneau), \href{mailto:<EMAIL>@concordia.ca} (Fr\'ed\'eric Godin).}}
\author[b]{Fr\'ed\'eric Godin}
\affil[a,b]{{\small Concordia University, Department of Mathematics and Statistics, Montr\'eal, Canada}}
\vspace{-10pt}
\date{
\today}
\maketitle \thispagestyle{empty}
\begin{abstract}
\vspace{-5pt}
This paper studies the equal risk pricing (ERP) framework for the valuation of European financial derivatives. This option pricing approach is consistent with global trading strategies by setting the premium as the value such that the residual hedging risk of the long and short positions in the option are equal under optimal hedging. The ERP setup of \cite{marzban2020equal} is considered where residual hedging risk is quantified with convex risk measures. The main objective of this paper is to assess through extensive numerical experiments the impact of including options as hedging instruments within the ERP framework. The reinforcement learning procedure developed in \cite{carbonneau2020equal}, which relies on the deep hedging algorithm of \cite{buehler2019deep}, is applied to numerically solve the global hedging problems by representing trading policies with neural networks. Among other findings, numerical results indicate that in the presence of jump risk, hedging long-term puts with shorter-term options entails a significant decrease of both equal risk prices and market incompleteness as compared to trading only the stock. Monte Carlo experiments demonstrate the potential of ERP as a fair valuation approach providing prices consistent with observable market prices. Analyses exhibit the ability of ERP to span a large interval of prices through the choice of convex risk measures which is close to encompass the variance-optimal premium.
\noindent \textbf{Keywords:} Equal risk pricing, Deep hedging, Convex risk measure, Reinforcement learning.
\end{abstract}
\medskip
\thispagestyle{empty} \vfill \pagebreak
\setcounter{page}{1}
\pagenumbering{roman}
\doublespacing
\setcounter{page}{1}
\pagenumbering{arabic}
\section{Introduction}
In the famous setup of \cite{black1973pricing} and \cite{merton1973theory}, every contingent claim can be perfectly replicated through continuous trading in the underlying stock and a risk-free asset. These markets are said to be \textit{complete}, and derivatives are redundant securities with a unique arbitrage-free price equal to the initial value of the replicating portfolio. However, the gigantic size of the derivatives market demonstrates unequivalently that options are non-redundant and provide additional value above the exclusive trading of the underlying asset from the standpoint of speculation, risk management and arbitraging \citep{hull2003options}.
Such value-added of derivatives in the real world stems from market incompleteness which
arises from several stylized features of market dynamics such as discrete-time trading, equity risk (e.g. jump and volatility risks) and market impact (e.g. trading costs and imperfect liquidity). In contrast to the complete market paradigm, in incomplete markets, the price of a derivative cannot be uniquely specified by
a no-arbitrage argument since perfect replication is not always possible.
The problem of determining the value of a derivative is intrinsically intertwined with its corresponding hedging strategy. On the spectrum of derivative valuation procedures in incomplete markets, one extreme possibility is the so-called \textit{super-hedging strategy}, where the derivative premium is set as the value such that the residual hedging risk of the seller is nullified. However, the super-hedging premium is in general very large and is thus most often deemed impractical \citep{gushchin2002bounds}.
On the other hand, a more reasonable and practical derivative premium entails that some level of risk cannot be hedged away and is thus intrinsic to the contingent claim. An additional layer of complexity to the hedging problem in incomplete markets is in selecting not only the sequence of investments in trading instruments, but also the \textit{category} of hedging instruments in the design of optimal hedges. Indeed, some categories of instruments are more effective to mitigate certain risk factors than others. For instance, it is well-known that in the presence of random jumps, option hedges are much more effective than trading exclusively the underlying stock due to the convex property of derivatives prices (see, for instance, \cite{coleman2007robustly} and \cite{carbonneau2020deep}). More generally, the use of option hedges dampens tail risk stemming from different risk factors (e.g. jump and volatility risks). The focus of this paper lies precisely on studying a derivative valuation approach called \textit{equal risk pricing} (ERP) for pricing
European derivatives consistently with optimal hedging strategies trading in various categories of hedging instruments (e.g. vanilla calls and puts as well as the underlying stock).
The ERP framework introduced by \cite{guo2017equal} determines the \textit{equal risk price} (i.e. the premium) of a financial derivative as the value such that the long and short positions in the contingent claim have the same residual hedging risk under optimal trading strategies. An important application of ERP in the latter paper is for pricing derivatives in the presence of short-selling restrictions for the underlying stock. Various studies have since extended this approach: \cite{ma2019pricing} provide Hamilton-Jacobi-Bellman equations for the optimization problem and establish additional analytical pricing formulas for equal risk prices, \cite{he2020revised} generalize the problem of pricing derivatives with short-selling restriction for the underlying by allowing for short trades in a correlated asset and \cite{alfeus2019empirical} perform an empirical study of equal risk prices when short selling is banned. One crucial pitfall of the \cite{guo2017equal} framework considered in all of the aforementioned papers is that the optimization problem required to be solved for the computation of equal risk prices is very complex. Consequently, closed-form solutions are restricted to very specific setups (e.g. Black-Scholes market) and no numerical scheme has been proposed to account for more realistic market assumptions.
\cite{marzban2020equal} recently extended the ERP framework by considering the use of convex risk measures under the physical measure to quantify residual hedging risk. A major benefit of the ERP setup of the latter paper is that it does not require the specification of an equivalent martingale measure (EMM), which is arbitrary in incomplete markets since there is an infinite set of EMMs \citep{harrison1981martingales}. Also, using convex measures to quantify residual risk is shown in \cite{marzban2020equal} to significantly reduce the complexity of computing equal risk prices; the optimization problem essentially boils down to solving two distinct non-quadratic global hedging problems, one for the long and one for the short position in the option. Dynamic programming equations are provided in \cite{marzban2020equal} for the aforementioned global hedging problems. However, it is well-known that traditional dynamics programming procedures are prone to the curse of dimensionality when the state and action spaces gets too large \citep{powell2009you}. The main objective of this current paper consists in studying the impact of trading different and possibly multiple hedging instruments
on the ERP framework, which thus necessitates large action spaces. Furthermore, a specific focus of this study is on assessing the interplay between different equity risk factors (e.g. jump and volatility risks) and the use of options as hedging instruments. Consequently,
large state spaces are also required to model the dynamics of the underlying stock and to characterize the physical measure dynamics of the implied volatility of options used as hedging instruments. A feasible numerical procedure in high-dimensional state and action spaces is therefore essential to this paper.
\cite{carbonneau2020equal} expanded upon the work of \cite{marzban2020equal} by developing a tractable solution with reinforcement learning to compute equal risk prices in high-dimensional state and action spaces. The approach of the foremost study relies on the deep hedging algorithm of \cite{buehler2019deep} to represent the long and short optimal trading policy with two distinct neural networks. One of the most important benefits of parameterizing trading policies as neural networks is that the computational complexity increases marginally with the dimension of the state and action spaces. \cite{carbonneau2020equal} also introduce novel $\epsilon$-completeness metrics to quantify the level of market incompleteness which will be used throughout this current study. Several papers have studied different aspects of the class of deep hedging algorithms: \cite{buehler2019deep_2} extend upon the work of \cite{buehler2019deep} by hedging path-dependent contingent claims with neural networks, \cite{carbonneau2020deep} presents an extensive benchmarking of global policies parameterized with neural networks to mitigate the risk exposure of very long-term contingent claims, \cite{cao2020discrete} show that the deep hedging algorithm provides good approximations of optimal initial capital investments for variance-optimal hedging problems and \cite{horvath2021deep} deep hedge in a non-Markovian framework with rough volatility models for risky assets.
The main objective of this paper consist in the assessment of the impact of using multiple hedging instruments on the ERP framework through exhaustive numerical experiments. To the best of the authors' knowledge, this is the first study within the ERP literature that considers trades involving options in the design of optimal hedges.
The performance of these numerical experiments heavily relies on the use of reinforcement learning procedures to train neural networks representing trading policies and would be hardly reachable with other numerical methods. The first key contribution of this paper consists in providing a broad analysis of the impact of jump and volatility risks on equal risk prices and on our $\epsilon$-completeness metrics.
These assessments expand upon the work of \cite{carbonneau2020equal} in two ways. First, the latter paper conducted sensitivity analyses of the ERP framework under different risky assets dynamics by trading exclusively with the underlying stock, not with options. However, the use of options as hedging instruments in the presence of such risk factors allows for the mitigation of some portion of unattainable residual risk when trading exclusively with the stock. Second, this current paper examines the sensitivity of equal risk prices and residual hedging risk to different levels of jump and volatility risks through a range of empirically plausible model parameters for asset prices dynamics (e.g. frequent small jumps and rare extreme jumps). The motivation is to provide new qualitative insights into the interrelation of different stylized features of jump and volatility risks on the ERP framework that are more extensive than in previous work studies.
%
The main conclusions of these experiments of pricing $1$-year European puts are summarized below.
\begin{itemize}
\item [1)] In the presence of downward jump risk, numerical values indicate that hedging with options entails significant reduction of both equal risk prices and on the level of market incompleteness as compared to hedging solely with the underlying stock. The latter stems from the fact that while the residual hedging risk of both the long and short positions in $1$-year puts decreases when short-term option trades are used for mitigating the presence of jump risk, a larger decrease is observed for the short position due to jump risk dynamics entailing predominantly negative jumps. These results further demonstrate that options are non-redundant securities as it is the case in the Black-Scholes world.
\item [2)] In the presence of volatility risk, numerical experiments demonstrate that while the use of options as hedging instruments can entail smaller derivative premiums, the impact can also be marginal and is highly sensitive to the moneyness level of the put option being priced as well as to the maturity of the traded options. This observation stems from the fact that contrarily to jump risk, volatility risk impacts both upside and downside risk. Thus, the use of option hedges does not necessarily benefit more the short position with a larger decrease of residual hedging risk as observed in the presence of jump risk.
\item [3)] The average price level of short-term options (i.e. average implied volatility level) used as hedging instruments is effectively reflected into the equal risk price of longer-term options. This demonstrates the potential of the ERP framework as a fair valuation approach consistent with observable market prices, which could be used, for instance, to price over-the-counter or long-term less liquid derivatives with short-term highly liquid options.
\end{itemize}
The last contribution of this paper is in benchmarking equal risk prices to derivative premiums obtained with \textit{variance-optimal hedging} \citep{schweizer1995variance}. Variance-optimal hedging procedures solve jointly for the initial capital investment and a self-financing strategy minimizing the expected squared hedging error. The optimized initial capital investment can be viewed as the production cost of the derivative, since the resulting dynamic trading strategy replicates the derivative's payoff as closely as possible in a quadratic sense.\footnote{
%
Note that derivatives premiums prescribed by variance-optimal procedures coincide with the risk-neutral price obtained under the so-called \textit{variance-optimal martingale measure} \citep{schweizer1996approximation}.
%
}
The main motivation for these experiments is the popularity of variance-optimal hedging procedures in the literature for pricing derivatives. Furthermore, while these two derivative valuation procedures are both consistent with optimal trading criteria, the underlying global hedging problem of each approach treats hedging shortfall through a radically different scope. Indeed, equal risk prices obtained under the Conditional Value-at-Risk measure with large confidence level values, as considered in this paper, are the result of joint optimizations over hedging decisions to minimize tail risk of hedging shortfalls which penalize mainly (and most often exclusively) hedging losses, not gains. Conversely, variance-optimal procedures penalize equally hedging gains and losses, not solely losses. This benchmarking of equal risk prices to variance-optimal premiums highlights the flexibility of ERP procedures for derivatives valuation through the choice of convex risk measure. Indeed, numerical values show that the range of equal risk prices obtained with several convex measures can be very large and is close to encompass the variance-optimal premium.
The rest of the paper is as follows. \cref{section:equal_risk_pricing_framework} details the equal risk pricing framework considered in this study. \cref{section:neural_network} presents the numerical scheme to solve the optimization problem with the use of neural networks. \cref{section:numerical_results} performs extensive numerical experiments studying the equal risk pricing framework.
\cref{section:conclusion} concludes.
\section{Equal risk pricing framework}
\label{section:equal_risk_pricing_framework}
This section details the equal risk pricing (ERP) framework considered in this paper,
which is an extension of the derivative valuation scheme introduced
in \cite{marzban2020equal}
with the addition of multiple hedging instruments.
\subsection{Specification of the financial market}
The financial market is in discrete-time with a finite time horizon of $T$ years and $N+1$ observation dates characterized by the set $\mathcal{T}:=\{t_n: t_n = n \Delta_N, n=0,\ldots,N\}$ where $\Delta_N:=T/N$. The probability space $(\Omega, \mathbb{P}, \mathcal{F})$ is equipped with the filtration $\mathbb{F}:=\{\mathcal{F}_n\}_{n=0}^{N}$ satisfying the usual conditions, where $\mathcal{F}_n$ contains all information available to market participants at time $t_n$. Assume $\mathcal{F}=\mathcal{F}_{N}$. $\mathbb{P}$ is referred to as the physical probability measure.
%
On each observation date, a total of $D+2$ financial securities can be traded on the market, which includes a risk-free asset, a non-dividend paying stock and $D$ standard European calls and puts on the latter stock whose maturity dates fall within $\mathcal{T}$. Let $\{B_{n}\}_{n=0}^{N}$ be the price process of the risk-free asset, where $B_{n}:=e^{r t_n}$ for $n=0,\ldots,N$ with $r \in \mathbb{R}$ being the annualized continuously compounded risk-free rate.
The definition of the price process for the risky securities
is now outlined. Since some of the tradable options can mature before the final time horizon $T$, the set of options that can be traded at the beginning of two different observation periods could differ. To reflect this modeling feature and properly represent gains of trading strategies, two different stochastic processes are defined, namely the price of tradable assets at the beginning and at the end of each period. First, let $\{\bar{S}_{n}^{(b)}\}_{n=0}^{N}$
be the \textit{beginning-of-period risky price process} whose element $\bar{S}_{n}^{(b)}$ contains the time-$t_n$ price of all risky assets traded at time $t_n$. More precisely, $\bar{S}_{n}^{(b)}:=[S_{n}^{(0,b)}, \ldots, S_{n}^{(D,b)}]$ with $S_{n}^{(0,b)}$ and $S_{n}^{(j,b)}$ respectively being the time-$t_n$ price of the underlying stock and of the $j^{\text{th}}$ option that can be traded at time $t_n$ for $j=1,\ldots,D$. Similarly, let $\{\bar{S}_{n}^{(e)}\}_{n=0}^{N-1}$ be the \textit{end-of-period risky price process} where $\bar{S}_{n}^{(e)}:=[S_{n}^{(0,e)}, \ldots, S_{n}^{(D,e)}]$ with $S_{n}^{(0,e)}$ and $S_{n}^{(j,e)}$ respectively being the time $t_{n+1}$ price of the underlying stock and $j^{\text{th}}$ option that can be traded at time $t_n$.
Since the underlying asset is denoted as the risky asset with index $0$, $S_{n}^{(0,e)}=S_{n+1}^{(0,b)}$ for $n=0,\ldots,N-1$.
Also, if the $j^{\text{th}}$ option that can be traded at $t_n$ matures at time $t_{n+1}$, then $S_n^{(j,e)}$ is the payoff of that option. In that case, $S_{n+1}^{(j,b)}$ is the price of a new contract with the same characteristics in terms of payoff function, moneyness level and time-to-maturity.
Otherwise, $S_{n+1}^{(j,b)}= S_{n}^{(j,e)}$ holds for all time steps and all risky assets (i.e. for $j=0,\ldots,D$ and $n=0,\ldots,N-1$). An implicit assumption stemming from the latter equality is that trading in risky assets does not impact their prices. Moreover, for convenience, it is assumed throughout the current work that only options with a single-period time-to-maturity are traded, i.e. options are traded once and held until expiry.\footnote{
%
Note that the optimization procedure for global policies described in \cref{section:neural_network} can naturally be generalized
for the case of rebalancing multiple times option contracts prior to their expiry.
}
This paper studies the problem of pricing a simple European-type derivative providing a time-$T$ payoff denoted by $\Phi(S_N^{(0,b)}) \geq 0$.\footnote{
%
The derivative valuation approach presented in this paper can easily be adapted for European options whose payoff is of the form $\Phi(S_N^{(0,b)}, Z_N) \geq 0$ with $\{Z_n\}_{n=0}^{N}$ as some $\mathbb{F}$-adapted potentially multidimensional random process encompassing the path-dependence property of the payoff function.
For examples of such exotic derivatives, the reader is referred to \cite{carbonneau2020equal}.
}
For such purposes, the equal risk pricing scheme is considered, which entails optimizing two distinct self-financing dynamic trading strategies separately for both the long and short positions on the derivative, and then determining the premium which equates the residual hedging risk of the two hedged positions.
The mathematical formalism used for trading strategies in the current study is now outlined.
A \textit{trading strategy}
$\{\delta_{n}\}_{n=0}^{N}$ is an $\mathcal{F}$-predictable process\footnote{
%
A process $X=\{X_{n}\}_{n=0}^{N}$ is said to be $\mathcal{F}$-predictable if $X_0$ is $\mathcal{F}_0$-measurable and $X_n$ is $\mathcal{F}_{n-1}$-measurable for $n=1,\ldots,N$.
%
}
where $\delta_{n}:=[\delta_{n}^{(0)},\ldots,\delta_{n}^{(D)}, \delta_{n}^{(B)}]$ with $\delta_{n}^{(B)}$ and $\delta_{n}^{(j)}$, $j=0,\ldots,D$, respectively denoting the number of shares of the risk-free asset and the $j^{\text{th}}$ risky asset traded at time $t_{n-1}$ held in the hedging portfolio throughout the period $(t_{n-1},t_n]$, except for the case $n=0$ which represents the hedging portfolio composition exactly at time $t_0$. The notation $\delta_n^{(0:D)}:=[\delta_n^{(0)},\ldots,\delta_n^{(D)}]$ is used to define the vector containing exclusively positions in the risky assets. Furthermore, the initial capital investment of the trading strategy is always assumed to be completely invested in the risk-free asset, i.e. $\delta_0^{(B)}$ is the initial investment amount and $\delta_0^{(0:D)}:=[0,\ldots,0]$.
In this work, the trading strategies considered to hedge $\Phi$
are obtained through a joint optimization over all trading decisions to minimize
global risk exposure.
Before formally describing the optimization problem, some well-known prerequisites from the mathematical finance literature are now provided; the reader is referred to \cite{lamberton2011introduction} for additional details. Let $\{V_{n}^{\delta}\}_{n=0}^{N}$ be the hedging portfolio value process associated with the trading strategy $\delta$, where $V_n^{\delta}$ is the time-$t_n$ portfolio value prior to rebalancing with $V_0^{\delta} := \delta_0^{(B)}$ and
\begin{align}
V_n^{\delta}:=\delta_n^{(0:D)} \bigcdot \bar{S}_{n-1}^{(e)} + \delta_n^{(B)}B_n, \quad n =1,\ldots, N,\label{eq:ref_hedging_port_val}
\end{align}
where $\bigcdot$ is the dot product operator.\footnote{
%
For $X:=[X_1,\ldots,X_K]$ and $Y:=[Y_1,\ldots,Y_K]$, $X \bigcdot Y:=\sum_{j=1}^{K}X_j Y_j$.
%
} Furthermore, denote as $\{G_{n}^{\delta}\}_{n=0}^{N}$ the discounted gain process associated with $\delta$ where $G_{n}^{\delta}$ is the time-$t_n$ discounted gain prior to rebalancing with $G_{0}^{\delta} := 0$ and
\begin{align}
G_{n}^{\delta}:= \sum_{k=1}^{n} \delta_{k}^{(0:D)} \bigcdot (B_{k}^{-1} \bar{S}_{k-1}^{(e)} - B_{k-1}^{-1} \bar{S}_{k-1}^{(b)}), \quad n = 1,\ldots,N. \label{eq:ref_disc_gain_process_def}
\end{align}
The trading strategies considered in this paper are always \textit{self-financing}: they require no cash infusion nor withdrawal at intermediate times except possibly at the initialization of the strategy. More formally, a trading strategy is said to be self-financing if it is predictable
%
%
and if the following equality holds $\mathbb{P}$-a.s. for $n=0,\ldots,N-1$:
\begin{align}
\delta_{n+1}^{(0:D)} \bigcdot \bar{S}_n^{(b)} + \delta_{n+1}^{(B)}B_n = V_n^{\delta}. \label{eq:ref_self_financing}
\end{align}
Lastly, denote $\Pi$ as the set of accessible trading strategies, which includes all trading strategies that are self-financing and sufficiently well-behaved.
\begin{remark}
It can be shown that $\delta \in \Pi$ is self-financing if and only if $V_{n}^{\delta} = B_n(V_0^{\delta} + G_n^{\delta})$ holds $\mathbb{P}$-a.s. for $n=0,\ldots,N$; see for instance \cite{lamberton2011introduction}. The latter representation of portfolio values implies the following useful recursive equation \eqref{eq:ref_update_rule_port_value} to compute $V_{n}^{\delta}$ for $n=1,\ldots,N$ given $V_0^{\delta}$:
\begin{align}
V_{n}^{\delta} &= B_{n}(V_0^{\delta}+G_{n}^{\delta}) \nonumber
%
\\ &= B_{n}(V_0^{\delta}+G_{n-1}^{\delta}+\delta_{n}^{(0:D)} \bigcdot (B_{n}^{-1} \bar{S}_{n-1}^{(e)} - B_{n-1}^{-1} \bar{S}_{n-1}^{(b)})) \nonumber
\\ &= \frac{B_{n}}{B_{n-1}}V_{n-1}^{\delta} + \delta_{n}^{(0:D)} \bigcdot (\bar{S}_{n-1}^{(e)} - \frac{B_{n}}{B_{n-1}}\bar{S}_{n-1}^{(b)}) \nonumber
%
\\ &= e^{r \Delta_N}V_{n-1}^{\delta} + \delta_{n}^{(0:D)}\bigcdot(\bar{S}_{n-1}^{(e)} - e^{r \Delta_N}\bar{S}_{n-1}^{(b)}). \label{eq:ref_update_rule_port_value}
\end{align}
\end{remark}
\subsection{Equal risk pricing framework}
\label{subsec:ERP_framework}
The financial market setting considered in this paper implies incompleteness stemming from discrete-time trading and equity risk factors (e.g. jump risk and volatility risk).
For the hedger, these many sources of incompleteness entail that most contingent claims are not attainable through dynamic hedging. Following the work of \cite{marzban2020equal} and \cite{carbonneau2020equal}, this study quantifies the level of residual hedging risk with convex risk measures as defined in \cite{follmer2002convex}.
\begin{Def}{(Convex risk measure)}
For a set of random variables $\mathcal{X}$ representing liabilities and $X_1, X_2 \in \mathcal{X}$, $\rho : \mathcal{X} \rightarrow \mathbb{R}$ is a \textit{convex risk measure} if it satisfies the following properties:
\begin{itemize}
\item [0)] Normalized: $\rho(0) = 0$ (empty portfolio has no risk).
\item [1)] Monotonicity: $X_1 \leq X_2 \Longrightarrow \rho(X_1) \leq \rho(X_2)$ (larger liability is riskier).
\item [2)] Translation invariance: for $c \in \mathbb{R}$ and $X \in \mathcal{X}$, $\rho(X+c) = \rho(X) + c$ (borrowing amount $c$ increases the risk by that amount).
\item [3)] Convexity: for $c \in [0,1]$, $\rho(c X_1 + (1-c)X_2) \leq c \rho(X_1) + (1-c)\rho(X_2)$ (diversification does not increase risk).
\end{itemize}
\end{Def}
The hedging problem underlying the ERP framework is now formally defined.
\begin{Def}{(Long- and short-sided risk)}
\label{HedgingProbDef}
For a given convex risk measure $\rho$, define $\epsilon^{(\mathcal{L})}(V_0)$ and $\epsilon^{(\mathcal{S})}(V_0)$ respectively as the measured risk exposure of a long and short position in $\Phi$ under the optimal hedge if the value of the initial hedging portfolio is $V_0 \in \mathbb{R}$:
\begin{align}
\epsilon^{(\mathcal{L})}(V_0) &:= \underset{\delta\in \Pi}{\min} \, \rho \left(-\Phi(S_{N}^{(0,b)}) -B_{N}(V_0 + G_{N}^{\delta})\right), \label{eq:risk_long}
%
\\ \epsilon^{(\mathcal{S})}(V_0) &:= \underset{\delta \in \Pi}{\min} \, \rho \left(\Phi(S_{N}^{(0,b)}) - B_{N}(V_0 + G_{N}^{\delta})\right). \label{eq:risk_short}
\end{align}
\end{Def}
\begin{remark}
\label{remark:inf_vs_min}
As noted in \cite{carbonneau2020equal}, an assumption implicit to \cref{HedgingProbDef} is that the minimum in \eqref{eq:risk_long} or \eqref{eq:risk_short} is indeed attained by some trading strategy, i.e. that the infimum is in fact a minimum.
\end{remark}
Note that the same risk measure $\rho$ is used for both the long and short positions global hedging problems. The rationale for this choice is threefold. First, considering the same convex risk measure for long and short positions
is in line with the trading activities of some market participants that both buy and sell options with no directional view of the market. One example of such participant is a market maker of derivatives which typically expects to make a profit on bid-ask spreads, not by speculating \citep{basak2012dynamic}.
Another motivation for using the same convex measure is for cases where a price quote must be given prior to knowing if the derivative is being purchased or sold. For instance, a client asks his broker to provide a quote for a derivative without revealing his intention of buying or selling the option. A similar argument is made in \cite{bertsimas2001hedging} to motivate the use of a quadratic loss function for hedging shortfalls, which entails the same derivative price for the long and short position. Lastly, as shown in \cite{carbonneau2020equal}, using the same risk measure for both positions guarantees, under some specific conditions, that the ERP derivative premium is arbitrage-free.\footnote{
%
Nevertheless, the authors want to emphasize that the numerical scheme developed in \cref{section:neural_network} for the global hedging problems \eqref{eq:risk_long} and \eqref{eq:risk_short} could easily be extended to include two distinct convex measures respectively for the long and short position hedges (see Remark $3.4$ of \cite{carbonneau2020equal} for additional details).
%
}
It is interesting to note that the translation invariance property of $\rho$ entails that the optimal strategies solving \eqref{eq:risk_long}-\eqref{eq:risk_short}, denoted respectively by $\delta^{(\mathcal{L})}$ and $\delta^{(\mathcal{S})}$, are invariant to the initial capital investment amount $V_0$. The latter significantly enhances the tractability of the solution:
\begin{align}
\delta^{(\mathcal{L})} &:= \underset{\delta\in \Pi}{\argmin} \, \rho \left(-\Phi(S_{N}^{(0,b)}) -B_{N}(V_0 + G_{N}^{\delta})\right) = \underset{\delta\in \Pi}{\argmin} \, \rho \left(-\Phi(S_{N}^{(0,b)}) -B_{N}G_{N}^{\delta}\right), \label{eq:hedging_risk_long}
\\ \delta^{(\mathcal{S})} &:= \underset{\delta \in \Pi}{\argmin} \, \rho \left(\Phi(S_{N}^{(0,b)}) - B_{N}(V_0 + G_{N}^{\delta})\right) = \underset{\delta \in \Pi}{\argmin} \, \rho \left(\Phi(S_{N}^{(0,b)}) - B_{N}G_{N}^{\delta}\right). \label{eq:ref_hedging_risk_short}
%
\end{align}
Based on the aforementioned global hedging problems, the \textit{equal risk price} of a derivative is defined as the initial hedging portfolio value equating the measured risk exposures for both the long and short positions.
\begin{Def}{(Equal risk price)}
\label{def:ERP_option_price}
The equal risk price $C_0^{\star}$ of $\Phi$ is defined as the real number $C_0$ such that
\begin{align}
\epsilon^{(\mathcal{L})}(-C_0) = \epsilon^{(\mathcal{S})}(C_0). \label{eq:ref_ERP_def}
\end{align}
\end{Def}
As shown for instance in \cite{marzban2020equal}, equal risk prices have the following representation
which is used throughout the rest of the paper:
\begin{align}
C_{0}^{\star} = \frac{\epsilon^{(\mathcal{S})}(0) - \epsilon^{(\mathcal{L})}(0)}{2 B_{N}}. \label{eq:ref_equal_risk_prices}
\end{align}
\cite{carbonneau2020equal} introduced the market incompleteness metric $\epsilon^{\star}$ defined as the level of residual risk faced by the hedgers of $\Phi$ if the hedged derivative price is set to $C_0^{\star}$ and optimal trading strategies are used by both the long and short position hedgers:\footnote{
%
The last equality of \eqref{eq:ref_epsilon_def} can easily be obtained with the translation invariance property of $\rho$, see equation (8) of \cite{carbonneau2020equal} for the details.
%
}
\begin{align}
\epsilon^{\star}&:=\epsilon^{(\mathcal{L})}(-C_0^{\star}) = \epsilon^{(\mathcal{S})}(C_0^{\star}) = \frac{\epsilon^{(\mathcal{L})}(0) + \epsilon^{(\mathcal{S})}(0)}{2}. \label{eq:ref_epsilon_def}
\end{align}
%
Consistently with the terminology of \cite{carbonneau2020equal}, $\epsilon^{\star}$ and $\epsilon^{\star}/C_0^{\star}$ are referred respectively as the measured residual risk exposure per derivative contract and per dollar invested. These $\epsilon^{\star}$-metrics will be extensively studied in numerical experiments conducted in \cref{section:numerical_results} to assess, for instance, the impact of the use of options as hedging instruments on the level of market incompleteness.
\section{Deep equal risk pricing}
\label{section:neural_network}
The problem of solving the ERP framework, that is evaluating equal risk prices and $\epsilon$-completeness measures, boils down to the computation of the measured risk exposures $\epsilon^{(\mathcal{S})}(0)$ and $\epsilon^{(\mathcal{L})}(0)$. This section presents a reinforcement learning method to compute such quantities. The approach was first proposed in \cite{carbonneau2020equal} and relies on approximating optimal trading strategies with the \textit{deep hedging} algorithm of \cite{buehler2019deep}
through the representation of
the long and short global trading policy with two distinct neural networks. In its essence, neural networks are a class of composite functions mapping \textit{feature vectors} (i.e. input vectors) to \textit{output vectors} through multiple \textit{hidden layers}, with the latter being functions applying successive
affine and nonlinear transformations to input vectors.
%
In this paper, the type of neural network considered to represent global hedging policies is the \textit{long short-term memory} (LSTM, \cite{hochreiter1997long}). LSTMs belong to the class of \textit{recurrent neural networks} (RNNs, \cite{rumelhart1986learning}), which have self-connections in hidden layers: the output of the time-$t_n$ hidden layer is a function of both the time-$t_n$ feature vector as well as the output of the time-$t_{n-1}$ hidden layer. The periodic computation of long short-term memory neural networks is done with so-called \textit{LSTM cells}, which are similar to but more complex than the typical hidden layer of RNNs.
%
LSTMs have recently been applied with success to approximate global hedging policies in several studies: \cite{buehler2019deep_2}, \cite{cao2020discrete} and \cite{carbonneau2020deep}.
Additional remarks are made in subsequent sections to motivate this choice of neural networks for the specific setup of this paper.
\subsection{Neural networks representing trading policies}
\label{subsec:neural_net_hedging}
The following formally defines the architecture of long-short term memory neural networks. For convenience, a very similar notation for neural networks as the one of \cite{carbonneau2020deep} is used. Note that the time steps of the feature and output vectors coincide with the set of financial market trading dates $\mathcal{T}$. For additional general information about LSTMs, the reader is referred to Chapter $10.10$ of \cite{goodfellow2016deep} and the many references therein.
\begin{Def}{(LSTM)}
\label{def:LSTM}
For $H, d_0, \ldots, d_{H+1} \in \mathbb{N}$, let $F_{\theta}:\mathbb{R}^{N \times d_0} \rightarrow \mathbb{R}^{N \times d_{H+1}}$ be an LSTM which maps the sequence of feature vectors $\{X_{n}\}_{n=0}^{N-1}$ to output vectors $\{Y_{n}\}_{n=0}^{N-1}$ where $X_{n} \in \mathbb{R}^{d_0}$ and $Y_{n} \in \mathbb{R}^{d_{H+1}}$ for $n=0,\ldots,N-1$.
The computation of $Y_n$, the subset of outputs of $F_{\theta}$ associated with time $t_n$, is achieved through $H$ LSTM cells,
each of which
outputs a vector of $d_{j}$ neurons denoted as $h_{n}^{(j)} \in \mathbb{R}^{d_{j} \times 1}$ for $j=1,\ldots,H$. More precisely, the computation applied by the $j^{\text{th}}$ LSTM cell for the time-$t_{n}$ output is as
follows:\footnote{
%
At time $0$ (i.e. $n=0$), the computation of the LSTM cells is the same as in \eqref{eq:ref_LSTM_cell} with $h_{-1}^{(j)}$ and $c_{-1}^{(j)}$ defined as vectors of zeros of dimension $d_{j}$ for $j=1,\ldots,H$.
}
\begin{align}
i_{n}^{(j)} &= \text{sigm}(U_i^{(j)} h_n^{(j-1)} + W_i^{(j)} h_{n-1}^{(j)} + b_i^{(j)}), \nonumber
%
\\ f_{n}^{(j)} &= \text{sigm}(U_f^{(j)} h_n^{(j-1)} + W_f^{(j)} h_{n-1}^{(j)} + b_f^{(j)}), \nonumber
%
\\ o_{n}^{(j)} &= \text{sigm}(U_o^{(j)} h_n^{(j-1)} + W_o^{(j)} h_{n-1}^{(j)} + b_o^{(j)}), \nonumber
%
\\ c_{n}^{(j)} &= f_{n}^{(j)} \odot c_{n-1}^{(j)} + i_{n}^{(j)} \odot \text{tanh}(U_c^{(j)} h_n^{(j-1)} + W_c^{(j)} h_{n-1}^{(j)} + b_c^{(j)}), \nonumber
%
\\ h_{n}^{(j)} &= o_{n}^{(j)} \odot \text{tanh}(c_{n}^{(j)}), \label{eq:ref_LSTM_cell}
\end{align}
where $\odot$ denotes the Hadamard product (the element-wise product), $\text{sigm}(\cdot)$ and $\text{tanh}(\cdot)$ are the sigmoid and hyperbolic tangent functions applied element-wise to each scalar given as input\footnote{
%
For $X:=[X_1,\ldots,X_K]$, $\text{sigm}(X) := \left[\frac{1}{1 + e^{-X_1}}, \ldots, \frac{1}{1 + e^{-X_K}}\right]$ and $\text{tanh}(X) := \left[\frac{e^{X_1}-e^{-X_1}}{e^{X_1}+e^{-X_1}}, \ldots, \frac{e^{X_K}-e^{-X_K}}{e^{X_K}+e^{-X_K}}\right]$.
%
} and
\begin{itemize}
\item $U_i^{(j)}, U_f^{(j)}, U_o^{(j)}, U_c^{(j)} \in \mathbb{R}^{d_j \times d_{j-1}}$, $W_i^{(j)}, W_f^{(j)}, W_o^{(j)}, W_c^{(j)} \in \mathbb{R}^{d_j \times d_{j}}$ and $b_i^{(j)}, b_f^{(j)}, b_o^{(j)}, b_c^{(j)} \in \mathbb{R}^{d_j \times 1}$ for $j=1,\ldots,H$.
\end{itemize}
At each time-step, the input of the first LSTM cell is the feature vector $($i.e. $h_{n}^{(0)}:=X_{n}$$)$ and the final output is an affine transformation of the output of the last LSTM cell:
\begin{align}
Y_{n} = W_{y}h_{n}^{(H)} + b_{y}, \quad n = 0,\ldots,N-1, \label{eq:ref_output_LSTM}
\end{align}
where $W_y \in \mathbb{R}^{d_{H+1} \times d_{H}}$ and $b_y \in \mathbb{R}^{d_{H+1} \times 1}$. Lastly, the set of trainable parameters denoted as $\theta$ consists of all weight matrices and bias vectors:
\begin{align}
\theta := \left\{\{U_i^{(j)}, U_f^{(j)}, U_o^{(j)}, U_c^{(j)}, W_i^{(j)}, W_f^{(j)}, W_o^{(j)}, W_c^{(j)}, b_i^{(j)}, b_f^{(j)}, b_o^{(j)}, b_c^{(j)}\}_{j=1}^{H}, W_y, b_y\right\}. \label{eq:ref_trainable_params}
\end{align}
\end{Def}
In this study, the computation of hedging positions is done through the mapping of a sequence of relevant financial market observations into the periodic number of shares held in each hedging instrument with an LSTM. One of the main objectives of this paper is to analyze the impact of including vanilla options as hedging instruments on the ERP framework.
For the numerical experiments conducted in the subsequent \cref{section:numerical_results}, the hedging instruments consist of either only the underlying asset (without options) or exclusively options (without the underlying asset). The case of using both the stock and options is not considered since the options can always replicate positions in the underlying asset with calls and puts by relying on the put-call parity.
%
In what follows, let $\{X_n\}_{n=0}^{N-1}$ and $\{Y_n\}_{n=0}^{N-1}$ be respectively the sequence of feature vectors and output vectors of an LSTM as in \cref{def:LSTM}. When hedging only with the underlying, the time-$t_n$ feature vector considered is\footnote{
%
The use of $\log(S_{n}^{(0,b)}/K)$ instead of $S_{n}^{(0,b)}$ in feature vectors was found to improve the learning speed of the neural networks (i.e. time taken to find a good set of trainable parameters). Note that log transformation for risky asset prices was also considered in \cite{carbonneau2020deep}, \cite{buehler2019deep} and \cite{buehler2019deep_2}.
%
}
\begin{align}
X_{n} = [\log(S_{n}^{(0,b)}/K), V_{n}^{\delta}, \varphi_n], \quad n =0,\ldots,N-1, \label{eq:ref_feature_vect_stock}
\end{align}
where $K$ is the strike price of $\Phi$ and $\{\varphi_n\}_{n=0}^{N-1}$ is a sequence of additional relevant state variables associated with the dynamics of asset prices. For instance, if the underlying log-returns are modeled with a GARCH process, it is well-known that the bivariate process of the underlying price and the GARCH volatility has the Markov property under $\mathbb{P}$ with respect to the market filtration $\mathbb{F}$. The time-$t_n$ volatility of the GARCH process is thus added to the feature vectors through $\varphi_n$.
Furthermore, in that same case where the underlying stock is considered as the only hedging instrument, the output vectors of the LSTM consist of the number of underlying asset shares to be held in the portfolio for all time steps, i.e. $Y_{n} = \delta_{n+1}^{(0)}$ for $n=0,\ldots,N-1$.
Conversely, when hedging is performed with options as hedging instruments, the \textit{implied volatilities} (IVs) of such options denoted as $\{IV_{n}\}_{n=0}^{N-1}$ are added to feature vectors with $IV_n$ encompassing every implied volatilities needed to price the $D$ options used for hedging:\footnote{
%
Note that the bijection relation between implied volatilities and option prices entails that either values could theoretically be used in feature vectors as one is simply a nonlinear transformation of the other.
%
}
\begin{align}
X_{n} = [\log(S_{n}^{(0,b)}/K), V_{n}^{\delta}, \varphi_n, IV_{n}], \quad n =0,\ldots,N-1. \label{eq:ref_feature_vect_opts}
\end{align}
In that case, the output vectors are the number of option contracts held in the portfolio for the various time steps: $Y_{n}=[\delta_{n+1}^{(1)}, \ldots, \delta_{n+1}^{(D)}]$ for $n=0,\ldots,N-1$. Recall that when options are used as hedging instruments, $\delta_{n+1}^{(0)}=0$ for $n=0,\ldots,N-1$.
\begin{remark}
\label{remark:adding_port_value}
Although the portfolio value $V_{n}^{\delta}$
is in theory
a redundant feature in the context of LSTMs since it can be retrieved as a function of previous times inputs and outputs of the neural network (see \eqref{eq:ref_update_rule_port_value}), incorporating it to feature vectors was found to significantly improve upon the hedging effectiveness of the LSTMs in the numerical experiments conducted in \cref{section:numerical_results}.
\end{remark}
\subsection{Equal risk pricing with neural networks}
\label{subsec:deep_ERP}
To numerically solve the underlying global hedging problems of the ERP framework, \cite{carbonneau2020equal} propose to use two distinct neural networks denoted as $F_{\theta}^{(\mathcal{L})}$ and $F_{\theta}^{(\mathcal{S})}$ to approximate the global trading policies of respectively the long and short positions in $\Phi$. This is the approach considered in the current paper. As illustrated below, the procedure consists in solving the alternative problems of optimizing the neural networks trainable parameters so as to minimize the corresponding hedging shortfall:
\begin{align}
\epsilon^{(\mathcal{L})}(V_0) &\approx \underset{\theta \in \mathbb{R}^{q}}{\min} \, \rho \left(-\Phi(S_{N}^{(0,b)}) -B_{N}(V_0 + G_{N}^{\delta^{(\mathcal{L},\theta)}})\right), \label{eq:risk_long_NNET}
%
\\
\epsilon^{(\mathcal{S})}(V_0) &\approx \underset{\theta \in \mathbb{R}^{q}}{\min} \, \rho \left(\Phi(S_{N}^{(0,b)}) - B_{N}(V_0 + G_{N}^{\delta^{(\mathcal{S},\theta)}})\right), \label{eq:risk_short_NNET}
\end{align}
where $\delta^{(\mathcal{L},\theta)}$ and $\delta^{(\mathcal{S},\theta)}$ are to be understood respectively as the output sequences of $F_{\theta}^{(\mathcal{L})}$ and $F_{\theta}^{(\mathcal{S})}$, and $q \in \mathbb{N}$ is the total number of trainable parameters of $F_{\theta}^{(\mathcal{L})}$ and $F_{\theta}^{(\mathcal{S})}$.
The approximated measured risk exposures obtained through \eqref{eq:risk_long_NNET} and \eqref{eq:risk_short_NNET} are subsequently used to compute equal risk prices and $\epsilon$-completeness measures with \eqref{eq:ref_equal_risk_prices} and \eqref{eq:ref_epsilon_def}.
%
One implicit assumption associated with \eqref{eq:risk_long_NNET} and \eqref{eq:risk_short_NNET} is that the architecture of all neural networks in terms of the number of LSTM cells and neurons per cell is always fixed; the hyperparameter tuning step of the optimization problem is not considered in this paper. \cref{subsec:SGD} that follows presents the procedure considered in this study to optimize the trainable parameters of the LSTMs.
\begin{remark}
\label{remark:why_LSTM}
\cite{carbonneau2020equal} show that when relying on feedforward neural networks (FFNNs\footnote{FFNNs are another class of neural networks which map input vectors into output vectors, in contrast to LSTMs which map input vector sequences to output vector sequences.}) instead of LSTMs, the alternative problems \eqref{eq:risk_long_NNET}-\eqref{eq:risk_short_NNET} allow for arbitrarily precise approximations of the measured risk exposures \eqref{eq:risk_long}-\eqref{eq:risk_short} due to results from \cite{buehler2019deep}. Despite this theoretical ability of FFNNs to approximate arbitrarily well global hedging policies in such context, the authors of the current paper found that LSTMs are able to learn significantly better trading policies than FFNNs in the numerical experiments carried out in \cref{section:numerical_results}, which motivates their use over FFNNs. The theoretical justifications for the outperformance of LSTMs over FFNNs in the financial market settings of this paper are out-of-scope and are left-out as interesting potential future research.
\end{remark}
\subsection{Training neural networks}
\label{subsec:SGD}
The numerical scheme to optimize the trainable parameters of neural networks as entailed by the global hedging optimization problems \eqref{eq:risk_long_NNET}-\eqref{eq:risk_short_NNET} is now described. The procedure first proposed in \cite{buehler2019deep} uses minibatch stochastic gradient descent (SGD) to approximate the gradient of the cost function with Monte Carlo sampling. For convenience, the notation used for the optimization procedure is similar to the one from \cite{carbonneau2020equal}. Without loss of generality, the numerical procedure is only presented for the short measured risk exposure; the corresponding procedure for the long position is simply obtained through modifying the objective function \eqref{eq:ref_cost_func_short} that follows. Let $J:\mathbb{R}^{q} \rightarrow \mathbb{R}$ be the cost function to be minimized for the short position in $\Phi$, where $\theta$ is the set of trainable parameters of $F_{\theta}^{(\mathcal{S})}$:\footnote{
%
Minimizing $J$ with respect to $\theta$ corresponds to the alternative problem \eqref{eq:risk_short_NNET} with zero initial capital. Recall that $\epsilon^{(\mathcal{L})}(0)$ and $\epsilon^{(\mathcal{S})}(0)$ are required for the computation of $C_0^{\star}$ and $\epsilon^{\star}$. Consequently, hedging portfolio values used in LSTM feature vectors are equal to hedging gains, i.e. $V_n^{\delta} = B_nG_n^{\delta}$.
%
}
\begin{align}
J(\theta):= \rho \left(\Phi(S_{N}^{(0,b)}) - B_{N}G_{N}^{\delta^{(\mathcal{S},\theta)}}\right), \quad \theta \in \mathbb{R}^{q}. \label{eq:ref_cost_func_short}
\end{align}
A typical stochastic gradient descent procedure entails adapting the trainable parameters iteratively and incrementally in the opposite direction of the cost function gradient with respect to $\theta$:
\begin{align}
\theta_{j+1} &= \theta_{j} - \eta_j \nabla_{\theta}J(\theta_{j}), \label{eq:ref_SGD_iterative_param}
\end{align}
where $\theta_0$ is the initial values for the trainable parameters,
$\eta_j$ is a small deterministic positive real value commonly called the \textit{learning rate} and $\nabla_{\theta}$ denotes the gradient operator. In the current study, the \textit{Glorot uniform initialization} of \cite{glorot2010understanding} is always used to select initial parameters in $\theta_0$.
Since closed-form solutions for the gradient of the cost function with respect to the trainable parameters are
unavailable
in the general market setting considered in this work, the approach relies instead on Monte Carlo sampling
to provide an estimate. Thus, let $\mathbb{B}_j:=\{\pi_{i,j}\}_{i=1}^{N_{\text{batch}}}$ be a minibatch of simulated hedging errors of size $N_{\text{batch}} \in \mathbb{N}$ where $\pi_{i,j}$ is the $i^{\text{th}}$ simulated hedging error when $\theta = \theta_j$:
\begin{align}
\pi_{i,j}:= \Phi(S_{N,i}^{(0,b)}) - B_{N}G_{N,i}^{\delta^{(\mathcal{S},\theta_j)}}, \label{eq:ref_simulated_hedging_error}
\end{align}
where $S_{N,i}^{(0,b)}$ and $G_{N,i}^{\delta^{(\mathcal{S},\theta_j)}}$ are the $i^{th}$ random realization among the minibatch of the terminal underlying asset price and discounted hedging portfolio gains, respectively.
Furthermore, denote $\hat{\rho}:\mathbb{R}^{N_{\text{batch}}} \rightarrow \mathbb{R}$ as the empirical estimator of $\rho(\Phi(S_{N}^{(0,b)}) - B_{N}G_{N}^{\delta^{(\mathcal{S},\theta)}})$ evaluated with minibatches of hedging errors.
Minibatch SGD consists in approximating the gradient of the cost function $\nabla_{\theta}J(\theta_{j})$ with $\nabla_{\theta}\hat{\rho}(\mathbb{B}_j)$ in the update rule for trainable parameters:
\begin{align}
\theta_{j+1} &= \theta_{j} - \eta_j \nabla_{\theta}\hat{\rho}(\mathbb{B}_j). \label{eq:ref_SGD_iterative_param_approx}
\end{align}
For the numerical experiments conducted in \cref{section:numerical_results}, the convex risk measure considered is the Conditional Value-at-Risk (CVaR, \cite{rockafellar2002conditional}). For an absolutely continuous integrable random variable\footnote{
%
In \cref{section:numerical_results}, the only dynamics considered for the risky assets produce integrable and absolutely continuous hedging errors.
%
}, the CVaR has the representation
\begin{align}
\text{CVaR}_{\alpha}(X):=\mathbb{E}[X|X \geq \text{VaR}_{\alpha}(X)], \quad \alpha \in (0,1), \label{eq:ref_CVaR_def}
\end{align}
where $\text{VaR}_{\alpha}(X) := \min \left\{x : \mathbb{P}(X \leq x) \geq \alpha \right\}$ is the Value-at-Risk (VaR) with confidence level $\alpha$ of the liability $X$.
Let $\{\pi_{[i],j}\}_{i=1}^{N_{\text{batch}}}$ be the order statistics (i.e. values sorted by increasing order) of $\mathbb{B}_{j}$. For $\tilde{N}:= \ceil*{\alpha N_{\text{batch}}}$ where $\ceil{x}$ is the ceiling function (i.e. the smallest integer greater or equal to $x$), the empirical estimator of the CVaR used in this study is from the work of \cite{hong2014monte}
and has the representation
$$\reallywidehat{\text{VaR}}_{\alpha}(\mathbb{B}_{j}) := \pi_{[\tilde{N}], j},$$
$$\reallywidehat{\text{CVaR}}_{\alpha}(\mathbb{B}_{j}):= \reallywidehat{\text{VaR}}_{\alpha}(\mathbb{B}_{j}) + \frac{1}{(1-\alpha)N_{\text{batch}}}\sum_{i=1}^{N_{\text{batch}}}\max(\pi_{i,j}-\reallywidehat{\text{VaR}}_{\alpha}(\mathbb{B}_{j}),0).$$
The gradient of the empirical estimator of the Conditional Value-at-Risk with respect to the trainable parameters (i.e. $\nabla_{\theta}\reallywidehat{\text{CVaR}}_{\alpha}(\mathbb{B}_{j})$) required for the update rule \eqref{eq:ref_SGD_iterative_param_approx} can be computed exactly without discretization or other numerical approximations. Such computations can be implemented with modern deep learning libraries such as Tensorflow \citep{abadi2016tensorflow}.
Furthermore, algorithms which dynamically adapt the learning rate $\eta_j$
in \eqref{eq:ref_SGD_iterative_param_approx} such as \textit{Adam} \citep{kingma2014adam} have been shown to improve upon the effectiveness of SGD procedures for neural networks. For all numerical experiments conducted in \cref{section:numerical_results}, an implementation of Tensorflow with the Adam algorithm is used to optimize neural networks; the reader is referred to the online Github repository for samples of codes in Python.\footnote{
%
\href{https://github.com/alexandrecarbonneau}{github.com/alexandrecarbonneau}.
%
} Also, \cref{appendix:pseudo_code} presents a pseudo-code of the training procedure for $F_{\theta}^{\mathcal{(S)}}$.
\section{Numerical experiments}
\label{section:numerical_results}
This section performs various numerical experimentations of the ERP approach for derivatives valuation. The main goal is to study the impact of including options as hedging instruments on equal risk prices and on the level of market incompleteness. A special case assessed throughout this section is
the trading of short-term vanilla options for the pricing and hedging of longer-term derivatives. The conduction of these experiments heavily relies on the neural network scheme described in \cref{section:neural_network} to solve the underlying global hedging problems of the ERP framework. Such exhaustive numerical study would have been hardly accessible with traditional methods (e.g. conventional dynamic programming algorithms) due to the high-dimensional continuous state and action spaces of the hedging problem stemming from the use of multiple short-term options as hedging instruments and from the asset price dynamics considered. As a result, the use of neural networks enables us to provide novel qualitative insights into the ERP framework.
%
The analysis begins in \cref{subsec:sensitivity_MJD} and \cref{subsec:sensitivity_GARCH} with the assessment of the sensitivity of equal risk prices and residual hedging risk to the presence of two salient equity stylized features: jump and volatility risks. The impact of the choice of convex risk measure on the ERP framework when trading exclusively options is examined in \cref{subsec:sensitivity_risk_aversion}. Lastly, \cref{subsec:benchmark_QDH_ERP} presents the benchmarking of equal risk prices to derivative premiums obtained with variance-optimal hedging. The specific financial market setup and asset dynamics models considered for all numerical experiments are described in \cref{subsec:financial_market} that follows.
\subsection{Market setup and asset dynamics models}
\label{subsec:financial_market}
For the rest of the paper, the derivative to price is a European vanilla put option of payoff function $\Phi(S_N^{(0,b)}) = \max(K-S_N^{(0,b)},0)$ with $K = 90, 100$ and $110$ corresponding respectively to an out-of-the-money (OTM), an at-the-money (ATM) and an in-the-money (ITM) option. The maturity of the derivative is set to $1$ year (i.e. $T=1$) with $252$ days. The annualized continuously compounded risk-free rate is $r=0.03$. In addition to the risk-free asset, the hedging instruments consist of either only the underlying stock, or exclusively shorter-term ATM European calls and puts. When hedging is performed with the underlying stock, daily and monthly rebalancing are considered, corresponding to respectively $N=252$ and $N=12$ trading periods per year. When hedging with options, all options are assumed to have a single-period time-to-maturity, i.e. they are traded once and held until expiration. We consider either 1-month or 3-months maturities ATM calls and puts as hedging instruments, which respectively entails $N=12$ or $N=4$. Less frequent rebalancing when hedging with options rather than only with the underlying stock is consistent with market practices; such hedging instruments are commonly embedded in semi-static type of trading strategies, see for instance \cite{carr2014static}. Lastly, note that daily variations for the underlying log-returns and implied volatilities are always considered throughout the rest of the paper, even with non-daily rebalancing periods (i.e. when hedging with the underlying stock on a monthly basis or with $1$-month and $3$-months maturities options) by aggregating daily variations over the rebalancing period.
\subsubsection{Asset price dynamics}
\label{subsec:dynamics_underlying}
The asset price dynamics models considered in stochastic simulations are now introduced. To characterize jump risk, the Merton jump-diffusion model (MJD, \cite{Merton1976}) is considered. Furthermore, the impact of volatility risk is assessed with the GJR-GARCH model of \cite{glosten1993relation}. Several sets of parameters are tested for each model to conduct a sensitivity analysis and highlight the impact of various model features on both equal risk prices and residual hedging risk.
%
Denote $y_n:=\log (S_n^{(0,b)}/S_{n-1}^{(0,b)})$ as the periodic underlying stock log-return between the trading periods $t_{n-1}$ and $t_{n}$. Since our modeling framework assumes daily variations for asset prices and possibly non-daily rebalancing, let $\{\tilde{y}_{j,n}\}_{j=1}^{M}$ be the $M$ daily stock log-returns in the time interval $[t_{n-1}, t_{n}]$ such that\footnote{
%
For completeness, let $\{\tilde{S}^{(0,b)}_{j,n}\}_{j=0,n=1}^{M,N}$ be the daily underlying stock prices where $\{\tilde{S}^{(0,b)}_{j,n}\}_{j=0}^{M}$ corresponds to the $M+1$ daily prices during the period $[t_{n-1}, t_n]$. Also, let $\mathbb{G}:=\{\mathcal{G}_{j,n}\}_{j=0, n=1}^{M,N}$ be a filtration satisfying the usual conditions with $\mathcal{G}_{j,n}$ containing all information available to market participants at the $j^{th}$ day of the time period $[t_{n-1},t_n]$. The filtration used to optimize trading strategies $\mathbb{F}$ with time steps $t_0, t_1,\ldots,t_N$ is a subset of $\mathbb{G}$ by construction. However, since the risky asset dynamics considered in this paper have the Markov property, optimizing trading strategies with the filtration $\mathbb{F}$ or $\mathbb{G}$ results in the same trading policy.
%
}
\begin{align}
y_n = \sum_{j=1}^{M}\tilde{y}_{j,n}, \quad n = 1,\ldots,N, \quad N \times M = 252, \label{eq:ref_agg_return}
\end{align}
where $N$ corresponds to the number of trading dates to hedge the $1$ year maturity derivative $\Phi$ and $M$ to the number of days between two trading dates. Thus, daily stock hedges corresponds to the case of $N=252$ and $M=1$, monthly stock and $1$-month option hedges to $N=12$ and $M=21$, and $3$-months option hedges to $N=4$ and $M=63$.
The asset price dynamics are now formally defined for the daily log-returns. For the rest of the section, let $\{\epsilon_{j,n}\}_{j=1, n=1}^{M,N}$ be a sequence of independent standardized Gaussian random variables where the subsequence $\{\epsilon_{j,n}\}_{j=1}^{M}$ will be used to model the $M$ daily innovations of log-returns in the time interval $[t_{n-1}, t_n]$.
\subsubsection{Discrete-time Merton-Jump diffusion model \citep{Merton1976}}
\label{subsubsec:MJD_model}
The Merton-jump diffusion dynamics expands upon the ideal market conditions of the Black-Scholes model by incorporating random Gaussian jumps along stock paths. Let $\{N_{j,n}\}_{j=0,n=1}^{M,N}$ be a discrete-time sampling from a Poisson process of intensity parameter $\lambda > 0$ where the subsequence $\{N_{j,n}\}_{j=0}^{M}$ corresponds to the $M+1$ daily values of the Poisson process occurring during the time interval $[t_{n-1}, t_n]$. $N_{0,1}:=0$ is the initial value of the process and $N_{0,n+1} := N_{M,n}$ for $n=1,\ldots,N-1$. Furthermore, denote $\{\xi_k\}_{k=1}^{\infty}$ as a sequence of random Gaussian variables corresponding to the jumps of mean $\mu_{J}$ and variance $\sigma_{J}^{2}$. $\{N_{j,n}\}_{j=0,n=1}^{M,N}$, $\{\xi_k\}_{k=1}^{\infty}$ and $\{\epsilon_{j,n}\}_{n=1,j=1}^{N,M}$ are independent. For $n=1,\ldots,N$ and $j=1,\ldots, M$, the daily log-return dynamics can be specified as\footnote{
%
This paper adopts the convention that if $N_{j,n}=N_{j-1,n}$, i.e. that no jumps occurred on that day, then:
$$\sum_{k=N_{j-1,n} + 1}^{N_{j,n}}\xi_{k} = 0.$$
%
}
\begin{align}
\tilde{y}_{j,n} = \frac{1}{252}\left(\nu - \lambda \left(e^{\mu_J + \sigma_J^{2}/2}-1\right) - \frac{\sigma^{2}}{2}\right) + \sigma \sqrt{\frac{1}{252}}\epsilon_{j,n} + \sum_{k=N_{j-1,n} + 1}^{N_{j,n}}\xi_{k}, \label{eq:ref_MJD}
\end{align}
where $\{\nu, \mu_J, \sigma_J, \lambda, \sigma\}$ are the model parameters with $\{\nu, \lambda, \sigma\}$ being on a yearly scale, $\nu \in \mathbb{R}$ and $\sigma > 0$. Furthermore, since $\{S_n^{(0,b)}\}_{n=0}^{N}$ has the Markov property with respect to the filtration $\mathbb{F}$ generated by the trading dates observations, no additional state associated to the risky asset dynamics is required to be added to the feature vectors of neural networks
(i.e. $\varphi_n = 0$ for all time steps $n$ in \eqref{eq:ref_feature_vect_stock} and \eqref{eq:ref_feature_vect_opts}).
\subsubsection{GJR-GARCH(1,1) model \citep{glosten1993relation}}
\label{subsubsec:GARCH_model}
GARCH processes also expand upon the Black-Scholes ideal framework by exhibiting well-known empirical features of risky assets such as time-varying volatility, volatility clustering and the leverage effect (i.e. negative correlation between underlying returns and its volatility). Daily log-returns modeled with a GJR-GARCH(1,1) dynamics have the representation
\begin{align}
\tilde{y}_{j,n} &= \mu + \tilde{\sigma}_{j,n} \epsilon_{j,n}, \nonumber
\\
\tilde{\sigma}_{j+1, n}^{2} &= \omega + \upsilon \tilde{\sigma}_{j,n}^{2}(|\epsilon_{j,n}| - \gamma \epsilon_{j,n})^{2} + \beta \tilde{\sigma}_{j,n}^{2}, \label{eq:ref_GARCH}
\end{align}
where $\{\tilde{\sigma}_{j,n}^2\}_{j=1,n=1}^{M+1,N}$ are the daily conditional variances of log-returns. More precisely, $\{\tilde{\sigma}_{j,n}^2\}_{j=1}^{M+1}$ are the $M+1$ daily conditional variances in the time interval $[t_{n-1}, t_n]$. Also, $\tilde{\sigma}_{1,n+1}^2 := \tilde{\sigma}_{M+1,n}^2$ for $n=1,\ldots,N-1.$ Model parameters consist of $\{\mu, \omega, \upsilon, \gamma, \beta\}$ with $\{\omega, \upsilon, \beta\}$ being positive real values and $\gamma, \mu \in \mathbb{R}$. Note that if the starting value of the GARCH process $\tilde{\sigma}_{1,1}^{2}$ is deterministic, then $\{\tilde{\sigma}_{j,n}^2\}_{j=1,n=1}^{M+1,N}$ can be computed recursively with the observed daily log-returns. In this paper, $\tilde{\sigma}_{1,1}^{2}$ is set as the stationary variance: $\tilde{\sigma}_{1,1}^{2} := \frac{\omega}{1 - \upsilon(1+\gamma^{2})-\beta}$. Also, contrarily to the MJD model, the GJR-GARCH(1,1) requires adding at each trading time $t_n$ the current stochastic volatility value to the feature vectors of the neural networks, i.e. $\varphi_n = \tilde{\sigma}_{1,n+1}$ for $n=0,\ldots,N-1$.
\subsubsection{Implied volatility dynamics}
\label{subsec:dynamics_IV}
This work proposes to model the daily variations of the logarithm of ATM implied volatilities with 1-month and 3-months maturities as a discrete-time version of the Ornstein-Uhlenbeck (OU) process.\footnote{
%
It is important to note that implied volatilities are used strictly for pricing options used as hedging instruments. They are not used to price the derivative $\Phi$.
%
}
The choice of an OU type of dynamics for IVs is motivated by the work of \cite{cont2002dynamics} which shows that for S\&P 500 index options,
the first principal component
of the daily variations of the logarithm of the IV surface accounts for the majority of its variance and can be interpreted as a level effect. Also, this first principal component can be well represented by a low-order autoregressive (AR) model.
The OU dynamics considered in this study therefore has the representation of an AR model of order $1$.
The dynamics for the daily evolution of IVs is now formally defined. For convenience, this paper assumes that 1-month and 3-months IVs are the same.\footnote{
It is worth highlighting that since trading strategies allow for the use of either $1$-month or $3$-months maturities ATM calls and puts, but not both maturities within the same strategy, $1$-month and $3$-months IVs are never used at the same time.
}
Using a similar notation as for daily log-returns, let $\{\widetilde{IV}_{j,n}\}_{j=0,n=1}^{M,N}$ be the daily ATM IV process for both 1-month and 3-months maturities where $\{\widetilde{IV}_{j,n}\}_{j=0}^{M}$ are the $M+1$ daily observations during the time interval $[t_{n-1}, t_n]$ with $\widetilde{IV}_{0,n+1}:=\widetilde{IV}_{M,n}$ for $n=1,\ldots,N-1$.
Furthermore, let $\{Z_{j,n}\}_{j=1,n=1}^{M,N}$ be an additional sequence of independent standardized Gaussian random variables characterizing shocks in the IV dynamics.
In order to incorporate the stylized feature of strong negative correlation between implied volatilities and asset returns (\cite{cont2002dynamics}), the modeling framework assumes that the daily innovations of log-returns and IVs are correlated with parameter $\varrho := corr(\epsilon_{j,n},Z_{j,n})$ set at $-0.6$ for all time steps.
The dynamics for the evolution of the logarithm of IVs, which is referred from now on as the log-AR(1) model, has the following representation for $n=1,\ldots,N$ and $j=0,\ldots,M-1$:
\begin{align}
\log \widetilde{IV}_{j+1,n} &= \log \widetilde{IV}_{j,n} + \kappa(\vartheta - \log \widetilde{IV}_{j,n}) + \sigma_{IV} Z_{j+1,n}, \label{eq:ref_IV_model}
\end{align}
where $\{\kappa, \vartheta , \sigma_{IV}\}$ are the model parameters with $\kappa, \vartheta \in \mathbb{R}$ and $\sigma_{IV} > 0$. The initial value of the process is set as $\log \widetilde{IV}_{0,1} = \vartheta$. Also, recall that when trading options, their corresponding implied volatilities at each trading date are added to the feature vectors of neural networks, i.e. $IV_{n-1} = \widetilde{IV}_{0,n}$ in \eqref{eq:ref_feature_vect_opts} for $n=1,\ldots,N$.
The pricing of calls and puts used as hedging instruments is done with the well-known Black-Scholes formula hereby stated with the annual volatility term set at the implied volatility value. For the underlying price $S$, implied volatility $IV$, strike price $K$ and time-to-maturity $\Delta T$, the Black-Scholes pricing formulas for calls and puts are respectively
\begin{align}
C(S, IV, \Delta T, K)&:=S \mathcal{N}(d_1) - e^{-r\Delta T}K \mathcal{N}(d_2), \label{eq:ref_BSM_optprice_call}
\\ P(S, IV, \Delta T, K)&:=e^{-r\Delta T} K \mathcal{N}(-d_2) - S \mathcal{N}(-d_1), \label{eq:ref_BSM_optprice_put}
\end{align}
where $\mathcal{N}(\cdot)$ denotes the standard normal cumulative distribution function and
$$d_1:=\frac{\log(\frac{S}{K}) + (r+\frac{IV^2}{2})\Delta T}{IV\sqrt{\Delta T}}, \quad d_2:=d_1 - IV\sqrt{\Delta T}.$$
\subsubsection{Hyperparameters}
\label{subsubsec:hyperparameters}
The set of hyperparameters for the LSTMs are two LSTM cells (i.e. $H=2$) and $24$ neurons per cell (i.e. $d_j = 24$ for $j=1,2$). A training set of $400,\!000$ paths is used to optimize the trainable parameters with a total of $50$ epochs and a minibatch size of $1,\!000$ sampled exclusively from the training set.\footnote{
%
One epoch consists of a complete iteration of SGD on the training set. For a training set of $400,\!000$ paths and a minibatch of size $1,\!000$, a total of $400$ updates of the trainable parameters as in \eqref{eq:ref_SGD_iterative_param_approx} is performed within an epoch.
%
} The deep learning library Tensorflow \citep{abadi2016tensorflow} is used to implement the stochastic gradient descent procedure with the Adam optimizer of \cite{kingma2014adam} with a learning rate hyperparameter value of $0.01/6$. All numerical results presented throughout this section are computed based on a test set (i.e. out-of-sample dataset) of $100,\!000$ paths. Lastly, unless specified otherwise, the convex risk measure chosen for all experiments is the CVaR with confidence level $\alpha = 0.95$. Sensitivity analyses of equal risk prices and residual hedging risk with respect to the confidence level parameter of the CVaR measure are performed in \cref{subsec:sensitivity_risk_aversion} and \cref{subsec:benchmark_QDH_ERP}.
\subsection{Sensitivity of equal risk pricing to jump risk}
\label{subsec:sensitivity_MJD}
This section examines the sensitivity of the ERP solution to equity jump risk.
The analysis is carried out by considering three different sets of parameters for the MJD dynamics which induce different levels of jump frequency and severity. While maintaining empirical plausibility, this is done by modifying the intensity parameter $\lambda$ controlling the expected frequency of jumps as well as parameters $\mu_J$ and $\sigma_J$ controlling the severity component of jumps. In order to better isolate the impact of different stylized features of jump risk on the ERP framework, the diffusion parameter\footnote{
%
The parameter $\sigma$ in \eqref{eq:ref_MJD} corresponds to the diffusion parameter of the MJD dynamics.
%
} is fixed for all three sets of parameters. Also, the parameters $\{\lambda, \mu_J, \sigma_J, \nu\}$ are chosen such that the yearly expected value and standard deviation of log-returns are respectively $10\%$ and $15\%$ for all three cases. To facilitate the analysis, the three sets of parameters are referred to as scenario $1$, scenario $2$ and scenario $3$ for jump risk. Model parameter values for the three scenarios are presented in \cref{table:all_MJD}.
\begin{table}[ht]
\caption {Parameters of the Merton jump-diffusion model for the three scenarios.} \label{table:all_MJD}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccc}
\hline
$ $ & $\nu$ & $\sigma$ & $\lambda$ & $\mu_J$ & $\sigma_J$
\\
\hline\noalign{\medskip}
%
$\text{Scenario 1}$ & $0.1112$ & $0.1323$ & $1$ & $-0.05$ & $0.05$
\\
%
$\text{Scenario 2}$ & $0.1111$ & $0.1323$ & $0.25$ & $-0.10$ & $0.10$
\\
%
$\text{Scenario 3}$ & $0.1110$ & $0.1323$ & $0.08$ & $-0.20$ & $0.15$
\\
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
\centering{Notes: $\nu$, $\sigma$ and $\lambda$ are on an annual basis.}
\end{table}
Scenario $1$ represents relatively smaller but more frequent jumps with on average one jump per year of mean $-5\%$ and standard deviation $5\%$. Scenario $2$ entails more severe, but less frequent jumps with on average one jump every four years of mean $-10\%$ and standard deviation $10\%$. Lastly, scenario $3$ depicts the most extreme case with rare but very severe jumps with on average one jump every twelve and a half years of mean $-20\%$ and standard deviation $15\%$.
Moreover, parameter values for the log-AR(1) implied volatility model are kept fixed for all three scenarios and are presented in \cref{table:all_OU_params}. Note that the long-run parameter $\vartheta$ is set at the logarithm of the yearly standard deviation of log-returns with $\vartheta = \log 0.15$, and other parameters are chosen in an ad hoc fashion so as to produce reasonable values for implied volatilities.
\begin{table} [ht]
\caption {Parameters of the log-AR(1) model for the evolution of implied volatilities.}
\label{table:all_OU_params}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{cccc}
\hline
$\kappa$ & $\vartheta $ & $\sigma_{\text{IV}}$ & $\varrho$
\\
\hline\noalign{\medskip}
%
$0.15$ & $\log(0.15)$ & $0.06$ & $-0.6$
\\
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
\end{table}
\subsubsection{Benchmarking results in the presence of jump risk}
\label{subsubsec:jump_risk}
\cref{table:ERP_one_year_put_MJD_analysis}
presents equal risk prices $C_0^{\star}$ and residual hedging risk $\epsilon^{\star}$ across the three scenarios of jump parameters and different trading instruments.
\begin{table}
\caption {Sensitivity analysis of equal risk prices $C_{0}^{\star}$ and residual hedging risk $\epsilon^{\star}$ to jump risk for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$.} \label{table:ERP_one_year_put_MJD_analysis}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$\text{OTM}$} & & \multicolumn{3}{c}{$\text{ATM}$} & & \multicolumn{3}{c}{$\text{ITM}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Jump Scenario & $(1)$ & $(2)$ & $(3)$ & & $(1)$ & $(2)$ & $(3)$ & & $(1)$ & $(2)$ & $(3)$ \\
%
\hline\noalign{\medskip}
%
\multicolumn{4}{l}{\underline{\emph{$C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $1.89$ & $2.58$ & $3.36$ & & $5.21$ & $6.01$ & $6.81$ & & $10.81$ & $11.68$ & $12.13$ \\
$\text{Monthly stock}$ & $1.97$ & $2.60$ & $3.31$ & & $5.04$ & $5.77$ & $6.38$ & & $10.73$ & $11.44$ & $11.86$ \\
$\text{1-month options}$ & $1.82$ & $2.24$ & $2.55$ & & $4.99$ & $5.36$ & $5.60$ & & $10.48$ & $10.86$ & $10.83$ \\
$\text{3-months options}$ & $1.74$ & $2.08$ & $2.39$ & & $4.87$ & $5.12$ & $5.28$ & & $10.43$ & $10.51$ & $10.57$ \\
%
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$\epsilon^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $1.09$ & $1.98$ & $2.67$ & & $1.76$ & $2.74$ & $3.54$ & & $1.82$ & $2.78$ & $3.27$ \\
$\text{Monthly stock}$ & $1.82$ & $2.52$ & $3.26$ & & $3.00$ & $3.88$ & $4.57$ & & $3.07$ & $3.91$ & $4.37$ \\
$\text{1-month options}$ & $0.76$ & $1.18$ & $1.52$ & & $1.14$ & $1.53$ & $1.78$ & & $1.17$ & $1.56$ & $1.54$ \\
$\text{3-months options}$ & $1.03$ & $1.37$ & $1.68$ & & $1.59$ & $1.82$ & $2.02$ & & $1.70$ & $1.79$ & $1.88$ \\
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$\epsilon^{\star}/C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $0.58$ & $0.77$ & $0.79$ & & $0.34$ & $0.46$ & $0.52$ & & $0.17$ & $0.24$ & $0.27$ \\
$\text{Monthly stock}$ & $0.92$ & $0.97$ & $0.99$ & & $0.60$ & $0.67$ & $0.72$ & & $0.29$ & $0.34$ & $0.37$ \\
$\text{1-month options}$ & $0.42$ & $0.53$ & $0.60$ & & $0.23$ & $0.28$ & $0.32$ & & $0.11$ & $0.14$ & $0.14$ \\
$\text{3-months options}$ & $0.59$ & $0.66$ & $0.70$ & & $0.33$ & $0.36$ & $0.38$ & & $0.16$ & $0.17$ & $0.18$ \\
%
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the Merton Jump-Diffusion model for the underlying (see \cref{subsubsec:MJD_model} for model description). Three different sets of parameters values are considered with $\lambda = \{1, 0.25, 0.08\}$, $\mu_J = \{-0.05, -0.10, -0.20\}$ and $\sigma_J = \{0.05, 0.10, 0.15\}$ respectively for jump scenario $1$, $2$, and $3$ (see \cref{table:all_MJD} for all parameters values). \textit{Hedging instruments}: daily or monthly rebalancing with the underlying stock and 1-month or 3-months options with ATM calls and puts. Options used as hedging instruments are priced with implied volatility modeled with a log-AR(1) dynamics (see \cref{subsec:dynamics_IV} for model description and \cref{table:all_OU_params} for parameters values). The training of neural networks is done as described in \cref{subsubsec:hyperparameters}. The confidence level of the CVaR measure is $\alpha = 0.95$.
\end{table}
Numerical values indicate that in the presence of jump risk, hedging with options entails significant reduction of both equal risk prices and market incompleteness as compared to hedging solely with the underlying stock across all moneyness levels and jump risk scenarios. The reduction in hedging residual risk by trading options is obtained despite less frequent rebalancing than when only the stock is used.
These results add additional evidence that options are indeed non-redundant as prescribed by the Black-Scholes world:
the equal risk pricing framework dictates that hedging with options in the presence of jump risk can significantly impact both derivative premiums and hedging risk as quantified by our incompleteness metrics.
The relative reduction achieved in $C_0^{\star}$ with 1-month and 3-months options as compared to hedging with the stock is most important for OTM puts, followed by ATM and ITM contracts. For instance, the relative reduction obtained with 3-months options hedging over daily stock hedging ranges across the three jump risk scenarios between $8\%$ to $29\%$ for OTM, $6\%$ to $22\%$ for ATM and $4\%$ to $13\%$ for ITM puts.\footnote{
%
If $C_0^{\star}(\text{daily stock})$ and $C_0^{\star}(\text{3-months options})$ are equal risk prices obtained respectively by hedging with the stock on a daily basis and with 3-months options, the relative reduction is computed as $1 - \frac{C_0^{\star}(\text{3-months options})}{C_0^{\star}(\text{daily stock})}$ for all examples.
%
}
This reduction in $C_0^{\star}$ when using options as hedging instruments can be explained by the following observations. As pointed out in \cite{carbonneau2020equal}, the fact that a put option payoff is bounded below at zero entails that the short position hedging error has a thicker right tail than the long position hedging error. Also, it is widely documented in the literature that hedging jump risk with options significantly dampens tail risk as compared to using only the underlying stock (see for instance \cite{coleman2007robustly} and \cite{carbonneau2020deep}).\footnote{
\cite{horvath2021deep} deep hedge derivatives under a rough Bergomi volatility model by trading the underlying stock and a variance swap. The latter paper shows that this dynamics exhibits jump-like behaviour when discretized. As results presented in this current paper highlights the fact that global hedging jump risk with option hedges is very effective, deep hedging with options could also potentially be effective under such rough volatility models.
}
Consequently,
the choice of trading options to mitigate jump risk reduces the measured risk exposure of both the long and short positions, but the thicker right tail for the short position hedging error entails a larger decrease for the latter than for the long position. In such situations, the ERP framework dictates that the long position should be compensated with a lower derivative premium $C_0^{\star}$ to equalize residual hedging risk of both positions.
Moreover, values for both $\epsilon^{\star}$-metrics indicate that in the presence of jump risk, the use of options contributes significantly to the reduction of market incompleteness as both the long and short position hedges achieve risk reduction when compared to trading only with the stock.
The latter conclusion is in itself not novel, and is widely documented in the literature (see, for instance, \cite{tankov2003financial} and the many references therein). Indeed, this is a consequence of the well-known convex property of put option prices, which implies that hedging random jumps solely with the underlying stock is ineffective. Our $\epsilon^{\star}$-metrics have the advantage of allowing for a precise quantification of such reduction in residual hedging risk achieved through the use of options as hedging instruments.
The sensitivity of equal risk prices and residual hedging risk across the three jump risk scenarios for each set of hedging instruments is now examined.
Numerical results presented in \cref{table:ERP_one_year_put_MJD_analysis} indicate that for a fixed set of hedging instruments, both the equal risk price and the level of incompleteness increases with the severity of jumps across all moneyness levels.
Indeed, the relative increase of equal risk prices observed under scenario $3$ as compared to scenario $1$ respectively for OTM, ATM and ITM puts is $78\%, 31\%$ and $12\%$ with the daily stock, $68\%$, $27\%$ and $11\%$ with the monthly stock, $40\%$, $12\%$ and $3\%$ with 1-month options and $38\%$, $8\%$ and $1\%$ with 3-months options.\footnote{
%
For a fixed hedging instrument and moneyness level, if $C_0^{\star}(\text{scenario 1})$ and $C_0^{\star}(\text{scenario 3})$ are respectively the equal risk price obtained under jump risk scenario $1$ and $3$, the relative increase is computed as $\frac{C_0^{\star}(\text{scenario 3})}{C_0^{\star}(\text{scenario 1})} - 1$.
%
} Similar observations can be made for both incompleteness metrics: increases in jump severity leads to larger $\epsilon^{\star}$ and $\epsilon^{\star}/C_0^{\star}$. This positive association between both equal risk prices and the level of market incompleteness to jump severity can be explained by the following observations. For a fixed hedging instrument and moneyness level, the long measured risk exposure is closed to invariant to jump severity (i.e. similar values across the three jump risk scenarios). The latter stems from the fact that jump dynamics considered in this paper predominantly entail negative jumps, which result in a thicker left tail for the long position hedging error (i.e. hedging gains) as jump severity increases, but in close to no impact on the right tail of the long position hedging error. In contrast, since the right tail weight of the short position hedging error increases with the expected (negative) magnitude and volatility of jumps, the short measured risk exposure always increases going from scenario $1$ to scenario $3$, which consequently increases both the equal risk price and the level of market incompleteness.
\subsection{Sensitivity of equal risk pricing to volatility risk}
\label{subsec:sensitivity_GARCH}
Having examined the impact of jump risk on the ERP framework, the impact of volatility risk is now studied.
In the same spirit as analyses done for jump risk, three different sets of parameters are considered for the GARCH dynamics which imply annualized stationary (expected) volatilities of $10\%, 15\%$ and $20\%$.\footnote{
%
The annualized stationary volatility with $252$ days per year is computed as
$$\sqrt{\frac{252\omega}{1 - \upsilon (1+\gamma^{2})-\beta}}.$$
%
}
The three sets of parameters are presented in \cref{table:all_GARCH}.
Note that every parameter is fixed for all three sets, except for the level parameter $\omega$, which is adjusted to attain the wanted stationary volatility. The value of the drift parameter $\mu$ is set such that the yearly expected value of log-returns is $10\%$. Also, values for $\{\upsilon , \gamma, \beta\}$ are inspired from parameters estimated with maximum likelihood on a time series of daily log-returns on the S\&P 500 index for the period 1986-12-31 to 2010-04-01 used in \cite{carbonneau2020equal}. The same setup is considered as in \cref{subsec:sensitivity_MJD} in terms of the derivative to be priced ($1$-year maturity European puts) and for the choice of hedging instruments (underlying stock traded on a daily or monthly basis and $1$-month or $3$-months maturities ATM calls and puts).
The same parameters as in the study of jump risk conducted in \cref{subsec:sensitivity_MJD} are used for $\{\kappa, \sigma_{\text{IV}}, \varrho\}$ of the log-AR(1) dynamics for the evolution of $1$-month and $3$-months ATM IVs (i.e. $\kappa = 0.15, \sigma_{\text{IV}}=0.06$ and $\varrho = -0.6$), except for the long-run parameter $\vartheta$, which is set to be in line with the underlying GARCH process as $\log(0.10), \log(0.15)$ and $\log(0.20)$ when the stationary volatility is $10\%$, $15\%$ and $20\%$, respectively. It is worth highlighting that the choice of modeling implied volatilities for short-term options with higher and smaller average levels enables us to assess the impact of larger and smaller average costs for trading options on the equal risk price and residual hedging risk of longer-term options.
\begin{table} [ht]
\caption {Parameters of the GJR-GARCH model for $10\%, 15\%$ and $20\%$ stationary yearly volatilities.} \label{table:all_GARCH}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{cccccc}
\hline
$\text{Stationary volatility}$ & $\mu$ & $\omega$ & $\upsilon$ & $\gamma$ & $\beta$
\\
\hline\noalign{\medskip}
%
$\text{10\%}$ & $3.968\text{e-}04$ & $8.730\text{e-}07$ & $0.05$ & $0.6$ & $0.91$
\\
%
$\text{15\%}$ & $3.968\text{e-}04$ & $1.964\text{e-}06$ & $0.05$ & $0.6$ & $0.91$
\\
%
$\text{20\%}$ & $3.968\text{e-}04$ & $3.492\text{e-}06$ & $0.05$ & $0.6$ & $0.91$
\\
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
\end{table}
\subsubsection{Benchmarking results with volatility risk}
\cref{table:ERP_one_year_put_vol_analysis} presents equal risk prices $C_0^{\star}$ and $\epsilon^{\star}$-metrics for put options of $1$ year maturity across the three sets of volatility risk parameters and hedging instruments.
\begin{table}
\caption {Sensitivity analysis of equal risk prices $C_{0}^{\star}$ and residual hedging risk $\epsilon^{\star}$ to volatility risk for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$.} \label{table:ERP_one_year_put_vol_analysis}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$\text{OTM}$} & & \multicolumn{3}{c}{$\text{ATM}$} & & \multicolumn{3}{c}{$\text{ITM}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Stationary volatility & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ \\
%
\hline\noalign{\medskip}
%
\multicolumn{4}{l}{\underline{\emph{$C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $1.01$ & $2.35$ & $3.85$ & & $3.23$ & $5.36$ & $7.24$ & & $8.56$ & $10.55$ & $12.51$ \\
$\text{Monthly stock}$ & $1.17$ & $2.65$ & $4.23$ & & $3.37$ & $5.44$ & $7.58$ & & $8.85$ & $10.82$ & $12.88$ \\
$\text{1-month options}$ & $0.56$ & $1.74$ & $3.27$ & & $2.86$ & $4.87$ & $6.89$ & & $8.46$ & $10.32$ & $12.35$ \\
$\text{3-months options}$ & $0.76$ & $2.07$ & $3.65$ & & $3.01$ & $5.08$ & $7.16$ & & $8.51$ & $10.38$ & $12.44$ \\
%
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$\epsilon^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $0.77$ & $1.51$ & $2.21$ & & $1.28$ & $2.12$ & $2.67$ & & $1.12$ & $1.98$ & $2.65$ \\
$\text{Monthly stock}$ & $1.15$ & $2.44$ & $3.67$ & & $2.15$ & $3.41$ & $4.62$ & & $1.92$ & $3.29$ & $4.56$ \\
$\text{1-month options}$ & $0.26$ & $0.65$ & $1.06$ & & $0.59$ & $1.00$ & $1.36$ & & $0.62$ & $1.03$ & $1.39$ \\
$\text{3-months options}$ & $0.59$ & $1.32$ & $2.02$ & & $1.10$ & $1.77$ & $2.41$ & & $1.04$ & $1.68$ & $2.31$ \\
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$\epsilon^{\star}/C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $0.77$ & $0.64$ & $0.57$ & & $0.40$ & $0.40$ & $0.37$ & & $0.13$ & $0.19$ & $0.21$ \\
$\text{Monthly stock}$ & $0.99$ & $0.92$ & $0.87$ & & $0.64$ & $0.63$ & $0.61$ & & $0.22$ & $0.30$ & $0.35$ \\
$\text{1-month options}$ & $0.45$ & $0.37$ & $0.32$ & & $0.21$ & $0.20$ & $0.20$ & & $0.07$ & $0.10$ & $0.11$ \\
$\text{3-months options}$ & $0.77$ & $0.64$ & $0.55$ & & $0.36$ & $0.35$ & $0.34$ & & $0.12$ & $0.16$ & $0.19$ \\
%
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the GJR-GARCH(1,1) model for the underlying with three sets of parameters implying stationary yearly volatilities of $10\%, 15\%$ and $20\%$ (see \cref{subsubsec:GARCH_model} for model description and \cref{table:all_GARCH} for parameters values). \textit{Hedging instruments}: daily or monthly rebalancing with the underlying stock and 1-month or 3-months options with ATM calls and puts. Options used as hedging instruments are priced with implied volatility modeled as a log-AR(1) dynamics with $\kappa = 0.15, \sigma_{\text{IV}} = 0.06$ and $\varrho = -0.6$ for all cases, and $\vartheta$ set to $\log(0.10), \log(0.15)$ and $\log(0.20)$ when the GARCH stationary volatility is $10\%, 15\%$ and $20\%$, respectively (see \cref{subsec:dynamics_IV} for the log-AR(1) model description). The training of neural networks is done as described in \cref{subsubsec:hyperparameters}. The confidence level of the CVaR measure is $\alpha = 0.95$.
\end{table}
Numerical results indicate that in the presence of volatility risk, the use of options as hedging instruments can reduce $C_0^{\star}$ as compared to daily stock hedging. However, this impact on $C_0^{\star}$ when trading options can be marginal and is highly sensitive to the moneyness level of the put option being priced as well as to the maturity of the traded options. Furthermore, the impact on $C_0^{\star}$ of the use of options within hedges tends to diminish when traded options are more costly (i.e. as the average level of implied and GARCH volatility increases). Indeed, the relative reduction in equal risk prices achieved with 1-month options hedging as compared to daily stock hedging with $10\%, 15\%$ and $20\%$ stationary volatility is respectively $44\%, 26\%$ and $15\%$ for OTM puts, $12\%, 9\%$ and $5\%$ for ATM and $1\%, 2\%$ and $1\%$ for ITM options.\footnote{
%
If $C_0^{\star}(\text{daily stock})$ and $C_0^{\star}(\text{1-month options})$ are respectively the equal risk price obtained by hedging with the stock on a daily basis and with 1-month options, the relative reduction is computed as $1 - \frac{C_0^{\star}(\text{1-month options})}{C_0^{\star}(\text{daily stock})}$ for all examples.
%
} However, the relative reduction in $C_0^{\star}$ with $3$-months option hedges as compared to using the stock on a daily basis is overall much more marginal, with the notable exceptions of OTM and ATM puts with $10\%$ stationary volatility which achieve respectively $25\%$ and $7\%$ reduction as well as for the OTM moneyness under $15\%$ stationary volatility with a $12\%$ reduction.
Also, as expected, values presented in \cref{table:ERP_one_year_put_vol_analysis} confirm that the level of market incompleteness as measured by the $\epsilon^{\star}$ metric has a positive relationship with the average level of stationary volatility for all hedging instruments.
The previously described observations about the impact of option hedges on both equal risk prices and residual hedging risk all stem from the realized reduction in measured risk exposure by the long and short positions. However, contrarily to results obtained with jump risk, the reduction in measured risk exposure when hedging volatility risk with options can be very similar for both the long and short positions, whereas with jump risk, the reduction is asymmetric by always favoring the short position with a larger reduction. The latter can be explained by the fact that volatility risk impacts both upside and downside risk, while the impact of jump risk dynamics considered in this paper is very asymmetric by entailing significantly more weight on the right (resp. left) tail of the short (resp. long) hedging error with predominantly negative jumps.
Values presented in \cref{table:ERP_one_year_put_vol_analysis} confirm this analysis of the interrelation between volatility risk and the choice of hedging instruments. For instance, for ITM puts, the measured risk exposure of the long and short positions decreases by a similar amount when trading $1$-month or $3$-months options as compared to daily stock hedges, which explains the significant decrease in $\epsilon^{\star}$, but also the insensitivity of $C_0^{\star}$ to the choice of hedging instruments and rebalancing frequency. On the other hand, for OTM puts, 1-month and 3-months option hedges results in larger decreases of measured risk exposure for the short position than for the long position, which explains the reduction in $C_0^{\star}$ and $\epsilon^{\star}$ as compared to daily stock hedges.
Lastly, it is very interesting to observe that the average price level of short-term options used as hedging instruments is effectively reflected into the equal risk price of longer-term options. Indeed, numerical results for $C_0^{\star}$ presented in \cref{table:ERP_one_year_put_vol_analysis} highlight the fact that higher hedging options implied volatilities for $1$-month and $3$-months ATM calls and puts leads to higher equal risk prices for $1$-year maturity puts. Furthermore, to isolate the idiosyncratic contribution of the variations of option prices used as hedging instruments on the equal risk price from the impact of the stationarity volatility of the GARCH process, the authors also tested fixing the stationarity volatility of the GARCH process to $15\%$ and setting the long-run parameter of the IV process to $14\%$ and $16\%$. These results presented in the Supplementary Material, Table SM2, confirm that higher implied volatilities for options used as hedging instruments leads to higher equal risk prices.
All of these benchmarking results demonstrate the potential of the ERP framework as a fair valuation approach consistent with observable market prices. For instance, the ERP framework could be used to price and optimally hedge over-the-counter derivatives with vanilla options. An additional potential application is the marking-to-market of less liquid long-term derivatives (e.g. Long-Term Equity AnticiPation Securities (LEAPS)) consistently with highly liquid shorter-term option hedges. The ERP framework could also be used for the fair valuation of segregated funds guarantees, which are equivalent to very long-term (up to 40 years) derivatives sold by insurers.\footnote{
%
Note that \cite{carbonneau2020deep} demonstrates the potential of the deep hedging algorithm for global hedging long-term lookback options embedded in segregated funds guarantees with multiple hedging instruments. It is also worth highlighting that \cite{barigou2020insurance} developed a pricing scheme consistent with local non-quadratic hedging procedures for insurance liabilities which relies on neural networks.
%
}
Indeed, International Financial Reporting Standards 17 (IFRS 17, \cite{IFRS17}) mandates a market consistent valuation of options embedded in segregated funds guarantees with readily available observable market prices at the measurement date. The ERP framework could potentially be applied to price such very long-term options consistently with shorter-term implied volatility surface dynamics, with the latter being much less challenging to calibrate due to the higher liquidity of short-term options.\footnote{
In the context of segregated funds, the short position of the embedded option is assumed to be held by an insurance company who has to provide a quote and mitigate its risk exposure. The long position is held by an unsophisticated investor who will not be hedging his risk exposure. Nevertheless, as IFRS 17 mandates the use of a fair valuation approach for embedded options consistent with observable market prices, the ERP framework could potentially be used in this context.
}
\subsection{Sensitivity analyses to the confidence level of $\text{CVaR}_{\alpha}$}
\label{subsec:sensitivity_risk_aversion}
This section conducts sensitivity analyses with respect to the choice of convex risk measure on the ERP framework when trading exclusively options. Similarly to the work of \cite{carbonneau2020equal}, values for equal risk prices and $\epsilon^{\star}$-metrics are examined across the confidence levels $0.90, 0.95$ and $0.99$ for the $\text{CVaR}_{\alpha}$ measure. As argued in the latter paper, higher confidence levels corresponds to more risk averse agents by concentrating more relative weight on losses of larger magnitude. The main finding of the sensitivity analysis conducted in \cite{carbonneau2020equal} is that when trading exclusively the underlying stock, higher confidence levels leads to larger values for $C_0^{\star}$ and $\epsilon^{\star}$ metrics. The objective of this section is to assess if this finding is robust to the use of short-term option hedges instead of the underlying stock.
%
For each confidence level, the authors of the current paper computed both equal risk prices and residual hedging risk obtained by trading $3$-months ATM calls and puts with the same setup as in \cref{subsec:sensitivity_MJD} and \cref{subsec:sensitivity_GARCH}, i.e. for all three jump and volatility scenarios of parameters.\footnote{
%
Unreported tests performed by the authors show that values lower than $0.90$ for the confidence level of $\text{CVaR}_{\alpha}$ with $1$-month and $3$-months option hedges lead to trading policies with significantly larger tail risk in a way which would deem such policies as inadmissible by hedgers. Using the $\text{CVaR}_{0.90}$ measure with $1$-month options also resulted in trading policies with significantly larger tail risk. However, this large increase in tail risk was not observed with the $\text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ measures when trading $1$-month options, nor with $\text{CVaR}_{0.90}$, $\text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ when trading $3$-months options. These observations motivated the choice of performing sensitivity analysis for $\text{CVaR}_{\alpha}$ with $\alpha = 0.90, 0.95$ and $0.99$ exclusively when trading $3$-months options.
%
} Overall, the main conclusions are found to be qualitatively similar for all of the different setups. Thus, to save space, values for equal risk prices and residual hedging risk are only reported under the MJD dynamics with jump risk scenario $2$ by trading $3$-months options; these results are presented in \cref{table:sensitivity_analysis_CVAR_090_095_099}. The interested reader in numerical results obtained under jump risk scenarios $1$ and $3$ as well as under the three sets of volatility risk parameters is referred to Table SM3 and Table SM4 of the Supplementary Material.
\begin{table} [ht]
\caption {Sensitivity analysis of equal risk prices $C_{0}^{\star}$ and residual hedging risk $\epsilon^{\star}$ for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$ under the MJD dynamics with jump risk scenario $2$.} \label{table:sensitivity_analysis_CVAR_090_095_099}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$C_{0}^{\star}$} & & \multicolumn{3}{c}{$\epsilon^{\star}$} & & \multicolumn{3}{c}{$\epsilon^{\star}/C_{0}^{\star}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Moneyness & $\text{OTM}$ & $\text{ATM}$ & $\text{ITM}$ & & $\text{OTM}$ & $\text{ATM}$ & $\text{ITM}$ & & $\text{OTM}$ & $\text{ATM}$ & $\text{ITM}$ \\
\hline\noalign{\medskip}
%
$\text{CVaR}_{0.90}$ & $1.86$ & $4.93$ & $10.40$ & & $0.99$ & $1.43$ & $1.50$ & & $0.53$ & $0.29$ & $0.14$ \\
$\text{CVaR}_{0.95}$ & $12\%$ & $4\%$ & $1\%$ & & $39\%$ & $28\%$ & $20\%$ & & $24\%$ & $23\%$ & $18\%$ \\
$\text{CVaR}_{0.99}$ & $40\%$ & $10\%$ & $4\%$ & & $116\%$ & $76\%$ & $65\%$ & & $54\%$ & $60\%$ & $59\%$ \\
%
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the Merton Jump-Diffusion model for the underlying (see \cref{subsubsec:MJD_model} for model description) with parameters $\nu = 0.1111, \sigma = 0.1323, \lambda = 0.25, \mu_J = -0.10$ and $\sigma_J = 0.10$ corresponding to jump risk scenario $2$ of \cref{table:all_MJD}. Hedging instruments consist of 3-months ATM calls and puts priced with implied volatility modeled with a log-AR(1) dynamics (see \cref{subsec:dynamics_IV} for model description and \cref{table:all_OU_params} for parameters values). The training of neural networks is done as described in \cref{subsubsec:hyperparameters}. Values for the $\text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ measures are expressed relative to $\text{CVaR}_{0.90}$ (\% increase).
\end{table}
Numerical values reported in \cref{table:sensitivity_analysis_CVAR_090_095_099} indicate that with option hedges, an increase in the confidence level parameter of the $\text{CVaR}_{\alpha}$ measure leads to larger equal risk prices $C_0^{\star}$ and residual hedging risk $\epsilon^{\star}$ across all examples. These results confirm that the finding of \cite{carbonneau2020equal} with respect to the sensitivity of $C_0^{\star}$ and $\epsilon^{\star}$ to the risk aversion of the hedger is robust to using exclusively options as hedging instruments. Furthermore, values for equal risk prices $C_0^{\star}$ show a largest increase when using $\text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ as compared to $\text{CVaR}_{0.90}$ for OTM puts, followed by ATM and ITM moneyness levels; the same conclusion was observed in \cite{carbonneau2020equal} when trading the underlying stock.
The increase in $C_0^{\star}$ with the risk aversion level of the hedger stems from the thicker right tail of the short position hedging error than for the long position hedging error. The latter observation is consistent with previous analyses: while option hedges are more effective than stock hedges in the presence of equity jump risk as demonstrated in \cref{subsec:sensitivity_MJD}, their inclusion within hedging portfolios does not fully mitigate the asymmetry in tail risk of the residual hedging error.
\subsection{Benchmarking of equal risk prices to variance-optimal premiums}
\label{subsec:benchmark_QDH_ERP}
This section presents the benchmarking of equal risk prices to derivative premiums obtained with \textit{variance-optimal hedging} procedures (VO, \cite{schweizer1995variance}), also commonly called \textit{global quadratic hedging}. Variance-optimal hedging solves jointly for the initial capital investment and a self-financing strategy minimizing the expected value of the squared hedging error:
\begin{align}
\underset{\delta\in \Pi, V_0 \in \mathbb{R}}{\min} \, \mathbb{E} \left[\left(\Phi(S_{N}^{(0,b)}) -B_{N}(V_0 + G_{N}^{\delta})\right)^{2}\right]. \label{eq:ref_QDH_optimization}
\end{align}
The optimized initial capital investment denoted hereafter as $C_0^{(VO)}$ can be viewed as the production cost of $\Phi$, since the resulting dynamic trading strategy replicates the derivative's payoff as closely as possible in a quadratic sense. The optimization problem \eqref{eq:ref_QDH_optimization} can also be solved in a similar fashion as the non-quadratic global hedging problems embedded in the ERP framework, but with two distinctions: the initial capital investment is treated as an additional trainable parameter and a single neural network is considered since the optimal trading strategy is the same for the long and short position due to the quadratic penalty.\footnote{
%
\cite{cao2020discrete} showed that the deep hedging algorithm for variance-optimal hedging problems provides good approximations of optimal initial capital investments by comparing the optimized values to known formulas.
%
} The reader is referred to \cref{appendix:QDH_price} for a complete description of the numerical scheme for variance-optimal hedging implemented in this study.
The setup considered for the examination of this benchmarking is the same as in \cref{subsec:sensitivity_MJD} with the MJD dynamics under the three jump risk scenarios, with the exception of the confidence level of the $\text{CVaR}_{\alpha}$ measure, which is studied at first with $\alpha = 0.95$ fixed as in \cref{subsec:sensitivity_MJD} and \cref{subsec:sensitivity_GARCH}, and subsequently across $\alpha = 0.90, 0.95$ and $0.99$ as in \cref{subsec:sensitivity_risk_aversion}. Note that the authors also conducted the same experiments under the setup of \cref{subsec:sensitivity_GARCH} with volatility risk, and found that the main qualitative conclusions are very similar. The reader is referred to Table SM$6$ and Table SM$8$ of the Supplementary Material for the benchmarking of ERP to VO procedures in the presence of volatility risk.
\subsubsection{Benchmarking results}
\cref{table:ERP_RN_prices_one_year_JUMPS} presents benchmarking results of equal risk prices $C_0^{\star}$ to variance-optimal prices $C_0^{(VO)}$ under the MJD dynamics with the $\text{CVaR}_{0.95}$ measure.
%
\begin{table} [ht]
\caption {Equal risk prices $C_{0}^{\star}$ and variance-optimal (VO) prices $C_0^{(VO)}$ with jump risk for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$.} \label{table:ERP_RN_prices_one_year_JUMPS}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$\text{OTM}$} & & \multicolumn{3}{c}{$\text{ATM}$} & & \multicolumn{3}{c}{$\text{ITM}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Jump Scenario & $(1)$ & $(2)$ & $(3)$ & & $(1)$ & $(2)$ & $(3)$ & & $(1)$ & $(2)$ & $(3)$ \\
%
\hline\noalign{\medskip}
%
\multicolumn{4}{l}{\underline{\emph{$C_0^{(VO)}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $1.62$ & $1.77$ & $1.92$ & & $4.71$ & $4.79$ & $4.80$ & & $10.20$ & $10.18$ & $10.12$ \\
$\text{Monthly stock}$ & $1.55$ & $1.73$ & $1.86$ & & $4.62$ & $4.72$ & $4.72$ & & $10.14$ & $10.11$ & $10.05$ \\
$\text{1-month options}$ & $1.71$ & $2.04$ & $2.38$ & & $4.82$ & $5.11$ & $5.31$ & & $10.27$ & $10.45$ & $10.52$ \\
$\text{3-months options}$ & $1.58$ & $1.83$ & $2.08$ & & $4.64$ & $4.82$ & $4.97$ & & $10.11$ & $10.15$ & $10.21$ \\
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $17\%$ & $45\%$ & $75\%$ & & $11\%$ & $25\%$ & $42\%$ & & $6\%$ & $15\%$ & $20\%$ \\
$\text{Monthly stock}$ & $27\%$ & $51\%$ & $78\%$ & & $9\%$ & $22\%$ & $35\%$ & & $6\%$ & $13\%$ & $18\%$ \\
$\text{1-month options}$ & $6\%$ & $10\%$ & $7\%$ & & $3\%$ & $5\%$ & $6\%$ & & $2\%$ & $4\%$ & $3\%$ \\
$\text{3-months options}$ & $10\%$ & $14\%$ & $15\%$ & & $5\%$ & $6\%$ & $6\%$ & & $3\%$ & $4\%$ & $4\%$ \\
%
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the Merton Jump-Diffusion model for the underlying (see \cref{subsubsec:MJD_model} for model description). Three different sets of parameters values are considered with $\lambda = \{1, 0.25, 0.08\}$, $\mu_J = \{-0.05, -0.10, -0.20\}$ and $\sigma_J = \{0.05, 0.10, 0.15\}$ respectively for the jump scenario $1$, $2$, and $3$ (see \cref{table:all_MJD} for all parameters values). \textit{Hedging instruments}: daily or monthly rebalancing with the underlying stock and 1-month or 3-months options with ATM calls and puts. Options used as hedging instruments are priced with implied volatility modeled with a log-AR(1) dynamics (see \cref{subsec:dynamics_IV} for model description and \cref{table:all_OU_params} for parameters values). The training of neural networks for ERP and VO hedging is done as described in \cref{subsubsec:hyperparameters} and \cref{appendix:QDH_price}, respectively. The confidence level of the CVaR measure is $\alpha = 0.95$. $C_0^{\star}$ are expressed relative to $C_0^{(VO)}$ (\% increase).
\end{table}
%
Numerical experiments show that $C_0^{\star}$ is at least larger than $C_0^{(VO)}$ for all examples, but the relative increase is always smaller and less sensitive to jump severity when trading options. Furthermore, the relative increase in derivative premiums observed with the ERP framework over VO hedging is the largest for OTM puts, followed by ATM and ITM options across all jump risk scenarios and hedging instruments. For instance, the relative increase in $C_0^{\star}$ as compared to $C_0^{(VO)}$ when trading the daily stock ranges from jump scenario $1$ to scenario $3$ between $17\%$ to $75\%$ for OTM puts, $11\%$ to $42\%$ for ATM and $6\%$ to $20\%$ for ITM options.\footnote{
%
The relative increase is computed as $\frac{C_0^{\star}}{C_0^{(VO)}}-1$ for all examples.
%
} On the other hand, the relative increase in $C_0^{\star}$ as compared to $C_0^{(VO)}$ is much less sensitive to jump severity when trading 1-month options by ranging from scenario $1$ to scenario $3$ between $6\%$ to $10\%$ for OTM puts, $3\%$ to $6\%$ for ATM and $2\%$ to $4\%$ for ITM. Based on these results, we can assert that although both derivative valuation schemes are consistent with optimal trading criteria, the choice of hedging instrument and pricing procedure (hence implicitly of the treatment of hedging gains and losses) has a material impact on resulting derivative premiums and must thus be carefully chosen.
This smaller disparity between equal risk and variance-optimal prices with option hedges is in line with previous analyses: in the presence of jump or volatility risk, hedging with options entails significant reduction of the level market incompleteness as compared to trading solely the underlying stock. In such cases, premiums obtained with both derivative valuation approaches should be closer with the limiting case of being the same in a complete market.\footnote{
%
To further illustrate this phenomenon, the authors also performed the same benchmarking with the Black-Scholes dynamics under which market incompleteness solely stems from discrete-time trading. The latter results are presented in the Supplementary Material. Numerical values show that under the Black-Scholes dynamics, trading the underlying stock on a daily basis leads for most combinations of moneyness level and yearly volatility to the closest derivative premiums between ERP and VO procedures as compared to the other hedging instruments (see Table SM5). Also, as expected under the Black-Scholes dynamics, daily stock hedging entails the smallest level of residual hedging risk across the different hedging instruments (see Table SM1).
%
} These observations expand upon the work of \cite{carbonneau2020equal}, which shows that equal risk prices of puts obtained by hedging solely with the underlying stock are always larger than risk-neutral prices computed under convential change of measures. Indeed, benchmarking results presented in this current paper provide important novel insights into this price inflation phenomenon observed with the ERP framework: the disparity between equal risk and variance-optimal prices is always significantly smaller and less sensitive to stylized features of risky assets (e.g. jump or volatility risk) when option hedges are considered instead of trading exclusively the underlying stock.
Moreover, \cref{table:ERP_vs_VO_sensitivity_alpha_JUMP} presents benchmarking results of $C_0^{\star}$ to $C_0^{(VO)}$ with $\text{CVaR}_{0.90}, \text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ measures with $3$-months option hedges.
\begin{table}
\caption {Sensitivity analysis of equal risk prices $C_{0}^{\star}$ with $\text{CVaR}_{0.90}, \text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ measures to variance-optimal (VO) prices $C_0^{(VO)}$ under jump risk for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$.} \label{table:ERP_vs_VO_sensitivity_alpha_JUMP}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$\text{OTM}$} & & \multicolumn{3}{c}{$\text{ATM}$} & & \multicolumn{3}{c}{$\text{ITM}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Jump Scenario & $(1)$ & $(2)$ & $(3)$ & & $(1)$ & $(2)$ & $(3)$ & & $(1)$ & $(2)$ & $(3)$ \\
%
\hline\noalign{\medskip}
%
%
$C_0^{(VO)}$ & $1.58$ & $1.83$ & $2.08$ & & $4.64$ & $4.82$ & $4.97$ & & $10.11$ & $10.15$ & $10.21$ \\
%
$C_0^{\star}(\text{CVaR}_{0.90})$ & $3\%$ & $2\%$ & $0\%$ & & $2\%$ & $2\%$ & $2\%$ & & $2\%$ & $2\%$ & $2\%$ \\
$C_0^{\star}(\text{CVaR}_{0.95})$ & $10\%$ & $14\%$ & $15\%$ & & $5\%$ & $6\%$ & $6\%$ & & $3\%$ & $4\%$ & $4\%$ \\
$C_0^{\star}(\text{CVaR}_{0.99})$ & $32\%$ & $43\%$ & $55\%$ & & $10\%$ & $12\%$ & $16\%$ & & $6\%$ & $6\%$ & $8\%$ \\
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the Merton Jump-Diffusion model for the underlying (see \cref{subsubsec:MJD_model} for model description). Three different sets of parameters values are considered with $\lambda = \{1, 0.25, 0.08\}$, $\mu_J = \{-0.05, -0.10, -0.20\}$ and $\sigma_J = \{0.05, 0.10, 0.15\}$ respectively for jump scenario $1$, $2$, and $3$ (see \cref{table:all_MJD} for all parameters values). Hedging instruments consist of 3-months ATM calls and puts priced with implied volatility modeled as a log-AR(1) dynamics (see \cref{subsec:dynamics_IV} for model description and \cref{table:all_OU_params} for parameters values). The training of neural networks for ERP and VO hedging is done as described in \cref{subsubsec:hyperparameters} and \cref{appendix:QDH_price}, respectively. $C_0^{\star}$ with $\text{CVaR}_{0.90}, \text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ are expressed relative to $C_0^{(VO)}$ (\% increase).
\end{table}
Values presented in this benchmarking demonstrate the ability of ERP, through the choice of convex risk measures, to span a large interval of prices which is close to encompass the variance-optimal premium. Indeed, under the $\text{CVaR}_{0.90}$ measure, we observe that $C_0^{\star}$ values are very close to $C_0^{(VO)}$ where the relative difference ranges between $0\%$ and $3\%$ across all moneynesses and jump risk scenarios. On the other hand, optimizing trading policies with more risk averse agents, i.e. with $\text{CVaR}_{0.95}$ or $\text{CVaR}_{0.99}$, provides a very wide range of derivative premiums with the ERP framework, especially for the OTM moneyness level. It is very interesting to note that this added flexibility of ERP procedures for pricing derivatives does not come at the expense of less effective hedging policies. Indeed, a major drawback of variance-optimal hedging lies in penalizing equally gains and losses through a quadratic penalty for hedging shortfalls. Conversely, the long and short trading policies solving the non-quadratic global hedging problems of the ERP framework are optimized to minimize a loss function which is possibly more in line with the financial objectives of the hedger by mainly (and most often exclusively) penalizing hedging losses, not gains.
\section{Conclusion}
\label{section:conclusion}
This paper studies the equal risk pricing (ERP) framework for pricing and hedging European derivatives in discrete-time with multiple hedging instruments. The ERP approach sets derivative prices as the value such that the optimally hedged residual risk of the long and short positions in the contingent claim are equal. The ERP setup of \cite{marzban2020equal} is considered where residual hedging risk is quantified through convex measures. The main objective of this current paper is in assessing the impact of including options within hedges on the equal risk price $C_0^{\star}$ and on the level of market incompleteness quantified by our $\epsilon^{\star}$-metrics. A specific focus is on the examination of the interplay between different stylized features of equity jump and volatility risks and the use of options as hedging instruments within the ERP framework.
The numerical scheme of \cite{carbonneau2020equal}, which relies on the deep hedging algorithm of \cite{buehler2019deep}, is used to solve the embedded global hedging problems of the ERP framework through the representation of the long and short trading policies with two distinct long-short term memory (LSTM) neural networks.
%
Sensitivity analyses with Monte Carlo simulations are performed under several empirically plausible sets of parameters for the jump and volatility risk models in order to highlight the impact of different stylized features of the models on $C_0^{\star}$ and $\epsilon^{\star}$. Numerical values indicate that in the presence of jump risk, hedging with options entails a significant reduction of both equal risk prices and market incompleteness as compared to hedging solely with the underlying stock. The latter stems from the fact that using options as hedging instruments rather than only the underlying stock shrinks the asymmetry of tail risk, which tends to both shrink option prices and reduce market incompleteness. On the other hand, in the presence of volatility risk, while option hedges can reduce equal risk prices as compared to stock hedges, the impact can be marginal and is highly sensitive to the moneyness level of the put option being priced as well as to the maturity of traded options. This can be explained by the fact that while the impact of jump risk dynamics considered in this paper is asymmetric by entailing significantly more weight on the right (resp. left) tail of the short (resp. long) hedging error through predominantly negative jumps, volatility risk impacts both upside and downside risk. Furthermore, additional experiments conducted show that the average price level of short-term options used as hedging instruments is effectively reflected into the equal risk price of longer-term options. The latter highlights the potential of the ERP framework as a fair valuation approach providing prices consistent with observable market prices. Thus, ERP could be applied for instance in the context of pricing over-the-counter derivatives with vanilla calls and puts hedges or pricing less liquid long-term derivatives (e.g. LEAPS contracts) with shorter-term liquid options.
Moreover, the benchmarking of equal risk prices to variance-optimal derivative premiums $C_0^{(VO)}$ is performed. The deep hedging algorithm is also used as the numerical scheme to solve the variance-optimal hedging problems. Numerical results show that while $C_0^{\star}$ tends to be larger than $C_0^{(VO)}$, trading options entails much smaller disparity between equal risk and variance-optimal prices as compared to trading only the underlying stock in the presence of jump or volatility risk. The latter is due to the market incompleteness being significantly smaller when option hedges are used to mitigate jump and volatility risks. Furthermore, additional experiments conducted demonstrate the ability of ERP to span a large interval of prices through the choice of convex risk measures, which is close to encompass the variance-optimal premium.
\section{Acknowledgements}
Alexandre Carbonneau gratefully acknowledges financial support from the Fonds de recherche du Qu\'ebec - Nature et technologies (FRQNT, grant number 205683) and The Montreal Exchange. Fr{\'e}d{\'e}ric Godin gratefully acknowledges financial support from Natural Sciences and Engineering Research Council of Canada (NSERC, grant number RGPIN-2017-06837).
\bibliographystyle{apalike}
\section{0pt}{12pt plus 4pt minus 2pt}{6pt plus 2pt minus 2pt}
\titlespacing\subsection{0pt}{12pt plus 4pt minus 2pt}{3pt plus 2pt minus 3pt}
\titlespacing\subsubsection{0pt}{12pt plus 4pt minus 2pt}{0pt plus 2pt minus 3pt}
\singlespacing
\def\boxit#1{\vbox{\hrule\hbox{\vrule\kern6pt
\vbox{\kern6pt#1\kern6pt}\kern6pt\vrule}\hrule}}
\def\bfred#1{{\color{red}\bf#1}}
\def\bfblue#1{{\color{blue}\bf#1}}
\def\red#1{{\color{red}#1}}
\def\blue#1{{\color{blue}#1}}
\definecolor{orange}{rgb}{1,0.5,0}
\definecolor{MyDarkBlue}{rgb}{0,0.08,0.45}
\def\orange#1{{\color{orange}#1}}
\def\fredcomment#1{\vskip 2mm\boxit{\vskip 2mm{\color{MyDarkBlue}\bf#1}
{\color{MyDarkBlue}\bf -- Fred\vskip 2mm}}\vskip 2mm}
\def\alexcomment#1{\vskip 2mm\boxit{\vskip 2mm{\color{red}\bf#1}
{\color{red}\bf -- Alex\vskip 2mm}}\vskip 2mm}
\newtheorem{corollary}{Corollary}[section]
\newtheorem{definition}{Definition}
\newtheorem{example}{Example}[section]
\newtheorem{lemma}{Lemma}[section]
\newtheorem{proposition}{Proposition}[section]
\newtheorem{remark}{Remark}[section]
\newtheorem{theorem}{Theorem}[section]
\newtheorem{problem}{Problem}[section]
\newtheorem{assumption}{Assumption}[section]
\newtheorem{Def}{Definition}[section]
\renewcommand{\baselinestretch}{1.5}
\newcommand\blfootnote[1]{%
\begingroup
\renewcommand\thefootnote{}\footnote{#1}%
\addtocounter{footnote}{-1}%
\endgroup
}
\setcounter{table}{0}
\renewcommand{\thetable}{SM\arabic{table}}
\setcounter{section}{0}
\renewcommand{\thesection}{A\arabic{section}}
\begin{document}
\title{\Large \bfseries Supplementary Material: Deep Equal Risk Pricing of Financial Derivatives with Multiple Hedging Instruments}
\author[a] {Alexandre Carbonneau\thanks{Corresponding author\vspace{0.2em} \newline
{\it Email addresses:} \href{mailto:<EMAIL>@mail.concordia.ca} (Alexandre Carbonneau), \href{mailto:<EMAIL>@concordia.ca} (Fr\'ed\'eric Godin).}}
\author[b]{Fr\'ed\'eric Godin}
\affil[a,b]{{\small Concordia University, Department of Mathematics and Statistics, Montr\'eal, Canada}}
\vspace{-10pt}
\date{
\today}
\maketitle \thispagestyle{empty}
This supplementary material presents additional numerical experiments of the ERP framework.
%
\cref{sec:BSM_ERP_analysis} presents values for $C_0^{\star}$ and $\epsilon^{\star}$ under the Black-Scholes model with the same setup as in Table $3$ and Table $5$.
%
\cref{sec:sens_avg_IV} presents the sensitivity analysis of $C_0^{\star}$ and $\epsilon^{\star}$ to implied volatility risk with a similar setup as in Table $5$.
%
\cref{sec:add_sens_analysis_convex} presents the corresponding sensitivity analysis of Table $6$ under jump risk scenario $1$ and $3$ as well as under the three sets of parameters for volatility risk.
%
\cref{sec:add_benchmark_ERP_VO} presents the benchmarking of $C_0^{\star}$ to $C_0^{(VO)}$ under the BSM and GARCH dynamics with the same setup as in Table $7$ and Table $8$.
\section{Sensitivity of equal risk pricing under Black-Scholes}
\label{sec:BSM_ERP_analysis}
The Black-Scholes dynamics models log-returns as independent and identically distributed Gaussian random variables of daily periodic mean $\frac{1}{252}(\mu - \frac{\sigma^{2}}{2})$ and variance $\frac{\sigma^{2}}{252}$. Daily log-returns under this model have the representation
\begin{align}
\tilde{y}_{j,n} = \frac{1}{252}\left(\mu - \frac{\sigma^{2}}{2}\right) + \sigma \sqrt{\frac{1}{252}}\epsilon_{j,n}, \label{eq:ref_BSM_log_ret}
\end{align}
where $\mu$ and $\sigma$ are the parameters on a yearly scale. Under the BSM, feature vectors of the neural networks do not require the inclusion of additional state variables (i.e. $\varphi_n = 0$ for all time steps $n$ in $(3.4)$ and $(3.5)$).
\cref{table:ERP_one_year_put_BSM_analysis} presents the equal risk prices $C_0^{\star}$ and $\epsilon^{\star}$-metrics
under the BSM with $\mu=0.1$ and $\sigma = 0.10, 0.15$ and $0.20$.
\begin{table}
\caption {Equal risk prices $C_{0}^{\star}$ and residual hedging risk $\epsilon^{\star}$ under the Black-Scholes model for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$.} \label{table:ERP_one_year_put_BSM_analysis}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$\text{OTM}$} & & \multicolumn{3}{c}{$\text{ATM}$} & & \multicolumn{3}{c}{$\text{ITM}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Annual volatility ($\sigma$) & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ \\
%
\hline\noalign{\medskip}
%
\multicolumn{4}{l}{\underline{\emph{$C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $0.43$ & $1.48$ & $2.83$ & & $2.65$ & $4.60$ & $6.68$ & & $8.37$ & $10.19$ & $12.08$ \\
$\text{Monthly stock}$ & $0.49$ & $1.67$ & $3.15$ & & $2.71$ & $4.70$ & $6.72$ & & $8.56$ & $10.42$ & $12.39$ \\
$\text{1-month options}$ & $0.44$ & $1.56$ & $3.00$ & & $2.74$ & $4.70$ & $6.71$ & & $8.48$ & $10.30$ & $12.27$ \\
$\text{3-months options}$ & $0.44$ & $1.59$ & $3.08$ & & $2.81$ & $4.80$ & $6.82$ & & $8.58$ & $10.47$ & $12.51$ \\
%
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$\epsilon^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $0.19$ & $0.41$ & $0.65$ & & $0.38$ & $0.64$ & $1.00$ & & $0.39$ & $0.67$ & $0.88$ \\
$\text{Monthly stock}$ & $0.50$ & $1.52$ & $2.59$ & & $1.54$ & $2.53$ & $3.48$ & & $1.51$ & $2.59$ & $3.60$ \\
$\text{1-month options}$ & $0.19$ & $0.57$ & $0.94$ & & $0.57$ & $0.91$ & $1.30$ & & $0.64$ & $1.01$ & $1.37$ \\
$\text{3-months options}$ & $0.29$ & $0.93$ & $1.55$ & & $0.96$ & $1.55$ & $2.11$ & & $1.09$ & $1.76$ & $2.35$ \\
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$\epsilon^{\star}/C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $0.44$ & $0.28$ & $0.23$ & & $0.14$ & $0.14$ & $0.15$ & & $0.05$ & $0.07$ & $0.07$ \\
$\text{Monthly stock}$ & $1.01$ & $0.91$ & $0.82$ & & $0.57$ & $0.54$ & $0.52$ & & $0.18$ & $0.25$ & $0.29$ \\
$\text{1-month options}$ & $0.44$ & $0.37$ & $0.31$ & & $0.21$ & $0.19$ & $0.19$ & & $0.07$ & $0.10$ & $0.11$ \\
$\text{3-months options}$ & $0.67$ & $0.59$ & $0.50$ & & $0.34$ & $0.32$ & $0.31$ & & $0.13$ & $0.17$ & $0.19$ \\
%
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the Black-Scholes model for the underlying with $\mu=0.1$ and $\sigma=0.1, 0.15$ and $0.20$ (see \eqref{eq:ref_BSM_log_ret} for model description). \textit{Hedging instruments}: daily or monthly rebalancing with the underlying stock and 1-month or 3-months options with ATM calls and puts. Options used as hedging instruments are priced with implied volatility modeled as a log-AR(1) dynamics with $\kappa = 0.15, \sigma_{\text{IV}} = 0.06$ and $\varrho = -0.6$ for all cases, and $\vartheta$ set to $\log(0.10), \log(0.15)$ and $\log(0.20)$ when $\sigma = 0.10, 0.15$ and $0.20$, respectively (see Section $4.1.4$ for log-AR(1) model description). The training of neural networks is done as described in Section $4.1.5$. The confidence level of the CVaR measure is $\alpha = 0.95$.
\end{table}
\section{Sensitivity of equal risk pricing to implied volatility}
\label{sec:sens_avg_IV}
\cref{table:ERP_one_year_put_IV_analysis} presents the sensitivity analysis of the ERP framework to the long-run implied volatility when trading $1$-month or $3$-months options. These options are priced with implied volatility modeled as a log-AR(1) dynamics with parameters of Table $2$ for $\{\kappa, \sigma_{\text{IV}}, \varrho\}$ (i.e. $\kappa = 0.15, \sigma_{\text{IV}} = 0.06$ and $\varrho = -0.6$), and the long-run parameter $\vartheta$ set to $\log(0.14), \log(0.15)$ and $\log(0.16)$, respectively. The underlying stock dynamics is the GARCH model with the set of parameters corresponding to the stationary yearly volatility of $15\%$ in Table $4$.
\begin{table}
\caption {Sensitivity analysis of equal risk prices $C_{0}^{\star}$ and residual hedging risk $\epsilon^{\star}$ to implied volatility (IV) risk for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$.} \label{table:ERP_one_year_put_IV_analysis}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$\text{OTM}$} & & \multicolumn{3}{c}{$\text{ATM}$} & & \multicolumn{3}{c}{$\text{ITM}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Long-run IV & $14\%$ & $15\%$ & $16\%$ & & $14\%$ & $15\%$ & $16\%$ & & $14\%$ & $15\%$ & $16\%$ \\
%
\hline\noalign{\medskip}
%
\multicolumn{4}{l}{\underline{\emph{$C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{1-month options}$ & $1.52$ & $1.74$ & $1.98$ & & $4.52$ & $4.87$ & $5.20$ & & $10.01$ & $10.32$ & $10.66$ \\
$\text{3-months options}$ & $1.86$ & $2.07$ & $2.28$ & & $4.75$ & $5.08$ & $5.39$ & & $10.08$ & $10.38$ & $10.66$ \\
%
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$\epsilon^{\star}$}}} \\\noalign{\smallskip}
%
$\text{1-month options}$ & $0.61$ & $0.65$ & $0.66$ & & $0.96$ & $1.00$ & $0.96$ & & $0.99$ & $1.03$ & $1.01$ \\
$\text{3-months options}$ & $1.28$ & $1.32$ & $1.36$ & & $1.72$ & $1.77$ & $1.78$ & & $1.63$ & $1.68$ & $1.72$ \\
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$\epsilon^{\star}/C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{1-month options}$ & $0.40$ & $0.37$ & $0.34$ & & $0.21$ & $0.20$ & $0.18$ & & $0.10$ & $0.10$ & $0.09$ \\
$\text{3-months options}$ & $0.69$ & $0.64$ & $0.60$ & & $0.36$ & $0.35$ & $0.33$ & & $0.16$ & $0.16$ & $0.16$ \\
%
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the GJR-GARCH(1,1) model for the underlying with $\mu=3.968\text{e-}04,$ $\omega=1.964\text{e-}06$, $\upsilon = 0.05$, $\gamma = 0.6$ and $\beta = 0.91$ which entails stationary yearly volatility of $15\%$ (see Section $4.1.3$ for model description). Hedging instruments consist of 1-month or 3-months options with ATM calls and puts. The latter options are priced with implied volatility modeled as a log-AR(1) dynamics with $\kappa = 0.15, \sigma_{\text{IV}} = 0.06$ and $\varrho = -0.6$ for all cases, and the long-run parameter $\vartheta$ set to $\log(0.14), \log(0.15)$ or $\log(0.16)$ (see Section $4.1.4$ for the log-AR(1) model description). The training of neural networks is done as described in Section $4.1.5$. The confidence level of the CVaR measure is $\alpha = 0.95$.
\end{table}
\section{Additional sensitivity analysis convex risk measure}
\label{sec:add_sens_analysis_convex}
\cref{table:sensitivity_analysis_CVAR_090_095_099_jump_risk} presents the sensitivity analysis of the ERP framework to the choice of convex risk measure when trading $3$-months options under jump risk scenario $1$ and $3$ of Table $1$. \cref{table:sensitivity_analysis_CVAR_090_095_099_vol_risk} presents this sensitivity analysis under volatility risk with the three sets of parameters described in Table $4$.
%
\begin{table} [ht]
\caption {Sensitivity analysis of equal risk prices $C_{0}^{\star}$ and residual hedging risk $\epsilon^{\star}$ for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$ under the MJD dynamics with jump risk scenario $1$ and scenario $3$.} \label{table:sensitivity_analysis_CVAR_090_095_099_jump_risk}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$C_{0}^{\star}$} & & \multicolumn{3}{c}{$\epsilon^{\star}$} & & \multicolumn{3}{c}{$\epsilon^{\star}/C_{0}^{\star}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Moneyness & $\text{OTM}$ & $\text{ATM}$ & $\text{ITM}$ & & $\text{OTM}$ & $\text{ATM}$ & $\text{ITM}$ & & $\text{OTM}$ & $\text{ATM}$ & $\text{ITM}$ \\
%
\hline\noalign{\medskip}
%
\multicolumn{4}{l}{\underline{\textbf{{\emph{Jump risk scenario $1$}}}}} \\\noalign{\smallskip}
$\text{CVaR}_{0.90}$ & $1.63$ & $4.76$ & $10.32$ & & $0.78$ & $1.28$ & $1.42$ & & $0.48$ & $0.27$ & $0.14$ \\
$\text{CVaR}_{0.95}$ & $6\%$ & $2\%$ & $1\%$ & & $32\%$ & $24\%$ & $20\%$ & & $24\%$ & $21\%$ & $19\%$ \\
$\text{CVaR}_{0.99}$ & $28\%$ & $7\%$ & $4\%$ & & $103\%$ & $70\%$ & $65\%$ & & $59\%$ & $58\%$ & $58\%$ \\
\\
\multicolumn{4}{l}{\underline{\textbf{{\emph{Jump risk scenario $3$}}}}} \\\noalign{\smallskip}
$\text{CVaR}_{0.90}$ & $2.07$ & $5.08$ & $10.44$ & & $1.21$ & $1.61$ & $1.57$ & & $0.58$ & $0.32$ & $0.15$ \\
$\text{CVaR}_{0.95}$ & $15\%$ & $4\%$ & $1\%$ & & $39\%$ & $25\%$ & $20\%$ & & $21\%$ & $21\%$ & $19\%$ \\
$\text{CVaR}_{0.99}$ & $56\%$ & $13\%$ & $5\%$ & & $128\%$ & $76\%$ & $68\%$ & & $47\%$ & $56\%$ & $60\%$ \\
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the Merton Jump-Diffusion model for the underlying (see Section $4.1.2$ for model description). Parameters for jump risk scenario $1$ and $3$ are presented in Table $1$.
Hedging instruments consist of 3-months ATM calls and puts priced with implied volatility modeled with a log-AR(1) dynamics (see Section $4.1.4$ for model description and Table $2$ for parameters values). The training of neural networks is done as described in Section $4.1.5$. Values for the $\text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ measures are expressed relative to $\text{CVaR}_{0.90}$ (\% increase).
\end{table}
%
\begin{table} [ht]
\caption {Sensitivity analysis of equal risk prices $C_{0}^{\star}$ and residual hedging risk $\epsilon^{\star}$ for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$ under the GARCH dynamics.} \label{table:sensitivity_analysis_CVAR_090_095_099_vol_risk}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$C_{0}^{\star}$} & & \multicolumn{3}{c}{$\epsilon^{\star}$} & & \multicolumn{3}{c}{$\epsilon^{\star}/C_{0}^{\star}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Moneyness & $\text{OTM}$ & $\text{ATM}$ & $\text{ITM}$ & & $\text{OTM}$ & $\text{ATM}$ & $\text{ITM}$ & & $\text{OTM}$ & $\text{ATM}$ & $\text{ITM}$ \\
%
\hline\noalign{\medskip}
%
\multicolumn{4}{l}{\underline{\textbf{{\emph{Stationary volatility: $10\%$}}}}} \\\noalign{\smallskip}
$\text{CVaR}_{0.90}$ & $0.65$ & $2.89$ & $8.42$ & & $0.40$ & $0.83$ & $0.86$ & & $0.61$ & $0.29$ & $0.10$ \\
$\text{CVaR}_{0.95}$ & $17\%$ & $4\%$ & $1\%$ & & $46\%$ & $32\%$ & $21\%$ & & $25\%$ & $27\%$ & $20\%$ \\
$\text{CVaR}_{0.99}$ & $70\%$ & $13\%$ & $3\%$ & & $159\%$ & $94\%$ & $66\%$ & & $52\%$ & $72\%$ & $60\%$ \\
\\
\multicolumn{4}{l}{\underline{\textbf{{\emph{Stationary volatility: $15\%$}}}}} \\\noalign{\smallskip}
$\text{CVaR}_{0.90}$ & $1.88$ & $4.91$ & $10.28$ & & $0.97$ & $1.39$ & $1.40$ & & $0.51$ & $0.28$ & $0.14$ \\
$\text{CVaR}_{0.95}$ & $10\%$ & $4\%$ & $1\%$ & & $36\%$ & $27\%$ & $21\%$ & & $24\%$ & $23\%$ & $19\%$ \\
$\text{CVaR}_{0.99}$ & $38\%$ & $11\%$ & $4\%$ & & $122\%$ & $81\%$ & $69\%$ & & $61\%$ & $63\%$ & $63\%$ \\
\\
\multicolumn{4}{l}{\underline{\textbf{{\emph{Stationary volatility: $20\%$}}}}} \\\noalign{\smallskip}
$\text{CVaR}_{0.90}$ & $3.40$ & $6.95$ & $12.27$ & & $1.54$ & $1.93$ & $1.89$ & & $0.45$ & $0.28$ & $0.15$ \\
$\text{CVaR}_{0.95}$ & $7\%$ & $3\%$ & $1\%$ & & $31\%$ & $25\%$ & $22\%$ & & $22\%$ & $21\%$ & $20\%$ \\
$\text{CVaR}_{0.99}$ & $28\%$ & $10\%$ & $5\%$ & & $103\%$ & $77\%$ & $71\%$ & & $58\%$ & $61\%$ & $64\%$ \\
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the GJR-GARCH(1,1) model for the underlying with three sets of parameters implying stationary yearly volatilities of $10\%, 15\%$ and $20\%$ (see Section $4.1.3$ for model description and Table $4$ for parameters values). Hedging instruments consist of 3-months ATM calls and puts priced with implied volatility modeled with a log-AR(1) dynamics with $\kappa = 0.15, \sigma_{\text{IV}} = 0.06$ and $\varrho = -0.6$ for all cases, and $\vartheta$ set to $\log(0.10), \log(0.15)$ and $\log(0.20)$ when the GARCH stationary volatility is $10\%, 15\%$ and $20\%$, respectively (see Section $4.1.4$ for log-AR(1) model description).
The training of neural networks is done as described in Section $4.1.5$. Values for the $\text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ measures are expressed relative to $\text{CVaR}_{0.90}$ (\% increase).
\end{table}
\section{Benchmarking of equal risk prices to variance-optimal premiums}
\label{sec:add_benchmark_ERP_VO}
\cref{table:ERP_RN_prices_one_year_BSM} and \cref{table:ERP_RN_prices_one_year_VOL} present the benchmarking of $C_0^{\star}$ with the $\text{CVaR}_{0.95}$ risk measure to $C_0^{(VO)}$ with the same setup as Table $7$ under the Black-Scholes and GARCH dynamics, respectively. Furthermore, \cref{table:ERP_vs_VO_sensitivity_alpha_BSM} and \cref{table:ERP_vs_VO_sensitivity_alpha_VOL} perform the benchmarking of $C_0^{\star}$ to $C_0^{(VO)}$ with the $\text{CVaR}_{0.90}, \text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ measures and $3$-months option hedges under the BSM and GARCH dynamics, respectively. The latter experiments are performed with the same setup as Table $8$ under the MJD dynamics.
\begin{table}[ht]
\caption {Equal risk prices $C_{0}^{\star}$ and variance-optimal (VO) prices $C_0^{(VO)}$ under the Black-Scholes model for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$.} \label{table:ERP_RN_prices_one_year_BSM}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$\text{OTM}$} & & \multicolumn{3}{c}{$\text{ATM}$} & & \multicolumn{3}{c}{$\text{ITM}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} $\text{Annual volatility } (\sigma)$ & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ \\
%
\hline\noalign{\medskip}
%
\multicolumn{4}{l}{\underline{\emph{$C_0^{(VO)}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $0.38$ & $1.39$ & $2.77$ & & $2.62$ & $4.53$ & $6.46$ & & $8.34$ & $10.12$ & $12.04$ \\
$\text{Monthly stock}$ & $0.35$ & $1.36$ & $2.73$ & & $2.55$ & $4.47$ & $6.39$ & & $8.27$ & $10.07$ & $11.98$ \\
$\text{1-month options}$ & $0.42$ & $1.47$ & $2.84$ & & $2.67$ & $4.58$ & $6.52$ & & $8.35$ & $10.13$ & $12.06$ \\
$\text{3-months options}$ & $0.39$ & $1.44$ & $2.84$ & & $2.65$ & $4.56$ & $6.50$ & & $8.32$ & $10.12$ & $12.04$ \\
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $12\%$ & $6\%$ & $2\%$ & & $1\%$ & $2\%$ & $3\%$ & & $0\%$ & $1\%$ & $0\%$ \\
$\text{Monthly stock}$ & $40\%$ & $23\%$ & $15\%$ & & $6\%$ & $5\%$ & $5\%$ & & $3\%$ & $4\%$ & $3\%$ \\
$\text{1-month options}$ & $5\%$ & $6\%$ & $6\%$ & & $3\%$ & $3\%$ & $3\%$ & & $2\%$ & $2\%$ & $2\%$ \\
$\text{3-months options}$ & $11\%$ & $10\%$ & $9\%$ & & $6\%$ & $5\%$ & $5\%$ & & $3\%$ & $3\%$ & $4\%$ \\
%
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the Black-Scholes model for the underlying with $\mu=0.1$ and $\sigma=0.1, 0.15$ and $0.20$ (see \eqref{eq:ref_BSM_log_ret} for model description). \textit{Hedging instruments}: daily or monthly rebalancing with the underlying stock and 1-month or 3-months options with ATM calls and puts. Options used as hedging instruments are priced with implied volatility modeled as a log-AR(1) dynamics with $\kappa = 0.15, \sigma_{\text{IV}} = 0.06$ and $\varrho = -0.6$ for all cases, and $\vartheta$ set to $\log(0.10), \log(0.15)$ and $\log(0.20)$ when $\sigma = 0.10, 0.15$ and $0.20$, respectively (see Section $4.1.4$ for log-AR(1) model description). The training of neural networks for ERP and VO hedging is done as described in Section $4.1.5$ and Appendix A, respectively. The confidence level of the CVaR measure is $\alpha = 0.95$. $C_0^{\star}$ are expressed relative to $C_0^{(VO)}$ (\% increase).
\end{table}
\begin{table}[ht]
\caption {Equal risk prices $C_{0}^{\star}$ and variance-optimal (VO) prices $C_0^{(VO)}$ with volatility risk for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$.} \label{table:ERP_RN_prices_one_year_VOL}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$\text{OTM}$} & & \multicolumn{3}{c}{$\text{ATM}$} & & \multicolumn{3}{c}{$\text{ITM}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Stationary volatility & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ \\
%
\hline\noalign{\medskip}
%
\multicolumn{4}{l}{\underline{\emph{$C_0^{(VO)}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $0.78$ & $1.88$ & $3.24$ & & $2.94$ & $4.75$ & $6.60$ & & $8.29$ & $9.98$ & $11.81$ \\
$\text{Monthly stock}$ & $0.81$ & $1.91$ & $3.27$ & & $2.97$ & $4.80$ & $6.62$ & & $8.30$ & $10.01$ & $11.83$ \\
$\text{1-month options}$ & $0.50$ & $1.60$ & $3.01$ & & $2.75$ & $4.66$ & $6.60$ & & $8.32$ & $10.10$ & $12.04$ \\
$\text{3-months options}$ & $0.61$ & $1.74$ & $3.19$ & & $2.79$ & $4.72$ & $6.68$ & & $8.23$ & $10.05$ & $11.99$ \\
%
%
& & & & & & & & & & & \\
%
\multicolumn{4}{l}{\underline{\emph{$C_{0}^{\star}$}}} \\\noalign{\smallskip}
%
$\text{Daily stock}$ & $29\%$ & $25\%$ & $19\%$ & & $10\%$ & $13\%$ & $10\%$ & & $3\%$ & $6\%$ & $6\%$ \\
$\text{Monthly stock}$ & $45\%$ & $39\%$ & $30\%$ & & $13\%$ & $13\%$ & $14\%$ & & $7\%$ & $8\%$ & $9\%$ \\
$\text{1-month options}$ & $12\%$ & $9\%$ & $9\%$ & & $4\%$ & $4\%$ & $4\%$ & & $2\%$ & $2\%$ & $3\%$ \\
$\text{3-months options}$ & $24\%$ & $19\%$ & $14\%$ & & $8\%$ & $8\%$ & $7\%$ & & $3\%$ & $3\%$ & $4\%$ \\
%
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the GJR-GARCH(1,1) model for the underlying with three sets of parameters implying stationary yearly volatilities of $10\%, 15\%$ and $20\%$ (see Section $4.1.3$ for model description and Table $4$ for parameters values). \textit{Hedging instruments}: daily or monthly rebalancing with the underlying stock and 1-month or 3-months options with ATM calls and puts. Options used as hedging instruments are priced with implied volatility modeled as a log-AR(1) dynamics with $\kappa = 0.15, \sigma_{\text{IV}} = 0.06$ and $\varrho = -0.6$ for all cases, and $\vartheta$ set to $\log(0.10), \log(0.15)$ and $\log(0.20)$ when the GARCH stationary volatility is $10\%, 15\%$ and $20\%$, respectively (see Section $4.1.4$ for log-AR(1) model description). The training of neural networks for ERP and VO hedging is done as described in Section $4.1.5$ and Appendix A, respectively. The confidence level of the CVaR measure is $\alpha = 0.95$. $C_0^{\star}$ are expressed relative to $C_0^{(VO)}$ (\% increase).
\end{table}
\begin{table}[ht]
\caption {Sensitivity analysis of equal risk prices $C_{0}^{\star}$ with $\text{CVaR}_{0.90}, \text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ measures to variance-optimal (VO) prices $C_0^{(VO)}$ under the Black-Scholes model for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$.} \label{table:ERP_vs_VO_sensitivity_alpha_BSM}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$\text{OTM}$} & & \multicolumn{3}{c}{$\text{ATM}$} & & \multicolumn{3}{c}{$\text{ITM}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Annual volatility ($\sigma$) & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ \\
%
\hline\noalign{\medskip}
%
%
$C_0^{(VO)}$ & $0.39$ & $1.44$ & $2.84$ & & $2.65$ & $4.56$ & $6.50$ & & $8.32$ & $10.12$ & $12.04$ \\
%
$C_0^{\star}(\text{CVaR}_{0.90})$ & $-2\%$ & $4\%$ & $4\%$ & & $3\%$ & $3\%$ & $3\%$ & & $3\%$ & $2\%$ & $2\%$ \\
$C_0^{\star}(\text{CVaR}_{0.95})$ & $11\%$ & $10\%$ & $9\%$ & & $6\%$ & $5\%$ & $5\%$ & & $3\%$ & $3\%$ & $4\%$ \\
$C_0^{\star}(\text{CVaR}_{0.99})$ & $44\%$ & $26\%$ & $22\%$ & & $12\%$ & $13\%$ & $9\%$ & & $6\%$ & $7\%$ & $6\%$ \\
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the Black-Scholes model for the underlying with $\mu=0.1$ and $\sigma=0.1, 0.15$ and $0.20$ (see \eqref{eq:ref_BSM_log_ret} for model description). Hedging instruments consist of 3-months ATM calls and puts priced with implied volatility modeled as a log-AR(1) dynamics with $\kappa = 0.15, \sigma_{\text{IV}} = 0.06$ and $\varrho = -0.6$ for all cases, and $\vartheta$ set to $\log(0.10), \log(0.15)$ and $\log(0.20)$ when $\sigma = 0.10, 0.15$ and $0.20$, respectively (see Section $4.1.4$ for log-AR(1) model description). The training of neural networks for ERP and VO hedging is done as described in Section $4.1.5$ and Appendix A, respectively. $C_0^{\star}$ with $\text{CVaR}_{0.90}, \text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ are expressed relative to $C_0^{(VO)}$ (\% increase).
\end{table}
\begin{table}[ht]
\caption {Sensitivity analysis of equal risk prices $C_{0}^{\star}$ with $\text{CVaR}_{0.90}, \text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ measures to variance-optimal (VO) prices $C_0^{(VO)}$ under volatility risk for OTM ($K=90$), ATM ($K=100$) and ITM ($K=110$) put options of maturity $T=1$.} \label{table:ERP_vs_VO_sensitivity_alpha_VOL}
\renewcommand{\arraystretch}{1.15}
\begin{adjustwidth}{-1in}{-1in}
\centering
\begin{tabular}{lccccccccccc}
\hline\noalign{\smallskip}
& \multicolumn{3}{c}{$\text{OTM}$} & & \multicolumn{3}{c}{$\text{ATM}$} & & \multicolumn{3}{c}{$\text{ITM}$} \\
\cline{2-4}\cline{6-8}\cline{10-12} Stationary volatility & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ & & $10\%$ & $15\%$ & $20\%$ \\
%
\hline\noalign{\medskip}
%
%
$C_0^{(VO)}$ & $0.61$ & $1.74$ & $3.19$ & & $2.79$ & $4.72$ & $6.68$ & & $8.23$ & $10.05$ & $11.99$ \\
%
$C_0^{\star}(\text{CVaR}_{0.90})$ & $6\%$ & $8\%$ & $6\%$ & & $4\%$ & $4\%$ & $4\%$ & & $2\%$ & $2\%$ & $2\%$ \\
$C_0^{\star}(\text{CVaR}_{0.95})$ & $24\%$ & $19\%$ & $14\%$ & & $8\%$ & $8\%$ & $7\%$ & & $3\%$ & $3\%$ & $4\%$ \\
$C_0^{\star}(\text{CVaR}_{0.99})$ & $81\%$ & $49\%$ & $37\%$ & & $17\%$ & $16\%$ & $15\%$ & & $6\%$ & $6\%$ & $7\%$ \\
\noalign{\medskip}\hline
\end{tabular}%
\end{adjustwidth}
Notes: Results are computed based on $100,\!000$ independent paths generated from the GJR-GARCH(1,1) model for the underlying with three sets of parameters implying stationary yearly volatilities of $10\%, 15\%$ and $20\%$ (see Section $4.1.3$ for model description and Table $4$ for parameters values). Hedging instruments consist of 3-months ATM calls and puts priced with implied volatility modeled as a log-AR(1) dynamics with $\kappa = 0.15, \sigma_{\text{IV}} = 0.06$ and $\varrho = -0.6$ for all cases, and $\vartheta$ set to $\log(0.10), \log(0.15)$ and $\log(0.20)$ when the GARCH stationary volatility is $10\%, 15\%$ and $20\%$, respectively (see Section $4.1.4$ for log-AR(1) model description). The training of neural networks for ERP and VO hedging is done as described in Section $4.1.5$ and Appendix A, respectively. $C_0^{\star}$ with $\text{CVaR}_{0.90}, \text{CVaR}_{0.95}$ and $\text{CVaR}_{0.99}$ are expressed relative to $C_0^{(VO)}$ (\% increase).
\end{table}
\end{document} |
\section{\label{sec:intro}Introduction}
Hadronic time-like form factors will be measured in upcoming experiments to an unprecedented precision.
An understanding of these quantities which are displaying pronounced structures originating from hadron
resonances will contribute significantly to our knowledge on the relation between the hadrons' substructure
and the hadron spectrum. On the one hand, based on respective experimental and theoretical progress in the
last decades, it is by now evident that hadron resonances can be, at least in principle, described in terms of quarks
and gluons. The latter, being the QCD degrees of freedom, are considered to be complete in the sense that
they allow for a computation of every hadronic observable. Changing the perspective in an attempt to understand
the Strong Interaction starting from the low-energy regime, a possible way of phrasing the phenomenon of confinement
in QCD is the statement that all possible hadronic degrees of freedom will form also a complete set of physical states.
Therefore, the equivalence of descriptions of observables in either quark and glue or hadronic degrees of freedom
is a direct consequence of confinement and unitarity. This is the gist of a chain of arguments which can be
sophisticated and applied to many different phenomena involving hadrons. The related picture is known under the
name ``quark-hadron duality'', and its consequences have been verified on the qualitative as well as the
semi-quantitative level, for a review see, e.g., Ref.\ \cite{Melnitchouk:2005zr}. A verification of this duality is, beyond
the trivial fact of the absence of coloured states, the clearest experimental signature for confinement. To appreciate the
scope of such a scenario it is important to note that a perfect orthogonality of the quark-glue degrees of freedom
on the one hand and hadronic states on the other hand, and thus the perfect absence of ‘‘double-counting’’ in any of
the two ‘‘languages’’, is nothing else but another way to express confinement.
Gaining insight into the interplay between formation of hadronic bound states, consisting of quarks and gluons, and
the open decay channels of the respective resonance is an essential element of every study of time-like form factors
in kinematic regions close to a resonance. Here, attention should be paid to the fact that the hadron whose form
factor is investigated and the hadronic resonance which is apparent in the form factor are both to be described
as composite objects of quarks and gluons. The same is true for the hadronic decay products in a hadronic or
semi-leptonic decay of the resonance. This makes evident that an approach to calculate a time-like form factor from
QCD, or from a microscopic model based on QCD degrees of freedom, faces the challenging task to treat all elements
appearing in the calculation on the same footing and to a sufficient degree of sophistication if the result is intended to allow for conclusions on the dynamics underlying such a form factor.
Herein, we will report on an exploratory study of the
time-like pion electromagnetic form factor using functional methods. More precisely, we will employ a combination
of Bethe-Salpeter and Dyson-Schwinger equations (for recent reviews on this and related approaches see, {\it e.g.}, \cite{Cloet:2013jya, Eichmann:2016yit,Huber:2018ned,Sanchis-Alepuz:2017jjd}). Although such an approach is
capable of allowing a first-principle calculation (see, {\it e.g.}, the computation of the glueball spectrum reported
in ref.\ \cite{Huber:2020ngt}), for the task at hand this is yet out of reach.
To grasp all essential features of the pions' time-like form factor\footnote{There are several investigations of the
space-like pion electromagnetic form factor in the Dyson-Schwinger--Bethe-Salpeter approach, early examples
include \cite{Langfeld:1989en,Roberts:1994hh,Maris:2000sk}. A calculation of this form factor spanning the entire
domain of space-like momentum transfers is described in ref.\ \cite{Chang:2013nia}. }
one needs to describe at least
(i) the pion as bound state of quark and antiquark thereby at the same time taking
into account its special role as would-be Goldstone boson of the dynamically broken chiral symmetry of QCD,
(ii) the mixing, respectively, the interference of the $\rho$-meson, being described also as a quark-antiquark bound
state, with a virtual photon when this photon is in turn coupled to a quark-antiquark pair via the fully renormalized
quark-photon vertex, and
(iii) the dominant decay channel of the $\rho$-meson, namely, $\rho \to \pi \pi$. The study presented here is now in
two aspects exploratory. First of all, the interaction between quarks and antiquarks is modelled in such a way that
the essential features, as implied by QCD and phenomenology, are taken into account but that it is on the other hand
still manageable in such an involved calculation. Second, in several places we will make technical simplifications,
especially when the such introduced error can for good reasons assumed to be small and the reduction in the
computing time needed is substantial. We thus aim here more for an understanding of how the different features of the form
factor arise from the QCD degrees of freedom than for a quantitative agreement with the experimental data.
In the chosen model for the quark-antiquark interaction, besides a gluon-mediated interaction also pions will be
included explicitly. The reason for this is as follows: if one were able to take into account the fully renormalized
quark-gluon vertex exactly within this approach, hadronic degrees of freedom will effectively emerge and
thus be included in the interaction between quarks and antiquarks, respectively, they
will back-feed on the quarks' dynamics. Due to the pions' Goldstone boson nature,
and especially due to the implied small pion mass, the pions are the most important low-energy degrees of
freedom within the Strong Interaction, as, {\it e.g.}, also elucidated by chiral perturbation theory. And as
in order to describe the physics of decays non-valence effects need to be taken into account, it is some minimal
requirement for the investigation reported here to include pions as the most important non-valence-type interaction
mediator in the sub-GeV region.
The interaction model herein is also chosen in view of a possible generalisation to the study of baryon form factors,
and hereby especially the nucleons' time-like form factor. In this respect one can build on existing
calculations of space-like form factors from bound state amplitudes, see, {\it e.g.}, refs.\
\cite{Eichmann:2016yit,Nicmorus:2010sd,Eichmann:2011vu,Eichmann:2011pv,Sanchis-Alepuz:2013iia,Sanchis-Alepuz:2017mir}
for some recent respective work. A thorough understanding of the proton time-like form factor at very low $Q^2$ is
a very timely subject as the upcoming PANDA experiment possesses the unique possibility to measure the proton's
electromagnetic form factors in the so-called unphysical region through the process $\bar p p \to l^+l^-\pi^0$, $l=e, \mu$
\cite{Fischer:2021kcr}. At large $Q^2$ the question of the onset of the convergence scale between the space-like and
the time-like form factors arises.
However, also the pions' time-like electromagnetic form factor will be studied further by upcoming experiments, among other reasons because recently the consistency of the available data sets has been questioned \cite{Ananthanarayan:2020vum}. Based on the long-known fact that the $\tau$ radiative decay allows to extract the
pion form factor \cite{Kim:1979hx} and that a very large number of $\tau$-leptons are produced at B-meson factories
further high-precision data in sub-GeV region will become available.\footnote{In this work we will compare to the dataset of ref.\ \cite{Akhmetshin:2006bx} available at https://www.hepdata.net/record/ins728302.}
Besides earlier lattice QCD calculations of the pions' space-like electromagnetic form factor
\cite{Brommel:2006ww,Frezzotti:2008dr,Boyle:2008yd,Aoki:2009qn,Nguyen:2011ek,Brandt:2013dua,Koponen:2013boa}
recently calculations of the time-like pion form factor have become available
\cite{Meyer:2011um,Feng:2014gba,Bulava:2015qjz,Erben:2019nmx}. These results typically show a good
agreement with the experimental data. In those calculations, the extraction of the time-like form factor employs a parameterisation
based on the vector meson dominance (VMD) picture to extract from the lattice data at discrete energies the time-like
pion form factor.
The exploratory calculation presented herein is done in the isospin symmetric limit. One of the effects of isospin
breaking clearly visible in the time-like pion form factor is $\rho$-$\omega$ mixing, see, {\it e.g.}, the review
\cite{OConnell:1995nse}. To this end we will employ the VMD-based fit given in \cite{OConnell:1995nse} and modify
it such that an expected form of the pion form factor without this mixing effect is extracted.\footnote{In
ref.\ \cite{Jegerlehner:2011ti} another method has been used to remove the $\rho$-$\omega$ mixing effects from the data.}
Furthermore, we will present a simplified but numerically quite accurate VMD parameterisation of the time-like form
factor which then serves as a basis for a detailed analysis of our results. Here, the focus is more on a comparison of
our results with the form expected on the basis of the VMD parameterisation than on numerical agreement.\footnote{A precise representation of the time-like form factor requires parameterisations including excited $\rho$-mesons,
respective examples can be found in refs.\ \cite{Schael:2005am,Davier:2005xq}. Two-photon effects, on the other
hand, can very likely be safely neglected, for a corresponding study of the form factor at large
momentum transfer see \cite{Chen:2018tch}.}
This paper is organized as follows: In Sect.\ \ref{sec:PionFF} we review some facts about the electromagnetic
pion form factor,
employ the VMD-based fit given in \cite{OConnell:1995nse} to remove the $\rho$-$\omega$ mixing effects, and
provide an expected form of the pion form factor.
In Sect.\ \ref{sec:DSE_BSE} we present our approach based on Bethe-Salpeter and Dyson-Schwinger equations.
Our results are presented and analyzed in Sect.~\ref{sec:results}. In Sect.~\ref{sec:outlook} we present conclusions and
an outlook. Some technical details are deferred to two appendices.
\section {\label{sec:PionFF} The time-like electromagnetic pion form factor}
\begin{figure}[ht]
\begin{center}
\setlength{\unitlength}{0.5mm}
\begin{picture}(150,60)
\thicklines
\multiput(32.5,30)(10,0){6}{\oval(5,5)[t]}
\multiput(37.5,30)(10,0){6}{\oval(5,5)[b]}
\put(60,40){$\gamma^\star $}
\put(10,50){\vector(1,-1){10}}
\put(20,40){\line(1,-1){10}}
\put(20,45){\Large $e^-$}
\put(30,30){\vector(-1,-1){10}}
\put(20,20){\line(-1,-1){10}}
\put(20,15){\Large $e^+$}
\put(100,30){\circle{20}}
\put(95,28.5){\Large $F_\pi$}
\put(125,45){\Large $\pi^+$}
\put(125,15){\Large $\pi^-$}
\multiput(108,35)(10,10){2}{\line(1,1){8}}
\multiput(108,25)(10,-10){2}{\line(1,-1){8}}
\end{picture}
\end{center}
\caption{Electron-positron pair annihilating to a virtual photon with time-like momentum which then
decays to a pion pair. }
\label{Fig:PiFF}
\end{figure}
The pion, being a composite object, does not have a point-like interaction with the electromagnetic field, and
the related substructure, the pion being a pseudoscalar, is related to one form factor. Considering, for example,
the scattering of an electron off a pion $\pi^+$ one can describe the leptonic part of the interaction quite
precisely in lowest-order perturbation theory, {\it i.e.}, one considers the process in which the electron emits a
virtual photon, and the latter couples to the pion. Defining the form factor $F_\pi (s)$ via the relation
\begin{equation}
\langle \pi^+(p_1) | j^\mu _{e.m.} | \pi^+(p_2) \rangle = e (p_1+p_2)^\mu F_\pi (q^2) \, ,
\end{equation}
where $q^\mu = p_1^\mu - p_2^\mu$ is the virtual photon momentum,
and $e$ is the elementary electric charge. The $S$-matrix element for electron-pion scattering is then
proportional to the form factor,
\begin{eqnarray}
i {\cal M}_{e\pi \to e \pi } &=& e (p_1+p_2)^\mu F_\pi (q^2) D_{\rm photon}^{\mu \nu } (q)
\nonumber \\ && \qquad
\left( ie \bar u (k_1,s_1) \gamma _\nu u(k_2,s_2) \right) \, ,
\end{eqnarray}
where $D_{\rm photon}^{\mu \nu } (q)$ is the photon propagator, and $u(k_i,s_i)$ is the electron spinor,
see, {\it e.g.}, Sect.\ 8.4 of \cite{Aitchison:2003tq} for more details. The kinematics of this
scattering process is such that
the photon momentum is space-like, and without loss of generality one can assume the form factor
$F_\pi (s)$ to be real. Charge conservation requires that for a real photon one has
$ F_\pi (0) = 1$.
Turning to the process of electron-positron annihilation into a pion pair the corresponding $S$-matrix
element is again proportional to the form factor, see , {\it e.g.}, Sect.\ 8.5 of \cite{Aitchison:2003tq},
\begin{eqnarray}
i {\cal M}_{e^+e^-\to \pi^+ \pi^- } &=& e (p_1+p_2)^\mu F_\pi (q^2) D_{\rm photon}^{\mu \nu } (q)
\nonumber \\ && \qquad
\left(- ie \bar v (k_1,s_1) \gamma _\nu u(k_2,s_2) \right) \, ,
\label{eq:Smatrix}
\end{eqnarray}
with appropriately redefined momenta. Especially, the virtual photon momentum is now time-like,
{\it cf.} Fig.\ \ref{Fig:PiFF}, and
one measures in such an annihilation process to a pion pair the form factor for time-like momenta. Above
the two-pion production threshold, {\it i.e.}, in the physical region, the corresponding cut in the amplitude \eqref{eq:Smatrix} necessitates to treat the time-like pion form factor as a complex quantity, it fulfils the
dispersion relation
\begin{equation}
F_\pi (q^2) = 1 + \frac {q^2}\pi \int _{4m_\pi^2} ds \frac {{\cal I}m \, F_\pi (s) }{s(s-q^2-i \epsilon )} \, .
\label{eq:DispRel}
\end{equation}
Especially, it is expected that the phase of the pion form factor varies strongly in a two-pion resonance
region. Below the inelastic threshold, {\it i.e.}, for $s< 4 m_\pi^2$,
the time-like pion form factor is, via Watson's final state theorem,
related to the isovector P-wave scattering phase shift $\delta_{1,1} (s) $:
\begin{eqnarray}
{\cal I}m \, F_\pi (q^2) &=& \frac 1 {2i} \left( 1 - e^{-2i \delta_{1,1}(s)} \right) F_\pi (s+i \epsilon )
\nonumber \\
& = & \sin (\delta_{1,1}(s)) e^{-i \delta_{1,1}(s) } F_\pi (s+i \epsilon ) \, .
\end{eqnarray}
As depicted in Fig.\ \ref{Fig:PiFF}, the pion form factor contains, for time-like as well as for space-like
photon virtualities, all kind of interaction processes turning a photon into a pion pair. As strong-interaction
processes dominate the corresponding amplitude, and as gluons do not couple directly to photons, one
decisive element of the pion form factor is the amplitude describing how the photon couples to a quark,
taking hereby all possible contributing QCD processes into account. This amplitude is exactly the
full quark-photon vertex. As we will see in the following, this correlation function carries information about
the virtual photon's hadronic substructure. And as will be described in detail in the next section,
the other quantities needed then for calculating the pion
form factor are the fully renormalized quark propagator and the pion bound state amplitude,
the latter describing how an antiquark and a quark form a pionic bound state.
While our calculation is based on QCD degrees of freedom it is capable of providing an understanding
why in the resonance region a vector-meson dominance (VMD) picture provides very good results for the
time-like pion form factor.
In addition, it will elucidate to which extent a VMD picture might be applicable in other kinematic regions.
In the VMD picture, the hadronic contribution to the photon propagator is given by mixing with
electrically neutral vector mesons. Restricting to light-quark mesons, the corresponding vector meson
is the $\rho^0$, {\it i.e.}, the uncharged member of the isotriplet of vector mesons. In a would-be isospin
symmetric world this would be the only vector meson below one GeV with which a virtual photon
mixes because the isosinglet $\omega$ will not mix with the photon due to $G$-parity.
In the real world, isospin symmetry is broken, and one of the many effects of isospin breaking is
$\rho$-$\omega$ mixing, see, {\it e.g.}, the review \cite{OConnell:1995nse}.
This mixing and the resulting interference of states lead
to quite some pronounced structure in the time-like pion form factor around
$m_\omega^2$. Indeed, it is by now well understood that the sharp dip in the experimental data stems from a combination of $\rho$-$\omega$ mixing and interference effects between the decays
$\rho\rightarrow\pi\pi$ and $\omega\rightarrow \pi\pi$, the latter being isospin breaking, see, {\it e.g.},
Ref.\ \cite{OConnell:1995nse}. Moreover, it has been estimated that this combination of mixing and interference effects decreases the height of the bump of the order of up to $10\%$ (see \cite{Jegerlehner:2011ti} and references therein).
As the here presented exploratory calculation is performed in the isospin
limit\footnote{Isospin breaking and thus the effect of $\rho$-$\omega$ mixing on the pion form
factor is currently under investigation, the corresponding results will be published elsewhere.}
and thus $\rho$-$\omega$ mixing is neglected
we estimate its effect by comparing the VMD-based fit to the pion form
factor given in ref.\ \cite{OConnell:1995nse} with a plot of the same expression but the mixing
matrix element put to zero, $\Pi_{\rho\omega}=0$, see Fig.\ \ref{fig:FpiVMDplot}.
As expected the resulting curve is much smoother
than the one including the $\rho$-$\omega$ interference effect. Around the $\omega$ mass the deviation
of the two curves can be as large as almost ten percent, however, this effect is limited to a small
interval, and the two curves are practically indistinguishable outside this small interval. Therefore
we expect our calculation to reproduce all qualitative features of the curve representing the case
with $\rho$-$\omega$ mixing switched off, and to be in a reasonable quantitative agreement with it.
\begin{figure}[ht]
\includegraphics[width=0.49\textwidth]{FpiVMDplot_v3
\caption{Absolute value of the pion form factor in the time-like ($Q^2<0$) domain from the VMD-based
fit given in ref.\ \cite{OConnell:1995nse} (full line) in comparison to the experimental data \cite{Ananthanarayan:2020vum}.
The dashed line is based on the same expression
but the mixing matrix element put to zero, $\Pi_{\rho\omega}=0$. }
\label{fig:FpiVMDplot}
\end{figure}
It is instructive to analyze the momentum behaviour of a simplified version of the fit given in ref.\ \cite{OConnell:1995nse}.
In order to be in agreement with the notation in the following presentation we
introduce the photon virtuality with the convention that $Q^2<0$ corresponds to the time-like region.
In the above used fit a momentum dependent expression for the width of the $\rho$-meson is used.
It displays the two-pion cut,
\begin{equation}
\Gamma_\rho (Q^2) \propto \left( -Q^2 - 4 m_\pi^2 \right)^{3/2} \Theta ( -Q^2 - 4 m_\pi^2)\, ,
\end{equation}
which is important for a qualitatively correct analytic structure of the pion form factor. However, especially close
to the maximum of the pion form factor, the pion mass is quantitatively negligible, and for vanishing
pion mass the momentum dependent width assumes for time-like $Q^2<0$ the relatively simple form
\begin{equation}
\Gamma_\rho (Q^2) = \bar \Gamma_\rho |Q^2| / m_\rho^2\, ,
\end{equation}
where $ \bar \Gamma_\rho = \Gamma_\rho (-m_\rho^2)$. This then leads to the simplified but still quite accurate
form for the fit\footnote{In contrast to a constant width approximation the pole is in this parameterization and thus in the
employed fit not located at
$Q^2=-m_\rho^2+i m_\rho \bar \Gamma_\rho $ but at
$Q^2=(-m_\rho^2+i m_\rho \bar \Gamma_\rho)/(1+\bar \Gamma_\rho^2/m_\rho^2) $, {\it i.e.}, real and imaginary
part of the pole position are decreased by 3.7\% \ if the same values for the $\rho$'s mass and width are used.}
\begin{eqnarray}
F_\pi (Q^2+i\epsilon ) &=& 1 - \frac{g_{\rho\pi\pi}}{g_\rho} \frac {Q^2(Q^2+m_\rho^2)}{(Q^2+m_\rho^2)^2 + Q^4
\bar \Gamma^2 / m_\rho^2 } \nonumber \\
&+& i \frac{g_{\rho\pi\pi}}{g_\rho} \frac {Q^4 \bar \Gamma / m_\rho }
{(Q^2+m_\rho^2)^2 + Q^4 \bar \Gamma^2 / m_\rho^2 } \, ,
\label{eq:VMDsimplified}
\end{eqnarray}
where $\epsilon \to 0^+$ has been introduced to fix the sign of the imaginary part. Hereby, the coupling constant
$g_{\rho\pi\pi}$ is determined from $\Gamma _{\rho\pi\pi} = \bar \Gamma = 149$ MeV to be
$g_{\rho\pi\pi} \approx 6$, and $g_\rho$ is a parameter reflecting the strength of the
$\gamma$-$\rho^0$-mixing, which is then described by the effective Lagrangian
$$
{\cal L}_{\rho \gamma} = - \frac {e\, m_\rho^2}{g_\rho} \rho_\mu^0 A^\mu \, .
$$
From the partial width $\Gamma _{\rho e^+e^-} = 7$ keV one infers $g_\rho \approx 5$.
The form given in Eq.\ \eqref{eq:VMDsimplified} motivates to compare the obtained results for the real and the
imaginary part of the time-like form factor from our calculation to a rational, resp., Pad\'e fit. Phrased otherwise,
Eq.\ \eqref{eq:VMDsimplified} formalises the expectation for the time-like form factor based on
the VMD picture, and our results based on the quark-photon vertex function and the pion bound state
amplitude will be analysed by discussing how much they extend beyond to this form.
\section{\label{sec:DSE_BSE}Dyson-Schwinger and Bethe-Salpeter formalism}
We determine the necessary input for the calculation of the pion electromagnetic form factor using a combination of
Bethe-Salpeter (BSE) and Dyson-Schwinger equations (DSE). To make this presentation
self-contained we summarise in this section the most relevant aspects of the approach, for more details see,
{\it e.g.}, the recent reviews \cite{Cloet:2013jya, Eichmann:2016yit,Huber:2018ned,Sanchis-Alepuz:2017jjd}
as well as references therein. All expressions in the following are understood to be formulated in
Euclidean momentum space, {\it i.e.}, after a Wick rotation.
In the DSE/BSE formalism, the fully-dressed quark-photon vertex $\Gamma^{\mu}$, which describes the interaction bet\-ween quarks and photons in a quantum field theory, can be obtained as the solution of an inhomogeneous BSE
\begin{eqnarray}
\label{eq:inhomBSE_vector}
\bracket{\Gamma^{\mu}}_{a\alpha,b\beta}\bracket{p,Q}&=&
Z_2 \bracket{\gamma^\mu}_{ab} t_{\alpha\beta}\\
&+&\int_q K^{r\rho,s\sigma}_{a\alpha,b\beta}\bracket{Q,p,q} \, S_{r\rho,e\epsilon}\bracket{k_1}
\nonumber \\
&\times& \bracket{\Gamma^{i,\mu}}_{e\epsilon,n\nu}\bracket{Q,q}S_{n\nu,s\sigma}\bracket{k_2} \, .
\nonumber
\end{eqnarray}
Here, $Q$ is the photon momentum, $p$ is the relative momentum between quark and antiquark, $q$ is an internal relative momentum which is integrated over, the internal quark and antiquark momenta are defined as $k_1=q+Q/2$ and $k_2=q-Q/2$, respectively, such that $Q=k_1-k_2$ and $q=(k_1+k_2)/2$. Latin letters represent Dirac indices, and Greek letters represent flavour indices.
The isospin structure of the vertex is given by
$t_{\alpha\beta}= \textrm{diag}\bracket{\nicefrac{2}{3},\nicefrac{-1}{3}}$. The Dirac structure of the vertex can be expanded in a basis consisting of twelve elements \cite{Miramontes:2019mco}, and all of them are considered in our calculation.
Similarly, mesons as bound states of two quarks are described in this framework by Bethe-Salpeter amplitudes $\Gamma$ which are obtained as solutions of a homogeneous BSE,
\begin{dmath}
\label{eq:homogeneousBSE}
\bracket{\Gamma}_{a\alpha,b\beta}\bracket{p,P}= \int_q K^{r\rho,s\sigma}_{a\alpha,b\beta}\bracket{P,p,q} \times S_{r\rho,e\epsilon}\bracket{k_1}\bracket{\Gamma}_{e\epsilon,n\nu}\bracket{q,P}S_{n\nu,s\sigma}\bracket{k_2}~,
\end{dmath}
where for clarity we have here used $P$ for the total meson momentum (instead of $Q$ as above). For pions, the Dirac part of the Bethe-Salpeter amplitude $\Gamma$ can be expanded in a tensorial basis with four elements.
In the equations above, the interaction kernel $K$ describes the interaction between quark and antiquark, and $S$ is the fully-dressed quark propagator. We will discuss in detail the interaction kernels below. The quark propagator $S(p)$ is obtained as the solution of the quark DSE,
\begin{dmath}
S^{-1} = S_0^{-1} - Z_{1f} \int_q \gamma_\mu S(q) \Gamma^{qgl}_\nu(q,k) D_{\mu \nu}(k)~,
\label{eq:quarkDSE}
\end{dmath}
with $S_0^{-1}$ the renormalised bare propagator,
\begin{equation}
S_0^{-1}(p) = Z_2\left(i \Slash{p} + Z_m m\right)~,
\end{equation}
and $Z_{1f}$, $Z_2$ and $Z_m$ are renormalisation constants, $m$ is the (renormalisation-point dependent) current quark mass, $\Gamma^{qgl}$ is the full quark-gluon vertex and $D_{\mu \nu}$ is the full gluon propagator which, in the Landau gauge, is parametrised as
\begin{equation}
D_{\mu \nu}(k) = \left( \delta_{\mu \nu} - \frac{k_\mu k_\nu}{k^2}\right) \frac{Z(k^2)}{k^2}~,
\end{equation}
with $Z(p^2)$ being the gluon dressing function. For simplicity, we have suppressed the color indices.
\begin{figure*}[t!]
\centerline{%
\includegraphics[width=17cm]{FF_pion4}}
\caption{Diagrams relevant for the calculation of the pion form factor in the truncation employed herein, as determined by the gauging method. The impulse approximation implies considering the first diagram and the corresponding permutation only.}
\label{Fig:BI_diagrams}
\end{figure*}
\subsection{Interaction Kernels}
The interaction kernel $K$ in eqs.\ \eqref{eq:inhomBSE_vector} and \eqref{eq:homogeneousBSE}
encodes all possible interactions processes between a quark and an antiquark. In a diagrammatic representation, it contains a sum of infinitely many terms. In practical calculations, the expansion of the interaction kernel must be truncated to a sum of a finite number of terms, chosen such that the relevant dynamics and global symmetries are correctly implemented. Chiral symmetry and its dynamical breaking ensures that pions are massless bound states in the chiral limit as a consequence of Goldstone’s theorem. On the other hand, U(1) vector symmetry ensures charge conservation in electromagnetic processes. Chiral symmetry will be correctly implemented in the DSE/BSE formalism only if the kernel fulfills the axial-vector Ward-Takahashi identity (Ax-WTI)
\begin{dmath}
i\Sigma_{ar}\bracket{p_+}\gamma^5_{rb}t^i_{\alpha\beta}
+i\gamma^5_{ar}\Sigma_{rb}\bracket{p_-}t^i_{\alpha\beta}=\int_q K^{r\rho,s\sigma}_{a\alpha,b\beta}\bracket{Q,p,q}\left[i~t^i_{\rho\nu}\gamma^5_{rn}S_{n\nu,s\sigma}\bracket{q_-} +i~S_{r\rho,e\epsilon}\bracket{q_+}\gamma^5_{es}~t^i_{\epsilon\sigma}\right]~,\label{eq:AxWTI_BSEkernel}
\end{dmath}
with $\Sigma$ the quark self-energy and $v_\pm=v\pm Q/2$. Similarly, vector symmetry will be correctly implemented if the kernel satisfies the vector Ward-Takahashi identity (V-WTI)
\begin{dmath}
i\Sigma_{ab}\bracket{p_+}t^i_{\alpha\beta}
-i\Sigma_{ab}\bracket{p_-}t^i_{\alpha\beta}=\int_q K^{r\rho,s\sigma}_{a\alpha,b\beta}\bracket{Q,p,q}\left[i~t^i_{\rho\nu}S_{r\nu,s\sigma}\bracket{q_-}-i~S_{r\rho,s\epsilon}\bracket{q_+}~t^i_{\epsilon\sigma}\right]~.\label{eq:VWTI_BSEkernel}
\end{dmath}
In DSE/BSE studies the most widely used truncation is the so-called rainbow-ladder (RL) truncation, whereby the BSE kernel consists of a vector-vector gluon exchange, namely (omitting again color indices)
\begin{dmath}
K^{r\rho,s\sigma}_{a\alpha,b\beta}\bracket{Q,p,q}=\alpha\bracket{k^2}\gamma^\mu_{ar}\gamma^\nu_{sb}D^{\mu\nu}\bracket{k}\delta^{\alpha\rho}\delta^{\sigma\beta}~,\label{eq:RLkernel}
\end{dmath}
with $k=p-q$ the gluon momentum. In order to preseve the Ax-WTI and V-WTI the kernel \eqref{eq:RLkernel} is used in combination with a truncated quark DSE, defined by the replacement
\begin{equation}
Z_{1f} \gamma_{\mu} Z(k^2)\Gamma_{\nu}^{\text{qgl}}(q,p) \rightarrow Z_2^2 \gamma_{\mu} 4\pi \alpha(k^2) \gamma_{\nu}
\end{equation}
such that $\alpha(k^2)$ provides an effective coupling that describes the strength of the quark-antiquark interaction.
To parametrise this effective interaction we use the Maris-Tandy model \cite{Maris:1997tm, Maris:1999nt}
\begin{dmath}\label{eq:MTmodel}
\alpha(q^2)=
\pi\eta^7\left(\frac{q^2}{\Lambda^2}\right)^2
e^{-\eta^2\frac{q^2}{\Lambda^2}}+\frac{2\pi\gamma_m
\big(1-e^{-q^2/\Lambda_{t}^2}\big)}{\textnormal{ln}[e^2-1+(1+q^2/\Lambda_{QCD}
^2)^2]}~,
\end{dmath}
where the second term on the right-hand side reproduces the one-loop QCD behavior of the quark propagator in the ultraviolet, and the Gaussian term provides enough interaction strength for dynamical chiral symmetry breaking to take place. The model parameters $\Lambda$ and $\eta$ are determined as explained in the next section. The scale $\Lambda_t=1$~GeV is introduced for technical reasons and has no impact on the results. For the anomalous dimension we use $\gamma_m=12/(11N_C-2N_f)=12/25$ with $N_f=4$ flavours and $N_c=3$ colours. For the QCD scale we use $\Lambda_{QCD}=0.234$ GeV.
In the RL truncation, bound states cannot develop a decay width, correspondingly their masses are real numbers. Note that bound states, determined as solutions of BSEs, appear as poles in Green's functions with the corresponding quantum numbers. In the RL approximation these poles occur for real
(and in the convention employed here, negative) momentum-squared values in certain kinematic configurations. In particular, the photon being described by a vector field, electrically neutral vector mesons appear as poles of the quark-photon vertex. In the RL approximation these poles are located at negative and real values of $Q^2$ (for which $M^2=-Q^2$ with $M$ being the mass of the vector meson). Such poles in the quark-photon vertex also manifest as poles in the calculation of time-like form factors, in contradiction with phenomenology. Generally speaking, any physical phenomenon that is triggered by the presence of virtual intermediate particles, as, {\it e.g.}, decays, will be absent from any calculation using the RL truncation only.
It is possible to improve the RL truncation in this respect by re-introducing\footnote{Note
that effects of intermediate virtual states like the decay of hadrons are in principle present in the full quark-gluon
vertex which, however, is drastically simplified in the RL truncation.}
the presence of intermediate particles explicitly. The simplest implementation of such an idea was introduced in \cite{Fischer:2007ze,Fischer:2008sp}\footnote{See, however, ref.\ \cite{Alkofer:1993gu} for considering pion loop
contributions to the electromagnetic pion radius in the DSE/BSE approach.
An alternative way is to introduce two-pion states via an explicit two-pion component in the bound state amplitude,
see ref.\ \cite{Santowsky:2020pwd} and references therein.}, where,
based on the role of the pion as lightest
hadron, explicit pion-quark interactions were introduced in the truncated quark DSE and in the BSE kernel $K$,
with the pion-quark interaction vertex given by the pion Bethe-Salpeter amplitude
$\Gamma$, calculated consistently via a truncated BSE.
The corresponding additional BSE kernels are given in Appendix \ref{kernels} and shown in Fig.~\ref{fig:kernels}, and the technical difficulties arising for time-like momenta when those kernels are used in BSEs have been described in detail in \cite{Miramontes:2019mco}. This type of kernels enable the possibility of intermediate virtual decays in the BSE interaction kernel to occur. As a consequence, certain BSE solutions signal a finite decay width and thus represent (a) hadron resonance(s). {\it E.g.}, the description of the $\rho$-meson as a finite-width resonance is then mostly due to the intermediate process $\rho \rightarrow \pi\pi$ (see \cite{Williams:2018adr} for a treatment in the here discussed approach as well as \cite{Jarecke:2002xd,Mader:2011zf} and references therein for respective calculations based on DSEs and
BSEs),
the partial decay width to the latter process representing more than 99\% \ of the total $\rho$ decay width.
Additionally, and thereby completing the physical effects of intermediate virtual $\pi\pi$-states, in this truncation the quark-photon vertex develops a multi-particle branch cut along the negative real $Q^2$ axis, starting as expected at the two-pion production threshold \cite{Miramontes:2019mco}.
Clearly, including these two effects of the intermediate virtual $\pi\pi$-states is especially important when it comes to the calculation of the pion form factor for time-like momenta in the sub-GeV kinematic region.
We wish to stress here that, for computational feasibility, for the pion vertices in \eqref{eq:BSEkernel_tchannel}--\eqref{eq:BSEkernel_uchannel} we used the leading $\gamma_5$ component of the pion Bethe-Salpeter amplitude in the chiral limit, given by $B/f_\pi$ with $B$ one of the quark's dressing functions, see Eq.~\eqref{eq:pion_chiral}. On the other hand, the pion amplitudes used in the form factor calculation of Eq.~\eqref{eq:Current} are considered in full, including their leading and sub-leading contributions. In that regard, our calculations contain \textit{two types of treatments of pions}.
Even though the ``pionic'' kernels \eqref{eq:BSEkernel_tchannel}--\eqref{eq:BSEkernel_uchannel} are phenomenologically justified and, as we will see in the next section, constitute a first step in the correct direction, it must be noted that they have not been (yet) rigorously derived from QCD. Lacking a solid quantum-field theoretical basis the use of these kernels comes with some shortcomings, especially it implies that the Ax-WTI and V-WTI are not fulfilled simultaneously. Indeed, one can choose to preserve either the Ax-WTI (and hence chiral symmetry) or the V-WTI (and hence charge conservation), but not both \cite{Fischer:2007ze,Miramontes:2019mco}. It turns out, however, that the respective violation of either of them induce typically only small errors in physical observables, as we will demonstrate for some quantities in the next section.
\subsection{Form factor calculation}
Meson form factors are extracted from a current $J^{\mu}$ encoding the coupling of a meson to an external electromagnetic current. In the BSE framework, the current is calculated by means of the coupling of an external photon to each of the constituents of the bound state, as specified by a procedure known as \textit{gauging} and developed in \cite{Haberzettl:1997jg,Kvinikhidze:1998xn,Kvinikhidze:1999xp,Oettel:1999gc,Oettel:2000jj}. The conserved current $J^{\mu}$ that describes the coupling of a single photon with a quark-antiquark, a three-quark or other multi-quark system is given by,
\begin{equation}
J^{\mu} = \bar{\Psi}_f G_0 (\mathbf{\Gamma}^{\mu} - K^{\mu}) G_0 \Psi_i ~,
\label{eq:Current}
\end{equation}
with $\Psi_{i,f}$ the incoming and outgoing Bethe-Salpeter amplitudes of the meson, the baryon or some other multi-quark state, and $G_0$ represents the appropriate product of dressed quark propagators. This equation is shown diagrammatically for the quark-antiquark--meson case in Fig.~\ref{Fig:BI_diagrams}. The term $\Gamma^{\mu}$ represents the impulse approximation diagrams where the photon couples to the valence quarks only
\begin{equation}
\mathbf{\Gamma}^{\mu} = \left(S^{-1} \otimes S^{-1} \right)^{\mu} = \Gamma^{\mu}\otimes S^{-1} + S^{-1}\otimes \Gamma^{\mu}~,
\end{equation}
with $\Gamma^{\mu}$ the quark-photon vertex. The term $K^{\mu}$ describes the interaction of the photon with the
Bethe-Salpeter kernel, which in our truncation includes the coupling of the photon to the quark-pion vertex and to the
propagating pions
(second and third diagram in Fig.\ \ref{Fig:BI_diagrams}).
Including both terms in \eqref{eq:Current} is necessary in order to implement current conservation precisely.
Note that the s- and u-channel pion decay terms (most right diagram in Fig.\ \ref{fig:kernels}) do not contribute to
the term $K^{\mu}$, also not via seagulls, the reason being that trying to include them would leave a $\pi$ $\pi$ $\pi$
amplitude on one side of the respective full diagram.
However, the new vertices appearing in the term $K^{\mu}$,
the coupling of the photon to the quark-pion vertex and to the propagating pions,
represent an enormous computational challenge.
Given the exploratory purpose of the present calculation, we thus decided to omit the coupling of the photon to the quark-pion vertex and to the propagating pions, and to consider the impulse diagram only (first diagram in Fig.~\ref{Fig:BI_diagrams}). As we will show in the next section, the thereby implied violation of charge conservation is at the level of approximately one percent.
\section{\label{sec:results} Results}
Following the formalism sketched above, we have calculated the pion electromagnetic form factor in the time-like $Q^2<0$ domain. For comparison with previous calculations we will also show results in the space-like $Q^2>0$ domain.
In our discussion in the preceding section, it remained to be explained how the parameters $\eta$ and $\Lambda$ of our interaction model \eqref{eq:MTmodel} as well as the value of the current quark mass are fixed. It is customary in studies using the RL truncation to adjust those parameters such that the pion decay constant agrees with the experimental value\footnote{Note that, for this observable, the result is quite independent of the value of $\eta$ around $\eta=1.8$ and thus, effectively, only the parameter $\Lambda$ has to be adjusted.}.
In ref.\ \cite{Miramontes:2019mco} the quark-photon vertex has been calculated for the first time with the above discussed interaction kernels taken into account and also used herein. The parameters, including the isospin symmetric light quark current mass $m_q$, were adjusted such that the pion mass and decay constant as well as the $\rho$-meson mass have been correctly reproduced in the employed approximation. Here, only the gluon- and pion-exchange kernels need to be used for fixing the parameters because the pion decay kernels \eqref{eq:BSEkernel_schannel} and \eqref{eq:BSEkernel_uchannel} do not contribute to the pion BSE.
For the present exploratory calculation, we choose to adjust the parameters in a slightly different and simpler manner, especially as we aim at a qualitative understanding of the physical mechanisms involved in determining the shape of the pion form factor, and not so much at achieving an accurate quantitative agreement with experiment.
First, we set initially $\eta=1.5$. Second, although we assume (confirmed by our calculation) that the pion-decay kernels will not only move the $\rho$-meson pole into the complex plane but also shift down its real value, we nevertheless adjust the parameter $\Lambda$ such that we obtain a $\rho$-meson mass close to the phenomenological value already in the calculations with gluon- and pion-exchange kernels only. Third, we require to reproduce a quite accurate value for the pion decay constant.
Finally, we adjust $m_q$ to obtain the correct value for the pion mass as well. In this way we chose the model parameters to be $\eta=1.5$, $\Lambda=0.78$~GeV and $m_q=6.8$~MeV at a renormalisation scale $\mu=19$~GeV. As a rudimentary test of model dependence we additionally perform the calculations for $\eta=1.6$ as well (keeping $\Lambda$ and $m_q$ unchanged).
The results for the pion mass and decay constant as well as for the $\rho$-meson and $\omega$-meson masses, without the pion decay kernels being taken into account, are shown in Tab.~\ref{tab:masses}.
\begin{table}[t]\caption{\label{tab:masses}
The pion mass $m_\pi$, the pion decay constant $f_\pi$, the $\rho$-meson and $\omega$-meson masses $m_\rho$ and $m_\omega$ for the two different parameterizations of the model used herein
and for the case with rainbow-ladder and pion-exchange kernels but without decay kernels are shown.
The light quark mass has been set to $m_q=0.0068~$GeV.
The rightmost column displays the extracted $\rho$-meson pole position defined as
$M_{pole}^2=M_\rho^2-iM_\rho \Gamma_\rho $, as discussed in the text.
All values for dimensionful quantities are given in GeV.}
\begin{ruledtabular}
\begin{tabular}{l|cccc||cc}
$\Lambda=0.78$&
$m_\pi$&
$f_\pi$&
$m_\rho$&
$m_\omega$&
$M_\rho$&
$\Gamma_\rho$\\
\colrule
$\eta=1.5$ & 0.139 & 0.138 & 0.768 & 0.778 & 0.750 & 0.100\\
$\eta=1.6$ & 0.126 & 0.138 & 0.774 & 0.784 & 0.759 & 0.105\\
\end{tabular}
\end{ruledtabular}
\end{table}
As mentioned in the previous section, the full QCD quark-photon vertex possesses poles reflecting
the masses and widths of the electrically neutral vector meson resonances.
Correspondingly, and as discussed in detail in refs.\ \cite{Miramontes:2019mco,Williams:2018adr}, a solution for the quark-photon vertex in the DSE/BSE framework allows to extract the $\rho$-meson mass and width via the position of the poles of the vertex dressing functions. For the RL truncation
as well as for the RL plus pion exchange approximation this pole is located on the real negative $Q^2$ axis indicating that the $\rho$ mesons were stable for those truncations. Including the $s$- and $u$-channel decay kernels \eqref{eq:BSEkernel_schannel} and \eqref{eq:BSEkernel_uchannel}, the pole of the dressing functions moves into the complex plane and can be extracted from the data on the real axis via a Pad\'e fit. Parametrising the pole position as
$M_{pole}^2=M^2_\rho -iM_\rho \Gamma_\rho $, we extract the corresponding results for the
$\rho$-meson mass and width, $ M_\rho$ and $\Gamma_\rho $ for this truncation (see Tab.~\ref{tab:masses}) in reasonable agreement with the experimental values. Hereby, it has to be noted that underestimating the $\rho$-meson width does not
come unexpected because taking into account only the leading $\gamma_5$ component of the pion Bethe-Salpeter amplitude
in the kernels \eqref{eq:BSEkernel_tchannel}--\eqref{eq:BSEkernel_uchannel} misses some strengths therein. Whether
considering in addition also the sub-leading pion amplitudes will provide a much better result for the $\rho$-meson width
can only be answered by performing the corresponding calculation. This computation is then, however, an order of magnitude
more expensive than the present exploratory calculation.
It is interesting to note that the pion exchange kernels lift the degeneracy in between the isovector
$\rho$- and the isosinglet $\omega$-meson present at the level of RL calculations the interaction kernels of which are flavour blind and thus flavour U(2) (resp., flavour U$(N_f)$) symmetric. The pion
exchange kernels lead to a splitting such that $m_\omega - m_\rho = 10$ MeV which compares favorably with the experimental splitting of 7 - 8 MeV.
We turn now to the calculation of the pion form factor.
As already indicated in the previous section, there are, besides restricting to the leading pion amplitude in the
kernels\eqref{eq:BSEkernel_tchannel}--\eqref{eq:BSEkernel_uchannel},
two major approximations that we must perform in order to keep the calculation technically manageable. First, using the
decay kernels as described in this work entails that one must choose whether the axial-vector or the vector WTI are
preserved while the other one is violated. Following \cite{Miramontes:2019mco} we choose to preserve the vector identity.
A violation of the axial-vector WTI is manifested, among others, in the pion not being massless in the chiral limit, and
therefore the value of the current mass for which the pion becomes massless allows for a quantification of the violation
of the axial-vector WTI. In Fig.~\ref{fig:mpiFif} we therefore show the evolution of the pion mass with varying $m_q$ in the
employed truncation. First, the relation is linear as expected from the Gell-Mann--Oakes--Renner relation which
is a direct consequence of the dynamical breaking of chiral symmetry. Second, the pion does not
become massless in the chiral limit but for a value of the current mass
$m_q^{(0)}$($\mu=19$~GeV) = 3 MeV. On the one hand, this explains the relatively large value of
$m_q$($\mu=19$~GeV) = 6.8 MeV we needed to obtain the correct pion mass: The related
explicitly chiral-symmetry-breaking term is $m_q - m_q^{(0)}$ = 3.8 MeV, and thus much closer to what is expected from the known parameters of QCD. Second, as the masses of the vector mesons
depend linearly on the current mass the induced error on $m_\rho$ and $m_\omega$ is of the order
of $m_q^{(0)}$ = 3 MeV, and thus it is as small or even smaller than other uncertainties in our calculation of the vector meson masses.
\begin{figure}[ht]
\includegraphics[width=0.49\textwidth]{mpi_plot
\caption{The pion mass squared $m_\pi^2$ vs.\ the current mass $m_q$($\mu=19$~GeV) for the employed truncation.}
\label{fig:mpiFif}
\end{figure}
Second, in the calculation of the form factor we use the impulse approximation which, in this context, implies neglecting the second and third diagrams in Fig.~\ref{Fig:BI_diagrams}. The consequence of discarding diagrams is the violation of charge conservation or, equivalently, a deviation from $F_\pi(0)=1$. As can be seen in the inset in Fig.~\ref{fig:FF_spacelike}, this effect is of the order of $\sim 1\%$ only.
\begin{figure}[hb]
\includegraphics[width=0.49\textwidth]{FFspacelike_v3
\caption{Pion form factor in the spacelike $Q^2>0$ domain for the model parameters $\eta=1.5$ and $\eta=1.6$ as described in the text and compared to experimental data.}
\label{fig:FF_spacelike}
\end{figure}
\begin{figure*}[ht]
\centerline{%
\includegraphics[width=0.8\textwidth]{FF_timelike_v2}}
\caption{Absolute value of the pion form factor in the time-like $Q^2<0$ domain for the model parameters $\eta=1.5$ and $\eta=1.6$ as described in the text and compared to experimental data.}
\label{fig:FF_abs_timelike}
\end{figure*}
As can be also seen from Fig.~\ref{fig:FF_spacelike}, the results for the pion form factor in the space-like $Q^2>0$ regime are practically independent from the value of the $\eta$ parameter of the model.
Even more remarkable is the fact that our calculation shows a very good agreement with experimental data in the space-like domain even though, as evident from the above discussion, we aimed at including all physical effects which are important in the time-like regime. An interpretation of this result in view of the
dispersion relation \eqref{eq:DispRel} provides an indication that the imaginary part in the time-like region is
precisely enough reproduced to provide very good results for the space-like form factor.
\begin{figure*}[ht]
\centerline{%
\includegraphics[width=0.8\textwidth]{phase_v3}}
\caption{Phase of the pion form factor in the time-like $Q^2<0$ domain for the model parameters $\eta=1.5$ and $\eta=1.6$ as described in the text and compared to experimental data on pion-pion phase shift.}
\label{fig:FF_phase_timelike}
\end{figure*}
We show our results for the pion form factor for time-like ($Q^2<0$) virtualities in Figs.~\ref{fig:FF_abs_timelike} and \ref{fig:FF_phase_timelike}. As discussed in the previous section, as a consequence of the decay kernels in our truncation, the pion form factor develops a branch cut along the real negative axis starting from the two-pion threshold $Q^2=-4m_\pi^2$, induced by the corresponding cut in the quark-photon vertex \cite{Miramontes:2019mco}. Hence, in that region the (complex) form factor is defined from its analytic continuation as $F(Q^2+i\epsilon)$. In the numerical calculations we have typically chosen
$\epsilon=0.0001$~GeV$^2$ after verifying that this value is small enough to not disturb the presented results. In Fig.~\ref{fig:FF_abs_timelike} we present the absolute value for the model parameter $\eta=1.5$ and $\eta=1.6$, with the remaining model parameters kept constant, as discussed above. As a manifestation of the fact that the $\rho$-meson pole in the quark-photon vertex moves into the complex plane when the decay kernels are included in the calculation, the pion form factor develops a bump on the real and negative $Q^2$ axis with an approximately correct height and width. Therefore our calculation overcomes a major
deficiency of the RL truncation, without or even with the pion exchange term, for which the form factor diverges instead at the $Q^2$-value corresponding to the $\rho$-meson mass in those truncations,
see e.g.~\cite{Maris:1999bh,Krassnigg:2004if,Eichmann:2019tjk}.
This constitutes already one main result of the
here presented investigation.
We note, however, that, contrary to the results for space-like regime, the position and height of the bump of the form factor depends strongly on the value of the $\eta$ parameter shape of the form factor. Of course, this reflects the different positions of the $\rho$-meson pole, {\it cf.} Tab.~\ref{tab:masses}. Nevertheless, there are features that appear to be independent of $\eta$, most prominently that the height of the bump is underestimated. As expected from the discussion in Sec.~\ref{sec:PionFF} the form factor behaves smoothly after the bump, in contrast to the sharp dip in the experimental data. Even though an unambiguous analysis of the origin of such discrepancies can only result from the inclusion of all relevant physical mechanisms in our calculations, it is evident from the analysis performed in Sec.~\ref{sec:PionFF}
that one of the main missing elements is the $\rho$-$\omega$ mixing due to isospin breaking,
which is completely absent in the present study.
Particularly sensitive to the deficiencies of our truncation is the phase of the form factor, shown in Fig.~\ref{fig:FF_phase_timelike}. Even though our data shows the expected behaviour near the resonance value, it severely underestimates the experimental data, particularly in the elastic region. This is a manifestation of the absence of some hadronic effects in our approximation scheme, which only includes those stemming from the resonance complex pole in the quark-photon vertex and the $\rho\rightarrow\pi\pi$ induced branch cut. In addition to the isospin breaking effects discussed above, and which would be more relevant in the region above the resonance, the impulse approximation used in our calculation of the form factor entails that effects coming from the coupling of the photon to the intermediate hadrons, via its coupling to the exchanged pion or to the quark-pion vertex (see Fig.~\ref{Fig:BI_diagrams}), are missing. It has been shown \cite{Cotanch:2002vj} that considering only impulse-like diagrams leads to a very small pion-pion scattering amplitude in the elastic region in the isospin $I=1$ channel. This, due to unitarity, implies a very small value of the imaginary part of the pion form factor in the elastic region (which is, in fact, what we observe) which translates into a very small phase, as seen in Fig.~\ref{fig:FF_phase_timelike} (and as could be inferred from Watson's theorem).
Last but not least, we are comparing Pad\'e fits, resp.\ rational fits of the order (3,3), for the real part and
for the imaginary part of the form factor to the expression \eqref{eq:VMDsimplified}. Trying first,
\begin{eqnarray}
Re \, F_\pi (Q^2) - F_\pi (0) & \approx &
- \frac {a_0 + a_1 Q^2 + a_2 (Q^2)^2 + a_3 (Q^2)^3}{b_0 + b_1 Q^2 + b_2 (Q^2)^2 + b_3 (Q^2)^3} \nonumber \\
Im \, F_\pi (Q^2) & \approx & \frac {c_0 + c_1 Q^2 + c_2 (Q^2)^2 + c_3 (Q^2)^3}
{d_0 + d_1 Q^2 + d_2 (Q^2)^2 + d_3 (Q^2)^3} \, , \nonumber \\
\end{eqnarray}
we obtain tiny values for the coefficients $a_0$, $a_3$, $b_3$, $c_0$, $c_3$ and $d_3$. Note that this {\em confirms the
structure expected from the VMD form} \eqref{eq:VMDsimplified}. We repeated the fits for
\begin{eqnarray}
Re \, F_\pi (Q^2) - F_\pi (0) & \approx &
- \frac {a_1 Q^2 + a_2 (Q^2)^2 }{b_0 + b_1 Q^2 + b_2 (Q^2)^2} \nonumber \\
Im \, F_\pi (Q^2) & \approx & \frac { c_1 Q^2 + c_2 (Q^2)^2 }
{d_0 + d_1 Q^2 + d_2 (Q^2)^2 } \, .
\end{eqnarray}
The coefficients resulting from these fits as well as the ones resulting from expression \eqref{eq:VMDsimplified}
are given in table~\ref{tab:Coeffs}. From this we conclude that the expression based on the VMD is an
astonishingly good representation of our results. Therefore, our investigation makes it plausible that the VMD picture
can be derived from QCD. At least, the results of the here presented microscopic approach give a strong hint into this
direction.
\begin{table}[h]\caption{\label{tab:Coeffs}
The coefficients of the rational fits to the pion form factor as discussed in the text.
All values for dimensionful quantities are given in GeV.}
\begin{ruledtabular}
\begin{tabular}{l|ll||l}
& $\eta$ =1.5 & $\eta$ =1.6 & Eq.\ \eqref{eq:VMDsimplified} \\
\hline \hline
$a_1$ & 0.5587 & 0.4149 & 0.72\\
$a_2$ & 0.8828 & 0.6827 & 1.2 \\
\hline
$b_0$ & 0.3600 & 0.3600 & 0.36\\
$b_1$ & 1.2307 & 1.2517 & 1.2\\
$b_2$ & 1.0722 & 1.1000 & 1.037\\
\hline \hline
$c_1$ & 0.0591 & 0.0997 & 0\\
$c_2$ & 0.1295 & 0.2383 & 0.2308\\
\hline
$d_0$ & 0.3600 & 0.3600 & 0.36\\
$d_1$ & 1.1924 & 1.2464 & 1.2\\
$d_2$ & 0.9973 & 1.0916 & 1.0037\\
\end{tabular}
\end{ruledtabular}
\end{table}
\section{\label{sec:outlook} Conclusions and outlook}
In this work we have presented an exploratory study of the pion form factor in the DSE/BSE approach. Our focus has been
to explore how the interplay between hadron structure (as described by form factors) and the hadron spectrum (as described
by resonance masses and widths) can be realised in the microscopic approach presented herein. In particular, we focused
on the effect of intermediate pions in the BSE interaction kernel, the inclusion of which is sufficient to describe the $\rho$ meson as a resonance \cite{Williams:2018adr}. As elucidated by the detailed analysis in the last section our calculation represents
a verification of the vector meson dominance picture and provides an explanation how at the quark level vector meson dominance becomes effective. A more complete calculation than the one presented here might then actually provide a
derivation of vector meson dominance from QCD.
Despite the fairly drastic approximations used in this preliminary study, the agreement with experiment is remarkable. On the space-like side our calculations agree with experimental data at the quantitative level. For time-like momentum transfers
the agreement is mostly qualitative and consistent with the fact that in our approximation scheme time-like physics is
dominated by the lowest-lying $\rho$-meson resonance only. The absolute value of the calculated form factor features a
bump at approximately the correct $Q^2$ region as caused by a resonance pole. Our result lacks, however, other features
such as those caused by the isospin-breaking $\rho$-$\omega$ mixing and interference. The phase of the form factor also shows deficiencies caused by the employed impulse approximation. However, the overall qualitatively correct behaviour is, nevertheless, very encouraging for future studies on time-like phenomenology with BSE methods as it shows that the
necessary computational techniques are getting more and more under control, and that within a functional-method-based
bound-state approach to QCD direct calculations in the time-like regime are becoming feasible.
Among the different physical mechanisms absent in our calculation, the most relevant one appears to be isospin breaking
by the light quarks' masses and electric charges, and the different phenomena associated with it.
The presented exploratory calculation paved the way to include in
a bound-state approach formulated in QCD degrees of freedom the effects of isospin violation, and hereby most
prominently $\rho$-$\omega$ mixing, on the time-like pion electromagnetic form factor. Thus including isospin violation
in a BSE approach is the topic of ongoing work.
A further related topic is the study of the form factor for the coupling of a photon to three pions. On the one hand, this
process is of special theoretical interest because the related form factor is at the soft point completely determined
by the Abelian chiral anomaly. On the other hand, data of the COMPASS experiment are currently analysed \cite{Dominik},
and therefore experimental data for this form factor in the time-like region will become available. Combing previous studies
in the DSE/BSE approach for the space-like $\gamma \pi \pi \pi $ form factor
\cite{Alkofer:1995jx,Bistrovic:1999dy,Cotanch:2003xv,Benic:2011rk,Eichmann:2011ec}
with the techniques of the here presented calculation will thus enable a respective investigation of this form factor.
Another aspect we want to investigate is how to realise the idea of decay kernels like \eqref{eq:BSEkernel_schannel} and \eqref{eq:BSEkernel_uchannel} in a baryon bound state equation. This is a necessary step in order to tackle
time-like nucleon form factors in the DSE/BSE approach, which is one of our major goals due to the increased effort
and interest from the experimental side in highly precise measurements over a wide kinematical domain of the
nucleon form factors.
\section*{Acknowledgements}
We thank Gernot Eichmann and Christian Fischer for a critical reading of the manuscript and helpful discussions.
This work was partially supported by the the Austrian Science Fund (FWF) under project number P29216-N36.\\
A.S.\ Miramontes acknowledges CONACyT for financial support.
The numerical computations have been performed at the high-performance compute cluster of the University of Graz.
|
\section{Experimental Implementation}\label{sec:appendix-experiments}
We discuss the implementation details of the experiments in this section. Because the setups of the domains are analogous, we discuss the delivery domain first in every section and then briefly relate how the same formulation applies to the reacher and pick-and-place domains as well. In this section, we use the simpler formulation of the main paper and not the more general formulation discussed in Appendix~\ref{sec:appendix-lof}.
\subsection{Propositions}
The delivery domain has 7 propositions plus 4 composite propositions. The subgoal propositions are $\mathcal{P}_G = \{a, b, c, h\}$. Each of these propositions is associated with a single state in the environment (see Fig.~\ref{fig:appendix-discrete-domain}). The safety propositions are $\mathcal{P}_S = \{o, e\}$. $o$ is the obstacle proposition. It is associated with many states -- the black squares in Fig.~\ref{fig:appendix-discrete-domain}. $e$ is the empty proposition, associated with all of the white squares in the domain. This is the default proposition for when there are no other active propositions. The event proposition is $\mathcal{P}_E = \{can\}$. $can$ is the ``cancelled'' proposition, representing when one of the subgoals has been cancelled.
To simplify the FSAs and the implementation, we make an assumption that multiple propositions cannot be true at the same state. However, it is reasonable for $can$ to be true at the subgoals, and therefore we introduce 4 composite propositions, $ca = a \land can$, $cb = b \land can$, $cc = c \land can$, $ch = h \land can$. These can be counted as event propositions without affecting the operation of the algorithm.
The reacher domain has analogous propositions. The subgoals are $r, g, b, y$ and correspond to $a, b, c, h$. The environment does not contain obstacles $o$ but does have safety proposition $e$, and it also has the event proposition $can$ and the composite propositions $cr, cg, cb, cy$ for when $can$ is true at the same time that a subgoal proposition is true. Another difference is that the subgoal propositions are associated with a small spherical region instead of a single state as in the delivery domain; this is a necessity for continuous domains and unfortunately breaks one of our conditions for optimality because the subgoals are now associated with multiple states instead of a single state. However, the LOF meta-policy will still converge to a hierarchically optimal policy.
The pick-and-place domain has subgoals $r, g, b, y$ like the reacher domain, and event proposition $can$. Like the reacher domain, the pick-and-place domain's subgoals become true in a region around the goal state, breaking one of the necessary conditions for optimality. However, the LOF meta-policy still converges to a hierarchically optimal policy.
\subsection{Reward Functions}
Next, we define the reward functions of the physical environment $R_E$, safety propositions $R_S$, and FSA states $R_F$. We realize that often in reinforcement learning, the algorithm designer has no control over the reward functions of the environment. However, in our case, there are no publicly available environments such as OpenAI Gym or the DeepMind Control Suite that we know of that have a high-level FSA built-in. Therefore, anyone implementing our algorithm will likely have to implement their own high-level FSA and define the rewards associated with it.
For the delivery domain, the low-level environment reward function $R_E : \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$ is defined to be $-1 \; \forall s, a$. In other words, it is a time/distance cost.
We assign costs to the safety propositions by defining the reward function $R_S : \mathcal{P}_S \rightarrow \mathbb{R}$. All of the costs are $0$ except for the obstacle cost, $R_S(o) = -1000$. Therefore, there is a very high penalty for encountering an obstacle.
We define the environment reward function $R_\mathcal{E} : \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$ to be $R_\mathcal{E}(s, a) = R_E(s, a) + R_S(T_P(s))$. In other words, it is the sum of $R_E$ and $R_S$. This reward function meets Condition~\ref{lemma:subgoals} for the optimal option policies to always converge to their subgoals.
Lastly, we define $R_F : \mathcal{F} \rightarrow \mathbb{R}$ to be $R_F(f) = 1 \; \forall f$. Therefore the SMDP cost $R_SMDP(f, s, o) = R_o(s)$ and meets Condition~\ref{lemma:satisfaction} so that the LOF meta-policy converges to the optimal policy.
The reacher environment has analogous reward functions. The safety reward function $R_S(p) = 0 \; \forall p \in \mathcal{P}_S$ because there is no obstacle proposition. Also, the physical environment reward function differs during option training and meta-policy learning. For meta-policy learning, the reward function is $R_E(s, a) = -a^\top a - 0.1$ -- a time cost and an actuation cost. During option training, we speed learning by adding the distance to the goal state as a cost, instead of a time cost: $R_E(s, a) = -a^\top a - \vert\vert s - s_g \vert \vert^2$. Although the reward functions and value functions are different, the costs are analogous and lead to good performance as seen in the results. Note that this method can't be used for Reward Machines, because it trains sub-policies for FSA states, and the subgoals for FSA states are not known ahead of time, so distance to subgoal cannot be calculated.
The pick-and-place domain has reward functions analogous to the reacher domain's.
\subsection{Algorithm for \texttt{LOF-QL}}\label{sec:appendix-lof-ql}
The \texttt{LOF-QL} baseline uses Q-learning to learn the meta-policy instead of value iteration. We therefore use ``Logical Q-Learning'' equations in place of the Logical Value Iteration equations described in Eqs.~\ref{eq:q-update} and~\ref{eq:v-update} in the main text. The algorithm is described in Alg.~\ref{alg:lof-ql}. A benefit of using Q-learning instead of value iteration is that the transition function $T_F$ of the FSA $\mathcal{T}$ does not have to be explicitly known, as the algorithm samples from the transitions rather than using $T_F$ explicitly in the formula. However, as described in the main text, this comes at the expense of reduced composability, as \texttt{LOF-QL} takes around $5x$ more iterations to converge to a new meta-policy than \texttt{LOF-VI} does. Let $Q_0(f, s, o)$ be initialized to be all $0$s. The Q update formulas are given in Alg.~\ref{alg:lof-ql} lines~\ref{eq:lof-ql-update1} and~\ref{eq:lof-ql-update2}.
\begin{algorithm}[!ht]
\caption{LOF with $\epsilon$-greedy Q-learning}\label{alg:lof-ql}
\begin{algorithmic}[1]
\STATE \textbf{Given:} \par
Propositions $\mathcal{P}$ partitioned into subgoals $\mathcal{P}_G$, safety propositions $\mathcal{P}_S$, and event propositions $\mathcal{P}_E$ \par
Environment MDP $\mathcal{E} = (\mathcal{S}, \mathcal{A}, T_E, R_\mathcal{E}, \gamma)$ \par
Logical options $\mathcal{O}$ with reward models $R_o(s)$ and transition models $T_o(s' \vert s)$ \par
Liveness property $\mathcal{T} = (\mathcal{F}, \mathcal{P}_G \cup \mathcal{P}_E, T_F, R_F, f_0, f_g)$ ($T_F$ does not have to be explicitly known if it can be sampled from a simulator) \par
Learning rate $\alpha$, exploration probability $\epsilon$ \par
Number of training episodes $n$, episode length $m$
\STATE \textbf{To learn:}
\STATE Meta-policy $\mu(f, s, o)$ along with $Q(f, s, o)$ and $V(f, s)$
\STATE \textbf{Find a meta-policy $\mu$ over the options:}
\STATE Initialize $Q : \mathcal{F} \times \mathcal{S} \times \mathcal{O} \rightarrow \mathbb{R}$ and $V : \mathcal{F} \times \mathcal{S} \rightarrow \mathbb{R}$ to $0$
\FOR{$k \in [1, \dots, n] $:}
\STATE Initialize FSA state $f \leftarrow 0$, $s$ a random initial state from $\mathcal{E}$
\STATE Draw $\bar{p}_e \sim T_{P_E}()$
\FOR{$j \in [1, \dots, m]$:}
\STATE With probability $\epsilon$ let $o$ be a random option; otherwise, $o \leftarrow \argmax\limits_{o' \in \mathcal{O}} Q(f, s, o')$
\STATE $s' \sim T_o(s)$
\STATE $f' \sim T_F(T_{P_G}(s'), \bar{p}_e, f)$
\STATE $Q_k(f, s, o) \leftarrow Q_{k-1}(f, s, o) + \alpha \big( R_F(f)R_o(s) + \gamma V(f', s') - Q_{k-1}(f, s, o) \big)$ \label{eq:lof-ql-update1}
\STATE $V_k(f, s) \leftarrow \max\limits_{o' \in \mathcal{O}} Q_k(f, s, o')$ \label{eq:lof-ql-update2}
\STATE $f \leftarrow f'$
\ENDFOR
\ENDFOR
\STATE $\mu(f, s, o) = \argmax\limits_{o \in \mathcal{O}} Q(f, s, o)$
\STATE \textbf{Return: } Options $\mathcal{O}$, meta-policy $\mu(f, s, o)$ and Q- and value functions $Q(f, s, o), V(f, s)$
\end{algorithmic}
\end{algorithm}
\subsection{Comparison of LOF and Reward Machines}\label{sec:appendix-lof-vs-rm}
Figs.~\ref{fig:lof-vs-rm-1},~\ref{fig:lof-vs-rm-rm},~\ref{fig:lof-vs-rm-3}, and~\ref{fig:lof-vs-rm-4} give a visual overview of how LOF and Reward Machines work, and illustrate how they differ.
\begin{figure*}[!th]
\centering
\begin{subfigure}[t]{.35\textwidth}
\centering
\includegraphics[width=0.6\textwidth]{figures/lof_vs_rm/rm_vs_lof-cropped_Part1.pdf}
\caption{Environment MDP $\mathcal{E}$.}
\label{fig:lof-vs-rm-environment}
\end{subfigure} \hfill
\begin{subfigure}[t]{.6\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/lof_vs_rm/rm_vs_lof-cropped_Part2.pdf}
\caption{Liveness property $\mathcal{T}$. The natural language rule can be represented as an LTL formula which can be translated into an FSA.}
\label{fig:lof-vs-rm-fsa}
\end{subfigure}
\caption{LOF and \texttt{RM} both require an environment MDP $\mathcal{E}$ and an automaton $\mathcal{T}$ that specifies a task.}
\label{fig:lof-vs-rm-1}
\end{figure*}
\begin{figure*}[!th]
\centering
\includegraphics[width=0.8\textwidth]{figures/lof_vs_rm/rm_vs_lof-cropped_Part3.pdf}
\caption{In \texttt{RM}, sub-policies are learned for each state of the automaton. In this case, in state $S0$, a sub-policy is learned that goes either to the shopping cart of the kid, whichever is closer. In state $S1$, the sub-policy goes to the house.}
\label{fig:lof-vs-rm-rm}
\end{figure*}
\begin{figure*}[!th]
\centering
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics[width=0.85\textwidth]{figures/lof_vs_rm/rm_vs_lof-cropped_Part4.pdf}
\caption{Step 1 of LOF: Learn a logical option for each subgoal.}
\label{fig:lof-vs-rm-lof1}
\end{subfigure}
~
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics[width=0.85\textwidth]{figures/lof_vs_rm/rm_vs_lof-cropped_Part5.pdf}
\caption{Step 2 of LOF: Use Logical Value Iteration to find a meta-policy that satisfies the liveness property. In this image, the boxed subgoals indicate that the corresponding option is the optimal option to take from that low-level state. The policy ends up being the same as \texttt{RM}'s policy -- in state $S0$, the optimal meta-policy chooses the ``grocery shoppping'' option if the grocery cart is closer and the ``pick up kid'' option if the kid is closer. In the state $S1$, the optimal meta-policy is to always choose the ``home'' option.}
\label{fig:lof-vs-rm-lof2}
\end{subfigure}
\caption{LOF has two steps. In (a) the first step, logical options are learned for each subgoal. In (b) the second step, a meta-policy is found using Logical Value Iteration.}
\label{fig:lof-vs-rm-3}
\end{figure*}
\begin{figure*}[!th]
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics[width=0.5\textwidth]{figures/lof_vs_rm/rm_vs_lof-cropped_Part6.pdf}
\caption{LOF can easily solve this new liveness property without training new options.}
\label{fig:lof-vs-rm-new-task}
\end{subfigure}
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics[width=0.85\textwidth]{figures/lof_vs_rm/rm_vs_lof-cropped_Part7.pdf}
\caption{Logical Value Iteration can be used to find a meta-policy on the new task without the need to retrain the logical options. A new meta-policy can be found in 10-50 iterations. The new policy finds that in state $S0$, ``home'' option is optimal if the agent is closer to ``home'', and the ``kid'' option is optimal if the agent is closer to ``kid''. In state $S1$, the ``grocery shopping'' option is optimal everywhere.}
\label{fig:lof-vs-rm-compose}
\end{subfigure}
\caption{What distinguishes LOF from \texttt{RM} is that the logical options of LOF can be easily composed to solve new tasks. In this example, the new task is to go home or pick up the kid, then go grocery shopping. Logical Value Iteration can find a new meta-policy in 10-50 iterations without needing to relearn the options.}
\label{fig:lof-vs-rm-4}
\end{figure*}
\subsection{Tasks}\label{appendix:tasks}
We test the environments on four tasks, a ``sequential'' task (Fig.~\ref{fig:fsa-sequential}), an ``IF'' task (Fig.~\ref{fig:fsa-if}), an ``OR'' task (Fig.~\ref{fig:fsa-or}), and a ``composite'' task (Fig.~\ref{fig:fsa-composite}). The reacher domain has the same tasks, expect $r, g, b, y$ replace $a, b, c, h$, and there are no obstacles $o$. Note that in the LTL formulae, $\Box !o$ is the safety property $\phi_{safety}$; the preceding part of the formula is the liveness property $\phi_{liveness}$ used to construct the FSA.
\begin{figure}[!htb]
\centering
\resizebox{.95\linewidth}{!}{\input{figures/fsa/sequential_delivery}}
\caption{FSA for the sequential task. The LTL formula is $\diamondsuit(a \land \diamondsuit(b \land \diamondsuit(c \land \diamondsuit h))) \land \Box ! o$. The natural language interpretation is ``Deliver package $a$, then $b$, then $c$, and then return home $h$. And always avoid obstacles $o$''.}
\label{fig:fsa-sequential}
\end{figure}
\begin{figure}[!htb]
\centering
\resizebox{.7\linewidth}{!}{\input{figures/fsa/if_delivery}}
\caption{FSA for the IF task. The LTL formula is $(\diamondsuit (c \land \diamondsuit a) \land \Box ! can) \lor (\diamondsuit a \land \diamondsuit can) \land \Box ! o$. The natural language interpretation is ``Deliver package $c$, and then $a$, unless $a$ gets cancelled. And always avoid obstacles $o$''.}
\label{fig:fsa-if}
\end{figure}
\begin{figure}[!htb]
\centering
\resizebox{.6\linewidth}{!}{\input{figures/fsa/or_delivery}}
\caption{FSA for the OR task. The LTL formula is $\diamondsuit ((a \lor b) \land \diamondsuit c) \land \Box ! o$. The natural language interpretation is ``Deliver package $a$ or $b$, then $c$, and always avoid obstacles $o$''.}
\label{fig:fsa-or}
\end{figure}
\begin{figure}[!htb]
\centering
\resizebox{.95\linewidth}{!}{\input{figures/fsa/composite_delivery}}
\caption{FSA for the composite task. The LTL formula is $(\diamondsuit((a \lor b) \land \diamondsuit(c \land \diamondsuit h)) \land \Box ! can) \lor (\diamondsuit((a \lor b) \land \diamondsuit h) \land \diamondsuit can) \land \Box ! o$. The natural language interpretation is ``Deliver package $a$ or $b$, and then $c$, unless $c$ gets cancelled, and then return to home $h$. And always avoid obstacles''.}
\label{fig:fsa-composite}
\end{figure}
\pagebreak
\subsection{Full Experimental Results}\label{sec:appendix-results}
For the satisfaction experiments for the delivery domain, 10 policies were trained for each task and for each baseline. Training was done for 1600 episodes, with 100 steps per episode. Every 2000 training steps, the policies were tested on the domain and the returns recorded. For this discrete domain, we know the minimum and maximum possible returns for each task, and we normalized the returns using these minimum and maximum returns. The error bars are the standard deviation of the returns over the 10 policies' rollouts.
For the satisfaction experiments for the reacher domain, a single policy was trained for each task and for each baseline. The baselines were trained for 900 epochs, with 50 steps per epoch. Every 2500 training steps, each policy was tested by doing 10 rollouts and recording the returns. For the \texttt{RM} baseline, training was for 1000 epochs with 800 steps per epoch, and the policy was tested every 8000 training steps. Because we don't know the minimum and maximum rewards for each task, we did not normalize the returns. The error bars are the standard deviation over the 10 rollouts for each baseline.
For the composability experiments, a set of options was trained once, and then meta-policing training using \texttt{LOF-VI}, \texttt{LOF-QL}, and \texttt{Greedy} was done for each task. Returns were recorded at every training step by rolling out each baseline 10 times. The error bars are the standard deviations on the 10 rollouts.
For the pick-and-place domain, 1 policy was trained for the satisfaction experiments, and experimental results were evaluated over 2 rollouts. Training was done for 7500 epochs with 1000 steps per epoch. Every 250,000 training steps, the policy was tested by doing 2 rollouts and recording the returns. For the composability experiments, returns were recorded by rolling out each baseline 2 times. The \texttt{RM} baseline was trained over 10000 epochs with 1000 steps per epoch.
Code and videos of the domains and tasks are in the supplement.
\begin{figure*}[!th]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=3cm]{figures/delivery_environment-cropped.pdf}
\caption{Delivery domain.}
\label{fig:appendix-discrete-domain}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/satisfaction/results_averaged_over_tasks.png}
\caption{Averaged.}
\label{fig:appendix-discrete-satisfaction-average}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/satisfaction/results_complex.png}
\caption{Composite.}
\label{fig:appendix-discrete-satisfaction-composite}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/satisfaction/results_OR.png}
\caption{OR.}
\label{fig:appendix-discrete-satisfaction-or}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/satisfaction/results_IF.png}
\caption{IF.}
\label{fig:appendix-discrete-satisfaction-if}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/satisfaction/results_sequential.png}
\caption{Sequential.}
\label{fig:appendix-discrete-satisfaction-sequential}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[height=0.6cm]{figures/discrete/satisfaction/legend.png}
\label{fig:appendix-discrete-satisfaction-legend}
\end{subfigure}
\caption{All satisfaction experiments on the delivery domain. Notice how for the composite and OR tasks (Figs.~\ref{fig:appendix-discrete-satisfaction-composite} and~\ref{fig:appendix-discrete-satisfaction-or}), the \texttt{Greedy} baseline plateaus before \texttt{LOF-VI} and \texttt{LOF-QL}. This is because \texttt{Greedy} chooses a suboptimal path through the FSA, whereas \texttt{LOF-VI} and \texttt{LOF-QL} find an optimal path. Also, notice that \texttt{RM} takes many more training steps to achieve the optimal cumulative reward. This is because for \texttt{RM}, the only reward signal is from reaching the goal state. It takes a long time for the agent to learn an optimal policy from such a sparse reward signal. This is particularly evident for the sequential task (Fig.~\ref{fig:appendix-discrete-satisfaction-sequential}), which requires the agent to take a longer sequence of actions/FSA states before reaching the goal. The options-based algorithms train much faster because when training the options, the agent receives a reward for reaching each subgoal, and therefore the reward signal is much richer.}
\label{fig:appendix-discrete-satisfaction-experiments}
\end{figure*}
\begin{figure*}[!th]
\centering
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/reacher_environment-cropped.pdf}
\caption{Reacher domain.}
\label{fig:appendix-continuous-domain}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/results_averaged_over_tasks.png}
\caption{Averaged.}
\label{fig:appendix-continuous-satisfaction-average}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/results_composite.png}
\caption{Composite.}
\label{fig:appendix-continuous-satisfaction-composite}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/results_OR.png}
\caption{OR.}
\label{fig:appendix-continuous-satisfaction-or}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/results_IF.png}
\caption{IF.}
\label{fig:appendix-continuous-satisfaction-if}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/results_sequential.png}
\caption{Sequential.}
\label{fig:appendix-continuous-satisfaction-sequential}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[height=0.6cm]{figures/continuous/satisfaction/legend.png}
\label{fig:appendix-continuous-satisfaction-legend}
\end{subfigure}
\caption{Satisfaction experiments for the reacher domain, without \texttt{RM} results. The results are equivalent to the results on the delivery domain.}
\label{fig:appendix-continuous-satifaction-experiments}
\end{figure*}
\begin{figure*}[!th]
\centering
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/reacher_environment-cropped.pdf}
\caption{Reacher domain.}
\label{fig:appendix-continuous-domain-with-rm}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/with_rm/results_averaged_over_tasks.png}
\caption{Averaged.}
\label{fig:appendix-continuous-satisfaction-average-with-rm}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/with_rm/results_composite.png}
\caption{Composite.}
\label{fig:appendix-continuous-satisfaction-composite-with-rm}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/with_rm/results_OR.png}
\caption{OR.}
\label{fig:appendix-continuous-satisfaction-or-with-rm}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/with_rm/results_IF.png}
\caption{IF.}
\label{fig:appendix-continuous-satisfaction-if-with-rm}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/with_rm/results_sequential.png}
\caption{Sequential.}
\label{fig:appendix-continuous-satisfaction-sequential-with-rm}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[height=0.6cm]{figures/continuous/satisfaction/with_rm/legend.png}
\label{fig:appendix-continuous-satisfaction-legend-with-rm}
\end{subfigure}
\caption{Satisfaction experiments for the reacher domain, including \texttt{RM} results. \texttt{RM} takes significantly more training steps to train than the other baselines, although it eventually reaches and surpasses the cumulative reward of the other baselines. This is because for the continuous domain, we violate some of the conditions required for optimality when using the Logical Options Framework -- in particular, the condition that each subgoal is associated with a single state. In a continuous environment, this condition is impossible to meet, and therefore we made the subgoals small spherical regions, and we only made the subgoals associated with specific Cartesian coordinates and not velocities (which are also in the state space). Meanwhile, the optimality conditions of \texttt{RM} are looser and were not violated, which is why it achieves a higher final cumulative reward.}
\label{fig:appendix-continuous-satifaction-experiments-with-rm}
\end{figure*}
\begin{figure*}[!th]
\centering
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=3.4cm]{figures/arm_domain-cropped.pdf}
\caption{Reacher domain.}
\label{fig:appendix-arm-domain}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/results_averaged_over_tasks.png}
\caption{Averaged.}
\label{fig:appendix-arm-satisfaction-average}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/results_composite.png}
\caption{Composite.}
\label{fig:appendix-arm-satisfaction-composite}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/results_OR.png}
\caption{OR.}
\label{fig:appendix-arm-satisfaction-or}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/results_IF.png}
\caption{IF.}
\label{fig:appendix-arm-satisfaction-if}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/results_sequential.png}
\caption{Sequential.}
\label{fig:appendix-arm-satisfaction-sequential}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[height=0.6cm]{figures/arm/satisfaction/legend.png}
\label{fig:appendix-arm-satisfaction-legend}
\end{subfigure}
\caption{Satisfaction experiments for the pick-and-place domain, without \texttt{RM} results. The results are equivalent to the results on the delivery and reacher domains.}
\label{fig:appendix-arm-satifaction-experiments}
\end{figure*}
\begin{figure*}[!th]
\centering
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=3.4cm]{figures/arm_domain-cropped.pdf}
\caption{Reacher domain.}
\label{fig:appendix-arm-domain-with-rm}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/with_rm/results_averaged_over_tasks.png}
\caption{Averaged.}
\label{fig:appendix-arm-satisfaction-average-with-rm}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/with_rm/results_composite.png}
\caption{Composite.}
\label{fig:appendix-arm-satisfaction-composite-with-rm}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/with_rm/results_OR.png}
\caption{OR.}
\label{fig:appendix-arm-satisfaction-or-with-rm}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/with_rm/results_IF.png}
\caption{IF.}
\label{fig:appendix-arm-satisfaction-if-with-rm}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/with_rm/results_sequential.png}
\caption{Sequential.}
\label{fig:appendix-arm-satisfaction-sequential-with-rm}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[height=0.6cm]{figures/arm/satisfaction/with_rm/legend.png}
\label{fig:appendix-arm-satisfaction-legend-with-rm}
\end{subfigure}
\caption{Satisfaction experiments for the pick-and-place domain, including \texttt{RM} results. For the pick-and-place domain, \texttt{RM} did not converge to a solution within the training time allotted for it (10 million training steps).}
\label{fig:appendix-arm-satifaction-experiments-with-rm}
\end{figure*}
\begin{figure*}[!th]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=3cm]{figures/delivery_environment-cropped.pdf}
\caption{Delivery domain.}
\label{fig:appendix-discrete-domain-composability}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/composability/results_averaged_over_tasks.png}
\caption{Averaged.}
\label{fig:appendix-discrete-composability-average}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/composability/results_composite.png}
\caption{Composite.}
\label{fig:appendix-discrete-composability-composite}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/composability/results_OR.png}
\caption{OR.}
\label{fig:appendix-discrete-composability-or}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/composability/results_IF.png}
\caption{IF.}
\label{fig:appendix-discrete-composability-if}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/composability/results_sequential.png}
\caption{Sequential.}
\label{fig:appendix-discrete-composability-sequential}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[height=0.6cm]{figures/discrete/composability/legend.png}
\label{fig:appendix-discrete-composability-legend}
\end{subfigure}
\caption{All composability experiments for the delivery domain.}
\label{fig:appendix-discrete-composability-experiments}
\end{figure*}
\begin{figure*}[!th]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/reacher_environment-cropped.pdf}
\caption{Delivery domain.}
\label{fig:appendix-continuous-domain-composablity}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/composability/results_averaged_over_tasks.png}
\caption{Averaged.}
\label{fig:appendix-continuous-composability-average}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/composability/results_composite.png}
\caption{Composite.}
\label{fig:appendix-continuous-composability-composite}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/composability/results_OR.png}
\caption{OR.}
\label{fig:appendix-continuous-composability-or}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/composability/results_IF.png}
\caption{IF.}
\label{fig:appendix-continuous-composability-if}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/composability/results_sequential.png}
\caption{Sequential.}
\label{fig:appendix-continuous-composability-sequential}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[height=0.6cm]{figures/continuous/composability/legend.png}
\label{fig:appendix-continuous-composability-legend}
\end{subfigure}
\caption{All composability experiments for the reacher domain.}
\label{fig:appendix-continuous-composability-experiments}
\end{figure*}
\begin{figure*}[!th]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=3.4cm]{figures/arm_domain-cropped.pdf}
\caption{Delivery domain.}
\label{fig:appendix-arm-domain-composablity}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/composability/results_averaged_over_tasks.png}
\caption{Averaged.}
\label{fig:appendix-arm-composability-average}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/composability/results_composite.png}
\caption{Composite.}
\label{fig:appendix-arm-composability-composite}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/composability/results_OR.png}
\caption{OR.}
\label{fig:appendix-arm-composability-or}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/composability/results_IF.png}
\caption{IF.}
\label{fig:appendix-arm-composability-if}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/composability/results_sequential.png}
\caption{Sequential.}
\label{fig:appendix-arm-composability-sequential}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[height=0.6cm]{figures/arm/composability/legend.png}
\label{fig:appendix-arm-composability-legend}
\end{subfigure}
\caption{All composability experiments for the pick-and-place domain.}
\label{fig:appendix-arm-composability-experiments}
\end{figure*}
\section{Further Discussion}
\paragraph{What happens when incorrect rules are used?} One benefit of representing the rules of the environment as LTL formulae/automata is that these forms of representing rules are much more interpretable than alternatives (such as neural nets). Therefore, if an agent's learned policy has bad behavior, a user of LOF can inspect the rules to see if the bad behavior is a consequence of a bad rule specification. Furthermore, one of the consequences of composability is that any modifications to the FSA will alter the resulting policy in a direct and predictable way. Therefore, for example, if an incorrect human-specified task yields undesirable behavior, with our framework it is possible to tweak the task and test the new policy without any additional low-level training (however, tweaking the safety rules would require retraining the logical options).
\paragraph{What happens if there is a rule conflict?} If the specified LTL formula is invalid, the LTL-to-automaton translation tool will either throw an error or return a trivial single-state automaton that is not an accepting state. Rollouts would terminate immediately.
\paragraph{What happens if the agent can't satisfy a task without violating a rule?} The solution to this problem depends on the user's priorities. In our formulation, we have assigned finite costs to rule violations and an infinite cost to not satisfying the task (see Appendix~\ref{sec:appendix-proofs}). We have prioritized task satisfaction over safety satisfaction. However, it is possible to flip the priorities around by terminating training/rollouts if there is a safety violation. In our proofs, we have assumed that the agent can reach every subgoal from any state, implying either that it is always possible to avoid safety violations or that safety violations are allowed.
\paragraph{Why is the safety property not composable?} The safety property is not composable because we allow safety propositions to be associated with more than one state in the environment (unlike subgoals). The fact that there can be multiple instances of a safety proposition in the environment means that it is impossible to guarantee that a new option policy will be optimal if retraining is done only at the level of the safety automaton and not also over the low-level states. In order to guarantee optimality, retraining would have to be done over both the high and low levels (the safety automaton and the environment). Our definition of composability involves only replanning over the high level of the FSA. Therefore, safety properties are not composable. Furthermore, rewards/costs of the safety property can be associated with propositions and not just with states (as with the liveness property). This is because a safety violation via one safety proposition (e.g., a car going onto the wrong side of the road) may incur a different penalty than a violation via a different proposition (a car going off the road). The propositions are associated with low-level states of the environment. Therefore any retraining would have to involve retraining at both the high and low levels, once again violating our definition of composability.
\paragraph{Simplifying the option transition model:} In our experiments, we simplify the transition model by setting $\gamma = 1$, an assumption that does not affect convergence to optimality. In the case where $\gamma = 1$, Eq.~\ref{eq:option-transitions} reduces to $T_o(s' \vert s) = \sum_k p(s', k)$. Assuming that the option terminates only at state $s_g$, then Eq.\ref{eq:option-transitions} further reduces to $T_o(s_g | s) = 1$ and $T_o(s' | s) = 0$ for all other $s' \neq s_g$. Therefore no learning is required for the transition model. For cases where the assumption that $\gamma = 1$ does not apply, \citep{abel2019expected} contains an interesting discussion.
\paragraph{Learning the option reward model:} The option reward model $R_o(s)$ is the expected reward of carrying out option $o$ to termination from state $s$. It is equivalent to a value function. Therefore, it is convenient if the policy-learning algorithm used to learn the options learns a value function as well as a policy (e.g., Q-learning and PPO). However, as long as the expected return can be computed between pairs of states, it is not necessary to learn a complete value function. This is because during Logical Value Iteration, the reward model is only queried at discrete points in the state space (typically corresponding to the initial state and the subgoals). So as long as expected returns between the initial state and subgoals can be computed, Logical Value Iteration will work.
\paragraph{Why is \texttt{LOF-VI} so much more efficient than the \texttt{RM} baseline?} In short, \texttt{LOF-VI} is more efficient than \texttt{RM} because \texttt{LOF-VI} has a dense reward function during training and \texttt{RM} has a sparse reward function. During training, \texttt{LOF-VI} trains the options independently and rewards the agent for reaching the subgoals associated with the options. This is in effect a dense reward function. The generic reward function for \texttt{RM} only rewards the agent for reaching the goal state. There are no other high-level rewards to guide the agent through the task. This is a very sparse reward that results in less efficient training. \texttt{RM}'s reward function could easily be made dense by rewarding every transition of the automaton. In this case, \texttt{RM} would probably train as efficiently as \texttt{LOF-VI}. However, imagine an FSA with two paths to the goal state. One path has only 1 transition but has much lower low-level cost, and one path has 20 transitions and a much higher low-level cost. \texttt{RM} might learn to prefer the reward-heavy 20-transition path rather than the reward-light 1-transition path, even if the 1-transition path results in a lower low-level cost. In theory it might be possible to design an \texttt{RM} reward function that adjusts the automaton transition reward depending on the length of the path that the state is in, but this would not be a trivial task when accounting for branching and merging paths. We therefore decided that it would be a fairer comparison to use a trivial \texttt{RM} reward function, just as we use a trivial reward function for the LOF baselines. However, we were careful to not list increased efficiency in our list of contributions; although increased efficiency was an observed side effect of LOF, LOF is not inherently more efficient than other algorithms besides the fact that it automatically imposes a dense reward on reaching subgoals.
\section{Formulation of Logical Options Framework with Safety Automaton}\label{sec:appendix-lof}
In this section, we present a more general formulation of LOF than that presented in the paper. In the paper, we make two assumptions that simplify the formulation. The first assumption is that the LTL specification can be divided into two independent formulae, a liveness property and a safety property: $\phi = \phi_{liveness} \land \phi_{safety}$. However, not all LTL formulae can be factored in this way. We show how LOF can be applied to LTL formulae that break this assumption. The second assumption is that the safety property takes a simple form that can be represented as a penalty on safety propositions. We show how LOF can be used with arbitrary safety properties.
\subsection{Automata and Propositions}
All LTL formulae can be translated into B\"uchi automata using automatic translation tools such as SPOT \citep{Duret16}. All B\"uchi automata can be decomposed into liveness and safety properties \citep{alpern1987recognizing}, so that automaton $\mathcal{W} = \mathcal{W}_{liveness} \times \mathcal{W}_{safety}$. This is a generalization of the assumption that all LTL formulae can be divided into liveness and safety properties $\phi_{liveness}$ and $\phi_{safety}$. The liveness property $\mathcal{W}_{liveness}$ must be an FSA, although this assumption could also be loosened to allow it to be a deterministic B\"uchi automaton via some minor modifications (allowing multiple goal states to exist and continuing episodes indefinitely, even once a goal state has been reached).
As in the main text, we assume that there are three types of propositions -- subgoals $\mathcal{P}_G$, safety propositions $\mathcal{P}_S$, and event propositions $\mathcal{P}_E$. The event propositions have set values and can occur in both $\mathcal{W}_{liveness}$ and $\mathcal{W}_{safety}$. Safety propositions only appear in $\mathcal{W}_{safety}$. Subgoal propositions only appear in $\mathcal{W}_{liveness}$. Each subgoal may only be associated with one state. Note that after writing a specification and decomposing it into $\mathcal{W}_{liveness}$ and $\mathcal{W}_{safety}$, it is possible that some subgoals may unexpectedly appear in $\mathcal{W}_{safety}$. This can be dealt with by creating ``safety twins'' of each subgoal -- safety propositions that are associated with the same low-level states as the subgoals and can therefore substitute for them in $\mathcal{W}_{safety}$.
Subgoals are propositions that the agent must achieve in order to reach the goal state of $\mathcal{W}_{liveness}$. Although event propositions can also define transitions in $\mathcal{W}_{liveness}$, we assume that ``achieving'' them is not necessary in order to reach the goal state. In other words, we assume that from any state in $\mathcal{W}_{liveness}$, there is a path to the goal state that involves only subgoals. This is because in our formulation, the event propositions are meant to serve as propositions that the agent has no control over, such as receiving a phone call. If satisfaction of the liveness property were to depend on such a proposition, then it would be impossible to guarantee satisfaction. However, if the user is unconcerned with guaranteeing satisfaction, then specifying a liveness property in which satisfaction depends on event propositions is compatible with LOF.
Safety propositions may only occur in $\mathcal{W}_{safety}$ and are associated with things that the agent ``must avoid''. This is because every state of $\mathcal{W}_{safety}$ is an accepting state \citep{alpern1987recognizing}, so all transitions between the states are non-violating. However, any undefined transition is not allowed and is a violation of the safety property. In our formulation, we assign costs to violations, so that violations are allowed but come at a cost. In practice, it also may be the case that the agent is in a low-level state from which it is impossible to reach the goal state without violating the safety property. In our formulation, satisfaction of the liveness property (but not the safety property) is still guaranteed in this case, as the finite cost associated with violating the rule is less than the infinite cost of not satisfying the liveness property, so the optimal policy for the agent will be to violate the rule in order to satisfy the task (see the proofs, Appendix~\ref{sec:appendix-proofs}). This scenario can be avoided in several ways. For example, do not specify an environment in which it is only possible for the agent to satisfy the task by violating a rule. Or, instead of prioritizing satisfaction of the task, it is possible to instead prioritize satisfaction of the safety property. In this case, satisfaction of the liveness property would not be guaranteed but satisfaction of the safety property would be guaranteed. This could be accomplished by terminating the rollout if a safety violation occurs.
We assume that event propositions are observed -- in other words, that we know the values of the event propositions from the start of a rollout. This is because we are planning in a fully observable setting, so we must make this assumption to guarantee convergence to an optimal policy. However, the partially observable case is much more interesting, in which the values of the event propositions are not known until the agent checks or the environment randomly reveals their values. This case is beyond the scope of this paper; however, LOF can still guarantee satisfaction and composability in this setting, just not optimality.
Proposition labeling functions relate states to propositions: $T_{P_G} : \mathcal{S} \rightarrow 2^{\mathcal{P}_G}$, $T_{P_S} : \mathcal{S} \rightarrow 2^{\mathcal{P}_S}$, and $T_{P_E} : 2^{\mathcal{P}_E} \rightarrow \{0, 1\}$.
Given these definitions of propositions, it is possible to define the liveness and safety properties formally. $\mathcal{W}_{liveness} = (\mathcal{F}, \mathcal{P}_G \cup \mathcal{P}_E, T_F, R_F, f_0, f_g)$. $\mathcal{F}$ is the set of states of the liveness property. The propositions can be either subgoals $\mathcal{P}_G$ or event propositions $\mathcal{P}_E$. The transition function relates the current FSA state and active propositions to the next FSA state, $T_F : \mathcal{F} \times 2^{\mathcal{P}_G} \times 2^{\mathcal{P}_E} \times \mathcal{F} \rightarrow [0, 1]$. The reward function assigns a reward to the current FSA state, $R_F : \mathcal{F} \rightarrow \mathbb{R}$. We assume there is one initial state $f_0$ and one goal state $f_g$.
The safety property is a B\"uchi automaton $\mathcal{W}_{safety} = (\mathcal{F}_S, \mathcal{P}_S \cup \mathcal{P}_E, T_S, R_S, F_0)$. $\mathcal{F}_S$ are the states of the automaton. The propositions can be safety propositions $\mathcal{P}_S$ or event propositions $\mathcal{P}_E$. The transition function $T_S$ relates the current state and active propositions to the next state, $T_S : \mathcal{F}_S \times 2^{\mathcal{P}_S} \times 2^{\mathcal{P}_E} \times \mathcal{F}_S \rightarrow [0, 1]$. The reward function relates the automaton state and safety propositions to rewards (or costs), $R_S : \mathcal{F}_S \times 2^{\mathcal{P}_S} \rightarrow \mathbb{R}$. $F_0$ defines the set of initial states. We do not specify an accepting condition because for safety properties, every state is an accepting state.
\subsection{The Environment MDP}
There is a low-level environment MDP $\mathcal{E} = (\mathcal{S}, \mathcal{A}, R_E, T_E, \gamma)$. $\mathcal{S}$ is the state space and $\mathcal{A}$ is the action space. They can be either discrete or continuous. $R_E$ is the low-level reward function that characterizes, for example, time, distance, or actuation costs. $T_E : \mathcal{S} \times \mathcal{A} \times \mathcal{S} \rightarrow [0, 1]$ is the transition function and $\gamma$ is the discount factor. Unlike in the simpler formulation in the paper, we do not combine $R_E$ and the safety automaton reward function $R_S$ in the MDP formulation $\mathcal{E}$.
\begin{algorithm}[!t]
\caption{Learning and Planning with Logical Options}\label{alg:lof-full}
\begin{algorithmic}[1]
\STATE \textbf{Given:} \par
Propositions $\mathcal{P}$ partitioned into subgoals $\mathcal{P}_G$, safety propositions $\mathcal{P}_S$, and event propositions $\mathcal{P}_E$ \par
$\mathcal{W}_{liveness} = (\mathcal{F}, \mathcal{P}_G \cup \mathcal{P}_E, T_F, R_F, f_0, f_g)$ \par
$\mathcal{W}_{safety} = (\mathcal{F}_S, \mathcal{P}_S \cup \mathcal{P}_E, T_S, R_S, F_0)$ \par
Low-level MDP $\mathcal{E} = (\mathcal{S}, \mathcal{A}, R_E, T_E, \gamma)$ \par
Proposition labeling functions $T_{P_G} : \mathcal{S} \rightarrow 2^{\mathcal{P}_G}$, $T_{P_S} : \mathcal{S} \rightarrow 2^{\mathcal{P}_S}$, and $T_{P_E} : 2^{\mathcal{P}_E} \rightarrow \{0, 1\}$ \par
\STATE \textbf{To learn:}
\STATE Set of options $\mathcal{O}$, one for each subgoal proposition $p \in \mathcal{P}_G$
\STATE Meta-policy $\mu(f, f_s, s, o)$ along with $Q(f, f_s, s, o)$ and $V(f, f_s, s)$
\STATE \textbf{Learn logical options:}\label{alg:logical-options-appendix}
\STATE For every $p$ in $\mathcal{P}_G$, learn an option for achieving $p$, $o_p = (\mathcal{I}_{o_p}, \pi_{o_p}, \beta_{o_p}, R_{o_p}, T_{o_p})$ \par
\STATE $ \mathcal{I}_{o_p} = \mathcal{S} $
\STATE $ \beta_{o_p} = \begin{cases} 1 &\quad\text{if } p \in T_{P_G}(s)\\
0 &\quad\text{otherwise} \\
\end{cases} $
\STATE $\pi_{o_p} = $ optimal policy on $\mathcal{E} \times \mathcal{W}_{safety}$ with rollouts terminating when $p \in T_{P_G}(s)$
\STATE $T_{o_p}(f'_s, s' \vert f_s, s) = \begin{cases} \sum\limits_{k=1}^{\infty} p(f'_s, k) \gamma^k &\quad\text{if } p \in T_P(s') \\
0 &\quad\text{otherwise} \\
\end{cases}$
\STATE $R_{o_p}(f_s, s) = \mathbb{E} [ \mathcal{R}_\mathcal{E}(f_{s}, s, a_1) + \gamma \mathcal{R}_\mathcal{E}(f_{s, 1}, s_1, a_2) + \dots + \gamma^{k-1} \mathcal{R}_\mathcal{E}(f_{s, k-1}, s_{k-1}, a_{k}) ]$
\STATE \textbf{Find a meta-policy $\mu$ over the options:}
\STATE Initialize $Q : \mathcal{F} \times \mathcal{F}_S \times \mathcal{S} \times \mathcal{O} \rightarrow \mathbb{R}$ and $V : \mathcal{F} \times \mathcal{F}_S \times \mathcal{S} \rightarrow \mathbb{R}$ to $0$
\FOR{$(k, f, f_s, s) \in [1, \dots, n] \times \mathcal{F} \times \mathcal{F}_S \times \mathcal{S}$:}
\FOR{$o \in \mathcal{O}$:}
\STATE $Q_k(f, f_s, s, o) \leftarrow R_F(f)R_o(f_s, s) + $ \par $\sum\limits_{f' \in \mathcal{F}} \sum\limits_{f'_s \in \mathcal{F}_S} \sum\limits_{\bar{p}_e \in 2^{\mathcal{P}_E}} \sum\limits_{s' \in \mathcal{S}} T_F(f' \vert f, T_{P_G}(s'), \bar{p}_e)$ \par $ T_S(f'_s \vert f_s, T_{P_S}(s'), \bar{p}_e) T_{P_E}(\bar{p}_e)$ \par $T_o(s' \vert s) V_{k-1}(f', f'_s, s')$
\ENDFOR
\STATE $V_k(f, f_s, s) \leftarrow \max\limits_{o \in \mathcal{O}} Q_k(f, f_s, s, o)$
\ENDFOR
\STATE $\mu(f, f_s, s, o) = \argmax\limits_{o \in \mathcal{O}} Q(f, f_s, s, o)$
\STATE \textbf{Return: } Options $\mathcal{O}$, meta-policy $\mu(f, f_s, s, o)$, and $Q(f, f_s, s, o), V(f, f_s, s)$
\end{algorithmic}
\end{algorithm}
\subsection{Logical Options}
We associate every subgoal $p_g$ with an option $o_{p_g} = (\mathcal{I}_{p_g}, \pi_{p_g}, \beta_{p_g}, R_{p_g}, T_{p_g})$. Every $o_{p_g}$ has a policy $\pi_{p_g}$ whose goal is to reach the state $s_{p_g}$ where $p_g$ is true. Option policies are learned by training on the product of the environment and the safety automaton, $\mathcal{E} \times \mathcal{W}_{safety}$ and terminating training only when $s_{p_g}$ is reached. $R_\mathcal{E} : \mathcal{F}_S \times \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$ is the reward function of the product MDP $\mathcal{E} \times \mathcal{W}_{safety}$. There are many reward-shaping policy-learning algorithms that specify how to define $R_\mathcal{E}$. In fact, learning a policy for $\mathcal{E} \times \mathcal{W}_{safety}$ is the sort of hierarchical learning problem that many reward-shaping algorithms excel at, including Reward Machines \citep{icarte2018using} and \citep{li2017reinforcement}. This is because in LOF, safety properties are not composable, so using a learning algorithm that is satisfying and optimal but not composable to learn the safety property is appropriate. Alternatively, there are many scenarios where $\mathcal{W}_{safety}$ is a trivial automaton in which each safety proposition is associated with its own state, as we describe in the main paper, so penalties can be assigned to propositions and the state of the agent in $\mathcal{W}_{safety}$ can be ignored.
Note that since the options are trained independently, one limitation of our formulation is that the safety properties cannot depend on the liveness state. In other words, when an agent reaches a new subgoal, the safety property cannot change. However, the workaround for this is not too complicated. First, if the liveness state affects the safety property, this implies that liveness propositions such as subgoals may be in the safety property. In this case, as we described above, the subgoals present in the safety property need to be substituted with ``safety twin'' propositions. Then during option training, a policy-learning algorithm must be chosen that will learn sub-policies for all of the safety property states, even if those states are only reached after completing a complicated task (for example, all of the sub-policies could be trained in parallel as in \citep{icarte2018using}). Lastly, during meta-policy learning and during rollouts, when a new option is chosen, the current state of the safety property must be passed to the new option.
The components of the logical options are defined starting at Alg.~\ref{alg:lof-full} line~\ref{alg:logical-options-appendix}. Note that for stochastic low-level transitions, the number of time steps $k$ at which the option terminates is stochastic and characterized by a distribution function. In general this distribution function must be learned, which is a challenging problem. However, there are many approaches to solving this problem; \citep{abel2019expected} contains an excellent discussion.
The most notable difference between the general formulation and the formulation in the paper is that the option policy, transition, and reward functions are functions of the safety automaton state $f_s$ as well as the low-level state $s$. This makes Logical Value Iteration more complicated, because in the paper, we could assume we knew the final state of each option (i.e., the state of its associated subgoal $s_g$). But now, although we still assume that the option will terminate at $s_g$, we do not know which safety automaton state it will terminate in, so the transition model must learn a distribution over safety automaton states, and Logical Value Iteration must account for this uncertainty.
\subsection{Hierarchical SMDP}
Given a low-level environment $\mathcal{E}$, a liveness property $\mathcal{W}_{liveness}$, a safety property $\mathcal{W}_{safety}$, and logical options $\mathcal{O}$, we can define a hierarchical semi-Markov Decision Process (SMDP) $\mathcal{M} = \mathcal{E} \times \mathcal{W}_{liveness} \times \mathcal{W}_{safety}$ with options $\mathcal{O}$ and reward function $R_{SMDP}$. This SMDP differs significantly from the SMDP in the paper in that the safety property $\mathcal{W}_{safety}$ is now an integral part of the formulation. $R_{SMDP}(f, f_s, s, o) = R_F(f)R_o(f_s, o)$.
\subsection{Logical Value Iteration}
A value function and Q-function are found for the SMDP using the Bellman update equations:
\begin{align}
\begin{split}\label{eq:q-update-appendix}
Q_k(&f, f_s, s, o) \leftarrow R_F(f)R_o(f_s, s) + \sum_{f' \in \mathcal{F}} \sum_{f'_s \in \mathcal{F}_S} \\
&\sum_{\bar{p}_e \in 2^{\mathcal{P}_E}} \sum_{s' \in \mathcal{S}} T_F(f' \vert f, T_{P_G}(s'), \bar{p}_e) \\
&T_S(f'_s \vert f_s, T_{P_S}(s'), \bar{p}_e) T_{P_E}(\bar{p}_e) T_o(s' \vert s) V_{k-1}(f', f'_s, s')
\end{split}\\
V_k(&f, f_s, s) \leftarrow \max_{o \in \mathcal{O}} Q_k(f, f_s, s, o) \label{eq:v-update-appendix}
\end{align}
\section{Proofs and Conditions for Satisfaction and Optimality}\label{sec:appendix-proofs}
The proofs are based on the more general LOF formulation of Appendix~\ref{sec:appendix-lof}, as results on the more general formulation also apply to the simpler formulation used in the paper.
\begin{definition}
\label{definition:options}
Let the reward function of the environment be $R_\mathcal{E}(f_s, s, a)$, which is some combination of $R_E(s, a)$ and $R_S(f_s, \bar{p}_s) = R_S(f_s, T_{P_S}(s))$. Let $\pi' : \mathcal{F}_S \times \mathcal{S} \times \mathcal{A} \times \mathcal{S} \rightarrow [0, 1]$ be the optimal goal-conditioned policy for reaching a state $s'$. In the case of a goal-conditioned policy, the reward function is $R_\mathcal{E}$, and the objective is to maximize the expected reward with the constraint that $s'$ is reached in a finite amount of time. We assume that every state $s'$ is reachable from any state $s$, a standard regularity assumption in MDP literature. Let $V^{\pi'}(f_s, s \vert s')$ be the optimal expected cumulative reward for reaching $s'$ from $s$ with goal-conditioned policy $\pi'$. Let $s_g$ be the state associated with the subgoal, and let $\pi_g$ be the optimal goal-conditioned policy associated with reaching $s_g$. Let $\pi^*$ be the optimal policy for the environment $\mathcal{E}$.
\end{definition}
\begin{condition}
The optimal policy for the option must be the same as the goal-conditioned policy that has subgoal $s_g$ as its goal: $\pi^*(f_s, s) = \pi_g(f_s, s \vert s_g)$. In other words, $V^{\pi_g}(f_s, s \vert s_g) > V^{\pi'}(f_s, s \vert s') \;\; \forall f_s, s, s' \neq s_g$.
\label{lemma:subgoals}
\end{condition}
This condition guarantees that the optimal option policy will always reach the subgoal $s_g$.
It can be achieved by setting all rewards $-\infty < R_\mathcal{E}(f_s, s, a) < 0$ and terminating the episode only when the agent reaches $s_g$. Therefore the expected return for reaching $s_g$ is a bounded negative number, and the expected return for all other states is $-\infty$.
\begin{lemma}
Given that the goal state of $\mathcal{W}_{liveness}$ is reachable from any other state using only subgoals and that there is an option for every subgoal and that all the options meet Condition~\ref{lemma:subgoals}, there exists a meta-policy that can reach the FSA goal state from any non-trap state in the FSA.
\label{lemma:satisfaction-exists}
\end{lemma}
\begin{proof}
This follows from the fact that transitions in $\mathcal{W}_{liveness}$ are determined by achieving subgoals, and it is given that there exists an option for achieving every subgoal. Therefore, it is possible for the agent to execute any sequence of subgoals, and at least one of those sequences must satisfy the task specification since the FSA representing the task specification is finite and satisfiable, and the goal state $f_g$ is reachable from every FSA state $f \in \mathcal{F}$ using only subgoals.
\end{proof}
\begin{definition}
From \citet{dietterich2000hierarchical}: A \textbf{hierarchically optimal} policy for an MDP or SMDP is a policy that achieves the highest cumulative reward among all policies consistent with the given hierarchy.
\label{def:hierarchical-optimality}
\end{definition}
In our case, this means that the hierarchically optimal meta-policy is optimal over the available options.
\begin{definition}
Let the expected cumulative reward function of an option $o$ started at state $(f_s, s)$ be $R_o(f_s, s)$. Let the reward function on the SMDP be $R_{SMDP}(f, f_s, s, o) = R_F(f)R_o(f_s, s)$ with $R_F(f) \geq 0$\footnote{The assumption that $R_{SMDP}(f, f_s, s, o) = R_F(f)R_o(f_s, s)$ and $R_{HMDP}(f, f_s, s, a) = R_F(f)R_\mathcal{E}(f_s, s, a)$ can be relaxed so that $R_{SMDP}$ and $R_{HMDP}$ are functions that are monotonic increasing in the low-level rewards $R_o$ and $R_\mathcal{E}$, respectively.}. Let $\mu' : \mathcal{F} \times \mathcal{F}_S \times \mathcal{S} \times \mathcal{O} \times \mathcal{F} \rightarrow [0, 1] $ be the hierarchically optimal goal-conditioned meta-policy for achieving liveness state $f'$. The objective of the meta-policy is to maximize the reward function $R_{SMDP}$ with the constraint that it reaches $f'$ in a finite number of time steps. Let $V^{\mu'}(f, f_s, s \vert f')$ be the hierarchically optimal return for reaching $f'$ from $(f, f_s, s)$ with goal-conditioned meta-policy $mu'$. Let $\mu^*$ be the hierarchically optimal policy for the SMDP. Let $f_g$ be the goal state, and $\mu_g$ be the hierarchically optimal goal-conditioned meta-policy for achieving the goal state.
\end{definition}
\begin{condition}
The hierarchically optimal meta-policy must be the same as the goal-conditioned meta-policy that has the FSA goal state $f_g$ as its goal: $\mu^*(f,f_s,s) = \mu_g(f,f_s,s \vert f_g)$. In other words, $V^{\mu_g}(f,f_s, s \vert f_g) > V^{\mu'}(f,f_s, s \vert f') \;\; \forall f, f_s, s, f' \neq f_g$.
\label{lemma:satisfaction}
\end{condition}
This condition guarantees that the hierarchically optimal meta-policy will always go to the FSA goal state $f_g$ (thereby satisfying the specification). Here is an example of how this condition can be achieved: If $-\infty < R_\mathcal{E}(f_s, s, a) < 0 \;\; \forall s$, then $R_o(f_s, s) < 0 \;\; \forall f_s, o, s$. Then if $R_F(f) > 0$ (in our experiments, we set $R_F(f) = 1 \;\; \forall f$), $R_{SMDP}(f, f_s, s, o) = R_F(f)R_o(f_s, s) < 0$, and if the episode only terminates when the agent reaches the goal state, then the expected return for reaching $f_g$ is a bounded negative number, and the expected return for all other states is $-\infty$.
\begin{lemma}
From \citep{sutton1999between}: Value iteration on an SMDP converges to the hierarchically optimal policy.
\label{lemma:converge-hierarchical}
\end{lemma}
Therefore, the meta-policy found using the Logical Options Framework converges to a hierarchically optimal meta-policy that satisfies the task specification as long as Conditions~\ref{lemma:subgoals} and~\ref{lemma:satisfaction} are met.
\begin{definition}
Consider the SMDP where planning is allowed over the low-level actions instead of the options. We will call this the hierarchical MDP (HMDP), as this MDP is the product of the low-level environment $\mathcal{E}$, the liveness property $\mathcal{W}_{liveness}$, and the safety property $\mathcal{W}_{safety}$. Let $R_F(f) > 0 \; \forall f$, and let $R_{HMDP}(f, f_s, s, a) = R_F(f)R_\mathcal{E}(f_s, s, a)$, and let $\pi^*_{HMDP}$ be the optimal policy for the HMDP.
\end{definition}
\begin{theorem}
Given Conditions~\ref{lemma:subgoals} and~\ref{lemma:satisfaction}, the hierarchically optimal meta-policy $\mu_g$ with optimal option policies $\pi_g$ has the same expected returns as the HMDP optimal policy $\pi^*$ and satisfies the task specification.
\label{theorem:converge-optimal}
\end{theorem}
\begin{proof}
By Condition~\ref{lemma:subgoals}, every subgoal has an option associated with it whose optimal policy is to go to the subgoal. By Condition~\ref{lemma:satisfaction}, the hierarchically optimal meta-policy will reach the FSA goal state $f_g$. The meta-policy can only accomplish this by going to the subgoals in a sequence that satisfies the task specification. It does this by executing a sequence of options that correspond to a satisfying sequence of subgoals and are optimal in expectation. Therefore, since $R_F(f) > 0 \; \forall f$ and $R_{SMDP}(f, f_s, s, o) = R_F(f)R_o(f_s, s)$, and since the event propositions that affect the order of subgoals necessary to satisfy the task are independent random variables, the expected cumulative reward is a positive linear combination of the expected option rewards, and since all option rewards are optimal with respect to the environment and the meta-policy is optimal over the options, our algorithm attains the optimal expected cumulative reward.
\end{proof}
\section{Background}
\textbf{Linear Temporal Logic:} We use linear temporal logic (LTL) to formally specify rules \citep{Clark01}. LTL can express tasks and rules using temporal operators such as ``eventually'' and ``always.'' LTL formulae are used only indirectly in LOF, as they are converted into automata that the algorithm uses directly. We chose to use LTL to represent rules because LTL corresponds closely to natural language and has proven to be a more natural way of expressing tasks and rules for engineers than designing FSAs by hand \citep{kansou2019converting}. Formulae $\phi$
have the syntax grammar
\begin{align*}
\phi := p \;|\; \neg \phi \;|\; \phi_1 \vee \phi_2 \;|\; \bigcirc \phi \;|\; \phi_1 \until \phi_2
\end{align*}
where $p$ is a \textit{proposition} (a boolean-valued truth statement that can correspond to objects or events in the world), $\neg$ is negation, $\vee$ is disjunction, $\bigcirc$ is ``next'', and $\until$ is ``until''. The derived rules are conjunction $(\wedge)$, implication $(\implies)$, equivalence $(\leftrightarrow)$, ``eventually'' ($\diamondsuit \phi \equiv \texttt{True} \until \phi$) and ``always'' ($\Box \phi \equiv \neg \diamondsuit \neg \phi$) \citep{Baier08}. $\phi_1 \until \phi_2$ means that $\phi_1$ is true until $\phi_2$ is true, $\diamondsuit \phi$ means that there is a time where $\phi$ is true and $\Box \phi$ means that $\phi$ is always true.
\textbf{The Options Framework:} The options framework is a framework for defining and solving semi-Markov Decision Processes (SMDPs) with a type of macro-action called an option \citep{sutton1999between}.
The inclusion of options in an MDP problem turns it into an SMDP problem, because actions are dependent not just on the previous state but also on the identity of the currently active option, which could have been initiated many time steps before the current time.
An option $o$ is a variable-length sequence of actions defined as $o = (\mathcal{I}, \pi, \beta, R_o(s), T_o(s' \vert s))$. $\mathcal{I} \subseteq \mathcal{S}$ is the initiation set of the option. $\pi : \mathcal{S} \times \mathcal{A} \rightarrow [0, 1]$ is the policy of the option. $\beta : \mathcal{S} \rightarrow [0, 1]$ is the termination condition. $R_o(s)$ is the reward model of the option. $T_o(s' \vert s)$ is the transition model. A major challenge in option learning is that, in general, the number of time steps before the option terminates, $k$, is a random variable. With this in mind, $R_o(s)$ is defined as the expected cumulative reward of option $o$ given that the option is initiated in state $s$ at time $t$ and ends after $k$ time steps. Letting $r_t$ be the reward received by the agent at $t$ time steps from the beginning of the option,
\begin{equation}
\label{eq:option-reward}
R_o(s) = \mathbb{E} \big[ r_{1} + \gamma r_{2} + \dots \gamma^{k-1}r_{k} \big]
\end{equation}
$T_o(s' \vert s)$ is the combined probability $p_o(s', k)$ that option $o$ will terminate at state $s'$ after $k$ time steps:
\begin{equation}
\label{eq:option-transitions}
T_o(s' \vert s) = \sum_{k=1}^\infty p_o(s', k)\gamma^k
\end{equation}
A crucial benefit of using options is that they can be composed in arbitrary ways. In the next section, we describe how LOF composes them to satisfy logical specifications.
\section{Discussion and Conclusion}
In this work, we claim that LOF has a unique combination of three properties: satisfaction, optimality, and composability. We state and prove the conditions for satisfaction and optimality in Sec.~\ref{sec:proofs}. The experimental results confirm our claims while also pointing out some weaknesses. \texttt{LOF-VI} achieves optimal or near-optimal policies and trains an order of magnitude faster than the existing work most similar to it, \texttt{RM}. However, the optimality condition that each subgoal be associated with one state cannot be met for continuous domains, and therefore \texttt{RM} eventually outperforms \texttt{LOF-VI}. But even when optimality is not guaranteed, \texttt{LOF-VI} is hierarchically optimal, which is why it outperforms \texttt{Greedy} in the composite and OR tasks.
Next, the composability experiments show that \texttt{LOF-VI} can compose its learned options to perform new tasks in about 10-50 iterations. Although \texttt{Greedy} requires no retraining steps, it is a tiny fraction of the tens of thousands of steps required to learn the original policy.
Lastly, we have shown that LOF learns policies efficiently, and that it can be used with a variety of domains and policy-learning algorithms.
\section{Experiments \& Results}\label{sec:experiments}
\textbf{Experiments:} We performed experiments to demonstrate satisfaction and composability. For satisfaction, we measure cumulative reward over training steps. Cumulative reward is a proxy for satisfaction, as the environments can only achieve the maximum reward when they satisfy their tasks. For the composability experiments, we take the trained options and record how many meta-policy retraining steps it takes to learn an optimal meta-policy for a new task.
\textbf{Environments:} We measure the performance of LOF on three environments. The first environment is a discrete gridworld (Fig.~\ref{fig:discrete-domain}) called the ``delivery domain,'' as it can represent a delivery truck delivering packages to three locations ($a$, $b$, $c$) and having a home base $h$. There are also obstacles $o$ (the black squares). The second environment is called the reacher domain, from OpenAI Gym (Fig.~\ref{fig:continuous-domain}). It is a two-link arm that has continuous state and action spaces. There are four subgoals represented by colored balls: red $r$, green $g$, blue $b$, and yellow $y$. The third environment is called the pick-and-place domain, and it is a continuous 3D environment with a robotic Panda arm from CoppeliaSim and PyRep \cite{james2019pyrep}. It is inspired by the lunchbox-packing experiments of \citet{araki2019learning} in which subgoals $r$, $g$, and $b$ are food items that must be packed into lunchbox $y$. All environments also have an event proposition called $can$, which represents when the need to fulfill part of a task is cancelled.
\textbf{Tasks:} We test satisfaction and composability on four tasks. The first task is a ``sequential'' task. For the delivery domain, the LTL formula is $\diamondsuit(a \land \diamondsuit(b \land \diamondsuit(c \land \diamondsuit h))) \land \Box \neg o$ -- ``deliver package $a$, then $b$, then $c$, and then return to home $h$. And always avoid obstacles.'' The next task is the ``IF'' task (equivalent to the task shown in Fig.~\ref{fig:demo-ltl-fsa}): $(\diamondsuit (c \land \diamondsuit a) \land \Box \neg can) \lor (\diamondsuit c \land \diamondsuit can) \land \Box \neg o$ -- ``deliver package $c$, and then $a$, unless $a$ gets cancelled. And always avoid obstacles''. We call the third task the ``OR'' task, $\diamondsuit ((a \lor b) \land \diamondsuit c) \land \Box \neg o$ -- ``deliver package $a$ or $b$, then $c$, and always avoid obstacles''. The ``composite'' task has elements of all three of the previous tasks: $(\diamondsuit((a \lor b) \land \diamondsuit(c \land \diamondsuit h)) \land \Box \neg can) \lor (\diamondsuit((a \lor b) \land \diamondsuit h) \land \diamondsuit can) \land \Box \neg o$. ``Deliver package $a$ or $b$, and then $c$, unless $c$ gets cancelled, and then return to home $h$. And always avoid obstacles''. The tasks for the reacher and pick-and-place environments are equivalent, except that there are no obstacles for the reacher and arm to avoid.
The sequential task is meant to show that planning is efficient and effective even for long-time horizon tasks. The ``IF'' task shows that the agent's policy can respond to event propositions, such as being alerted that a delivery is cancelled. The ``OR'' task is meant to demonstrate the optimality of our algorithm versus a greedy algorithm, as discussed in Fig.~\ref{fig:lof-vs-greedy}. Lastly, the composite task shows that learning and planning are efficient and effective even for complex tasks.
\textbf{Baselines:} We test four baselines against our algorithm. Our algorithm is \texttt{LOF-VI}, short for ``Logical Options Framework with Value Iteration,'' because it uses value iteration for high-level planning. \texttt{LOF-QL} uses Q-learning instead (details are in App.~\ref{sec:appendix-lof-ql}).
Unlike \texttt{LOF-VI}, \texttt{LOF-QL} does not need explicit knowledge of $T_F$, the FSA transition function.
\texttt{Greedy} is a naive implementation of task satisfaction; it uses its knowledge of the FSA to select the next subgoal with the lowest cost to attain.
This leaves it vulnerable to choosing suboptimal paths through the FSA, as shown in Fig.~\ref{fig:lof-vs-greedy}.
\texttt{Flat Options} uses the options framework with no knowledge of the FSA. Its SMDP formulation is not hierarchical -- the state space and transition function do not contain high-level states $\mathcal{F}$ or transition function $T_F$.
The last baseline is \texttt{RM}, short for Reward Machines \citep{icarte2018using}.
Whereas LOF learn options to accomplish subgoals, \texttt{RM} learns sub-policies for every FSA state. App.~\ref{sec:appendix-lof-vs-rm} discusses the differences between \texttt{RM} and LOF in detail.
\begin{figure}
\centering
\includegraphics[width=0.75\columnwidth]{figures/lof_vs_greedy2-cropped.pdf}
\caption{In this environment, the agent must either pick up the kid or go grocery shopping, and then go home (the ``OR'' task). Starting at $S0$, the greedy algorithm picks the next step in the FSA with the lowest cost (picking up the kid), which leads to a higher overall cost. LOF finds the optimal path through the FSA.}
\label{fig:lof-vs-greedy}
\end{figure}
\begin{figure*}[!th]
\centering
\begin{subfigure}[b]{0.3\textwidth}
\centering
\includegraphics[width=3cm]{figures/delivery_environment-cropped.pdf}
\caption{Delivery domain.}
\label{fig:discrete-domain}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/satisfaction/results_averaged_over_tasks.png}
\caption{Satisfaction performance.}
\label{fig:discrete-ave-perf}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/discrete/composability/results_averaged_over_tasks.png}
\caption{Composability performance.}
\label{fig:discrete-composability}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/reacher_environment-cropped.pdf}
\caption{Reacher domain.}
\label{fig:continuous-domain}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/satisfaction/results_averaged_over_tasks.png}
\caption{Satisfaction performance.}
\label{fig:continuous-ave-perf}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/continuous/composability/results_averaged_over_tasks.png}
\caption{Composability performance.}
\label{fig:continuous-composability}
\end{subfigure}
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=3.4cm]{figures/arm_domain-cropped.pdf}
\caption{Pick-and-place domain.}
\label{fig:arm-domain}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/satisfaction/results_averaged_over_tasks.png}
\caption{Satisfaction performance.}
\label{fig:arm-ave-perf}
\end{subfigure} \hfill
\begin{subfigure}[b]{.3\textwidth}
\centering
\includegraphics[width=4cm]{figures/arm/composability/results_averaged_over_tasks.png}
\caption{Composability performance.}
\label{fig:arm-composability}
\end{subfigure}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[height=0.6cm]{figures/discrete/satisfaction/legend.png}
\label{fig:satisfaction-legend}
\end{subfigure}
\caption{Performance on the satisfaction and composability experiments, averaged over all tasks. Note that \texttt{LOF-VI} composes new meta-policies in just 10-50 retraining steps. The first row is the delivery domain, the second row is the reacher domain, and the third row is the pick-and-place domain.
All results, including \texttt{RM} performance on the reacher and pick-and-place domains, are in App.~\ref{sec:appendix-results}.}
\label{fig:satisfaction-experiments}
\end{figure*}
\textbf{Implementation:} For the delivery domain, options were learned using Q-learning with an $\epsilon$-greedy exploration policy.
\texttt{RM} was learned using the Q-learning for Reward Machines (QRM) algorithm described in \citep{icarte2018using}. For the reacher and pick-and-place domains, options were learned by using proximal policy optimization (PPO) \citep{schulman2017proximal} to train goal-oriented policy and value functions, which were represented using $128 \times 128$ and $128 \times 128 \times 128$ fully connected neural networks, respectively. Deep-QRM was used to train \texttt{RM}. The implementation details are discussed more fully in App.~\ref{sec:appendix-experiments}.
\subsection{Results}
\textbf{Satisfaction:} Results for the satisfaction experiments, averaged over all four tasks, are shown in Figs.~\ref{fig:discrete-ave-perf},~\ref{fig:continuous-ave-perf}, and~\ref{fig:arm-ave-perf}. (Results on all tasks are in App.~\ref{sec:appendix-results}). As expected, \texttt{Flat Options} shows no ability to satisfy tasks, as it has no knowledge of the FSAs. \texttt{Greedy} trains as quickly as \texttt{LOF-VI} and \texttt{LOF-QL}, but its returns plateau before the others because it chooses suboptimal paths in the composite and OR tasks. The difference is small in the continuous domains but still present. \texttt{LOF-QL} achieves as high a return as \texttt{LOF-VI}, but it is less composable (discussed below). \texttt{RM} learns much more slowly than the other methods. This is because for \texttt{RM}, a reward is only given for reaching the goal state, whereas in the LOF-based methods, options are rewarded for reaching their subgoals, so during training LOF-based methods have a richer reward function than \texttt{RM}. For the continuous domains, \texttt{RM} takes an order of magnitude more steps to train, so we left it out of the figures for clarity (see App. Figs.~\ref{fig:appendix-continuous-satifaction-experiments-with-rm} and \ref{fig:appendix-arm-satifaction-experiments-with-rm}). However, in the continuous domains, \texttt{RM} eventually achieves a higher return than the LOF-based methods. This is because for those domains, we define the subgoals to be spherical regions rather than single states, violating one of the conditions for optimality. Therefore, for example, it is possible that the meta-policy does not take advantage of the dynamics of the arm to swing through the subgoals more efficiently. \texttt{RM} does not have this condition and learns a single policy that can take advantage of inter-subgoal dynamics to learn a more optimal policy.
\textbf{Composability:} The composability experiments were done on the three composable baselines, \texttt{LOF-VI}, \texttt{LOF-QL}, and \texttt{Greedy}. App.~\ref{sec:appendix-lof-vs-rm} discusses why \texttt{RM} is not composable. \texttt{Flat Options} is not composable because its formulation does not include the FSA $\mathcal{T}$. Therefore it is completely incapable of recognizing and adjusting to changes in the FSA. The composability results are shown in Figs.~\ref{fig:discrete-composability},~\ref{fig:continuous-composability}, and~\ref{fig:arm-composability}. \texttt{Greedy} requires no retraining steps to ``learn'' a meta-policy on a new FSA -- given its current FSA state, it simply chooses the next available FSA state that has the lowest cost to achieve. However, its meta-policy may be arbitrarily suboptimal. \texttt{LOF-QL} learns optimal (or in the continuous case, close-to-optimal) policies, but it takes $\sim$50-250 retraining steps, versus $\sim$10-50 for \texttt{LOF-VI}. Therefore \texttt{LOF-VI} strikes a balance between \texttt{Greedy} and \texttt{LOF-QL}, requiring far fewer steps than \texttt{LOF-QL} to retrain, and achieving better performance than \texttt{Greedy}.
\section{Introduction}
To operate in the real world, intelligent agents must be able to make long-term plans by reasoning over symbolic abstractions while also maintaining the ability to react to low-level stimuli in their environment \citep{zhang2020survey}. Many environments obey rules that can be represented as logical formulae; e.g., the rules a driver follows while driving, or a recipe a chef follows to cook a dish. Traditional motion and path planning techniques struggle to plan over these long-horizon tasks, but hierarchical approaches such as hierarchical reinforcement learning (HRL) can solve lengthy tasks by planning over both the high-level rules and the low-level environment. However, solving these problems involves trade-offs among multiple desirable properties, which we identify as \textit{satisfaction}, \textit{optimality}, and \textit{composability} (described below). Today's hierarchical planning algorithms lack at least one of these objectives. For example, Reward Machines \citep{icarte2018using} are satisfying and optimal, but not composable; the options framework \citep{sutton1999between} is composable and hierarchically optimal, but cannot satisfy specifications. An algorithm that achieves all three of these properties would be very powerful because it would enable a model learned on one set of rules to generalize to arbitrary rules. We introduce the \textit{Logical Options Framework}, which builds upon the options framework and aims to combine symbolic reasoning and low-level control to achieve satisfaction, optimality, and composability with as few compromises as possible. Furthermore, we demonstrate that models learned with our framework generalize to arbitrary sets of rules without any further learning, and we also show that our framework is compatible with arbitrary domains and planning algorithms, from discrete domains and value iteration to continuous domains and proximal policy optimization (PPO).
\textbf{Satisfaction:} An agent operating in an environment governed by rules must be able to satisfy the specified rules. Satisfaction is a concept from formal logic, in which the input to a logical formula causes the formula to evaluate to \texttt{True}. Logical formulae can encapsulate rules and tasks like the ones described in Fig.~\ref{fig:demo-domain}, such as ``pick up the groceries'' and ``do not drive into a lake''. In this paper, we state conditions under which our method is guaranteed to learn satisfying policies.
\begin{figure*}[!th]
\centering
\begin{subfigure}[b]{0.85\textwidth}
\centering
\includegraphics[width=0.99\textwidth]{figures/demo_nl-cropped.pdf}
\caption{These natural language instructions can be transformed into an FSA, shown in (b).}
\label{fig:demo-nl}
\end{subfigure}
\begin{subfigure}[t]{0.49\textwidth}
\centering
\includegraphics[width=.8\textwidth]{figures/demo_fsa_props-cropped.pdf}
\caption{The FSA representing the natural language instructions. The propositions are divided into ``subgoal'', ``safety'', and ``event.''}
\label{fig:demo-ltl-fsa}
\end{subfigure} \hfill
\begin{subfigure}[t]{.44\textwidth}
\centering
\includegraphics[width=.6\textwidth]{figures/demo_mdp-cropped.pdf}
\caption{The low-level MDP and corresponding policy that satisfies the instructions.}
\label{fig:demo-mdp}
\end{subfigure}
\caption{Many parents face this task after school ends -- who picks up the kid, and who gets groceries? The pictorial symbols represent propositions, which are true or false depending on the state of the environment.
The arrows in (c) represent sub-policies, and the colors of the arrows match the corresponding transition in the FSA. The boxed phone at the beginning of some of the arrows represents how these sub-policies can occur only after the agent receives a phone call.}
\label{fig:demo-domain}
\end{figure*}
\textbf{Optimality:} Optimality requires that the agent maximize its expected cumulative reward for each episode. In general, satisfaction can be achieved by rewarding the agent for satisfying the rules of the environment. In hierarchical planning there are several types of optimality, including hierarchical optimality (optimal with respect to the hierarchy) and optimality (optimal with respect to everything). We prove in this paper that our method is hierarchically optimal and, under certain conditions, optimal.
\textbf{Composability:} Our method is also composable -- once it has learned the low-level components of a task, the learned model can be rearranged to satisfy arbitrary tasks. More specifically, the rules of an environment can be factored into liveness and safety properties, which we discuss in Sec.~\ref{sec:lof}. The learned model has high-level actions called options that can be composed to satisfy new liveness properties. A shortcoming of many RL models is that they are not composable -- trained to solve one specific task, they are incapable of handling even small variations in the task structure. However, the real world is a dynamic and unpredictable place, so the ability to use a learned model to automatically reason over as-yet-unseen tasks is a crucial element of intelligence.
Fig.~\ref{fig:demo-domain} gives an example of how LOF works. The environment is a world with a grocery store, your (hypothetical) kid, your house, and some lakes, and in which you, the agent, are driving a car. The propositions are divided into ``subgoals'', representing events that can be achieved, such as going grocery shopping; ``safety'' propositions, representing events that you must avoid (driving into a lake); and ``event'' propositions, corresponding to events that you have no control over (receiving a phone call) (Fig.~\ref{fig:demo-ltl-fsa}). In this environment, you have to follow rules (Fig.~\ref{fig:demo-nl}). These rules can be converted into a logical formula, and from there into a finite state automaton (FSA) (Fig.~\ref{fig:demo-ltl-fsa}). LOF learns an option for each subgoal (illustrated by the arrows in Fig.~\ref{fig:demo-mdp}), and a meta-policy for choosing amongst the options to reach the goal state of the FSA. After learning, the options can be recombined to fulfill arbitrary tasks.
\subsection{Contributions}
This paper introduces the Logical Options Framework (LOF) and makes four contributions to the hierarchical reinforcement learning literature:
\begin{enumerate}
\item The definition of a hierarchical semi-Markov Decision Process (SMDP) that is the product of a logical FSA and a low-level environment MDP.
\item A planning algorithm for learning options and meta-policies for the SMDP that allows the options to be composed to solve new tasks with only 10-50 retraining steps and no additional samples from the environment.
\item Conditions and proofs for satisfaction and optimality.
\item Experiments on a discrete delivery domain, a continuous 2D reacher domain, and a continuous 3D pick-and-place domain on four tasks demonstrating satisfaction, optimality, and composability.
\end{enumerate}
\section{Logical Options Framework}\label{sec:lof}
Here is a brief overview of how we will present our formulation of LOF:
\begin{enumerate}
\item The LTL formula is decomposed into liveness and safety properties. The liveness property defines the task specification and the safety property defines the costs for violating rules.
\item The propositions are divided into subgoals, safety propositions, and event propositions. Each subgoal is associated with its own option, whose goal is to achieve that subgoal. Safety propositions are used to define rules. Event propositions serve as control flow variables that affect the task.
\item We define an SMDP that is the product of a low-level MDP and a high-level logical FSA.
\item We define the logical options.
\item We present an algorithm for finding the hierarchically optimal policy on the SMDP.
\item We state conditions under which satisfaction of the LTL specification is guaranteed, and we prove that the planning algorithm converges to an optimal policy by showing that the hierarchically optimal SMDP policy is the same as the optimal MDP policy.
\end{enumerate}
\textbf{The Logic Formula:} LTL formulae can be translated into B\"uchi automata using automatic translation tools such as SPOT \citep{Duret16}. All B\"uchi automata can be decomposed into liveness and safety properties \citep{alpern1987recognizing}. We assume here that the LTL formula itself can be divided into liveness and safety formulae, $\phi = \phi_{liveness} \land \phi_{safety}$. For the case where the LTL formula cannot be factored, see App.~\ref{sec:appendix-lof}.
The liveness property describes ``things that must happen'' to satisfy the LTL formula. It is a task specification and is used in planning to determine which subgoals the agent must achieve. The safety property describes ``things that can never happen'' and is used to define costs for violating the rules. In LOF, the liveness property is written using a finite-trace subset of LTL called syntactically co-safe LTL \citep{bhatia2010sampling}, in which $\Box$ (``always'') is not allowed and $\bigcirc$, $\until$, and $\diamondsuit$ are only used in positive normal form. This way, the liveness property can be satisfied by finite sequences of propositions, so the property can be represented as an FSA.
\textbf{Propositions:} Propositions are boolean-valued truth statements corresponding to goals, objects, and events in the environment.
We distinguish between three types of propositions: subgoals $\mathcal{P}_G$, safety propositions $\mathcal{P}_S$, and event propositions $\mathcal{P}_E$.
Subgoals must be achieved in order to satisfy the liveness property. They are associated with goals such as ``the agent is at the grocery store''. They only appear in $\phi_{liveness}$. Each subgoal may only be associated with one state. Note that in general, it may be impossible to avoid having subgoals appear in $\phi_{safety}$. App.~\ref{sec:appendix-lof} describes how to deal with this scenario.
Safety propositions are propositions that the agent must avoid -- for example, driving into a lake. They only appear in $\phi_{safety}$.
Event propositions are not goals, but they can affect the task specification -- for example, whether or not a phone call is received. They may occur in $\phi_{liveness}$, and, with extensions described in App.~\ref{sec:appendix-lof}, in $\phi_{safety}$. In the fully observable setting, event propositions are somewhat trivial because the agent knows exactly when/if the event will occur, but in the partially observable setting, they enable complex control flow.
Our optimality guarantees only apply in the fully observable setting; however, LOF's properties of satisfaction and composability still apply in the partially observable setting. The goal state of the liveness FSA must be reachable from every other state using only subgoals. This means that no matter what event propositions occur, it must be possible for the agent to satisfy the liveness property.
$T_{P_G} : \mathcal{S} \rightarrow 2^{\mathcal{P}_G}$ and $T_{P_S} : \mathcal{S} \rightarrow 2^{\mathcal{P}_S}$ relate states to the subgoal and safety propositions that are true at that state. $T_{P_E} : 2^{\mathcal{P}_E} \rightarrow \{0, 1\}$ assigns truth labels to the event propositions.
\begin{algorithm}[!t]
\caption{Learning and Planning with Logical Options}\label{alg:lof}
\begin{algorithmic}[1]
\STATE \textbf{Given:} \par
Propositions $\mathcal{P}$ partitioned into subgoals $\mathcal{P}_G$, safety propositions $\mathcal{P}_S$, and event propositions $\mathcal{P}_E$ \par
Logical FSA $\mathcal{T} = (\mathcal{F}, \mathcal{P}_G \times \mathcal{P}_E, T_F, R_F, f_0, f_g)$ derived from $\phi_{liveness}$ \par
Low-level MDP $\mathcal{E} = (\mathcal{S}, \mathcal{A}, R_\mathcal{E}, T_E, \gamma)$, where $R_\mathcal{E}(s, a) = R_E(s, a) + R_S(T_{P_S}(s))$ combines the environment and safety rewards \par
Proposition labeling functions $T_{P_G} : \mathcal{S} \rightarrow 2^{\mathcal{P}_G}$, $T_{P_S} : \mathcal{S} \rightarrow 2^{\mathcal{P}_S}$, and $T_{P_E} : 2^{\mathcal{P}_E} \rightarrow \{0, 1\}$ \par
\STATE \textbf{To learn:}
\STATE Set of options $\mathcal{O}$, one for each subgoal $p \in \mathcal{P}_G$
\STATE Meta-policy $\mu(f, s, o)$, $Q(f, s, o)$, and $V(f, s)$
\STATE \textbf{Learn logical options:}\label{alg:logical-options1}
\FOR{ $p \in \mathcal{P}_G$}
\STATE Learn an option that achieves $p$, \par $o_p = (\mathcal{I}_{o_p}, \pi_{o_p}, \beta_{o_p}, R_{o_p}(s), T_{o_p}(s' \vert s))$
\STATE $ \mathcal{I}_{o_p} = \mathcal{S} $
\STATE $ \beta_{o_p} = \begin{cases} 1 &\quad\text{if } p \in T_{P_G}(s)\\
0 &\quad\text{otherwise} \\
\end{cases} $
\STATE $\pi_{o_p} = $ optimal policy on $\mathcal{E}$ with rollouts terminating when $p \in T_{P_G}(s)$
\STATE $T_{o_p}(s' \vert s) = \begin{cases} \mathbb{E} \gamma^k &\quad \text{\parbox{11em}{if $p \in T_{P_G}(s')$; $k$ is number of time steps to reach $p$}}\\
0 &\quad\text{otherwise} \\
\end{cases}$\label{alg:lof-transition-model}
\STATE $R_{o_p}(s) = \mathbb{E} [ R_\mathcal{E}(s, a_{1}) + \gamma R_\mathcal{E}(s_{1}, a_{2}) + \dots$ \par
\quad\quad\quad\quad\;\;\;$ + \gamma^{k-1} R_\mathcal{E}(s_{k-1}, a_{k}) ]$ \label{alg:logical-options2}
\ENDFOR
\STATE \textbf{Find a meta-policy $\mu$ over the options:} \label{alg:lof-metapolicy}
\STATE Initialize $Q : \mathcal{F} \times \mathcal{S} \times \mathcal{O} \rightarrow \mathbb{R}$, $V : \mathcal{F} \times \mathcal{S} \rightarrow \mathbb{R}$ to $0$
\FOR {$(k, f, s) \in [1, \dots, n] \times \mathcal{F} \times \mathcal{S}$}
\FOR{$o \in \mathcal{O}$}
\STATE $Q_k(f, s, o) \leftarrow R_F(f)R_o(s) +$ \par $ \sum\limits_{f' \in \mathcal{F}} \sum\limits_{\bar{p}_e \in 2^{\mathcal{P}_E}} \sum\limits_{s' \in \mathcal{S}} T_F(f' \vert f, T_P(s'), \bar{p}_e) T_{P_E}(\bar{p}_e)$ \par
\quad\quad\quad\quad\quad\quad\quad$T_o(s' \vert s) V_{k-1}(f', s')$
\ENDFOR
\STATE $V_k(f, s) \leftarrow \max\limits_{o \in \mathcal{O}} Q_k(f, s, o)$
\ENDFOR
\STATE $\mu(f, s, o) = \argmax\limits_{o \in \mathcal{O}} Q(f, s, o)$
\STATE \textbf{Return:} Options $\mathcal{O}$, meta-policy $\mu(f, s, o)$ and Q- and value functions $Q(f, s, o), V(f, s)$
\end{algorithmic}
\end{algorithm}
\textbf{Hierarchical SMDP:} LOF defines a hierarchical semi-Markov Decision Process (SMDP), learns the options, and plans over them. The high level of the SMDP is an FSA specified with LTL. The low level is an environment MDP.
We assume that the LTL specification $\phi$ can be decomposed into a liveness property $\phi_{liveness}$ and a safety property $\phi_{safety}$. The propositions $\mathcal{P}$ are the union of the subgoals $\mathcal{P}_G$, safety propositions $\mathcal{P}_S$, and event propositions $\mathcal{P}_E$. We assume that the liveness property can be translated into an FSA $\mathcal{T} = (\mathcal{F}, \mathcal{P}, T_F, R_F, f_0, f_g)$. $\mathcal{F}$ is the set of automaton states; $\mathcal{P}$ is the set of propositions; $T_F$ is the transition function relating the current state and proposition to the next state, $T_F: \mathcal{F} \times \mathcal{P} \times \mathcal{F} \rightarrow [0, 1]$. In practice, $T_F$ is deterministic despite our use of probabilistic notation. We assume that there is a single initial state $f_0$ and final state $f_g$, and that the goal state $f_g$ is reachable from every state $f \in \mathcal{F}$ using only subgoals. The reward function assigns a reward to every FSA state, $R_F : \mathcal{F} \rightarrow \mathbb{R}$. In our experiments, the safety property takes the form $\bigwedge_{p_s \in \mathcal{P}_S} \Box \neg p_s$, which implies that no safety proposition is allowed, and that they have associated costs, $R_S : 2^{\mathcal{P}_S} \rightarrow \mathbb{R}$. $\phi_{safety}$ is not limited to this form; App.~\ref{sec:appendix-lof} covers the general case.
There is a low-level environment MDP $\mathcal{E} = (\mathcal{S}, \mathcal{A}, R_\mathcal{E}, T_E, \gamma)$. $\mathcal{S}$ is the state space and $\mathcal{A}$ is the action space. They can be discrete or continuous. $R_E: \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}$ is a low-level reward function that characterizes, for example, distance or actuation costs. $R_\mathcal{E}$ is a combination of the safety reward function $R_S$ and $R_E$, e.g. $R_\mathcal{E}(s, a) = R_E(s, a) + R_S(T_{P_S}(s))$.
The transition function of the environment is $T_E : \mathcal{S} \times \mathcal{A} \times \mathcal{S} \rightarrow [0 ,1]$.
From these parts we define a hierarchical SMDP $\mathcal{M} = (\mathcal{S} \times \mathcal{F}, \mathcal{A}, \mathcal{P}, \mathcal{O}, T_E \times T_P \times T_F, R_{SMDP}, \gamma)$. The hierarchical state space contains two elements: low-level states $\mathcal{S}$ and FSA states $\mathcal{F}$. The action space is $\mathcal{A}$. The set of propositions is $\mathcal{P}$. The set of options (one option associated with each subgoal in $\mathcal{P}_G$) is $\mathcal{O}$. The transition function consists of the low-level environment transitions $T_E$ and the FSA transitions $T_F$. $T_P = T_{P_G} \times T_{P_S} \times T_{P_E}$. We call $T_P$, relating states to propositions, a transition function because it determines when FSA transitions occur. The transitions are applied in the order $T_E$, $T_P$, $T_F$. The reward function $R_{SMDP}(f, s, o) = R_F(f)R_o(s)$, so $R_F(f)$ is a weighting on the option rewards. The SMDP has the same discount factor $\gamma$ as $\mathcal{E}$. Planning is done on the SMDP in two steps: first, the options $\mathcal{O}$ are learned over $\mathcal{E}$ using an appropriate policy-learning algorithm such as PPO or Reward Machines. Next, a meta-policy over the task specification $\mathcal{T}$ is found using the learned options and the reward function $R_{SMDP}$.
\textbf{Logical Options:} The first step of Alg.~\ref{alg:lof} is to learn the logical options. We associate every subgoal $p$ with an option $o_{p} = (\mathcal{I}_{o_p}, \pi_{o_p}, \beta_{o_p}, R_{o_p}, T_{o_p})$. These terms are defined starting at Alg.~\ref{alg:lof} line~\ref{alg:logical-options1}. Every $o_{p}$ has a policy $\pi_{o_p}$ whose goal is to reach the state $s_{p}$ where $p$ is true. Options are learned by training on the environment MDP $\mathcal{E}$ and terminating only when $s_{p}$ is reached. As we discuss in Sec.~\ref{sec:proofs}, under certain conditions the optimal option policy is guaranteed to always terminate at the subgoal. This allows us to simplify the transition model of Eq.~\ref{eq:option-transitions} to the form in Alg.~\ref{alg:lof} line~\ref{alg:lof-transition-model}. In the experiments, we further simplify this expression by setting $\gamma = 1$.
\textbf{Logical Value Iteration:} After finding the logical options, the next step is to find a meta-policy for FSA $\mathcal{T}$ over the options (see Alg.~\ref{alg:lof} line~\ref{alg:lof-metapolicy}). Q- and value functions are found for the SMDP using the Bellman update equations:
\begin{align}
\begin{split}\label{eq:q-update}
Q_k(& f, s, o) \leftarrow R_F(f)R_o(s) + \sum_{f' \in \mathcal{F}} \sum_{\bar{p}_e \in 2^{\mathcal{P}_E}} \sum_{s' \in \mathcal{S}} \\
& T_F(f' \vert f, T_{P_G}(s'), \bar{p}_e) T_{P_E}(\bar{p}_e ) T_o(s' \vert s) V_{k-1}(f', s')
\end{split}\\
V_k&( f, s) \leftarrow \max_{o \in \mathcal{O}} Q_k(f, s, o) \label{eq:v-update}
\end{align}
Eq.~\ref{eq:q-update} differs from the generic equations for SMDP value iteration in that the transition function has two extra components, $\sum_{f' \in \mathcal{F}} T_F(f' \vert f, T_P(s'), \bar{p}_e)$ and $\sum_{\bar{p}_e \in 2^{\mathcal{P}_E}} T_{P_E}(\bar{p}_e)$. The equations are derived from \citet{araki2019learning} and the fact that, on every step in the environment, three transitions are applied: the option transition $T_o$, the event proposition ``transition'' $T_{P_E}$, and the FSA transition $T_F$. Note that $R_o(s)$ and $T_o(s' \vert s)$ compress the consequences of choosing an option $o$ at a state $s$ from a multi-step trajectory into two real-valued numbers, allowing for more efficient planning.
\subsection{Conditions for Satisfaction and Optimality}\label{sec:proofs}
Here we give an overview of the proofs and necessary conditions for satisfaction and optimality. The full proofs and definitions are in App.~\ref{sec:appendix-proofs}.
First, we describe the condition for an optimal option to always reach its subgoal. Let $\pi'(s \vert s')$ be the optimal goal-conditioned policy for reaching a goal $s'$. If the optimal option policy equals the goal-conditioned policy for reaching the subgoal $s_g$, i.e. $\pi^*(s) = \pi_g(s \vert s_g)$, then the option will always reach the subgoal. This can be stated in terms of value functions: let $V^{\pi'}(s \vert s')$ be the expected return of $\pi'(s \vert s')$. If $V^{\pi_g}(s \vert s_g) > V^{\pi'}(s \vert s') \; \forall s, s' \neq s_g$, then $\pi^*(s) = \pi_g(s \vert s_g)$. This occurs for example if $-\infty < R_\mathcal{E}(s, a) < 0$ and if the episode terminates when the agent reaches $s_g$. Then $V^{\pi_g}$ is a bounded negative number, and $V^{\pi'}$ for all other states is $-\infty$. We show that if every option is guaranteed to achieve its subgoal, then there must exist at least one sequence of options that satisfies the specification.
We then give the condition for the hierarchically optimal meta-policy $\mu^*(s)$ to always achieve the FSA goal state $f_g$. In our context, hierarchical optimality means that the meta-policy is optimal over the available options.
Let $\mu'(f, s \vert f')$ be the hierarchically optimal goal-conditioned meta-policy for reaching FSA state $f'$.
If the hierarchically optimal meta-policy equals the goal-conditioned meta-policy for reaching the FSA goal state $f_g$, i.e. $\mu^*(f, s) = \mu_g(f, s \vert f_g)$, then $\mu^*(f, s)$ will always reach $f_g$. In terms of value functions: let $V^{\mu'}(f, s \vert f')$ be the expected return for $\mu'$. If $V^{\mu_g}(f, s \vert f_g) > V^{\mu'}(f, s \vert f') \forall f, s, f' \neq f_g$, then $\mu^* = \mu_g$. This occurs if all FSA rewards $R_F(f) > 0$, all environment rewards $-\infty < R_\mathcal{E}(s, a) < 0$, and the episode only terminates when the agent reaches $f_g$. Then $V^{\mu_g}$ is a bounded negative number, and $V^{\mu'}$ for all other states is $-\infty$. Because LOF uses the Bellman update equations to learn the meta-policy, the LOF meta-policy will converge to the hierarchically optimal meta-policy.
Consider the SMDP where planning is allowed over low-level actions, and let us call it the ``hierarchical MDP'' (HMDP) with optimal policy $\pi^*_{HMDP}$.
Our result is:
\begin{theorem}\label{theorem}
Given that the conditions for satisfaction and hierarchical optimality are met, the LOF hierarchically optimal meta-policy $\mu_g$ with optimal option sub-policies $\pi_g$ has the same expected returns as the optimal policy $\pi^*_{HMDP}$ and satisfies the task specification.
\end{theorem}
\subsection{Composability}
The results in Sec.~\ref{sec:proofs} guarantee that LOF's learned model can be composed to satisfy new tasks. Furthermore, the composed policy has the same properties as the original policy -- satisfaction and optimality. LOF's possession of composability along with satisfaction and optimality derives from two facts: 1) Options are inherently composable because they can be executed in any order. 2) If the conditions of Thm.~\ref{theorem} are met, LOF is guaranteed to find a (hierarchically) optimal policy over the options that will satisfy any liveness property that uses subgoals associated with the options. The composability of LOF distinguishes it from other algorithms that can achieve satisfaction and optimality.
\section*{Acknowledgments}
Toyota Research Institute provided funds to support this work. This material is also supported by the Office of Naval Research grant ONR N00014-18-1-2830, and the Under Secretary of Defense for Research and Engineering under Air Force Contract No. FA8702-15-D-0001. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Under Secretary of Defense for Research and Engineering.
\section{Related Work}
We distinguish our work from related work in HRL by its possession of three desirable properties -- composability, satisfaction, and optimality. Most other works possess two of these properties at the cost of the other.
\textbf{Not Composable:} The previous work most similar to ours is \citet{icarte2018using, icarte2019learning}, which introduces a method to solve tasks defined by automata called Reward Machines. Their method learns a sub-policy for every state of the automaton that achieves satisfaction and optimality. However, the learned sub-policies have limited composability because they end up learning a specific path through the automaton, and if the structure of the automaton is changed, there is no guarantee that the sub-policies will be able to satisfy the new automaton without re-training. By contrast, LOF learns a sub-policy for every subgoal, independent of the automaton, and therefore the sub-policies can be arranged to satisfy arbitrary tasks. Another similar work is Logical Value Iteration (LVI) \citep{araki2019learning, araki2020deep}. LVI defines a hierarchical MDP and value iteration equations that find satisfying and optimal policies; however, the algorithm is limited to discrete domains and has limited composability. A number of HRL algorithms use reward shaping to guide the agent through the states of an automaton \citep{li2017reinforcement, li2019formal, camacho2019ltl, hasanbeig2018logically, jothimurugan2019composable, shah2020planning, yuan2019modular}. While these algorithms can guarantee satisfaction and sometimes optimality, they cannot be composed because their policies are not hierarchical. Another approach is to use a symbolic planner to find a satisfying sequence of tasks and use an RL agent to learn and execute that sequence of tasks \citep{gordon2019should, illanes2020symbolic, lyu2019sdrl}. However, the meta-controllers of \citet{gordon2019should} and \citet{lyu2019sdrl} are not composable as they are trained together with the low-level controllers. Although the work of \citet{illanes2020symbolic} is amenable to transfer learning, it is not composable. \citet{paxton2017combining, mason2017assured} use logical constraints to guide exploration, and while these approaches are also satisfying and optimal, they are not composable as the agent is trained for a specific set of rules.
LOF is composable unlike the above methods because it has a hierarchical action space with high-level options. Once the options are learned, they can be composed arbitrarily.
\textbf{Not Satisfying:} Most hierarchical frameworks cannot satisfy tasks. Instead, they focus on using state and action abstractions to make learning more efficient \citep{dietterich2000hierarchical, dayan1993feudal, parr1998reinforcement, diuk2008object, oh2019planning}.
The options framework \citep{sutton1999between} stands out because of its composability and its guarantee of hierarchical optimality, which is why we based our work off of it. There is also a class of HRL algorithms
that builds on the idea of goal-oriented policies that can navigate to nearby subgoals \citep{eysenbach2019search, ghosh2018learning, faust2018prm}. By sampling sequences of subgoals and using a goal-oriented policy to navigate between them, these algorithms can travel much longer distances than a policy can travel on its own. Although these algorithms are ``composable'' in that they can navigate to far-away goals without further training, they are not able to solve tasks.
\citet{andreas2017modular} present an algorithm for solving simple policy ``sketches'' which is also composable; however, sketches are considerably less expressive than automata and linear temporal logic, which we use.
Unlike the above methods, LOF is satisfying because it has a hierarchical state space with low-level MDP states and high-level FSA states. Therefore LOF can satisfy tasks by learning policies that reach the FSA goal state.
\textbf{Not Optimal:} In HRL, there are at least three types of optimality -- hierarchical, recursive, and overall. As defined in \citet{dietterich2000hierarchical}, the hierarchically optimal policy is the optimal policy given the constraints of the hierarchy, and recursive optimality is when a policy is optimal given the policies of its children. For example, the options framework is hierarchically optimal, while MAXQ and abstract MDPs \citep{gopalan2017planning} are recursively optimal. The method described in \citet{kuo2020encoding} is fully composable, but not optimal as it uses a recurrent neural network to generate a sequence of high-level actions and is therefore not guaranteed to find optimal policies. LOF is hierarchically optimal because it finds an optimal meta-policy over the high-level options, and as we state in the paper, there are also conditions under which the overall policy is optimal.
|
\section{Introduction}
GitHub is one of the leading platforms of open source software projects \cite{gousios2014lean}.
GitHub facilitates several developers to collaborate and contribute to projects by performing several actions such as updating projects using commits, viewing changes made to projects through commit history \cite{tsay2014influence}, logging issues or defects through issue reports and contributing to other projects through pull requests and so on \cite{jiang2017and, jiang2017understanding}. Open source projects originating from well known organizations such as \textit{vscode} from Microsoft, \textit{react-native} from Facebook and \textit{tensorflow} from Google, have more than 1K\footnote{\url{https://github.com/microsoft/vscode}}, 2K\footnote{\url{https://github.com/facebook/react-native}} and 2.8K\footnote{\url{https://github.com/tensorflow/tensorflow}} contributors respectively (as of January 2021). Developers also contribute to and reuse several other publicly available repositories on GitHub \cite{zagalsky2015emergence}. Documentation in the repositories facilitates developers to understand about a project and consequently helps them in deciding on projects they wish to contribute \cite{borges2018s}. Repositories contain \textit{Readme} files that provide information on the purpose, requirements, usage instructions and various other information about the repositories \cite{prana2019categorizing, perez2016ten}. Several other files of the repository such as \textit{License} files, \textit{UML} files and so on also present different types of repository information such as license permissions and design decisions \cite{hebig2016quest, vendome2015license}.
Though many developers are interested contribute to GitHub repositories, they face multiple hurdles during this process, resulting in reduced motivation towards contributing to repositories \cite{mendez2018open, steinmacher2016overcoming}. The existing insufficient and scattered documentation on these repositories makes it difficult for developers to understand about the repositories, consequently reducing the advantages of huge number of contributions \cite{aghajani2019software}. Developers tend to visit other artifacts in the repositories to better understand about the repository, which is effort-intensive \cite{fronchetti2019attracts, robles2006beyond}. We believe that contribution efforts from wide range of developers could be well leveraged if documentation is improved and consolidated.
Researchers have explored multiple dimensions of understanding, usage and ways to improve software documentation, as it plays a significant role in performing various tasks. These tasks include software development, testing, integration, maintenance and so on \cite{garousi2015usage, mahmood2011industrial}, but, the documentation is insufficient in majority of the projects \cite{aghajani2020software, aghajani2019software}.
For example, documentation helps in improving software development and maintenance by providing necessary information to users in different roles such as system integrators, quality analysts, developers and so on \cite{kipyegen2013importance}. Good documentation helps in re-engineering existing software during maintenance and migration \cite{de2005study, aghajani2020software}. It has also been observed that developers regard documentation as an important aspect, even in projects with faster release cycles such as agile projects \cite{stettina2011necessary}.
Considering the wide usage of GitHub and activities performed by developers on GitHub, documentation of software repositories that are hosted on GitHub also plays a major role in various stages of the project such as development, deployment, maintenance and so on. It has been observed that projects having better popularity tend to have better documentation \cite{cosentino2017systematic}. Developers of popular projects tend to improve documentation by regular updates, to provide better insights to users \cite{aggarwal2014co}. Also, documentation is one of the factors considered by developers before contributing to repositories on GitHub \cite{borges2018s}. Currently, documentation is spread across multiple files such as \textit{source code}, \textit{design diagrams}, \textit{readme files}, \textit{license files} and so on \cite{hebig2016quest, prana2019categorizing, ma2018automatic}. We hypothesize that other artifacts could also provide valuable information about a repository, that could be considered as documentation, without documentation being explicitly present. Researchers have also observed that developers spend considerable effort on multiple artifacts, other than source code, in end-user oriented projects, supporting our idea of considering documentation from multiple sources \cite{robles2006beyond, fronchetti2019attracts}. Currently, sources of documentation are diversified in a repository. The documentation present in these sources is also unclear. \textit{Unifying the information from multiple sources could help in enhancement of documentation, and eventually reduce effort for developers, thus motivating them to contribute to GitHub repositories.} Though there are studies to identify different types of documentation present in software repositories \cite{prana2019categorizing, hebig2016quest, chimalakonda2020software}, we are not aware of any work that integrates documentation from multiple artifacts, thus, motivating our work. Hence, in this paper, we aim to identify and gather different artifacts of software repositories, that could contain information relevant to documentation. The main contributions of this paper are:
\begin{itemize}
\item A taxonomy of documentation sources in GitHub Repositories based on card-sorting approach and developer-based interviews with 20 developers.
\item An empirical analysis on 950 GitHub repositories, of four programming languages, C++, C\#, Python and Java, to understand the types of documentation that could be extracted from multiple software artifacts.
\item Results of the empirical study, along with the percentage of available information that could contribute to documentation, in each of the software artifacts that are identified as potential sources of documentation. The results of the study and dataset used for the study can be accessed here\footnote{Results \& Dataset - \url{https://osf.io/dfx9r/?view_only=0954174b44054893a3dfbe6cc7dd5db5}}.
\end{itemize}
\vspace{1mm} \fbox{\begin{minipage}{25em}
\textit{\textbf{Documentation Type}} - We define type of documentation based on content present in the documentation. For example, documentation that refers to reporting or fixing an error is considered as \textit{Error-related} documentation type.
\end{minipage}
}
\vspace{1mm}
\newline
\newline
\fbox{\begin{minipage}{25em}
\textit{\textbf{Documentation Source} - } We define documentation source as a software artifact of the project, capable of providing potentially useful information related to software documentation.
\end{minipage}
}
\section{Research Methodology}
In this paper, we aim to analyze contents in GitHub repositories from a software documentation perspective, specifically focusing on various software artifacts that could serve as sources of documentation.
We followed an approach that comprises of the following six phases.
\begin{itemize}
\begin{figure*}
\centering
\includegraphics[width = \linewidth]{tax.png}
\caption{Proposed Taxonomy of Documentation Sources}
\label{fig:taxonomy}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width = \linewidth]{typetax.png}
\caption{Proposed Taxonomy of Documentation Types}
\label{fig:typestax}
\end{figure*}
\item \textit{Research Questions Definition} - Four research questions aimed to identify possible sources of software documentation have been defined.
\item \textit{Card Sorting Approach and Developer Interviews} - For manual analysis, 60 of the 950 GitHub repositories, which include 15 top trending repositories from C++, C\#, Java and Python programming languages, were selected analyzed by three individuals using card sorting approach to narrow down onto potential categories of documentation and sources that could be leveraged in answering the research questions. Twenty professional software developers were interviewed to understand different types of documentation and sources as perceived by developers.
\item \textit{Data Extraction} - 950 public GitHub repositories, which include 233, 241, 253 and 223 repositories from C++, C\#, Java and Python programming languages respectively, were scraped and processed for further analysis.
\item \textit{Topic Modelling and Manual Inspection} - LDA topic modelling techniques were applied to 20\% of the scraped data to identify topics in the data. The topics identified are manually walked through to label the categories.
\item \textit{Automated Analysis} - Models to analyze extracted data against research questions are developed
\item \textit{Result Comprehension} - Research questions are answered as a result of automated analysis.
\end{itemize}
\subsection{Research Questions}
The main goal of this study is to propose a taxonomy of documentation sources in GitHub, by identifying different potential sources of software documentation in a GitHub repository, including the sources in which documentation is not explicitly specified. We conduct this study by answering the following research questions:
\begin{itemize}
\item \textbf{RQ1: What are the types of documentation present in GitHub repositories ?}
To answer this question, we employed card-sorting approach and developer interviews with 20 developers to understand different documentation types present in GitHub repositories, as perceived by GitHub users.
\textbf{Summary:} We observe that majority of the GitHub repositories comprise of six documentation types, of which at least three of them are not explicitly mentioned as documentation in the repositories.
\item \textbf{RQ2: What are the sources of documentation in GitHub repositories? }
We answer this question by identifying various sources that are perceived to contribute to different documentation types through developer interviews with 20 developers and card-sorting approach with 3 participants.
\textbf{Summary:} We observe that majority of the developers use six artifacts of GitHub Repositories, to understand about a repository, and further contribute to it. This also included \textit{issues} and \textit{pull requests}, which do not explicitly convey documentation related information.
\item \textbf{RQ3: What is the distribution of different documentation types in each of the identified sources?}
To answer this question, we categorize the information extracted from each source of 950 GitHub repositories into one of the documentation types identified, and calculate the percentage of each type in the sources.
\textbf{Summary:} Of the six documentation types identified, \textit{error-related} documentation is present to a larger extent in majority of the documentation sources.
\item \textbf{RQ4: What is the contribution of identified sources to different documentation types?}
To answer this question, we calculate the percentage contribution of a source by finding the ratio of amount of information present in a source with respect to the documentation type to the total amount of information of the documentation type, present across different sources.
\textbf{Summary:} Following a well-established notion, majority of the documentation was observed to be present in \textit{source-code} comments and \textit{textual documents}. However, \textit{commit logs} and \textit{issues} were also observed to contribute to documentation, with a minor variation in the percentage, in comparison to the \textit{textual} documentation.
\end{itemize}
\subsection{Card Sorting Approach and Developer Interviews}
We observe that card-sorting \cite{spencer2004card} approach has been used widely towards arriving at various taxonomies and classifications in the software engineering literature, such as classifying requirements change \cite{nurmuliani2004using} and research ideas\cite{lo2015practitioners}, which motivated us to employ card-sorting as an initial step towards arriving at different types of documentation and their sources.
We have downloaded 60 of the 950 GitHub repositories, such that they include 15 top trending repositories, from each of the four programming languages being considered, for manual analysis. Three individuals, comprising two researchers and one under-graduate student were assigned the task of identifying different types and sources of documentation present in the repositories. Initially, all the three individuals have explored \textit{readme files} of the repositories, considering them to be a basic and easily available source of documentation. We employed open card sorting approach, without pre-defining possible groups of documentation. In the first step, each of the three individuals came up with different number of groups (4, 6 and 10).
On further iterative discussions on labels of each of the groups, we observed similarities between the information content present among the 10 groups identified by the third individual, resulting in a set of 5 groups (T1 to T4 and T6 documentation types presented in Table \ref{tab:typedefs}). We further compared and discussed about the data content among other groups, identified by each individual and finally arrived at a common decision, with 6 possible types of documentation, accompanied by labels to these types, as mentioned in Table \ref{tab:typedefs}. As a next step, all the three individuals navigated through various artifacts within the repositories, and attempted to identify documentation in those artifacts. On identification of documentation, each artifact was labelled with the corresponding documentation type (identified in the previous step) it would be representing. While one individual narrowed down to five artifacts, the other two individuals pointed out six possible artifacts containing documentation. Each of the three individuals have explained the reasons for selecting respective number of artifacts, based on which, we finally decided to consider six different artifacts (mentioned in Fig. \ref{fig:taxonomy}) as part of the proposed taxonomy.
We later organised developer interviews with 20 professional developers to validate the types and sources of documentation obtained as a result of the card sorting approach. We approached 20 developers working in different organizations, with work experience ranging from 2 to 30 years. All the 20 developers interviewed belonged to organizations with more than 1000 employees and with a team size of at least 10. We selected developers with this profile to ensure that they are well-equipped with the collaborative project development and management. All the 18 developers with less than 10 years of experience mentioned that they actively contribute to open-source projects.
We asked them open-ended questions to understand their perception about types of documentation and the sources of documentation present on GitHub. 12 of the developers, with work experience ranging between 3 to 10 years explained that they find six documentation types (presented in Table \ref{tab:typedefs}) in the projects they interact with, on GitHub. 2 developers with work experience of 29 and 30 years pointed out that they find seven types of documentation in the projects, which included \textit{migration guidelines}, apart from the six types identified in Table \ref{tab:typedefs}. \textit{Migration guidelines} could however be considered as \textit{project-related} documentation, at a higher level of abstraction, as they specify details necessary in migrating the project. The rest of the developers with less than three years of experience have expressed that they have observed only three types of documentation (T1, T2 and T5 in Table \ref{tab:typedefs}). This varied insights from developers could also be due to varied abstractions of projects they are exposed to, in their respective organizations. We then discussed with them the documentation types identified previously, through card-sorting approach. 18 of the 20 developers have later agreed to the identified six documentation types, while the other two developers have suggested inclusion of one more category, which corresponds to \textit{migration guidelines} of the project.
Considering the majority responses, we decided to proceed further, with six documentation types. We further queried the developers to understand different artifacts on GitHub that they use to understand about a repository, which could be helpful, if collated as a single documentation file. All the 20 developers have pointed out six sources of documentation, as shown in Figure \ref{fig:taxonomy}, in GitHub repositories, and also mentioned that five (\textit{Source code}, \textit{textual documents}, \textit{commits}, \textit{pull requests} and \textit{issue reports}) of these six documentation sources, are frequently referred by them.
\begin{table}[]
\caption{Observed Documentation Types and corresponding definitions }
\vspace{-2mm}
\centering
\begin{tabular}{|l|l|l|}
\hline
\textbf{S.}&
\textbf{Types of} & \textbf{Definition}\\
\textbf{no.}& \textbf{Documentation} & \\
\hline
T1 & API-related & Documentation capable of providing information \\
&documentation& about APIs used in the project\\
\hline
T2 & File-related & Documentation capable of providing \\
&Documentation & file-level information such as updates made to \\
&&the files and dependencies of files\\
\hline
T3 & Project-related & Documentation capable of providing \\
&Documentation & project-level information such as installation \\
& & instructions, branches in projects, enhancements \\
& & to the projects and so on\\
\hline
T4 & License-related & Documentation capable of providing information \\
&Documentation & about licenses in the project
\\
\hline
T5 & Error/Bug-related & Documentation capable of providing information \\
&documentation& about errors or bugs encountered \\
& & in multiple files of project\\
\hline
T6 & Architecture- & Documentation capable of providing information \\
&related &about architecture of the project, such as \\
&Documentation&module interactions.\\
\hline
\end{tabular}
\label{tab:typedefs}
\vspace{-2mm}
\end{table}
Based on the card sorting approach and developer interviews, we propose six potential documentation sources, that includes 3 software artifacts which do not have documentation explicitly specified, as shown in Fig. \ref{fig:taxonomy}. Thus, we observe six types of documentation and six sources of documentation, as mentioned in Table \ref{tab:typedefs} and Fig. \ref{fig:taxonomy} respectively.
\newline
\newline
\fbox{\begin{minipage}{25em}
\textbf{RQ1:} We observed six different documentation types in GitHub Repositories. \textit{API-related}, \textit{file-related}, \textit{project-related}, \textit{license-related}, \textit{error/bug-related} and \textit{architecture-related} documentations are the six different documentation types identified.
\textbf{Insights-} Developers perceived \textit{file-related} and \textit{project-related} documentation types to be more useful in contributing to a project, among the six types.
\end{minipage}}
\vspace{1mm}
\newline
\newline
\fbox{\begin{minipage}{25em}
\textbf{RQ2:} We observed six different sources containing potential information related to documentation in GitHub repositories.\textit{ Source code, textual documents, commits, pull requests, issue reports and design diagrams} are the six artifacts, observed to be potential sources of documentation in GitHub repositories.
\textbf{Insights-} Apart from the widely accepted documentation sources - \textit{source code comments},\textit{design diagrams} and \textit{textual documents}, our interviews with developers revealed that \textit{pull requests}, \textit{commits} and \textit{issues} can also contribute to documentation.
\end{minipage}}
\subsection{Data Extraction}
We identified top-starred 300 GitHub repositories, written in each of the four different programming languages, C++, Python, C\# and Java. The choice of languages was based on random numbers generated from 1 to 10, and the languages corresponding to these random numbers among the top 10 active languages on GitHub\footnote{https://githut.info/}, as of 01
January 2021. We further filtered out forked repositories and other repositories with zero pull requests, to exclude tutorial-based repositories. This resulted in 233, 241, 253 and 223 repositories corresponding to C++, C\#, Java and Python programming languages respectively, leading to a total of 950 repositories. We analyzed each of the repositories to extract information from multiple documentation sources. We have manually walked through top 15 repositories in each of the programming languages to identify file formats for different documentation types. Based on the findings from developer interviews and card-sorting approach, we classified all files in the repositories into 5 categories - Textual Documentation, Images, Design Diagrams, Source Code and Others. The rules employed for this categorization are presented in the Table \ref{tab:fileclass}.
\begin{table}[]
\caption{File Classification rules}
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Category } & \textbf{File Format} \\
\hline
Textual & `.txt' , `.md' , `readme' , `license'\\
Images & `.png', `.jpg', `.jpeg'\\
Design Diagrams & `.xmi',`.uml'\\
Source Code & `.cpp', `.cs', `.py', `.java'\\
Others & Any other extensions\\
\hline
\end{tabular}
\label{tab:fileclass}
\end{table}
Files with extensions - {`.txt', `.md', } are categorized to \textit{textual} documentation.
All files with extensions - {`.xmi', `.uml'} are categorized to UML diagrams.
Files with {`.jpg',`.jpeg',`.png'} extensions are primarily categorized as Images, which could be categorized as design diagrams through further analysis. Files containing extension of programming languages considered, i.e., .cpp, .py, .java and .cs are classified as source code files. Though the repositories were extracted based on the programming language in which they are written, a majority of the repositories also contain files written in other programming languages.
Considering the vast number of programming languages supported by GitHub, it is difficult to analyze files of other programming languages, as programming constructs differ from one language to another. Hence, files with formats not belonging to the above four classes are categorized as \textit{Others}. Comments from files in the \textit{source code} category have been extracted. This exploration of repository files has extracted information that could contribute to documentation, from two documentation sources mentioned in the taxonomy, namely - \textit{textual documents} and \textit{source code files}. Also, \textit{design diagrams} for each of the repositories have been identified and information from these diagrams could be extracted using image processing techniques. In view of the technical effort and difficulty involved in choosing and applying appropriate image processing techniques, we consider data extraction from design diagrams to be out of the scope of this study.
For each of the repositories, other software artifacts, i.e., \textit{pull requests, issues} and \textit{commits}, that were identified as sources of documentation in the proposed taxonomy, not older than three years from January 01, 2021, were extracted using GitHub API. However, blank data was returned for 46 python repositories, leaving us with pull request data of only 177 of the 223 Python repositories considered. This blank data could be due to the large number of pull requests in these repositories, which could not be extracted due to rate limits of GitHub API, or those repositories with pull requests older than three years, from 01 January, 2021.
We have identified fields that are capable of containing documentation related information through manual inspection of extracted artifacts in 50 of the 950 repositories, selected randomly. These fields are presented in Table \ref{tab:fields}, and the data of these fields is stored as text documents. This accomplishes the task of extracting information useful for documentation from three sources of documentation, mentioned in the proposed taxonomy, namely - \textit{Pull Requests, Issues} and \textit{Commits}.
\begin{table}[]
\caption{Documentation Sources and fields considered}
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Documentation Source} & \textbf{Fields Identified}\\
\hline
Issues & title , body, comments\\
Pull Requests & title, body, comments\\
Commits & message, comments\\
\hline
\end{tabular}
\label{tab:fields}
\end{table}
The fields that contain textual data and contribute to documentation types are identified (as presented in Table \ref{tab:fields}) and corresponding text is stored for further analysis.
\subsection{Topic Modelling and Manual Inspection}
We observed that software artifacts that are identified as sources of documentation contain information that could support different documentation types. Hence, we analyze this information and categorize into different documentation types. Topic modelling technique can identify keywords, that belong to different number of topics specified, in a given document. This feature of topic modelling fits well with our requirement of categorizing text into different documentation types. We performed Latent Dirichlet Allocation (LDA) topic modelling technique to categorize the information. Though we have consolidated different types of documentation, all sources of documentation need not contribute to all types of documentation. Hence, we primarily tried to understand the optimal number of topics into which information from each of the software artifacts could be classified. Towards this, LDA models with topics in the range of 2 to 20 have been generated\footnote{LDA has been applied on data of each artifact individually, for each repository}, for textual data from each of the software artifacts for 50 randomly selected repositories in each programming language. Coherence scores for these models have been calculated and the number of topics of model with highest coherence score are identified as optimal number of topics for the corresponding data source. It is to be noted that data of different repositories or different artifacts was not combined during this process, to ensure that the topic number is not influenced by the data of other repositories. We observed that the number of optimal topics for issue data varied between 4 and 5, with 5 having higher frequency than 4. The optimal topics for commit data and pull request data were also observed to be varying between 4 and 5, but, with 4 being more frequently repeated, than 5. Hence, the optimal number of topics for commit data, issue data and pull request data are 4, 5 and 4 respectively. LDA topic modelling was applied on the same data\footnote{Same data used to identify optimal number of topics} with optimal number of topics and the top 10 keywords for each of the topics were obtained.
A manual inspection of the keywords obtained for each artifact has been performed repository wise, for the 200 repositories (50 from each programming language) being considered. We have manually walked through keywords in each topic, compared them to source code of the repository and other files in the repository to identify the object of reference, and labelled the topics accordingly. During the manual inspection of keywords, we did not find any keywords that explicitly correspond to architecture-related documentation. Considering the design diagrams to contain more relevant architecture related documentation, we assumed such documentation to be absent in the documentation sources being considered for this study. Based on the identified keywords, it could also be observed that though these sources might contain architecture-related documentation, its presence is almost negligible, thus, resulting in a set of 5 documentation types being present in the considered documentation sources.
Though the optimal number of keywords for \textit{pull request} data and \textit{commit} data was observed to be 4 through LDA approach, we observed the data to contain keywords from the 5\textsuperscript{th} category as well during our manual analysis. This indicates that though the information present in \textit{pull requests} and \textit{commit} data can be classified into 4 categories, the set of these four categories, differ across repositories. For example, some repositories might comprise of \textit{API-related}, \textit{File-related}, \textit{Project-related} and \textit{Error-related} documentation, while some repositories might contain \textit{License-related, File-related, Project-related} and \textit{Error-related} documentation in the \textit{pull request} and \textit{commit} data.
While labelling, in some cases, more than one topics were observed to have same labels, which indicate that different repositories have varied distributions of information in view of documentation types. Keywords of similar topics identified for all the artifact data are compared and 10 most frequently repeated keywords for each documentation type are identified.
\subsection{Automated Analysis}
A rule based classification model is built by considering the 10 most frequent keywords of each topic and its corresponding label.
The topics obtained as a result of topic modelling for each software artifact are labelled using the rule based classifier.
The rule based classifier includes rules to label topics based on similarity score of the topic with respect to keyword sets of each of the 5 identified categories(\textit{error-related}, \textit{file-related}, \textit{project-related}, \textit{license-related} and \textit{API-related}). Topics that have almost equal similarity scores (difference less than 0.05) for all categories are labelled into \textit{others} category.
Thus, artifact data having almost equal possibility to belong to more than one categories are classified into \textit{others} category.
The percentage of each documentation type is calculated based on the topic frequency in the artifact. These two features - identifying documentation types and calculating percentages, are integrated into a result generator script written in python programming language. This result generator takes as input list of information extracted from multiple software artifacts of all 950 repositories and automatically generates percentage of different documentation types present in each of the software artifacts for all the 950 repositories. Thus, the percentages generated imply frequency of related-keywords of each documentation type, in each of the artifacts.
\section{Results}
The distribution of documentation types obtained as a result of automated analysis of each artifact data is presented in Fig. \ref{fig:types}. The contribution of sources to all documentation types is presented in Fig. \ref{fig:sources}.
Also the results of artifact contributions to specific documentation types and distribution of documents among specific sources across multiple programming languages are presented in the form of plots.
\subsection{Distribution of Documentation Types in Documentation Sources}
\begin{itemize}
\item \textbf{Source Code Comments -} The percentage distribution of all documentation types in \textit{Source code comments} is presented in Fig. \ref{fig:comments}. It has been observed that majority of the information present in source code comments contribute to license based documentation, in repositories of all programming languages considered. Also, we observed that information present in source code comments contribute the least to \textit{error-related} documentation, in repositories of C\# and Java programming languages, while that in repositories of C++ and Python programming languages contribute the least to \textit{API-related} documentation.
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{comments.png}
\caption{Distribution of document types in source code comments in repositories of the four programming languages considered.}
\label{fig:comments}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{texts.png}
\caption{Distribution of document types in textual documents in repositories of the four programming languages considered.}
\label{fig:texts}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{commits.png}
\caption{Distribution of document types in commits in repositories of the four programming languages considered.}
\label{fig:commits}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{issues.png}
\caption{Distribution of document types in issues in repositories of the four programming languages considered.}
\label{fig:issues}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{prs.png}
\caption{Distribution of document types in pull requests in repositories of the four programming languages considered.}
\label{fig:prs}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{allSources.png}
\caption{Average distribution of document types in all sources, across all 950 repositories.}
\label{fig:all}
\vspace{-4mm}
\end{figure}
\item \textbf{Textual Documents-} The plot shown in Fig. \ref{fig:texts} indicates that most of the information in \textit{textual documents} of the repositories contribute to \textit{license-related} documentation, followed by \textit{file-related} documentation in repositories of C++ and Python programming languages.
\item \textbf{Commits-} The plot displayed in Fig. \ref{fig:commits} indicates that most of the information present in textual sentences of commit logs contribute to \textit{project-related} documentation in majority of the repositories, followed by \textit{error-related} documentation. For repositories in C++ programming language, majority of the information in commit logs contributes to \textit{error-related} documentation, followed by \textit{project-related} documentation.
\item \textbf{Issues-} The plot shown in Fig. \ref{fig:issues} indicates that percentage distribution of \textit{error-related} documentation and \textit{project-related} documentation are prominent in textual sentence information of issues, with \textit{error-related} documentation having highest percentage distribution across repositories in all 4 programming languages.
\item \textbf{Pull Requests-} Fig. \ref{fig:prs} shows that most of the information in textual sentences of pull requests could potentially contribute to \textit{error-related} documentation, with all repositories having an average of more than 30\% information that contributes to \textit{error-related} documentation.
\end{itemize}
A consolidated distribution of documentation types across all identified sources of documentation is presented in Fig. \ref{fig:all}, which indicates that minimal amount of information in \textit{commits}, \textit{issues} and \textit{pull requests}, and majority of information in \textit{textual} \textit{documents} and\textit{ source code comments} contributes to \textit{license based} documentation.
Fig. \ref{fig:types} indicates that \textit{error-related} documentation largely exists in the information extracted.
We further observed that about 12.75\% of the total information did not contribute to any of the identified documentation types.
\newline
\newline
\fbox{\begin{minipage}{25em}
\textbf{RQ3:} Majority of the total extracted documentation from multiple sources, consists of \textit{error-related} documentation (25.9\%), followed by \textit{project-related} documentation (23.6\%). \textit{File-related} documentation and \textit{License-related} documentation account to 16.04\% and 15.99\% of the total extracted documentation respectively. 5.63\% of the total extracted documentation consists of \textit{API-related} documentation.
\textbf{Insights-} Tagging information across documentation sources, based on corresponding documentation types could help developers in better comprehension of the project.
\end{minipage}}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{types.png}
\caption{Percentage of documentation types present in information extracted from all sources of documentation for all 950 repositories}
\label{fig:types}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{file.png}
\caption{Contribution of documentation sources to \textit{file-related} documentation in repositories of the four programming languages considered}
\label{fig:file}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{error.png}
\caption{Contribution of documentation sources to error/bug-related documentation in repositories of the four programming languages considered}
\label{fig:error}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{project.png}
\caption{Contribution of documentation sources to \textit{project-related} documentation in repositories of the four programming languages considered}
\label{fig:project}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{api.png}
\caption{Contribution of documentation sources to \textit{API-related} documentation in repositories of the four programming languages considered}
\label{fig:api}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{lic.png}
\caption{Contribution of documentation sources to \textit{license-related} documentation in repositories of the four programming languages considered}
\label{fig:lic}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{allTypes.png}
\caption{Average contribution of documentation sources to all documentation types, across all 950 repositories}
\label{fig:types1}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{sources.png}
\caption{Percentage of contribution of documentation sources to all documentation types for all 950 repositories.}
\label{fig:sources}
\vspace{-4mm}
\end{figure}
\subsection{ Contribution of Documentation Sources to Documentation types}
\begin{itemize}
\item \textbf{File-related Documentation-} The plot in Fig. \ref{fig:file} represents contribution of each documentation source to \textit{file-related} documentation, which varied across repositories of the 4 programming languages. We observe that \textit{pull requests} contribute to most of the \textit{file-related} documentation in repositories of C\#, while \textit{textual documents} contribute the most in repositories of C++ programming language. We also observed that source-code comments contribute the least to \textit{file-related} documentation in repositories of C++, C\# and Java programming languages.
\item \textbf{Error-related Documentation-} Fig. \ref{fig:error} shows that maximum contribution of \textit{error-related} documentation is from \textit{pull requests} and \textit{issues}. It can also be observed that \textit{textual} documents and \textit{source-code comments} contain the least amount of potential information with respect to \textit{error-related} documentation.
\item \textbf{Project-related Documentation-} The plot in Fig. \ref{fig:project} indicates that most of the \textit{project-related} documentation is contributed by information in \textit{commit logs}, when compared to other artifacts being considered.
\item \textbf{API-related Documentation-} We deduce that source-code comments in repositories contain more information about \textit{API-related} documentation, succeeded by textual documents in the repositories from the plot in Fig. \ref{fig:api}.
\item \textbf{License-related Documentation-} It could be observed from the plot shown in Fig. \ref{fig:lic}, that textual documents in majority of the repositories in all four programming languages contribute to \textit{license-related} documentation, followed by source code comments in repositories.
\end{itemize}
The consolidated contribution of documentation sources towards all documentation types is presented in Fig. \ref{fig:types1}, which indicates that most of the potential information for \textit{file-related} documentation
is obtained by data in \textit{textual} documents.
Fig. \ref{fig:sources} indicates that \textit{commit logs} contribute the most towards documentation, succeeding the \textit{source code comments} and \textit{textual documents}.
\newline
\newline
\fbox{\begin{minipage}{25em}
\textbf{RQ4:} Majority of the total documentation is extracted from source code comments (23.04\%), followed by textual documents (22.58\%). Commit logs and issues contributed to 18.5\% and 18.21\% of the total extracted documentation respectively. 17.63\% of the total extracted documentation was contributed by issues in the repositories.
\textbf{Insights-} Structured formats for each of these sources, could help in easy extraction and towards generation of unified documentation for a repository.
\end{minipage}}
\section{Discussion}
Towards answering the research questions proposed, we realised that different documentation types could be derived from other software artifacts in the GitHub repositories. Researchers could explore the direction of consolidating this information across multiple artifacts to arrive at documentation for the repository. Establishing traceability links among documentation types and documentation sources is an important future direction. Tools towards generating documentation from multiple sources could be developed to ease the efforts of software practitioners. It might be difficult to generate meaningful documentation from the sources identified, requiring software practitioners to improve the textual information being logged through these artifacts. Also, there is a need for approaches and tools to migrate existing unstructured documentation into a structured form, for eliciting valuable information from multiple sources of software documentation. Furthermore, researchers could also explore specific formats for each of the identified documentation sources, to ensure extraction of information that could be easily converted to a meaningful documentation from the sources.
\section{Threats to Validity}
\textbf{Internal Validity - }The categories in proposed taxonomy are based on developer interviews and card-sorting approach. Thus, there is a possibility of missing categories in the proposed taxonomy of types and sources of documentation, that might not have been observed during our interviews with 20 developers or during the execution of card-sorting approach.
Moreover, as a first step towards considering documentation sources, we considered only pull requests, commits, issues, source code files and textual files which could be obtained through GitHub API or by downloading the repositories. Other sources such as Wiki, which are specific to GitHub projects, and, which require separate cloning, other than that of repository were not considered in this study. Such sources could be considered in the future versions of this study to improve the documentation being extracted.
Except for the top 50 repositories in each programming language, topics of artifact data are labelled by comparing topic keywords with most frequently occurring keywords of respective labelled-topics, in top 50 repositories in each language. This generalization of keywords might not consider prominent keywords that could occur in other repositories, and thus could be biased towards the top 20\% repositories in each programming language. In addition, as we only analyzed textual sentences of data from \textit{issues}, \textit{pull requests} and \textit{commits}, we might have missed information from other data present in these artifacts such as \textit{date} and \textit{status}.
Also, only comments in source code files with extensions .cpp, .py, .cs and .java were analyzed, considering the complexity involved in addressing files of all programming languages. Also, text of files in the format specified in Table \ref{tab:fileclass} has been analyzed. As a result, useful insights on documentation that might be present in files of other formats such as \textit{.html, .tex, .pdf}, and comments in other programming language files are compromised.
\textbf{Construct Validity - }
All the results obtained are valid only for the dataset of considered 950 repositories. Performing this study on a different dataset might yield different results, However, considering the presence of repositories of varied programming languages in the dataset, indicating a representative sample of all repositories in the four programming languages, similar distribution of documentation types and sources could be identified in other repositories across GitHub, suggesting (but not proving) generalizability of the results.
\textbf{External Validity - }
The results obtained are confined to version of repositories and their and corresponding software artifacts as on 01 January 2021, as this empirical study is performed on locally downloaded repositories.
The accuracy of categorizing the extracted information into different documentation type classes is dependant on the efficiency of LDA approach. Different set of keywords might be identified for each class, consequently resulting in some sentences to be classified into different classes, if a different Natural Language Processing technique is employed.
\section{Related Work}
\begin{table}[]
\caption{Documentation Sources and Types Observed from Literature}
\vspace{-2mm}
\centering
\begin{tabular}{|l|l|l|}
\hline
\textbf{Documentation Type} & \textbf{Source of Documentation} & \textbf{Literature} \\
\hline
API-related & example files, readme file & \cite{fowkes2016parameter} \cite{prana2019categorizing}
\\
\hline
Architecture-related & UML files, non-textual files & \cite{hebig2016quest}\\
\hline
License-related & commit messages & \cite{vendome2017license} \cite{prana2019categorizing}\\
& issues, License files, readme & \\
\hline
Error/Bug-related & issue comments & \cite{rastkar2010summarizing}\\
\hline
Project installation-related & readme& \cite{prana2019categorizing}\\
\hline
Project Enhancement-related & readme&\cite{prana2019categorizing}\cite{rastkar2010summarizing}\\
\hline
Project Background-related & readme&\cite{prana2019categorizing}\\
\hline
Project Status-related & readme, commits, issues, &\cite{prana2019categorizing}\cite{coelho2018identifying}\\
&pull requests.&\\
\hline
Project Team-related & readme, commits, issues, &\cite{prana2019categorizing}\cite{rahman2014insight}\\
&pull requests.&\cite{coelho2018identifying}\\
\hline
Project Advantage-related & readme &\cite{prana2019categorizing}\\
\hline
References-related & readme&\cite{prana2019categorizing}\\
\hline
Contribution Guidelines & readme &\cite{prana2019categorizing}\\
\hline
\end{tabular}
\label{tab:lit}
\end{table}
Repositories on GitHub comprises of various artifacts such as \textit{pull requests}, \textit{issue reports}, \textit{star-count}, \textit{fork-count}, \textit{number of watchers} and so on, that provide different types of information about the repository. Stars, forks and watchers are numeric in nature, indicating information on popularity of repositories \cite{borges2018s} whereas software artifacts such as \textit{issue reports}, \textit{pull requests}, \textit{commit logs} contain information related to development and maintenance of the repositories \cite{dabbish2012social, sheoran2014understanding, rahman2014insight}.
Researchers have conducted several studies to understand the role of these software artifacts as an aid for developers in improving quality and better handling of the project. A consolidated list of documentation types observed from the literature, along with their sources is presented in Table \ref{tab:lit}.
\subsection{Using Individual Artifact Data}
\textit{Issues}, \textit{commits} and pull \textit{requests} data is used in the literature for multiple research studies. Kikas et al. have observed various features present in \textit{issue reports}, as an aid in prediction of \textit{issue lifetimes} \cite{kikas2016using}. They have observed that analysis of contextual information from commits of the repository such as \textit{last commit date}, \textit{recent project activity}, \textit{date of issue creation} and other dynamic features such as number of actions performed on the issue, number of users who have worked on the issue, \textit{comment count} and so on, could help in predicting life time of the \textit{issue} \cite{kikas2016using}.
Liao et al. have observed that users who comment on \textit{issues} and who create \textit{pull requests} related to the \textit{issues} contribute more towards project development.
Zhang et al. have linked multiple related issues to support easy resolution and easy querying of issues by applying a combination of information retrieval technique and deep learning techniques \cite{zhang2020ilinker}. They also developed a tool to calculate frequency, word and document similarity scores between the queried issue and pending issues, based on the mentioned techniques to recommend related issues \cite{zhang2020ilinker}.
Michaud et al. have attempted to identify branch of commits in the repository based on commit messages and types of merges to track the evolution history of a repository \cite{michaud2016recovering}.
Tsay et al. have presented a study that analyzed various factors of \textit{pull requests} that contribute towards their acceptance or rejection \cite{tsay2014influence}. It has been observed that pull requests with large number of changed files and more number of comments have lesser probability to be accepted, depending on other social factors \cite{tsay2014influence}.
\subsection{Using Integrated Data from Multiple Software Artifacts}
Zhou et al. have considered various features from issue reports and commits to detect security threats or vulnerabilities in the repository \cite{zhou2017automated}. An optimal machine learning model has been built using commit messages from commits and multiple features of issue reports such as title, description, comments and so on, to detect unidentified security threats in the repository \cite{zhou2017automated}.
Considering the importance of information related to issue reports and commit history in assessing software quality and other factors of a repository, RCLinker has been proposed to link issue reports to their corresponding commits by comparing source code of two consequent commits and summarizing the difference in source code \cite{le2015rclinker}.
Coelho et al. have considered features of multiple software artifacts such as issues, pull requests, commits and so on to identify status of maintenance of a project \cite{coelho2018identifying}. In an attempt to identify various issues in software documentation, Aghajani et al. have studied multiple sources that discuss about software documentation, which also include issue reports and pull requests on GiHhub, apart from developer email lists and discussions on knowledge sharing platforms \cite{aghajani2019software}.
\subsection{Documentation in Software Projects}
A study has been performed by Borges et al. to identify various software artifacts that contribute to star count of GitHub repositories \cite{borges2018s}. It was observed that code quality and documentation largely influence star count. This study also re-emphasizes the importance of software documentation in GitHub repositories \cite{borges2018s}.
Considering this importance of documentation, several attempts to improve and generate different types of software documentation are being developed.
\textit{Quasoledo} has been proposed to evaluate quality of documentation in \textit{Eclipse} projects, based on quality metrics related to completeness and readability \cite{schreck2007documentation}.
Fowkes et al. have proposed an approach using probabilistic model of sequences to generate patterns of API, based on multiple usages in a project \cite{fowkes2016parameter}.
Source code examples have been linked to official API documentation based on method calls and references to enable better API usage and understanding \cite{subramanian2014live}. A study has been performed by Hebig et al. to understand the number of projects in GitHub, that use uml diagrams \cite{hebig2016quest}.
In a survey with 146 software practitioners, Aghajani et al. have presented 13 documentation types, that were observed to be used by practitioners, for accomplishing multiple tasks \cite{aghajani2020software}. Of these 13 documentation types, around 6 of the documentation types correspond to end-user interactions with the project, such as \textit{video tutorials}, \textit{installation guides}, and so on. These set of documentation types also included \textit{Community Knowledge}, which could correspond to artifacts such as pull requests, issues and commits \cite{aghajani2020software} .
Prana et al. have identified 8 categories of information present in ReadMe files through manual analysis of 50 ReadMe files, and annotated 150 ReadMe files based on these categories \cite{prana2019categorizing}.
Information in artifacts of the repository such as commit messages and discussions on issue trackers have been analyzed to obtain insights about changes in licenses used in the software \cite{vendome2015license}. Vendome et al. have attempted to trace the reasons for change in licenses through investigation of commit history and discussions in issue trackers and consequently obtaining insights on when (from commit history) and why (from issue tracker discussions) the licenses have been changed \cite{vendome2015license}.
The existing literature has emphasized the need for software documentation, analyzed the availability of documentation types in GitHub repositories and has also presented approaches that consider information present in software artifacts, to generate respective documentation. In our analysis of the literature, we observe that the generation of documentation considers information from individual artifacts and is mostly limited to information present in \textit{bug reports}, \textit{source code} and \textit{readme files}. We also observe that documentation is present in many software artifacts, but is not explicitly mentioned. Extracting documentation information present in various software artifacts could help in enhancing the documentation. It is important to identify software artifacts that serve as potential sources of documentation.
However, to the best of our knowledge, we are not aware of any work in the literature that identifies and comprehends all the potential sources of different types of documentation. Hence, as a first step towards identifying potential documentation sources, we present a taxonomy of sources of documentation in GitHub.
\section{Conclusion and Future Work}
In this paper, we proposed a taxonomy of documentation sources in GitHub repositories. We identified sources of documentation for different documentation types through card-sorting approach and developer interviews. This resulted in a taxonomy of six different documentation sources that could provide potential information for different documentation types identified. We also perform an empirical analysis to understand distribution of documentation types across multiple software artifacts that are identified as documentation sources, and the contribution of each source to the type of documentation. 950 GitHub repositories, with 233, 241, 253 and 223 repositories from C++, C\#, Java and Python programming languages respectively, have been scraped to perform the study.
Software artifacts such as \textit{issues}, \textit{commits} and \textit{pull requests} of these repositories, not present explicitly in scraped data were fetched using GitHub API. Topic modelling has been applied on data extracted from all documentation sources and the obtained topics are labelled accordingly. The total extracted documentation from multiple sources, consisted 25.9\% of \textit{error-related} documentation and 23.6\% of \textit{project-related} documentation. \textit{File-related} documentation and \textit{license-related} documentation account to 16.04\% and 15.99\% of the total extracted documentation respectively. 5.63\% of the total extracted documentation consisted of \textit{API-related} documentation. \textit{Source code comments} contributed to 23.04\% and \textit{textual documents} contributed to 22.58\% of the total extracted documentation. \textit{Commit logs} and \textit{issues} contributed to 18.5\% and 18.21\% of the total extracted documentation respectively. 17.63\% of the total extracted documentation was contributed by \textit{issues} in the repositories.
As a part of future work, we plan to identify other software artifacts that could be considered as potential sources of documentation. We also plan to improve the current taxonomy by adding more levels to the taxonomy. In future, information obtained from sources in the proposed taxonomy could be processed and used to generate documentation of different types. Also, techniques to extract text from design diagrams could be implemented, to leverage other documentation information that could be present in design diagrams.
\balance
\bibliographystyle{IEEEtran}
\section{Introduction}
GitHub is one of the leading platforms of open source software projects \cite{gousios2014lean}.
GitHub facilitates several developers to collaborate and contribute to projects by performing several actions such as updating projects using commits, viewing changes made to projects through commit history \cite{tsay2014influence}, logging issues or defects through issue reports and contributing to other projects through pull requests and so on \cite{jiang2017and, jiang2017understanding}. Open source projects originating from well known organizations such as \textit{vscode} from Microsoft, \textit{react-native} from Facebook and \textit{tensorflow} from Google, have more than 1K\footnote{\url{https://github.com/microsoft/vscode}}, 2K\footnote{\url{https://github.com/facebook/react-native}} and 2.8K\footnote{\url{https://github.com/tensorflow/tensorflow}} contributors respectively (as of January 2021). Developers also contribute to and reuse several other publicly available repositories on GitHub \cite{zagalsky2015emergence}. Documentation in the repositories facilitates developers to understand about a project and consequently helps them in deciding on projects they wish to contribute \cite{borges2018s}. Repositories contain \textit{Readme} files that provide information on the purpose, requirements, usage instructions and various other information about the repositories \cite{prana2019categorizing, perez2016ten}. Several other files of the repository such as \textit{License} files, \textit{UML} files and so on also present different types of repository information such as license permissions and design decisions \cite{hebig2016quest, vendome2015license}.
Though many developers are interested contribute to GitHub repositories, they face multiple hurdles during this process, resulting in reduced motivation towards contributing to repositories \cite{mendez2018open, steinmacher2016overcoming}. The existing insufficient and scattered documentation on these repositories makes it difficult for developers to understand about the repositories, consequently reducing the advantages of huge number of contributions \cite{aghajani2019software}. Developers tend to visit other artifacts in the repositories to better understand about the repository, which is effort-intensive \cite{fronchetti2019attracts, robles2006beyond}. We believe that contribution efforts from wide range of developers could be well leveraged if documentation is improved and consolidated.
Researchers have explored multiple dimensions of understanding, usage and ways to improve software documentation, as it plays a significant role in performing various tasks. These tasks include software development, testing, integration, maintenance and so on \cite{garousi2015usage, mahmood2011industrial}, but, the documentation is insufficient in majority of the projects \cite{aghajani2020software, aghajani2019software}.
For example, documentation helps in improving software development and maintenance by providing necessary information to users in different roles such as system integrators, quality analysts, developers and so on \cite{kipyegen2013importance}. Good documentation helps in re-engineering existing software during maintenance and migration \cite{de2005study, aghajani2020software}. It has also been observed that developers regard documentation as an important aspect, even in projects with faster release cycles such as agile projects \cite{stettina2011necessary}.
Considering the wide usage of GitHub and activities performed by developers on GitHub, documentation of software repositories that are hosted on GitHub also plays a major role in various stages of the project such as development, deployment, maintenance and so on. It has been observed that projects having better popularity tend to have better documentation \cite{cosentino2017systematic}. Developers of popular projects tend to improve documentation by regular updates, to provide better insights to users \cite{aggarwal2014co}. Also, documentation is one of the factors considered by developers before contributing to repositories on GitHub \cite{borges2018s}. Currently, documentation is spread across multiple files such as \textit{source code}, \textit{design diagrams}, \textit{readme files}, \textit{license files} and so on \cite{hebig2016quest, prana2019categorizing, ma2018automatic}. We hypothesize that other artifacts could also provide valuable information about a repository, that could be considered as documentation, without documentation being explicitly present. Researchers have also observed that developers spend considerable effort on multiple artifacts, other than source code, in end-user oriented projects, supporting our idea of considering documentation from multiple sources \cite{robles2006beyond, fronchetti2019attracts}. Currently, sources of documentation are diversified in a repository. The documentation present in these sources is also unclear. \textit{Unifying the information from multiple sources could help in enhancement of documentation, and eventually reduce effort for developers, thus motivating them to contribute to GitHub repositories.} Though there are studies to identify different types of documentation present in software repositories \cite{prana2019categorizing, hebig2016quest, chimalakonda2020software}, we are not aware of any work that integrates documentation from multiple artifacts, thus, motivating our work. Hence, in this paper, we aim to identify and gather different artifacts of software repositories, that could contain information relevant to documentation. The main contributions of this paper are:
\begin{itemize}
\item A taxonomy of documentation sources in GitHub Repositories based on card-sorting approach and developer-based interviews with 20 developers.
\item An empirical analysis on 950 GitHub repositories, of four programming languages, C++, C\#, Python and Java, to understand the types of documentation that could be extracted from multiple software artifacts.
\item Results of the empirical study, along with the percentage of available information that could contribute to documentation, in each of the software artifacts that are identified as potential sources of documentation. The results of the study and dataset used for the study can be accessed here\footnote{Results \& Dataset - \url{https://osf.io/dfx9r/?view_only=0954174b44054893a3dfbe6cc7dd5db5}}.
\end{itemize}
\vspace{1mm} \fbox{\begin{minipage}{25em}
\textit{\textbf{Documentation Type}} - We define type of documentation based on content present in the documentation. For example, documentation that refers to reporting or fixing an error is considered as \textit{Error-related} documentation type.
\end{minipage}
}
\vspace{1mm}
\newline
\newline
\fbox{\begin{minipage}{25em}
\textit{\textbf{Documentation Source} - } We define documentation source as a software artifact of the project, capable of providing potentially useful information related to software documentation.
\end{minipage}
}
\section{Research Methodology}
In this paper, we aim to analyze contents in GitHub repositories from a software documentation perspective, specifically focusing on various software artifacts that could serve as sources of documentation.
We followed an approach that comprises of the following six phases.
\begin{itemize}
\begin{figure*}
\centering
\includegraphics[width = \linewidth]{tax.png}
\caption{Proposed Taxonomy of Documentation Sources}
\label{fig:taxonomy}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width = \linewidth]{typetax.png}
\caption{Proposed Taxonomy of Documentation Types}
\label{fig:typestax}
\end{figure*}
\item \textit{Research Questions Definition} - Four research questions aimed to identify possible sources of software documentation have been defined.
\item \textit{Card Sorting Approach and Developer Interviews} - For manual analysis, 60 of the 950 GitHub repositories, which include 15 top trending repositories from C++, C\#, Java and Python programming languages, were selected analyzed by three individuals using card sorting approach to narrow down onto potential categories of documentation and sources that could be leveraged in answering the research questions. Twenty professional software developers were interviewed to understand different types of documentation and sources as perceived by developers.
\item \textit{Data Extraction} - 950 public GitHub repositories, which include 233, 241, 253 and 223 repositories from C++, C\#, Java and Python programming languages respectively, were scraped and processed for further analysis.
\item \textit{Topic Modelling and Manual Inspection} - LDA topic modelling techniques were applied to 20\% of the scraped data to identify topics in the data. The topics identified are manually walked through to label the categories.
\item \textit{Automated Analysis} - Models to analyze extracted data against research questions are developed
\item \textit{Result Comprehension} - Research questions are answered as a result of automated analysis.
\end{itemize}
\subsection{Research Questions}
The main goal of this study is to propose a taxonomy of documentation sources in GitHub, by identifying different potential sources of software documentation in a GitHub repository, including the sources in which documentation is not explicitly specified. We conduct this study by answering the following research questions:
\begin{itemize}
\item \textbf{RQ1: What are the types of documentation present in GitHub repositories ?}
To answer this question, we employed card-sorting approach and developer interviews with 20 developers to understand different documentation types present in GitHub repositories, as perceived by GitHub users.
\textbf{Summary:} We observe that majority of the GitHub repositories comprise of six documentation types, of which at least three of them are not explicitly mentioned as documentation in the repositories.
\item \textbf{RQ2: What are the sources of documentation in GitHub repositories? }
We answer this question by identifying various sources that are perceived to contribute to different documentation types through developer interviews with 20 developers and card-sorting approach with 3 participants.
\textbf{Summary:} We observe that majority of the developers use six artifacts of GitHub Repositories, to understand about a repository, and further contribute to it. This also included \textit{issues} and \textit{pull requests}, which do not explicitly convey documentation related information.
\item \textbf{RQ3: What is the distribution of different documentation types in each of the identified sources?}
To answer this question, we categorize the information extracted from each source of 950 GitHub repositories into one of the documentation types identified, and calculate the percentage of each type in the sources.
\textbf{Summary:} Of the six documentation types identified, \textit{error-related} documentation is present to a larger extent in majority of the documentation sources.
\item \textbf{RQ4: What is the contribution of identified sources to different documentation types?}
To answer this question, we calculate the percentage contribution of a source by finding the ratio of amount of information present in a source with respect to the documentation type to the total amount of information of the documentation type, present across different sources.
\textbf{Summary:} Following a well-established notion, majority of the documentation was observed to be present in \textit{source-code} comments and \textit{textual documents}. However, \textit{commit logs} and \textit{issues} were also observed to contribute to documentation, with a minor variation in the percentage, in comparison to the \textit{textual} documentation.
\end{itemize}
\subsection{Card Sorting Approach and Developer Interviews}
We observe that card-sorting \cite{spencer2004card} approach has been used widely towards arriving at various taxonomies and classifications in the software engineering literature, such as classifying requirements change \cite{nurmuliani2004using} and research ideas\cite{lo2015practitioners}, which motivated us to employ card-sorting as an initial step towards arriving at different types of documentation and their sources.
We have downloaded 60 of the 950 GitHub repositories, such that they include 15 top trending repositories, from each of the four programming languages being considered, for manual analysis. Three individuals, comprising two researchers and one under-graduate student were assigned the task of identifying different types and sources of documentation present in the repositories. Initially, all the three individuals have explored \textit{readme files} of the repositories, considering them to be a basic and easily available source of documentation. We employed open card sorting approach, without pre-defining possible groups of documentation. In the first step, each of the three individuals came up with different number of groups (4, 6 and 10).
On further iterative discussions on labels of each of the groups, we observed similarities between the information content present among the 10 groups identified by the third individual, resulting in a set of 5 groups (T1 to T4 and T6 documentation types presented in Table \ref{tab:typedefs}). We further compared and discussed about the data content among other groups, identified by each individual and finally arrived at a common decision, with 6 possible types of documentation, accompanied by labels to these types, as mentioned in Table \ref{tab:typedefs}. As a next step, all the three individuals navigated through various artifacts within the repositories, and attempted to identify documentation in those artifacts. On identification of documentation, each artifact was labelled with the corresponding documentation type (identified in the previous step) it would be representing. While one individual narrowed down to five artifacts, the other two individuals pointed out six possible artifacts containing documentation. Each of the three individuals have explained the reasons for selecting respective number of artifacts, based on which, we finally decided to consider six different artifacts (mentioned in Fig. \ref{fig:taxonomy}) as part of the proposed taxonomy.
We later organised developer interviews with 20 professional developers to validate the types and sources of documentation obtained as a result of the card sorting approach. We approached 20 developers working in different organizations, with work experience ranging from 2 to 30 years. All the 20 developers interviewed belonged to organizations with more than 1000 employees and with a team size of at least 10. We selected developers with this profile to ensure that they are well-equipped with the collaborative project development and management. All the 18 developers with less than 10 years of experience mentioned that they actively contribute to open-source projects.
We asked them open-ended questions to understand their perception about types of documentation and the sources of documentation present on GitHub. 12 of the developers, with work experience ranging between 3 to 10 years explained that they find six documentation types (presented in Table \ref{tab:typedefs}) in the projects they interact with, on GitHub. 2 developers with work experience of 29 and 30 years pointed out that they find seven types of documentation in the projects, which included \textit{migration guidelines}, apart from the six types identified in Table \ref{tab:typedefs}. \textit{Migration guidelines} could however be considered as \textit{project-related} documentation, at a higher level of abstraction, as they specify details necessary in migrating the project. The rest of the developers with less than three years of experience have expressed that they have observed only three types of documentation (T1, T2 and T5 in Table \ref{tab:typedefs}). This varied insights from developers could also be due to varied abstractions of projects they are exposed to, in their respective organizations. We then discussed with them the documentation types identified previously, through card-sorting approach. 18 of the 20 developers have later agreed to the identified six documentation types, while the other two developers have suggested inclusion of one more category, which corresponds to \textit{migration guidelines} of the project.
Considering the majority responses, we decided to proceed further, with six documentation types. We further queried the developers to understand different artifacts on GitHub that they use to understand about a repository, which could be helpful, if collated as a single documentation file. All the 20 developers have pointed out six sources of documentation, as shown in Figure \ref{fig:taxonomy}, in GitHub repositories, and also mentioned that five (\textit{Source code}, \textit{textual documents}, \textit{commits}, \textit{pull requests} and \textit{issue reports}) of these six documentation sources, are frequently referred by them.
\begin{table}[]
\caption{Observed Documentation Types and corresponding definitions }
\vspace{-2mm}
\centering
\begin{tabular}{|l|l|l|}
\hline
\textbf{S.}&
\textbf{Types of} & \textbf{Definition}\\
\textbf{no.}& \textbf{Documentation} & \\
\hline
T1 & API-related & Documentation capable of providing information \\
&documentation& about APIs used in the project\\
\hline
T2 & File-related & Documentation capable of providing \\
&Documentation & file-level information such as updates made to \\
&&the files and dependencies of files\\
\hline
T3 & Project-related & Documentation capable of providing \\
&Documentation & project-level information such as installation \\
& & instructions, branches in projects, enhancements \\
& & to the projects and so on\\
\hline
T4 & License-related & Documentation capable of providing information \\
&Documentation & about licenses in the project
\\
\hline
T5 & Error/Bug-related & Documentation capable of providing information \\
&documentation& about errors or bugs encountered \\
& & in multiple files of project\\
\hline
T6 & Architecture- & Documentation capable of providing information \\
&related &about architecture of the project, such as \\
&Documentation&module interactions.\\
\hline
\end{tabular}
\label{tab:typedefs}
\vspace{-2mm}
\end{table}
Based on the card sorting approach and developer interviews, we propose six potential documentation sources, that includes 3 software artifacts which do not have documentation explicitly specified, as shown in Fig. \ref{fig:taxonomy}. Thus, we observe six types of documentation and six sources of documentation, as mentioned in Table \ref{tab:typedefs} and Fig. \ref{fig:taxonomy} respectively.
\newline
\newline
\fbox{\begin{minipage}{25em}
\textbf{RQ1:} We observed six different documentation types in GitHub Repositories. \textit{API-related}, \textit{file-related}, \textit{project-related}, \textit{license-related}, \textit{error/bug-related} and \textit{architecture-related} documentations are the six different documentation types identified.
\textbf{Insights-} Developers perceived \textit{file-related} and \textit{project-related} documentation types to be more useful in contributing to a project, among the six types.
\end{minipage}}
\vspace{1mm}
\newline
\newline
\fbox{\begin{minipage}{25em}
\textbf{RQ2:} We observed six different sources containing potential information related to documentation in GitHub repositories.\textit{ Source code, textual documents, commits, pull requests, issue reports and design diagrams} are the six artifacts, observed to be potential sources of documentation in GitHub repositories.
\textbf{Insights-} Apart from the widely accepted documentation sources - \textit{source code comments},\textit{design diagrams} and \textit{textual documents}, our interviews with developers revealed that \textit{pull requests}, \textit{commits} and \textit{issues} can also contribute to documentation.
\end{minipage}}
\subsection{Data Extraction}
We identified top-starred 300 GitHub repositories, written in each of the four different programming languages, C++, Python, C\# and Java. The choice of languages was based on random numbers generated from 1 to 10, and the languages corresponding to these random numbers among the top 10 active languages on GitHub\footnote{https://githut.info/}, as of 01
January 2021. We further filtered out forked repositories and other repositories with zero pull requests, to exclude tutorial-based repositories. This resulted in 233, 241, 253 and 223 repositories corresponding to C++, C\#, Java and Python programming languages respectively, leading to a total of 950 repositories. We analyzed each of the repositories to extract information from multiple documentation sources. We have manually walked through top 15 repositories in each of the programming languages to identify file formats for different documentation types. Based on the findings from developer interviews and card-sorting approach, we classified all files in the repositories into 5 categories - Textual Documentation, Images, Design Diagrams, Source Code and Others. The rules employed for this categorization are presented in the Table \ref{tab:fileclass}.
\begin{table}[]
\caption{File Classification rules}
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Category } & \textbf{File Format} \\
\hline
Textual & `.txt' , `.md' , `readme' , `license'\\
Images & `.png', `.jpg', `.jpeg'\\
Design Diagrams & `.xmi',`.uml'\\
Source Code & `.cpp', `.cs', `.py', `.java'\\
Others & Any other extensions\\
\hline
\end{tabular}
\label{tab:fileclass}
\end{table}
Files with extensions - {`.txt', `.md', } are categorized to \textit{textual} documentation.
All files with extensions - {`.xmi', `.uml'} are categorized to UML diagrams.
Files with {`.jpg',`.jpeg',`.png'} extensions are primarily categorized as Images, which could be categorized as design diagrams through further analysis. Files containing extension of programming languages considered, i.e., .cpp, .py, .java and .cs are classified as source code files. Though the repositories were extracted based on the programming language in which they are written, a majority of the repositories also contain files written in other programming languages.
Considering the vast number of programming languages supported by GitHub, it is difficult to analyze files of other programming languages, as programming constructs differ from one language to another. Hence, files with formats not belonging to the above four classes are categorized as \textit{Others}. Comments from files in the \textit{source code} category have been extracted. This exploration of repository files has extracted information that could contribute to documentation, from two documentation sources mentioned in the taxonomy, namely - \textit{textual documents} and \textit{source code files}. Also, \textit{design diagrams} for each of the repositories have been identified and information from these diagrams could be extracted using image processing techniques. In view of the technical effort and difficulty involved in choosing and applying appropriate image processing techniques, we consider data extraction from design diagrams to be out of the scope of this study.
For each of the repositories, other software artifacts, i.e., \textit{pull requests, issues} and \textit{commits}, that were identified as sources of documentation in the proposed taxonomy, not older than three years from January 01, 2021, were extracted using GitHub API. However, blank data was returned for 46 python repositories, leaving us with pull request data of only 177 of the 223 Python repositories considered. This blank data could be due to the large number of pull requests in these repositories, which could not be extracted due to rate limits of GitHub API, or those repositories with pull requests older than three years, from 01 January, 2021.
We have identified fields that are capable of containing documentation related information through manual inspection of extracted artifacts in 50 of the 950 repositories, selected randomly. These fields are presented in Table \ref{tab:fields}, and the data of these fields is stored as text documents. This accomplishes the task of extracting information useful for documentation from three sources of documentation, mentioned in the proposed taxonomy, namely - \textit{Pull Requests, Issues} and \textit{Commits}.
\begin{table}[]
\caption{Documentation Sources and fields considered}
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Documentation Source} & \textbf{Fields Identified}\\
\hline
Issues & title , body, comments\\
Pull Requests & title, body, comments\\
Commits & message, comments\\
\hline
\end{tabular}
\label{tab:fields}
\end{table}
The fields that contain textual data and contribute to documentation types are identified (as presented in Table \ref{tab:fields}) and corresponding text is stored for further analysis.
\subsection{Topic Modelling and Manual Inspection}
We observed that software artifacts that are identified as sources of documentation contain information that could support different documentation types. Hence, we analyze this information and categorize into different documentation types. Topic modelling technique can identify keywords, that belong to different number of topics specified, in a given document. This feature of topic modelling fits well with our requirement of categorizing text into different documentation types. We performed Latent Dirichlet Allocation (LDA) topic modelling technique to categorize the information. Though we have consolidated different types of documentation, all sources of documentation need not contribute to all types of documentation. Hence, we primarily tried to understand the optimal number of topics into which information from each of the software artifacts could be classified. Towards this, LDA models with topics in the range of 2 to 20 have been generated\footnote{LDA has been applied on data of each artifact individually, for each repository}, for textual data from each of the software artifacts for 50 randomly selected repositories in each programming language. Coherence scores for these models have been calculated and the number of topics of model with highest coherence score are identified as optimal number of topics for the corresponding data source. It is to be noted that data of different repositories or different artifacts was not combined during this process, to ensure that the topic number is not influenced by the data of other repositories. We observed that the number of optimal topics for issue data varied between 4 and 5, with 5 having higher frequency than 4. The optimal topics for commit data and pull request data were also observed to be varying between 4 and 5, but, with 4 being more frequently repeated, than 5. Hence, the optimal number of topics for commit data, issue data and pull request data are 4, 5 and 4 respectively. LDA topic modelling was applied on the same data\footnote{Same data used to identify optimal number of topics} with optimal number of topics and the top 10 keywords for each of the topics were obtained.
A manual inspection of the keywords obtained for each artifact has been performed repository wise, for the 200 repositories (50 from each programming language) being considered. We have manually walked through keywords in each topic, compared them to source code of the repository and other files in the repository to identify the object of reference, and labelled the topics accordingly. During the manual inspection of keywords, we did not find any keywords that explicitly correspond to architecture-related documentation. Considering the design diagrams to contain more relevant architecture related documentation, we assumed such documentation to be absent in the documentation sources being considered for this study. Based on the identified keywords, it could also be observed that though these sources might contain architecture-related documentation, its presence is almost negligible, thus, resulting in a set of 5 documentation types being present in the considered documentation sources.
Though the optimal number of keywords for \textit{pull request} data and \textit{commit} data was observed to be 4 through LDA approach, we observed the data to contain keywords from the 5\textsuperscript{th} category as well during our manual analysis. This indicates that though the information present in \textit{pull requests} and \textit{commit} data can be classified into 4 categories, the set of these four categories, differ across repositories. For example, some repositories might comprise of \textit{API-related}, \textit{File-related}, \textit{Project-related} and \textit{Error-related} documentation, while some repositories might contain \textit{License-related, File-related, Project-related} and \textit{Error-related} documentation in the \textit{pull request} and \textit{commit} data.
While labelling, in some cases, more than one topics were observed to have same labels, which indicate that different repositories have varied distributions of information in view of documentation types. Keywords of similar topics identified for all the artifact data are compared and 10 most frequently repeated keywords for each documentation type are identified.
\subsection{Automated Analysis}
A rule based classification model is built by considering the 10 most frequent keywords of each topic and its corresponding label.
The topics obtained as a result of topic modelling for each software artifact are labelled using the rule based classifier.
The rule based classifier includes rules to label topics based on similarity score of the topic with respect to keyword sets of each of the 5 identified categories(\textit{error-related}, \textit{file-related}, \textit{project-related}, \textit{license-related} and \textit{API-related}). Topics that have almost equal similarity scores (difference less than 0.05) for all categories are labelled into \textit{others} category.
Thus, artifact data having almost equal possibility to belong to more than one categories are classified into \textit{others} category.
The percentage of each documentation type is calculated based on the topic frequency in the artifact. These two features - identifying documentation types and calculating percentages, are integrated into a result generator script written in python programming language. This result generator takes as input list of information extracted from multiple software artifacts of all 950 repositories and automatically generates percentage of different documentation types present in each of the software artifacts for all the 950 repositories. Thus, the percentages generated imply frequency of related-keywords of each documentation type, in each of the artifacts.
\section{Results}
The distribution of documentation types obtained as a result of automated analysis of each artifact data is presented in Fig. \ref{fig:types}. The contribution of sources to all documentation types is presented in Fig. \ref{fig:sources}.
Also the results of artifact contributions to specific documentation types and distribution of documents among specific sources across multiple programming languages are presented in the form of plots.
\subsection{Distribution of Documentation Types in Documentation Sources}
\begin{itemize}
\item \textbf{Source Code Comments -} The percentage distribution of all documentation types in \textit{Source code comments} is presented in Fig. \ref{fig:comments}. It has been observed that majority of the information present in source code comments contribute to license based documentation, in repositories of all programming languages considered. Also, we observed that information present in source code comments contribute the least to \textit{error-related} documentation, in repositories of C\# and Java programming languages, while that in repositories of C++ and Python programming languages contribute the least to \textit{API-related} documentation.
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{comments.png}
\caption{Distribution of document types in source code comments in repositories of the four programming languages considered.}
\label{fig:comments}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{texts.png}
\caption{Distribution of document types in textual documents in repositories of the four programming languages considered.}
\label{fig:texts}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{commits.png}
\caption{Distribution of document types in commits in repositories of the four programming languages considered.}
\label{fig:commits}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{issues.png}
\caption{Distribution of document types in issues in repositories of the four programming languages considered.}
\label{fig:issues}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{prs.png}
\caption{Distribution of document types in pull requests in repositories of the four programming languages considered.}
\label{fig:prs}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{allSources.png}
\caption{Average distribution of document types in all sources, across all 950 repositories.}
\label{fig:all}
\vspace{-4mm}
\end{figure}
\item \textbf{Textual Documents-} The plot shown in Fig. \ref{fig:texts} indicates that most of the information in \textit{textual documents} of the repositories contribute to \textit{license-related} documentation, followed by \textit{file-related} documentation in repositories of C++ and Python programming languages.
\item \textbf{Commits-} The plot displayed in Fig. \ref{fig:commits} indicates that most of the information present in textual sentences of commit logs contribute to \textit{project-related} documentation in majority of the repositories, followed by \textit{error-related} documentation. For repositories in C++ programming language, majority of the information in commit logs contributes to \textit{error-related} documentation, followed by \textit{project-related} documentation.
\item \textbf{Issues-} The plot shown in Fig. \ref{fig:issues} indicates that percentage distribution of \textit{error-related} documentation and \textit{project-related} documentation are prominent in textual sentence information of issues, with \textit{error-related} documentation having highest percentage distribution across repositories in all 4 programming languages.
\item \textbf{Pull Requests-} Fig. \ref{fig:prs} shows that most of the information in textual sentences of pull requests could potentially contribute to \textit{error-related} documentation, with all repositories having an average of more than 30\% information that contributes to \textit{error-related} documentation.
\end{itemize}
A consolidated distribution of documentation types across all identified sources of documentation is presented in Fig. \ref{fig:all}, which indicates that minimal amount of information in \textit{commits}, \textit{issues} and \textit{pull requests}, and majority of information in \textit{textual} \textit{documents} and\textit{ source code comments} contributes to \textit{license based} documentation.
Fig. \ref{fig:types} indicates that \textit{error-related} documentation largely exists in the information extracted.
We further observed that about 12.75\% of the total information did not contribute to any of the identified documentation types.
\newline
\newline
\fbox{\begin{minipage}{25em}
\textbf{RQ3:} Majority of the total extracted documentation from multiple sources, consists of \textit{error-related} documentation (25.9\%), followed by \textit{project-related} documentation (23.6\%). \textit{File-related} documentation and \textit{License-related} documentation account to 16.04\% and 15.99\% of the total extracted documentation respectively. 5.63\% of the total extracted documentation consists of \textit{API-related} documentation.
\textbf{Insights-} Tagging information across documentation sources, based on corresponding documentation types could help developers in better comprehension of the project.
\end{minipage}}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{types.png}
\caption{Percentage of documentation types present in information extracted from all sources of documentation for all 950 repositories}
\label{fig:types}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{file.png}
\caption{Contribution of documentation sources to \textit{file-related} documentation in repositories of the four programming languages considered}
\label{fig:file}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{error.png}
\caption{Contribution of documentation sources to error/bug-related documentation in repositories of the four programming languages considered}
\label{fig:error}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{project.png}
\caption{Contribution of documentation sources to \textit{project-related} documentation in repositories of the four programming languages considered}
\label{fig:project}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{api.png}
\caption{Contribution of documentation sources to \textit{API-related} documentation in repositories of the four programming languages considered}
\label{fig:api}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{lic.png}
\caption{Contribution of documentation sources to \textit{license-related} documentation in repositories of the four programming languages considered}
\label{fig:lic}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{allTypes.png}
\caption{Average contribution of documentation sources to all documentation types, across all 950 repositories}
\label{fig:types1}
\vspace{-4mm}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{sources.png}
\caption{Percentage of contribution of documentation sources to all documentation types for all 950 repositories.}
\label{fig:sources}
\vspace{-4mm}
\end{figure}
\subsection{ Contribution of Documentation Sources to Documentation types}
\begin{itemize}
\item \textbf{File-related Documentation-} The plot in Fig. \ref{fig:file} represents contribution of each documentation source to \textit{file-related} documentation, which varied across repositories of the 4 programming languages. We observe that \textit{pull requests} contribute to most of the \textit{file-related} documentation in repositories of C\#, while \textit{textual documents} contribute the most in repositories of C++ programming language. We also observed that source-code comments contribute the least to \textit{file-related} documentation in repositories of C++, C\# and Java programming languages.
\item \textbf{Error-related Documentation-} Fig. \ref{fig:error} shows that maximum contribution of \textit{error-related} documentation is from \textit{pull requests} and \textit{issues}. It can also be observed that \textit{textual} documents and \textit{source-code comments} contain the least amount of potential information with respect to \textit{error-related} documentation.
\item \textbf{Project-related Documentation-} The plot in Fig. \ref{fig:project} indicates that most of the \textit{project-related} documentation is contributed by information in \textit{commit logs}, when compared to other artifacts being considered.
\item \textbf{API-related Documentation-} We deduce that source-code comments in repositories contain more information about \textit{API-related} documentation, succeeded by textual documents in the repositories from the plot in Fig. \ref{fig:api}.
\item \textbf{License-related Documentation-} It could be observed from the plot shown in Fig. \ref{fig:lic}, that textual documents in majority of the repositories in all four programming languages contribute to \textit{license-related} documentation, followed by source code comments in repositories.
\end{itemize}
The consolidated contribution of documentation sources towards all documentation types is presented in Fig. \ref{fig:types1}, which indicates that most of the potential information for \textit{file-related} documentation
is obtained by data in \textit{textual} documents.
Fig. \ref{fig:sources} indicates that \textit{commit logs} contribute the most towards documentation, succeeding the \textit{source code comments} and \textit{textual documents}.
\newline
\newline
\fbox{\begin{minipage}{25em}
\textbf{RQ4:} Majority of the total documentation is extracted from source code comments (23.04\%), followed by textual documents (22.58\%). Commit logs and issues contributed to 18.5\% and 18.21\% of the total extracted documentation respectively. 17.63\% of the total extracted documentation was contributed by issues in the repositories.
\textbf{Insights-} Structured formats for each of these sources, could help in easy extraction and towards generation of unified documentation for a repository.
\end{minipage}}
\section{Discussion}
Towards answering the research questions proposed, we realised that different documentation types could be derived from other software artifacts in the GitHub repositories. Researchers could explore the direction of consolidating this information across multiple artifacts to arrive at documentation for the repository. Establishing traceability links among documentation types and documentation sources is an important future direction. Tools towards generating documentation from multiple sources could be developed to ease the efforts of software practitioners. It might be difficult to generate meaningful documentation from the sources identified, requiring software practitioners to improve the textual information being logged through these artifacts. Also, there is a need for approaches and tools to migrate existing unstructured documentation into a structured form, for eliciting valuable information from multiple sources of software documentation. Furthermore, researchers could also explore specific formats for each of the identified documentation sources, to ensure extraction of information that could be easily converted to a meaningful documentation from the sources.
\section{Threats to Validity}
\textbf{Internal Validity - }The categories in proposed taxonomy are based on developer interviews and card-sorting approach. Thus, there is a possibility of missing categories in the proposed taxonomy of types and sources of documentation, that might not have been observed during our interviews with 20 developers or during the execution of card-sorting approach.
Moreover, as a first step towards considering documentation sources, we considered only pull requests, commits, issues, source code files and textual files which could be obtained through GitHub API or by downloading the repositories. Other sources such as Wiki, which are specific to GitHub projects, and, which require separate cloning, other than that of repository were not considered in this study. Such sources could be considered in the future versions of this study to improve the documentation being extracted.
Except for the top 50 repositories in each programming language, topics of artifact data are labelled by comparing topic keywords with most frequently occurring keywords of respective labelled-topics, in top 50 repositories in each language. This generalization of keywords might not consider prominent keywords that could occur in other repositories, and thus could be biased towards the top 20\% repositories in each programming language. In addition, as we only analyzed textual sentences of data from \textit{issues}, \textit{pull requests} and \textit{commits}, we might have missed information from other data present in these artifacts such as \textit{date} and \textit{status}.
Also, only comments in source code files with extensions .cpp, .py, .cs and .java were analyzed, considering the complexity involved in addressing files of all programming languages. Also, text of files in the format specified in Table \ref{tab:fileclass} has been analyzed. As a result, useful insights on documentation that might be present in files of other formats such as \textit{.html, .tex, .pdf}, and comments in other programming language files are compromised.
\textbf{Construct Validity - }
All the results obtained are valid only for the dataset of considered 950 repositories. Performing this study on a different dataset might yield different results, However, considering the presence of repositories of varied programming languages in the dataset, indicating a representative sample of all repositories in the four programming languages, similar distribution of documentation types and sources could be identified in other repositories across GitHub, suggesting (but not proving) generalizability of the results.
\textbf{External Validity - }
The results obtained are confined to version of repositories and their and corresponding software artifacts as on 01 January 2021, as this empirical study is performed on locally downloaded repositories.
The accuracy of categorizing the extracted information into different documentation type classes is dependant on the efficiency of LDA approach. Different set of keywords might be identified for each class, consequently resulting in some sentences to be classified into different classes, if a different Natural Language Processing technique is employed.
\section{Related Work}
\begin{table}[]
\caption{Documentation Sources and Types Observed from Literature}
\vspace{-2mm}
\centering
\begin{tabular}{|l|l|l|}
\hline
\textbf{Documentation Type} & \textbf{Source of Documentation} & \textbf{Literature} \\
\hline
API-related & example files, readme file & \cite{fowkes2016parameter} \cite{prana2019categorizing}
\\
\hline
Architecture-related & UML files, non-textual files & \cite{hebig2016quest}\\
\hline
License-related & commit messages & \cite{vendome2017license} \cite{prana2019categorizing}\\
& issues, License files, readme & \\
\hline
Error/Bug-related & issue comments & \cite{rastkar2010summarizing}\\
\hline
Project installation-related & readme& \cite{prana2019categorizing}\\
\hline
Project Enhancement-related & readme&\cite{prana2019categorizing}\cite{rastkar2010summarizing}\\
\hline
Project Background-related & readme&\cite{prana2019categorizing}\\
\hline
Project Status-related & readme, commits, issues, &\cite{prana2019categorizing}\cite{coelho2018identifying}\\
&pull requests.&\\
\hline
Project Team-related & readme, commits, issues, &\cite{prana2019categorizing}\cite{rahman2014insight}\\
&pull requests.&\cite{coelho2018identifying}\\
\hline
Project Advantage-related & readme &\cite{prana2019categorizing}\\
\hline
References-related & readme&\cite{prana2019categorizing}\\
\hline
Contribution Guidelines & readme &\cite{prana2019categorizing}\\
\hline
\end{tabular}
\label{tab:lit}
\end{table}
Repositories on GitHub comprises of various artifacts such as \textit{pull requests}, \textit{issue reports}, \textit{star-count}, \textit{fork-count}, \textit{number of watchers} and so on, that provide different types of information about the repository. Stars, forks and watchers are numeric in nature, indicating information on popularity of repositories \cite{borges2018s} whereas software artifacts such as \textit{issue reports}, \textit{pull requests}, \textit{commit logs} contain information related to development and maintenance of the repositories \cite{dabbish2012social, sheoran2014understanding, rahman2014insight}.
Researchers have conducted several studies to understand the role of these software artifacts as an aid for developers in improving quality and better handling of the project. A consolidated list of documentation types observed from the literature, along with their sources is presented in Table \ref{tab:lit}.
\subsection{Using Individual Artifact Data}
\textit{Issues}, \textit{commits} and pull \textit{requests} data is used in the literature for multiple research studies. Kikas et al. have observed various features present in \textit{issue reports}, as an aid in prediction of \textit{issue lifetimes} \cite{kikas2016using}. They have observed that analysis of contextual information from commits of the repository such as \textit{last commit date}, \textit{recent project activity}, \textit{date of issue creation} and other dynamic features such as number of actions performed on the issue, number of users who have worked on the issue, \textit{comment count} and so on, could help in predicting life time of the \textit{issue} \cite{kikas2016using}.
Liao et al. have observed that users who comment on \textit{issues} and who create \textit{pull requests} related to the \textit{issues} contribute more towards project development.
Zhang et al. have linked multiple related issues to support easy resolution and easy querying of issues by applying a combination of information retrieval technique and deep learning techniques \cite{zhang2020ilinker}. They also developed a tool to calculate frequency, word and document similarity scores between the queried issue and pending issues, based on the mentioned techniques to recommend related issues \cite{zhang2020ilinker}.
Michaud et al. have attempted to identify branch of commits in the repository based on commit messages and types of merges to track the evolution history of a repository \cite{michaud2016recovering}.
Tsay et al. have presented a study that analyzed various factors of \textit{pull requests} that contribute towards their acceptance or rejection \cite{tsay2014influence}. It has been observed that pull requests with large number of changed files and more number of comments have lesser probability to be accepted, depending on other social factors \cite{tsay2014influence}.
\subsection{Using Integrated Data from Multiple Software Artifacts}
Zhou et al. have considered various features from issue reports and commits to detect security threats or vulnerabilities in the repository \cite{zhou2017automated}. An optimal machine learning model has been built using commit messages from commits and multiple features of issue reports such as title, description, comments and so on, to detect unidentified security threats in the repository \cite{zhou2017automated}.
Considering the importance of information related to issue reports and commit history in assessing software quality and other factors of a repository, RCLinker has been proposed to link issue reports to their corresponding commits by comparing source code of two consequent commits and summarizing the difference in source code \cite{le2015rclinker}.
Coelho et al. have considered features of multiple software artifacts such as issues, pull requests, commits and so on to identify status of maintenance of a project \cite{coelho2018identifying}. In an attempt to identify various issues in software documentation, Aghajani et al. have studied multiple sources that discuss about software documentation, which also include issue reports and pull requests on GiHhub, apart from developer email lists and discussions on knowledge sharing platforms \cite{aghajani2019software}.
\subsection{Documentation in Software Projects}
A study has been performed by Borges et al. to identify various software artifacts that contribute to star count of GitHub repositories \cite{borges2018s}. It was observed that code quality and documentation largely influence star count. This study also re-emphasizes the importance of software documentation in GitHub repositories \cite{borges2018s}.
Considering this importance of documentation, several attempts to improve and generate different types of software documentation are being developed.
\textit{Quasoledo} has been proposed to evaluate quality of documentation in \textit{Eclipse} projects, based on quality metrics related to completeness and readability \cite{schreck2007documentation}.
Fowkes et al. have proposed an approach using probabilistic model of sequences to generate patterns of API, based on multiple usages in a project \cite{fowkes2016parameter}.
Source code examples have been linked to official API documentation based on method calls and references to enable better API usage and understanding \cite{subramanian2014live}. A study has been performed by Hebig et al. to understand the number of projects in GitHub, that use uml diagrams \cite{hebig2016quest}.
In a survey with 146 software practitioners, Aghajani et al. have presented 13 documentation types, that were observed to be used by practitioners, for accomplishing multiple tasks \cite{aghajani2020software}. Of these 13 documentation types, around 6 of the documentation types correspond to end-user interactions with the project, such as \textit{video tutorials}, \textit{installation guides}, and so on. These set of documentation types also included \textit{Community Knowledge}, which could correspond to artifacts such as pull requests, issues and commits \cite{aghajani2020software} .
Prana et al. have identified 8 categories of information present in ReadMe files through manual analysis of 50 ReadMe files, and annotated 150 ReadMe files based on these categories \cite{prana2019categorizing}.
Information in artifacts of the repository such as commit messages and discussions on issue trackers have been analyzed to obtain insights about changes in licenses used in the software \cite{vendome2015license}. Vendome et al. have attempted to trace the reasons for change in licenses through investigation of commit history and discussions in issue trackers and consequently obtaining insights on when (from commit history) and why (from issue tracker discussions) the licenses have been changed \cite{vendome2015license}.
The existing literature has emphasized the need for software documentation, analyzed the availability of documentation types in GitHub repositories and has also presented approaches that consider information present in software artifacts, to generate respective documentation. In our analysis of the literature, we observe that the generation of documentation considers information from individual artifacts and is mostly limited to information present in \textit{bug reports}, \textit{source code} and \textit{readme files}. We also observe that documentation is present in many software artifacts, but is not explicitly mentioned. Extracting documentation information present in various software artifacts could help in enhancing the documentation. It is important to identify software artifacts that serve as potential sources of documentation.
However, to the best of our knowledge, we are not aware of any work in the literature that identifies and comprehends all the potential sources of different types of documentation. Hence, as a first step towards identifying potential documentation sources, we present a taxonomy of sources of documentation in GitHub.
\section{Conclusion and Future Work}
In this paper, we proposed a taxonomy of documentation sources in GitHub repositories. We identified sources of documentation for different documentation types through card-sorting approach and developer interviews. This resulted in a taxonomy of six different documentation sources that could provide potential information for different documentation types identified. We also perform an empirical analysis to understand distribution of documentation types across multiple software artifacts that are identified as documentation sources, and the contribution of each source to the type of documentation. 950 GitHub repositories, with 233, 241, 253 and 223 repositories from C++, C\#, Java and Python programming languages respectively, have been scraped to perform the study.
Software artifacts such as \textit{issues}, \textit{commits} and \textit{pull requests} of these repositories, not present explicitly in scraped data were fetched using GitHub API. Topic modelling has been applied on data extracted from all documentation sources and the obtained topics are labelled accordingly. The total extracted documentation from multiple sources, consisted 25.9\% of \textit{error-related} documentation and 23.6\% of \textit{project-related} documentation. \textit{File-related} documentation and \textit{license-related} documentation account to 16.04\% and 15.99\% of the total extracted documentation respectively. 5.63\% of the total extracted documentation consisted of \textit{API-related} documentation. \textit{Source code comments} contributed to 23.04\% and \textit{textual documents} contributed to 22.58\% of the total extracted documentation. \textit{Commit logs} and \textit{issues} contributed to 18.5\% and 18.21\% of the total extracted documentation respectively. 17.63\% of the total extracted documentation was contributed by \textit{issues} in the repositories.
As a part of future work, we plan to identify other software artifacts that could be considered as potential sources of documentation. We also plan to improve the current taxonomy by adding more levels to the taxonomy. In future, information obtained from sources in the proposed taxonomy could be processed and used to generate documentation of different types. Also, techniques to extract text from design diagrams could be implemented, to leverage other documentation information that could be present in design diagrams.
\balance
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:introduction}
The process of forming hypotheses, testing them on data, analysing the results, and forming new hypotheses is the foundation of the scientific method \cite{popper2002conjectures}.
For instance, imagine that Alice is a chemist trying to synthesise a vial of the compound \emph{octiron} from substances \emph{thaum} and \emph{slood}.
To do so, Alice can perform actions, such as fill a vial with a substance (\emph{fill(Vial,Sub)}) or mix two vials (\emph{mix(V1,V2,V3)}).
One such hypothesis is:
\begin{center}
\emph{synth(A,B,C) $\leftarrow$ fill(V1,A), fill(V1,B), mix(V1,V1,C)}
\end{center}
\noindent
This hypothesis says to synthesise compound \emph{C}, fill vial \emph{V1} with substance \emph{A}, fill vial \emph{V1} with substance \emph{B}, and then mix vial \emph{V1} with itself to form \emph{C}.
When Alice experimentally tests this hypothesis she finds that it \emph{fails}.
From this failure, Alice deduces that hypotheses that add more actions (i.e.~literals) will also fail (\textbf{C1}).
Alice can, however, go further and \emph{explain} the failure as ``vial \emph{V1} cannot be filled a second time'', which allows her to deduce that any hypothesis that includes \emph{fill(V1,A)} and \emph{fill(V1,B)} will fail (\textbf{C2}).
Clearly, conclusion \textbf{C2} allows Alice to eliminate more hypotheses than \textbf{C1}, that is, by explaining failures Alice can better form new hypotheses.
Our main contribution is to introduce similar explanation techniques for inductive program synthesis, where the goal is to machine learn computer programs from data \cite{mis}.
We build on the inductive logic programming (ILP) approach \emph{learning from failures} and its implementation called \textsc{Popper}{} \cite{popper}.
\textsc{Popper}{} learns logic programs by iteratively generating and testing hypotheses.
When a hypothesis fails on training examples, \textsc{Popper}{} examines the failure to learn constraints that eliminate hypotheses that will provably fail as well.
A limitation of \textsc{Popper}{} is that it only derives constraints based on \emph{entire} hypotheses (as Alice does for \textbf{C1}) and cannot explain \emph{why} a hypothesis fails (cannot reason as Alice does for \textbf{C2}).
We address this limitation by \emph{explaining} failures.
The idea is to analyse a failed hypothesis to identify \emph{sub-programs} that also fail.
We show that, by identifying failing sub-programs and generating constraints from them, we can eliminate more hypotheses, which can in turn improve learning performance.
By the Blumer bound \shortcite{blumer:bound}, searching a smaller hypothesis space should result in fewer errors compared to a larger space, assuming a solution is in both spaces.
Our approach builds on algorithmic debugging \cite{ADsurvey}.
We identify sub-programs of hypotheses by analysing paths in SLD-trees.
In similar work \cite{mis,ilasp3}, only entire clauses can make up these sub-programs.
By contrast, we can identify literals responsible for a failure \emph{within} a clause.
We extend \textsc{Popper}{} with failure explanation and experimentally show that failure explanation can significantly improve learning performance.
Our contributions are:
\begin{itemize}
\item We relate logic programs that fail on examples to their failing sub-programs.
For wrong answers we identify clauses.
For missing answers we additionally identify literals within clauses.
\item We show that hypotheses that are specialisations and generalisations of failing sub-programs can be eliminated.
\item We prove that hypothesis space pruning based on sub-programs is more effective than pruning without them.
\item We introduce an SLD-tree based technique for failure explanation.
We introduce \name{}, which adds the ability to explain failures to the \textsc{Popper}{} ILP system.
\item We experimentally show that failure explanation can drastically reduce (i) hypothesis space exploration and (ii) learning times.
\end{itemize}
\section{Related Work}
\label{sec:related}
Inductive program synthesis systems automatically generate computer programs from specifications, typically input/output examples \cite{mis}.
This topic interests researchers from many areas of machine learning, including Bayesian inference \cite{silver:aaai20} and neural networks \cite{dreamcoder}.
We focus on ILP techniques, which induce logic programs \cite{mugg:ilp}.
In contrast to neural approaches, ILP techniques can generalise from few examples \cite{ilp30}.
Moreover, because ILP uses logic programming as a uniform representation for background knowledge (BK), examples, and hypotheses, it can be applied to arbitrary domains without the need for hand-crafted, domain-specific neural architectures.
Finally, due to logic's similarity to natural language, ILP learns comprehensible hypotheses.
Many ILP systems \cite{progol,tilde,aleph,atom,lfit,inspire,fastlas} either cannot or struggle to learn recursive programs.
By contrast, \name{} can learn recursive programs and thus programs that generalise to input sizes it was not trained on.
Compared to many modern ILP systems \cite{ilasp3,dilp,hexmil,apperception}, \name{} supports large and infinite domains, which is important when reasoning about complex data structures, such as lists.
Compared to many state-of-the-art systems \cite{metaho,dilp,hexmil,celine:bottom,topmil} \name{} does not need metarules (program templates) to restrict the hypothesis space.
Algorithmic debugging \cite{ADsurvey} explains failures in terms of sub-programs.
Similarly, in databases \emph{provenance} is used to explain query results \cite{CheneyProvenance}.
In seminal work on logic program synthesis, Shapiro \shortcite{mis} analysed \emph{debugging trees} to identify failing clauses.
By contrast, our failure analysis reasons about concrete SLD-trees.
Both ILASP3 \cite{ilasp3} and the remarkably similar ProSynth \cite{prosynth} induce logic programs by precomputing every possible clause and then using a select-test-and-constrain loop.
This precompute step is infeasible for clauses with many literals and restricts their failure explanation to clauses.
By contrast, \name{} does not precompute clauses and can identify clauses and literals \emph{within} clauses responsible for failure.
\textsc{Popper}{} \cite{popper} learns first-order constraints, which can be likened to conflict-driven clause learning \cite{CDCLhandbook}.
Failure explanation in \name{} can therefore be viewed as enabling \textsc{Popper}{} to detect smaller conflicts, yielding smaller yet more general constraints that prune more effectively.
\section{Problem Setting}
\label{sec:framework}
We now reiterate the LFF problem \cite{popper} as well as the relation between constraints and failed hypotheses.
We then introduce failure explanation in terms of sub-programs.
We assume standard logic programming definitions \cite{lloyd:book}.
\subsection{Learning From Failures}
To define the LFF problem, we first define \emph{predicate declarations} and \emph{hypothesis constraints}.
LFF uses predicate declarations as a form of language bias, defining which predicate symbols may appear in a hypothesis.
A predicate declaration is a ground atom of the form $\emph{head\_pred}(\emph{p},a)$ or $\emph{body\_pred}(\emph{p},a)$ where \emph{p} is a predicate symbol of arity $a$.
Given a set of predicate declarations $D$, a definite clause $C$ is \emph{declaration consistent} when two conditions hold (i) if $\emph{p}/m$ is the predicate in the head of $C$, then $\emph{head\_pred}(\emph{p},m)$ is in $D$, and (ii) for all $\emph{q}/n$ predicate symbols in the body of $C$, $\emph{body\_pred}(\emph{q},n)$ is in $D$.
To restrict the hypothesis space, LFF uses hypothesis constraints.
Let $\mathcal{L}$ be a language that defines hypotheses, i.e.~a meta-language.
Then a hypothesis constraint is a constraint expressed in $\mathcal{L}$.
Let $C$ be a set of hypothesis constraints written in a language $\mathcal{L}$.
A set of definite clauses $H$ is \emph{consistent} with $C$ if, when written in $\mathcal{L}$, $H$ does not violate any constraint in $C$.
We now define the LFF problem, which is based on the ILP learning from entailment setting \cite{luc:book}:
\begin{definition}[\textbf{LFF input}]
\label{def:probin}
A \emph{LFF} input is a tuple $(E^+, E^-, B, D, C)$ where $E^+$ and $E^-$ are sets of ground atoms denoting positive and negative examples respectively; $B$ is a Horn program denoting background knowledge; $D$ is a set of predicate declarations; and $C$ is a set of hypothesis constraints.
\end{definition}
\noindent
A definite program is a \emph{hypothesis} when it is consistent with both $D$ and $C$.
We denote the set of such hypotheses as $\mathcal{H}_{D,C}$.
We define a LFF solution:
\begin{definition}[\textbf{LFF solution}]
\label{def:solution}
Given an input tuple $(E^+, E^-, B, D, C)$, a hypothesis $H \in \mathcal{H}_{D,C}$ is a \emph{solution} when $H$ is \emph{complete} ($\forall e \in E^+, \; B \cup H \models e$) and \emph{consistent} ($\forall e \in E^-, \; B \cup H \not\models e$).
\end{definition}
\noindent
If a hypothesis is not a solution then it is a \emph{failure} and a \emph{failed} hypothesis.
A hypothesis $H$ is \emph{incomplete} when $\exists e^+ \in E^+, \; H \cup B \not \models e^+$.
A hypothesis $H$ is \emph{inconsistent} when $\exists e^- \in E^-, \; H \cup B \models e^-$.
A worked example of LFF is included in Appendix \ref{app:lff_example}.
\subsection{Specialisation and Generalisation Constraints}
\label{sec:constraints}
The key idea of LFF is to learn constraints from failed hypotheses.
\citeauthor{popper} \shortcite{popper} introduce constraints based on subsumption \cite{plotkin:thesis} and theory subsumption \cite{midelfart}.
A clause $C_1$ \emph{subsumes} a clause $C_2$ if and only if there exists a substitution $\theta$ such that $C_1\theta \subseteq C_2$.
A clausal theory $T_1$ subsumes a clausal theory $T_2$, denoted $T_1 \preceq T_2$, if and only if $\forall C_2 \in T_2, \exists C_1 \in T_1$ such that $C_1$ subsumes $C_2$.
Subsumption implies entailment, i.e. if $T_1 \preceq T_2$ then $T_1 \models T_2$.
A clausal theory $T_1$ is a \emph{specialisation} of a clausal theory $T_2$ if and only if $T_2 \preceq T_1$.
A clausal theory $T_1$ is a \emph{generalisation} of a clausal theory $T_2$ if and only if $T_1 \preceq T_2$.
Hypothesis constraints prune the hypothesis space.
\emph{Generalisation} constraints only prune generalisations of inconsistent hypotheses.
\emph{Specialisation} constraints only prune specialisations of incomplete hypotheses.
Generalisation and specialisation constraints are \emph{sound} in that they do not prune solutions \cite{popper}.
\subsection{Missing and Incorrect Answers}
We follow Shapiro \shortcite{mis} in identifying examples as responsible for the failure of a hypothesis $H$ given background knowledge $B$.
A positive example $e^+$ is a \emph{missing answer} when $B \cup H \not\models e^+$.
Similarly, a negative example $e^-$ is an \emph{incorrect answer} when $B \cup H \models e^-$.
We relate missing and incorrect answers to specialisations and generalisations.
If $H$ has a missing answer $e^+$, then each specialisation of $H$ has $e^+$ as a missing answer, so the specialisations of $H$ are incomplete and can be eliminated.
If $H$ has an incorrect answer $e^-$, then each generalisation of $H$ has $e^-$ as an incorrect answer, so the generalisations of $H$ are inconsistent and can be eliminated.
\begin{example}[\textbf{Missing answers and specialisations}]
Consider the following \emph{droplast} hypothesis:
\[
\emph{H$_1$} = \left\{
\begin{array}{l}
\emph{droplast(A,B) $\leftarrow$ empty(A),tail(A,B)}
\end{array}
\right\}
\]
Both $\emph{droplast}([1,2,3],[1,2])$ and $\emph{droplast}([1,2],[1])$ are missing answers of $H_1$, so $H_1$ is incomplete and we can prune its specialisations, e.g.~programs that add literals to the clause.
\end{example}
\begin{example}[\textbf{Incorrect answers and generalisations}]
Consider the hypothesis $H_2$:
\[
\emph{H$_2$} = \left\{
\begin{array}{l}
\emph{droplast(A,B) $\leftarrow$ tail(A,C),tail(C,B)} \\
\emph{droplast(A,B) $\leftarrow$ tail(A,B)}
\end{array}
\right\}
\]
\noindent
In addition to being incomplete, $H_2$ is inconsistent because of the incorrect answer $\emph{droplast}([1,2],[])$, so we can prune the generalisations of $H_2$, e.g.~programs with additional clauses.
\end{example}
\subsection{Failing Sub-programs}
\label{sec:answerstoconstraints}
We now extend LFF by explaining failures in terms of failing sub-programs.
The idea is to identify sub-programs that cause the failure.
Consider the following two examples:
\begin{example}[\textbf{Explain incompleteness}]
\label{ex:explain_failures}
Consider the positive example $e^+ = \emph{droplast}([1,2],[1])$ and the previously defined hypothesis $H_1$.
An explanation for why $H_1$ does not entail $e^+$ is that \emph{empty([1,2])} fails.
It follows that the program
$
\emph{H$_1'$} = \left\{
\begin{array}{l}
\emph{droplast(A,B) $\leftarrow$ empty(A)}\\
\end{array}
\right\}
$
has $e^+$ as a missing answer and is incomplete, so we can prune all specialisations of it.
\end{example}
\begin{example}[\textbf{Explain inconsistency}]
Consider the negative example $e^- = \emph{droplast}([1,2],[])$ and the previously defined hypothesis $H_2$.
The first clause of $H_2$ always entails $e^-$ regardless of other clauses in the hypothesis.
It follows that the program $
\emph{H$_2'$} = \left\{
\begin{array}{l}
\emph{droplast(A,B) $\leftarrow$ tail(A,C),tail(C,B)}\\
\end{array}
\right\}
$ has $e^-$ as an incorrect answer and is inconsistent, so we can prune all generalisations of it.
\end{example}
\noindent
We now define a \emph{sub-program}:
\begin{definition}[\textbf{Sub-program}]
The definite program $P$ is a \emph{sub-program} of the definite program $Q$ if and only if either:
\begin{itemize}
\setlength\itemsep{0pt}
\setlength\parskip{0pt}
\item $P$ is the empty set
\item there exists $C_p \in P$ and $C_q \in Q$ such that $C_p \subseteq C_q$ and $P \setminus \{ C_p \}$ is a sub-program of $Q \setminus \{C_q\}$
\end{itemize}
\end{definition}
\noindent
In functional program synthesis, sub-programs (sometimes called \emph{partial programs}) are typically defined by leaving out nodes in the parse tree of the original program \cite{feng:neo}.
Our definition generalises this idea by allowing for arbitrary ordering of clauses and literals.
We now define the failing sub-programs problem:
\begin{definition}[\textbf{Failing sub-programs problem}]
\label{def:subprogprob}
Given the definite program $P$ and sets of examples $E^+$ and $E^-$, the \emph{failing sub-programs problem} is to find all sub-programs of $P$ that do not entail an example of $E^+$ or entail an example of $E^-$.
\end{definition}
\noindent
By definition, a failing sub-program is incomplete and/or inconsistent, so, by Section \ref{sec:constraints}, we can always prune specialisations and/or generalisations of a failing sub-program.
\begin{remark}[\textbf{Undecidability}]
\label{prop:undecidable}
The failing sub-programs problem is undecidable in general as deciding entailment can be reduced to it.
\end{remark}
\noindent
We show that sub-programs are effective at pruning:
\begin{theorem}[\textbf{Better pruning}]
\label{prop:pruning}
Let $H$ be a definite program that fails and $P$ ($\neq H$) be a sub-program of $H$ that fails.
Specialisation and generalisation constraints for $P$ can always achieve additional pruning versus those only for $H$.
\end{theorem}
\begin{proof}
Suppose $H$ is a specialisation of $P$.
If $P$ is incomplete, then among the specialisations of $P$, which are all prunable, is $H$ and its specialisations.
If $P$ is inconsistent, $P$'s generalisations do not completely overlap with $H$'s generalisations and specialisations (using that $P \neq H$).
Hence, pruning $P$'s generalisations prunes programs not pruned by $H$.
The case where $H$ is a generalisation of $P$ is analogous.
In the remaining case, where $H$ and $P$ are not related by subsumption, it is immediate that the constraints derived for $P$ prune a distinct part of the hypothesis space.
\end{proof}
\iffalse
\subsection{Pruning based on Failing Sub-Programs}
\rolf{The following can be expressed better}
By definition, a failing sub-program of a hypothesis has missing and/or incorrect answers.
Thus, as with failing hypotheses, we can eliminate specialisations and generalisations of failing sub-programs.
Note that in our running example, $H_1$ is a specialisation of both $H_1'$ and $H_1''$ and that specialisations of both $H_1'$ and $H_1''$ can be eliminated.
Similarly, hypothesis $H_1 \cup H_2$ has $H_2$ as failing sub-program, while all generalisations of $H_2$ can be eliminated.
This identification of failing sub-programs also works for recursive hypotheses.
Consider the hypothesis $H_3$:
\[
\emph{H$_3$} = \left\{
\begin{array}{l}
\emph{droplast(A,B) $\leftarrow$ tail(A,B).}\\
\emph{droplast(A,B) $\leftarrow$ tail(A,C),droplast(C,D),tail(D,B).}
\end{array}
\right\}
\]
\noindent
This hypothesis has $e^+ = \emph{droplast}([1,2],[1])$ as a missing answer.
Its sub-program \emph{H$_3'$} also misses out $e^+$:
\[
\emph{H$_3'$} = \left\{
\begin{array}{l}
\emph{droplast(A,B) $\leftarrow$ tail(A,B).}\\
\emph{droplast(A,B) $\leftarrow$ tail(A,C),droplast(C,D).}
\end{array}
\right\}
\]
\noindent
We can therefore eliminate specialisations of $H_3'$, including $H_3$ and its specialisations.
It is important to note that although $H_2'$ is a failing sub-program of $H_3$, $H_3$ is neither a specialisation nor a generalisation of $H_2'$.
In general, a program and its sub-program need not subsume each other.
\subsection{Sub-programs Eliminate More Programs}
\label{sec:smallprunemore}
\rolf{Consider doing an appendix}
\rolf{@AC: this is still a WIP.}
We finish off the theory of failing sub-programs by showing their effectiveness at eliminating programs.
\begin{prop}
Let $P$ be a $n$ literal program and $Q$ a sub-program of $P$ with $m < n$ literals.
Suppose $P$ is a specialisation of $Q$ and that there are $k$ $\emph{body\_pred}$ declarations.
Then $Q$ has at least $k^{n-m} - 1$ specialisations that are not specialisations of $P$.
\rolf{This (almost) completely disregards predicates occurring with different variables}
\end{prop}
\begin{proof}
\rolf{Showing this is not particularly nice. Need to decide what to do here.}
\end{proof}
This same statement and proof also hold for generalisations.
Consider eliminating all specialisations for $P$ and $Q$.
It is not only these $k^{n - m} - 1$ programs that can additionally be eliminated for $Q$.
By transitivity of subsumption, also all specialisations of these $k ^{n - m}$ specialisations can be eliminated, including the specialisations of $P$.
Note that the above focuses on $P$ and $Q$ being related by subsumption.
This is only for demonstration purposes.
In general we have that smaller programs have many more specialisations or generalisations than larger programs.
However, the characterising the increase in the effectiveness of $Q$ over $P$ becomes harder.
\rolf{Incorporate the following}
The Blumer bound (\citeyear{blumer:bound})\footnote{The bound is a reformulation of Lemma 2.1.} helps explain this performance improvement.
This bound states that given two hypothesis spaces, searching the smaller space will result in fewer errors compared to the larger space, assuming that the target hypothesis is in both spaces.
\fi
\section{Implementing Failure Explanation}
\label{sec:application}
We now describe our failure explanation technique, which identifies sub-programs by identifying both clauses and literals within clauses responsible for failure.
Subsequently we summarise the \textsc{Popper}{} ILP system before introducing our extension of it: \name{}.
\subsection{SLD-trees and Sub-programs}
In algorithmic debugging, missing and incorrect answers help characterise which parts of a \emph{debugging tree} are wrong \cite{ADsurvey}.
Debugging trees can be seen as generalising SLD-trees, with the latter representing the search for a refutation \cite{ilp:book}.
Exploiting their granularity, we analyse SLD-trees to address the failing sub-programs problem, only identifying a subset of them.
A \emph{branch} in a SLD-tree is a path from the root \emph{goal} to a leaf.
Each goal on a branch has a \emph{selected atom}, on which resolution is performed to derive child goals.
A branch that ends in an empty leaf is called \emph{successful}, as such a path represents a refutation.
Otherwise a branch is \emph{failing}.
Note that selected atoms on a branch identify a subset of the literals of a program.
Let $B$ be a Horn program, $H$ be a hypothesis, and $e$ be an atom.
The SLD-tree $T$ for $B \cup H \cup \{ \neg e \}$, with $\neg e$ as the root, proves $B \cup H \models e$ iff $T$ contains a successful branch.
Given a branch $\lambda$ of $T$, we define the $\lambda$-sub-program of $H$.
A literal $L$ of $H$ occurs in \emph{$\lambda$-sub-program} $H'$ if and only if $L$ occurs as a selected atom in $\lambda$ or $L$ was used to produce a resolvent that occurs in $\lambda$.
The former case is for literals in the body of clauses and the latter for head literals.
Now consider the SLD-tree $T'$ for $B\cup H' \cup \{ \neg e \}$ with $\neg e$ as root.
As all literals necessary for $\lambda$ occur in $B\cup H'$,
the branch $\lambda$ must occur in $T'$ as well.
Suppose $e^-$ is an incorrect answer for hypothesis $H$.
Then the SLD-tree for $B \cup H \cup \{\neg e^-\}$, with $\neg e^-$ as root, has a successful branch $\lambda$.
The literals of $H$ necessary for this branch are also present in $\lambda$-sub-program $H'$,
hence $e^-$ is also an incorrect answer of $H'$.
Now suppose $e^+$ is a missing answer of $H$.
Let $T$ be the SLD-tree for $B \cup H \cup \{\neg e^+\}$, with $\neg e^+$ as root, and $\lambda'$ be any failing branch of $T$.
The literals of $H$ in $\lambda'$ are also present in $\lambda'$-sub-program $H''$.
This is however insufficient for concluding that the SLD-tree for $H''$ has no successful branch.
Hence it is not immediate that $e^+$ is a missing answer for $H''$.
In case that $H''$ is a specialisation of $H$ we can conclude that $e^+$ is a missing answer.
\subsection{\textsc{Popper}{}}
\label{sec:popper}
\textsc{Popper}{} tackles the LFF problem (Definition \ref{def:probin}) using a \emph{generate}, \emph{test}, and \emph{constrain} loop.
A logical formula is constructed and maintained whose models correspond to Prolog programs.
The first stage is to generate a model and convert it to a program.
The program is tested on all positive and negative examples.
The number of missing and incorrect answers determine whether specialisations%
\footnote{\textsc{Popper}{} generates \emph{elimination constraints} when a hypothesis entails none of the positive examples \cite{popper}.}
and/or generalisations can be pruned.
When a hypothesis fails, new hypothesis constraints (Section \ref{sec:constraints}) are added to the formula, which eliminates models and thus prunes the hypothesis space.
\textsc{Popper}{} then loops back to the generate stage.
Smaller programs prune more effectively, which is partly why \textsc{Popper}{} searches for hypotheses by their size (number of literals)\footnote{The other reason is to find \emph{optimal} solutions, i.e.~those with the minimal number of literals.}%
.
Yet there are many small programs that \textsc{Popper}{} does not consider well-formed that achieve significant, sound pruning.
Consider the sub-program $H_1' = \{~\emph{droplast(A,B) $\leftarrow$ empty(A)}~\}$ from Example \ref{ex:explain_failures}.
\textsc{Popper}{} does not generate $H_1'$ as it does not consider it a well-formed hypothesis (as the head variable $\emph{B}$ does not occur in the body).
Yet precisely because this sub-program has so few body literals is why it is so effective at pruning specialisations.
\subsection{\name{}}
\label{sec:popper_x}
We now introduce \name{}, which extends \textsc{Popper}{} with SLD-based failure explanation.
Like \textsc{Popper}{}, any generated hypothesis $H$ is tested on the examples.
However, additionally, for each tested example we obtain the selected atoms on each branch of the example's SLD-tree, which correspond to sub-programs of $H$.
As shown, sub-programs derived from incorrect answers have the same incorrect answers.
For each such identified inconsistent sub-program $H'$ of $H$ we tell the constrain stage to prune generalisations of $H'$.
Sub-programs derived from missing answers are retested, now without obtaining their SLD-trees.
If a sub-program $H''$ of $H$ is incomplete we inform the constrain stage to prune specialisations%
\footnote{As in \textsc{Popper}{}, we prune by elimination constraints if no positive examples are entailed.}
of $H''$.
Pruning for sub-programs is in addition to the pruning that the constrain stage already does for $H$.
This is important as $H$'s failing sub-programs need not be specialisations/generalisations of $H$.
\subsection{Experiment 1: Robot Planning}
\label{sec:expr-robots}
The goal of this experiment is to evaluate whether failure explanation can improve performance when progressively increasing the size of the target program.
We therefore need a problem where we can vary the program size.
We consider a robot strategy learning problem.
There is a robot that can move in four directions in a grid world, which we restrict to being a corridor (dimensions $1 \times 10$).
The robot starts in the lower left corner and needs to move to a position to its right.
In this experiment, failure explanation should determine that any strategy that moves up, down, or left can never succeed and thus can never appear in a solution.
\paragraph{Settings.}
An example is an atom $f(s_1,s_2)$, with start ($s_1$) and end ($s_2$) states.
A state is a pair of discrete coordinates $(x,y)$.
We provide four dyadic relations as BK: $\emph{move\_right}$, $\emph{move\_left}$, $\emph{move\_up}$, and $\emph{move\_down}$, which change the state, e.g.~$\emph{move\_right((2,2),(3,2))}$.
We allow one clause with up to 10 body literals and 11 variables.
We use hypothesis constraints to ensure this clause is forward-chained \cite{hexmil}, which means body literals modify the state one after another.
\paragraph{Method.}
The start state is $(0,0)$ and the end state is $(n,0)$, for $n$ in $1,2,3,\ldots,10$.
Each trial has only one (positive) example: $f((0,0),(n,0))$.
We measure learning times and the number of programs generated.
We enforce a timeout of 10 minutes per task.
We repeat each experiment 10 times and plot the mean and (negligible) standard error.
\paragraph{Results.}
Figure \ref{fig:robots-time} shows that \name{} substantially outperforms \popper{} in terms of learning time.
Whereas \name{} needs around 80 seconds to find a 10 move solution, \popper{} exceeds the 10 minute timeout when looking for a six move solution.
The reason for the improved performance is that \name{} generates far fewer programs, as failure explanation will, for example, prune all programs whose first move is to the left.
For instance, to find a five literal solution, \popper{} generates 1300 programs, whereas \name{} only generates 62.
When looking for a 10 move solution, \name{} only generates 1404 programs in a hypothesis space of 1.4 million programs.
These results show that, compared to \popper{}, \name{} generates substantially fewer programs and requires less learning time.
The results from this experiment strongly suggest that the answer to questions \textbf{Q1} and \textbf{Q2} is yes.
\begin{figure}[ht]
\small
\centering
\begin{subfigure}[b]{0.225\textwidth}
\centering
\pgfplotsset{every tick label/.append style={font=\Large}}
\begin{tikzpicture}[scale=.48]
\begin{axis}[
xlabel=Program size,
ylabel=Learning time (seconds),
xmin=1,xmax=10,
ymin=0,ymax=600,
ylabel style={yshift=-2mm},
label style={font=\LARGE},
legend style={at={(0.42,0.97)},style={font=\Large,nodes={right}}}
]
\addplot[blue,mark=*,mark options={fill=blue},error bars/.cd,y dir=both,y explicit]
table [
x=size,
y=time,
y error plus expr=\thisrow{error},
y error minus expr=\thisrow{error},
] {
size time error
1 0.1 0.0
2 0.15 0.0
3 0.2 0.0
4 0.33 0.0
5 0.55 0.0
6 1.14 0.01
7 2.71 0.02
8 7.93 0.04
9 21.5 0.16
10 80.61 1.63
};
\addplot[black,mark=+,error bars/.cd,y dir=both,y explicit]
table [
x=size,
y=time,
y error plus expr=\thisrow{error},
y error minus expr=\thisrow{error},
] {
size time error
1 0.1 0.0
2 0.19 0.0
3 0.4 0.0
4 3.75 0.03
5 47.84 0.59
6 600.14 0.02
7 1000.14 0.02
};
\legend{\name{}, baseline}
\end{axis}
\end{tikzpicture}
\caption{
Learning time.
}
\label{fig:robots-time}
\end{subfigure}
\hspace{1ex}
\begin{subfigure}[b]{0.225\textwidth}
\centering
\pgfplotsset{every tick label/.append style={font=\Large}}
\begin{tikzpicture}[scale=.48]
\begin{axis}[
xlabel=Program size,
ylabel=Generated programs,
xmin=1,xmax=10,
ymin=0,ymax=6000,
ylabel style={yshift=1mm},
label style={font=\LARGE},
legend style={at={(0.43,0.97)},style={font=\Large,nodes={right}}}
]
\addplot[blue,dashed,mark=*,mark options={fill=blue},error bars/.cd,y dir=both,y explicit]
table [
x=size,
y=progs,
y error plus expr=\thisrow{error},
y error minus expr=\thisrow{error},
] {
size progs error
1 2.0 0.0
2 10.0 0.0
3 18.0 0.0
4 36.0 0.0
5 62.0 0.0
6 117.0 0.0
7 214.0 0.0
8 400.0 0.0
9 688.0 0.0
10 1404.0 0.0
};
\addplot[black,dashed,mark=+,error bars/.cd,y dir=both,y explicit]
table [
x=size,
y=progs,
y error plus expr=\thisrow{error},
y error minus expr=\thisrow{error},
] {
size progs error
1 2.0 0.0
2 20.0 0.0
3 57.0 0.0
4 319.0 0.0
5 1300.0 0.0
6 4673.0 7.67
7 8500 0
};
\legend{\name{},baseline}
\end{axis}
\end{tikzpicture}
\caption{
Number of programs.
}
\label{fig:robots-generated}
\end{subfigure}
\caption{Results of robot planning experiment.
The x-axes denote the number of body literals in the solution, i.e.~the number of moves required.
}
\label{fig:robots}
\end{figure}
\subsection{Experiment 3: Programming Puzzles}
\label{expr:lists}
This experiment evaluates whether failure explanation can improve performance when learning programs for recursive list problems, which are notoriously difficult for ILP systems.
Indeed, other state-of-the-art ILP system \cite{ilasp3,dilp,hexmil} struggle to solve these problems.
We use the same 10 problems used by \cite{popper} to show that \popper{} drastically outperforms \textsc{Metagol} \cite{metaho} and \textsc{Aleph} \cite{aleph}.
The 10 tasks include a mix of monadic (e.g.~\emph{evens} and \emph{sorted}), dyadic (e.g.~\emph{droplast} and \emph{finddup}), and triadic (\emph{dropk}) target predicates.
Some problems are functional (e.g.~\emph{last} and \emph{len}) and some are relational (e.g.~\emph{finddup} and \emph{member}).
\paragraph{Settings.}
We provide as BK the monadic relations \emph{empty}, \emph{zero}, \emph{one}, \emph{even}, \emph{odd},
the dyadic relations \emph{element}, \emph{head}, \emph{tail}, \emph{increment}, \emph{decrement}, \emph{geq}, and the triadic relation \emph{cons}.
We provide simple types and mark the arguments of predicates as either input or output.
We allow up to two clauses with five body literals and up to five variables per clause.
\paragraph{Method.}
We generate 10 positive and 10 negative examples per problem.
Each example is randomly generated from lists up to length 50, whose integer elements are sampled from 1 to 100.
We test on a 1000 positive and a 1000 negative randomly sampled examples.
We measure overall learning time, number of programs generated, and predictive accuracy.
We also measure the time spent in the three distinct stages of \popper{} and \name{}.
We repeat each experiment 25 times and record the mean and standard error.
\paragraph{Results.}
\begin{table}[ht]
\newcommand\narrowc{@{\hskip4.08pt}c@{\hskip4.08pt}}
\scriptsize
\centering
\begin{tabular}{
c
|@{\hskip 1mm}|
\narrowc|\narrowc|\narrowc
|@{\hskip 1mm}|
\narrowc|\narrowc|\narrowc
}
\multicolumn{1}{c}{} &
\multicolumn{3}{c}{\textbf{Number of programs}} &
\multicolumn{3}{c}{\textbf{Learning time} (seconds)}
\\
\toprule
\textbf{Name} &
\popper{} & \name{} & \textbf{ratio} &
\popper{} & \name{} & \textbf{ratio} \\
\midrule
len &
590 $\pm$ 4 & 60 $\pm$ 5 & \textbf{0.10} &
16 $\pm$ 0.2 & 1 $\pm$ 0.1 & \textbf{0.09} \\
dropk &
114 $\pm$ 0.7 & 13 $\pm$ 1 & \textbf{0.12} &
1 $\pm$ 0.01 & 0.3 $\pm$ 0.02 & \textbf{0.23} \\
finddup &
1223 $\pm$ 22 & 644 $\pm$ 14 & \textbf{0.53} &
53 $\pm$ 2 & 20 $\pm$ 0.7 & \textbf{0.38} \\
member &
57 $\pm$ 2 & 14 $\pm$ 0.7 & \textbf{0.24} &
0.6 $\pm$ 0.03 & 0.2 $\pm$ 0.01 & \textbf{0.41} \\
last &
232 $\pm$ 6 & 64 $\pm$ 6 & \textbf{0.28} &
3 $\pm$ 0.1 & 2 $\pm$ 0.1 & \textbf{0.48} \\
evens &
306 $\pm$ 2 & 278 $\pm$ 2 & \textbf{0.91} &
7 $\pm$ 0.06 & 7 $\pm$ 0.09 & \textbf{1.00} \\
threesame &
18 $\pm$ 4 & 15 $\pm$ 3 & \textbf{0.81} &
0.2 $\pm$ 0.04 & 0.2 $\pm$ 0.04 & \textbf{1.00} \\
droplast &
161 $\pm$ 9 & 148 $\pm$ 10 & \textbf{0.92} &
11 $\pm$ 0.9 & 11 $\pm$ 1 & 1.02 \\
addhead &
32 $\pm$ 3 & 31 $\pm$ 2 & \textbf{0.96} &
0.6 $\pm$ 0.04 & 0.6 $\pm$ 0.04 & 1.05 \\
sorted &
708 $\pm$ 40 & 599 $\pm$ 26 & \textbf{0.85} &
29 $\pm$ 3 & 31 $\pm$ 2 & 1.08 \\
\bottomrule
\end{tabular}
\caption{
Results for programming puzzles.
Left, the average number of programs generated by each system.
Right, the corresponding average time to find a solution.
We round values over one to the nearest integer.
The error is standard error.
}
\label{tab:puzzles}
\end{table}
Both systems are equally accurate, except on \emph{sorted} where \popper{} scores 98\% and \name{} 99\%.
Accuracy is 98\% on \emph{dropk} and 99\% on both \emph{finddup} and \emph{threesame}.
All other problems have 100\% accuracy.
Table \ref{tab:puzzles} shows the learning times in relation to the number of programs generated.
Crucially, it includes the ratio of the mean of \name{} over the mean of \popper{}.
On these 10 problems, \name{} always considers fewer hypotheses than \popper{}.
Only on three problems is over 90\% of the original number of programs considered.
On the \emph{len} problem, \name{} only needs to consider 10\% of the number of hypotheses.
As seen from the ratio columns, the number of generated programs correlates strongly with the learning time (0.96 correlation coefficient).
Only on three problems is \name{} slightly slower than \popper{}.
Hence \popper{} can be negatively impacted by failure explanation, however, when \name{} is faster, the speed-up can be considerable.
To illustrate how failure explanation can drastically improve pruning, consider the following hypothesis that \name{} considers in the \emph{len} problem:
\begin{center}
\emph{f(A,B):- element(A,D),odd(D),even(D),tail(A,C),element(C,B)}.
\end{center}
\noindent
Failure explanation identifies the failing sub-program:
\begin{center}
\emph{f(A,B):- element(A,D),odd(D),even(D).}
\end{center}
\noindent
As should be hopefully clear, generating constraints from this smaller failing program, which is \emph{not} a \textsc{Popper}{} hypothesis, leads to far more effective pruning.
Figure \ref{fig:relative_stages} shows the relative time spent in each stage of \name{} and \textsc{Popper}{} and that any of the stages can dominate the runtime.
For \emph{addhead}, it is hypothesis generation (predominantly spent searching for a model).
For \emph{finddup}, it is constraining (mostly spent grounding constraints).
More pertinently, \emph{droplast}, the only dyadic problem whose output is a list, is dominated by testing.
We can also infer the overhead of failure explanation by analysing SLD-trees from Figure \ref{fig:relative_stages}.
All problems from \emph{last} to \emph{sorted} have \name{} spend more time on testing than \textsc{Popper}{}.
On both \emph{last} and \emph{sorted}, \name{} incurs considerable testing overhead.
Whilst for \emph{last} this effort translates into more effective pruning constraints, for \emph{sorted} this is not the case.
Abstracting away from the implementation of failure explanation,
we see that \popper{} outfitted with zero-overhead sub-program identification would have been strictly faster.
Overall, these result strongly suggest that the answer to questions \textbf{Q1} and \textbf{Q2} is yes.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{lists_plot.png}
\vspace{-5ex}
\caption{
Relative time spent in three stages of \name{} and \popper{}.
From bottom to top: testing (in red), generating hypotheses (in blue), and imposing constraints (in orange).
Times are scaled by the total learning time of \popper{}, with \popper{}'s average time(s) on the left and \name{}'s on the right.
Bars are standard error.
}
\label{fig:relative_stages}
\end{figure}
\subsection{Experiment 2: String Transformations}
\label{sec:strings}
We now explore whether failure explanation can improve learning performance on real-world string transformation tasks.
We use a standard dataset \cite{metabias,playgol} formed of 312 tasks, each with 10 input-output pair examples.
For instance, task 81 has the following two input-output pairs:
\begin{center}
\begin{tabular}{l|l}
\textbf{Input} & \textbf{Output} \\
\hline
``Alex'',``M'',41,74,170 & M \\
``Carly'',``F'',32,70,155 & F
\end{tabular}
\end{center}
\paragraph{Settings.}
As BK, we give each system the monadic predicates \emph{is\_uppercase}, \emph{is\_empty}, \emph{is\_space}, \emph{is\_letter}, \emph{is\_number} and dyadic predicates \emph{mk\_uppercase}, \emph{mk\_lowercase}, \emph{skip1}, \emph{copyskip1}, \emph{copy1}.
For each monadic predicate we also provide a predicate that is its negation.
We allow up to 3 clauses with 4 body literals and up to 5 variables per clause.
\paragraph{Method.}
The dataset has 10 positive examples for each problem.
We perform cross validation by selecting 10 distinct subsets of 5 examples for each problem, using the other 5 to test.
We measure learning times and number of programs generated.
We enforce a timeout of 120 seconds per task.
We repeat each experiment 10 times, once for each distinct subset, and record means and standard errors.
\paragraph{Results.}
\begin{figure}[ht]
\centering
\begin{tikzpicture}[scale=0.60]
\begin{axis}[
xlabel=Ratio of generated programs,
ylabel=Ratio of learning time,
ylabel style={yshift=-3mm},
xlabel style={yshift=1mm},
xmin=0, xmax=1.3,
ymin=0, ymax=1.3
]
\addplot [only marks,mark=*] table {
0.18 0.14
0.32 0.28
0.35 0.18
0.30 0.25
0.94 0.91
0.43 0.41
0.30 0.26
0.90 0.88
0.34 0.31
0.32 0.27
0.32 0.28
0.32 0.28
0.32 0.28
0.32 0.28
0.32 0.28
0.32 0.27
0.88 0.88
0.33 0.28
0.25 0.22
0.32 0.28
0.32 0.28
0.42 0.39
0.26 0.21
0.41 0.22
0.23 0.22
0.32 0.28
0.17 0.12
0.32 0.28
0.32 0.28
0.32 0.28
0.34 0.31
0.09 0.05
0.32 0.29
0.37 0.19
0.32 0.27
1.00 1.07
0.32 0.28
0.35 0.32
0.32 0.28
0.32 0.29
0.48 0.28
0.32 0.28
1.10 0.93
0.35 0.18
0.26 0.21
0.34 0.31
0.41 0.38
0.38 0.22
0.94 0.75
0.32 0.28
0.21 0.17
0.32 0.28
};
\addplot[dashed, samples=100, smooth,domain=0:1] {x};
\end{axis}
\end{tikzpicture}
\caption{String transformation results.
The ratio of number of programs that \name{} needs versus \popper{} is plotted against the ratio of learning time needed on that problem.}
\label{fig:strings}
\end{figure}
For 52 problems both \popper{} and \name{} find solutions%
\footnote{Note that these problems are very difficult with many of them not having solutions given only our primitive BK and with the learned program restricted to defining a single predicate. Therefore, absolute performance should be ignored. The important result is the relative performance of the two systems.
}%
.
On 11 tasks \popper{} timeouts, and on 7 of these in all trials.
\name{} finds solutions on these same 11 tasks, with timeouts in some trials on only 6 tasks.
As relational solutions are allowed, many solutions are not ideal, e.g.~allowing for optionally copying over a character.
Figure \ref{fig:strings} plots ratios of generated programs and learning times.
Each point represents a single problem.
The x-axis is the ratio of programs that \name{} generates versus the number of programs that \popper{} generates.
The y-value is the ratio of learning time of \name{} versus \popper{}.
These ratios are acquired by dividing means, the mean of \name{} over that of \popper{}.
Looking at x-axis values, of the 52 problems plotted 50 require fewer programs when run with \name{}.
Looking at the y-axis, the learning times of 51 problems are faster on \name{}.
Note that either failure explanation is very effective or its influence is rather limited, which we explore more in the next experiment.
Overall, these results show that, compared to \popper{}, \name{} almost always needs fewer programs and less time to learn programs.
This suggests that the answer to questions \textbf{Q1} and \textbf{Q2} is yes.
\iffalse
Both systems achieve the same predictive accuracy, except for one problem where \popper{} has two timeouts while \name{} has none.
The mean learning time of \name{} is always lower than that of \popper{}.
On average, \name{} needs 15 seconds while \popper{} requires 22 seconds.
The mean number of programs that \name{} generates is 140 versus 193 for \popper{}.
Figure \ref{fig:strings} plots ratios of generated programs and mean learning times.
Each point represents a single problem.
The x-axis is the ratio of programs that \name{} generates versus the number of programs that \popper{} generates.
The y-value is the ratio of learning time of \name{} versus \popper{}.
These ratios are acquired by dividing means, the mean of \name{} over that of \popper{}.
Looking at x-axis values, the 31 problems plotted all require generating less programs when run with \name{}.
Looking at the y-axis, the learning times are also always lower for \name{}.
One of the 32 problems does not have a data point in figure \ref{fig:strings}.
For this problem \name{} generates more programs (mean of 43 versus 31), as the additional constraints steer the search to other hypotheses first.
Nevertheless, \name{}'s mean learning time of 2.21 seconds is faster than \popper{}'s 2.51 seconds.
We conclude that Figure \ref{fig:strings} does show a strong correlation of generating less programs and requiring less learning time.
Overall, these results show that, compared to \popper{}, \name{} almost always needs fewer programs and less time to learn programs, which suggests that the answer to questions \textbf{Q1} and \textbf{Q2} is yes.
\fi
\section{Experiments}
\label{sec:experiments}
We claim that failure explanation can improve learning performance.
Our experiments therefore aim to answer the questions:
\begin{itemize}
\item[\textbf{Q1}] Can failure explanation prune more programs?
\item[\textbf{Q2}] Can failure explanation reduce learning times?
\end{itemize}
\noindent
A positive answer to \textbf{Q1} does not imply a positive answer for \textbf{Q2} because of the potential overhead of failure explanation.
Identifying sub-programs requires computational effort and the additional constraints could potentially overwhelm a learner.
For example, as well as identifying sub-programs, \name{} needs to derive more constraints, ground them, and have a solver reason over them.
These operations are all costly.
To answer \textbf{Q1} and \textbf{Q2}, we compare \name{} against \textsc{Popper}{}.
The addition of failure explanation is the only difference between the systems and in all the experiments the settings for \name{} and \textsc{Popper}{} are identical.
We do not compare against other state-of-the-art ILP systems, such as Metagol \cite{metaho} and ILASP3 \cite{ilasp3} because such a comparison cannot help us answer \textbf{Q1} and \textbf{Q2}.
Moreover, \textsc{Popper}{} has been shown to outperform these two systems on problems similar to the ones we consider \cite{popper}.
We run the experiments on a 10-core server (at 2.2GHz) with 30 gigabytes of memory (note that \textsc{Popper}{} and \name{} only run on a single CPU).
When testing individual examples, we use an evaluation timeout of 33 milliseconds.
\input{05-01-robots}
\input{05-03-strings}
\input{05-02-lists}
\section{Conclusions and Limitations}
\label{sec:conclusion}
To improve the efficiency of ILP, we have introduced an approach for failure explanation.
Our approach, based on SLD-trees, identifies failing sub-programs, including failing literals in clauses.
We implemented our idea in \textsc{Popper}\textsubscript{$X$}{}.
Our experiments show that failure explanation can drastically reduce learning times.
\paragraph{Limitations.}
We have shown that identifying failing sub-programs will lead to more constraints and thus more pruning of the hypothesis space (Theorem \ref{prop:pruning}), which our experiments empirically confirm.
We have not, however, quantified the theoretical effectiveness of pruning by sub-programs, nor have we evaluated improvements in predictive accuracy, which are implied by the Blumer bound \cite{blumer:bound}.
Future work should address both of these limitations.
Although we have shown that failure explanation can drastically reduce learning times, we can still significantly improve our approach.
For instance, reconsider the failing sub-program \emph{f(A,B):- element(A,D),odd(D),even(D)} from Section \ref{expr:lists}.
We should be able to identify that the two literals \emph{odd(D)} and \emph{even(D)} can never both hold in the body of a clause, which would allow us to prune more programs.
Finally, in future work, we want to explore whether our inherently interpretable failure explanations can aid explainable AI and \emph{ultra-strong} machine learning \cite{usml,mugg:compmlj}.
\section{Appendix: LFF Example}
\label{app:lff_example}
\begin{example}
\label{ex:lff_example}
To illustrate LFF, consider learning a \emph{droplast/2} program.
Suppose our predicate declarations $D$ are \emph{head\_pred(droplast,2)}, denoting that we want to learn a \emph{droplast/2} relation, and \emph{body\_pred(empty,1)}, \emph{body\_pred(head,2)}, \emph{body\_pred(tail,2)}, and \emph{body\_pred(cons,3)}.
Suitable definitions for the provided body predicate declarations constitute our background knowledge $B$.
To allow for learning a recursive program, we also supply the predicate declaration \emph{body\_pred(droplast,2)}.
Let $e_1^+ = \emph{droplast}([1,2,3],[1,2])$,
$e_2^+ = \emph{droplast}([1,2],[1])$
and
$e_1^- = \emph{droplast}([1,2],[])$.
Then $E^+ = \{ e_1^+, e_2^+ \}$ and $E^- = \{ e_1^- \}$ are the positive and negative examples, respectively.
Our initial set of hypothesis constraints $C$ only ensure that hypotheses are well-formed,
e.g.~that each variable that occurs in the head of a rule occurs in the rule's body.
\begin{figure*}[ht!]
\centering
\begingroup
\newcommand{\hypo}[2]{{
\tw{#1} = \left\{
\begin{array}{l}
#2
\end{array}
\right\}
}\\
}
\[
\mathcal{H}_1 = \left\{
\begin{array}{l}
\hypo{h$_1$}{
\tw{droplast(A,B):- empty(A),tail(A,B).}\\
}
\hypo{h$_2$}{
\tw{droplast(A,B):- empty(A),cons(C,D,A),tail(D,B).}\\
}
\hypo{h$_3$}{
\tw{droplast(A,B):- tail(A,C),tail(C,B).}\\
\tw{droplast(A,B):- tail(A,B).}\\
}
\hypo{h$_4$}{
\tw{droplast(A,B):- empty(A),tail(A,B),head(A,C),head(B,C).}\\
}
\hypo{h$_5$}{
\tw{droplast(A,B):- tail(A,C),tail(C,B).}\\
\tw{droplast(A,B):- tail(A,B),tail(B,A).}\\
}
\hypo{h$_6$}{
\tw{droplast(A,B):- tail(A,B),empty(B).}\\
\tw{droplast(A,B):- cons(C,D,A),droplast(D,E),cons(C,E,B).}\\
}
\hypo{h$_7$}{
\tw{droplast(A,B):- tail(A,C),tail(C,B).}\\
\tw{droplast(A,B):- tail(A,B).}\\
\tw{droplast(A,B):- tail(A,C),droplast(C,B).}\\
}
\end{array}
\right\}
\]
\endgroup
\caption{LFF hypothesis space considered in Example \ref{ex:lff_example}.}
\label{fig:hypospace}
\end{figure*}
We now consider learning a solution for LFF input $(E^+, E^-, B, D, C)$, where, for demonstration purposes, we use the simplified hypothesis space $\mathcal{H}_1 \subseteq \mathcal{H}_{D,C}$ of figure \ref{fig:hypospace}.
The order the hypotheses are considered in is by their number of literals.
Pruning is achieved by adding additional hypothesis constraints.
First we learn by a generate-test-and-constrain loop \emph{without} failure explanation.
This first sequence is representative of \textsc{Popper}{}'s execution:
\begin{enumerate}
\item
\textsc{Popper}{} starts by generating \tw{h$_1$}.
$B \cup \tw{h$_1$}$ fails to entail $e_1^+$ and $e_2^+$ and correctly does not entail $e_1^-$.
Hence only specialisations of \tw{h$_1$} get pruned, namely \tw{h$_4$}.
\item
\textsc{Popper}{} subsequently generates \tw{h$_2$}.
$B \cup \tw{h$_2$}$ fails to entail $e_1^+$ and $e_2^+$ and is correct on $e_1^-$.
Hence specialisations of \tw{h$_2$} get pruned, of which there are none in $\mathcal{H}_1$.
\item
\textsc{Popper}{} subsequently generates \tw{h$_3$}.
$B \cup \tw{h$_3$}$ does not entail the positive examples, but does entail negative example $e_1^-$.
Hence specialisations and generalisations of \tw{h$_3$} get pruned, meaning only generalisation \tw{h$_7$}.
\item
\textsc{Popper}{} subsequently generates \tw{h$_5$}.
$B \cup \tw{h$_5$}$ is correct on none of the examples.
Hence specialisations and generalisations of \tw{h$_5$} get pruned, of which there are none in $\mathcal{H}_1$.
\item
\textsc{Popper}{} subsequently generates \tw{h$_6$}.
$B \cup \tw{h$_6$}$ is correct on all the examples and hence is returned.
\end{enumerate}
Now consider learning by a generate-test-and-constrain loop \emph{with} failure explanation.
The following execution sequence is representative of \textsc{Popper}\textsubscript{$X$}{}:
\begin{enumerate}
\item
\textsc{Popper}\textsubscript{$X$}{} starts by generating \tw{h$_1$}.
$B \cup \tw{h$_1$}$ fails to entail $e_1^+$ and $e_2^+$ and correctly does not entail $e_1^-$.
Failure explanation identifies sub-program $\tw{h$_1'$} = \{ \tw{droplast(A,B):- empty(A).} \}$.
\tw{h$_1'$} fails in the same way as \tw{h$_1$}.
Hence specialisations of both \tw{h$_1$} and \tw{h$_1'$} get pruned, namely \tw{h$_2$} and \tw{h$_4$}.
\item
\textsc{Popper}\textsubscript{$X$}{} subsequently generates \tw{h$_3$}.
$B \cup \tw{h$_3$}$ does not entail the positive examples, but does entail negative example $e_1^-$.
Failure explanation identifies sub-program $\tw{h$_3'$} = \{ \tw{droplast(A,B):- tail(A,C),tail(C,B).} \}$.
$B \cup \tw{h$_3'$}$ fails in the same way as \tw{h$_3$}.
Hence specialisations and generalisations of \tw{h$_3$} and \tw{h$_3'$} get pruned, meaning \tw{h$_5$} and \tw{h$_7$}.
\item
\textsc{Popper}\textsubscript{$X$}{} subsequently generates \tw{h$_6$}.
$B \cup \tw{h$_6$}$ is correct on all the examples and hence is returned.
\end{enumerate}
The difference in these two execution sequences is illustrative of how failure explanation can help prune away significant parts of the hypothesis space.
\end{example}
\end{document}
|
\subsubsection*{References}}
\usepackage{mathtools}
\usepackage{booktabs}
\usepackage{tikz}
\usepackage{amsmath, amssymb}
\usepackage{subcaption}
\usepackage[capitalize]{cleveref}
\usepackage{pgfplots, pgfplotstable}
\usetikzlibrary{pgfplots.groupplots}
\pgfplotsset{compat=1.16}
\usetikzlibrary{shapes.arrows, fadings, shapes}
\usetikzlibrary{arrows}
\usetikzlibrary{positioning}
\usetikzlibrary{patterns}
\definecolor{nice-red}{HTML}{E41A1C}
\definecolor{nice-orange}{HTML}{FF7F00}
\definecolor{nice-yellow}{HTML}{FFC020}
\definecolor{nice-green}{HTML}{4DAF4A}
\definecolor{nice-blue}{HTML}{377EB8}
\definecolor{nice-purple}{HTML}{984EA3}
\definecolor{nice-grey}{HTML}{6C7A89}
\definecolor{nice-pink}{HTML}{DB5A6B}
\newcommand\blfootnote[1]{%
\begingroup
\renewcommand\thefootnote{}\footnote{#1}%
\addtocounter{footnote}{-1}%
\endgroup
}
\newcommand{\todo}[1]{\textcolor{red}{[\textbf{TODO:} #1]}}
\newcommand\argmax{\ensuremath{\arg\max}}
\DeclareMathAlphabet{\mathcal}{OMS}{cmsy}{m}{n}
\title{Learning Emergent Discrete Message Communication \\ for Cooperative Reinforcement Learning}
\author[1]{Sheng~Li}
\author[2]{Yutai~Zhou}
\author[2]{Ross~Allen}
\author[1]{Mykel~J.~Kochenderfer}
\affil[1]{%
Aero/Astro Engineering\\
Stanford University\\
Stanford, CA, USA
}
\affil[2]{%
MIT Lincoln Lab\\
Lexington, MA, USA\\
}
\begin{document}
\maketitle
\input{sections/0_abstract}
\input{sections/1_intro}
\input{sections/2_background}
\input{sections/3_approach}
\input{sections/4_experiments}
\input{sections/5_conclusion}
\section{Introduction}\label{sec:intro}
Communication allows agents to share information so that they can perform tasks cooperatively. There has been existing work on using deep reinforcement learning (RL) to produce communication protocols. For example, CommNet~\citep{sukhbaatar2016learning}, a recurrent communication model, averages the hidden states for centralized communication.
IC3Net~\citep{singh2018learning}, an extension on CommNet, adopts a more complicated but similar centralized aggregation approach to communication.
Instead of centralized aggregation and averaging, TarMAC~\citep{das2019tarmac} use multi-headed attention to distribute information to other agents.
BiCNet~\citep{peng2017multiagent} and ATOC~\citep{jiang2018learning} both use a bidirectional recurrent network as a communication channel. They fix the positions of agents in the bidirectional recurrent network to specify their roles.
DICG~\citep{li2020deep} uses graph convolution to implicitly pass information between agents.
However, typically, existing multi-agent communication approaches use continuous messages to communicate. They use real-valued vectors to encode messages. Human languages, however, use discrete characters and words. An advantage of continuous messaging is its representational capacity, but it can be at the expense of interpretability from the perspective of human designers or other agents.
We propose a deep RL model for agents to learn to generate their own discrete message protocols. Our model produces discrete messages by identifying the maximum element in message vectors, resulting in greater stability than sampling. The model adopts a broadcast-and-listen procedure to send and receive messages. It uses self-attention mechanism~\citep{cheng2016long} to aggregate messages sent by other agents. The model is differentiable and therefore can be learned end-to-end. \citet{evtimova2017emergent} use bit-string messaging to learn emergent communication in referential games for two agents. Our approach is applicable to any number of agents.
We compare the performance of discrete message communication with continuous message communication in a variety of domains, showing that discrete message communication has comparable performance to continuous message communication with a much smaller vocabulary size. We also study the effects of communication bandwidth and vocabulary size on discrete message communication, using the metrics positive listening and positive signaling, where positive listening indicates received messages are influencing agents’ behaviors in some way, and positive signaling indicates an agent is sending messages that are related in some way with its own observations or actions~\citep{lowe2019pitfalls, jaques2019social}.
Furthermore, we propose an approach for human-agent interaction using discrete message communication, demonstrating its interpretability.
\section{Background}\label{sec:bg}
We represent the problem as a Dec-POMDP~\citep{oliehoek2016concise} defined by the tuple $\langle \mathcal{I}, \mathcal{S}, \{\mathcal{A}^i\}_{i=1}^n, \mathcal{V}, \mathcal{T}, \mathcal{Z}, R, \mathcal{O}, \gamma\rangle$, where $\mathcal{I} = \{1, \ldots, n\}$ is the set of agents, $\mathcal{S}$ is the global state space, $\mathcal{A}^i$ is the action space of the $i$th agent, and $\mathcal{Z}$ is the observation space for an agent.
The discrete communication vocabulary set is defined by $\mathcal{V} = \{0, 1\}^b$, where $b$ is the band width of communication. A message from $\mathcal{V}$ is therefore a binary vector.
The transition function defining the next state distribution is given by $\mathcal{T}: \mathcal{S} \times \prod_i \mathcal{A}^i \times \mathcal{S} \to [0,1]$.
The reward function is $R: \mathcal{S} \times \prod_i \mathcal{A}^i \to \mathbb{R}$, and the discount factor is $\gamma \in [0, 1)$.
The observation model defining the observation distribution from the current state is $\mathcal{O}: \mathcal{S} \times \mathcal{Z} \to [0, 1]$.
Each agent $i$ has a stochastic policy $\pi^i$ conditioned on its observations $o_i$.
The discounted return is $G_t = \sum_{l=0}^\infty \gamma^{l} r_{t+l}$, where $r_t$ is the joint reward at step $t$.
The joint policy $\mathbf{\pi}$ induces a value function $V^{\mathbf{\pi}}(s_t) = \mathbb{E}[G_t \mid s_t]$ and an action-value function $Q^\mathbf{\pi}(s_t, \mathbf{a}_t) = \mathbb{E}[G_t \mid s_t, \mathbf{a}_t]$, where $\mathbf{a}_t$ is the joint action.
The advantage function is $A^\mathbf{\pi}(s_t, \mathbf{a}_t)=Q^\mathbf{\pi}(s_t, \mathbf{a}_t) - V^{\mathbf{\pi}}(s_t)$.
\subsection{Policy Optimization}\label{sec:ppo}
We use policy optimization to maximize the expected discounted return.
Given policy $\pi_\theta$ parameterized by $\theta$, the surrogate policy optimization objective is~\citep{schulman2017proximal}:
\begin{equation}
\underset{\theta}{\mathrm{maximize}} \quad \hat{\mathbb{E}}_t\left[\frac{\pi_{\theta}(a_t \mid s_t)}{\pi_{\theta_{\text{old}}}(a_t \mid s_t)} \hat{A}_t\right],
\end{equation}
where we use the generalized advantage estimation (GAE)~\citep{schulman2015high} to estimate $\hat{A}_t$ at time step $t$, and the expectation $\hat{\mathbb{E}}_t[\cdot]$ indicates the empirical average over a finite batch of samples.
In practice, we use the clipped PPO objective \citep{schulman2017proximal} to limit the step size for stable updates. The entropy bonus is also added to encourage exploration. The objective to maximize becomes
\begin{equation}\label{eq:ppo_obj}
\small
\begin{aligned}
L^{\mathrm{PPO}}_t(\theta) = \hat{\mathbb{E}}_t \Big[&\min\big(r_t(\theta) \hat{A}_t,\,\, \mathrm{clip}(r_t(\theta), 1 - \epsilon, 1 + \epsilon)\hat{A}_t \big) \\
&+ \beta H[\pi_\theta](s_t)\Big],
\end{aligned}
\end{equation}
where $r_t(\theta) = \frac{\pi_{\theta}(a_t \mid s_t)}{\pi_{\theta_{\text{old}}}(a_t \mid s_t)}$, and $H[\pi_\theta](s_t)$ is the entropy of the policy given state $s_t$, the clipping parameter $\epsilon$ and the entropy coefficient $\beta$ are hyperparameters.
In the context of centralized training but decentralized execution, we employ a \emph{parameter sharing} strategy whereby each agent in a homogeneous team uses identical copies of policy parameters~\citep{gupta2017cooperative}.
\subsection{Self-attention}
Self-attention mechanism~\citep{cheng2016long} emerged from the natural language processing community.
It is used to relate different positions of a single sequence. The difference between self-attention and standard attention is that self-attention uses a single sequence as both its source and target sequence. It has been shown to be useful in image caption generation~\citep{liu2018show, yu2019multimodal} and machine reading comprehension~\citep{cheng2016long, yu2018qanet}.
The attention mechanism has also been adopted recently for multi-agent reinforcement learning. The relations between a group of agents can be learned through attention. \citet{iqbal2018actor} use attention to extract relevant information of each agent from the other agents. \citet{jiang2018learning} use self-attention to learn when to communicate with neighboring agents. \citet{wright2019attentional} use self-attention on the policy level to differentiate different types of connections between agents. \citet{Jiang2020Graph} use multi-head dot product attention to compute interactions between neighboring agents for the purpose of enlarging agents' receptive fields and extracting latent features of observations.
\citet{li2020deep} use multiplicative self-attention to implicitly build coordination graphs by weighting the graph edges with attention weights.
We use self-attention to learn the attention weights between agents. The attention weights are used to differentiate the importance of messages in the public communication channel for each agent.
\section{Approach}\label{sec:approach}
\begin{figure*}[t]
\centering
\input{figs/model}
\caption{Network architecture}
\label{fig:model}
\end{figure*}
We use a broadcast-and-listen mechanism to achieve agent communication. Instead of building agent-pair specific communication channels, our model has a public `chat room' to allow agents to share information. To selectively receive information, each agent differentiates the importance of messages from other agents by weighting them using attention weights. Then, agents aggregate the publicly broadcast messages using a weighted sum. The aggregated messages are concatenated with other agent-specific vectors to selection actions for agents. \cref{fig:model} shows the network architecture. It demonstrates the information flow between an agent pair, which can be easily vectorized for any number of agents.
In detail, we first pass $n$ observations $\{o_i\}_{i=1}^n$ of $n$ agents through a parameter sharing observation encoder $f_o$ parameterized by $\theta_o$. The observation encoder outputs observation embeddings $\{e_i\}_{i=1}^n$:
\begin{equation}
e_i = f_o(o_i; \theta_o), \text{ for } i = 1,\dots,n.
\end{equation}
We then compute the attention logits $\{c_i\}_{i=1}^n$ and the message logits $\{\mu_i\}_{i=1}^n$ with the parameter sharing attention encoder $f_a$ parameterized by $\theta_a$ and the parameter sharing message encoder $f_m$ parameterized by $\theta_m$ for all observation embeddings $\{e_i\}_{i=1}^n$:
\begin{equation}
c_i = f_a(e_i; \theta_a), \enspace \mu_i = f_m(e_i; \theta_m) \text{ for } i = 1,\dots,n.
\end{equation}
We use multiplicative attention~\citep{luong2015effective} to compute the attention scores and then softmax to obtain the attention weights. For an arbitrary pair of agents $i$ and $j$, the attention score $s_{ij}$ and the attention weight $w_{ij}$ of agent $i$ towards agent $j$ are:
\begin{equation}
s_{ij} = c_j^\top W_a c_i \qquad
w_{ij} = \frac{\exp(s_{ij})}{\sum_{k=1}^n \exp(s_{ik})}.
\end{equation}
The multiplicative attention operation is parameterized by a square matrix $W_a$.
The attention scores and weights can be efficiently computed for all $i, j$ pairs (including $i=j$, i.e. self-attention weights).
Instead of sampling from distributions, we use the $\argmax$ operation to extract discrete messages from message logits for the agents to broadcast.
We adopt two approaches to formulate discrete messages: one-hot and bit-string. For a given message bandwidth $b$:
\begin{itemize}
\item \textit{One-hot}: with a message logit vector $\mu_i \in \mathbb{R}^b$, the $k$th entry of $m_{\text{one-hot}, i} \in \{0, 1\}^b$ is given by
\begin{equation}
m_{\text{one-hot}, ik} =
\begin{cases}
1 & \text{if $k = \argmax_{l = 1,\dots,b} \mu_{il}$,}\\
0 & \text{otherwise.}
\end{cases}
\end{equation}
The resulting vocabulary size $|\mathcal{V}|$ of one-hot encoded messages is $b$. An example of an one-hot encoded message with bandwidth $b=5$ can be $m_{\text{one-hot}} = [0,0,1,0,0]$.
\item \textit{Bit-string}: we first need a message logit vector $\mu_i \in \mathbb{R}^{2b}$, the $k$th entry of $m_{\text{bit-string}, i} \in \{0, 1\}^b$ is given by
\begin{equation}
m_{\text{bit-string}, ik} =
\begin{cases}
1 & \text{if $k = \argmax_{l\in\{k, k+b\}} \mu_{il}$,}\\
0 & \text{otherwise.}
\end{cases}
\end{equation}
The resulting vocabulary size $|\mathcal{V}|$ of bit-string encoded messages is $2^b$. An example of a bit-string encoded message with bandwidth $b=5$ can be $m_{\text{bit-string}} = [1,0,0,1,1]$.
\end{itemize}
With the messages $m_i$ and attention weights $w_{ij}$, agents can aggregate the messages: $m_{aggr, i} = \sum_{j=1}^n w_{ij} m_j$.
For agent $i$, we then concatenate the observation embedding $e_i$, the attention weights $w_i=\{w_{ij}\}_{j=1}^n$, the message logits $\mu_i$ and the aggregated message $m_{aggr, i}$ to form an augmented embedding $\hat{e}_i = [e_i; w_i; \mu_i; m_{aggr, i}]$.
The concatenation creates several skip connections in the computation graph. They compensate for the gradient cut-off at the $\argmax$ operation and boost the gradients of the network components closer to the input head.
The augmented embedding $\hat{e}_i$ is then passed through a parameter sharing action policy $\pi$ parameterized by $\theta_\pi$ to infer action
\begin{equation}
a_i = \pi(\hat{e}_i;\theta_\pi).
\end{equation}
Continuous message communication can be achieved by circumventing the $\argmax$ operation in the network architecture and use message logits $\mu_i$ as the message to broadcast.
In summary, our model uses one round of communication, making the representational capacity of the communication protocols especially important.
\section{Experiments}
We perform experiments and analysis for discrete message communication by
(1) analyzing the effects of bandwidth and vocabulary size and comparing with continuous message communication;
(2) analyzing the importance of self-attention for discrete message communication;
(3) analyzing positive listening and signaling; and
(4) introducing human interaction with agents by using discrete message communication.
We show the results obtained from three environments: Pulling Levers~\citep{sukhbaatar2016learning}, Predator-Prey~\citep{Bohmer2019-zv, li2020deep}, and Multi-Walker~\citep{gupta2017cooperative, terry2020pettingzoo}. These environments have challenging tasks that must be performed cooperatively to achieve high returns. Using only local information cannot achieve high returns.
We use PPO~\citep{schulman2017proximal} for policy optimization (see Section~\ref{sec:ppo}). A multi-layer perceptron (MLP) baseline (value function) with global state is used for reducing the variance of advantage estimation with GAE~\citep{schulman2015high} during training.
\subsection{Analysis for Bandwidth and Vocabulary Size}
We compare the performance of communication protocols with various bandwidth and vocabulary size. The metric used is the average return. Average return indirectly measures the effects of communication on behaviors (i.e. positive listening). Continuous message communication is used as a baseline.
\subsubsection{Pulling Levers}\label{sec:levers_results}
\begin{figure}[t]
\input{figs/levers/n=5_tot=20_AverageReturn_bar}
\caption{Returns of Pulling Levers (5 levers, 20 total participants) with various communication protocols.}
\label{fig:levers_5_20}
\end{figure}
\begin{figure}[t]
\input{figs/levers/n=5_tot=100_AverageReturn_bar}
\caption{Returns of Pulling Levers (5 levers, 100 total participants) with various communication protocols.}
\label{fig:levers_5_100}
\end{figure}
Pulling Levers~\citep{sukhbaatar2016learning} is a simple multi-agent multi-armed bandit task that requires communication to achieve high a score. There are $n$ levers in the task. In each round, $n$ agents are randomly sampled from a total of $N$ participants. The action of each agent is to select a lever to pull. The reward of one round is given by the ratio between the number of unique levers pulled by the agents and the number of levers $n$. The maximum reward per round is therefore $1$, meaning each pulled lever is unique in that step. The theoretical expected reward of randomly pulling levers is $1 - (\frac{n-1}{n})^n$. The observation $o_i$ for each agent is the one-hot encoding of its ID number. Choosing levers without communication in this game will be similar to acting randomly.
A \textit{zero baseline} (value function) is used for training the pulling levers environment, since the global state (observation) is not related to return. Each episode consists of 50 rounds, and reward for one episode is summed over rewards from each round. Results are averaged over $5$ random seeds.
\cref{fig:levers_5_20} shows the normalized average returns of various communication protocols with $n=5$ and $N=20$. The communication protocols and their vocabulary sizes $|\mathcal{V}|$ are labeled on the horizontal axis: `\texttt{no}' stands for no communication, `\texttt{c}' means continuous message, `\texttt{b}' means bit-string, `\texttt{o}' means one-hot, and the number following the letter is the bandwidth $b$ of communication. A red dashed line is over-plotted to indicate the expected return with random action.
From \cref{fig:levers_5_20}, No communication shows the performance close to acting randomly. Continuous message works the best. The performance differences between continuous message bandwidths are not significant, since continuous message with any bandwidth can effectively encode an infinite number of meanings. Discrete message communication shows inferior performance due to its limited representational capacity. Generally, the performance of discrete message communication increases when the vocabulary size increases. This is particularly true for bit-string encoding. The 32-bit-string works the best among discrete communication, since its vocabulary size is $2^{32} \approx 4 \times 10^9$. This size is beyond human’s cognitive range. The average vocabulary size of native English speakers is around 20000, while 6000 to 7000 are sufficient for understanding most communication~\citep{rosenberg2008human}. The vocabulary sizes of the one-hot messages are within a human’s cognitive range, and their performance are not too far off comparing with continuous and bit-string messaging. The performance of one-hot messaging drops as the bandwidth increases beyond 128. This result indicates that overly large bandwidths are harmful for learning emergent communication protocols.
\cref{fig:levers_5_100} shows the normalized average returns for $n=5$ and $N=100$. The trend of the results is consistent with $N=20$.
\subsubsection{Predatory-Prey}\label{sec:predprey_results}
\begin{figure}
\centering
\includegraphics[width=0.18\textwidth]{figs/predprey/predprey_env.pdf}
\caption{Predators are marked in blue, and prey are marked in red. The cyan grids are the capture range of predators. An example of successful capture is predator 2 and 6 capturing prey 3. An example of a single-agent capture attempt that results in a penalty is predator 3 capturing prey 8 alone.}
\label{fig:predprey_env}
\end{figure}
We use a Predatory-Prey environment similar to that described by \citet{Bohmer2019-zv} and \citet{li2020deep}. The environment consists of a $7\times7$ grid world with 4 predators and 4 prey. We control the movement of predators to capture prey. The prey move by hard-coded and randomized rules to avoid predators. If a prey is captured, the agents receives a reward of $10$. However, the environment penalizes any single-agent attempt to capture prey with a negative reward $-0.5$; at least two agents are required to be present in the neighboring grid cells of a prey for a successful capture. \cref{fig:predprey_env} illustrates the environment and the reward mechanism. The agents have a vision range of 2 grids from itself. We engineer the agent's observation so that we can remove other agents positions away from an agent’s field of view. That means an agent can now only see the prey but not the other agents. The combination of agent invisibility and the single-agent capture-attempt penalty makes the task even harder. Cooperation is necessary to achieve a high return in this environment. We set the episode length to $50$ steps, and impose a step cost of $-0.1$.
\begin{figure}
\input{figs/predprey/penalty=-0.5_AverageReturn}
\caption{Returns of the 4-agent Predator-Prey with various communication protocols.}
\label{fig:predprey_results}
\end{figure}
The results are shown in \cref{fig:predprey_results}. Results are averaged over 5 random seeds. We ran experiments for various communication protocols. In particular, no communication \textbf{with} agent visibility (\texttt{no\_T}) and no communication \textbf{without} agent visibility (\texttt{no\_F}) form the performance upper-bound and lower-bound respectively. In the former setting, each agent has the richest information about the other agents, whereas in the latter setting, each agent has the scarcest information. Agents are set to invisible for all the other communication protocols. The dashed lines mark the upper and lower bounds. (The labels on the horizontal axis follow the convention defined in \cref{sec:levers_results} and \cref{fig:levers_5_20}.)
The performance of all the communicative methods falls within the upper and lower bounds. We can see similar increasing patterns for increasing bandwidth as in \cref{sec:levers_results} for Pulling Levers, which indicates that richer information transmission comes with higher bandwidth (up to a certain limit). Continuous messaging performs the best, followed by bit-string messaging. Bit-string messaging matches continuous messaging performance as its vocabulary size increases to $2^{32}$. One-hot messaging is slightly better than bit-string messaging when their vocabulary sizes are same. All the discrete communication protocols outperform no communication by a large margin.
\subsubsection{Multi-Walker}
\begin{figure}
\centering
\includegraphics[width=0.25\textwidth]{figs/multiwalker/multi_walker_env.png}
\caption{An illustration of the Multi-Walker environment with 3 agents.}
\label{fig:multi-walker_env}
\end{figure}
The previous two domains consist of discrete observation and action spaces. We would like to examine the consistency of discrete message communication in a more challenging continuous observation and action space task.
In the Multi-Walker environment~\citep{gupta2017cooperative, terry2020pettingzoo}, $n$ bipedal robot agents try to carry a bar-shaped package and move forward as far as possible as illustrated in \cref{fig:multi-walker_env}. The agents receive positive rewards for moving forward and negative reward for moving backward. Large negative rewards are given if the an agent falls or the package falls. The task is a combination of learning robotic locomotion and inter-agent cooperation.
In the original observation design, each agent receives an observation composed of physical properties of its legs and joints, as well as LIDAR readings from the space immediately in front and below the robot. The original observation also includes information about neighboring walkers, and the package. To emphasize the necessity of communication, we engineer the observation space to remove LIDAR readings and information about neighboring walkers. \textit{The resulting engineered observation therefore is only composed of the agent's own physical properties and information about the package.} The information about the other agents is now only accessible from communication.
\begin{figure}[t]
\input{figs/multiwalker/n_agents=3_AverageReturn_bar}
\caption{Returns of the 3-agent Multi-Walker with various communication protocol.}
\label{fig:multi_walker_results}
\end{figure}
The results are shown in \cref{fig:multi_walker_results} for $3$ agents averaged over 5 random seeds. We ran experiments for various communication protocols. In particular, we ran experiments with no communication \textbf{with} neighbor visibility and LIDAR turned on (original observation configuration, labeled with \texttt{no\_T}), and also no communication \textbf{without} neighbor visibility and LIDAR (observation configuration with scarcest info, labeled with \texttt{no\_F}).
The results show observe that communicative methods \emph{outperform} the original observation configuration \texttt{no\_T}. This can be explained by the fact that \texttt{no\_T} only has neighbor information available for each individual agent, while communicative methods can pass the information of non-neighboring agents to each other, providing a wider perception field. Continuous messaging with low bandwidth (\texttt{c4}) has poor performance. Continuous messaging with larger bandwidth ($\ge 8$) has much better performance, within which \texttt{c16} performs best. Bit-string messaging performs worse than continuous messaging and one-hot messaging. In contrast with Pulling Levers and Predator-Prey, the performance of bit-string messaging deteriorates when the bandwidth or vocabulary size increases. And the variance in performance also greatly increases accordingly. One-hot messaging has similar performance as continuous messaging in Multi-Walker, among which \texttt{o32} performs the best. The performance dramatically drops (below \texttt{no\_F}) as the bandwidth increases to 256. Similar to Pulling Levers, overly large bandwidths are harmful for learning emergent communication protocols.
\subsection{Ablation with Self-Attention}
\begin{figure}[t]
\input{figs/predprey/penalty=-0.5_AverageReturn_ua_bar}
\caption{Average return comparison between self-attention and uniform attention from the 4-agent Predator-Prey. Uniform attention bars are shaded and labeled with \texttt{ua} suffix.}
\label{fig:ua}
\end{figure}
In this analysis, we try to see the contribution of the self-attention to message aggregation. We run self-attention ablation experiments in the 4-agent Predator-Prey environment (the same as \cref{sec:predprey_results}). In \cref{fig:ua}, we show the performance comparison of communication protocols with self-attention (non-shaded) and their counterparts with uniform attention (shaded and with the \texttt{ua} suffix). The results are averaged over 5 random seeds. Uniform attention means that we circumvent the multiplicative attention operation and assign equal attention weight $1/n$ to each agent.
We observe that uniform attention poses a much more significant negative performance impact on the discrete messaging than on the continuous messaging. Possible explanations are that continuous messages can make up for the difference in attention weights by messages themselves, but different bit-string messages can have drastically different meanings and their limited representational power cannot replace the contributions of learned self-attention weights.
\subsection{Positive Listening and Signaling of Communication}
We design supervised learning tasks~\citep{li2020deep} to measure positive listening and signaling of communication.
\subsubsection{Positive Listening}
Positive listening quantifies the degree to which received messages are influencing an agent's behaviors ~\citep{lowe2019pitfalls, jaques2019social}.
It can be measured with a supervised learning task by using the masked aggregated message $\bar{m}_{aggr, i}$ \emph{received} by agent $i$ to predict agent $i$'s actions. The masking is done to remove agent $i$'s contribution to the aggregated message.
To compute the masked aggregated message $\bar{m}_{aggr, i}$, we first mask the self-attention weight $w_{ii}$ to zero $w_{ii} = 0$, then re-normalize the attention weights $\bar{w}_{ij} = w_{ij} / \sum_{k=1,\dots,n} w_{ik}$, and finally compute the masked aggregated message using the re-normalized attention weights $\bar{m}_{aggr, i} = \sum_{j=1,\dots,n} \bar{w}_{ij} m_j$.
We formulate the supervised learn task as $\hat{a}_i = f(\bar{m}_{aggr, i}; \phi)$ with loss $L = \text{CrossEntropy}(\hat{a}_i, a_i)$ \citep{li2020deep}.
Theoretically, with a finite amount of data, if $m_{aggr, i}$ is more correlated with $a_i$, i.e. if positive listening is strong, the classifier $f(\cdot; \phi)$ can produce a higher action prediction accuracy. We use a simple multi-layer perceptron (MLP) classifier parameterized by $\phi$, with a single 128-unit hidden layer and ReLU as activation.
\begin{figure}[t]
\input{figs/predprey/pos_listening_bar}
\caption{Positive listening analysis on the 4-agent Predator-Prey environment: predicting agent $i$'s actions $a_i$ using the masked aggregated messages $\bar{m}_{aggr,i}$ received by agent $i$. Data are collected from evaluation trajectories of five random runs.}
\label{fig:pos_lis}
\end{figure}
Results of five random runs from Predator-Prey are shown in \cref{fig:pos_lis}. Environment configuration is the same as that in \cref{sec:predprey_results}.
In general, higher bandwidths or larger vocabulary size can bring a stronger positive listening for all communication protocols. In terms of communication bandwidth, continuous messaging shows the best positive listening, closely followed by bit-string messaging. Bit-string messaging exhibits potential to yield positive listening on par with continuous messaging.
In terms of the vocabulary size instead of the bandwidth, one-hot messaging has a higher action prediction accuracy than bit-string messaging at low vocabulary sizes ($|\mathcal{V}| \le 32$).
A high action prediction accuracy is achieved because the broadcast-and-listen architecture allows individual agents to learn other agents' behaviors and intentions through messages they sent.
\subsubsection{Positive Signaling}
Positive signaling quantifies the degree to which an agent's sent messages are related to its observations or actions~\citep{lowe2019pitfalls, jaques2019social}.
It can be measured with a supervised learning task by using the message $m_i$ \emph{sent} by agent $i$ to predict agent $i$'s actions.
Similar to positive listening, we formulate the supervised learn task as $\hat{a}_i = f(m_i; \phi)$ with the same cross entropy loss and network architecture as used for positive listening.
With a finite amount of data, if $m_i$ is more correlated with $a_i$, i.e. if positive signaling is strong, the classifier $f(\cdot; \phi)$ can produce a higher action prediction accuracy.
\begin{figure}[t]
\input{figs/predprey/pos_signaling_bar}
\caption{Positive signaling analysis on the 4-agent Predator-Prey environment: predicting agent $i$'s actions $a_i$ using the message $m_i$ sent by agent $i$. Data are collected from evaluation trajectories of five random runs.}
\label{fig:pos_sig}
\end{figure}
Results of five random runs from Predator-Prey are shown in \cref{fig:pos_sig}. Continuous messaging shows the strongest positive signaling, while one-hot messaging shows the weakest positive signaling.
Bit-string messaging shows large gaps between continuous messaging when communication bandwidth is small. The gap decreases as the bandwidth increases. Bit-string messaging outperforms continuous messaging when bandwidth reaches $32$.
One-hot messaging under performs bit-string messaging by a large margin.
\subsection{Human Interpretation and Interaction}
Discrete messaging has a finite vocabulary size that human can easily interpret. We design a protocol to allow a human to interactively send messages to AI agents.
First, we collect raw observation and message pairs from trained agents. We cluster the raw observations using t-SNE~\citep{van2008visualizing, Policar731877}. Then, we label the raw observations with their corresponding discrete messages. \cref{fig:tsne} shows such clustering and labeling from Predator-Prey with \texttt{b4} (with a vocabulary size of $2^4=16$). Different colors in the clustering plot means different discrete messages. We use this clustering plot as a reference. We observe that the same cluster of raw observations tend to have the same discrete message labels. This provides an entry point for human-agent interaction. We can project new observations into the pre-known reference clusters as shown by the red cross in \cref{fig:tsne}, and empirically select the most probable messages to send to agents. \cref{fig:interaction} shows an outline for the human-agent interaction workflow.
\begin{figure}[t]
\centering
\includegraphics[trim={1cm 0 0 1.5cm}, clip, width=0.42\textwidth]{figs/predprey/tsne-generated_cluster.pdf}
\caption{T-SNE clustering of agent observations labeled by their corresponding messages (4-agent Predator-Prey with \texttt{b4}). The red cross represents an example of the projection of a new observation.}
\label{fig:tsne}
\end{figure}
\begin{figure}[t]
\centering
\input{figs/human_agent_interaction}
\caption{Human-agent interaction.}
\label{fig:interaction}
\end{figure}
\begin{figure}[t]
\centering
\input{figs/predprey/human_interaction_AverageReturn_bar}
\caption{Human-agent interaction results and comparison for Predator-Prey with \texttt{b4}.}
\label{fig:interaction_results}
\end{figure}
We design an experiment for letting human to select messages for \textbf{one} agent, with the rest of the agents selecting messages using the trained network (actual human-agent interaction experiments have text labels besides colors on the reference clustering points to assist in message selection).
We compare the performance with agent selecting messages and random messages for one agent.
\cref{fig:interaction_results} shows the results for the 4-agent Predator-Prey with 4-bit-string messaging (vocabulary size $2^4=16$). Environment configuration is the same as that in \cref{sec:predprey_results}. Agent and random selection are averaged over 200 episodes and human selection is averaged over 20 episodes. We can see human message selection has a bit worse performance than agent selection. The performance gap arises from the empirical prediction error when a human selects messages from their estimate using clustering and projection. Human message selection outperforms random selection.
In summary, the human-agent interaction protocol described in this analysis shows the interpretability of discrete messaging and demonstrates a way to integrate both human and AI agents.
\section{Conclusions}
In this work, we present a broadcast-and-listen model that enables end-to-end learning of emergent discrete message communication. We demonstrate that the bandwidth and the
vocabulary size of discrete messaging affects its performance. In some domains, discrete message communication can yield return performance and positive listening and signaling on par with or exceeding continuous message communication. Since discrete messages are easier to interpret by humans, we propose a human-agent interaction protocol that allows human to send discrete messages to agents. For future work, we would like to try multi-headed attention~\citep{vaswani2017attention} for the information aggregation process and new metrics other than average returns to more directly measure communication capabilities~\citep{lowe2019pitfalls}.
\section{Acknowledgements}
DISTRIBUTION STATEMENT A. Approved for public release. Distribution is unlimited.
This material is based upon work supported by the Under Secretary of Defense for Research and Engineering under Air Force Contract No. FA8702-15-D-0001. Any opinions, findings, conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the Under Secretary of Defense for Research and Engineering.
|
\section{INTRODUCTION}
\label{sec:intro}
In the inner disks ($r$~$<$~10 AU) around young stars, molecular gas emission and absorption
provide our best means of inferring the physical conditions at the radii where gas
giant and rocky planet cores are forming and accreting their nascent
atmospheres. The abundances of gas-phase species at the location of formation
and accretion are expected to affect the chemical abundances of exoplanet
atmospheres \citep{oberg11}. Recent surveys of molecular emission from
Classical T Tauri Stars (CTTSs) and Herbig Ae stars have provided new
constraints on the radial distribution, temperature, and composition of
planet-forming disks. Surveys of mid-IR emission from CO
\citep{salyk09,banzatti15}, H$_{2}$O and organic molecules
\citep{pontoppidan10,carr11}, and spectrally/spatially-resolved
near-IR observations \citep{pontoppidan11,brittain15} have
placed constraints on the CO and H$_{2}$O inventory at planet-forming radii and
the evolution of the inner molecular gas disk radius \citep{antonellini20}.
All of these studies require an accurate knowledge of the absolute abundances
to convert these measurements into local gas masses.
An emerging consensus from ALMA observations of protoplanetary disks is that on large
scales ($\sim 30-100$ AU), CO is heavily depleted relative to H$_2$\ \citep[e.g.,][]{miotello17,long17}
This depletion process may be fairly rapid, occurring within the first $\approx 1$ Myrs
of a proto\textit{stellar} disk's lifetime \citep{zhang20}.
When interpreted with physical-chemical models of disks that include CO freeze-out and
photodissociation \citep{miotello16}, the $^{13}$CO line fluxes are much
weaker than expected from dust emission and the few available detections of HD
\citep{bergin13,mcclure16}, implying that something other than freeze-out is
depleting the CO. The most plausible explanation is that complex
C-bearing molecules also freeze-out in the disk mid-plane, although chemical
reprocessing of CO in the cold disk mid-plane is also possible for
CTTS disks which are older than $\approx 1$ Myr \citep{bosman18}.
Once frozen out in a low-viscosity region, the ices will remain frozen out, with a time-dependence
that will gradually deplete the disk of C and therefore CO \citep{kama16,schwarz16}.
Our best opportunity to converge on a three-dimensional view of the
planet-forming regions around young stars, both inside and outside 10 AU, is to
assemble a broad spectrum of relevant disk tracers and diagnostics
\citep{sicilia16}. Among the most important items in this toolkit
of disk diagnostics are measurements of the abundances and the physical
conditions of the most important gas-phase species. Molecular absorption
spectroscopy, where a pencil beam through the circumstellar disk is observed
against the host star, provides a one-dimensional slice through the disk and
allows the column density, temperature, and radial velocity of the absorbing
gas to be quantified. This has historically been one of the most powerful
techniques for diagnosing the composition and temperature structure of the ISM,
and our approach in this work is to apply ultraviolet absorption spectroscopy
techniques to better understand planet-forming environments around young stars.
UV absorption line spectroscopy has been used to measure CO around a number of
CTTSs \citep{mcjunkin13} and resulted in the first circumstellar CO/H$_2$\ ratio for a pre-main sequence
star \citep{france14}. UV spectroscopy cannot probe the disk mid-plane,
but current models predict that carbon should rapidly get sequestered
in solids in the disk mid-plane at all radii and not return to the gas phase.
In the inner disk ($r \lesssim 4$ AU) the CO and H$_2$ in the warm molecular
layer accounts for the vast majority of the carbon and hydrogen respectively
\citep{adamkovics14}, which suggests that CO/H$_2$ measured through
this region of the disk is a good approximation of the true CO/H$_2$
value.
The focus of this work is the circumstellar gas around the Herbig Ae stars T Ori
and HK Ori. Both T Ori and HK Ori are well studied pre-main
sequence systems. T Ori is a UX Orionis (UXOR) variable star
which hosts a self-shadowed circumstellar disk \citep{hillenbrand92,bertelsen16}
and has an estimated accretion rate of $\text{log}(\dot{M}) = -6.58 \pm 0.40 \ifmmode {\rm\,\it{M_\odot}}\else ${\rm\,M_\odot}$\fi\ \text{yr}^{-1}$
\citep[][however, see \autoref{sec:disc}]{mendigutia11}. The leading explanation
for the UXOR phenomenon is the occultation of the central star by dust in
the warped or puffed up inner disk rim
\citep{dullemond03,kreplin16}. This requires that the disk be viewed nearly
edge-on ($i \gtrsim 70$). Although no estimates of T Ori's disk inclination angle exist,
its nature as a UXOR and the SED evidence of a self-shadowed disk suggests that its
disk is viewed at a large inclination angle. It also has a stellar rotational velocity
of $v$sin$i \approx 150 - 175$ km s$^{-1}$ \citep{mora01,alecian13,cauley15} implying
that the star is viewed equatorially rather than pole-on.
HK Ori is also actively accreting from its circumstellar disk \citep{hillenbrand92,mendigutia11}
with an accretion rate of $\text{log}(\dot{M}) = –5.24 \pm 0.12 \ifmmode {\rm\,\it{M_\odot}}\else ${\rm\,M_\odot}$\fi\ \text{yr}^{-1}$
and has a K4 binary companion at $\approx 150$ AU, or $\approx 0\overset{\ifmmode {^{\prime\prime}}\else $^{\prime\prime}$\fi}{.}3$
\citep{leinert97,baines04,smith05,wheelwright10}.
There are no published constraints on the inclination angle of HK Ori's circumstellar
disk, although \citet{blondel06} note that the low equivalent width of the \ion{Fe}{2} 2400 \AA\
absorption line suggests that we are viewing the star closer to pole-on since
an edge-on orientation would produce a large \ion{Fe}{2} equivalent width.
\citet{cauley15} derived a projected rotational velocity of $v$sin$i = 20$ km
s$^{-1}$ for HK Ori which is also indicative of a low inclination angle for
the system. However, the spectrum of HK Ori is contaminated by a plethora of
emission lines and there are few clean photospheric absorption lines which can
be used to fit spectral templates. Thus the value from \citet{cauley15} should
be viewed with caution. In addition, HK Ori shows significant spectroscopic
and photometric variability \citep{eiroa02,baines04,mendigutia11a} similar to the UXOR
phenomenon although with smaller amplitude. Finally, HK Ori on average shows
a double-peaked H$\alpha$ line profile with a strong central absorption feature
\citep{reipurth96,mendigutia11a,cauley15}. This profile shape was shown to be
qualitatively reproduced by accreting HAe stars viewed at high ($\gtrsim 70^\circ$)
inclination \citep{muzerolle04}. The H$\alpha$ line profile variability also
agrees well with the UXOR obscuration scenario tested by \citet{muzerolle04} for
UX Ori. Overall, the evidence for a disk with a high inclination angle (i.e., edge-on disk)
is stronger than for a low inclination angle.
In this paper we present new and archival \textit{Hubble Space Telescope} FUV data
for T Ori and HK Ori. We use the UV spectra to measure CO and
H$_2$\ column densities via pencil beam absorption through the circumstellar
material and then compare the derived column
densities to place constraints on the CO/H$_2$\ ratios in the gas. \autoref{sec:obs} describes
the UV spectroscopic data sets and \autoref{sec:models} describes the spectral synthesis
modeling used to derive molecular parameters. \autoref{sec:disc} places these results in
context of current disk and interstellar sight-line studies and a brief summary
of our findings is presented in \autoref{sec:conc}.
\section{OBSERVATIONS AND DATA REDUCTION}
\label{sec:obs}
The data were collected in 2012 and 2019 as part of the HST Guest Observing
Programs 12996 (P.I. Johns-Krull) and 15070 (P.I. France).
Both HK Ori and T Ori were observed with two different Cosmic Origins
Spectrograph (COS) settings in each program. The COS G130M $\lambda 1222$
setting from Program 15070 was selected to cover the wavelength range 1090 -
1360 \AA\, which includes two H$_2$ $v=0$ Lyman band transitions \citep{abgrall89}. We
note that the G130M $\lambda 1291$ setting from Program 12996 did not reach far
enough into the blue to cover the relevant H$_2$ lines. Thus the H$_2$ analysis
utilizes only the Program 15070 data. The G130M observations have a resolving
power of $R \approx 16,000$, or a velocity resolution of $\Delta v = 19$ km
s$^{-1}$. HK Ori was observed for a total of $t = 7758$ seconds in this mode,
while T Ori was observed for $t = 10089$ seconds.
The COS G160M $\lambda 1577$ setting covers the wavelength range $1400 - 1750$
\AA\ and contains the CO Fourth Positive \textit{A-X} system transitions from
the $0 \leq v' \leq 4$ vibrational states. The G160M observations have a
resolving power of $R \approx 18,000$, or a velocity resolution of $\Delta v =
17$ km s$^{-1}$. HK Ori was observed for a total of $t = 1376$ seconds and
T Ori was observed for $t = 1696$ seconds. All COS exposures were reduced
with the standard \texttt{calCOS} pipeline.
Quality G160M spectra are available for T Ori from both 2012 and 2019 through
both programs. The 2019 spectrum taken through Program 15070 has approximately
double the amount of flux compared with the 2012 spectrum and the CO absorption
is weaker by a factor of $\approx 2$. Interestingly, the flux ratio is
roughly constant at $\approx 2$ between the two epochs for wavelengths $\gtrsim 1300$
\AA\ where the photosphere dominates but drops to $\approx 1$ for wavelengths $\lesssim
1300$ \AA\ where the accretion continuum likely dominates.
We have opted to use the 2012 spectrum in our analysis because of the more
prominent CO feature. Although our choice to use the 2019 spectrum biases the CO analysis
because of the stronger absorption seen in that epoch, the lack of H$_2$ in T Ori's
spectrum (see \autoref{sec:h2models}) prohibits any useful constraints on
CO/H$_2$ in the system. Thus the bias introduced does not influence any of our
main conclusions concerning the circumstellar material around T Ori. We
provide a more detailed discussion of T Ori's FUV flux variability and how it
might be related to the variable CO absorption and lack of H$_2$ absorption
in \autoref{sec:tori_h2}.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.70,clip,trim=0mm 15mm 5mm 25mm,angle=0]{h2spectra_raw.pdf}
\figcaption{Extracted G130M spectra in the surrounding region of the H$_2$
Lyman band. The observed spectra are shown in black and the continuum fits in
red. Transitions to $v^{`} = 0$ are marked in green and transitions to $v^{`} = 1$
are marked in orange. Transition labels for (1-0)P(1) at 1094.05 \AA, (1-0)R(6) at 1109.86
\AA, (0-0)P(1) at 1110.06 \AA, (0-0)P(2) at 1112.50 \AA, and (0-0)P(3) at 1115.90 \AA\
are omitted for clarity. Clear absorption is visible in the HK Ori spectrum while
the T Ori spectrum shows no sign of absorption by H$_2$. The absorption features in the T Ori
spectrum are \ion{C}{1} lines, most of which are likely accretion-related.
\label{fig:h2specs}}
\end{figure*}
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.70,clip,trim=0mm 15mm 5mm 25mm,angle=0]{cospectra_raw.pdf}
\figcaption{Extracted G160 M spectra in the surrounding region of the CO
Fourth Positive system. The stellar spectra are shown in black and
scaled photosphere models from PHOENIX are shown in red. The
photospheric models are used as a guide to which features are stellar versus
circumstellar. The wavelength segments containing the CO $A - X$ bands
are shaded in gray. \label{fig:cospecs}}
\end{figure*}
\autoref{fig:h2specs} show the final co-added spectra surrounding the H$_2$
Lyman band transitions of interest \citep{abgrall89} and \autoref{fig:cospecs} shows the final
co-added spectra bracketing the CO $A - X$ band transitions. The spectra are
smoothed with a 5-pixel boxcar. The UV spectra of many Herbig Ae/Be stars are
littered with emission lines \citep[e.g.,][]{grady96,bouret98,cauley16}, which
are most clearly seen in HK Ori's spectrum. In order to isolate the H$_2$ and CO
absorption, we manually fit a high order spline, combined with a first or
second degree polynomial, to the contaminating emission lines to estimate the
``continuum'' emission. The continuum fits in the $1090 - 1130$ \AA\ region are
plotted as red lines in \autoref{fig:h2specs}.
For the $1400 - 1520$ \AA\ region containing the CO $A - X$ absorption bands,
we compared the observed spectra to PHOENIX stellar photosphere models
\citep{husser13} to aid in identifying which features are stellar versus
circumstellar. Scaled and reddened PHOENIX models are over-plotted in red in
\autoref{fig:cospecs}. Note that HK Ori is a binary system with an A-type
primary and G-type T Tauri secondary with a separation of $\approx 150$ AU
\citep{leinert97,smith05}. We only show the A-type photosphere model in
\autoref{fig:cospecs} due to the much weaker flux of
the G-type photosphere in the FUV. It is clear that the photosphere of
HK Ori is dwarfed by the emission lines so the photosphere is
relatively unimportant in this case. For T Ori, however, the spectrum
from $1400 - 1520$ \AA\ is almost entirely due
to the star's photosphere with little, if any, excess accretion emission
present. Thus we closely compared the photosphere model to the observed spectrum
during the spline continuum fitting for T Ori so as to minimize the inclusion
of stellar features in the final normalized CO spectrum. The excess CO
absorption is fairly clear in most of the bandheads in \autoref{fig:cospecs}.
\section{MODELING THE H$_2$ and CO ABSORPTION}
\label{sec:models}
Calculating the CO/H$_2$ ratio in the warm molecular layer of a disk requires a
measurement of the co-spatial CO and H$_2$ densities. To accomplish this we
model the column densities of H$_2$ and CO independently. The temperature of
each gas is also derived independently. For the CO models, the rotational
temperature is a model parameter; the H$_2$ rotational temperature is derived
from the best-fit column densities in a separate fitting procedure. All model
spectra are convolved with the COS linespread-function (LSF) \citep{kriss11},
which varies as a function of wavelength.
\subsection{H$_2$ models}
\label{sec:h2models}
The $v^{\prime\prime}=0,J^{\prime\prime}$ H$_2$ absorption lines are fit using
the \textit{h2ools} suite of optical depth templates from \citet{mccandliss03}, which
contain all of the Lyman and Werner transitions from the ground vibrational
level of the electronic ground state \citep{abgrall89}. Each template spectrum is calculated for
an integer value of the Doppler parameter $b$, where
$b = (2 k T_\text{rot}/m_\text{H2}+v_\text{turb}^2)^{1/2}$, and assuming a column density of
$N = 10^{21}$ cm$^{-2}$. The free parameters in the fit are the
line-of-sight radial velocity of the absorbing gas $v_\text{rad}$, the coverage
fraction of the stellar disk by the gas $f_\text{cov}$, the Doppler broadening
parameter $b$, and the column densities of the rotational transitions
$J^{\prime\prime} = 0 - 8$. We note that although the velocity resolution
of the observations is $\approx 17$ km s$^{-1}$, the value of $b$ can be constrained
at much higher precision since it has a strong effect on the optical
depth of the unsaturated H$_2$ transitions. Because the templates from \citet{mccandliss03}
are optical depth vectors, and the optical depth is proportional to the
column density, the templates can be adjusted for any column density
by multiplying by $N/10^{21}$. The absorption spectrum is then computed
as $e^{-\tau_i}$ where $\tau_i$ is the value of the scaled optical depth
template at each wavelength. Rotational transitions $J^{\prime\prime}>8$ are not
seen in the data and thus we exclude them from the modeling.
We perform the model fits using our custom MCMC routine based on the algorithm
of \citet{goodman10} \citep[see also][]{foreman13}. Uniform
priors are assumed for all parameters. We run the MCMC for 10,000 steps with
100 walkers, resulting in 10$^6$ samplings of the posterior distribution.
Because the template spectra are calculated for discrete values of $b$ we
interpolate between the two templates which bracket the model value of $b$. For
example, if the MCMC iteration requires a spectrum with $b = 3.2$ km s$^{-1}$
we linearly interpolate between the $b = 3.0$ km s$^{-1}$ and $b = 4.0$ km s$^{-1}$
templates. Once the interpolation is complete the resulting spectrum is shifted
by the radial velocity and scaled to find the best-fit column density. The model
spectrum, which is at the native velocity resolution, is convolved with the COS
LSF for comparison with the observed spectrum. The median values of the
marginalized posterior distributions are taken as the best-fit parameter values
and the $1\sigma$ confidence intervals are the 16$^\text{th}$ and 84$^\text{th}$
percentiles of the marginalized posteriors.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.80,clip,trim=5mm 15mm 35mm 20mm,angle=0]{h2_corner_hkori_corner.pdf}
\figcaption{Corner plot of the marginalized posterior distributions for the
H$_2$ MCMC of HK Ori. Note that all column densities $N$ are in units of
cm$^{-2}$. \label{fig:h2corner_hk}}
\end{figure*}
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.80,clip,trim=5mm 15mm 35mm 20mm,angle=0]{h2_corner_tori_corner.pdf}
\figcaption{Corner plot of the marginalized posterior distributions for the
H$_2$ MCMC of T Ori. The parameters are almost entirely unconstrained due to
the lack of absorption in the observed spectrum. \label{fig:h2corner_t}}
\end{figure*}
For T Ori there is no visual evidence of H$_2$ absorption (see
\autoref{fig:h2specs} and \autoref{fig:h2mods}) and the spectrum
from $1090 - 1130$ \AA\ is almost pure
continuum (surprising given the strong CO detection, discussed in the following section).
We apply the above H$_2$ fitting procedure to T Ori's spectrum, with some exceptions, to
measure the upper limit to T Ori's H$_{2}$ absorption properties. For T Ori we fix the
coverage fraction at $f_\text{cov} = 1.0$ and we freeze the radial velocity
near a literature value for the system of $v_\text{rad} = 56.0$ km s$^{-1}$
\citep{cauley15a}. This is necessary because the lack of absorption lines in
the spectrum allows a total degeneracy between $f_\text{cov}$ and the H$_2$
column densities. For the same reason $v_\text{rad}$ is entirely unconstrained.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.70,clip,trim=5mm 15mm 5mm 30mm,angle=0]{h2mods_final.pdf}
\figcaption{Normalized \textit{HST}\ COS spectra of HK Ori (top) and T Ori (bottom) from 1090
-- 1120 \AA. The best-fit H$_2$\ absorption models are over-plotted in red. No
significant H$_2$\ absorption is detected in the spectrum of T Ori. The magenta model
shows the expected H$_2$\ absorption assuming a conservative CO/H$_2$\ ratio of
10$^{-4}$ and the same temperature as that derived for the CO gas (see \autoref{sec:comodels}).
The blue model shows an assumed ratio of 10$^{-5}$.
\label{fig:h2mods}}
\end{figure*}
The marginalized posterior distributions from the MCMC are shown for HK Ori in
\autoref{fig:h2corner_hk} and for T Ori in \autoref{fig:h2corner_t}. Darker
blue colors indicate regions of higher posterior density. The fixed parameters
for T Ori $f_\text{cov}$ and $v_\text{rad}$ are excluded from
\autoref{fig:h2corner_t}. We also exclude the posteriors for the rotational
states $J^{\prime\prime} = 7,8$ for both objects due to severe blending of the
$J^{\prime\prime} = 7$ state and no detectable features corresponding to the
$J^{\prime\prime} = 8$ state.
The most likely model parameters and their $1\sigma$ confidence intervals are
given in \autoref{tab:h2fits} and the best-fit models are shown in
\autoref{fig:h2mods}. As expected, we are only able to place an upper
limit on the H$_2$\ column density for T Ori due to the lack of any discernible H$_2$
absorption in the spectrum. Also over-plotted for T Ori in \autoref{fig:h2mods}
is the expected H$_2$\ absorption assuming the canonical interstellar dense and translucent cloud ratios (CO/H$_2$$=10^{-4}$ and $=10^{-5}$)
and temperature $T_\text{rot} = 125$ K where the temperature and CO column
density are those derived in \autoref{sec:comodels}. A model with such a large
H$_2$\ column density is strongly ruled out by the MCMC posteriors. We will return to potential
explanations for the lack of circumstellar H$_2$ absorption around T Ori in \autoref{sec:disc}.
\begin{deluxetable}{lcc}
\tablewidth{0pt}
\tablecaption{H$_2$ model fit parameters\label{tab:h2fits}}
\tablehead{\colhead{Parameter}&\colhead{HK Ori}&\colhead{T Ori}}
\colnumbers
\tabletypesize{\scriptsize}
\startdata
$N$($J''=0$)$^\dagger$ (cm$^{-2}$) & 20.17$^{+0.04}_{-0.04}$ & 15.8$^{+1.0}_{-2.8}$ \\
$N$($J''=1$) (cm$^{-2}$) & 19.80$^{+0.04}_{-0.04}$ & $<$15.5 \\
$N$($J''=2$) (cm$^{-2}$) & 18.85$^{+0.13}_{-0.17}$ & $<$14.3 \\
$N$($J''=3$) (cm$^{-2}$) & 17.44$^{+0.30}_{-0.31}$ & $<$14.4 \\
$N$($J''=4$) (cm$^{-2}$) & 16.75$^{+0.35}_{-0.30}$ & $<$13.7 \\
$N$($J''=5$) (cm$^{-2}$) & 16.12$^{+0.48}_{-0.35}$ & $<$13.9 \\
$N$($J''=6$) (cm$^{-2}$) & 11.71$^{+1.33}_{-1.23}$ & $<$13.5 \\
$N$($J''=7$) (cm$^{-2}$) & \nodata & \nodata \\
$N$($J''=8$) (cm$^{-2}$) & 12.58$^{+1.59}_{-1.85}$ & $<$15.5 \\
$v_\text{rad}$ (km s$^{-1}$) & -19.9$^{+0.9}_{-0.9}$ & 56.0 \\
$b$ (km s$^{-1}$) & 4.9$^{+0.58}_{-0.60}$ & 2.1$^{+3.8}_{-0.8}$ \\
$f_\text{cov}$ & 0.995$^{+0.004}_{-0.007}$ & 1.0 \\
$T_\text{rot}$ (K) & 141$^{+6}_{-6}$ & 452$^{+649}_{-178}$ \\
\enddata
\tablenotetext{\dagger}{All column densities $N$ are log$_{10}$ values.}
\end{deluxetable}
The most likely parameter values for HK Ori, on the other hand, give a total
H$_2$ column density of log$_{10}$(N[H$_2$])$=20.34^{+0.03}_{-0.03}$ cm$^{-2}$.
The broadening parameter value is $b = 4.9^{+0.58}_{-0.60}$ km s$^{-1}$ suggesting a
significant contribution from turbulence or unresolved components to the line widths.
The model also indicates that the H$_2$ line-of-sight absorption covers the entire stellar
disk given the value of $f_\text{cov}$ so close to 1.0.
The temperature of the H$_2$ is not a free parameter in the spectrum fitting.
Instead, we use the derived column densities to find the most-likely rotational
temperature. The low-$J^{\prime\prime}$ H$_2$ level populations in the disk are expected to be determined by
collisions so we can calculate the temperature of the gas assuming a
Maxwell-Boltzmann distribution of the rotational states
\citep[see][]{france14}. We accomplish this by using the same MCMC routine to
fit a line to the H$_2$ densities as a function of the excitation temperature
of the various rotational states. The rotational, or kinetic, temperature
$T_\text{rot}$ of the gas is then the slope of the best-fit line.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.4,clip,trim=10mm 10mm 5mm 25mm,angle=0]{hkori_h2tempfit.pdf}
\figcaption{Fit to the H$_2$\ excitation temperature versus column density
for HK Ori. One hundred random samplings of the posterior distributions are shown in gray.
The best-fit value for the rotational excitation temperature is $T_\text{rot} =
141^{+6}_{-6}$ K, indicative of material in a warm disk layer.
\label{fig:hkori_tempfit}}
\end{figure}
The most-likely linear fit is shown with a red line in \autoref{fig:hkori_tempfit}
along with one hundred models from a random sampling of
the accepted posterior distributions (gray lines). We find that the H$_2$ gas
around HK Ori has a rotational temperature of $T_\text{rot} = 142^{+6}_{-6}$ K.
The small fit uncertainties on the $J^{\prime\prime} = 0,1,2$ dominate the slope fit
and thus the temperature determination. As expected from the lack of any H$_2$\
absorption around T Ori, the corresponding temperature fit is mostly unconstrained:
the same procedure yields $T_\text{rot} = 452^{+649}_{-178}$ K for T Ori.
\subsection{CO models}
\label{sec:comodels}
The CO fitting procedure is outlined in detail in \citet{mcjunkin13} \citep[see
also][]{france14} and we briefly summarize the salient points here. To derive
the circumstellar CO column densities we fit the CO Fourth Positive band system
shaded in gray in \autoref{fig:cospecs}. We include the $(4-0), (3-0),
\text{and} (2-0)$ bands in the fits for both objects but exclude the $(1-0)$
band for the HK Ori fit due to contamination by what is most likely an emission
line from atomic iron. The synthetic CO spectra are generated using literature
oscillator strengths and ground-state energy levels
\citep{haridass94,eidelsberg99,eidelsberg03}. The free parameters in the MCMC
fitting procedure are the column densities of $^{12}$CO and $^{13}$CO, the
radial velocity of the absorbing gas $v_\text{rad}$, the rotational temperature
of the gas $T_\text{rot}$, and the Doppler-broadening parameter $b$. Similar
to the case for H$_2$, $b$ has a strong affect on the optical depth of the CO
transitions and is somewhat degenerate with the column density. This makes
the value of $b$ sensitive to not only the line widths but the line depths as well.
We again adopt uniform, non-restrictive priors for most of the parameters and run the 100
walkers for 10,000 steps each. The one exception is the Doppler broadening
parameter $b$: we restrict the prior range to $0.0 < b \leq 5.0$ km s$^{-1}$.
For each model iteration we generate the CO spectrum using the input $T_\text{rot}$,
$b$, and column density values. The spectrum is then shifted in velocity space
according to $v_\text{rad}$ and then convolved with the COS LSF to produce the output model.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.40,clip,trim=15mm 15mm 35mm 20mm,angle=0]{co_corner_hkori_corner.pdf}
\figcaption{Corner plot of the marginalized posterior distributions for the CO MCMC of HK Ori.
\label{fig:cocorner_hk}}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=.40,clip,trim=15mm 15mm 35mm 10mm,angle=0]{co_corner_tori_corner.pdf}
\figcaption{Corner plot of the marginalized posterior distributions for the CO MCMC of T Ori.
\label{fig:cocorner_t}}
\end{figure}
The marginalized posterior distributions for the CO MCMC fits are shown in
\autoref{fig:cocorner_hk} and \autoref{fig:cocorner_t}. The best-fit models
(red) and fifty random samplings of the accepted posteriors (green) are shown
in \autoref{fig:comods}. The most noticeable difference in the CO spectra is
the width of the absorption bands, where HK Ori's are fairly narrow while T
Ori's are broad. This is manifested in the large difference in derived
$T_\text{rot}$ values: for HK Ori $T_\text{rot} = 19^{+11}_{-6}$ K and for T
Ori $T_\text{rot} = 124^{+53}_{-44}$ K. Note that T Ori's $b$-value
is just the median of the prior range, i.e., it is unconstrained. This is
a result of the high temperature: as the temperature increases and the optical
depth in the higher rotational states increases, the individual line broadening
becomes less important and the instrumental profile makes differences in
$b$ negligible. This is not the case for the low temperature fits for HK Ori.
At low temperatures larger values of $b$ increase the optical depth over
a narrow range of rotational states which significantly increases the absorption
depth. Thus there is some degeneracy between $N$ and $b$ at low $T$ and
$b$ is constrained even though it is much smaller than the instrumental resolution.
Neither spectrum shows any strong absorption due to $^{13}$CO, in contrast to some
CTTS disks \citep[e.g.][]{france12}.
We note that the $^{12}$CO/$^{13}$CO ratios indicated by our analysis
are much larger than is typically found for the ISM or the circumstellar
environments of young stellar objects (YSOs) \citep{smith15}. However, the
$^{13}$CO column densities are highly uncertain and it's likely that the
$^{13}$CO absorption is almost entirely lost in the noise for both objects.
Thus we caution against applying the derived $^{12}$CO/$^{13}$CO ratios
to further work on these systems.
\begin{figure*}[ht!]
\centering
\includegraphics[scale=.60,clip,trim=5mm 5mm 10mm 60mm,angle=0]{comods_final.pdf}
\figcaption{Normalized spectra surrounding the CO (\textit{A-X}) absorption
bands for HK Ori (top row) and T Ori (bottom row). The best-fit CO absorption
models are over-plotted in red and one hundred random draws from the accepted
posterior chains are shown in green. Note that the CO(1-0) band is excluded
from the model determination for HK Ori due to the contamination by the strong
emission line near 1511 \AA. \label{fig:comods}}
\end{figure*}
\begin{deluxetable}{lcc}
\tablecaption{CO model fit parameters\label{tab:cofits}}
\tablehead{\colhead{Parameter}&\colhead{HK Ori}&\colhead{T Ori}}
\colnumbers
\tabletypesize{\scriptsize}
\startdata
$N(^{12}\text{CO})^\dagger$ (cm$^{-2}$) & 15.5$^{+0.3}_{-0.3}$ & 14.9$^{+0.2}_{-0.1}$ \\
$N(^{13}\text{CO})$ (cm$^{-2}$) & 11.0$^{+2.6}_{-4.2}$ & 9.2$^{+3.1}_{-3.0}$ \\
$v_\text{rad}$ (km s$^{-1}$) & $-$14.9$^{+4.1}_{-4.6}$ & 44.5$^{+6.7}_{-5.1}$ \\
$b$ (km s$^{-1}$) & 1.1$^{+0.3}_{-0.2}$ & 2.5$^{+1.7}_{-1.4}$ \\
$T_\text{rot}$ (K) & 19$^{+11}_{-6}$ & 124$^{+53}_{-44}$ \\
\enddata
\tablenotetext{\dagger}{Column densities $N$ are log$_{10}$ values.}
\end{deluxetable}
\subsection{\ion{H}{1} Absorption}
\label{sec:HImodels}
Absorption line analyses of interstellar sight-lines often rely on a knowledge of the
total neutral hydrogen column density, N(\ion{H}{1}), to place molecular absorption
properties in context. Toward this end, we use the 2012 $HST$-COS G130M observations
of both stars to measure N(\ion{H}{1}) by fitting the Ly$\alpha$ absorption profile.
We show the observed profiles and reconstructed spectra from the model fits in \autoref{fig:lymmods}.
Both HK Ori and T Ori display broad Ly$\alpha$ emission features consistent with
accretion-generated \ion{H}{1} emission observed on lower mass T Tauri stars
\citep[e.g.,][]{france14b}. In analogy to the methodology employed for H$_{2}$ and CO,
we assume the observed profile is the superposition of an underlying stellar emission
spectrum and the Ly$\alpha$ absorber. For HK Ori, we employed a broad Gaussian as
the intrinsic Ly$\alpha$ profile shape, and implemented an iterative fitting routine
to find the best fit to the observed COS G130M Ly$\alpha$ spectra \citep{france12,mcjunkin13}.
The Ly$\alpha$ emission peak is less pronounced in T Ori, and here we fit a spline function
to the continuum and emission line as a baseline for fitting the \ion{H}{1} absorption.
The Doppler $b$-value, $b_{HI}$, is assumed to be 10 km s$^{-1}$ for both stars. For
more details on reconstructing intrinsic Ly$\alpha$ profiles see \citet{mcjunkin14}.
\begin{figure}[htbp]
\centering
\includegraphics[scale=.38,clip,trim=25mm 10mm 10mm 10mm,angle=0]{lyman_fits_both.pdf}
\figcaption{Observed (top panels) and reconstructed (bottom panels) Ly$\alpha$ profiles for T Ori and HK Ori. The best-fit neutral hydrogen ISM column densities are labeled in the bottom panels. \label{fig:lymmods}}
\end{figure}
For HK Ori, our fits find a baseline continuum of 8~$\times$~10$^{-15}$ erg
cm$^{-2}$ s$^{-1}$ \AA$^{-1}$, Gaussian peak amplitude of 6~$\times$~10$^{-12}$
erg cm$^{-2}$ s$^{-1}$ \AA$^{-1}$, a Gaussian FWHM = 1100 km s$^{-1}$, and
log$_{10}$N(\ion{H}{1}) = 20.75~$\pm$~0.10. Taking the spline function as the
underlying stellar spectrum for T Ori, we find log$_{10}$N(\ion{H}{1}) =
20.25~$\pm$~0.10. These results are summarized in Table 3.
\section{DISCUSSION}
\label{sec:disc}
\subsection{Comparing H$_{2}$, CO, and \ion{H}{1} with the Translucent ISM}
\label{sec:ismcomp}
We compared the absorption properties of the two targets against typical properties
of interstellar sight-lines to help solidify assignment of these components to
a circumstellar disk origin. The HK Ori CO/H$_{2}$ ratio and molecular
fraction\footnote{The molecular fraction, $f_{H2}$ is defined as
$f_{H2}$~=~2N (H$_{2}$)/(2N(H$_{2}$) + N(\ion{H}{1})).} are consistent with interstellar clouds
at the diffuse-to-translucent boundary \citep[e.g.,][]{burgh07}. However, the CO
and molecular fraction are considerably higher than expected for a star with HK
Ori's optical reddening $E(B - V)$ \citep[$=0.12$; assuming the N(\ion{H}{1}) to
reddening conversion from][]{diplas94}. Furthermore, the H$_{2}$ column density
to HK Ori is orders of magnitude higher than what is seen to typical hot star
targets near the Orion star-forming region \citep{savage77}. This points to a
circumstellar origin of the molecular gas in HK Ori.
Further support for a circumstellar origin of the molecular gas around HK Ori comes from an
analysis of the excitation temperatures. While T(H$_{2}$) = 141 K is consistent with an
interstellar origin, T(CO) = 19 K is much higher than what is seen for translucent clouds
\citep[e.g., see Figure 6 of][]{burgh07}. We note that the temperature discrepancy could be
due to the density of CO being below the critical density (see \autoref{sec:ismcomp} for
a more detailed explanation). Therefore, numerous lines of evidence suggest a
circumstellar, as opposed to interstellar, origin for the molecular gas observed toward HK Ori.
T Ori's hydrogen sightline is characteristic of hot stars near the Orion star
forming regions \citep[N(H$_{2}$)~$<$~10$^{16}$ cm$^{-2}$; N(\ion{H}{1}) = 10$^{20.0-20.5}$
cm$^{-2}$;][]{savage77}. However, the diffuse sightlines in Orion are typically
associated with very low CO upper limits \citep[N(CO)~$<$~10$^{13}$ cm$^{-2}$;][]{federman80}.
Similarly, the very low molecular fraction is incompatible with the large CO column density,
arguing that the CO we observe towards T Ori is almost certainly of a circumstellar origin.
This is corroborated by the high T(CO), which is a factor of $\sim$10~--~20 higher than
seen on translucent interstellar sight-lines \citep{burgh07,sheffer08}. The circumstellar
nature of T Ori's CO absorption is also bolstered by the fact that it shows a variable
column density likely related to inner disk material (see \autoref{sec:tori_h2}).
\begin{deluxetable*}{lccccc}
\tablecaption{Total column densities and N(CO)/N(H$_2$) ratios\label{tab:ratio}}
\tablehead{\colhead{Object}&\colhead{N(H$_2$)}&\colhead{N(CO)}& \colhead{N(\ion{H}{1})}& \colhead{N(CO)/N(H$_2$)} & \colhead{$f_{H2}$}}
\colnumbers
\startdata
HK Ori & 20.34$^{+0.03}_{-0.03}$ & 15.5$^{+0.5}_{-0.2}$ & 20.75~$\pm$~0.10 & 1.3$^{+1.6}_{-0.7} \times 10^{-5}$ & 0.44 \\
T Ori & $<$~15.9$^{+1.0}_{-0.4}$ & 14.9$^{+0.2}_{-0.1}$ & 20.25~$\pm$~0.10 & \nodata & $<$1~$\times$~10$^{-4}$ \\
\enddata
\end{deluxetable*}
\subsection{Where's the H$_2$ around T Ori?}
\label{sec:tori_h2}
The lack of H$_{2}$ absorption towards T Ori is surprising given the system's
young age and the high likelihood that we observe the disk closer to an edge-on
orientation. The properties of the CO gas we derive confirm that the gas is
circumstellar ($T \approx 125$ K) and not interstellar, where the gas typically
has temperatures $T < 10$ K. Furthermore, the radial velocity of the CO absorption,
$\approx 45$ km s$^{-1}$, is consistent with, although mildly blue-shifted relative
to, the system velocity of $\approx 56$ km s$^{-1}$ (noting that $HST$-COS has a
wavelength accuracy of 15 km s$^{-1}$).
\begin{figure*}[htbp]
\centering
\includegraphics[scale=.75,clip,trim=15mm 10mm 10mm 20mm,angle=0]{tori_fluxcomp.pdf}
\figcaption{Comparison of the 2012 (red) and 2019 (purple) FUV T Ori spectra.
The bottom panel displays the flux spectra from both epochs, as well as a
photospheric template scaled to match the 2019 flux. The Lyman-$\alpha$ emission
lines are interpolated over for clarity. Only wavelengths common to both
sets of observations are shown. The middle panel shows the ratio of the 2012
spectrum to the 2019 spectrum and the top panels show zoomed-in normalized
spectrum comparisons between strong spectral features from each epoch. The CO
bandheads are marked with gray shaded regions and the wavelength ranges shown
in the top panels are marked with light blue shading in the middle panel.
\label{fig:toriflux}}
\end{figure*}
We can look to the variability between the 2012 and 2019 observations for
clues about the lack of H$_2$ absorption towards the star. In \autoref{fig:toriflux}
we show a comparison between the \textit{HST} COS spectra from 2012 (red)
and 2019 (purple). The bottom panel shows the raw spectra for wavelengths
common to both epochs and the middle panel shows the ratio of the fluxes. The
top panel displays zoomed-in normalized comparisons between strong spectral
features. A few things are evident from \autoref{fig:toriflux}: 1. The flux
at wavelengths longer than $\approx 1300$ \AA\ was $\approx 2\times$ greater
in 2019 than in 2012; 2. The flux at wavelengths shorter than $\approx 1300$ \AA\
stayed relatively constant; 3. Most of the photospheric absorption lines are weaker
when the star is fainter; 4. The CO absorption is stronger when the star is
fainter (see the flux dips within the gray bands in the middle panel of
\autoref{fig:toriflux}). Additionally, we downloaded $V$-mag ASAS-SN light curves
\citep{shappee14,kochanek17} of T Ori and interpolated the observed magnitudes
onto our \textit{HST} COS observation dates: for the 2012 observations $V \approx 11.3$
and for the 2019 observations $V \approx 10.4$, which translates to a factor
of $\approx 2.3$ in flux. The $V$-band measurements confirm that the optical
flux varies at a similar level to the FUV flux.
We speculate that T Ori's nature as a rapid rotator and UXOR variable
can qualitatively explain the phenomena seen in \autoref{fig:toriflux}.
First, rapidly rotating stars are subject to non-negligible gravity darkening
where the oblateness of the star causes the poles to be hotter than the
equator \citep{vonzeipel24,espinosa11}. If the UXOR screen preferentially blocks
equatorial stellar latitudes, or is at least more optically thick across these
regions, the star will be dimmer but
the stellar spectrum will be weighted towards the hotter polar regions. This
can account for the observed shallower absorption lines in the 2012 faint phase.
Second, the increased CO absorption during the faint phase could be a result of the
increased gas column density produced by the occulting screen. Finally, the lack
of H$_2$ absorption in the Lyman transitions may be the result of the
sub-1300 \AA\ flux, which is related to material accreting onto the star, being
scattered into the line-of-sight over a large volume in the inner disk region.
If the flux is the result of scattering over a large volume, we are not probing
the same pencil beam through the disk that produces the CO absorption since the
flux from 1400 - 1500 \AA\ is largely photospheric in origin. This also explains
the relatively stable sub-1300 \AA\ flux since the UXOR screen does not occult
the entire scattering surface.
We stress that the scenario outlined above is speculative; there are
likely other plausible explanations for the various phenomena (e.g., an occulting
accretion column and related veiling of the stellar photosphere). In particular, the
proposed viewing geometry must be fine-tuned in order to preferentially occult
equatorial latitudes on the star, leaving the hotter poles visible. Additional
FUV time-series spectra will be useful in determining a more precise view of the
system geometry and we defer a more quantitative characterization of the variability
to future work.
\subsection{Large Doppler broadening parameters}
\label{sec:doppler}
The H$_2$ and CO analysis for HK Ori suggest large
($\approx 1-4$ km s$^{-1}$) non-thermal broadening components are necessary to
reproduce the observed H$_2$ and CO absorption. This would imply significant contributions
from turbulence or spatially unresolved gas motions responsible for the line broadening.
The turbulence interpretation seems unlikely given that CO radio observations suggest
that turbulence in protoplanetary disks is weak at distances of $\approx 10s$ of AU
\citep{hughes11,flaherty18,flaherty20}. On the other hand, the pencil beam H$_2$ observations,
if passing through the disk, likely probe higher surface layers than the thermal CO emission.
We caution against an over-interpretation of the large value for
the Doppler parameter for two reasons. First, absorption features with widths of a
few kilometers per second are not resolved in our spectra. The large $b$-values are
mainly a result of the models fitting the absorption line depths, which are
determined by both $b$ and the column density $N$. Without resolving the line widths
of individual H$_2$ and CO transitions it is difficult to say with confidence
what is driving the large values of $b$.
Second, there is considerable uncertainty around the viewing geometry for the
disk around HK Ori. More precise constraints on
our viewing angle through these circumstellar environments
could help clarify the interpretation of the large value for the Doppler parameter.
\subsection{Spatial Location and Distribution of Molecular Absorption Components around HK Ori}
\label{sec:ismcomp}
We have measured significant columns of H$_2$ and CO towards HK Ori. The
presence of the K4 companion at $0\overset{\ifmmode {^{\prime\prime}}\else $^{\prime\prime}$\fi}{.}3$ might suggest that the COS
observations contain flux from both components. While it is plausible that the
less massive companion contributes some coronal emission line flux to the FUV spectrum,
the absorption signatures are most likely produced by circumstellar material
around the primary: \citet{smith05} found no evidence for an IR excess around
the K4 companion whereas the IR excess around HK Ori A is well matched by
an accretion disk extending to $\approx 30$ AU. Given the lack of circumstellar
material around the secondary, we continue under the assumption that the absorption
is created by material occulting the primary star.
The most critical assumption necessary for deriving a CO/H$_2$\ ratio in
a disk via pencil beam absorption is that the CO and H$_2$\ populations are
co-spatial, i.e., we are sampling the same parcels of the disk in both
species. At first glance, the differences in the derived rotational temperatures between
the H$_{2}$ and CO populations suggest we might be sampling different gas components.
However, while the CO rotational levels are populated by collisions, radiative processes
can alter the observed distributions of $J$-levels. If the local number density is lower
than the CO critical density, these levels can radiatively depopulate fast enough that the
populations are not representative of the local kinetic temperature.
Referencing the density and temperature structures of Herbig Ae disks presented in \citet{agundez18},
we estimate the local H number density in the region probed by our HK Ori H$_{2}$ absorption spectra is
$n_{H}$~$\sim$~10$^{4-7}$ cm$^{-3}$. Assuming an H$_{2}$ ortho–para ratio
of 3 and that the T$_{H2}$~$\approx$~T$_{kin}$ = 150 K, the total ortho- + para-H2
collision rate, summed over all possible lower levels, is $\Gamma_{TOT}$ = 2.6 $\times$
10$^{-10}$ cm$^{3}$ s$^{-1}$ at $J = 10$ \citep{yang10}. The CO critical density for $J = 10$ at 150 K
is $n_{H2}$ $\approx$ 2.8 $\times$ 10$^{5}$ cm$^{-3}$. Therefore, it is possible that the high-J CO
populations that drive the determination of T(CO) are significantly sub-thermal, which would
explain the temperature difference between the observed H$_2$ and CO populations. To explore this
possibility, we computed RADEX models~\citep{vandertak07} using the online
interface\footnote{http://var.sron.nl/radex/radex.php}. For our measured values of T(H$_{2}$),
N(CO), $b_{CO}$, and $n_{H}$ = 10$^{4}$ cm$^{3}$, we find that CO rotational levels 3~$\leq$~$J$~$\leq$~30
are characterized by temperatures less than the H$_{2}$ rotational temperature. Finally, sub-thermal
CO rotational temperatures are almost always observed in co-spatial CO and H$_{2}$ populations
in the ISM \citep{burgh07,sheffer08}. We conclude that while there is uncertainty about the exact
spatial distribution of the gas and the viewing geometry, it is reasonable to infer that the H$_{2}$
and CO are co-spatial around HK Ori with CO/H$_{2}$ ratio $\approx 1.3 \times 10^{-5}$.
The system radial velocity for HK Ori is $\approx 20-25$ km s$^{-1}$
\citep{reipurth96,baines04} which means the best-fit radial velocities for both the H$_2$
and CO absorption are blue-shifted from the system velocity by $\approx 35 - 40$ km s$^{-1}$
since we fit the spectrum in the Earth's rest frame. Again, it is worth noting
that the absolute velocity precision of COS is $\approx 15$ km s$^{-1}$. Even taking
into account the velocity precision of COS, the measured blue-shifts for the H$_2$
and CO absorption are large and inconsistent with a pencil beam observed through
a Keplerian disk since material in a circular disk along any line-of-sight should have velocities
centered around $\approx 0$ km s$^{-1}$. However, it's possible that the absorption
arises at the base of a molecular disk wind launched magnetocentrifugally from the disk's upper
atmosphere. Evidence of such a wind has been detected in CO emission from the Herbig Ae star
HD 163296 by \citet{klaassen13} with characteristic velocities of $\approx 18$ km s$^{-1}$.
Thus depending on the wind launching angle and the disk inclination angle, blue-shifted
velocities of $\approx 40$ km s$^{-1}$ are not unreasonable if the absorption arises
in a disk wind.
Given the evidence for a disk or disk wind origin of the gas absorption towards
HK Ori, our derived CO/H$_2$ ratio of $\approx 1.3 \times 10^{-5}$ suggests significant
depletion of CO in the circumstellar material relative to the canonical molecular cloud
value of $10^{-4}$ \citep{bergin17}. Lower values of CO/H$_2$ are expected for cooler
disks with mid-plane temperatures $\lesssim 30-40$ K which is mainly driven by dust grain
surface chemistry \citep{reboussin15}. However, the depletion factor of CO is highly
dependent on age and location in the disk \citep[e.g.,][]{krijt20} which makes it difficult
to say anything definitive about HK Ori's disk based on our derived ratio given the uncertainties
around the line-of-sight probed by our observations. If our observations trace the
warm molecular layer of HK Ori's disk, the CO/H$_2$ ratio is line with the depleted
values reported for other protoplanetary disks \citep[e.g.,][]{favre13,mcclure16,schwarz16}.
\section{CONCLUSION}
\label{sec:conc}
We presented new FUV observations from the \textit{Hubble Space Telescope}
of H$_2$ and CO absorption towards the Herbig Ae stars HK Ori and T Ori. We find
a significant column of CO absorption and a notable lack of H$_2$ absorption
towards T Ori, which we posit is related to variable occultation
of the star by inner disk warping. HK Ori shows blue-shifted absorption in
both H$_2$ and CO with derived temperatures and column densities which
place the gas firmly in the circumstellar environment
around HK Ori. Although there is considerable uncertainty as to the exact
line-of-sight traced by our observations, the derived CO/H$_2$ value around
HK Ori is $\approx 1.3 \times 10^{-5}$ suggesting a depleted reservoir of CO.
Spatially-resolved molecular emission maps of both systems will help place our
pencil-beam observations in context and provide a more precise understanding
of the CO/H$_2$ value derived for the gas around HK Ori.
\bigskip
{\bf Acknowledgments:} We thank the referee for their suggestions, which
helped improve the quality of the manuscript. P. W. C. acknowledges useful
conversations with Nicole Arulanantham about the molecular components of
inner disks. The data in this paper were obtained through
\textit{HST} Guest Observing programs 12996 and 15070. This work has made use of
NASA's Astrophysics Data System and the SIMBAD database,
which is operated at CDS, Strasbourg, France.
\clearpage
|
\section{Appendix}
\subsection{Brief introduction to CET and MPK}
\label{app:cet}
\paragraph{CET} \cet is an instruction set architecture extension that helps
prevent Return-Oriented Programming and Call/Jmp-Oriented Programming
via use of a \emph{shadow stack}, and \emph{indirect branch tracking}
(IBT).
The shadow stack is a hardware-maintained stack used
exclusively to check the integrity of return addresses on the program stack.
To ensure the shadow stack cannot be tampered with, it is inaccessible via
standard load and store instructions.
The IBT allows the enforcement of coarse-grained control flow integrity
(CFI)~\cite{cfi} via a branch termination instruction, \mono{endbranch}.
Binaries that wish to use IBT place the \mono{endbranch} at all valid indirect
jump targets.
If an indirect jump instruction lands on any other instruction, the CPU
reports a control-flow protection fault.
Additionally, the IBT also supports a \emph{legacy bitmap}, which allows
programs to demarcate which code pages have IBT checking enabled.
Importantly, \cet guarantees that any shadow stack mismatches observed during
speculative execution of \mono{return} instruction immediately halts further
speculative execution.
Similarly, any indirect jump during speculative execution from an IBT enabled
code page to a page with IBT disabled also halts speculation.
\paragraph{MPK} \mpk uses four bits in each page-table entry to assign one of sixteen
"keys" to any given memory page, allowing for 16 different memory domains.
User mode instructions \mono{wrpkru} and \mono{rdpkru} allow setting read and
write permissions for each of these domains on a per-thread basis.
\mpk thus allows a process to partition its memory and selectively
enable/disable read and write access to any of regions without invoking the
kernel functions or switching page tables.
Importantly, \mono{wrpkru} does not execute speculatively - memory accesses
affected by the PKRU register will not execute (even speculatively) until all
prior executions of \mono{wrpkru} have completed execution and updated the PKRU
register and are also resistant to Meltdown style attacks~\cite{IntelSpecMit}.
\subsection{Testing Disclaimer}
\label{sec:app:testing-disclaimer}
Since we use a software development platform provided by Intel, we include the
following disclaimer from Intel:
\begin{myleftbar}
\small
Software and workloads used in performance tests may
have been optimized for performance only on Intel microprocessors. Performance
tests, such as SYSmark and MobileMark (in this paper SPEC CPU
2006 and Sightglass), are measured using specific computer systems, components,
software, operations and functions. Any change to any of those factors may cause
the results to vary. You should consult other information and performance tests
to assist you in fully evaluating your contemplated purchases, including the
performance of that product when combined with other products. For more complete
information visit www.intel.com/benchmarks. Performance results are based on
testing as of dates shown in configurations and may not reflect all publicly
available updates. See backup for configuration details. No product or component
can be absolutely secure. Your costs and results may vary. Intel technologies
may require enabled hardware, software or service activation.
\end{myleftbar}
\section{A brief overview of Wasm and Spectre}
\label{sec:background}
In this section, we give a brief overview of WebAssembly's use in multi-tenant
serverless and edge-cloud computing platforms---and more generally function as
a service (FaaS) platforms.
In particular, we describe how FaaS platforms use \wasm as an intermediate
compilation layer for isolating different tenants today.
We then briefly review Spectre attacks and describe how today's approach to
isolating \wasm code is vulnerable to this class of attacks.
\begin{figure}[t!]
\centering
\resizebox{.9\hsize}{!}{
\includegraphics[width=1\hsize]{figures/cdn-overview.pdf}
}
\caption{FaaS platform using Wasm to isolate mutually distrusting tenants.}
\label{fig:faas}
\vspace{-.1in}
\end{figure}
\subsection{WebAssembly}
\label{subsec:wasm_bg}
\wasm is a low-level 32-bit machine language explicitly designed to embed C/C++
and Rust code in host applications.
\wasm programs (which are simply a collection of functions) are (1)
deterministic and well-typed, (2) follow a structured control flow discipline,
and (3) separate the heap---the \emph{linear memory}---from the well-typed
stack and program code.
These properties make it easy for compilers like Lucet to sandbox \wasm code and
safely embed it within an application~\cite{wasm-paper}.
\paragraph{Control flow safety}
Lucet's code generator, Cranelift,
ensures that the control flow of the compiled code is restricted
to the sandbox and cannot be bent to bypass bounds checks (e.g., via
return-oriented programming).
This comes directly from preserving \wasm's semantics during
compilation.
For example, compiled code preserves \wasm's safe stack~\cite{wasm-paper,
safestack}, ensuring that stack frames (and thus return values on the stack)
cannot be clobbered.
The compiled code also enforces \wasm's coarse-grained CFI and, for example,
matches the type of each indirect call site with the type of the target.
\paragraph{Memory isolation}
When Lucet creates a \wasm sandbox, it reserves 4GB of virtual memory for the
\wasm heap and uses Cranelift to bound all heap loads and stores.
To this end, Cranelift (1) explicitly passes a pointer to the base of the heap
as the first argument to each function and (2) masks all pointers to be within
this 4GB range.
Like previous software-based isolation (SFI) systems~\cite{wahbe-sfi},
Cranelift avoids expensive bounds check operations by using guard pages to trap
any offsets that may reach beyond the 4GB heap space.
\paragraph{Embedding \wasm}
An application with embedded \wasm code will
typically require context switching between the \wasm code and host---e.g.,\xspace to
read data from a socket.
Lucet allows safe control and data flow across the host-sandbox
boundary via \emph{springboards} and \emph{trampolines}.
Springboards are used to enter \wasm code---they set up the program context
for \wasm execution---while trampolines are used to restore the host context
and resume execution in the host.
\paragraph{Using \wasm in FaaS platforms}
WebAssembly FaaS platforms like Fastly's Terrarium allow clents to deploy
scalable function-oriented Web and cloud applications written in any langauge
(that can be compiled to Wasm).
Clients compile their code to \wasm and upload the resulting module to the
platform; the platform handles scaling and isolation.
As shown in Figure~\ref{fig:faas}, FaaS platforms place thousands of client
\wasm modules within a single host process, and distribute these processes
across thousands of servers and multiple datacenters.
This is the key to scaling---it allows any host process, in any
datacenter, to spawn a fresh \wasm sandbox instance and run any client function
in response to a user request.
By using \wasm as an intermediate layer, FaaS platforms isolate the client
for free~\cite{wahbe-sfi, fastly-wasm, gobi, wasm-il, lucet-talk, rlbox}.
Unfortunately, this isolation does not hold in the presence of Spectre attacks.
\subsection{Spectre attacks}
\label{subsec:spectre-attacks}
Spectre attacks exploit hardware predictors to induce mispredictions and
speculatively execute instructions---\emph{gadgets}---that would not run
sequentially~\cite{Kocher2019spectre}.
Spectre attacks are classified by the hardware predictor they exploit~\cite{Canella2019}.
We focus on the three Spectre variants that hijack control flow:
\begin{CompactItemize}
\item{\bf Spectre-PHT}
Spectre-PHT~\cite{Kocher2019spectre} exploits the \emph{pattern history table} (PHT),
which is used as part of the \emph{conditional branch predictor} (CBP) to guess
the direction of a conditional branch while the condition is still being evaluated.
%
In a Spectre-PHT attack, the attacker (1) pollutes entries in the PHT so that a branch
is mispredicted to the wrong path.
%
The attacker can then use this wrong-path execution to
bypass memory isolation guards or control flow integrity.
\item{\bf Spectre-BTB}
Spectre-BTB~\cite{Kocher2019spectre} exploits the \emph{branch target buffer} (BTB),
which is used to predict the target of an indirect jump~\cite{branch-predictors}.
%
In a Spectre-BTB attack, the attacker pollutes entries in the BTB, redirecting
speculative control flow to an arbitrary target.
%
Spectre-BTB can thus be used to speculatively execute gadgets that are
not in the normal execution path (e.g.,\xspace to carry out a ROP-style attack).
\item{\bf Spectre-RSB}
Spectre-RSB~\cite{Maisuradze2018spectre5,Koruyeh2018spectre5} exploits the
\emph{return stack buffer} (RSB), which memorizes the
location of recently executed \asm{call} instructions to predict the targets
of \asm{ret} instructions.
%
In a Spectre-RSB attack, the attacker uses chains of \asm{call} or \asm{ret}
instructions to over- or underflow the RSB, and redirect speculative control
flow in turn.
\end{CompactItemize}
Spectre can be used \emph{in-place} or
\emph{out-of-place}~\cite{Canella2019}.
In an in-place attack, the attacker mistrains the prediction for a victim branch by
repeatedly executing the victim branch itself.
In an out-of-place attack, the attacker
finds a secondary branch that is \emph{congruent} to the victim branch---predictor
entries are indexed using a subset of address bits---and uses this
secondary branch to mistrain the prediction for the victim branch.
\subsection{Spectre attacks on FaaS platforms}
\label{sec:SpectreOnWasm}
A malicious FaaS platform client who can upload arbitrary \wasm code can force
the \wasm compiler to emit native code which is safe during sequential
execution, but uses Spectre to bypass \wasm's isolation guarantees
during speculative execution.
We identify three kinds of attacks (Figure~\ref{fig:wasm_spectre}):
\begin{figure}[t!]
\centering
\resizebox{.99\hsize}{!}{
\includegraphics[width=1\hsize]{figures/wasm_spectre.pdf}
}
\caption{A malicious tenant can fill branch predictors with invalid state
(red). In one scenario, the attacker causes its own branches to
speculatively execute code that access memory outside of the sandbox. In the
second and third scenarios, the attacker uses Spectre to respectively target
a victim sandbox or the host runtime to misspeculate and leak secret data.
}
\label{fig:wasm_spectre}
\end{figure}
\begin{CompactItemize}
\item \textbf{Scenario 1: \attackOneUpper attacks}
The attacker bends the speculative control flow of their own
module to access data outside the sandbox region.
For example, they can use Spectre-PHT to bypass conditional
bounds checks when accessing the indirect call table.
Alternatively, they can use Spectre-BTB to transfer the control flow into the
middle of instructions to execute unsafe code (e.g., code that bypasses
Wasm's implicit heap bounds checks).
\item \textbf{Scenario 2: \attackTwoUpper attacks}
The attacker uses an out-of-place Spectre attack to bend the control
flow of a victim sandbox and coerce the victim into leaking their own data.
Although this attack is considerably more sophisticated, we were still able to
implement proof of concept attacks following Canella et al.~\cite{Canella2019}.
Here, the attacker finds a (mispredicted) path in the victim sandbox that leads
to the victim leaking data, e.g.,\xspace through cache state.
They then force the victim to mispredict this path by using a congruent branch
within their own sandbox.
\item \textbf{Scenario 3: \attackThreeUpper attacks}
Instead of bending the control flow of the victim sandbox,
the attacker can use an out-of-place Spectre attack to bend the control flow of the
host runtime.
This
allows the attacker
to speculatively access data from the host as well as any other sandbox.
\end{CompactItemize}
\begin{figure}
\inputminted[linenos=true, breaklines=true]{myasm_lexer.py:MyAsmLexer -x}{code/example_attack.asm}
\vspace{-.2in}
\caption
{
A simplified snippet of the vulnerable code from our Spectre-PHT breakout attack.
%
This code is safe during sequential execution (it checks the index
\asmescaped{rcx} before using it to load a function table entry).
%
But, during speculative execution, control flow may bypass this check and
access memory outside the function table bounds.
}
\label{fig:attack_pinned_register}
\end{figure}
\noindent
\Cref{fig:attack_pinned_register} gives an example \attackOne gadget.
The gadget is in the implementation of the \wasm \asm{call_indirect}
instruction, which is used to call functions indexed in a module-level function
table. This code first compares the function index \asmescaped{rcx} to the length
of the function table (to ensure that \asmescaped{rcx} points to a valid
entry). If \asmescaped{rcx} is valid, it then jumps to \asmescaped{index_ok}, loads
the function from the corresponding entry in the table, and calls it; otherwise
the code traps.
An attacker can mistrain the conditional branch on line 3 and cause it to
speculatively jump to \asmescaped{index_ok} even when \asmescaped{rcx} is
out-of-bounds. By controlling the contents of \asmescaped{rcx}, the attacker
can thus execute arbitrary code locations outside the sandbox.
In Section~\ref{sec:pocs} we demonstrate several proof of concept attacks,
including a breakout attack that uses this gadget.
These attacks serve to highlight the importance of hardening \wasm against
Spectre.
\section{\sys: Hardening \wasm against Spectre}
\label{sec:design}
\begin{table}[tb!]
\centering
\footnotesize
\begin{tabular}{p{1.5cm}p{1.5cm}|p{0.6cm}p{0.6cm}p{0.6cm}p{0.6cm}}
\toprule
\multirow{2}{4cm}{\textbf{\pbox{\textwidth}{Attack variant}}}
&
& \multicolumn{2}{c|}{\textbf{\pbox{\textwidth}{\sysDesignOne}}}
& \multicolumn{2}{c}{\textbf{\pbox{\textwidth}{\sysDesignTwo}}}
\\\cline{3-6}
&
& \multicolumn{1}{c|}{\textbf{\pbox{\textwidth}{ASLR}}}
& \multicolumn{1}{c|}{\textbf{\pbox{\textwidth}{Det}}}
& \multicolumn{1}{c|}{\textbf{\pbox{\textwidth}{ASLR}}}
& \multicolumn{1}{c}{\textbf{\pbox{\textwidth}{Det}}} \\
\midrule
\multirow{2}{2cm}{Spectre-PHT}
& in-place & \fullcirc & \fullcirc & \fullcirc & \fullcirc \\
& out-of-place & \halfcirc & \fullcirc & \halfcirc & \fullcirc \\
\multirow{2}{2cm}{Spectre-BTB}
& in-place & \fullcirc & \fullcirc & \fullcirc & \fullcirc \\
& out-of-place & \fullcirc & \fullcirc & \fullcirc & \fullcirc \\
\multirow{2}{2cm}{Spectre-RSB}
& in-place & \fullcirc & \fullcirc & \fullcirc & \fullcirc \\
& out-of-place & \fullcirc & \fullcirc & \fullcirc & \fullcirc \\
\bottomrule
\end{tabular}
\caption{
Effectiveness of \sys against different Spectre variants.
A full circle indicates that \sys eliminates the attack while a half circle
indicates that \sys only mitigates the attack.
}
\label{tab:swivel-effectiveness}
\end{table}
\sys extends Lucet---and the underlying Cranelift code generator---to address
Spectre attacks on FaaS \wasm platforms.
We designed \sys with several goals in mind.
First, \emph{performance}: \sys minimizes the number of pipeline fences it
inserts, allowing Wasm to benefit from speculative execution as much as
possible.
Second, \emph{automation}: \sys does not rely on user annotations or source
code changes to guide mitigations; we automatically apply mitigations when
compiling \wasm.
Finally, \emph{modularity}: \sys offers configurable protection, ranging from
probabilistic schemes with high performance to thorough mitigations with
strong guarantees.
This allows \sys users to choose the most appropriate mitigations (see
Table~\ref{tab:swivel-effectiveness}) according to their application domain,
security considerations, or particular hardware platform.
In the rest of this section, we describe our attacker model and
introduce a core abstraction: \emph{linear blocks}.
We then show how linear blocks, together with several other techniques, are
used to address both sandbox breakout and poisoning attacks.
These techniques span two \sys designs: \sysDesignOne, a software-only scheme which
provides mitigations compatible with existing CPUs; and \sysDesignTwo, which
uses hardware extensions (\cet and \mpk) available in the 11th generation \Intel
CPUs.
\paragraph{Attacker model}
\label{subsec:attacker_model}
We assume that the attacker is a FaaS platform client who can upload arbitrary
\wasm code which the platform will then compile and run alongside other clients
using \sys.
The goal of the attacker is to read data sensitive to another (victim) client
using Spectre attacks.
In the \sysDesignTwo case, we only focus on exfiltration via the data
cache---and thus assume an attacker who can only exploit gadgets that leak via
the data cache.
We consider transient attacks that exploit the memory subsystem (e.g.,
Meltdown~\cite{Lipp2018meltdown}, MDS~\cite{Schwarz2019ZL,Canella2019Fallout,VanSchaik2019RIDL},
and LVI~\cite{vanbulck2020lvi}) out of scope and discuss this in detail
in~\cref{sec:otherattacks}.
We assume that our Wasm compiler and runtime are correct.
We similarly assume the underlying operating system is secure and the
latest CPU microcode updates are applied.
We assume hyperthreading is disabled for any \sys scheme except for
the deterministic variant of \sysDesignTwo.
Consistent with previous findings~\cite{branch-predictors}, we assume BTBs predict
the lower 32-bits of target addresses, while the upper 32-bits are inferred
from the instruction pointer.
\sys addresses attackers that intentionally extract information using Spectre.
We do not prevent clients from accidentally leaking secrets during sequential
execution and, instead, assume they use techniques like constant-time
programming to prevent such leaks~\cite{cauligi2020foundations}.
For all \sys schemes except the deterministic variant of \sysDesignTwo, we
assume that a sandbox cannot directly invoke function calls in other
sandboxes, i.e., it cannot control the input to another sandbox to perform an
in-place poisoning attack.
We lastly assume that host secrets can be protected by placing them in a \wasm
sandbox, and discuss this further in \Cref{subsec:host-secrets}.
\medskip
\subsection{Linear blocks: local \wasm isolation}
\label{subsec:linear_blocks}
To enforce \wasm's isolation sequentially and speculatively, \sysDesignOne and
\sysDesignTwo compile \wasm code to linear blocks (LBs).
Linear blocks are straight-line code blocks that do not contain
\emph{any} control flow instructions except for their terminators---this
is in contrast to traditional basic blocks, which typically do not
consider function calls as terminators.
This simple distinction is important: It allows us to ensure that \emph{all} control
flow transfers---both sequential and speculative---land on linear block boundaries.
Then, by ensuring that individual linear blocks are \emph{safe}, we can ensure
that whole \wasm programs, when compiled, are confined and cannot violate
\wasm's isolation guarantees.
A linear block is safe if, independent of the control flow into the block,
\wasm's isolation guarantees are preserved.
In particular, we cannot rely on safety checks (e.g.,\xspace bounds checks for memory
accesses) performed across linear blocks since, speculatively, blocks may not
always execute in sequential order (e.g.,\xspace because of Spectre-BTB).
When generating native code, \sys ensures that a linear block is safe by:
\paragraph{Masking memory accesses}
Since we cannot make any assumptions about the initial contents of registers,
\sys ensures that unconditional heap bounds checks (performed via masking)
are performed in the same linear block as the heap memory access itself.
We do this by modifying the Cranelift optimization passes which could lift
bounds checks (e.g., loop invariant code motion) to ensure that they don't
move masks across linear block boundaries.
Similarly, since we cannot trust values on the stack, \sys ensures
that any value that is unspilled from the stack and used in a bounds
check is masked again.
We use this \emph{mask-after-unspill} technique to replace
Cranelift's unsafe mask-before-spill approach.
\paragraph{Pinning the heap registers}
To properly perform bounds checks for heap memory accesses, a \sys linear
block must determine the correct value of the heap base.
Unfortunately, as described above, we cannot make any assumptions about the
contents of any register or stack slot.
\sys thus reserves one register, which we call the \emph{pinned heap
register}, to store the address of the sandbox heap.
Furthermore, \sys prevents any
instructions in the sandbox from altering the pinned heap register.
This allows each linear block to safely assume that the pinned heap register
holds the correct value of the heap base, even when the speculative control
flow of the program has gone awry due to
misprediction.
\paragraph{Hardening jump tables}
\wasm requires bounds checks on each access to indirect call tables and
switch tables.
\sys ensures that each of these bounds checks is local to the linear block
where the access is performed. Moreover, \sys implements the bounds check
using \emph{speculative load hardening}~\cite{RFC}, masking the index of
the table access with the length of the table.
This efficiently prevents the attacker from speculatively bypassing the bounds check.
\sys does not check the indirect jump targets
beyond what Cranelift does.
At the language level, \wasm already guarantees that the targets of indirect jumps (i.e.,\xspace
the entries in indirect call tables and switch tables) can only be the tops of
functions or switch-case targets.
Compiled, these correspond to the start of \sys linear blocks.
Thus, an attacker can only train the BTB with entries that point to linear
blocks, which, by construction, are safe to execute in any context.
\paragraph{Protecting returns}
\wasm's execution stack---in particular, return addresses on the stack---cannot
be corrupted during sequential execution.
Unfortunately, this does not hold speculatively:
An attacker can write to the stack (e.g.,\xspace with a buffer overflow) and speculatively execute a return
instruction which will divert the control flow to their location
of choice.
\sys ensures that return addresses on the stack cannot be corrupted as such,
even speculatively.
We do this using a \emph{separate stack}
or \emph{shadow stack}~\cite{sok-shadowstacks}, as we detail below.
\begin{figure*}[t!]
\centering
\resizebox{.90\hsize}{!}{
\includegraphics[width=1\hsize]{figures/swivel.pdf}
}
\caption{\sys hardens \wasm against spectre via compiler transformations.
In \sysDesignOne, we convert basic
blocks to linear blocks. Each linear block (e.g., \mono{lblock1\_1}
and
\mono{lblock1\_2}) maintains local security guarantees for speculative
execution.
Then, we protect the backward edge (\mono{block4}) by replacing the return
instructions
and using a separate return stack. To eliminate poisoning attacks, in
the deterministic version of \sysDesignOne, we further encode conditional
branches as indirect jumps.
%
In \sysDesignTwo, we similarly use linear blocks, but we allow return
instruction,
and protect returns using the hardware shadow stack. To reduce BTB
flushes,
we additionally use \cet's \mono{endbranch} to ensure that targets of
indirect branches
land at the beginning of linear blocks.
%
In the deterministic version, we avoid BTB flushing and instead use register
interlocking to prevent leakage on misspeculated paths.
}
\label{fig:design}
\end{figure*}
\subsection{\sysDesignOne}
\label{sec:sfi}
\sysDesignOne builds on top of linear blocks to address all three classes of
attacks.
\subsubsection{Addressing \attackOne attacks}
Compiling \wasm code to linear blocks eliminates most avenues for breaking out
of the sandbox.
The only way for an attacker to break out of the sandbox is to
speculatively jump into the middle of a linear block.
We prevent this with:
\paragraph{The separate stack}
We protect returns by preserving \wasm's safe return stack during
compilation.
Specifically, we create a separate stack in a fixed memory location, but outside
the sandbox stack and heap, to ensure that it cannot be overwritten by
sandboxed code.
We replace every call instruction with an instruction sequence that stores the
address of the subsequent instruction---the return address---to the next entry
in this separate stack. Similarly, we replace every return instruction with a
sequence that pops the address off the separate stack and jumps to that location.
To catch under- and over-flows, we surround the separate
stack with guard pages.
\paragraph{BTB flushing}
The other way an attacker can jump into the middle of a linear block is via a
mispredicted BTB entry.
Since all indirect jumps inside a sandbox can only point to the tops of linear
blocks, any such entries can only be set via a congruent entry outside any sandbox---i.e.,\xspace
an attacker must orchestrate the host runtime into mistraining a particular jump.
We prevent such attacks by flushing the BTB on transitions into and out
of the sandbox.\footnote{
In practice, BTB predictions are not absolute (as discussed in our attacker
model), instead they are 32-bit offsets relative to the instruction
pointer~\cite{branch-predictors}.
To ensure that this does not result in predictions at non linear block
boundaries, we restrict the sandbox code size to 4GB.
}
\subsubsection{Addressing sandbox and host poisoning attacks}
There are two ways for a malicious sandbox to carry out poisoning attacks:
By poisoning CBP or BTB entries.
Since we already flush the BTB to address \attackOne attacks, we trivially
prevent all BTB poisoning.
Addressing CBP poisoning is less straightforward.
We consider two schemes:
\paragraph{Mitigating CBP poisoning}
To mitigate CBP-based poisoning attacks, we use ASLR to randomize the layout of
sandbox code pages.
This mitigation is not sound---it is theoretically possible for an attacker
to influence a specific conditional branch outside of the sandbox.
As we discuss in Section~\ref{subsec:design_security},
this raises the bar to (at least) that of process isolation: The
attacker would would have to (1) de-randomize the ASLR of both their own
sandbox and the victim's and (2) find useful gadgets, which is itself an open problem (\S\ref{sec:related}).
\paragraph{Eliminating CBP poisoning}
\label{subsubsec:cbp-to-btb}
Clients that are willing to tolerate modest performance overheads
(\S\ref{sec:eval}) can opt to eliminate poisoning attacks.
We eliminate poisoning attacks by removing conditional branches from \wasm
sandboxes altogether.
Following~\cite{lee2017inferring}, we do this by using the \mono{cmov}
conditional move instruction to encode each conditional branch as an indirect
branch with only two targets (\Cref{fig:design}).
\input{design2}
\input{design_security}
\subsection{\sysDesignTwo}
\label{sec:cet}
\sysDesignOne avoids using fences to address Spectre attacks, but ultimately
bypasses all but the BTB predictors---and even then we flush the BTB on every
sandbox transition.
\sysDesignTwo
uses \Intel CET~\cite{intel-manual} and \Intel MPK~\cite{intel-manual}
to restore the use of the CBP and RSB, and avoid BTB flushing.\footnote{
Appendix~\ref{app:cet} gives a brief introduction to these new hardware features.
}
\subsubsection{Addressing \attackOne attacks}
Like \sysDesignOne, we build on linear blocks to address \attackOne attacks
(\Cref{fig:design}).
\sysDesignTwo, however, prevents an attacker from speculatively jumping into
the middle of a linear block using:
\paragraph{The shadow stack}
\sysDesignTwo uses \cet's \emph{shadow stack} to protect returns.
Unlike \sysDesignOne's separate stack, the shadow stack is a
hardware-maintained stack, distinct from the ordinary data stack and
inaccessible via standard load and store instructions.
The shadow stack allows us to use call and return instructions as
usual---the CPU uses the shadow stack to check the integrity of return
addresses on the program stack during both sequential and speculative
execution.
\paragraph{Forward-edge CFI}
Instead of flushing the BTB, \sysDesignTwo uses \cet's coarse-grained control
flow integrity (CFI)~\cite{cfi} to ensure that sandbox code can only jump to
the top of a linear block.
We do this by placing an \mono{endbranch} instruction at the beginning of every
linear block that is used as an indirect target (e.g.,\xspace the start of a function
that is called indirectly).
During speculative execution, if the indirect branch predictor targets
an instruction other than an \mono{endbranch} (e.g.,\xspace inside the host runtime),
the CPU stops speculating~\cite{intel-manual}.
\paragraph{Conditional BTB flushing}
When using ASLR to address \attackTwo attacks, we still need to
flush the BTB on transitions into each sandbox.
Otherwise, one sandbox could potentially jump to a linear block in
another sandbox.
Our deterministic approach to \attackTwo (described below), however,
eliminates BTB flushes altogether.
\subsubsection{Addressing host poisoning attacks}
To prevent \attackThree attacks, \sysDesignTwo uses \mpk.
\mpk exposes new user mode instructions that allow a process to partition its
memory into sixteen linear regions and to selectively enable/disable read/write
access to any of those regions.
\sysDesignTwo uses only two of these protection domains---one for the host and one shared by
all sandboxes---and switches domains during the transitions between host and sandbox.
When the host creates a new sandbox, \sysDesignTwo allocates the heap memory
for the new sandbox with the sandbox protection domain, and then relinquishes
its own access to that memory so that it is no longer accessible by the host.
This prevents \attackThree attacks by ensuring that the host cannot be coerced
into leaking secrets from another sandbox.
We describe how we safely copy data across the
boundary later (\S\ref{sec:impl}).
\subsubsection{Addressing sandbox poisoning attacks}
\label{subsubsec:cet_sandbox_poison}
By poisoning CBP or BTB entries, a malicious sandbox can coerce a
victim sandbox into executing a gadget that leaks sensitive data.
As with \sysDesignOne, we consider both a probabilistic and deterministic
design to addressing these attacks.
Since the probabilistic approach is like \sysDesignOne's, we describe only the
deterministic design.
\paragraph{Preventing leaks under poisoned execution}
\sysDesignTwo does \emph{not} eliminate cross-sandbox CBP or BTB poisoning.
Instead, we ensure that a victim sandbox cannot be coerced into leaking
data via the cache when executing a mispredicted path.
To leak secrets through the cache, the attacker must maneuver the secret
data to a gadget that will use it as an offset into a memory region.
In Cranelift, any such gadget will use the heap, as stack memory
is always accessed at constant offsets from the stack pointer (which itself
cannot be directly assigned).
We thus need only prevent leaks that are via the Wasm heap---we
do this using \emph{register interlocks}.
\paragraph{Register interlocking}
\label{subsubsec:interlocks}
Our register interlocking technique tracks the control flow of a \wasm program
and prevents it from accessing its stack or heap when the speculative path
diverges from the sequential path.
We first assign each non-trivial linear block a unique 64-bit
\emph{block label}.
We then calculate the expected block label of every direct or indirect branch
and assign this value to a reserved \emph{interlock register} prior to branching.
At the beginning of each linear block, we check that the value of the
interlock register corresponds to the static block label using \mono{cmov}
instructions.
If the two do not match, we zero out the heap base register as well as the stack
register.
Finally, we unmap pages from the address space to ensure that any access from a
zero heap or stack base will fault---and thus will not affect cache state.
The register interlock fundamentally introduces a data dependency
between memory operations and the resolution of
control flow.
In doing so, we prevent any memory operations that would result in cache based
leaks, but do not prevent all speculative execution.
In particular, any arithmetic operations may still be executed speculatively.
This is similar to hardware taint tracking~\cite{yu2019speculative}, but
enforced purely through compiler changes.
Finally, \wasm also stores certain data (e.g.,\xspace globals variables and internal
structures) outside the \wasm stack or heap.
To secure these memory accesses with the register interlock, we introduce an
artificial heap load in the address computation for this data.
\subsection{Security and performance trade-offs}
\label{subsec:design_security}
\sys offers two design points for protecting \wasm modules from Spectre
attacks: \sysDesignOne and \sysDesignTwo.
For each of these schemes we further consider probabilistic (ASLR) and
deterministic techniques.
In this section, we discuss the performance and security trade-offs when
choosing between these various \sys schemes.
\input{design_table}
\subsubsection{Probabilistic or deterministic?}
\Cref{tab:swivel-effectiveness} summarizes \sys's security guarantees.
\sys's deterministic schemes eliminate Spectre attacks, while
the probabilistic schemes eliminate Spectre attacks that exploit the BTB and
RSB, but trade-off security for performance when it comes to the CBP (\S\ref{sec:eval}):
Our probabilistic schemes only \emph{mitigate} Spectre attacks that exploit the
CBP.
To this end, (probabilistic) \sys hides branch offsets by randomizing code
pages.
Previously, similar fine-grain approaches to address randomization have been
proposed to mitigate attacks based on return-oriented
programming~\cite{crane2015s,gens2017lazarus}.
Specifically, when loading a module, \sys copies the code pages of the \wasm
module to random destinations, randomizing all but
the four least significant bits (LSBs) to keep 16-byte alignment.
This method is more fine-grained than page remapping, which would fail to randomize the
lower 12 bits for 4KB instruction pages.
Unfortunately, only a subset of address bits are typically used by hardware
predictors.
Zhang et. al~\cite{branch-predictors}, for example, found that only the 30
LSBs of the instruction address are used as input for BTB predictors.
Though a similar study has not been conducted for the CBP, if we
pessimistically assume that 30 LSBs are used for prediction then our
randomization offers at least 26 bits of entropy.
Since the attacker must de-randomize both their module and the victim
module, this is likely higher in practice.
As we show in Section~\ref{sec:eval}, the ASLR variants of \sys are faster than
the deterministic variants.
Using code page ASLR imposes less overhead than the deterministic techniques
(summarized in \Cref{tab:mitigations}).
This is not surprising: CBP conversion (in \sysDesignOne) and register
interlocking (in \sysDesignTwo) are the largest sources of performance
overhead.
For many application domains, this security-performance trade-off is
reasonable.
Our probabilistic schemes use ASLR only to mitigate \attackTwo
attacks---and
unlike \attackOne attacks, these attacks are significantly more challenging
for an attacker to carry out:
They must conduct an out-of-place attack on a specific target while accounting
for the unpredictable mapping of the branch predictor.
To our knowledge, such an attack has not been demonstrated, even
without the additional challenges of defeating ASLR.
Furthermore, on a FaaS platform, these attacks are even harder to pull off,
as the attacker has only a few hundred milliseconds to land an attack
on a victim sandbox instance before it finishes---and the
next victim instance will have entirely new mappings.
Previous work suggests that such an attack is not
practical in such a short time window~\cite{evtyushkin2018branchscope}.
For other application domains, the overhead of the deterministic \sys variants
may yet be reasonable. As we show in Section~\ref{sec:eval}, the average
(geometric) overhead of \sysDesignOne is \specSfiFullOverhead and that of
\sysDesignTwo is \specCetFullOverhead.
Moreover, users can choose to use \sysDesignOne and \sysDesignTwo according
to their trust model---\sys
allows sandboxes of both designs to coexist within a single process.
\subsubsection{Software-only or hardware-assisted?}
\sysDesignOne and \sysDesignTwo present two design points that have
different trade-offs beyond backwards compatibility.
We discuss their trade-offs, focusing on the deterministic variants.
\sysDesignOne eliminates Spectre attacks by allowing speculation only via the BTB predictor
and by controlling BTB entries through linear blocks and BTB flushing.
\sysDesignTwo, on the other hand, allows the other predictors.
To do this safely though, we use register interlocking to create data
dependencies (and thus prevent speculation) on certain operations after
branches.
Our interlock implementation only guards \wasm memory operations---this means
that, unlike \sysDesignOne, \sysDesignTwo only prevents \emph{cache-based}
\attackTwo attacks.
While non-memory instructions (e.g., arithmetic operations) can still
speculatively execute, register interlocks sink performance: Indeed, the
overall performance overhead of \sysDesignTwo is higher than
\sysDesignOne~(\S\ref{sec:eval}).
At the same time, \sysDesignTwo can be used to handle a more powerful
attacker model (than our FaaS model).
First, \sysDesignTwo eliminates poisoning attacks even in the presence of
attacker-controlled input.
This is a direct corollary of being able to safely execute code with poisoned
predictors.
Second, \sysDesignTwo (in the deterministic scheme) is safe in the presence of
hyperthreading; our other \sys schemes assume that hyperthreading is disabled
(\S\ref{subsec:attacker_model}).
\sysDesignTwo allows hyperthreading because it doesn't rely on BTB flushing;
it uses register interlocking to eliminate sandbox poisoning attacks.
In contrast, our SFI schemes require the BTB to be isolated for the host and each \wasm sandbox---an
invariant that may not hold if, for example,
hyperthreading interleaves host application and \wasm code on sibling threads.
\section{Limitations and discussion}
\label{sec:discussion}
In this section, we cover some of the current limitations of \sys, briefly
mention alternate design points, and address the generality of our
solutions.
\subsection{Limitations of \sys}
We discuss some limitations of \sys, both in general and for our
implementation in particular.
\paragraph{Implementation limitations}
For this paper, we have simplified some of the implementation
details for \sysDesignTwo to reduce the engineering burden of modifying
multiple compiler toolchains and standard libraries while still providing
accurate performance evaluations.
First, we do not ensure that interlock labels are unique to each linear block,
but rather reuse interlock labels;
while unique labels are critical for security, previous works have extensively
demonstrated the feasibility of assigning unique labels~\cite{cfi-survey}.
Our goal was to measure the performance of the instruction sequences for
interlock assignment (64-bit conditional moves) and checking (64-bit conditional checks).
Next, when disabling \mpk protections in \sysDesignTwo~(\S\ref{sec:cet}) in the host
calls, we must avoid using indirect branches; while we follow this principle
for hostcalls we expose (e.g., when marshaling data for web server requests), we
do not modify existing standard library hostcalls.
These additional modifications, while straightforward, would require significant
engineering effort in modifying the standard library (libc) used by \wasm.
Finally, we did not implement the required guard pages in the lower 4GB of
memory in our prototype of deterministic \sysDesignTwo.
Prior work~\cite{isboxing} has shown how to reserve the bottom 4GB of
memory---and that this does not impact performance.
\paragraph{Secretless host}
\label{subsec:host-secrets}
\sys assumes that the host (or runtime) doesn't contain secret information.
This assumption is sensible for some applications: in the CDN use case, the CDN
part of the process is lightweight and exists only to coordinate with the
sandboxes.
But not all.
As a counter-example, the Firefox web browser currently uses
\wasm to sandbox third-party libraries written in C/C++~\cite{rlbox,rlbox-blog}.
We could use \sys to ensure that Firefox is secure from Spectre attacks
conducted by a compromised third-party libraries.
To protect secrets in the host (Firefox), we could either place
the secrets into a separate \wasm sandbox, or apply one of our proposed
CBP protections to the host (e.g., CBP-to-BTB or interlocks).
\paragraph{Hyperthreading}
The only scheme in \sys that supports hyperthreading is the deterministic
\sysDesignTwo.
Alternately, instead of disabling hyperthreading, Intel suggests relying on
single-threaded indirect branch predictors (STIBP) to prevent a co-resident
thread from influencing indirect branch predictions~\cite{IntelSpeculation}.
STIBP could allow any \sys scheme to be used securely with hyperthreading.
\subsection{Other leakages and transient attacks} \label{sec:otherattacks}
\sysDesignTwo allows victim code to run with poisoned predictors but prevents
exfiltration via the data cache.
This, unfortunately, means that attackers may still be able to leak victim data
through other microarchitectural channels (e.g.,\xspace port contention or the
instruction cache~\cite{Kocher2019spectre}).
\sysDesignOne does not have this limitation; we can borrow techniques from
\sysDesignOne to eliminate such leaks (e.g., flushing the BTB).
The CPU's memory subsystem may also introduce other transient execution attacks.
Spectre-STL~\cite{Horn2018spectre4} can leak stale data (which may belong
to another security domain) before a preceding store could overwrite this data
due to speculative dependency checking of the load and the preceding stores.
\sys does not address Spectre-STL. However, Spectre-STL can been mitigated
through speculative store bypass disable~(SSBD)~\cite{ssbd}, which imposes a
small performance overhead (less than 5\% on most
benchmarks~\cite{phoronix-ssbd}) and is already enabled by default on most
systems.
Meltdown~\cite{Lipp2018meltdown} could leak privileged kernel memory from
userspace.
Variants of the Meltdown attack, e.g.,\xspace microarchitectural data sampling (MDS),
can be used to leak stale data from several microarchitectural
elements~\cite{Intel2019MDS,Canella2019Fallout,Schwarz2019ZL,moghimi2020Medusa,VanSchaik2019RIDL}.
Load Value Injection (LVI) exploits the same microarchitectural features as
Meltdown to inject data into the microarchitectural state~\cite{vanbulck2020lvi}.
More recent Intel CPUs (e.g., Tiger Lake) are designed to be resilient against
this class of attacks~\cite{intelMitigationList}, and we believe that these
attacks can efficiently be mitigated in hardware.
For this reason, recent research into secure speculation and architectural
defense against transient execution attacks are mostly focused on the Spectre class
of issues~\cite{guarnieri2018spectector,RFC,retpoline,wang2018oo7}.
For legacy systems, users should apply the latest microcode and software
patches to mitigate Meltdown and variants of MDS~\cite{kernelPTI,intelMitigationList}.
For variants of MDS that abuse hyperthreading on legacy systems,
Intel suggests safe scheduling of sibling CPU threads~\cite{Intel2019MDS}.
Since \wasm restricts what instructions are allowed in a \wasm module, this
makes some MDS attacks more challenging to execute.
For instance, \wasm modules cannot use \Intel TSX transactions or access
non-canonical or kernel addresses that are inherent to some of the MDS variants~\cite{TAA}.
LVI requires fine-grain control over inducing faults or microcode assists,
which is not available at the \wasm level.
Some legacy systems may still be vulnerable to LVI; however, the feasibility
of LVI attacks outside the Intel SGX environment is an open research
question~\cite{vanbulck2020lvi}.
\subsection{Alternate design points for \sys}
We next describe alternate designs for \sys and discuss the trade-offs of our
design choices.
\paragraph{CBP-to-BTB conversion in \sysDesignTwo}
Since register interlocking is more expensive than CBP-to-BTB conversion, a
reader may wonder whether the deterministic \sysDesignTwo could more
efficiently protect against \attackTwo using CBP-to-BTB conversion.
Unfortunately, since \sysDesignTwo does not flush the BTB in both directions,
CBP-to-BTB conversion is not sufficient to fully mitigate \attackTwo.
In addition to CBP-to-BTB conversion, \sysDesignTwo would also need to
use interlocking (without additional performance gain) or flush the BTB
both ways. In the latter case, we might as well use \sysDesignOne,
as the main advantage of using \cet in \sys is to avoid flushing the BTB.
\paragraph{Interlocking in \sysDesignOne}
Likewise, one may wonder about the benefits of using interlock in \sysDesignOne.
Unfortunately, for interlock to be useful, it requires hardware support.
First, we require \mpk to ensure that a sandbox can't confuse the host into
accessing (and then leaking) another sandbox's data.
Second, we require the \cet \mono{endbranch} instruction to
ensure that the sandbox cannot use BTB entries leftover from the host.
\paragraph{Partitioning shared resources}
A different approach to addressing Spectre would be to partition differ
hardware structures to ensure isolation.
For example, for the CBP, one approach would be to exploit the indexing
mechanism of branch predictors such that each sandbox uses an isolated
portion of the CBP.
Unfortunately doing this on existing CPUs is hard: superscalar CPUs use
complex predictors with multiple indexing functions,
and without knowledge of the underlying microarchitecture, we were unable to
experimentally find a way to partition the CBP.
Alternately, we could mitigate \attackThree---or even \attackTwo---attacks by
partitioning CPU cores, and preventing an attacker from running on the same core
as their victim.
This approach protects against \attackTwo and \attackThree, since branch
predictors are per-physical-core.
We tried this. Specifically, we implemented this mitigation in the \attackThree
context and measured its performance.
Unfortunately, requiring a core transition during every springboard and
trampoline is detrimental to performance, and this scheme was not competitive
with our chosen implementation.
\subsection{Generalizing \sys}
\sys's techniques are not specific to the Lucet compiler or runtime.
Our techniques can be applied to other \wasm compilers and runtimes, including
the just-in-time \wasm compilers used in the Chrome and Firefox browsers.
Our techniques can also be adopted to other software-based fault isolation
(SFI) compilers~\cite{tan-sfi-survey}.
Adopting \sys to the Native Client (NaCl) compiler~\cite{nacl, nacl64}, for
instance, only requires only a handful of changes.
For example, we wouldn't even need to add linear blocks: NaCl relies on
instruction bundles---32-byte aligned blocks of instructions---which are more
restrictive than our linear blocks (and satisfy our linear block invariants).
More generally, \sys can be adopted to other sandboxed languages and runtimes.
JavaScript just-in-time compilers are a particularly good fit.
Though JavaScript JITs are more complex than \wasm compilers, they
share a similar security model (e.g., JavaScript in the browser is untrusted)
and, in some cases, even share a common compilation pipeline.
For example, Cranelift---the backend used by Lucet and \sys---was designed to
replace Firefox's JavaScript and Wasm backend
implementations~\cite{cranelift-firefox}, and thus could transparently benefit
from our mitigations.
Beyond Cranelift, we think that adopting our linear blocks and code page
ASLR is relatively simple (e.g., compared to redesigning the browser to deal
with Spectre) and could make JavaScript Spectre attacks significantly more
difficult.
\subsection{Implementation bugs in \wasm}
Lehmann et al.~\cite{wasm-bugs} showed that some \wasm compilers and
runtimes, like prior SFI toolchains~\cite{tan-sfi-survey}, contain
implementation bugs.\footnote{
They also show that C memory safety bugs are still present within the Wasm
sandbox---this class of bugs is orthogonal and cannot alone be used to to
bypass Wasm's isolation guarantees.
}
For example, they showed that some Wasm runtimes fail to properly separate
the stack and heap.
Though they did not identify such bugs in Lucet, these classes of bugs are
inevitable---and, while identifying such bugs is important, this class of bugs
is orthogonal and well-understood in the SFI literature (and addressed, for
example, by VeriWasm~\cite{veriwasm}).
We focus on addressing Spectre attacks, which can fundamentally undermine the
guarantees of even bug-free \wasm toolchains.
\subsection{Future work}
\sys's schemes can benefit from extensions to compiler toolchains as well as
hardware to both simplify its mitigations and improve performance.
We briefly discuss some possible extensions and their benefits below.
\subsubsection{Compiler toolchain extensions}
We describe two performance optimizations for the \sysDesignTwo deterministic
scheme, and a way to improve the security of \sys's ASLR schemes.
\paragraph{Data dependent loops}
As discussed in Section~\ref{subsec:wasm_overhead}, the \sysDesignTwo
deterministic scheme imposes the greatest overheads in programs with
data-dependent loops---e.g., programs that iterate over strings or
linked lists (which loop until they find a null element).
\sys effectively serializes iterations of such data-dependent loops.
We expect that many other Spectre mitigation (see Section~\ref{sec:related}),
like speculative taint tracking~\cite{yu2019speculative},
would similarly slow down such programs.
One way to speed up such code is to replace the data-dependent loops with a
code sequence that first counts the expected number of iterations~($N$),
executes an \lfence, and then runs the original loop body for $N$ iterations.
This would introduce only a single stall in the loop and eliminate the
serialization between loop iterations.
\paragraph{Compiler secret tracking}
\sys currently assumes all locations in memory contain potentially secret data.
However, several works (e.g.,~\cite{wang2018oo7}) have proposed tracking
secrets in compiler passes.
This information can be used to optimize the \sysDesignTwo deterministic
scheme. In particular, any public memory access can be hoisted above the
register interlock to allow the memory to be accessed (and ``leaked'')
speculatively.
\paragraph{Software diversity}
\sys's ASLR variants randomize code pages.
We could additionally use software diversity to increase the entropy of our
probabilistic schemes~\cite{software-diversity, ilr}.
Software diversity techniques (e.g., \mono{nop} insertion) are
cheap~\cite{nop-diversity}, and since they do not affect the behavior of
branches, they can be used to specifically mitigate out-of-place Spectre-BTB
and Spectre-PHT attacks.
\subsubsection{Hardware extensions}
Hardware extensions can make \sys faster and simpler.
\paragraph{CBP flushing}
\sysDesignOne schemes rely on ASLR or CBP-to-BTB conversion to protect the CBP.
However, hardware support for CBP flushing could significantly speed up \sys.
Alternatively, hardware support for tagging predictor state (e.g., host code
and sandbox code) would allow \sys to isolate the CBP without flushing.
\paragraph{Dedicated interlock instructions}
The register interlocking used in deterministic \sysDesignTwo requires
several machine instructions in each linear block in order to assign and
check labels.
Dedicated hardware support for these operations could reduce
code bloat.
\paragraph{Explicit BTB prediction range registers}
The \sysDesignTwo deterministic scheme allocates unique 64-bit labels to each
linear block, which do not overlap across sandbox instances.
We could simplify and speed up this scheme with a hardware extension that can
be used to limit BTB predictions to a range of addresses.
With such an extension, \sys could set the prediction range during each
transition into the sandbox (to the sandbox region) and ensure that the BTB
could only predict targets inside the sandbox code pages.
This would eliminate out-of-place BTB attacks---and, with linear blocks, it
would eliminate breakout attacks in \wasm.
Finally, this would reduce code size: it would allow us to to reduce block
labels to, for example, 16 bits (since we only need labels to be unique within
the sandbox).
\section{Evaluation}
\label{sec:eval}
\begin{figure*}[htb!]
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=8.5cm]{figures/sightglass/baseline/combined.pdf}
\caption{Fence scheme overhead on Sightglass}
\label{fig:sightglass_fence}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=8.5cm]{figures/sightglass/ours/combined.pdf}
\caption{Swivel scheme overhead on Sightglass}
\label{fig:sightglass}
\end{subfigure}
\caption{
Performance overhead of \sys on the Sightglass benchmarks.
%
(a) On Sightglass, the baseline schemes \loadfence, \sysStrawman, and
\mincut incur geomean overheads of \sgLoadfenceOverhead,
\sgStrawmanOverhead, and \sgMincutOverhead respectively.
(b) In contrast, the Swivel schemes perform much better where the ASLR
versions of \sysDesignOne and \sysDesignTwo incur geomean
overheads of \sgSfiASLROverhead and \sgCetASLROverhead respectively.
%
With deterministic \attackTwo mitigations, these overheads are
\sgSfiFullOverhead and \sgCetFullOverhead.
%
}
\label{fig:wasm_overhead_sightglass}
\end{figure*}
\begin{figure*}[htb!]
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=8.5cm]{figures/spec/spec_results_baseline.pdf}
\caption{Fence scheme overhead on SPEC~2006}
\label{fig:spec_fence}
\end{subfigure}
\begin{subfigure}{0.49\textwidth}
\includegraphics[width=8.5cm]{figures/spec/spec_results_ours.pdf}
\caption{Swivel scheme overhead on SPEC~2006}
\label{fig:spec}
\end{subfigure}
\caption{
Performance overhead of \sys on SPEC~2006 benchmarks.
%
(a) On SPEC~2006, the baseline schemes \loadfence, \sysStrawman, and \mincut
incur overheads of \specLoadfenceMinOverhead--\specLoadfenceMaxOverhead,
\specStrawmanMinOverhead--\specStrawmanMaxOverhead, and
\specMincutMinOverhead--\specMincutMaxOverhead respectively.
%
(b) In contrast, the Swivel schemes perform much better where the ASLR
versions of \sysDesignOne and \sysDesignTwo incur overheads of at most
\specSfiASLRMaxOverhead and \specCetASLRMaxOverhead respectively.
%
With deterministic \attackTwo mitigations, these overheads are
\specSfiFullMinOverhead--\specSfiFullMaxOverhead and
\specCetFullMinOverhead--\specCetFullMaxOverhead respectively.
%
}
\label{fig:wasm_overhead_spec}
\end{figure*}
We evaluate \sys by asking four questions:
\begin{CompactItemize}
\item \textbf{What is the overhead of \wasm execution? (\S\ref{subsec:wasm_overhead})}
\sys's hardening schemes make changes to the code generated by the Lucet \wasm
compiler.
We examine the performance impact of these changes on Lucet's Sightglass
benchmark suite~\cite{sightglass} and Wasm-compatible
SPEC~2006~\cite{spec2006} benchmarks.
\item \textbf{What is the overhead of transitions? (\S\ref{subsec:transition_overhead})}
Swivel modifies the trampolines and springboards used to transition into and
out of Wasm sandboxes.
The changes vary across our different schemes---from adding {\lfence}s, or
flushing the BTB during one or both transition directions, to switching \mpk
domains.
We measure the impact of these changes on transition costs using a
microbenchmark.
\item \textbf{What is the end-to-end overhead of \sys? (\S\ref{subsec:application_overhead})}
We examine the impact of \wasm execution overhead and transition
overhead on a webserver that runs \wasm services.
We measure the impact of \sys protections on five different \wasm workloads
running on this webserver.
\item \textbf{Does \sys eliminate Spectre attacks? (\S\ref{sec:pocs})}
We evalute the security of \sys, i.e., whether \sys prevents sandbox breakout
and poisoning attacks, by implementing several
proof-of-concept Spectre attacks.
\end{CompactItemize}
\paragraph{Machine setup}
We run our benchmarks on a 4-core, 8-thread Tigerlake CPU software
development platform (2.7GHz with a turbo boost of 4.2GHz) supporting the \cet
extension.
The machine has 16~GB of RAM and runs 64-bit Fedora~32 with the 5.7.0 Linux
kernel modified to include \cet support~\cite{cet-kernel}.
Our \sys modifications are applied to Lucet version \mono{0.7.0-dev}, which
includes Cranelift version \mono{0.62.0}.
We perform benchmarks on standard SPEC~CPU~2006, and Sightglass version
\mono{0.1.0}.
Our webserver macrobenchmark relies on the Rocket webserver version
\mono{0.4.4}, and we use \mono{wrk} version \mono{4.1.0} for testing.
\subsection{\wasm execution overhead}
\label{subsec:wasm_overhead}
We measure the impact of \sys's Spectre mitigations on \wasm performance in
Lucet using two benchmark suites:
\begin{CompactItemize}
\item
The Sightglass benchmark suite~\cite{sightglass}, used by the Lucet compiler,
includes small functions such as cryptographic primitives (\mono{ed25519},
\mono{xchacha20}), mathematical functions (\mono{ackermann}, \mono{sieve}), and
common programming utilities (\mono{heapsort}, \mono{strcat}, \mono{strtok}).
\item
SPEC~CPU~2006 is a popular performance benchmark that includes various integer
and floating point workloads from typical programs.
We evaluate on only the subset of the benchmarks from SPEC~2006 that are
compatible with \wasm and Lucet. This excludes programs written in Fortran,
programs that rely on dynamic code rewriting, programs that require more
than 4GB of memory which \wasm does not support, and programs that use
exceptions, \mono{longjmp}, or multithreading.\footnote{Some Web-focused \wasm
platforms support some of these features.
Indeed, previous academic work evaluates these benchmarks on
\wasm~\cite{notsofast}, but non-Web \wasm platforms including Lucet do not
support them.}
We note that \sys does \emph{not} introduce new incompatibilities with SPEC~2006
benchmarks; all of \sys's schemes are compatible with the same benchmarks as
stock Lucet.
\end{CompactItemize}
\paragraph{Setup}
We compile both Sightglass and the SPEC~2006 benchmarks with our modified Lucet
\wasm compiler and run them with the default settings.
Sightglass repeats each test at least 10 times or for a total of 100ms
(whichever occurs first) and reports the median runtime.
We compare \sys's performance overhead
with respect to the performance of the same benchmarks when using the stock
Lucet compiler.
For increased measurement consistency on short-running benchmarks, while
measuring Sightglass we pin execution to a single core and disable CPU
frequency scaling.\footnote{Tests were performed on June 18, 2020; see testing disclaimer~\ref{sec:app:testing-disclaimer}.}
\paragraph{Baseline schemes}
In addition to our comparison against Stock Lucet, we also implement three
known Spectre mitigations using {\lfence}s and compare against these as a
reference point.
First, we implement \loadfence, which places an \lfence after every load,
similar to Microsoft's Visual Studio compiler's ``Qspectre-load''
mitigation~\cite{mscvcspectre}.
Next, we implement \sysStrawman, a scheme which restricts code to sequential
execution by placing an \lfence at all control flow targets (at the start of
all linear blocks)---this is similar to the Intel compiler's ``all-fix-lfence''
mitigation~\cite{icc-spectre}.
Finally, we implement \mincut, an \lfence insertion algorithm suggested by
Vassena et~al.\xspace\cite{blade} which uses a min-cut algorithm to minimize the
number of required {\lfence}s.
We further augment \mincut's \lfence insertion with several of our own
optimizations, including (1) only inserting a single \lfence per linear block;
and (3) unrolling loops to minimize branches, as we do for the register
interlock scheme and CBP conversions~(\S\ref{subsubsec:cbp-to-btb}).
Finally, to ensure that unrolling loops does not provide an unfair advantage,
we also present results for \mono{Stock-Unrolled}, which is stock Lucet with
the same loop unrolling as used in \sys's schemes.
\paragraph{Results}
We present the \wasm execution overhead of the various protection options
on the Sightglass benchmarks in \Cref{fig:sightglass}, and on the SPEC~2006
benchmarks in \Cref{fig:spec}.
The overheads of the ASLR versions of \sysDesignOne and \sysDesignTwo are
small: \sgSfiASLROverhead and \sgCetASLROverhead geomean overheads
respectively on Sightglass, and at most \specSfiASLRMaxOverhead (geomean:
\specSfiASLROverhead) and at most \specCetASLRMaxOverhead (geomean:
\specCetASLROverhead) respectively on SPEC.
The deterministic versions of \sys introduce modest
overheads: \sgSfiFullOverhead and \sgCetFullOverhead on Sightglass, and
\specSfiFullMinOverhead--\specSfiFullMaxOverhead (geomean:
\specSfiFullOverhead) and \specCetFullMinOverhead--\specCetFullMaxOverhead
(geomean: \specCetFullOverhead) on SPEC.
All four configurations outperform the baseline schemes by orders of magnitude:
\sysStrawman incurs geomean overheads of \sgStrawmanOverhead and
\specStrawmanOverhead on Sightglass and SPEC respectively, \loadfence incurs
\sgLoadfenceOverhead and \specLoadfenceOverhead overhead respectively, while
\mincut incurs \sgMincutOverhead and \specMincutOverhead respectively.
\paragraph{Breakdown}
Addressing CBP poisoning (CBP-to-BTB conversion in \sysDesignOne and register
interlocks in \sysDesignTwo) dominates the performance overhead of our
deterministic implementations.
We confirm this hypothesis with a microbenchmark: We measure the overheads of
these techniques individually on stock Lucet (with our loop
unrolling flags).
We find that the average (geomean) overhead of CBP conversion is
\sgPHTtoBTBOverhead on Sightglass and \specPHTtoBTBOverhead on SPEC.
The corresponding overheads for register interlocks are \sgInterlockOverhead
and \specInterlockOverhead.
Increasing loop unrolling thresholds does not significantly improve the
performance of stock Lucet (e.g., the speedup of our loop unrolling on stock
Lucet is \sgStockUnrollOverhead and \specStockUnrollOverhead on Sightglass and
SPEC, respectively).
It does impact the performance of our deterministic \sys variants though (e.g.,
we find that it contributes to a 15\%-20\% speed up).
This is not surprising since loop unrolling results in fewer conditional
branches (and thus reduces the effect of CBP conversions and register
interlocking).
To understand the outliers in \Cref{fig:wasm_overhead_sightglass} and
\Cref{fig:wasm_overhead_spec}, we inspect the source of the benchmarks.
Some of the largest overheads in \Cref{fig:wasm_overhead_sightglass} are on
Sightglass' string manipulation benchmarks, e.g.,\xspace \mono{strcat} and \mono{strlen}.
These microbenchmarks have lots of \emph{data-dependent loops}---tight loops
with data-dependent conditions---that cannot be unrolled at compile-time.
Since our register interlocking inserts a data dependence between the pinned
heap base register and the loop condition, this prevents the CPU from
speculatively executing instructions from subsequent iterations.
We believe that similar data-dependent loops are largely the cause for the
slowdowns on SPEC benchmarks, including \mono{429.mcf} and \mono{401.bzip2}.
Some of the other large overheads in Sightglass (e.g., \mono{fib2} and
\mono{nestedloop3}) are largely artifacts of the benchmarking suite:
These microbenchmarks test simple constructs like loops---and CBP-to-BTB
conversion naturally makes (almost empty) loops slow.
\subsection{Sandbox transition overhead}
\label{subsec:transition_overhead}
We evaluate the overhead of context switching.
As described in \Cref{sec:design}, \sys adds an {\lfence} instruction to
host-sandbox transitions to mitigate \attackOne attacks.
In addition to this:
\sysDesignOne flushes the BTB during each transition;
\sysDesignTwo, in deterministic mode, switches \mpk domains during each
transition;
and \sysDesignTwo, in ASLR mode, flushes the BTB in one direction and switches
\mpk domain in each transition.
We measure the time required for the host application to invoke a simple
no-op function call in the sandbox, as well as the time required for the
sandboxed code to invoke a permitted function in the application (i.e.,\xspace perform
a callback).
We compare the time required for \wasm code compiled by stock Lucet with the
time required for code compiled with our various protection schemes.
We measure the average performance overhead across 1000 such function call
invocations.\footnote{Tests were performed on June 18, 2020; see testing disclaimer~\ref{sec:app:testing-disclaimer}.}
These measurements are presented in
\Cref{tab:transitions}.
\begin{table}[t]
\centering
\footnotesize
\begin{tabular} {p{5cm}|c|c}
\toprule
\textbf{\pbox{\textwidth}{Transition Type}} &
\textbf{\pbox{\textwidth}{Function\\Invoke}} &
\textbf{\pbox{\textwidth}{Callback\\Invoke}} \\
\midrule
Stock &
\stockTransition &
\stockTransitionCb \\
\sysDesignOne (\lfence + BTB flush both ways) &
\sfiTransition &
\sfiTransitionCb \\
\sysDesignTwo ASLR (\lfence + BTB flush one way + MPK) &
\cetTransitionAslr &
\cetTransitionAslrCb \\
\sysDesignTwo deterministic (\lfence + MPK) &
\cetTransition &
\cetTransitionCb \\
\bottomrule
\end{tabular}
\caption{
Time taken for transitions between the application and sandbox---for
function calls into the sandbox and callback invocations from the
sandbox.
%
\sys overheads are generally modest, with the deterministic variant of
\sysDesignTwo in particular imposing very low overheads.
}
\label{tab:transitions}
\end{table}
\begin{table*}[t!]
\centering
\input{figures/wrk_table_1.tex}
\caption{
Average latency (\textbf{ALat}), 99\% tail latency (\textbf{TLat}), average throughput
(\textbf{Tput}) in requests/second and binary files size (\textbf{Size}) for the
webserver with different \wasm workloads
(1k = $10^3$, 1m = $10^6$).
}
\label{tab:webserver_latthrough_1}
\end{table*}
\begin{table}[h]
\centering
\input{figures/wrk_table_2.tex}
\caption{
Average latency (\textbf{ALat}), 99\% tail latency (\textbf{TLat}), average throughput
(\textbf{Tput}) in requests/second and binary files size (\textbf{Size}) for the
webserver for a long-running, compute-heavy \wasm workload
(1k = $10^3$, 1m = $10^6$).
}
\label{tab:webserver_latthrough_2}
\end{table}
First, we briefly note that function calls in stock Lucet take much longer
than callbacks.
This is because the Lucet runtime has not fully optimized the function call
transition, as these are relatively rare compared to callback transitions,
which occur during every syscall.
In general, \sys's overheads are modest, with the deterministic variant of
\sysDesignTwo in particular imposing very low overheads.
Flushing the BTB does increase transition costs, but the overall effect of
this increase depends on how frequently transitions occur between the
application and sandbox.
In addition, flushing the BTB affects not only transition performance but
also the performance of both the host application and sandboxed code.
Fully understanding these overheads requires that we evaluate the
overall performance impact on real world
applications, which we do next.
\subsection{Application overhead}
\label{subsec:application_overhead}
We now evaluate \sys's end-to-end performance impact on a
webserver which uses \wasm to host isolated web services.
\paragraph{Setup}
For this benchmark, we use the Rocket webserver~\cite{rocket}, which can
host web services written as \wasm modules.
Rocket operates very similarly to webservers used in previous academic papers
exploring \wasm modules~\cite{hall2019execution, faasm} as well as frameworks
used by CDNs such as Fastly.
We measure the webserver's performance while hosting five different web
services with varying CPU and IO profiles.
These services perform the following five tasks respectively:
(1) expanding an HTML template;
(2) converting XML input to JSON output;
(3) re-encoding a JPEG image to change image quality;
(4) computing the SHA-256 hash of a given input;
and (5) performing image classification using inference on a pretrained neural
network.
We measure the overall performance of the webserver by tracking the average
latency, 99\% tail-latency, and throughput for each of the five web services.
We also measure the size of the \wasm binaries produced.\footnote{Tests were performed on June 18, 2020; see testing disclaimer~\ref{sec:app:testing-disclaimer}.}
\paragraph{Results}
Tables~\ref{tab:webserver_latthrough_1} and~\ref{tab:webserver_latthrough_2}
show results of the webserver measurements.
From the table, we see any of sys's schemes only reduce geomean throughput
(across all workloads) between \cdnMinOverhead and \cdnMaxOverhead.
\sys also modestly increases \wasm binary sizes, particularly with its
deterministic schemes, due to additional instructions added for
separate stack, CBP-to-BTB, and interlock mechanisms.
For long-running, compute-heavy \wasm workloads such as JPEG re-encoding and
image classification, \sys's performance overhead is dominated by \wasm
execution overhead measured in \Cref{subsec:wasm_overhead}.
Thus, on these workloads the ASLR versions of \sys perform much better than
the deterministic versions, as their \wasm execution overhead is lower.
On the other hand, for short-running workloads such as templated HTML, we
observe that the deterministic schemes outperform the ASLR schemes.
This is because \sys's ASLR implementation must remap and \mono{memcpy} the
sandbox code pages during sandbox creation, effectively adding a fixed
overhead to each request.
For short-running requests, this fixed per-request cost dominates overall
overhead.
In contrast, Stock Lucet and \sys's deterministic schemes take advantage of
shared code pages in memory to create sandboxes more rapidly, incurring lower
overhead on short-running requests.
\subsection{Security evaluation}
\label{sec:pocs}
To evaluate the security of \sys, we implement several Spectre attacks in
\wasm and compile this attack code with both stock Lucet and \sys.
We find that stock Lucet produces code that is vulnerable to Spectre, i.e., our
proof of concept attacks (POCs) can be used to carry out both breakout and
poisoning attacks, and that \sys mitigates these attacks.
\paragraph{Attack assumptions}
Our attacks extend Google's Safeside~\cite{safeside} suite and, like the
Safeside POCs, rely on three low-level instructions: The \mono{rdtsc}
instruction to measure execution time, the \mono{clflush} instruction to evict
a particular cache line, and the \mono{mfence} instruction to wait for pending
memory operations to complete.
While these instructions are not exposed to \wasm code by default, we expose
these instructions to simplify our POCs.
Additionally, for cross \wasm module attacks, we manually specify the locations
where Wasm modules are loaded to simplify the task of finding partial address
collisions in the branch predictor.
Our simplifications are not fundamental and can be removed in an end-to-end
attack.
Previous work, for example, showed how to construct precise
timers~\cite{schwarz2017fantastic,frigo2018grand}, and how to control cache
contents~\cite{vila2019theory} in environments like JavaScript where these
instructions are not directly exposed.
The effects of the \mono{mfence} instruction can be achieved by
executing \mono{nop} instructions until all memory operations are drained.
And, in the style of heap and JIT spraying attacks~\cite{sintsov2010jit}, we
can increase the likelihood of partial address collision by deploying hundreds
to thousands of modules on the FaaS platform.
\paragraph{POC~1: Sandbox breakout via in-place Spectre-PHT}
Our first POC adopts the original Spectre-PHT bounds-check bypass
attack~\cite{Kocher2019spectre} to \wasm.
As mentioned in Section~\ref{sec:SpectreOnWasm}, in \wasm, indirect function
calls are expressed as indices into a function table.
Hence, the code emitted for the \asm{call_indirect} instruction performs a
bounds check, to ensure that the function index is within the bounds of the
table, before performing the lookup and call.
By inducing a misprediction on this check, our POC can read beyond the function
table boundary and treat the read value as a function pointer.
This effectively allows us to jump to any code location and speculatively
bypass \wasm's CFI (and thus isolation).
We demonstrate this by jumping to a host function that returns bytes of a
secret array.
\paragraph{POC~2: Sandbox breakout and poisoning via out-of-place Spectre-BTB}
Our second POC adopts the out-of-place Spectre-BTB attack of
Canella et al.~\cite{Canella2019} to Wasm.
Specifically, we mistrain an indirect jump in a victim or attacker-controlled module
by training a congruent indirect jump instruction in another attacker-controlled module.
We train the jump to land on a gadget of our choice.
To demonstrate the feasibility of a sandbox poisoning attack, we target a
double-fetch leak gadget.
To demonstrate a sandbox breakout attack, we jump in the middle of a basic
block to a memory load, skipping \wasm's heap bounds checks.
\paragraph{POC~3: Poisoning via out-of-place Spectre-RSB}
Our third POC compiles the Spectre-RSB attack from the Google Safeside
project~\cite{safeside} to Wasm.
This attack underflows the RSB to redirect speculative control flow.
We use this RSB underflow behavior to speculatively ``return'' to a gadget
that leaks module secrets.
We run this attack entirely within a single \wasm module.
However, on a FaaS platform this attack can be used across modules when the
FaaS runtime interleaves the execution of multiple modules, similar to the
Safeside cross-process Spectre-RSB attack.
\paragraph{Results}
We developed our POCs on a Skylake machine (Xeon Platinum 8160) and then tested
them on both this machine and the Tiger Lake \cet development platform we used
for our performance evaluation.
We found that stock Lucet on the Skylake machine was vulnerable to all three
POCs while \sysDesignOne, both the ASLR and deterministic versions, were not
vulnerable.
On the Tiger Lake machine, we found that stock Lucet was vulnerable to POC~3
while \sysDesignOne and \sysDesignTwo, both the ASLR and deterministic
versions, were not.
Although the Tiger Lake CPU is documented to be vulnerable to
all three Spectre variants~\cite{intelMitigationList}, we did not successfully
reproduce POC~1 and POC~2 on this machine.
Getting these attacks to work may require reverse engineering the branch
predictors used on these new CPUs. We thus leave the extensions of our
POCs to this microarchitecture to future work.
\section{Conclusion}
\label{sec:conclusion}
This work proposes a framework, \sys, which provides strong in-memory
isolation for \wasm modules by protecting against Spectre attacks.
We describe two \sys designs: \sysDesignOne, a software-only approach which
provides mitigations compatible with existing CPUs, and \sysDesignTwo, which
leverages \cet and \mpk.
Our evaluation shows that versions of \sys using ASLR incur low performance
overhead (at most \specASLRMaxOverhead on compatible SPEC~2006 benchmarks),
demonstrating that \sys can provide strong security guarantees for \wasm
modules while maintaining the performance benefits of in-process
sandboxing.
\section{Implementation}
\label{sec:impl}
We implement \sys on top of the Lucet \wasm compiler and
runtime~\cite{lucet-talk,fastly-wasm}.
In this section, we describe our modifications to Lucet.
We largely implement \sysDesignOne and \sysDesignTwo as passes in Lucet's code
generator Cranelift.
For both schemes, we add support for pinned heap registers and add direct
\mono{jmp} instructions to create linear block boundaries.
We modify Cranelift to harden switch-table and indirect-call accesses: Before
loading an entry from either table, we truncate the index to the length
of the table (or the next power of two) using a bitwise mask.
We also modify Lucet's stack overflow protection: Lucet emits conditional
checks to ensure that the stack does not overflow; these checks are rare and we
simply use \lfences.
We modify the springboard and trampoline transition functions in
the Lucet runtime.
Specifically, we add a single \lfence to each transition function since we must
disallow speculation from crossing the host-sandbox boundary.
The deterministic defenses for both \sysDesignOne and \sysDesignTwo---CBP
conversion and register interlocks---increase the cost of conditional
control flow.
To reduce the number of conditional branches, we thus enable explicit loop
unrolling flags when compiling the deterministic schemes.\footnote{
For simplicity, we do this in the Clang compiler when compiling applications to
Wasm and not in Lucet proper.
}
This is not necessary for the ASLR-based variants since they do not modify
conditional branches.
Indeed, the ASLR variants are straightforward modifications to the dynamic
library loader used by the Lucet runtime:
Since all sandbox code is position independent, we just copy a new sandbox
instance's code and data pages to a new randomized location in memory.
We also made changes specific to each \sys scheme:
\paragraph{\sysDesignOne}
We augment the Cranelift code generation pass to
replace \mono{call} and \mono{return} instructions with the \sysDesignOne
separate stack instruction sequences and we mask pointers when they are unspilled from the stack.
For the deterministic variant of \sysDesignOne, we also
replace conditional branches with indirect jump instructions,
as described in \Cref{sec:sfi}.
To protect against \attackTwo attacks~(\S\ref{sec:SpectreOnWasm}), we flush the
BTB during the springboard transition into any sandbox.
Since this is a privileged operation, we implement this using a custom Linux
kernel module.
\paragraph{\sysDesignTwo}
In the \sysDesignTwo code generation pass, we place \mono{endbranch} instructions at
each indirect jump target in \wasm to enable \cet protection.
These indirect jump targets include switch table entries and functions which
may be called indirectly.
We also use this pass to emit the register interlocks for the deterministic variant
of \sysDesignTwo.
We adapt the springboard and trampoline transition
functions to ensure that all uses of \mono{jmp}, \mono{call} and \mono{return}
conform to the requirements of \cet.
We furthermore use these transition functions to switch between the application
and sandbox \mpk domains.
Since \mpk blocks the application from accessing sandbox memory, we add
primitives that briefly turn off \mpk to copy memory into and out of sandboxes.
We implement these primitives using the \mono{rep} instruction prefix instead
of branching code, ensuring that the primitives are not vulnerable to Spectre attacks during
this window.
Finally, we add \cet support to both the Rust compiler---used to compile
Lucet---and to Lucet itself so that the resulting binaries are compatible with
the hardware.
\section{Introduction}
\label{sec:intro}
WebAssembly (\wasm) is a portable bytecode originally designed to safely run
native code (e.g., C/C++ and Rust) in the browser~\cite{wasm-paper}.
Since its initial design, though, \wasm has been increasingly used to
sandbox untrusted code outside the browser.
For example, Fastly and Cloudflare use \wasm to sandbox client applications
running on their edge clouds---where multiple client applications run within a
single process~\cite{fastly-wasm,cloudflare-workers}.
Mozilla uses \wasm to sandbox third-party C/C++ libraries in
Firefox~\cite{rlbox,rlbox-blog}.
Yet others use \wasm to isolate untrusted code in serverless
computing~\cite{hall2019execution}, IoT applications~\cite{microwasm},
games~\cite{flightsim}, trusted execution environments~\cite{enarxsgxwasm},
and even OS kernels~\cite{nebulet}.
In this paper, we focus on hardening Wasm against Spectre attacks---the class
of transient execution attacks which exploit control flow
predictors~\cite{Kocher2019spectre}.
Transient execution attacks which exploit features within the memory subsystem
(e.g., Meltdown~\cite{Lipp2018meltdown},
MDS~\cite{Schwarz2019ZL,Canella2019Fallout,VanSchaik2019RIDL}, and Load Value
Injection~\cite{vanbulck2020lvi}) are limited in scope and have already
been fixed in recent microarchitectures~\cite{intelMitigationList} (see
Section~\ref{sec:otherattacks}).
In contrast, Spectre can allow attackers to bypass \wasm's isolation boundary on
almost all superscalar
CPUs~\cite{AMDSpeculation,IntelSpeculation,AppleSpeculation}---and,
unfortunately, current mitigations for Spectre cannot be implemented entirely in
hardware~\cite{mcilroy2019spectre,yu2019speculative,barber2019specshield,koruyeh2019speccfi,venkman,RFC,jenkins2020ghostbusting,taram2019context}.
On multi-tenant serverless, edge-cloud, and function as a service (FaaS)
platforms, where Wasm is used as \emph{the} way to isolate mutually distursting
tenants, this is particulary concerning:\footnote{
Though our techniques are general, for simplicity we henceforth focus on Wasm
as used on FaaS platforms.
}
A malicious tenant can use Spectre to break out of the sandbox and read another
tenant's secrets in two steps (\S\ref{sec:pocs}).
First, they \emph{mistrain} different components of the underlying control flow
prediction---the conditional branch predictor (CBP), branch target buffer
(BTB), or return stack buffer (RSB)---to speculatively execute code that
accesses data outside the sandbox boundary.
Then, they \emph{reconstruct} the secret data from the underlying
microarchitectural state (typically the cache) using a side channel (cache
timing).
One way to mitigate such Spectre-based \emph{sandbox breakout attacks} is to
partition mutually distrusting code into separate processes.
By placing untrusted code in a separate process we can ensure that
the attacker cannot access secrets.
Chrome and Firefox, for example, do this by partitioning different \emph{sites}
into separate processes~\cite{chrome-site-isol, site-isolation-paper,
mozilla-sandbox}.
On a FaaS platform, we could similarly place tenants in separate processes.
Unfortunately, this would still leave tenants vulnerable to cross-process
\emph{\attackTwo
attacks}~\cite{IntelSpecMit,Canella2019,Koruyeh2018spectre5}.
Specifically, attackers can poison hardware predictors to coerce a victim
sandbox to speculatively execute gadgets that access secret data---from their
own memory region---and leak it via the cache (e.g., by branching on the
secret).
Moreover, using process isolation would sacrifice \wasm's scalability
(running many sandboxes within a process) and performance (cheap startup times
and context
switching)~\cite{fastly-wasm,cloudflare-workers,rlbox,hall2019execution}.
The other popular approach, removing speculation within the sandbox, is
also unsatisfactory.
For example, using pipeline fences to restrict \wasm code to sequential execution imposes
a \specStrawmanMinOverhead--\specStrawmanMaxOverhead slowdown on SPEC~2006
(\S\ref{sec:eval}).
Conservatively inserting pipeline fences before every dynamic load---an
approach inspired by the mitigation available in Microsoft's Visual Studio
compiler~\cite{mscvcspectre}---is even worse: it incurs
a \specLoadfenceMinOverhead--\specLoadfenceMaxOverhead overhead on SPEC
(\S\ref{sec:eval}).
In this paper, we take a compiler-based approach to hardening \wasm against
Spectre, without resorting to process isolation or the use of fences.
Our framework, \sys, addresses not only \attackOne and \attackTwo attacks, but
also \emph{\attackThree attacks}, i.e., Spectre attacks that coerce the process
hosting the \wasm sandboxes into leaking sensitive data.
That is, \sys ensures that a malicious \wasm tenant cannot
speculatively access data outside their sandbox nor coerce another tenant or
the host to divulge secrets of other sandboxes via poisoning.
We develop \sys via three contributions:
\paragraph{1. Software-only Spectre hardening (\S\ref{sec:sfi})}
Our first contribution, \sysDesignOne, is a software-only
approach to hardening \wasm against Spectre.
\sysDesignOne eliminates \attackOne attacks by compiling \wasm code to
\emph{linear blocks (LBs)}.
Linear blocks are straight-line x86 code blocks that satisfy two invariants:
(1) all transfers of control, including function calls, are at the block
boundary---to (and from) other linear blocks; and (2)
all memory accesses within a linear block are masked to the sandbox memory.
These invariants are necessary to ensure that the speculative control and data
flow of the \wasm code is restricted to the sandbox boundary.
They are not sufficient though: \sysDesignOne must also be tolerant of
possible RSB underflow.
We address this by (1) not emitting \mono{ret} instructions and therefore
completely bypassing the RSB and (2) using a \emph{separate stack} for return
addresses.
To address poisoning attacks, \sysDesignOne must still account for a poisoned
BTB or CBP.
Since these attacks are more sophisticated, we evaluate two different ways of
addressing them, and allow tenants to choose between them according to their
trust model.
The first approach uses address space layout randomization (ASLR) to randomize
the placement of each \wasm sandbox and flushes the BTB on each sandbox
boundary crossing.
This does not eliminate poisoning attacks; it only raises the bar of \wasm
isolation to that of process isolation.
Alternately, tenants can opt to eliminate these attacks altogether; to this
end, our deterministic \sysDesignOne rewrites conditional branches to indirect
jumps---thereby completely bypassing the CBP (which cannot be directly flushed)
and relying solely on the BTB (which can).
\paragraph{2. Hardware-assisted Spectre hardening (\S\ref{sec:cet})}
Our second contribution, \sysDesignTwo, restores the use of all predictors,
including the RSB and CBP, and partially obviates the need for BTB flushing.
It does this by sacrificing backwards compatibility and using new hardware security
extensions: Intel's Control-flow Enforcement Technology
(CET)~\cite{intel-manual} and Memory Protection Keys (MPK)~\cite{intel-manual}.
Like \sysDesignOne, \sysDesignTwo relies on linear blocks to address \attackOne
attacks.
But \sysDesignTwo does not avoid \mono{ret} instructions.
Instead, we use \cet's hardware \emph{shadow stack} to ensure that the RSB
cannot be misused to speculatively return to a location that is different from
the expected function return site on the stack~\cite{intel-manual}.
To eliminate \attackThree attacks, we use both \cet and \mpk.
In particular, we use \mpk to partition the application into two
domains---the host and \wasm sandbox(es)---and, on context switch,
ensure that each domain can only access its own memory regions.
We use \cet forward-edge control-flow integrity to ensure that application code
cannot jump, sequentially or speculatively, into arbitrary sandbox code (e.g.,
due to a poisoned BTB).
We do this by inserting \mono{endbranch} instructions in \wasm sandboxes to
demarcate valid jump targets, essentially partitioning the BTB into two
domains.
Our use of Intel's MPK and CET ensures that even if the host code
runs with poisoned predictors, it cannot read---and thus leak---sandbox data.
Since \mpk only supports 16 protection regions, we cannot use it to similarly prevent
\attackTwo attacks: serverless, edge-cloud, and FaaS platforms have thousands
of co-located tenants.
Hence, to address \attackTwo attacks, like for \sysDesignOne, we consider and evaluate
two designs.
The first (again) uses ASLR to randomize the location of each sandbox and
flushes the BTB on sandbox entry; we don't flush on sandbox exit since the
host can safely run with a poisoned BTB.
The second is deterministic and not only allows using conditional branches but
also avoids BTB flushes.
It does this using a new technique, \emph{register interlocking}, which tracks
the control flow of the \wasm sandbox and turns every misspeculated memory
access into an access to an empty guard page.
Register interlocking allows a tenant to run with a poisoned BTB or CBP
since any potentially poisoned speculative memory accesses will be invalidated.
\paragraph{3. Implementation and evaluation (\S\ref{sec:impl}--\ref{sec:eval})}
We implement both \sysDesignOne and \sysDesignTwo by modifying the Lucet
compiler's \wasm-to-x86 code generator (Cranelift) and runtime.
To evaluate \sys's security we implement proof of concept breakout and poisoning
attacks against stock Lucet (mitigated by \sys).
We do this for all three Spectre variants, i.e., Spectre attacks that abuse the
CBP, BTB, and RSB.
We evaluate \sys's performance against stock Lucet and several fence-insertion
techniques on several standard benchmarks.
On the \wasm compatible subset of the SPEC~2006 CPU benchmarking suite we find
the ASLR variants of \sysDesignOne and \sysDesignTwo impose little
overhead---they are at most \specSfiASLRMaxOverhead and \specCetASLRMaxOverhead
slower than Lucet, respectively.
Our deterministic implementations, which eliminate all three categories of
attacks, incur modest overheads: \sysDesignOne and \sysDesignTwo are respectively
\specSfiFullMinOverhead--\specSfiFullMaxOverhead (geomean:
\specSfiFullOverhead) and \specCetFullMinOverhead--\specCetFullMaxOverhead
(geomean: \specCetFullOverhead) slower than Lucet.
These overheads are smaller than the overhead imposed by state-of-the-art
fence-based techniques.
\paragraph{Open source and data}
We make all source and data available under an open source license at:
\url{https://swivel.programming.systems}.
\section*{Acknowledgment}
We thank Johnnie Birch, Jonathan Foote, Dan Gohman, Pat Hickey, Tyler McMullen,
Jan de Mooij, Vedvyas Shanbhogue, Jared Stark, Luke Wagner, and Andy Wortman for insightful discussions.
We thank Devdatta Akhawe and the anonymous reviewers for their valuable
comments for improving the quality of this paper.
We would also like to thank Hongjiu Lu and Yu-cheng Yu for their support on the \cet infrastructure.
This work was supported in part by gifts from Cisco, Fastly, Mozilla, and by
the CONIX Research Center, one of six centers in JUMP, a Semiconductor Research
Corporation (SRC) program sponsored by DARPA, by the NSF under grant
numbers CCF-1918573, CNS-1814406, CAREER CNS-2048262, and by NSF/Intel under
grant number CCF-1823444.
\fi
\ifUsenix
{
\fontsize{7}{8}\selectfont
\setlength{\bibsep}{2pt}
\bibliographystyle{abbrv}
\section{Related work}
\label{sec:related}
We give an overview of related work on mitigating Spectre
attacks by discussing microarchitectural proposals, software-based approaches
for eliminating Spectre gadgets, and previous approaches based on CFI or \mpk.
\paragraph{Thwarting covert channels}
Several works~\cite{kiriansky2018dawg,khasawneh2019safespec,barber2019specshield,yan2018invisispec,saileshwar2019cleanupspec}
propose making microarchitectural changes to block, isolate, or remove the
covert channels used to transfer transient secrets to architectural states.
For example, \emph{SafeSpec}~\cite{khasawneh2019safespec} proposes a
speculation-aware memory subsystem which ensures that microarchitectural
changes to the cache are not committed until predictions are validated.
Similarly, \emph{CleanupSpec}~\cite{saileshwar2019cleanupspec} proposes
an undo logic for the cache state.
Although these approaches remove the attacker's data leakage channel, they do
not address the root cause of Spectre vulnerabilities.
In contrast, \sys works with no hardware changes.
\paragraph{Safe speculation}
Intel has introduced hardware support to mitigate Spectre-BTB across separate
address spaces~\cite{IntelSpeculation,intelMitigationList}.
Specifically, the Indirect Branch Predictor Barrier (IBPB) allows the BTB to
be cleared across context switches, while Single Thread Indirect Branch
Predictors (STIBP) ensure that one thread's BTB entries will not be affected
by the sibling hyperthread.
These mitigations can be used by the OS as a coarse-grained mechanism for
safe speculation, but only apply to Spectre-BTB and have not been widely
adopted due to performance overhead~\cite{koruyeh2019speccfi}.
Other works propose microarchitectural changes to allow the software to
control speculation for security-critical
operations~\cite{taram2019context,weisse2019nda} or certain memory
pages~\cite{schwarz2020context,li2019conditional}.
Separately, STT~\cite{yu2019speculative} proposes speculative taint tracking
within the microarchitecture.
However, unlike \sys, these approaches require significant hardware changes
and do not offer a way to safely run code on existing CPUs.
\paragraph{Eliminating Spectre gadgets}
Another way to mitigate Spectre attacks is by inserting a
barrier instruction (e.g.,\xspace \lfence), which blocks speculative
execution~\cite{mscvcspectre,IntelSpecMit,AMDSoftwareTechnique}.
However, as we evaluated in \Cref{subsec:wasm_overhead}, insertion of
\lfence has a performance impact on the entire CPU pipeline
and undercuts the performance benefit of out-of-order and speculative execution.
In contrast, \sys makes little to no use of \lfence.
An optimized approach is to replace control flow instructions with
alternate code sequences that are safe to execute speculatively.
For instance, speculative load hardening (SLH) replaces conditional bounds
checks with an arithmetized form to avoid Spectre-PHT~\cite{RFC}.
Indeed, \sys uses SLH to protect the bounds checks for indirect call tables and
switch tables (\S\ref{subsec:linear_blocks}).
Alternatively, Oleksenko et al.~\cite{oleksenko2018you} propose inserting
artificial data dependencies between secret operations and pipeline serialization
instructions.
Finally, the \emph{retpoline} technique~\cite{retpoline} replaces indirect
branches with a specific code sequence using the \mono{ret}
instruction to avoid Spectre-BTB.
To reduce the overhead of such code transformations, researchers have
proposed several techniques to automatically locate Spectre
gadgets~\cite{wang2018oo7,guarnieri2018spectector,cauligi2020foundations} and
apply mitigations to risky blocks of code.
However, these techniques have to handle potential false positives or negatives; in
contrast, \sys focuses on defending against all possible Spectre attacks
from untrusted code by applying compile-time mitigations.
\paragraph{Speculative CFI}
\emph{SpecCFI}~\cite{koruyeh2019speccfi} has proposed hardware support
for speculative and fine-grained control-flow integrity (CFI), which
can be used to protect against attacks on indirect branches.
In comparison, \sysDesignTwo uses \cet, which only supports coarse-grained
CFI with speculative guarantees~\cite{shanbhogue2019security}.
\emph{Venkman}~\cite{venkman} uses a technique similar to \sys's
linear blocks to ensure that indirect branches always reach a
barrier instruction (e.g.,\xspace \lfence) by applying alignment to bundles similar to
classical software fault isolation~\cite{wahbe-sfi}.
In contrast, \sys is a fence-free approach that preserves the performance
benefits of speculative execution.
\paragraph{Intra-process isolation using \mpk}
Jenkins et~al.\xspace~\cite{jenkins2020ghostbusting} propose to provide intra-process
Spectre protection using \mpk.
They use \mpk to create separate isolation domains and use the relationship
between the code and secret data to limit speculative accesses.
However, since \mpk only provides 16 domains, relying fully on \mpk to isolate
many sandbox instances is infeasible for the CDN \wasm use case we consider.
|
\section{Introduction}
Approximating functions $f \colon \Omega \to \mathcal M$ with values in some
$D$-dimensional Riemannian manifold $\mathcal M$ has attracted lots of interest during the
last years. The central challenge is that with $\mathcal M$ not being linear,
the function spaces over $\Omega$ with values in $\mathcal M$ are not linear as
well and hence, all the well established linear approximation methods do not
have a straight forward generalization to manifold-valued functions.
One successful approach to manifold-valued approximation is to consider the
problem locally. Either one maps the function values locally to some linear
approximation space or one uses local averaging based on the geodesic
distance. These approaches allow to generalize subdivision schemes
\cite{Wallner05,XieYu,Gr09,Gro10,ShIt13}, moving least squares \cite{GrSpYu17},
quasi-interpolation \cite{Grohs2013ProjectionBasedQI} or splines
\cite{Shingel08} to the manifold-valued setting.
A different approach is to embed the manifold $\mathcal M$ into some higher
dimensional linear space $\ensuremath{\mathbb{R}}^{d}$ by a map
$\mathcal E \colon \mathcal M \to \ensuremath{\mathbb{R}}^{d}$. Note, that according to Nash's
embedding theorem \cite{Nash54} such a mapping always exists and can be guaranteed
to be locally isometric provided the dimension $d$ is at least $D(3D+11)/2$,
where $D$ denotes the dimension of $\mathcal M$. Embedding based approximation
methods can be summarized as follows
\begin{enumerate}
\item Transfer $f \in C(\Omega,\mathcal M$) via the embedding $ \mathcal E$
into the linear function space $C(\Omega,\ensuremath{\mathbb{R}}^{d})$. Since often a manifold is described by vectors in $\ensuremath{\mathbb{R}}^d$, we will use the notation
$f$ for the function in both function spaces.
\item Use a linear approximation operator
$I_{\ensuremath{\mathbb{R}}^{d}} \colon C(\Omega,\ensuremath{\mathbb{R}}^{d}) \to C(\Omega,\ensuremath{\mathbb{R}}^{d})$ to find an approximant
$I_{\ensuremath{\mathbb{R}}^{d}} f$ in the embedding.
\item Project the resulting $\ensuremath{\mathbb{R}}^{d}$-valued function back to the manifold
$I_{\mathcal M} f = P_{\mathcal M} \circ (I_{\ensuremath{\mathbb{R}}^{d}} f)$ using some
projection operator $P_{\mathcal M} \colon \ensuremath{\mathbb{R}}^{d} \to \mathcal M$.
\end{enumerate}
Because of its generality and simplicity this approach has already been widely
investigated \cite{Grohs2013ProjectionBasedQI,Ga18} and applied \cite{Mo02,SaSe08}. In
particular it has been shown in \cite{Ga18} that the approximation order of
the embedding based approximation operator $I_{\mathcal M} f$ is the same as
the approximation order of its linear counterpart $I_{\ensuremath{\mathbb{R}}^{d}}$. It is
important to note, that the projection operator $P_{\mathcal M}$ is in general
only defined in some neighborhood $U \supset \mathcal M$ of the
manifold. Hence, the pre-asymptotic behavior of the approximation operator
$I_{\mathcal M}$ strongly depends on the size of this neighborhood which is
directly related to the so-called reach of the embedded manifold.
The aim of this paper is to analyze the pre-asymptotic behavior of the
approximation operator $I_{\mathcal M}$ with respect to the reach of the
embedded manifold. While for the error $I_{\mathcal M} f - f$ the reach only
controls the required linear approximation error
$I_{\ensuremath{\mathbb{R}}^{d}} f - f$ that allows for a meaningful approximation
$I_{\mathcal M} f$, the situation is completely different for the error of the
derivatives $\d (I_{\mathcal M} f) - \d f$. In this case the chain rule has to be used for the derivative $\d (I_{\mathcal M} f)$
of a concatenation of approximation in $\ensuremath{\mathbb{R}}^d$ and projection on the manifold $\ensuremath{\mathcal{M}}$. The derivative of the projection on $\ensuremath{\mathcal{M}}$ leads
to pre-asymptotic constants, which depend on the reach of the manifold.
Our paper is organized as follows. In section~\ref{sec:submani} we will first
show some general differential geometric properties of submanifolds of
$\ensuremath{\mathbb{R}}^d$. Most importantly, we identify in Lemma \ref{thm:PM=piE} the projection
operator $P_{\mathcal M}$ with an orthogonal projection in the normal bundle
over the manifold $\mathcal M$. This is only possible within some tubular
neighborhood of the manifold $\ensuremath{\mathcal{M}}$ that is controlled by the reach of
$\mathcal M$. The relationship between the reach of the manifold $\mathcal M$
and its curvature or second fundamental form is addressed in
section~\ref{sec:embedd-based-prop}. In Theorem~\ref{thm:B} we make use of
these relationships to describe the differential of the projection operator
$P_{\mathcal M}$ in terms of the second fundamental form. In
section~\ref{sec:change-proj-oper} we end up with the main results of this
chapter, that is we show in Theorem~\ref{thm:Psi_bound} that the derivative
$dP_{\mathcal M}(\vec x)$ of the projection operator at some point
$\vec x \in \ensuremath{\mathbb{R}}^{d}$ satisfies a Lipschitz-condition with respect to $\vec x$. As
our Lipschitz bound is with respect to the Euclidean distance in the embedding
it is more sharp then the bound reported in \cite{BoLiWi19} that relies on the
geodesic distance.
Section~\ref{sec:manifold} is dedicated to manifold-valued approximation. We
show that the approach of using a linear approximation operator on an
embedded manifold $\ensuremath{\mathcal{M}}$ in $\ensuremath{\mathbb{R}}^d$ and then projecting back on the manifold inherits the
approximation order of the linear approximation. Our main result is stated in
Theorem~\ref{satz:wichtigableitung} and gives a pre-asymptotic bound for the
approximation error of the first derivative that relies exclusively on the
reach of the embedded manifold. This result is illustrated in
Theorem~\ref{thm:FPS} for a specific approximation operator, the Fourier
partial sum operator.
In the final section~\ref{sec:examples} we consider two real world examples
for approximating manifold valued data. The first example deals with functions
from the two-sphere into the two-dimensional projective space that describe
the dependency between the propagation direction and the polarization
directions of seismic waves. The second example is from crystallographic
texture analysis where the local alignment of the atom lattice is described by
a map with values in the quotient $\ensuremath{\mathrm{SO}(3)}/\mathcal{S}$ of the rotation group $\ensuremath{\mathrm{SO}(3)}$
modulo some finite symmetry group $\mathcal{S}$. The derivative of this map has
important connections microscopic and macroscopic material properties.
\section{Submanifolds}
\label{sec:submani}
In this section we will consider smooth compact Riemannian submanifolds $\ensuremath{\mathcal{M}}$
of $\ensuremath{\mathbb{R}}^d$. We will show some differential geometric properties of submanifolds
as well as some estimations for the projection $P_\ensuremath{\mathcal{M}}$ and the differential of
this projection. We will use these results for estimating some approximation
errors in section~\ref{sec:manifold}.
\subsection{The Projection Operator}
\label{sec:projection-operator-1}
Throughout our work we denote by $\mathcal M \subset \ensuremath{\mathbb{R}}^{d}$ a smooth compact
Riemannian submanifold of $\ensuremath{\mathbb{R}}^d$. For every point $\vec m\in \ensuremath{\mathcal{M}}$ we denote the tangent space $T_\vec m\ensuremath{\mathcal{M}}$ as well as the normal space $N_\vec m\ensuremath{\mathcal{M}}$. Furthermore, we denote by $P_{\mathcal M} \colon \ensuremath{\mathbb{R}}^{d} \to \mathcal M$
the projection operator onto $\mathcal M$ defined as the solution of the
minimization problem
\begin{equation}\label{eq:2}
P_{\mathcal M}(\vec x) = \argmin_{\vec m \in \mathcal M} \norm{\vec x-\vec m}_2.
\end{equation}
In general, this minimization problem does not posses a unique solution for
every $\vec x \in \ensuremath{\mathbb{R}}^{d}$, since there is an ambiguity to which branch of the manifold the point should be attributed.
However, if we restrict the domain of the definition of
$P_{\mathcal M}$ to some open neighborhood $U \subset \ensuremath{\mathbb{R}}^{d}$ of
$\mathcal M$ uniqueness can be granted.
In order to find such a neighborhood $U$ we define on the normal bundle
\begin{equation*}
N\ensuremath{\mathcal{M}}
=\{(\vec m,\vec v)\in \ensuremath{\mathbb{R}}^d\times \ensuremath{\mathbb{R}}^d:\vec m\in \ensuremath{\mathcal{M}},\vec v\in N_\vec m\ensuremath{\mathcal{M}}\}
\end{equation*}
of $\mathcal M$ the smooth map
\begin{equation*}
E \colon N\ensuremath{\mathcal{M}} \to \ensuremath{\mathbb{R}}^d, \quad
E(\vec m,\vec v)=\vec m + \vec v,
\end{equation*}
that maps every normal space $N_{\vec m} \mathcal M$ to an affine linear
subspace through $\vec m \in \ensuremath{\mathbb{R}}^{d}$. Since we assumed $\mathcal M$ to be
compact and smooth, there exist a maximum constant $\tau>0$ such that the
mapping $E$ restricted to the open subset
\begin{equation*}
V = \{ (\vec m,\vec v) \in N\ensuremath{\mathcal{M}}: \norm{\vec v}_2<\tau\}
\end{equation*}
of the normal bundle is injective, cf. \cite[6.24]{Lee12}. Setting $U = E(V)$
defines the so-called \emph{tubular neighborhood} of $\mathcal M$ and the
restriction $E \colon V \to U$ becomes a diffeomorphism. The constant $\tau$
is commonly called \emph{reach} and its inverse $1/\tau$ is the
\emph{condition number} of the manifold. The reach $\tau$ is affected by two
factors: the curvature of the manifold and the width of the narrowest
bottleneck-like structure of $\ensuremath{\mathcal{M}}$, which quantifies how far $\ensuremath{\mathcal{M}}$ is from
being self-intersecting. An estimate on the relationship between the reach and
the curvature of the manifold $\mathcal M$ will be given in
Lemma~\ref{lemma:curvatureReach}.
Using the mapping $E$ we may now give an explicit definition of the projection
operator $P_{\mathcal M}$.
\begin{Lemma}
\label{thm:PM=piE}
Let $\vec u \in U$ and let $\pi \colon N \mathcal M \to \mathcal M$,
$(\vec m,\vec n) \mapsto \vec m$ be the canonical projection operator. Then
\begin{equation*}
P_\ensuremath{\mathcal{M}}(\vec u) = \pi \circ E^{-1}(\vec u)
\end{equation*}
is the unique solution of the minimization problem \eqref{eq:2}.
\end{Lemma}
\begin{proof}
Let $\vec u\in U$ and $P_\ensuremath{\mathcal{M}}(\vec u)=\vec m\in \ensuremath{\mathcal{M}}$. We show that
$\vec u-\vec m\in N_\vec m\ensuremath{\mathcal{M}}$. We assume the opposite and decompose
$\vec u-\vec m$ in one part in $N_\vec m\ M$ and a part $\vec t$ in
$T_\vec mM$. Then there is a curve $\gamma(s)$ in $\ensuremath{\mathcal{M}}$ with
$\gamma(0)=\vec m$ and $\dot{\gamma}(0)=\vec t$. If we go along this curve,
we obtain for sufficient small $\epsilon >0$ that
$\vec u-\gamma(\epsilon)<\vec u-\gamma(0)=\vec u-\vec m$. That is a
contradiction to the definition of $P_\ensuremath{\mathcal{M}}$. Since the projection $P_\ensuremath{\mathcal{M}}$
should be unique, we have to show that $\pi\circ E^{-1}$ is also unique.
For that reason we assume that for $\vec u\in U$ there holds
$\pi\circ E^{-1}=\vec m\in \ensuremath{\mathcal{M}}$ and $\pi\circ E^{-1}=\vec m'\in \ensuremath{\mathcal{M}}$. This
would imply $\vec u=\vec m+\vec v=\vec m'+\vec v'$ with
$\vec v\in N_\vec m\ensuremath{\mathcal{M}}$ and $\vec v\in N_{\vec m'}\ensuremath{\mathcal{M}}$. That is a
contradiction to the uniqueness of $E^{-1}$ in the tubular neighborhood $U$.
\end{proof}
Let us illustrate this by a simple example.
\begin{example}\label{ex:Sd}
Let the manifold $\ensuremath{\mathcal{M}}$ be the $(d-1)$-dimensional unit sphere,
embedded in $\ensuremath{\mathbb{R}}^d$. These manifolds can be described by
\begin{equation*}
\S^{d-1}=\left\{\vec x\in \ensuremath{\mathbb{R}}^{d}:\|\vec x\|_2=1\right\}.
\end{equation*}
The projection $P_{\S^{d-1}}$ easily reads
\begin{equation*}
P_{\S^{d-1}}: \ensuremath{\mathbb{R}}^{d}\backslash\{\vec 0\} \rightarrow \S^{d-1}, \quad \quad P_{\S^{d-1}}(\vec x)=\frac{\vec x}{\|\vec x\|_2}.
\end{equation*}
This map is well-defined and smooth.
\end{example}
\subsection{Curvature and Reach of Submanifolds}
\label{sec:embedd-based-prop}
For any point $\vec m \in \mathcal M \subset \ensuremath{\mathbb{R}}^{d}$ on the manifold we can
decompose $\ensuremath{\mathbb{R}}^{d}$ as the direct sum $\ensuremath{\mathbb{R}}^d=T_\vec m\ensuremath{\mathcal{M}}\oplus N_\vec m\ensuremath{\mathcal{M}}$ of the
tangential space $T_\vec m\ensuremath{\mathcal{M}}$ and the normal space $N_\vec m\ensuremath{\mathcal{M}}$. Let us denote
by $P_T\colon\ensuremath{\mathbb{R}}^d \to T_\vec m\ensuremath{\mathcal{M}}$ and $P_N\colon \ensuremath{\mathbb{R}}^d \to N_\vec m\ensuremath{\mathcal{M}}$ the
corresponding orthogonal projections. Then the canonical connection $\nabla$
on $\ensuremath{\mathbb{R}}^{d}$ defines a connection $\nabla^{\mathcal M}$ on $\mathcal M$ by
\begin{equation}
\label{eq:7}
\nabla^{\mathcal M}_{\vec X}
= P_{T} \nabla_{\vec X}(P_{T} \vec Y) + P_{N} \nabla_{\vec X}(P_{N} \vec Y),
\end{equation}
where $\vec X \colon \mathcal M \to T \mathcal M$ is a tangential and
$\vec Y \colon \mathcal M \to \ensuremath{\mathbb{R}}^{d}$ a general vector field on
$\mathcal M$.
If $\vec Y$ is a tangential vector field as well, the first summand
$P_{T} \nabla_{\vec X}(P_{T} \vec Y) = P_{T} \nabla_{\vec X} \vec Y$ in
\eqref{eq:7} is just the Levi-Cevita-connection on $\mathcal M$, whereas its
orthogonal complement
\begin{equation*}
\rz{II}(\vec X,\vec Y) = P_{N}(\nabla_{\vec X} \vec Y)
\end{equation*}
is the \emph{second fundamental form} on $\mathcal M$.
We call a vector field $\vec Y \colon \mathcal M \to \ensuremath{\mathbb{R}}^{d}$ parallel along a
curve $\gamma$ if $\nabla^{\mathcal M}_{\dot \gamma} \vec Y = 0$. For a
geodesic $\gamma$ with $\gamma(0) = \vec m$,
$\dot \gamma(0) = \vec t \in T_{\vec m} \mathcal M$ and an arbitrary vector
$\vec y \in T_{\vec m} \mathcal M \oplus N_{\vec m} \mathcal M = \ensuremath{\mathbb{R}}^{d}$ we
shall use the abbreviation
\begin{equation*}
\nabla_{\vec t} \vec y = \nabla_{\vec t} \vec Y(0)
\end{equation*}
where $\vec Y$ is the parallel transport of the vector $\vec y$ along the
curve $\gamma$.
For a fixed point $\vec m \in \mathcal M$ and a normal direction
$\vec n \in N_{\vec m} \mathcal M$ we define the operator
$\vec B_{\vec n} \colon T_{\vec m}\mathcal M \to T_{\vec m}\mathcal M$ on the
tangent space by
\begin{equation}
\label{eq:Bn}
\dotprod{\vec B_{\vec n} \vec x}{\vec y} = \dotprod{\vec n}{\nabla_{\vec x} \vec y},
\quad \vec x,\vec y \in T_{\vec m} \mathcal M.
\end{equation}
We may also express $B_{\vec n} \vec x$ as the tangential part of the
covariant derivative of $\vec n$ in direction $\vec x$.
\begin{lemma}
\label{lem:BnX=nablaXn}
Let $\vec n \in N_{\vec m}\mathcal M$ be a normal and
$\vec x \in T_{\vec m} \mathcal M$ a tangential vector. Then
\begin{equation*}
\vec B_\vec n \vec x = - P_T\nabla_{\vec x}\vec n.
\end{equation*}
\end{lemma}
\begin{proof}
Let $\gamma$ be a geodesics in $\ensuremath{\mathcal{M}}$ with $\gamma(0)=\vec m$ and
$\dot{\gamma}(0) = \vec x$ and let $\vec N$ be the parallel transport of
$\vec n$ along $\gamma$. Let furthermore, $\vec Y$ be an arbitrary tangent
vector field parallel along $\gamma$. Then we have
\begin{equation*}
0
=\frac{\d}{\d s}\langle\vec N(s),\vec Y(s)\rangle|_{s=0}
=\langle\nabla_{\vec x}\vec N(0),\vec Y(0)\rangle
+\langle\vec N(0),\vec \nabla_{\vec x}\vec Y(0)\rangle.
\end{equation*}
This yields
\begin{equation*}
\langle\vec B_\vec n\vec x,\vec y\rangle
=\langle\vec n,\nabla_{\vec x}\vec y\rangle
=-\langle\nabla_{\vec x}\vec n,\vec y\rangle.
\end{equation*}
Since the vector field $\vec Y$ was arbitrary, this yields the assertion.
\end{proof}
The operator $\vec B_{\vec n}$ describes the extrinsic curvature of the
manifold in the point $\vec m $ and the normal direction $\vec n$. Its norm is
bounded by the condition number $\frac{1}{\tau}$ of $\mathcal M$. More
precisely the following result is shown in \cite[Proposition 6.1]{NSW08}.
\begin{lemma}
\label{lemma:curvatureReach}
Let $\tau$ be the reach of $\ensuremath{\mathcal{M}}$, $\vec m\in \ensuremath{\mathcal{M}}$ be an arbitrary point on the manifold and $\vec n \in N_{\vec m}\mathcal M$ be
a normal vector. Then the operator $B_{\vec n}$ defined in \eqref{eq:Bn} is
symmetric and bounded by $\frac{1}{\tau}$, i.e., we have for tangential
vectors $\vec x,\vec y \in T_{\vec m} \mathcal M$ the inequality
\begin{equation}
\label{eq:2formbound}
\dotprod{\vec B_{\vec n} \vec x}{\vec y}
\leq \frac{1}{\tau} \norm{\vec n} \norm{\vec x} \norm{\vec y}.
\end{equation}
\end{lemma}
The next lemma bounds the covariant derivative of parallel vector fields by
the condition number $\frac{1}{\tau}$ of the manifold.
\begin{lemma}\label{lem:bound_diff}
Let $\vec Y$ be a parallel vector field along a geodesic $\gamma$ in
$\ensuremath{\mathcal{M}}$. Then its covariant derivative in $\ensuremath{\mathbb{R}}^{d}$ is bounded by
\begin{equation*}
\norm{\nabla_{\dot{\gamma}}\vec Y}_2\leq \tfrac 1\tau \norm{\vec Y}_{2}\norm{\dot{\gamma}}_2.
\end{equation*}
\end{lemma}
\begin{proof}
Let $\vec Y = \vec T + \vec N$ be the decomposition of $\vec Y$ into a
tangent vector field $\vec T$ and a normal vector field $\vec N$. Since
$\vec Y$ is parallel along $\gamma$ we have
\begin{align*}
0
= \nabla_{\dot \gamma}^{\mathcal M} \vec Y
= P_{T} \nabla_{\dot \gamma} \vec T + P_{N} \nabla_{\dot \gamma} \vec N
\end{align*}
and, hence,
\begin{equation*}
\nabla_{\dot \gamma} \vec Y
= P_{N} \nabla_{\dot \gamma} \vec T
+ P_{T} \nabla_{\dot \gamma} \vec N.
\end{equation*}
Let $\vec n = P_{N} \nabla_{\dot \gamma(s)} \vec T(s)$. Then we obtain by Lemma
\ref{lemma:curvatureReach}
\begin{equation*}
\norm{\vec n}^{2}
= \dotprod{\vec n}{\nabla_{\dot \gamma(s)} \vec T(s)}
= \dotprod{\vec B_{\vec n} \dot \gamma(s)}{\vec T(s)}
\le \tfrac{1}{\tau} \norm{\dot \gamma} \norm{\vec n} \norm{\vec T(s)}.
\end{equation*}
For the tangential part we have by Lemma \ref{lem:BnX=nablaXn}
\begin{equation*}
\norm{P_{T} \nabla_{\dot{\gamma}(s)}\vec N(s)}_2 = \norm{\vec B_{\vec N(s)}
\dot{\gamma}(s)}_2\leq \tfrac 1\tau \norm{\dot \gamma}\norm{\vec N(s)}.
\end{equation*}
The assertion follows now from Parsevals inequality.
\end{proof}
\subsection{The Differential of the Projection Operator.}
\label{sec:diff-proj-oper}
The differential
$\d P_{\mathcal M}(\vec m) \colon \ensuremath{\mathbb{R}}^{d} \to T_{\vec m} \mathcal M$ of the
projection $P_{\mathcal M} \colon \ensuremath{\mathbb{R}}^{d} \to \mathcal M$ is especially easy to
compute at points $\vec m \in \mathcal M$ on the manifold. In this case it is
simply the linear projection
$P_{T_{\vec m} \mathcal M} \colon \ensuremath{\mathbb{R}}^{d} \to T_{\vec m} \mathcal M$ onto the
tangential space attached to $\vec m$, i.e.
\begin{equation}
\label{eq:dPMm}
\d P_{\mathcal M}(\vec m) = P_{T_{\vec m} \mathcal M}.
\end{equation}
We can verify this by observing that for normal vectors $\vec n\in N_\vec m\ensuremath{\mathcal{M}}$
we have
\begin{equation*}
\d P_\ensuremath{\mathcal{M}}(\vec m)\,\vec n
=\lim_{h\to 0}\frac{P_\ensuremath{\mathcal{M}}(\vec m+h\vec n)-P_\ensuremath{\mathcal{M}}(\vec m)}{h}
=\lim_{h\to 0}\frac{\vec m-\vec m}{h}=\vec 0,
\end{equation*}
while for tangent vectors $\vec t\in T_\vec m\ensuremath{\mathcal{M}}$ we obtain
\begin{align*}
\d P_\ensuremath{\mathcal{M}}(\vec m)\,\vec t
&=\lim_{h\to 0}\frac{P_\ensuremath{\mathcal{M}}(\vec m+h\vec t)-P_\ensuremath{\mathcal{M}}(\vec m)}{h}
=\lim_{h\to 0}\frac{\vec m+\exp(h\vec t)-\vec m}{h}\\
&=\lim_{h\to 0}\frac{\exp(h\vec t)}{h}=\vec t,
\end{align*}
where $\exp$ denotes the exponential map to the manifold.
The differential $\d P_{\mathcal M}(\vec m + \vec v)$,
$\vec v \in N_{\vec m} \mathcal M$ at a point not in the manifold is a little
bit more tricky. We start by observing that the tangential
$T_{(\vec m,\vec v)} N\mathcal M \subset \ensuremath{\mathbb{R}}^{2d}$ of the normal bundle at a point
$(\vec m,\vec v) \in N \mathcal M$ is
\begin{align*}
T_{(\vec m,\vec v)} N\mathcal M
&= \{ (\vec 0,\vec n) \mid \vec n\in N_{\vec m} \mathcal M \}
\oplus \{(\vec t,\nabla_{\vec t} \vec v) \mid \vec t \in T_{\vec m} \mathcal M\} \\
&= \{ (\vec t, \vec u) \mid
\vec t \in T_{\vec m} \mathcal M,
P_{T} \vec u = \nabla_{\vec t} \vec v\}.
\end{align*}
The following lemma describes the differential $\d P_\ensuremath{\mathcal{M}}(\vec m+\vec v)$.
\begin{lemma}
\label{lemma:dPM}
Let $\vec m \in \ensuremath{\mathcal{M}}$ be an arbitrary point on the manifold $\ensuremath{\mathcal{M}}$ and $\vec v \in N_{\vec m}\mathcal M$ be a normal
vector with $\norm{\vec v}_2<\tau$, i.e. $\vec m +\vec v$ is in the tubular
neighborhood of $\ensuremath{\mathcal{M}}$. Then the derivative $\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)$ satisfies for every
tangent direction $\vec t\in T_{\vec m}\ensuremath{\mathcal{M}}$,
\begin{equation*
\left(\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)\right)
\left(\vec t+\nabla_{\vec t}\vec v\right)
=\vec t.
\end{equation*}
while it vanishes for any normal direction $\vec n\in N_\vec m\ensuremath{\mathcal{M}}$, i.e.
\begin{equation*}
\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)\,\vec n=\vec 0.
\end{equation*}
\end{lemma}
\begin{proof}
According to Lemma~\ref{thm:PM=piE} we have $P_\ensuremath{\mathcal{M}} = \pi \circ E^{-1}$, where
$\pi \colon N\mathcal M \to \mathcal M$, $(\vec m,\vec v) \mapsto \vec m$ is
the projection operator. Its differential at the point $(\vec m,\vec v) \in
N\mathcal M$ is the projection
\begin{equation*}
\d \pi(\vec m,\vec v)
\colon T_{(\vec m,\vec v)} N \mathcal M \to T_{\vec m} \mathcal M, \quad
(\vec t, \vec u) \mapsto \vec t.
\end{equation*}
The differential of the mapping $E \colon N \mathcal M \to \ensuremath{\mathbb{R}}^{d}$ in a
point $(\vec m,\vec v) \in N\mathcal M$ is given by
\begin{equation*}
\d E(\vec m,\vec v) \colon T_{(\vec m,\vec v)} N \mathcal M \to \ensuremath{\mathbb{R}}^{d},
\quad
(\vec t, \vec u) \mapsto \vec t + \vec u.
\end{equation*}
Since $\vec m+\vec v$ is within the tubular neighborhood of
$\mathcal M$, $E$ is invertible in some neighborhood of
$\vec m+\vec v$. Then $\d E(\vec m,\vec v)$ is invertible as well and we
have for any normal vector $\vec n \in N_{\vec m} \mathcal M$
\begin{equation*}
\d E^{-1}(\vec m + \vec v)\, \vec n = (\vec 0, \vec n)
\end{equation*}
and for any tangent vector $\vec t \in T_{\vec m} \mathcal M$
\begin{equation*}
\d E^{-1}(\vec m + \vec v)\, (\vec t + \nabla_{\vec t} \vec v)
= (\vec t, \nabla_{\vec t} \vec v).
\end{equation*}
Together with the chain rule this implies the assertion.
\end{proof}
The image of $\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)$ is contained
in the tangential space $T_\vec m\ensuremath{\mathcal{M}}$, especially $\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)$ is
the projection $P_{T_\vec m\ensuremath{\mathcal{M}}}$ up to a factor matrix. We will write this
linear operator $\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)$ in another way, to see the
difference to the linear operator $\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m)$.
\begin{theorem}
\label{thm:B}
Let $\vec m \in \ensuremath{\mathcal{M}}$ be a point on the manifold, let
$\vec v \in N_{\vec m} \mathcal M$ be a normal vector with
$\norm{\vec v} < \tau$ and let
$\vec B_{\vec v} \colon T_{\vec m}\mathcal M \to T_{\vec m}\mathcal M$ be
the symmetric operator defined in \eqref{eq:Bn}, extended to
$\vec B_{\vec v} \colon \ensuremath{\mathbb{R}}^d \to \ensuremath{\mathbb{R}}^d$ by $\vec B_\vec v\vec n=\vec 0$ for
all normal vectors $\vec n\in N_\vec m\ensuremath{\mathcal{M}}$. Then the derivative of the
projection operator $P_\ensuremath{\mathcal{M}}$ satisfies
\begin{equation*}
\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)
=P_{T_{\vec m} \ensuremath{\mathcal{M}}}(\vec I - \vec B_{\vec v})^{-1}
=\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m) - \vec B_\vec v\left(\vec I+\vec B_{\vec v}\right)^{-1},
\end{equation*}
where $\vec I \colon \ensuremath{\mathbb{R}}^d \to \ensuremath{\mathbb{R}}^d$ is the identity.
\end{theorem}
\begin{proof}
Using Lemma \ref{lemma:dPM} we obtain for all tangential vectors
$\vec t\in T_\vec m\ensuremath{\mathcal{M}}$,
\begin{align*}
P_{T_\vec m\ensuremath{\mathcal{M}}}\vec t
&=\vec t=\left(\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)\right)\left(\vec t+\nabla_{\vec t} \vec v\right)\\
&=\left(\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)\right)\left(\vec t-\vec B_\vec v\vec t\right)
=\left(\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)\right)\left(\vec I-\vec B_\vec v\right)\vec t.
\end{align*}
and for all normal vectors $\vec n\in N_\vec m\ensuremath{\mathcal{M}}$,
\begin{equation*}
\vec 0 = P_{T_\vec m\ensuremath{\mathcal{M}}}\vec n=\left(\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)\right)\left(\vec I-\vec B_\vec v\right)\vec n.
\end{equation*}
Consequently, we have
\begin{equation*}
P_{T_\vec m\ensuremath{\mathcal{M}}} = \mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v) \left(\vec I-\vec B_\vec v\right).
\end{equation*}
By our assumption and \eqref{eq:2formbound} we have
$\norm{\vec B_\vec v}\leq\frac{1}{\tau} \norm{\vec v}<1$ and hence, the
operator $\vec I-\vec B_\vec v$ is invertible. This yields the first part of
the assertion. For the second part we use \eqref{eq:dPMm} and compute
\begin{align*}
\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)
&=P_{T_\vec m\ensuremath{\mathcal{M}}}\left(\vec I-\vec B_\vec v\right)^{-1}\\
&=P_{T_\vec m\ensuremath{\mathcal{M}}}\left(\vec I+\vec B_\vec v\left(\vec I-\vec B_\vec v\right)^{-1}\right)\\
&=\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m)+P_{T_\vec m\ensuremath{\mathcal{M}}}\vec B_\vec v\left(\vec I-\vec B_\vec v\right)^{-1}\\
&=\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m)+\vec B_\vec v\left(\vec I-\vec B_\vec v\right)^{-1},
\end{align*}
where the last equality follows from the fact that the image of
$\vec B_\vec n$ is in the tangent space $T_{\vec m}\ensuremath{\mathcal{M}}$, so the projection on
$T_{\vec m}\ensuremath{\mathcal{M}}$ is unnecessary.
\end{proof}
We consider again the manifold from example~\ref{ex:Sd}.
\begin{example}
For $\ensuremath{\mathcal{M}}=\S^{d-1}\subset \ensuremath{\mathbb{R}}^d$ any normal vector $\vec v \in N_{\vec
m}\S^{d-1}$ has the representation $\vec v = v\,\vec m$. Let
$\{\vec t_i\}_{i=1}^{d-1}\subset T_{\vec m}\S^{d-1}$ be an orthonormal basis of
$T_m\S^{d-1}$. Then $\nabla_{\vec t_i} \vec m = \vec t_{j}$ and hence
\begin{equation*}
\vec B_\vec v=-v\sum_{i=1}^{d-1}{\vec t_i \vec t_i^\top}.
\end{equation*}
By Theorem~\ref{thm:B} and the orthonormality of
$\{\vec m\}\cap \{\vec t_i\}_{i=1}^{d-1}$ of we obtain for $v > -1$ and
$\vec x = \vec m + v \vec m$,
\begin{equation*}
\label{eq:11}
\d P_{\S^{d-1}}(\vec x)
= \frac{1}{1+v}\sum_{i=1}^{d-1}{\vec t_i \vec t_i^\top}
= \frac{1}{\norm{\vec x}_2}\,
\left(\vec I_{d\times d}-\frac{\vec x}{\norm{\vec x}_2}\left(\frac{\vec x}{\norm{\vec x}_2}\right)^\top\right).
\end{equation*}
\end{example}
\subsection{Deviation of the Projection Operator}
\label{sec:change-proj-oper}
In this section we are interested in the change of the derivative
$\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m)$ of the projection operator for small deviations of
$\vec m$. We shall show that for two points $\vec m$ and $\vec z$ on $\ensuremath{\mathcal{M}}$ and
$\vec v\in N_\vec m\ensuremath{\mathcal{M}}$ with $\norm{\vec v}_2<\tau$ we can bound the difference
$\norm{\d P_\ensuremath{\mathcal{M}}(\vec m+\vec v)-\d P_\ensuremath{\mathcal{M}}(\vec z)}_2$ by a multiple of the
Euclidean distance $\norm{\vec m+\vec v-\vec z}_2$.
As usual we start with the case that both points are on the manifold.
According to \cite[Lemma 6]{BoLiWi19} the difference of the differentials is
then bounded by
\begin{equation*}
\norm{\d P_\ensuremath{\mathcal{M}}(\vec m)-\d P_\ensuremath{\mathcal{M}}(\vec z)}_2
\le \frac{1}{\tau}\, d(\vec m,\vec z),
\end{equation*}
where $d(\vec m,\vec z)$ denotes the geodesic distance between the points
$\vec m, \vec z \in \mathcal M$. If the Euclidean distance between the two
points is bounded by $\norm{\vec m - \vec z}_2 \le 2 \tau$ we have by
\cite[Lemma 3]{BoLiWi19} and the fact that $\arcsin(x)\leq \frac{\pi}{2}x$
for $0\leq x\leq 1$, the following estimate between geodesic distance and
Euclidean distance in the embedding
\begin{equation}
\label{eq:XX}
d(\vec m, \vec z) \le \frac\pi 2 \norm{\vec m - \vec z}_2,
\end{equation}
which leads to the local estimate
\begin{equation*}
\norm{\d P_\ensuremath{\mathcal{M}}(\vec m)-\d P_\ensuremath{\mathcal{M}}(\vec z)}_2
\le \frac{\pi}{2 \tau} \norm{\vec m - \vec z}_2.
\end{equation*}
In the following Theorem we prove a sharper and global bound for this
difference.
\begin{theorem}
\label{thm:C2Teil1}
For all $\vec m,\vec z\in \ensuremath{\mathcal{M}}$ the difference between the projection
operators $P_{T_\vec m\ensuremath{\mathcal{M}}}$ and $P_{T_\vec z \ensuremath{\mathcal{M}}}$ onto the respective
tangential spaces is bounded by
\begin{equation*}
\left\|P_{T_\vec m\ensuremath{\mathcal{M}}}-P_{T_\vec z\ensuremath{\mathcal{M}}}\right\|_2 \leq \frac 1\tau \norm{\vec m -\vec z}_2.
\end{equation*}
\end{theorem}
\begin{proof}
First of all we note that for $\norm{\vec m-\vec z}_2\ge 2\,\tau$ the
assertion is immediately satisfied since
$\left\|P_{T_\vec m\ensuremath{\mathcal{M}}}-P_{T_\vec z\ensuremath{\mathcal{M}}}\right\|_2 \leq 2$ independently of
$\vec m,\vec z\in \ensuremath{\mathcal{M}}$. We may therefore assume
$\norm{\vec m-\vec z}_2 < 2\,\tau$ for the rest of the proof.
In order to estimate the difference between the two projection operators we
consider a geodesic $\gamma$ with $\gamma(0)=\vec m$, $\gamma(t) = \vec z$
and $\norm{\dot \gamma}_2=1$. Furthermore, we consider an orthonormal basis
$\{\vec t_i\}_{i=1}^D$ in $T_\vec m\ensuremath{\mathcal{M}}$ and an orthonormal basis
$\{\vec n_j\}_{j=1}^{d-D}$ in $N_\vec m\ensuremath{\mathcal{M}}$. The parallel transport of these
basis vectors along $\gamma$ defines a rotation $\vec R \in \mathrm{SO}(d)$
that maps the tangent space $T_{\vec m}\ensuremath{\mathcal{M}}$ onto the tangent space
$T_{\vec z}\ensuremath{\mathcal{M}}$. Using the rotation $\vec R$ we may rewrite the difference
between the projection operators as
\begin{equation*}
P_{T_\vec m\ensuremath{\mathcal{M}}} - P_{T_\vec z\ensuremath{\mathcal{M}}}
= P_{T_\vec m\ensuremath{\mathcal{M}}} - \vec R P_{T_\vec m\ensuremath{\mathcal{M}}} \vec R^{T}.
\end{equation*}
By Lemma~\ref{lemma:commutator} in the appendix we obtain
\begin{equation}
\label{eq:5}
\norm{P_{T_\vec m\ensuremath{\mathcal{M}}} - P_{T_\vec z\ensuremath{\mathcal{M}}}}_{2} = \norm{P_{T_\vec m\ensuremath{\mathcal{M}}} \vec R -
\vec R P_{T_\vec z\ensuremath{\mathcal{M}}}}_{2} \le \norm{\vec I - \vec R}_{2}
\end{equation}
and hence, it suffices to bound for any normalized $\vec x \in \ensuremath{\mathbb{R}}^{d}$
\begin{equation}
\label{eq:6}
\norm{(\vec I - \vec R) \vec x}_{2}^{2}
= 2 - 2\dotprod{\vec x}{\vec R \vec x}.
\end{equation}
By definition $\vec R \vec x$ is the result of the parallel transport of
$\vec x$ along the curve $\gamma$ in $\gamma(t) = \vec z$. Let us denote by
$\vec X(s)$ the parallel transport of $\vec x$ along $\gamma$ for all
times $s \in [0,t]$. Viewing $s \mapsto \vec X(s)$ as a curve on
$\mathbb S^{d-1}$ with velocity bounded according to
Lemma~\ref{lem:bound_diff} by
$\norm{\dot{\vec X}(s)} = \norm{\nabla_{\dot{\gamma}(s)}\vec X(s)}_2 \leq
\frac{1}{\tau}$, we conclude that
\begin{equation}
\label{eq:3}
\angle(\vec X(\eta),\vec X(\xi))
\leq \frac{1}{\tau}|\eta-\xi|, \quad \eta,\xi \in [0,t].
\end{equation}
Since $\gamma$ is a geodesic we can set in \eqref{eq:3},
$\vec X = \dot \gamma$. As $\abs{\eta - \xi} \le t$ and $t$ is the geodesic
distance between $\vec z$ and $\vec m$ we can use \eqref{eq:XX} and our
assumption $\norm{\vec m-\vec z}_2 < 2 \tau$ to bound the right hand side of
\eqref{eq:3} by
\begin{equation*}
\angle(\dot{\gamma}(\xi),\dot{\gamma}(\eta))
\le \frac{1}{\tau} \abs{\eta - \xi}
\le \frac{t}{\tau}
\le \frac{\pi}{2\tau} \norm{\vec z - \vec m}_{2} \le \pi.
\end{equation*}
Making use of the monotonicity of the cosine this implies
\begin{equation}
\label{eq:10}
\cos \angle(\dot{\gamma}(\xi),\dot{\gamma}(\eta))
> \cos \tfrac{\xi-\eta}{\tau}, \quad
\xi, \eta \in [0,t].
\end{equation}
Considering again the general vector field $\vec X$ we use
\eqref{eq:3} and \eqref{eq:10} to bound \eqref{eq:6} by
\begin{align*}
2-2\dotprod{\vec X(0)}{\vec X(t)}
&= 2-2\,\cos(\angle (\vec X(0),\vec X(t)))\\
&\leq 2 - 2\,\cos \tfrac{t}{\tau}
=\frac{1}{\tau^2} \int_{0}^t\int_{0}^t
\cos \tfrac{\xi-\eta}{\tau}\,\mathrm{d} \eta\,\mathrm{d} \xi\\
&\leq \frac{1}{\tau^2} \int_{0}^t\int_{0}^t \cos
\angle(\dot{\gamma}(\xi),\dot{\gamma}(\eta)) \,\mathrm{d} \eta\,\mathrm{d} \xi\\
&= \frac{1}{\tau^2} \int_{0}^t\int_{0}^t\langle\dot{\gamma}(\xi),\dot{\gamma}(\eta)\rangle\,\mathrm{d} \eta\,\mathrm{d} \xi
= \frac{1}{\tau^2} \norm{\vec m-\vec z}_2^2.
\end{align*}
In combination with \eqref{eq:5} and \eqref{eq:6} this proves
\begin{equation*}
\left\|P_{T_\vec m\ensuremath{\mathcal{M}}}-P_{T_\vec z\ensuremath{\mathcal{M}}}\right\|_2 \leq \frac 1\tau \norm{\vec m -\vec z}_2.\qedhere
\end{equation*}
\end{proof}
Using the example of the unit circle it can be easily verified that our new
bound is sharp.
So far we bounded the variation of the projection operator for points on the
manifold. For the general case that only one point is on the manifold we have
the following result.
\begin{theorem}
\label{thm:Psi_bound}
Let $\vec m,\vec z\in \ensuremath{\mathcal{M}}$ and $\vec v\in N_\vec m\ensuremath{\mathcal{M}}$ with
$\norm{\vec v}_2<\tau$. Then
\begin{align*}
\norm{\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec m+\vec v)-\mathrm{d} P_\ensuremath{\mathcal{M}}(\vec z)}_2
&\le \frac{1}{\tau} \norm{\vec m - \vec z}_{2}
+ \frac{1}{\tau- \norm{\vec v}_{2}} \norm{\vec v}_{2}\\
&\leq\left(\frac 2\tau +\frac{1}{\tau-\norm{\vec v}_2}\right)
\norm{\vec m+\vec v-\vec z}_2.
\end{align*}
\end{theorem}
\begin{proof}
Using Theorem~\ref{thm:B} and Theorem~\ref{thm:C2Teil1} we find
\begin{align*}
\label{eq:split}
\norm{\mathrm{d} P_M(\vec m+\vec v) - \mathrm{d} P_M(\vec z)}_2
&\le \norm{\mathrm{d} P_M(\vec m+\vec v)-\mathrm{d} P_M(\vec m)}_2
+ \norm{ P_{T_\vec m\ensuremath{\mathcal{M}}}-P_{T_\vec z\ensuremath{\mathcal{M}}}}_2 \\
&\le \norm{\vec B_\vec v\left(\vec I+\vec B_\vec v\right)^{-1} }_2
+ \tfrac{1}{\tau} \norm{\vec m - \vec z}_{2}.
\end{align*}
From Lemma~\ref{lemma:curvatureReach} we know that $\norm{\vec B_\vec v} \leq
\frac{\norm{\vec v}_2}{\tau}<1$. This allows us to bound the second term by
\begin{equation*}
\norm{\vec B_{\vec v}\left(\vec I+\vec B_{\vec v}\right)^{-1}}_{2}
\le \frac{\norm{\vec B_{\vec v}}}{1-\norm{\vec B_{\vec v}}}
\le \frac{\norm{\vec v}_{2}}{\tau - \norm{\vec v}_{2}},
\end{equation*}
which implies the first inequality of the theorem
\begin{equation}
\label{eq:8}
\norm{\mathrm{d} P_M(\vec m+\vec v)-\mathrm{d} P_M(\vec z)}_{2}
\le \frac{1}{\tau} \norm{\vec m - \vec z}_{2}
+ \frac{1}{\tau- \norm{\vec v}_{2}} \norm{\vec v}_{2}.
\end{equation}
Since $\norm{\vec v}_{2}<\tau$ the point $\vec m + \vec v$ is within the tubular
neighborhood of $\mathcal M$ and, hence
$\norm{\vec v}_2 \leq \norm{\vec m+\vec v-\vec z}_2$. Together with the
triangle inequality this gives us
$\norm{\vec m -\vec z}_2\leq 2 \norm{\vec m +\vec v-\vec z}_2 $. Including
these two inequalities into \eqref{eq:8} we obtain the assertion.
\end{proof}
We observe that the constants in Theorem~\ref{thm:Psi_bound} become large if
either the reach of the manifold becomes small or the point $\vec m + \vec v$
is close to the boundary of the tubular neighborhood of $\ensuremath{\mathcal{M}}$.
\section{Manifold-valued Approximation}
\label{sec:manifold}
In this section we generalize arbitrary approximation operators for vector
valued functions to approximation operators for manifold-valued functions.
To this end we consider for an arbitrary domain $\Omega$
a generic approximation operator
$I_{\ensuremath{\mathbb{R}}^d} \colon C(\Omega,\ensuremath{\mathbb{R}}^d)\rightarrow
C(\Omega,\ensuremath{\mathbb{R}}^d)$. For an embedded manifold
$\ensuremath{\mathcal{M}} \subset \ensuremath{\mathbb{R}}^{d}$ with reach $\tau$ and projection operator
\begin{equation*}
P_{\ensuremath{\mathcal{M}}} \colon U \to \ensuremath{\mathcal{M}}, \quad
U = \{\vec y \in \ensuremath{\mathbb{R}}^{d} \mid \min_{\vec m \in \ensuremath{\mathcal{M}}} \norm{\vec y - \vec m} < \tau\},
\end{equation*}
we define the approximation operator
$I_{\ensuremath{\mathcal{M}}} \colon C(\Omega,\ensuremath{\mathcal{M}}) \to C(\Omega,\ensuremath{\mathcal{M}}) $ for manifold-valued functions as
\begin{equation*}
I_\ensuremath{\mathcal{M}} f = P_\ensuremath{\mathcal{M}}\circ I_{\ensuremath{\mathbb{R}}^d}f.
\end{equation*}
It is important to note that $I_{\ensuremath{\mathcal{M}}}$ is not defined for all functions
$f \in C(\Omega,\ensuremath{\mathcal{M}})$, but only for those for which $I_{\ensuremath{\mathbb{R}}^{d}} f(x)$
is within the reach of the manifold $\ensuremath{\mathcal{M}}$, i.e., $\norm{f(\vec x) - I_{\ensuremath{\mathbb{R}}^{d}}
f(x)}_{2}\le \tau$ for all $\vec x \in \Omega$.
It is straight forward to see that operator $I_{\ensuremath{\mathcal{M}}}$ has the same order of
approximation as $I_{\ensuremath{\mathbb{R}}^{d}}$, c.f.~\cite{Ga18}.
\begin{theorem}\label{satz:wichtig}
Let $f\in C(\Omega,\ensuremath{\mathcal{M}})$ be a continuous $\ensuremath{\mathcal{M}}$-valued function such that for
all $\vec x \in \Omega$, $I_{\ensuremath{\mathbb{R}}^d}f(\vec x)$ is contained in the reach of
$\ensuremath{\mathcal{M}}$. We then have for all $\vec x\in \Omega$
\begin{equation}\label{eq:err_func}
\norm{I_\ensuremath{\mathcal{M}} f(\vec x) - f(\vec x)}_2
\leq 2\,\norm{I_{\ensuremath{\mathbb{R}}^{d}} f(\vec x) - f(\vec x)}_2.
\end{equation}
\end{theorem}
\begin{proof}
Since $f$ has function values on $\ensuremath{\mathcal{M}}$, it follows from the definition of
$P_\ensuremath{\mathcal{M}}$ in equation~\eqref{eq:2} for all $\vec x\in \Omega$ that
\begin{equation*}
\left\|I_\ensuremath{\mathcal{M}} f(\vec x) - I_{\ensuremath{\mathbb{R}}^{d}} f(\vec x)\right\|_2
\leq \left\|f(\vec x) - I_{\ensuremath{\mathbb{R}}^{d}} f(\vec x) \right\|_2.
\end{equation*}
Because of the triangle inequality and the definition of $I_{\ensuremath{\mathcal{M}}}$ we have
\begin{equation*}
\norm{I_\ensuremath{\mathcal{M}} f(\vec x) - f(\vec x)}_2
\leq \norm{I_\ensuremath{\mathcal{M}} f(\vec x) - I_{\ensuremath{\mathbb{R}}^{d}} f(\vec x)}_2
+ \norm{I_{\ensuremath{\mathbb{R}}^{d}} f(\vec x) - f(\vec x)}_2,
\quad \vec x \in \Omega.\qedhere
\end{equation*}
\end{proof}
As we will see later, considering the error of the differential, things become more complicated.
\subsection{Approximation Order of the Differential}
\label{sec:app_diff}
In this section we are interested in the approximation error
$\norm{\d I_{\ensuremath{\mathcal{M}}} f - \d f}_2$ between the differential of the manifold-valued
approximation $\d I_{\ensuremath{\mathcal{M}}} f$ and the original differential $\d f$. To this end
we assume from now on that both, $f \colon \Omega \to \ensuremath{\mathbb{R}}^{d}$ and the vector-valued
approximation $\tilde f = I_{\ensuremath{\mathbb{R}}^{d}} f$, are differentiable.
While the error bound for $I_{\ensuremath{\mathcal{M}}} f$ is independent of the geometry of the
manifold $\ensuremath{\mathcal{M}}$, we will see that this is not true for the
differential $\d I_{\ensuremath{\mathcal{M}}} f$ of the manifold-valued approximation. Moreover, it
is not sufficient to ensure that $\tilde f$ is contained in the reach of
$\ensuremath{\mathcal{M}}$, but instead, it must be bounded away from the reach by some positive
constant.
\begin{theorem}\label{satz:wichtigableitung}
Let $\tau$ be the reach of the manifold $\mathcal M$, $\varepsilon<\tau$ and
$f\in C^1(\Omega, \ensuremath{\mathcal{M}})$, such that $\tilde{f}(\vec x) = I_{\ensuremath{\mathbb{R}}^{d}} f(\vec x)$
satisfies for all $\vec x\in \Omega$,
\begin{equation*}
\norm{f(\vec x)-\tilde{f}(\vec x)}_2\leq \varepsilon
\end{equation*}
and, consequently, is contained in the $\varepsilon$-tubular neighborhood of
$\ensuremath{\mathcal{M}}$. Then we have for all $\vec x \in \Omega$ the following upper bound on the
approximation error of the differential $\d I_\ensuremath{\mathcal{M}} f$,
\begin{equation}
\label{eq:error_diff}
\begin{split}
\norm{\d I_\ensuremath{\mathcal{M}} f(\vec x) - \d \,f(\vec x)}_{2}
\leq \norm{\d \tilde{f}(\vec x) - \d f(\vec x)}_{2} + C \norm{f(\vec x)-\tilde{f}(\vec x)}_2
\end{split}
\end{equation}
where $C$ is given by
\begin{equation*}
C = \left(\frac 2\tau+\frac{1}{\tau-\varepsilon}\right)
\left(\norm{\d \tilde{f}(\vec x) - \d f(\vec x)}_{2}
+\norm{\d f(\vec x)}_{2}\right).
\end{equation*}
\end{theorem}
\begin{proof}
By the chain rule we obtain for all $\vec x \in \Omega$,
\begin{equation*}
\d I_\ensuremath{\mathcal{M}} f(\vec x)=\mathrm{d} P_\ensuremath{\mathcal{M}}(\tilde{f}(\vec x))\circ\d\tilde{f}(\vec x)
\end{equation*}
and from $P_\ensuremath{\mathcal{M}} f= f$,
\begin{equation*}
\d f(\vec x)
= \d (P_\ensuremath{\mathcal{M}} f)(\vec x)=\mathrm{d} P_\ensuremath{\mathcal{M}} (f(\vec x))\circ \d f(\vec x).
\end{equation*}
Using the expansion
\begin{align*}
\d (I_\ensuremath{\mathcal{M}} f)(\vec x) - \d f(\vec x)
=&\mathrm{d} P_\ensuremath{\mathcal{M}}(\tilde{f}(\vec x))\circ \d \tilde{f}(\vec x)-
\mathrm{d} P_\ensuremath{\mathcal{M}} (f(\vec x))\circ \d f(\vec x)\\
=& (\mathrm{d} P_\ensuremath{\mathcal{M}}(\tilde{f}(\vec x))-\mathrm{d} P_\ensuremath{\mathcal{M}} (f(\vec x)))
\circ \d \tilde{f}(\vec x)\\
&+ \mathrm{d} P_\ensuremath{\mathcal{M}} f(\vec x) \circ
( \d\tilde{f}(\vec x)-\d f(\vec x))
\end{align*}
we conclude that
\begin{align*}
\norm{\d I_\ensuremath{\mathcal{M}} f(\vec x) - \d f(\vec x)}_2
\le& \norm{\mathrm{d} P_\ensuremath{\mathcal{M}}(\tilde{f}(\vec x))-\mathrm{d} P_\ensuremath{\mathcal{M}} (f(\vec x))}_2
\norm{\d \tilde{f}(\vec x)}_2\\
&+ \norm{\mathrm{d} P_\ensuremath{\mathcal{M}} (f(\vec x))}_2
\norm{\d\tilde{f}(\vec x)-\d f(\vec x)}_2.
\end{align*}
Since $f(\vec x) \in \ensuremath{\mathcal{M}}$ and $\norm{f(\vec x) - \tilde f(\vec x)}_2 <
\varepsilon$ we have by Theorem~\ref{thm:Psi_bound}
\begin{equation*}
\norm{\mathrm{d} P_\ensuremath{\mathcal{M}}(\tilde{f}(\vec x))-\mathrm{d} P_\ensuremath{\mathcal{M}} (f(\vec x))}
\le \left(\frac 2\tau+\frac{1}{\tau -\varepsilon}\right)
\norm{f(\vec x)-\tilde{f}(\vec x)}_2.
\end{equation*}
Together with the fact that $\norm{\mathrm{d} P_\ensuremath{\mathcal{M}} f(\vec x)}_2=1$ we obtain
\begin{align*}
\norm{\d (I_\ensuremath{\mathcal{M}} f)(\vec x) - \d f(\vec x)}_2
\le &\norm{f(\vec x)-\tilde{f}(\vec x)}_2
\left(\frac 2\tau+\frac{1}{\tau -\varepsilon}\right)
\norm{\d \tilde{f}(\vec x)}_2\\
&+ \norm{\d\tilde{f}(\vec x)-\d f(\vec x)}_2.
\end{align*}
This implies the assertion by triangle inequality.
\end{proof}
Asymptotically, as $\norm{f(\vec x)-\tilde{f}(\vec x)}_2 \to 0$ we
have $C \to \frac{3}{\tau} \norm{\d f(x)}_{2}$. Since for most approximation
methods the decay of the differential
$\norm{\d f(\vec x)-\d \tilde{f}(\vec x)}_2$ is one order slower than the
decay of $\norm{f(\vec x)-\tilde{f}(\vec x)}_2$, we conclude that the right
hand bound in \eqref{eq:error_diff} is dominated by the first summand and,
hence, the approximation error of the differential $\d I_\ensuremath{\mathcal{M}} f$ of the
manifold-valued approximant coincides asymptotically with the approximation
error of the vector-valued approximant, as it was already reported in
\cite{Ga18}. However, the pre-asymptotic behavior depends strongly on the
reach of the embedding of the manifold $\mathcal M$.
\subsection{Fourier Interpolation}
\label{sec:four-interp}
In this section we want to illustrate Theorem~\ref{satz:wichtigableitung}
using Fourier-Interpolation as the approximation operator $I_{\ensuremath{\mathbb{R}}^{d}}$.
More precisely, we define for a function $f \in C(\ensuremath{\mathbb{T}}, \ensuremath{\mathbb{R}}^{d})$ on the torus
$\ensuremath{\mathbb{T}}$ the Fourier partial sum
\begin{equation*}
I_{\ensuremath{\mathbb{R}}^{d}} f(t) = S_n f(t) = \sum_{k=-n}^n c_k(f) e^{2\pi i k t}
\end{equation*}
with the vector-valued Fourier coefficients
\begin{equation*}
c_{k}(f) = \int_{0}^{1} f(x) e^{-2\pi i k x} \d{x}.
\end{equation*}
The Fourier-Interpolation satisfies the following well known approximation
inequalities, cf.~\cite{PlPoStTa18}.
\begin{theorem}
\label{thm:Sn}
Let $r\in \ensuremath{\mathbb{N}}$ with $r\geq2$ and $f\in C^r(\ensuremath{\mathbb{T}},\ensuremath{\mathbb{R}}^d)$. Then
\begin{align*}
\norm{f(x)-S_nf(x)}_{2}
&\leq \frac{\sqrt{2d}}{(2\pi)^r}\,\frac{\sqrt n}{n^r}\,\norm{f^{(r)}}_{L^2(\ensuremath{\mathbb{T}}),2},\\
\norm{\d f( x)-\d (S_nf)(x)}_2
&\leq \frac{\sqrt{2d}}{(2\pi)^{r-1}}\,\frac{\sqrt n}{n^{r-1}}\,\norm{f^{(r)}}_{L^2(\ensuremath{\mathbb{T}}),2},
\end{align*}
with the norm $\norm{f}_{L^2(\ensuremath{\mathbb{T}}),2}^{2} = \int_{\ensuremath{\mathbb{T}}} \norm{f(x)}_2^{2} \d x.$
\end{theorem}
\begin{proof}
The first bound can be found in \cite[Theorem 4.3]{Co16}. The second bound
follows from $\d (S_nf)=S_n(\d f)$ and the fact that the regularity of
$\d f$ is one less than the regularity of $f$. The factor $\sqrt{d}$ comes
from the fact that the function $f$ maps in the d-dimensional space.
\end{proof}
In \cite[Theorem 1.39]{PlPoStTa18} a similar bound for the $L^\infty(\ensuremath{\mathbb{T}})$-norm
can be found. Using the Fourier partial sum operator as the approximation
operator $I_{\ensuremath{\mathbb{R}}^{d}}$ Theorem~\ref{satz:wichtigableitung} becomes the
following.
\begin{theorem}
\label{thm:FPS}
Let $\mathcal M \subset \ensuremath{\mathbb{R}}^{d}$ be a submanifold with reach $\tau>0$ and
$f\in C^r(\ensuremath{\mathbb{T}},\ensuremath{\mathcal{M}})$ an $r \geq 2$ times differentiable function with values in
$\mathcal M$. Let, furthermore, $\varepsilon < \tau$ be an auxiliary
constant and the bandwidth $n$ of the Fourier partial sum $S_{n} f$ at least
such that $\norm{f( x)-S_{n} f(x)}_{2} \le \varepsilon$ for all $x\in \ensuremath{\mathbb{T}}$, i.e.,
\begin{equation}
\label{eq:error_in_reach}
n^{r-\tfrac 12} \ge \frac{C_{1} \norm{f^{(r)}}_{L^2(\ensuremath{\mathbb{T}}),2}}{\varepsilon}
\text{ with }
C_{1} = \frac{ \sqrt{2d}}{(2\pi)^r}.
\end{equation}
Then the projection $P_{\mathcal M} \circ S_{n} f$ of the Fourier partial
sum satisfies for all $x \in \ensuremath{\mathbb{T}}$,
\begin{align}
\label{eq:error_fourier}
\left\|P_{\mathcal M} \circ S_{n} f(x) - f(x)\right\|_2
&\leq 2 \,C_{1} \,n^{\tfrac12 -r}\,\norm{f^{(r)}}_{L^2(\ensuremath{\mathbb{T}}),2},
\end{align}
whereas for its differential $\d ( P_{\mathcal M}
\circ S_{n} f)$ we obtain
\begin{equation}
\label{eq:error_diff_fourier}
\begin{split}
\left\|\d ( P_{\mathcal M} \circ S_{n} f)(x) - \d f(x)\right\|_2
&\leq
2\pi\, C_{1}\, \,n^{\tfrac 32-r}\, \norm{f^{(r)}}_{L^2(\ensuremath{\mathbb{T}}),2}\\
&+C_{2}\, n^{\tfrac 12-r} \,\norm{f^{(r)}}_{L^2(\ensuremath{\mathbb{T}}),2}^{2},
\end{split}
\end{equation}
with the constant
\begin{align*}
C_{2} = C_{1}\,(\tfrac{2}{\tau} + \tfrac{1}{\tau - \varepsilon})\,(1 + 2\pi\, C_{1}^{2}\, n^{\frac32-r}).
\end{align*}
\end{theorem}
\begin{proof}
Together with Theorem~\ref{thm:Sn} condition~\eqref{eq:error_in_reach}
ensures that
\begin{equation*}
\norm{f(x) - S_{n} f(x)}_{2} \le \varepsilon < \tau, \quad x \in \ensuremath{\mathbb{T}}
\end{equation*}
and, hence, $S_nf(x)$ has distance less than the reach to $\ensuremath{\mathcal{M}}$ for all
$x\in \ensuremath{\mathbb{T}}$. This allows us to apply Theorem~\ref{satz:wichtig} in
conjunction with Theorem~\ref{thm:Sn} to conclude \eqref{eq:error_fourier}.
For the approximation error of the derivative we have by
Theorem~\ref{satz:wichtigableitung}
\begin{equation*}
\norm{\d ( P_{\mathcal M} \circ S_{n} f)(x) - \d f(x)}_{2}
\le \norm{\d (S_{n} f)(x) - \d f(x)}_{2} + C \norm{S_{n} f(x) - f(x)}_{2}
\end{equation*}
with
\begin{equation*}
C = \left(\tfrac{\tau}2 + \tfrac{1}{\tau-\varepsilon}\right) \, \left(\norm{\d (S_{n} f)(x) -
\d f(x)}_{2} + \norm{\d f(x)}_{2}\right).
\end{equation*}
Together with Theorem~\ref{thm:Sn} this yields
\begin{equation*}
\norm{\d ( P_{\mathcal M} \circ S_{n} f)(x) - \d f(x)}_{2}
\le 2\pi\, C_{1} \,n^{\frac{3}{2}-r} \,\norm{f^{(r)}}_{L^2(\ensuremath{\mathbb{T}}),2} + C \,C_{1}\, n^{\frac{1}{2}-r}\,\norm{f^{(r)}}_{L^2(\ensuremath{\mathbb{T}}),2}
\end{equation*}
and
\begin{align*}
C &\le (\tfrac{\tau}2 + \tfrac{1}{\tau-\varepsilon})\,
(2\pi\,C_{1}\, n^{\frac{3}{2}-r} \norm{f^{(r)}}_{L^{2}(\ensuremath{\mathbb{T}}),2} + \norm{\d f(x)}_{2})\\
& \le (\tfrac{\tau}2 + \tfrac{1}{\tau-\varepsilon})\,
(2 \pi\, C_{1} \,n^{\frac{3}{2}-r} +1 )\norm{f^{(r)}}_{L^{2}(\ensuremath{\mathbb{T}}),2},
\end{align*}
where we have used the fact that
$\norm{\d f(x)}_2 \le \norm{f^{(r)}}_{L^{2}(\ensuremath{\mathbb{T}}),2}$ for periodic functions. Setting
\begin{equation*}
C_{2} = C_{1} (\tfrac{\tau}2 + \tfrac{1}{\tau-\varepsilon}) (1+2 \pi \,C_{1}\,
n^{\frac{3}{2}-r} )
\end{equation*}
yields the assertion.
\end{proof}
Since for $n \to \infty$ we have
$C_{2} \to C_{1}(\frac{2}{\tau}+\frac{1}{\tau-\varepsilon})$.
Theorem~\ref{thm:FPS} states that the approximation order of the differential
of the manifold-valued Fourier partial sum operator coincides with the
approximation order of the vector-valued operator. In the
preasymptotic setting, however, also the second summand with the faster rate
$n^{\frac{1}{2}-r}$ is relevant. The constant of this second summand becomes
large if the point-wise approximation error is close to the reach $\tau$ of
the embedding.
\section{Examples}\label{sec:examples}
In this section we apply our findings to two real world examples of
manifold-valued approximation. Both examples are related to the analysis of
crystalline materials. In the first example we consider functions that relate
propagation directions of waves to polarization directions and in the second
example we consider functions that relate points within crystalline specimen
to the local orientation of its crystal lattice. Both examples have
been realized using Matlab Toolbox \texttt{MTEX 5.8},
cf.~\cite{MaHiSc11}. The corresponding scripts and data files can be found
at
\url{https://github.com/mtex-toolbox/mtex-paper/tree/master/manifoldValuedApproximation}.
\subsection{Wave Velocities}
\label{sec:wave-velocities}
In crystalline materials the propagation velocity and polarization direction
of waves is often isotropic, i.e., it depends on the propagation direction
relative to the crystal lattice. This posses an important issue in seismology
where one analyzes the distribution of earthquake waves in order to get a
deeper understanding of the core of the earth,
cf. \cite{mainprice:hal-00408321}. Each earthquake wave decomposes into a
p-wave and two perpendicular shear-wave components. The polarization vectors
of p-wave components as well as of the two s-wave components depend on the
propagation direction of the wave relative to the crystal,
cf. \cite{nye85, MaHiSc11}. Mathematically, the directional dependency of the
polarization directions from the propagation direction is modeled as function
\begin{equation*}
f \colon \mathbb S^{2} \to \mathbb RP^{2}
\end{equation*}
from the two-sphere $\mathbb S^{2}$ into the two--dimensional projective space
$\mathbb RP^{2}$. Our goal is to approximate this function from finite
measurements $\vec y_{n} = f(\vec x_{n}) \in \mathbb RP^{2}$, $n=1,\ldots,N$.
To this end, we identify the two dimensional projective space $\mathbb RP^{2}$
with the quotient $\mathbb S^{2}/{\sim}$ with respect to the equivalence
relation $\vec x \sim -\vec x$ and consider the embedding
$\mathcal E \colon \mathbb S^{2}/{\sim} \to \ensuremath{\mathbb{R}}^{3 \times 3}$,
$\mathcal E(\vec x) = \vec x \vec x^{\top}$. The reach of this embedding is
$\tau = \tfrac{1}{\sqrt{2}}$ as we show in the following lemma.
\begin{lemma}
The two dimensional projective space $\mathbb RP^{2}$ embedded into the
space of symmetric $3 \times 3$ matrices
$\mathcal E(\mathbb S^{2}/\!\! \sim) \subset \ensuremath{\mathbb{R}}^{3 \times 3}$ has the reach
$\tau = \frac{1}{\sqrt{2}}$.
\end{lemma}
\begin{proof}
Following~\cite[Thm. 2.2]{aamari19}, we can estimate the reach by the
following infimum
\begin{equation}\label{eq:tau}
\tau
= \inf_{\vec x\neq \vec y\in \mathbb S^{2}/{ \sim}}
\frac{\norm{\ensuremath{\mathcal{E}}(\vec x)-\ensuremath{\mathcal{E}}(\vec y)}_2^2}
{2\, d(\ensuremath{\mathcal{E}}(\vec x)-\ensuremath{\mathcal{E}}(\vec y),T_{ \vec y}\mathbb S^{2}/{\sim})}.
\end{equation}
Since in our setting in both spaces, $\mathbb S^{2}/{ \sim}$ and
$\ensuremath{\mathbb{R}}^{3\times 3}$ the metric is invariant with respect to the action of
$\ensuremath{\mathrm{SO}(3)}$, it suffices to take the infimum for $\vec y=\vec e_1=(1, 0, 0)^\top$.
We define the other canonical basis vectors in $\ensuremath{\mathbb{R}}^3$ as
$\vec e_2=(0, 1, 0)^\top$ and $\vec e_3=(0, 0, 1)^\top$ The tangent vector
space in $\vec e_1$ is then given by these tangent vectors:
\begin{equation*}
T_{(1, 0,0)^\top}\mathbb S^{2}/{ \sim}
= \lin\left\{\frac{1}{\sqrt 2}(\vec e_2\vec e_1^\top
+ \vec e_1\vec e_2^\top),\frac{1}{\sqrt 2}(\vec e_3\vec e_1^\top
+ \vec e_1\vec e_3^\top)\right\}.
\end{equation*}
Hence, we write
$\vec v = \ensuremath{\mathcal{E}}(\vec x)-\ensuremath{\mathcal{E}}(\vec e_1)=\vec x\vec x^\top-\vec e_1\vec e_1^\top$ and therefore we
can calculate the reach by
\begin{align*}
\tau
&=\inf_{\vec x\neq \vec e_1\in \mathbb S^{2}/{\sim}}
\frac{\norm{\vec v}_2^2}{2\, \norm{\vec v-\langle \vec v,\frac{1}{\sqrt
2}\vec e_2\vec e_2^\top\rangle
-\langle \vec v,\frac{1}{\sqrt 2}\vec e_3\vec e_3^\top\rangle}_2}\\
&=\inf_{\vec x\neq \vec e_1\in \mathbb S^{2}/{\sim}}\frac{2-2\,x_1^2}{2\,\sqrt{2-2\,x_1^2-2x_1^2x_2^2-2x_1^2x_3^2}}\\
&=\inf_{\vec x\neq \vec e_1\in \mathbb
S^{2}/{\sim}}\frac{1-x_1^2}{\sqrt{2}\,\sqrt{(1-x_1^2)^2}}=\frac{1}{\sqrt 2},
\end{align*}
which finishes the proof.
\end{proof}
The calculation of the reach gives us the constants in
Theorem~\ref{satz:wichtig} and \ref{satz:wichtigableitung} for this specific
manifold.
Let $\mathcal E \circ f \colon \mathbb S^{2} \to \ensuremath{\mathbb{R}}^{3 \times 3}$ be the
embedded function. A common method, cf.~\cite{Mic13}, of approximating the
spherical function $\mathcal E \circ f$ is by linear combinations of spherical
harmonics $Y_{\ell,k}$, $\ell=0,\ldots,L$, $k = -\ell,\ldots,\ell$ up to a
fixed bandwidth $L$,
\begin{equation*}
S_{L} (\mathcal E \circ f)(\vec x)
= \sum_{\ell=0}^{L} \sum_{k=-\ell}^{\ell} c_{\ell,k} Y_{\ell,k}(\vec x),
\end{equation*}
where the coefficients $c_{\ell,k} \in \ensuremath{\mathbb{R}}^{3 \times 3}$,
$k=-\ell,\ldots,\ell$, $\ell=0,\ldots,L$ are elements of the embedding space.
This expansion in spherical harmonics coincides with the linear
Fourier operator in Section~\ref{sec:four-interp} defined for functions over
the sphere. In our little example we simply assume that the measurement
points $\vec x_{n}$ together with some weights $\omega_{n}$ form a spherical
quadrature rule up to degree $2L$ which allows us to determine the Fourier
coefficients $c_{\ell,k}$, by
\begin{equation*}
c_{k,\ell} = \sum_{n=1}^{N} \omega_{n} \mathcal E(\vec y_{n}) \overline{Y_{\ell,k}(\vec x_{n})}.
\end{equation*}
Fig.~\ref{fig:wave_a} displays the theoretical polarization directions of an
Olivine crystal in dependency of the propagation direction. We observe the
points of singularity, marked by the black squares. In order to
approximate this non-smooth function we fixed the bandwidth $L=8$
and used 144 Chebyshev quadrature nodes
$\vec x_{1},\ldots,\vec x_{144} \in \mathbb S^{2}$ as sampling points,
cf.~\cite{Gr_pointsquad}. These quadrature nodes are approximately
equispaced and are displayed as red lines in Fig.~\ref{fig:wave_a}.
The approximated function $P_{\ensuremath{\mathbb{R}} P^{2}} \circ S_{8}(\mathcal E \circ f)$ is
depicted in Fig.~\ref{fig:wave_b} and shows good approximation with the
original function away from the singularity points. This is supported
by a plot of the point-wise error
$\norm{ f(\vec x) - P_{\ensuremath{\mathbb{R}} P^2}\circ S_8 (\ensuremath{\mathcal{E}} \circ f)(\vec x)}_2$ in
Fig.~\ref{fig:wave_c}. Note that we measure the error in the Euclidean norm
of the $9$-dimensional embedding space $\ensuremath{\mathbb{R}}^9$, which is for
$\vec x,\vec y\in \ensuremath{\mathbb{R}} P^2 $ equal to the Frobenius norm
$\norm{\vec x\vec x^\top -\vec y\vec y^\top}_F$. Compared to this, Figure~\ref{fig:wave_d}
shows the error of the linear approximation $\norm{ f(\vec x) - S_8 (\ensuremath{\mathcal{E}} \circ f)(\vec x)}$,
which is half of the error bound from~\ref{satz:wichtig}. Additionally, we marked the areas where the
residual is bigger than the reach blue. In this regions our Theorems are not
applicable, since there the assumption $I_{\ensuremath{\mathbb{R}}^d}f$ is within the reach is not met.
\begin{figure}[t]
\centering
\begin{subfigure}[T]{0.4\textwidth}
\includegraphics[width=0.8\textwidth]{ps1Data}
\centering
\subcaption{$f \colon \mathbb S^{2} \to \ensuremath{\mathbb{R}} P^{2}$}
\label{fig:wave_a}
\end{subfigure}
\begin{subfigure}[T]{0.4\textwidth}
\includegraphics[width=0.8\textwidth]{ps1Approx}
\centering
\subcaption{$P_{\ensuremath{\mathbb{R}} P^2}\circ S_8 (\ensuremath{\mathcal{E}} \circ f)$}
\label{fig:wave_b}
\end{subfigure}
\begin{subfigure}[T]{0.4\textwidth}
\includegraphics[height=0.8\textwidth]{ps1error.pdf}
\centering
\subcaption{$\norm{f-P_{\ensuremath{\mathbb{R}} P^2}\circ S_8 (\ensuremath{\mathcal{E}} \circ f)}_{2}$.}
\label{fig:wave_c}
\end{subfigure}
\begin{subfigure}[T]{0.4\textwidth}
\includegraphics[height=0.8\textwidth]{ps1error8_lin.pdf}
\centering
\subcaption{$\norm{f- S_8 (\ensuremath{\mathcal{E}} \circ f)}_{2}$.}
\label{fig:wave_d}
\end{subfigure}
\caption{Polarization directions of the fastest shear wave with
respect to the propagation direction $\vec x \in \mathbb S^{2}$ plotted as
vector fields on the upper hemisphere. The left upper image (a) displays the true
polarization directions $f(\vec x) \in \ensuremath{\mathbb{R}} P^{2}$. The right upper image (b)
is the harmonic approximation $P_{\ensuremath{\mathbb{R}} P^2}\circ S_8 (\ensuremath{\mathcal{E}} \circ f)(\vec x)$
using the sampling points marked red in (a). The lower left image (c)
displays the norm of the point-wise residual. The lower right image (d)
displays the error of the linear approximation, i.e. exactly half of the
upper bound from Theorem~\ref{satz:wichtig}. We marked red the regions
where this residual is bigger than the reach of the manifold.}
\label{fig:wave}
\end{figure}
We determined the derivatives of $f$ numerically by choosing a basis
${\vec t_1,\vec t_2}$ in the tangent space $T_{\vec x}\S^2$ and
approximating the columns $\vec v_i \in \ensuremath{\mathbb{R}}^{9}$, $i=1,2$ of
$\d f(\vec x)\in \ensuremath{\mathbb{R}}^{9\times 2}$ by the difference quotients
\begin{equation*}
\vec v_i=\frac{1}{h} \left(f\bigl(\tfrac{\vec x+h\,\vec t_i}{\norm{\vec x+h\,\vec t_i}_2}\bigr)-f(\vec x)\right),
\end{equation*}
with $h=10^{-6}$. The norm of the derivative is depicted in
Fig.~\ref{fig:dwave_a} and clearly shows the position of the singularities.
In Fig.~\ref{fig:dwave_b} the error between $\mathrm d f(\vec x)$ and the
differential of harmonic approximation
$\d (P_{\ensuremath{\mathbb{R}} P^{2}} \circ S_{64})( \ensuremath{\mathcal{E}} \circ f)(\vec x)$ is plotted as a
function of the propagation direction $\vec x\in \mathbb S^{2}$. Since the
differential $\mathrm d f(\vec x) $ is a matrix in $\ensuremath{\mathbb{R}}^{9\times 2}$, we
consider here the spectral norm of the error matrix. In order to illustrate
our theoretical result of Theorem~\ref{satz:wichtigableitung} we plotted our
theoretical upper bound on that approximation error of the derivative in
Fig.~\ref{fig:dwave_c}. It should be noted that for the differential we
needed to increased the polynomial degree to $L=64$ with $21000$ sample
points in order to obtain a reasonable approximation at some distance to the
singularities.
\begin{figure}[tb]
\begin{subfigure}[T]{0.29\textwidth}
\includegraphics[height=\textwidth]{dps1.png}
\subcaption{$\norm{d f(x)}_2$}
\label{fig:dwave_a}
\end{subfigure}
\qquad
\begin{subfigure}[T]{0.29\textwidth}
\includegraphics[height=\textwidth]{dps1Error.png}
\subcaption{$\norm{d f(x) - d P_{\ensuremath{\mathbb{R}} P^2}\circ \tilde f}_2$}
\label{fig:dwave_b}
\end{subfigure}
\qquad
\begin{subfigure}[T]{0.29\textwidth}
\includegraphics[height=\textwidth]{dps1Bound.png}
\subcaption{Bound from Thm.~\ref{satz:wichtigableitung}.}
\label{fig:dwave_c}
\end{subfigure}
\caption{The left image (a) shows the point-wise norm of the
differential $\d f$. The middle image (b) depicts the approximation
error between the differential of the original function $f$ and the
differential of its harmonic approximation
$P_{\ensuremath{\mathbb{R}} P^{2}} \circ \tilde f = P_{\ensuremath{\mathbb{R}} P^{2}} \circ S_{64}(\mathcal E
\circ f)$. The right images (c) gives the upper bound for (b) from
Theorem~\ref{satz:wichtigableitung}.}
\label{fig:dwave}
\end{figure}
\subsection{Electron Back Scatter Diffraction}
\label{sec:so3subset-r3times-3}
The subject of crystallographic texture analysis is the microstructure of
polycrystalline materials. Locally the microstructure is described by the
orientation of the atom lattice with respect to some specimen fixed reference
frame. More specifically, one describes the local orientation of the atom
lattice by a coset $[\vec R]_{\mathcal S} \in \ensuremath{\mathrm{SO}(3)} / \mathcal S$ of the
rotation group $\ensuremath{\mathrm{SO}(3)}$ modulo the finite subgroup $\mathcal S \subset \ensuremath{\mathrm{SO}(3)}$,
called point group. The point group of a crystal consists of all symmetries
of its atom lattice and is either one of the cyclic groups $C_{1}$, $C_{2}$,
$C_{3}$, $C_{4}$, $C_{6}$, the dihedral groups $D_{2}$, $D_{3}$, $D_{4}$,
$D_{6}$, the tetragonal group $T$ or the octahedral group $O$. Assuming a
monophase material, i.e., a material consisting only of a single type of
crystals, the variation of the local orientation of the atom lattice at the
surface $\Omega \subset \ensuremath{\mathbb{R}}^{2}$ of the specimen is modeled by the map
\begin{equation*}
f \colon \Omega \to \ensuremath{\mathrm{SO}(3)}/\mathcal S.
\end{equation*}
The gradient of the function $f$, also called lattice curvature tensor
$\vec \kappa(\vec x)$, is closely related to elastic and plastic deformations
the specimen has been exposed to. More specifically, it is related via the Nye
equation to the dislocation density tensor $\vec \alpha(\vec x)$,
\cite{Pa08,KoZaRa15} that describes how many lattice dislocations are
geometrically necessary in order to preserve the compatibility of the lattice
for a given deformation. Hence, estimating $f$ and its derivatives from
experimental data is a central problem in material science.
Electron back scatter diffraction (EBSD) is an experimental technique
\cite{ASWK93,KuWrAdDi93} for determining the local lattice orientations
$f(\vec x_{\ell}) \in \ensuremath{\mathrm{SO}(3)}/\mathcal S$ at discrete sampling points
$\vec x_{i,j} \in \Omega$. An example of such EBSD data is the
$\ensuremath{\mathrm{SO}(3)}/\mathcal S$ - valued image displayed in Fig.~\ref{fig:EBSDRAW}. It
describes the variation of lattice orientation at the surface of an Aluminum
alloy of size $200\mu m \times 150 \mu m$ at an resolution of $0.4\mu m$. The
symmetry group in this case is the octahedral group $O$.
\begin{figure}[t]
\begin{subfigure}{0.42\linewidth}
\includegraphics[height=3.9cm]{ebsdRAW2}
\subcaption{full map with a global color key}
\label{fig:EBSDRAWa}
\end{subfigure}
\quad
\begin{subfigure}{0.5\linewidth}
\includegraphics[height=3.9cm]{ebsdAxisAngle87}
\subcaption{single grain with a local color key}
\label{fig:EBSDRAWb}
\end{subfigure}
\centering
\caption{The raw EBSD data. Each of the $410$ x $547$ pixels
corresponds to a single orientation measurement at the surface of the
specimen. The color is computed by the procedure described in
\cite{NoHi17}. The $5\%$ white pixels in Fig. (a) correspond to
corrupted measurements with no data. The data has been segmented into
$92$ grains as outlined by the black boundaries.}
\label{fig:EBSDRAW}
\end{figure}
The data is displayed with respect to two different color keys. In
Fig.~\ref{fig:EBSDRAWa} the colors are assigned globally to the cosets
$f(\vec x) \in \ensuremath{\mathrm{SO}(3)}/O$ as described in \cite{NoHi17}. Regions of similar
lattice orientation form so-called grains as outlined by the black
boundaries. In Fig.~\ref{fig:EBSDRAWb} only the single grain outlined
by the red boundary in Fig.~\ref{fig:EBSDRAWa} is displayed. For this grain
we computed an average lattice orientation $[\vec M] \in \ensuremath{\mathrm{SO}(3)}/O$ and selected
for each coset $f(\vec x) \in \ensuremath{\mathrm{SO}(3)}/O$ the rotation
$\vec R(x) \in [\vec M^{-1} f(\vec x)]$ with the smallest rotational
angle. Next we associated a color to $\vec R(x)$ according to a spherical
color representation where the rotational angle of $\vec R(x)$ determines
the saturation and the rotational axis hue and value. More details on this
orientation coloring can be found in \cite{Thomsen17}.
Estimating the derivative from such a noisy map of lattice orientations is
usually not a good idea as we will see later.
Reducing the noise by means of local approximation methods has been discussed
in \cite{HiSiSc19,Seret19}. In order to demonstrate our embedding based
approximation approach we make use of the locally isometric embedding
$\mathcal E_{O} \colon \ensuremath{\mathrm{SO}(3)}/O \to \ensuremath{\mathbb{R}}^{9}$ described in \cite{HiLi20} and
proceed as follows
\begin{enumerate}
\item Compute an $\ensuremath{\mathbb{R}}^{9}$-valued image $\vec u_{i,j} = \mathcal E_{O}(f(\vec
x_{i,j}))$.
\item Approximate the $\ensuremath{\mathbb{R}}^{9}$-valued image using a cosine series
$\tilde u \colon \Omega \to \ensuremath{\mathbb{R}}^{9}$ computed by robust, penalized least squares
\cite{Gar10}.
\item Evaluate the function $\tilde{u}$ at the grid points $\vec x_{i,j}$ to obtain a
noise reduced $\ensuremath{\mathbb{R}}^{9}$-valued image $\tilde{\vec u}_{i,j}$.
\item Compute the projection of $\vec{\tilde u}_{i,j}$ onto the
embedding $\mathcal E_{O}(\ensuremath{\mathrm{SO}(3)}/O)$ of the quotient and apply the inverse map
$\mathcal E_{O}^{-1}$ to end up with a noise reduced $\ensuremath{\mathrm{SO}(3)}/O$-valued image
$\tilde f(\vec x_{i,j})$.
\end{enumerate}
The resulting image is depicted in Figure~\ref{fig:EBSDSmooth}. We
observe that all no data pixels have been inpainted and that the magnified
part~\ref{fig:EBSDSmoothb} is much less noisy in comparison to
Fig.~\ref{fig:EBSDRAWb}.
\begin{figure}[t]
\centering
\begin{subfigure}{0.42\linewidth}
\includegraphics[height=3.9cm]{ebsdSRAW2}
\subcaption{full map of the approximated data}
\label{fig:EBSDSmootha}
\end{subfigure}
\quad
\begin{subfigure}{0.5\linewidth}
\includegraphics[height=3.9cm]{ebsdSAxisAngle87}
\subcaption{single grain with a local color key}
\label{fig:EBSDSmoothb}
\end{subfigure}
\centering
\caption{EBSD map from Fig.~\ref{fig:EBSDRAW} after the embedding based
approximation approach.}
\label{fig:EBSDSmooth}
\end{figure}
For the computation of the lattice curvature tensor $\vec \kappa$ we use
the skew symmetric matrices
\begin{equation*}
\vec s^{(1)} =
\begin{pmatrix}
0&0&0\\
0&0&-1\\
0&1&0
\end{pmatrix},\quad
\vec s^{(2)} =
\begin{pmatrix}
0&0&-1\\0&0&0\\1&0&0
\end{pmatrix},\quad
\vec s^{(3)} =
\begin{pmatrix}
0&-1&0\\1&0&0\\0&0&0
\end{pmatrix},
\end{equation*}
to fix the basis $\vec R \vec s^{(1)},\vec R \vec s^{(2)},\vec R \vec s^{(3)}$
in the tangential space $T_{\vec R} \ensuremath{\mathrm{SO}(3)}/ O$ at some rotation
$\vec R \in \ensuremath{\mathrm{SO}(3)}$. With respect to this basis the differential
$D\mathcal E(\vec R) \colon T_{\vec R} \ensuremath{\mathrm{SO}(3)}/O \to \ensuremath{\mathbb{R}}^{9}$ of the
embedding $\mathcal E \colon \ensuremath{\mathrm{SO}(3)}/O \to \ensuremath{\mathbb{R}}^{9}$ can be represented as a full
rank $3 \times 9$ matrix. Furthermore, we obtain for the differential
$D \tilde u \colon \ensuremath{\mathbb{R}}^{2} \to \ensuremath{\mathbb{R}}^{9}$ of the embedded image
$\tilde u = \mathcal E \circ \tilde f \colon \Omega \to \ensuremath{\mathbb{R}}^{9}$ at some point
$\vec x \in \Omega$ the matrix product
$ D \tilde u(\vec x) = D \mathcal E(\tilde f(\vec x)) D f(\vec x)$. Hence, the
lattice curvature tensor $\tilde{\vec \kappa}$ of the noise reduced EBSD map
evaluates to
\begin{equation*}
\tilde{\vec \kappa}(\vec x)
= D \tilde f(\vec x)
= \left(D \mathcal E(\tilde f(\vec x)) D \mathcal E(\tilde f(\vec x))^{\top}\right)^{-1}
D \mathcal E(\tilde uf(\vec x))^{\top} D \tilde u(\vec x).
\end{equation*}
The map of the first component $\tilde \kappa_{1,1}$ of the lattice curvature tensor
obtained from the approximating function $\tilde u$ is depicted in
Fig.~\ref{fig:EBSDKappaB}. For comparison we plotted in
Fig.~\ref{fig:EBSDKappaA} a finite difference approximation
\begin{equation*}
\kappa(\vec x_{i,j})_{1,1} = \frac{\log_{f(\vec x_{i,j})}(f(\vec x_{i+1,j}) )}{[x_{i+1,j}-x_{i,j}]_{1}},
\end{equation*}
derived from the discrete data $f(\vec x_{i,j})$. Here, we denoted by
$\log_{\vec R} \colon \ensuremath{\mathrm{SO}(3)}/O \to T_{\vec R} \ensuremath{\mathrm{SO}(3)}/O$ the logarithmic
mapping with respect to the base point $\vec R \in \ensuremath{\mathrm{SO}(3)}/O$. As expected, we observe that
the lattice curvature tensor $\tilde \kappa$ derived from the approximated map
$\tilde u$ is much less noisy.
\begin{figure}[htbp]
\centering
\begin{subfigure}{0.45\linewidth}
\includegraphics[width=\textwidth]{ebsdKappa11}
\caption{$\kappa_{1,1}$ of noisy map.}
\label{fig:EBSDKappaA}
\end{subfigure}
\quad
\begin{subfigure}{0.45\linewidth}
\includegraphics[width=\textwidth]{ebsdSKappa11}
\caption{$\kappa_{1,1}$ of noise reduced map.}
\label{fig:EBSDKappaB}
\end{subfigure}
\caption{First coefficient $\tilde\kappa_{1,1}(\vec x)$ of the lattice curvature
tensor of the $SO(3)/O$-valued map depicted in Fig.~\ref{fig:EBSDRAW}
(left) and Fig.~\ref{fig:EBSDSmooth} (right).}
\label{fig:EBSDKappa}
\end{figure}
\newpage
\section*{Conclusion and further directions}
We proposed a method for approximating a manifold-valued function using an
embedding approach and a generic approximation operator $I_{\ensuremath{\mathbb{R}}^{d}}$ into the
Euclidean space. Our main result are the Theorems~\ref{satz:wichtig}
and~\ref{satz:wichtigableitung} which give upper bounds on the approximation
error for the function values as well as for the derivatives. The central
requirement of the Theorems is that generic approximation $I_{\ensuremath{\mathbb{R}}^d} f(\vec x)$
has distance less than $\tau$ to the manifold $\ensuremath{\mathcal{M}}$. For the approximation error of
the function values it is only important that $I_{\ensuremath{\mathbb{R}}^d} f(\vec x)$ is within
the reach of the manifold $\mathcal M$ and has no impact on the convergence
rate or the constants. However, for the approximation error of the derivatives
the constant of the upper bound becomes arbitrary large for $I_{\ensuremath{\mathbb{R}}^d} f(\vec
x)$ close to the reach. This stresses the importance of finding embeddings
with a large reach.
\medskip
The basis of our approximation approach is to find a suitable embedding of the
manifold $\ensuremath{\mathcal{M}}$ into $\ensuremath{\mathbb{R}}^d$. For an arbitrary manifold this can be a difficult
challenge. However, for many important manifolds low dimensional embeddings
are well known. A further challenge is the numerical realization of the
projection operator $P_\ensuremath{\mathcal{M}}$ on the manifold needed for our embedding based
approximation method. This leads to a problem of manifold optimization.
\medskip
So far we did not deal with noisy data. The main challenge here is to guaranty
that $I_{\ensuremath{\mathbb{R}}^d} f(\vec x)$ is sufficiently close to the manifold even for noisy
data. Up to this point it is not clear how strongly noise that keeps the data on
the manifold can increase the distance of $I_{\ensuremath{\mathbb{R}}^d} f(\vec x)$ to
the manifold.
\bigskip
\section*{Acknowledgments}
The authors would like to thank Prof. Dr. Philipp Reiter for the nice hint for completing Theorem \ref{thm:C2Teil1}.
Furthermore, we thank the anonymous reviewers for providing helpful comments and suggestions to improve this article.
The second author acknowledges funding by Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - Project-ID 416228727 - SFB 1410.
|
\section{Transformation Selection}
\label{sec:trans_selection}
As discussed in Section~\ref{sec:cal}, the choice of transformations greatly affects the benefit of the result presented in Theorem~\ref{thm:sufficiency}. In particular, if a certain transformation results in a label switch for all examples, then it does not provide any useful confidence information. Thus, the most beneficial transformations are those that separate different examples into different partitions, as measured by the proportion of label switches caused by those transformations.
\paragraph{Choosing a class of transformations.} The first consideration when selecting a transformation is whether to apply it to the input $x$ or to some internal classifier representation, \emph{e.g., } the logits in last layer of a neural network. The benefits of applying input transformations are that they are independent of the classifier and can be chosen based on physical characteristics (\emph{e.g., } a small rotation should not affect an image's class). Applying transformations to the logits is also appealing due to the reduced dimensionality: this results in improved scalability and makes it easier to find useful transformations.%
Another consideration when choosing the transformations is what family to select them from. Input transformations offer a wide range of possibilities, especially in the case of images, e.g., rotation, translation, zoom out. On the other hand, logit transformations do not necessarily have a physical interpretation, so a more natural choice is to add noise selected from a known probability distribution, \emph{e.g., } Gaussian or uniform. In this paper, we explore the space of uniform noise \revised{and Gaussian noise}, as applied to the neural network's logits, in order to benefit from the scalability improvements due to logit transformations.
\paragraph{Parameter selection.} \revised{As discussed above, the noise parameters need to be chosen so as to maximize the benefit of using transformations. One way of measuring the effect of a given transformation is by computing the standard deviations of (non-calibrated) confidences over the entire validation set. Intuitively, if a transformation results in a large standard deviation of confidences, that means this transformation is correlated with the classifier's sensitivity to input perturbations and hence with the confidence in the classifier's correctness.
Therefore, we aim to identify transformations that maximize the standard deviation of confidences over the validation data.}
\revised{To compute the variance in predicted confidences for a specific transformation distribution $\mathcal{D}_T$, we use the sufficient condition presented in Theorem~\ref{thm:sufficiency}. In particular, suppose we are given a validation set $Z = (x_1, y_1), \dots, (x_N, y_N)$ and a sampled set of transformations $T = \{t_1, \dots, t_M\} \sim [\mathcal{D}_T]^M$.
Let $\mathcal{I}_1, \dots, \mathcal{I}_J$ be an index partition.\footnote{Note that, when choosing the noise parameters, we use a single bin for all data. Equations~\eqref{eq:alphah}-\eqref{eq:ph} are written for an arbitrary partition since they are referenced in Section~\ref{sec:implementation} as well.} Then, estimates of $\alpha$, $\beta$, $\gamma$, and calibrated confidence, $p$, can be calculated as}
\begin{align}
\label{eq:alphah}
\hat{\alpha}_{j} =& \frac{ \displaystyle\sum_{n \in \mathcal{I}_j} \sum_{m=1}^M \mathbbm{1}_{\{f(x_n) = y_n\}} \mathbbm{1}_{\{f(t_m(x_n)) = f(x_n)\}} }{ \displaystyle \sum_{n \in \mathcal{I}_j} \sum_{m=1}^M \mathbbm{1}_{\{f(t_m(x_n)) = f(x_n)\}} } \\
\label{eq:betah}
\hat{\beta}_{j} =& \frac{ \displaystyle \sum_{n \in \mathcal{I}_j} \sum_{m=1}^M \mathbbm{1}_{\{f(x_n) = y_n\}} \mathbbm{1}_{\{f(t_m(x_n)) \neq f(x_n)\}} }{ \displaystyle \sum_{n \in \mathcal{I}_j} \sum_{m=1}^M \mathbbm{1}_{\{f(t_m(x_n)) \neq f(x_n)\}} } \\
\hat{\gamma}_{j,n} =& \frac{1}{M}\sum_{m=1}^M \mathbbm{1}_{ \{f(t_m(x_n)=f(x_n) \} }\\
\label{eq:ph}
\hat{p}_{j,n} =& (\hat{\alpha}_{j} - \hat{\beta}_{j}) \hat{\gamma}_{j,n} + \hat{\beta}_{j}.
\end{align}
\revised{
To choose the transformation for each dataset-model combination (please refer to Section~\ref{sec:experiments} for a full description of the datasets), we perform grid search over Gaussian and uniform noise parameters and choose the setting that results in the largest standard deviation of $\hat{p}_{j,n}$. In the Gaussian case, we search over the space [-20, 20] for the mean and (0, 20] for the standard deviation.
For uniform noise, we explore the space [-20, 20], by varying both the minimum noise as well as the range of the noise.}
\revised{
Table \ref{tab:selected_transformation} shows the selected transformation parameters and corresponding values for $\hat{\alpha}$, $\hat{\beta}$, and standard deviation of $\hat{p}_{j,n}$, denoted by $\hat{\sigma}$, as computed over the different datasets and models used in our experiments.
As shown in the table, $\hat{\sigma}$ varies between 0.0395 and 0.0967, which illustrates the challenge of finding an appropriate transformation.
}
\begin{table}[!t]
\centering
\caption{Selected transformation parameters over the different datasets and models. The number of transformation is $M=1000$. We use $U(a,b)$ to denote uniform noise with a range of $[a, b]$ and $G(a,b)$ to denote Gaussian noise with a mean of $a$ and standard deviation of $b$.}
\label{tab:selected_transformation}
\vskip 0.15in
\begin{small}
\begin{sc}
\resizebox{0.5\textwidth}{!}{
\begin{tabular}{llcccc}
\toprule
Dataset & Model& Parameters & $\hat{\alpha}$ & $\hat{\beta}$ & $\hat{\sigma}$\\
\midrule
MNIST & LeNet 5 & U(-2, 4) & 0.9910 & 0.6358 & 0.0167\\
\midrule
\multirow{5}{*}{CIFAR10} & DenseNet 40 & $U(5, 14)$ & 0.9399 & 0.6046 & 0.0456 \\
& LeNet 5 & $U(16, 19)$ & 0.7795 & 0.4826 & 0.0616\\
& ResNet 110 & $U(-6, 3)$ & 0.9567 & 0.6320 & 0.0395 \\
& ResNet 110 SD & $U(-16, -8)$ & 0.9286 & 0.5990 & 0.0500 \\
& WRN 28-10 & $U(-16, -8)$ & 0.9715 & 0.6529 & 0.0332\\
\midrule
\multirow{5}{*}{CIFAR100} & DenseNet 40 & $G(-20, 2)$ & 0.7615 & 0.4082 & 0.0800\\
& LeNet 5 & $G(-5, 1)$ & 0.4817 & 0.2542 & 0.0553\\
& ResNet 110 & $G(-4, 2)$ & 0.7837 & 0.4457 & 0.0801\\
& ResNet 110 SD & $G(-19, 2)$ & 0.7870 & 0.4558 & 0.0782\\
& WRN 28-10 & $G(4, 2)$ & 0.8706 & 0.5038 & 0.0967\\
\midrule
\multirow{4}{*}{ImageNet} & DenseNet 161 & $G(16, 2)$ & 0.8464 & 0.5190 & 0.0806\\
& MobileNet V2 & $G(3, 2)$ & 0.8222 & 0.5013 & 0.0871 \\
& ResNet 152 & $G(0, 2)$ & 0.8551 & 0.5268 & 0.0803 \\
& WRN 101-2 & $G(3, 2)$ & 0.8596 & 0.5236 & 0.0820\\
\bottomrule
\end{tabular}
}
\end{sc}
\end{small}
\end{table}
\section{Conclusion}
\label{sec:conclusion}
\revised{This work proposed a confidence calibration algorithm based on the intuition that we can partition examples based on the neural network's sensitivity to transformations.}
Based on this intuition, we provided a sufficient condition for perfect calibration in terms of ECE.
We performed an extensive experimental comparison and demonstrated that Hoki{} outperforms state-of-the-art approaches in multiple datasets and models, and the benefits are especially pronounced on the challenging ImageNet. For future work, we plan to explore the benefits of combining different transformations, particularly a mix of input and logit transformations.
If those transformations are chosen carefully in order to identify input sensitivity, we expect that more accurate calibration is possible.
\section{Related work}
\label{sec:related_work}
Various approaches have been tried for obtaining accurate confidences, and in this paper, we consider the papers most related to our approach.
We review post-hoc calibration techniques, transformation based calibration, and also calibration with a theoretical guarantee.
\textbf{Post-hoc Calibration.}
Many researchers have proposed calibration methods for a neural network classifier so that the predicted probabilities match the empirical probabilities using a validation set \cite{gupta2020calibration, patel2020multi, platt1999probabilistic, guo2017calibration, kull2019beyond, rahimi2020intra, tran2019calibrating, zhang2020mix}.
\revised{These post-hoc approaches learn a (simpler) mapping function for the calibration without re-training the given (complex) classifier, and different functions for the different types of inputs have been proposed for the mapping.
For example, temperature scaling \citep{guo2017calibration} uses a linear function on logits, Platt scaling \citep{platt1999probabilistic} employs an affine function on logits, and Dirichlet calibration \citep{kull2019beyond} trains an affine function on log logits.}
\textbf{Transformation-based calibration.}
Approaches using input transformations for calibration have also been proposed.
\citet{bahat2020classification} apply semantic preserving transformations such as contrast change, rotation and zoom to augment given inputs and compute confidence using the augmented set of inputs.
\citet{jang2020improving} introduce a lossy label-invariant transformation for calibration. They define a lossy label-invariant transformation and use it to group inputs and apply group-wise temperature scalings.
While our method is also based on transformations, the choice of random transformations allows us to obtain a theoretical guarantee for perfect calibration.
\textbf{Calibration with theoretical guarantee. }
\citet{kumar2019verified} propose a scaling-binning calibrator which combines Platt Scaling and histogram binning, and provide a bound on the calibration error.
\citet{park2020pac} propose a calibrated prediction method which provides per-prediction confidence bound using Clopper-Pearson interval based on histogram binning.
With this method, examples in the same bin are assigned the same confidence, whereas Hoki{} assigns a different confidence to each example, while also aiming to achieve the sufficient condition for perfect calibration within the bin.
\section{Introduction}
\label{sec:introduction}
\begin{wrapfigure}{r}{0.43\textwidth}
\vskip -0.28in
\centering
\includegraphics[width=0.35\textwidth]{imgs/trans.png}
\caption{Random noise transformations may lead to a different number of label prediction changes for different images. Here we apply six different transformations sampled from $N(0, 0.16^2)$.}
\label{fig:ex_trans}
\end{wrapfigure}
Deep neural networks have proven useful in various fields such as image classification \cite{he2016deep, zagoruyko2016wide, tan2019efficientnet}, object detection \cite{redmon2018yolov3, bochkovskiy2020yolov4}, and speaker verification \cite{li2017deep}.
Motivated by these successes, deep neural networks are now being integrated into safety-critical systems such as medical systems~\cite{ronneberger2015u, arcadu2019deep}.
However, as observed by \citet{guo2017calibration}, neural networks are often over-confident on their predictions. This over-confidence can be a critical problem in the safety-critical applications where over-confident neural networks can be wrong with high confidence.
Various calibration techniques have been proposed to alleviate the miscalibration issue.
A common approach is to map uncalibrated logits or confidences to calibrated ones ~\cite{guo2017calibration, gupta2020calibration, jang2020improving, patel2020multi, platt1999probabilistic, kull2019beyond, rahimi2020intra, wenger2020non}. Another option is to train a neural network with a modified loss function to produce calibrated confidence~\cite{tran2019calibrating, kumar2018trainable, seo2019learning}.
\revised{Although existing methods successfully improve the confidence in terms of calibration metrics such as the expected calibration error (ECE)~\cite{naeini2015obtaining}, there is further space for improvement, especially in high-dimensional settings where vast amounts of data are required for good calibration.}
In this work, we propose Hoki{}, a calibration algorithm
that achieves strong empirical performance as compared with state-of-the-art techniques.
The intuition behind Hoki{} is illustrated in Figure~\ref{fig:ex_trans}.
Suppose we are given two images of Persian cats that are initially correctly classified and suppose that we randomly perturb each image a number of times.
\begin{wrapfigure}{l}{0.5\textwidth}
\vskip 0.2in
\centering
\resizebox{0.5\textwidth}{!}{
\includegraphics{imgs/logits_flows.png}
}
\caption{Proposed calibration method. We use logit transformations and corresponding label changes to calibrate the original confidence.}
\label{fig:overall_process}
\end{wrapfigure}
If all perturbed versions of the first image are still correctly classified whereas the second image transformations lead to label switches (\emph{e.g., }{} to Siamese cat), then intuitively we would have higher confidence in the first image's label.
More generally, the idea, as shown in Figure~\ref{fig:overall_process}, is to use logit transformations to test the network's sensitivity and group examples according to the proportion of observed label switches.
Based on this intuition, we first present a sufficient condition for perfect calibration in terms of ECE through leveraging the information from changes after transformations.
An added benefit of this theoretical result is that it also leads to a natural implementation, through minimizing the empirical calibration error (including transformations). The proposed algorithm is also efficient in terms of runtime, especially in the case of logit transformations due to their reduced dimensionality (as compared to the input space dimensionality).
To evaluate Hoki{}, we perform experiments on MNIST, CIFAR10/100, and ImageNet, and we use several standard models per dataset, including LeNet5, DesneNet, ResNet, and ResNet SD.
On these datasets, we compare Hoki{} with multiple state-of-the-art calibration algorithms, namely temperature scaling~\cite{guo2017calibration}, MS-ODIR, Dir-ODIR~\cite{kull2019beyond},
ETS, IRM, IROvA-TS~\cite{zhang2020mix}, and ReCal~\cite{jang2020improving}.
In terms of ECE~\cite{naeini2015obtaining}, Hoki{} outperforms other calibration algorithms on \revised{8} out of 15 benchmarks; we emphasize that Hoki{} achieves
lower ECE than other methods on all ImageNet models. Finally, in terms of learning time, Hoki{} achieves similar performance to temperature scaling, the fastest algorithm we compared with.
The main contributions of this paper are as follows:
\begin{itemize}
\item we propose Hoki, an iterative calibration algorithm using logit transformations;
\item we provide a sufficient condition for perfect calibration;
\item we show that Hoki\ outperforms other calibration algorithms on 8
%
out of 15 benchmarks, and achieves the lowest ECE on all ImageNet models;
\item we demonstrate that Hoki{} is also efficient in terms of learning time, achieving performance comparable to temperature scaling.
\end{itemize}
This paper is structured as follows.
In Section \ref{sec:related_work}, we summarize the related works, and in Section \ref{sec:problem_statement}, we describe the problem statement.
Then, we present the theory of calibration through transformations in Section \ref{sec:cal}.
In Section \ref{sec:trans_selection}, we demonstrate the transformation selection process, and we illustrate the calibration algorithm using transformations in Section \ref{sec:implementation}.
In Section \ref{sec:experiments}, we show the experimental results, and we conclude this paper in Section \ref{sec:conclusion}.
\section{Calibration using Transformations}
\label{sec:cal}
This section provides the intuition and theory of using transformations for the purpose of calibration.
We begin by providing high level intuition, followed by a sufficient condition for perfect calibration in expectation, which leads to a natural implementation as well.
\paragraph{High-level intuition.} Suppose that $f$ and $g$ form a classifier-calibrator pair. If we take a correctly classified image of a cat, for example, we would expect that the classification confidence would drop as we apply random transformations to the image (e.g., add noise, zoom out). Conversely, if the confidence does not decrease, we would conclude that $f$ and $g$ are not properly calibrated.
More generally, the goal of applying transformations is to group examples in bins of similar confidence. In particular, if a certain set of examples exhibits similar transformation patterns (e.g., label switching, misclassification), then the calibrator should learn to assign such examples a similar confidence value. Of course, this approach would only work for a good choice for transformations -- we discuss a number of options in Section~\ref{sec:trans_selection}.
\paragraph{Sufficiency for perfect calibration.}
We now investigate calibrator properties that ensure perfect calibration. Suppose we are given a class of transformations $\mathcal{T} = \{t: \mathcal{X} \to \mathcal{X}\}$, e.g., functions that add random noise, and a corresponding probability distribution $\mathcal{D}_T$ over $\mathcal{T}$. Then, for each example $(x,y)$, we can apply a number of transformations and observe how many transformations lead to a label switch. Specifically, the following result is key to achieving perfect calibration.
\begin{theorem}[Sufficiency for Perfect Calibration]
\label{thm:sufficiency}
Let $\mathcal{P}_1, \dots, \mathcal{P}_J$ be a confidence bin partition. A calibrator $g \in \mathcal{G}$ is perfectly calibrated, i.e., $CE(g) = 0$, if it satisfies, $\forall j \in \{1, \dots, J\}$,
\begin{align*}
E_{\mathcal{D}}\left[ g(X) \mid g(X) \in \mathcal{P}_{j} \right] = \alpha_{j} \gamma_j + \beta_{j} (1 - \gamma_j) &\\
\end{align*}
where
\begin{align*}
\alpha_{j} =& P_{\mathcal{D} \times \mathcal{D}_T}[f(X) = Y \mid f(T(X)) = f(X), g(X) \in \mathcal{P}_j ] \\
\beta_{j} =& P_{\mathcal{D} \times \mathcal{D}_T}[f(X) = Y \mid f(T(X)) \neq f(X), g(X) \in \mathcal{P}_j ] \\
\gamma_j =& P_{\mathcal{D} \times \mathcal{D}_T}\left[f(T(X)) = f(X) \mid g(X) \in \mathcal{P}_{j} \right] &\\
\end{align*}
\end{theorem}
\begin{proof}
Proof provided in the supplementary material.
\end{proof}
Intuitively, Theorem~\ref{thm:sufficiency} states that the label switching that we observe (in each bin) due to added transformations must be consistent with the confidence and accuracy in that bin. In particular, the average confidence in the bin must be equal to the weighted sum of accuracies over the two groups of examples: 1) examples whose label is changed by some transformation; 2) examples whose label is not changed due to transformations. The benefit of Theorem~\ref{thm:sufficiency} is that it leads to a natural implementation by estimating all probabilities given a validation set (as discussed in Section~\ref{sec:implementation}). In the Supplementary Material, we also provide a theoretical bound (Theorem 2) on the generalization ECE (given new data) of Hoki{} in a probably approximately correct sense.
\section{Problem Statement}
\label{sec:problem_statement}
Let $\mathcal{X}$ be a feature space, $\mathcal{Y} = \{1, \dots, C\}$ be a set of labels, and $\mathcal{D}$ be a distribution over $\mathcal{X} \times \mathcal{Y}$.
We are given a classifier $f: \mathcal{X} \to \mathcal{Y}$ and a corresponding calibrator $g: \mathcal{X} \to [0,1]$ such that, for a given example $x$, $(f(x), g(x)) = (\hat{y},\hat{p})$ is the label prediction $\hat{y}$ with a corresponding confidence $\hat{p}$. In what follows, we say that the sets $\mathcal{P}_1, \dots, \mathcal{P}_J$ form a bin partition of the confidence space, $[0,1]$, if $\cup_{i=1}^J \mathcal{P}_i = [0,1]$ and $\forall i \neq j, \; \mathcal{P}_i \cap \mathcal{P}_j = \emptyset$. Furthermore, given a dataset $Z = \{(x_1, y_1), \dots, (x_N, y_N)\}$, we say $g$ induces an index partition $\mathcal{I}_1, \dots, \mathcal{I}_J$ of $\{1,\dots,N\}$ such that $g(x_n) \in \mathcal{P}_j \iff n \in \mathcal{I}_j, \forall (x_n,y_n) \in Z$.
Before formally stating the problem considered in this work, we define calibration error (CE) and expected calibration error (ECE) \cite{naeini2015obtaining}.
\begin{definition}[Calibration Error (CE)]
\label{def:ce}
For any calibrator $g$ and confidence partitions $\mathcal{P}_1, \dots, \mathcal{P}_J$, the calibration error (CE) is defined as
\begin{equation*}
CE(g) = \sum_{j = 1}^{J} w_j \left| e_j \right|,
\end{equation*}
where
\begin{align*}
& \; e_{j} := P_{\mathcal{D}}\left[Y = f(X)\; \middle| \; g(X) \in \mathcal{P}_j \right] - E_{\mathcal{D}}\left[g(X) \mid g(X) \in \mathcal{P}_j \right]\\
& \; w_{j} := P_{\mathcal{D}}\left[g(X) \in \mathcal{P}_j\right].
\end{align*}
\end{definition}
Intuitively, the CE of a classifier-calibrator pair in a given partition is the expected difference between the classifier's accuracy and the calibrator's confidence. To get the CE over the entire space, we sum up all the individual partition CEs, weighted by the probability mass of each partition (i.e., the probability of an example falling in that partition).
\begin{definition}[Expected Calibration Error (ECE)]
\label{def:ece}
For any calibrator $g$, confidence partitions $\mathcal{P}_1, \dots, \mathcal{P}_J$, sampled dataset $Z \in (\mathcal{X} \times \mathcal{Y})^N$, and induced index partition $\{ \mathcal{I}_1, \dots , \mathcal{I}_J \}$,
we define the expected calibration error (ECE) as
\begin{equation*}
ECE(g) = \sum_{j = 1}^{J} \hat{w}_j \left| \hat{e}_j \right|,
\qquad
\text{where}\qquad
\hat{e}_j := \sum_{n \in \mathcal{I}_j} \frac{\mathbbm{1}_{\{y_n = \hat{y}_n\}} - g(x_n) }{| \mathcal{I}_j| } \quad \mbox{and} \quad \hat{w}_j := \frac{|\mathcal{I}_j|}{N}.
\end{equation*}
\end{definition}
Thus, the ECE is the sampled version of the CE. Note that Definition~\ref{def:ece} is equivalent to the standard ECE definition, as used in prior work~\cite{guo2017calibration}. We are now ready to state the problem addressed in this work, namely find a calibrator $\hat{g}$ that minimizes the ECE over a validation set.
\begin{problem}
\label{ps:main_problem}
Let $\mathcal{G} = \{g : \mathcal{X} \to [0,1]\}$ be the set of all calibrators. We aim to find $\hat{g} \in \mathcal{G}$ that minimizes the expected calibration error,
\begin{equation*}
\hat{g} = \arg \min_{g \in \mathcal{G}} ECE(g). \\
\end{equation*}
\end{problem}
\section{Implementation}
\label{sec:implementation}
Based on the theory described in Section \ref{sec:cal}, we propose Hoki{}, an iterative algorithm for confidence calibration.
Hoki{} operates differently during design time and runtime.
During design time, Hoki{} samples random transformations and learns the $\hat{\alpha}_j$ and $\hat{\beta}_j$ parameters for each bin. These parameters are then used at runtime, on test data, to estimate the confidence for new examples.
\paragraph{Design Time Algorithm.}
The design time algorithm is described in Algorithm~\ref{alg:design_time}. %
The high-level idea of the algorithm is to achieve the sufficient condition outlined in Theorem~\ref{thm:sufficiency} on the validation set. In particular, we first sample transformations $\{t_1, \dots, t_M\}$ from $\mathcal{T}$ and observe the corresponding predictions $\bar{y}_{n,m}$. Based on $\bar{y}_{n,m}$, we compute the fraction of transformed examples that have the same label as the original image, $\gamma_n$. We emphasize that, as a special case of Theorem~\ref{thm:sufficiency}, $\gamma_n$ is computed separately for each example, as opposed to averaged over the entire partition. This modification ensures that the data is spread across multiple partitions, while still satisfying the sufficient condition for perfect calibration.
After the initialization step, Hoki{} recursively estimates the parameters $\alpha_j$ and $\beta_j$ and computes the calibrated confidence using those two values based on Theorem \ref{thm:sufficiency}. Since the original partitioning may change after reestimating the parameters, we repeat this process until there is no change in the data partitioning (or a maximum number of iterations is reached). Note that for the corner case where the transformations result in empty sets (i.e., either all labels change or all remain the same), we set $\hat{\alpha}_j^k$, $\hat{\beta}_j^k$ to the bin accuracy. Ultimately, at design time, Hoki{} returns the set of calibration pairs for all partitions for all iterations learned in the design time algorithm as a set $\mathcal{H}$, the set of transformations $\hat{\mathcal{T}}$, and the validation data accuracy, $p$.
\begin{algorithm}[tb]
\caption{Design Time Algorithm}
\label{alg:design_time}
\begin{algorithmic}
\STATE {\bfseries Input:} validation set $Z = (x_1, y_1), \dots, (x_N, y_N)$, transformation set $\mathcal{T}$, number of transformations $M$, classifier $f$, confidence space partition $\mathcal{P}_1, \dots, \mathcal{P}_J$, maximum number of iterations $K$
\STATE
\STATE Sample transformations $\hat{\mathcal{T}} = \{t_1, \dots, t_M\}$ from $\mathcal{T}$
\STATE $\hat{p} = \frac{1}{N} \sum_{n = 1}^N \mathbbm{1}_{\{y_n = f(x_n)\}}$
\FOR{$n=1$ {\bfseries to} $N$}
%
%
%
\STATE $\hat{\gamma}_n = \frac{1}{M} \sum_{m = 1}^M \mathbbm{1}_{\{f(t_m(x_n)) = f(x_n)\}}$
%
\STATE $p_n = \hat{p}$
\ENDFOR
\STATE
%
%
%
\FOR{$k=1$ {\bfseries to} $K$}
\STATE Compute sets $\mathcal{I}_1^k, \dots, \mathcal{I}_J^k$ s. t. $n \in \mathcal{I}_j^k \iff p_n \in \mathcal{P}_j$
\IF{$k > 1 \wedge \mathcal{I}_1^k = \mathcal{I}_1^{k-1} \wedge \dots \wedge \mathcal{I}_J^k = \mathcal{I}_J^{k-1}$}
\STATE \textbf{break}
\ENDIF
\FOR{$j=1$ {\bfseries to} $J$}
\STATE $c =\sum_{n \in \mathcal{I}_j^k} \sum_{m=1}^M \mathbbm{1}_{\{f(x_n) = f(t_m(x_n))\}} $
%
%
\IF {$c = 0 \vee c = M|\mathcal{I}_j^k|$}
\STATE $\hat{\alpha}^k_j = \hat{\beta}^k_j = \frac{1}{|\mathcal{I}_j^k|} \sum_{n \in \mathcal{I}_j^k} \mathbbm{1}_{\{y_n = f(x_n)\}}$
\ELSE
\STATE Compute $\hat{\alpha}_j^k$ according to Equation~\eqref{eq:alphah}, using $\mathcal{I}_j^k$
\STATE Compute $\hat{\beta}_j^k$ according to Equation~\eqref{eq:betah}, using $\mathcal{I}_j^k$
%
%
\ENDIF
\STATE $p_n = \left(\hat{\alpha}^k_j - \hat{\beta}^k_j \right) \hat{\gamma}_n + \hat{\beta}^k_j, \; \forall n \in \mathcal{I}_j^k$
\ENDFOR
\STATE $K^* = k$
\ENDFOR
\STATE $\mathcal{H} = \{ (\hat{\alpha}_j^k, \hat{\beta}_j^k) \; | \; 1 \leq j \leq J, \; 1 \leq k \leq K^*\}$
\STATE return $\mathcal{H}$, $\hat{p}$, $\hat{\mathcal{T}}$
\end{algorithmic}
\end{algorithm}
\paragraph{Runtime Algorithm.}
The runtime algorithm is described in Algorithm~\ref{alg:run_time}. %
Once the parameters for each step, $\hat{\alpha}_j^k$ and $\hat{\beta}_j^k$, are learned in Algorithm~\ref{alg:design_time}, Hoki{} can calibrate the confidence for a new input $x$ using the calibration parameter pairs in $\mathcal{H}$.
Hoki{} first observes the original prediction $f(x)$ and the transformed data prediction $f(t_m(x))$ for all transformations $\{t_1, t_2, \dots, t_M\}$.
Hoki{} computes $\gamma$ based on these values and iteratively updates the confidence according to the calibration parameters learned in Algorithm~\ref{alg:design_time}.
\begin{algorithm}[tb]
\caption{Runtime Algorithm}
\label{alg:run_time}
\begin{algorithmic}
\STATE {\bfseries Input:} test sample $x \in \mathcal{X}$, original classifier $f$, outputs of Algorithm~\ref{alg:design_time}: $\mathcal{H}$, $\hat{p}$, $\hat{\mathcal{T}}$.
\STATE $\gamma= \frac{1}{M} \sum_{m = 1}^M \mathbbm{1}_{\{f(x)=f(t_m(x))\}}, \; t_m \in \hat{\mathcal{T}}$
\STATE $p = \hat{p}$
\FOR{$k=1$ {\bfseries to} $\frac{|\mathcal{H}|}{J}$}
\STATE Identify partition index $j'$ for $x$ such that $p \in \mathcal{P}_{j'}$
\STATE Identify calibration parameters $(\hat{\alpha}^k_{j'}, \hat{\beta}^k_{j'}) \in \mathcal{H}$
\STATE $p=\left(\alpha^k_{j'} - \beta^k_{j'} \right) \gamma + \beta^k_{j'}$
\ENDFOR
\STATE return $p$
\end{algorithmic}
\end{algorithm}
\paragraph{Limitations.}
\label{sec:limitations}
There are two main limitations of our approach. First of all, since Hoki{} uses transformations for calibration, choosing the appropriate transformations has a significant effect on performance.
To address this issue, we propose a transformation selection based on the maximization of confidence variance as described in Section \ref{sec:trans_selection}. The second limitation is the utility of the ECE metric itself -- it is possible to minimize the ECE by outputting the network accuracy as confidence for all examples, which defeats the purpose of calibration. We argue that by maximizing the variance of our algorithm, we ensure that the data is spread across multiple bins, as demonstrated in Section~\ref{sec:experiments}. Thus, one can use the ECE in multiple ways, e.g., in autonomous systems by making an informed decision through taking into account the calibrated confidence in a new example's predicted label.
\section{Experiments}
\label{sec:experiments}
We compare Hoki{} with state-of-the-art calibration algorithms using several standard datasets and models.
For each model and dataset, we compute the ECE for the uncalibrated model and the calibrated confidence by the algorithms.
The experimental setup, baseline algorithms and the evaluation metrics are explained in the following subsections.
\subsection{Experimental Setup}
This subsection provides the details about the datasets, models, baseline algorithms, and the evaluation metrics.
\textbf{Datasets and Models.}
\revised{We perform experiments on MNIST \cite{lecun1998gradient}, CIFAR 10/100 \cite{krizhevsky2009learning}, and ImageNet \cite{deng2009imagenet}}.
We use the following models for each dataset.
For MNIST, we use one model, LeNet5 \cite{lecun1998gradient}.
For CIFAR 10/100, we use five different models, DenseNet 40 \cite{huang2017densely}, LeNet5 \cite{lecun1998gradient}, ResNet110 \cite{he2016deep}, ResNet110 SD \cite{huang2016deep}, and WRN-28-10 \cite{zagoruyko2016wide}.
For ImageNet, we use four models, DenseNet161 \citep{huang2017densely}, MobileNetV2 \cite{sandler2018mobilenetv2}, ResNet152 \citep{he2016deep}, and WRN-101-2 \cite{zagoruyko2016wide}.
We implement LeNet5, ResNet110 SD, and obtain code for DenseNet40\footnote{https://github.com/andreasveit/densenet-pytorch under BSD 3-Clause License}, ResNet110\footnote{\label{fn:resnet110github}https://github.com/bearpaw/pytorch-classification, under MIT License} and WRN 28-10\footnotemark[\getrefnumber{fn:resnet110github}] from the corresponding github repositories.
We also obtained the pre-trained model for all models on ImageNet from PyTorch.\footnote{https://pytorch.org/docs/stable/torchvision/models.html, under BSD License}
\textbf{Baselines.}
We compare Hoki{} with several state-of-the-art calibration algorithms, namely temperature scaling, vector scaling \cite{guo2017calibration}, Dir-ODIR, MS-ODIR \cite{kull2019beyond}, \revised{ETS, IRM, IROvA-TS \cite{zhang2020mix}}, and ReCal \cite{jang2020improving}.
They calibrate confidence by learning a mapping function for uncalibrated logits or confidences.
We obtain other calibration algorithms from their papers except for temperature scaling and vector scaling which we obtain from \citet{kull2019beyond}.
For ReCal, the authors provide three different setups for their algorithm, and we choose ('zoom-out', 0.1, 0.9, 20) because it shows the best results on ImageNet.
\textbf{Evaluation Metric.}
As described in Problem Statement~\ref{ps:main_problem}, we evaluate all algorithms based on ECE (with $J=15$ bins of equal width~\cite{guo2017calibration, nixon2019measuring}), as defined in Definition \ref{def:ece}.
Additionally, we calculate the learning time during design time to investigate each algoritm's practical utility.
If a calibration algorithm is too slow on real datasets, it may not be appropriate to use the algorithm in practice.
\subsection{Results}
The experimental results have two parts.
The first part is a comparison on calibration performance in terms of ECE, and the second part is an analysis on time efficiency during design time.
\paragraph{ECE Results.}
\begin{table*}[bt]
\caption{ECE for the different calibration algorithms on different datasets and models. The number with the bold face and the underline denote the best and the second best result, respectively.}
\label{tab:exp_res_ECE}
\begin{center}
\resizebox{\textwidth}{!}{
\begin{sc}
\begin{tabular}{llcccccccccccc}
\toprule
Dataset & Model & \thead{Val Acc.\\(\%)} & \thead{Test Acc.\\(\%)} & Uncal. & TempS & VecS & \thead{MS-\\ODIR} & \thead{Dir-\\ODIR} & ETS & IRM & \thead{IROvA-\\TS} & ReCal & Hoki\\
\midrule
MNIST & LeNet5 & 98.85 & 98.81 & 0.0076 & 0.0018 & \underline{0.0015} & 0.0024 & 0.0022 & 0.0019 & 0.0019 & 0.0033 & 0.0021 & \textbf{0.0008}\\
\midrule
\multirow{5}{*}{CIFAR10} & DenseNet 40 & 91.92 & 91.75 & 0.0520 & 0.0070 & \underline{0.0044} & 0.0052 & \textbf{0.0039} & 0.0069 & 0.0095 & 0.0107 & 0.0101 & 0.0057 \\
& LeNet5 & 72.00 & 72.77 & 0.0182 & 0.0120 & \textbf{0.0092} & 0.0141 & \underline{0.0105} & 0.0115 & 0.0167 & 0.0229 & 0.0118 & 0.0110 \\
& ResNet110 & 94.12 & 93.10 & 0.0456 & 0.0088 & 0.0094 & 0.0088 & 0.0084 & \textbf{0.0066} & 0.0103 & 0.0133 & 0.0090 & \underline{0.0071}\\
& ResNet110 SD & 90.28 & 90.38 & 0.0538 & 0.0114 & \underline{0.0086} & 0.0102 & 0.0094 & 0.0112 & 0.0113 & 0.0156 & 0.0120 & \textbf{0.0044}\\
& WRN 28-10 & 96.06 & 95.94 & 0.0251 & 0.0097 & 0.0096 & 0.0092 & 0.0094 & 0.0157 & \underline{0.0049} & 0.0088 & 0.0091 & \textbf{0.0026} \\
\midrule
\multirow{5}{*}{CIFAR100} & DenseNet 40 & 68.82 & 68.16 & 0.1728 & 0.0154 & 0.0266 & 0.0296 & 0.0189 & 0.0136 & \underline{0.0135} & 0.0377 & 0.0154 & \textbf{0.0073}\\
& LeNet5 & 37.82 & 37.66 & \textbf{0.0100} & 0.0211 & 0.0155 & 0.0131 & 0.0142 & \underline{0.0120} & 0.0125 & 0.0363 & 0.0192 & 0.0123\\
& ResNet 110 & 70.60 & 69.52 & 0.1422 &\textbf{0.0091} & 0.0300 & 0.0345 & 0.0231 & 0.0155 & 0.0202 & 0.0457 & \underline{0.0121} & 0.0127\\
& ResNet 110 SD & 70.62 & 70.10 & 0.1229 & \underline{0.0089} & 0.0358 & 0.0355 & 0.0207 & \textbf{0.0086} & 0.0142 & 0.0425 & 0.0100 & 0.0098\\
& WRN 28-10 & 79.62 & 79.90 & 0.0534 & 0.0437 & 0.0452 & 0.0355 & 0.0346 & 0.0370 & \textbf{0.0108} & 0.0336 & 0.0373 & \underline{0.0112}\\
\midrule
\multirow{2}{*}{ImageNet} & DenseNet 161 & 76.83 & 77.45 & 0.0564 & 0.0199 & 0.0233 & 0.0368 & 0.0477 & 0.0100 & \underline{0.0090} & 0.0487 & 0.0133 & \textbf{0.0043}\\
& MobileNet V2 & 71.69 & 72.01 & 0.0274 & 0.0164 & 0.0153 & 0.0212 & 0.0269 & 0.0087 & \underline{0.0075} & 0.0477 & 0.0153 & \textbf{0.0011}\\
& ResNet 152 & 77.93 & 78.69 & 0.0491 & 0.0201 & 0.0207 & 0.0347 & 0.0397 & 0.0112 & \underline{0.0080} & 0.0457 & 0.0139 & \textbf{0.0052}\\
& WRN 101-2 & 78.67 & 79.15 & 0.0524 & 0.0307 & 0.0330 & 0.0418 & 0.0279 & 0.0165 & \underline{0.0086} & 0.0426 & 0.0258 & \textbf{0.0067}\\
\bottomrule
\end{tabular}
\end{sc}
}
\end{center}
\end{table*}
Table~\ref{tab:exp_res_ECE} displays ECE values for each algorithm, along with each model's validation set and test set accuracy (the Supplementary Material provides an extensive evaluation where we also vary the number of bins in the ECE evaluation, in order to test each algorithm's robustness to more fine-grained bins).
As discussed in Section~\ref{sec:trans_selection}, Hoki{} uses the transformations shown in Table~\ref{tab:selected_transformation}. As shown in Table~\ref{tab:exp_res_ECE}, Hoki{} achieves the lowest ECE on \revised{8}
out of 15 benchmarks. The benefit of using transformations is especially pronounced in the large-dimensional ImageNet dataset where Hoki{} consistently achieves the lowest ECE on all models.
Interestingly, the ECE produced by Hoki{} closely tracks the difference between the validation and test set accuracy. In some sense, one cannot hope to do better than this difference as it reflects the variance within each dataset. Thus, the benchmarks where Hoki{} does not achieve the best performance are settings with large differences in generalization accuracy. For example, in the case of ResNet110 on CIFAR10 and CIFAR100, the gaps are 1.02 and 1.08 percentage points, respectively.
Another reason for our strong performance on ImageNet is the dataset size (there are 25,000 images in the ImageNet validation set compared to 10,000 images in MNIST and 5,000 images in CIFAR10/100).
A larger validation set means that each partition is likely to have more samples, which in turn results in more accurate estimation of $\alpha$, $\beta$, and $\gamma$. %
\begin{figure}
\centering
\includegraphics[width=\textwidth]{imgs/ece_var_imgnet.png}
\caption{$\log$ ECE vs. Variance on the various ImageNet models.}
\label{fig:log_ece_var}
\end{figure}
\paragraph{ECE Variance.} For further evaluation, we also explore the variance in confidences produced by each algorithm. As noted in Section~\ref{sec:trans_selection}, a larger variance is preferred because it provides some indication that examples with low true confidence are indeed separated from those with high confidence. Figure~\ref{fig:log_ece_var} provides a plot of the ECE variance vs. the ECE \revised{(in log scale)} for all algorithms on the ImageNet models (plots for the other benchmarks are provided in the Supplementary Material). As shown in the figure, Hoki{} has comparable variance to other algorithms that also achieve low ECE on ImageNet. Overall, there appears to be a trade-off between achieving \revised{low}
ECE and high variance -- we leave exploring this phenomenon for future work.
\paragraph{Time Efficiency.}
Table \ref{tab:learng_time} displays the learning time of each calibration algorithm during design time.
The main reason for the proposed method's good scalability is that we apply transformations to the logits -- thus, we avoid the need to perform the input transformations that are needed in ReCal, for example.
Furthermore, Hoki{} is comparable to temperature scaling, which is a fairly simple approach in the sense that it only needs to learn one parameter.
In summary, Hoki{} not only achieves low ECE on most benchmarks but is also fast to execute.
\begin{table}[!hb]
\centering
\caption{Learning time (sec). The number with the bold face and the underline are the best and the second best result, respectively.}
\label{tab:learng_time}
\vskip 0.15in
\resizebox{1.0\textwidth}{!}{
%
\begin{sc}
\begin{tabular}{llccccccccc}
\toprule
Dataset & Model & TempS & VecS & MS-ODIR & Dir-ODIR & ETS & IRM & IROvA-TS & ReCal & Hoki\\
\midrule
MNIST & LeNet5 & 0.16 & 43.31 & 112.04 & 207.88 & 0.31 & \textbf{0.02} & \underline{0.12} & 57.67 & 0.53\\
\midrule
\multirow{5}{*}{CIFAR10}& DenseNet40 & 0.08 & 31.33 & 222.93 & 92.34 & \underline{0.04} & \textbf{0.01} & 0.06 & 84.04 & 0.34\\
& LeNet5 & 0.05 & 11.86 & 79.62 & 74.33 & \underline{0.04} & \textbf{0.01} & 0.06 & 110.79 & 0.32\\
& ResNet110 & 0.07 & 27.17 & 193.04 & 87.73 & \underline{0.05} & \textbf{0.01} & 0.07 & 38.85 & 0.28\\
& ResNet110 SD & 0.07 & 21.39 & 189.27 & 93.17 & \underline{0.04} & \textbf{0.01} & 0.09 & 58.74 & 0.28\\
& WRN 28-10 & 0.05 & 22.71 & 123.46 & 92.80 & \underline{0.04} & \textbf{0.01} & 0.08 & 49.62 & 0.39\\
\midrule
\multirow{5}{*}{CIFAR100} & DenseNet40 & \underline{0.51} & 23.17 & 1211.68 & 626.00 & 0.54 & \textbf{0.11} & 0.57 & 136.23 & 1.09\\
& LeNet5 & 0.42 & 24.17 & 459.59 & 236.87 & \underline{0.29} & \textbf{0.12} & 0.45 & 97.77 & 0.99\\
& ResNet110 & \underline{0.30} & 25.49 & 1459.71 & 510.12& 0.59 & \textbf{0.12} & 0.57 & 97.29 & 1.02\\
& ResNet110 SD & \underline{0.24} & 25.51 & 1696.10 & 495.23 & 0.61 & \textbf{0.12} & 0.63 & 604.12 & 1.01\\
& WRN 28-10 & \underline{0.30} & 26.71 & 1110.11 & 611.52 & 0.64 & \textbf{0.11} & 0.50 & 125.84 & 1.03\\
\midrule
\multirow{4}{*}{ImageNet} & DenseNet161 & \underline{18.79} & 179.61 & 13901.38 & 6891.19 & 29.34 & \textbf{8.76} & 31.16 & 50730.17 & 37.68\\
& MobileNet V2 & \underline{18.74} & 423.48 & 3899.07 & 12695.64& 27.56 & \textbf{8.80} & 26.99 & 3139.60 & 32.97\\
& ResNet152 & \underline{18.79} & 169.72 & 12401.58 & 5402.85 & 29.28 & \textbf{9.01} & 31.14 & 71254.34 & 33.78\\
& WRN 101-2 & \underline{18.73} & 182.39 & 16989.40 & 11378.40 & 29.17 & \textbf{8.67} & 31.46 & 31545.77 & 34.26\\
\bottomrule
\end{tabular}
\end{sc}
%
}
\end{table}
\section{Proof of Theorem 1}
We begin by observing for any $g$ and $\mathcal{P}_j \in \{ \mathcal{P}_1, \dots, \mathcal{P}_j\}$, the law of total probability states
\begin{align*}
P\left[Y = f(X) \mid g(X)\in \mathcal{P}_j \right] = &\\
%
= & \;P\left[Y = f(X) \mid f(T(X)) = f(X), g(X) \in \mathcal{P}_j \right]
P\left[f(T(X)) = f(X) \mid g(X) \in \mathcal{P}_j \right] \\
&+ P\left[Y = f(X) \mid f(T(X)) \neq f(X), g(X) \in \mathcal{P}_j \right]
P\left[f(T(X)) \neq f(X) \mid g(X) \in \mathcal{P}_j \right] \\
%
= & \;\alpha_{j} \gamma_j
+ \beta_j P\left[f(T(X)) \neq f(X) \mid g(X) \in \mathcal{P}_j \right] \\
%
= & \;\alpha_j \gamma_j + \beta_j (1 - \gamma_j) \\
\end{align*}
Then, from Definition~\ref{def:ece}, %
\begin{align*}
CE(g) = \sum_{j = 1}^{J} w_j \left|e_j \right|
= & \sum_{j = 1}^{J} w_j \Big| P\left[Y = f(X) \mid g(X)\in \mathcal{P}_j \right] - E\left[ g(X) \mid g(X) \in \mathcal{P}_j \right] \Big|\\
= & \sum_{j = 1}^{J} w_j \Big| \alpha_j \gamma_j + \beta_j (1 - \gamma_j) - E\left[ g(X) \mid g(X) \in \mathcal{P}_j \right] \Big| \\
\end{align*}
Thus, the following is a sufficient property for $CE(g) = 0$:
\begin{align*}
\forall j \in \{1, \dots, J\}, \; E\left[ g(X) \mid g(X) \in \mathcal{P}_j \right] = \alpha_j \gamma_j + \beta_j (1 - \gamma_j)
\end{align*}
\section{Generalization bounds on the ECE}
\label{sec:app_thm2}
This section presents a bound on the generalization ECE, given a new dataset, in a probably approximately correct (PAC) sense. Theorem 2 states that if a calibrator $g$ achieves a low ECE on a test set, $Z$, then the expected calibration error of $g$ can be bounded, in a PAC sense.
\begin{customthm}{2}[Bounded Calibration Error]
\label{thm:generalization}
Suppose a calibrator $g$ was evaluated on a test set $Z = \{(x_1, y_1), \dots, (x_N, y_N)\}$, achieving $ECE_Z(g)$.
For any $\delta$, the CE is bounded,~i.e.,
\begin{equation*}
P\left[CE(g) \leq \epsilon \right] \geq 1 - \delta,
\end{equation*}
when
\begin{align*}
& \epsilon = ECE_Z(g) + \frac{J\sqrt{2}}{\sqrt{N}} \sqrt{ 2 \ln(2) - \ln(\delta)}
\end{align*}
\end{customthm}
\begin{proof}
\begin{align*}
P\left[CE(g) \geq \epsilon \right] = &\\
= & \; P\left[ \sum_{j =1}^J \left| e_j \right| w_j \geq \epsilon \right] \\
= & \; P\left[ \sum_{j =1}^J \left| e_j \right| (w_j - \hat{w}_j + \hat{w}_j) \geq \epsilon \right] \\
\leq & \; P\left[ \sum_{j =1}^J \left| e_j \right| \left|w_j - \hat{w}_j \right| + \left| e_j \right| \hat{w}_j \geq \epsilon \right] \\
\leq & \; P\left[ \sum_{j =1}^J \left|w_j - \hat{w}_j \right| + \left| e_j \right| \hat{w}_j \geq \epsilon \right] \\
\leq & \; P\left[ \sum_{j =1}^J \left|w_j - \hat{w}_j \right| + \left| e_j - \hat{e}_j \right| \hat{w}_j + \left| \hat{e}_j \right| \hat{w}_j \geq \epsilon \right] \\
= & \; P\left[ \sum_{j =1}^J \left|w_j - \hat{w}_j \right| + \left| e_j - \hat{e}_j \right|\hat{w}_j \geq \epsilon - ECE_Z(g) \right] \\
\leq & \; \max_j P\left[ \left|w_j - \hat{w}_j \right| + \left| e_j - \hat{e}_j \right| \hat{w}_j \geq \frac{\epsilon - ECE_Z(g)}{J} \right] \\
\leq & \; \max_j P\left[ \left|w_j - \hat{w}_j \right| \geq \frac{\epsilon - ECE_Z(g)}{2J} \right] + P\left[ \left| e_j - \hat{e}_j \right| \geq \frac{\epsilon - ECE_Z(g)}{2 J \hat{w}_j} \right] \\
\leq & \; \max_j 2 \exp\left\{ - 2 N \left(\frac{\epsilon - ECE_Z(g)}{2J} \right)^2 \right\} + 2 \exp\left\{ - 2 N \hat{w}_j \left( \frac{\epsilon - ECE_Z(g)}{2 J \hat{w}_j} \right)^2 \right\} \\
= & \; \max_j 2 \exp\left\{ - 2 N \left(\frac{\epsilon - ECE_Z(g)}{2 J}\right)^2 \right\} + 2 \exp\left\{ - 2 N \frac{1}{\hat{w}_j} \left(\frac{\epsilon - ECE_Z(g)}{2 J} \right)^2 \right\} \\
\leq & \; 2 \exp\left\{ - 2 N \left(\frac{\epsilon - ECE_Z(g)}{2J}\right)^2 \right\} + 2 \exp\left\{ - 2 N \left(\frac{\epsilon - ECE_Z(g)}{2J} \right)^2 \right\} \\
= & \; 4 \exp\left\{ - \frac{ N (\epsilon - ECE_Z(g))^2}{2 J^2} \right\} \\
\end{align*}
We complete the proof by observing
\begin{align*}
& \; 4 \exp\left\{ - \frac{N (\epsilon - ECE_Z(g))^2}{2 J^2} \right\}\leq \delta
\iff \; \epsilon \geq ECE_Z(g) + \frac{J\sqrt{2}}{\sqrt{N}} \sqrt{\left( 2 \ln(2) - \ln(\delta) \right) }
\end{align*}
\end{proof}
%
%
%
%
%
%
\section{Additional Experiments}
In this section, we present additional experimental results.
We show more plots for ECE variance, comparisons using ECE with different number of bins, and Hoki's ECE with the different initialization.
\subsection{ECE Variance}
In addition to the ImageNet result in the main paper, we show the same ECE variance results for other benchmarks on MNIST (Figure~\ref{fig:ece_var_mnist}) and CIFAR10/100 (Figure~\ref{fig:ece_var_cifar10} and \ref{fig:ece_var_cifar100}).
The range of variance are different depending on dataset, but the widths of the range are equal.
Similar to ImageNet case in the main text, Hoki has comparable variance for MNIST (Figure~\ref{fig:ece_var_mnist}) and CIFAR10 (Figure~\ref{fig:ece_var_cifar10}), but with the smaller ECE.
As shown in Figure~\ref{fig:ece_var_cifar100}, Hoki has a similar pattern with ImageNet case on CIFAR100, \emph{i.e., }{}it has comparable variance with better ECE.
Note that uncalibrated classifier is not shown for DenseNet40~(Figure~\ref{fig:ece_var_cifar100_densenet40}), ResNet110~(Figure~\ref{fig:ece_var_cifar100_resnet110}), and ResNet110 SD~(Figure~\ref{fig:ece_var_cifar100_resnet110sd}), because the uncalibrated ECEs are high compared to other algorithms as shown in Table~\ref{tab:exp_res_ECE} and the variances are low.
\begin{figure}[!hbt]
\centering
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_MNIST_lenet5.png}
\caption{LeNet5}
\label{fig:ece_var_mnist_lenet5}
\end{subfigure}
\caption{log(ECE) vs. Variance on MNIST.}
\label{fig:ece_var_mnist}
\end{figure}
\begin{figure}[!hbt]
\centering
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_CIFAR10_densenet40.png}
\caption{DenseNet40}
\label{fig:ece_var_cifar10_densenet40}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_CIFAR10_lenet5.png}
\caption{LeNet5}
\label{fig:ece_var_cifar10_lenet5}
\end{subfigure}
\\
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_CIFAR10_resnet110.png}
\caption{ResNet110}
\label{fig:ece_var_cifar10_resnet110}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_CIFAR10_resnet110sd.png}
\caption{ResNet110 SD}
\label{fig:ece_var_cifar10_resnet110sd}
\end{subfigure}
\\
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_CIFAR10_wrn28-10.png}
\caption{WRN 28-10}
\label{fig:ece_var_cifar10_wrn28-10}
\end{subfigure}
\caption{log(ECE) vs. Variance on CIFAR10.}
\label{fig:ece_var_cifar10}
\end{figure}
\begin{figure}[!hbt]
\centering
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_CIFAR100_densenet40.png}
\caption{DenseNet40}
\label{fig:ece_var_cifar100_densenet40}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_CIFAR100_lenet5.png}
\caption{LeNet5}
\label{fig:ece_var_cifar100_lenet5}
\end{subfigure}
\\
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_CIFAR100_resnet110.png}
\caption{ResNet110}
\label{fig:ece_var_cifar100_resnet110}
\end{subfigure}
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_CIFAR100_resnet110sd.png}
\caption{ResNet110 SD}
\label{fig:ece_var_cifar100_resnet110sd}
\end{subfigure}
\\
\begin{subfigure}{.49\textwidth}
\includegraphics[width=0.9\textwidth]{imgs/ece_var_CIFAR100_wrn28-10.png}
\caption{WRN 28-10}
\label{fig:ece_var_cifar100_wrn28-10}
\end{subfigure}
\caption{log(ECE) vs. Variance on CIFAR100.}
\label{fig:ece_var_cifar100}
\end{figure}
\clearpage
\subsection{Number of bins}
In addition to the general setting for computing ECE (15 bins), we also use different number of bins (5, 10, 30, 50, 100) for better evaluation. The purpose of this evaluation is to show sensitive each algorithm is to the bin size -- a larger sensitivity would imply that an algorithm might not generalize as well on new data.
As shown in Tables~\ref{tab:ece_5_bins}-\ref{tab:ece_100_bins}, Hoki outperforms other algorithms in many benchmarks except the extreme cases, using 5 bins and 100 bins.
We emphasize that Hoki always produces the best or the second-best performance on the challenging ImageNet except for one case -- ECE using 100 bins with the ResNet152 model.
This result shows that, although Hoki is calibrated with 15 bins, the calibration is shown to be effective with various bin sizes, which highlights the benefit of calibration using transformations.
\begin{table*}[!hbt]
\caption{ECE using 5 bins}
\label{tab:ece_5_bins}
\centering
\resizebox{\textwidth}{!}{
\begin{sc}\begin{tabular}{llcccccccccc}
\toprule
Dataset & Model & Uncal. & TempS & VecS & \thead{MS-\\ODIR} & \thead{Dir-\\ODIR} & ETS & IRM & \thead{IROvA-\\TS} & ReCal & Hoki{}\\
\midrule
MNIST & LeNet 5 & 0.0074 & 0.0014 & 0.0013 & 0.0011 & 0.0016 & 0.0015 & \underline{0.0011} & 0.0015 & 0.0030 & \textbf{0.0008}\\
\midrule
CIFAR10 & DenseNet 40 & 0.0519 & 0.0043 & \textbf{0.0023} & 0.0040 & \underline{0.0035} & 0.0040 & 0.0037 & 0.0038 & 0.0064 & 0.0039\\
CIFAR10 & LeNet 5 & 0.0169 & 0.0074 & \underline{0.0065} & 0.0067 & 0.0068 & 0.0088 & \textbf{0.0065} & 0.0124 & 0.0192 & 0.0110\\
CIFAR10 & ResNet 110 & 0.0450 & 0.0066 & 0.0074 & \underline{0.0046} & 0.0083 & 0.0081 & 0.0051 & 0.0057 & 0.0103 & \textbf{0.0045}\\
CIFAR10 & ResNet 110 SD & 0.0534 & 0.0067 & 0.0053 & 0.0050 & \underline{0.0046} & 0.0091 & 0.0058 & 0.0063 & 0.0137 & \textbf{0.0044}\\
CIFAR10 & WRN 28-10 & 0.0248 & 0.0089 & 0.0095 & 0.0092 & 0.0092 & 0.0087 & 0.0115 & \underline{0.0027} & 0.0085 & \textbf{0.0026}\\
\midrule
CIFAR100 & DenseNet 40 & 0.1728 & 0.0126 & 0.0253 & 0.0285 & 0.0182 & 0.0106 & \underline{0.0085} & 0.0107 & 0.0344 & \textbf{0.0073}\\
CIFAR100 & LeNet 5 & \underline{0.0070} & 0.0206 & 0.0129 & 0.0105 & 0.0097 & 0.0180 & 0.0080 & \textbf{0.0060} & 0.0338 & 0.0123\\
CIFAR100 & ResNet 110 & 0.1422 & \textbf{0.0072} & 0.0300 & 0.0340 & 0.0190 & \underline{0.0109} & 0.0110 & 0.0147 & 0.0371 & 0.0127\\
CIFAR100 & ResNet 110 SD & 0.1229 & 0.0071 & 0.0358 & 0.0354 & 0.0207 & \underline{0.0067} & \textbf{0.0051} & 0.0125 & 0.0418 & 0.0084\\
CIFAR100 & WRN 28-10 & 0.0521 & 0.0422 & 0.0436 & 0.0354 & 0.0327 & 0.0364 & 0.0140 & \textbf{0.0088} & 0.0319 & \underline{0.0111}\\
\midrule
ImageNet & DenseNet 161 & 0.0564 & 0.0191 & 0.0211 & 0.0367 & 0.0477 & 0.0126 & 0.0093 & \underline{0.0068} & 0.0482 & \textbf{0.0043}\\
ImageNet & MobileNet V2 & 0.0266 & 0.0150 & 0.0135 & 0.0191 & 0.0204 & 0.0134 & 0.0060 & \underline{0.0038} & 0.0471 & \textbf{0.0010}\\
ImageNet & ResNet 152 & 0.0490 & 0.0201 & 0.0207 & 0.0347 & 0.0397 & 0.0139 & 0.0112 & \textbf{0.0031} & 0.0455 & \underline{0.0052}\\
ImageNet & WRN 101-2 & 0.0524 & 0.0307 & 0.0300 & 0.0413 & 0.0194 & 0.0245 & 0.0155 & \textbf{0.0065} & 0.0425 & \underline{0.0067}\\
\bottomrule
\end{tabular}
\end{sc}
}
\end{table*}
\begin{table*}[!hbt]
\caption{ECE using 10 bins}
\label{tab:ece_10_bins}
\centering
\resizebox{\textwidth}{!}{
\begin{sc}\begin{tabular}{llcccccccccc}
\toprule
Dataset & Model & Uncal. & TempS & VecS & \thead{MS-\\ODIR} & \thead{Dir-\\ODIR} & ETS & IRM & \thead{IROvA-\\TS} & ReCal & Hoki{}\\
\midrule
MNIST & LeNet 5 & 0.0076 & 0.0014 & 0.0013 & \underline{0.0011} & 0.0016 & 0.0016 & 0.0014 & 0.0019 & 0.0030 & \textbf{0.0008}\\
\midrule
CIFAR10 & DenseNet 40 & 0.0519 & 0.0068 & \textbf{0.0029} & 0.0040 & \underline{0.0035} & 0.0056 & 0.0073 & 0.0073 & 0.0094 & 0.0057\\
CIFAR10 & LeNet 5 & 0.0177 & 0.0079 & 0.0076 & \textbf{0.0067} & \underline{0.0068} & 0.0110 & 0.0071 & 0.0142 & 0.0192 & 0.0110\\
CIFAR10 & ResNet 110 & 0.0457 & 0.0079 & 0.0082 & 0.0075 & 0.0083 & 0.0084 & \textbf{0.0054} & 0.0091 & 0.0106 & \underline{0.0071}\\
CIFAR10 & ResNet 110 SD & 0.0535 & 0.0088 & 0.0077 & 0.0050 & \underline{0.0046} & 0.0109 & 0.0072 & 0.0078 & 0.0145 & \textbf{0.0044}\\
CIFAR10 & WRN 28-10 & 0.0251 & 0.0090 & 0.0095 & 0.0092 & 0.0092 & 0.0087 & 0.0147 & \underline{0.0046} & 0.0085 & \textbf{0.0026}\\
\midrule
CIFAR100 & DenseNet 40 & 0.1728 & 0.0152 & 0.0258 & 0.0285 & 0.0182 & 0.0134 & \underline{0.0098} & 0.0124 & 0.0380 & \textbf{0.0073}\\
CIFAR100 & LeNet 5 & 0.0118 & 0.0208 & 0.0129 & \underline{0.0105} & \textbf{0.0097} & 0.0204 & 0.0131 & 0.0119 & 0.0357 & 0.0282\\
CIFAR100 & ResNet 110 & 0.1422 & \textbf{0.0074} & 0.0301 & 0.0340 & 0.0190 & \underline{0.0121} & 0.0138 & 0.0181 & 0.0419 & 0.0127\\
CIFAR100 & ResNet 110 SD & 0.1229 & \underline{0.0077} & 0.0358 & 0.0354 & 0.0207 & 0.0079 & \textbf{0.0066} & 0.0140 & 0.0418 & 0.0098\\
CIFAR100 & WRN 28-10 & 0.0524 & 0.0424 & 0.0446 & 0.0354 & 0.0327 & 0.0388 & 0.0303 & \textbf{0.0106} & 0.0319 & \underline{0.0112}\\
\midrule
ImageNet & DenseNet 161 & 0.0564 & 0.0203 & 0.0236 & 0.0367 & 0.0477 & 0.0126 & 0.0097 & \underline{0.0077} & 0.0486 & \textbf{0.0043}\\
ImageNet & MobileNet V2 & 0.0266 & 0.0156 & 0.0149 & 0.0191 & 0.0204 & 0.0138 & 0.0068 & \underline{0.0064} & 0.0474 & \textbf{0.0010}\\
ImageNet & ResNet 152 & 0.0490 & 0.0201 & 0.0207 & 0.0347 & 0.0397 & 0.0139 & 0.0112 & \underline{0.0055} & 0.0457 & \textbf{0.0052}\\
ImageNet & WRN 101-2 & 0.0524 & 0.0307 & 0.0310 & 0.0413 & 0.0194 & 0.0245 & 0.0155 & \underline{0.0076} & 0.0425 & \textbf{0.0067}\\
\bottomrule
\end{tabular}
\end{sc}
}
\end{table*}
\begin{table*}[!hbt]
\caption{ECE using 30 bins}
\label{tab:ece_30_bins}
\centering
\resizebox{\textwidth}{!}{
\begin{sc}\begin{tabular}{llcccccccccc}
\toprule
Dataset & Model & Uncal. & TempS & VecS & \thead{MS-\\ODIR} & \thead{Dir-\\ODIR} & ETS & IRM & \thead{IROvA-\\TS} & ReCal & Hoki{}\\
\midrule
MNIST & LeNet 5 & 0.0078 & \underline{0.0025} & 0.0034 & 0.0037 & 0.0033 & 0.0035 & 0.0030 & 0.0032 & 0.0036 & \textbf{0.0008}\\
\midrule
CIFAR10 & DenseNet 40 & 0.0525 & 0.0100 & \textbf{0.0052} & 0.0088 & 0.0083 & 0.0108 & 0.0103 & 0.0110 & 0.0136 & \underline{0.0057}\\
CIFAR10 & LeNet 5 & 0.0230 & 0.0144 & 0.0153 & 0.0165 & 0.0145 & 0.0158 & \underline{0.0142} & 0.0181 & 0.0283 & \textbf{0.0110}\\
CIFAR10 & ResNet 110 & 0.0458 & 0.0098 & 0.0097 & 0.0098 & \underline{0.0091} & 0.0095 & \textbf{0.0086} & 0.0103 & 0.0156 & 0.0110\\
CIFAR10 & ResNet 110 SD & 0.0538 & 0.0135 & \underline{0.0110} & 0.0133 & \textbf{0.0103} & 0.0136 & 0.0130 & 0.0145 & 0.0172 & 0.0229\\
CIFAR10 & WRN 28-10 & 0.0255 & 0.0110 & 0.0103 & 0.0100 & 0.0105 & 0.0098 & 0.0164 & \underline{0.0051} & 0.0100 & \textbf{0.0026}\\
\midrule
CIFAR100 & DenseNet 40 & 0.1728 & 0.0201 & 0.0317 & 0.0324 & 0.0224 & 0.0179 & 0.0219 & \underline{0.0150} & 0.0386 & \textbf{0.0075}\\
CIFAR100 & LeNet 5 & \underline{0.0174} & 0.0224 & 0.0183 & 0.0192 & 0.0232 & 0.0231 & 0.0197 & \textbf{0.0144} & 0.0373 & 0.0282\\
CIFAR100 & ResNet 110 & 0.1423 & \textbf{0.0122} & 0.0319 & 0.0351 & 0.0231 & 0.0132 & 0.0190 & 0.0223 & 0.0470 & \underline{0.0127}\\
CIFAR100 & ResNet 110 SD & 0.1230 & 0.0119 & 0.0367 & 0.0362 & 0.0208 & 0.0173 & \underline{0.0108} & 0.0187 & 0.0449 & \textbf{0.0098}\\
CIFAR100 & WRN 28-10 & 0.0538 & 0.0449 & 0.0453 & 0.0378 & 0.0358 & 0.0392 & 0.0397 & \underline{0.0162} & 0.0366 & \textbf{0.0112}\\
\midrule
ImageNet & DenseNet 161 & 0.0564 & 0.0203 & 0.0244 & 0.0373 & 0.0477 & 0.0139 & 0.0111 & \underline{0.0105} & 0.0500 & \textbf{0.0085}\\
ImageNet & MobileNet V2 & 0.0280 & 0.0188 & 0.0193 & 0.0230 & 0.0275 & 0.0183 & 0.0110 & \underline{0.0089} & 0.0479 & \textbf{0.0011}\\
ImageNet & ResNet 152 & 0.0494 & 0.0211 & 0.0223 & 0.0359 & 0.0399 & 0.0144 & 0.0150 & \underline{0.0099} & 0.0460 & \textbf{0.0052}\\
ImageNet & WRN 101-2 & 0.0532 & 0.0321 & 0.0330 & 0.0420 & 0.0305 & 0.0262 & 0.0234 & \underline{0.0105} & 0.0428 & \textbf{0.0067}\\
\bottomrule
\end{tabular}
\end{sc}
}
\end{table*}
\begin{table*}[!hbt]
\caption{ECE using 50 bins}
\label{tab:ece_50_bins}
\centering
\resizebox{\textwidth}{!}{
\begin{sc}\begin{tabular}{llcccccccccc}
\toprule
Dataset & Model & Uncal. & TempS & VecS & \thead{MS-\\ODIR} & \thead{Dir-\\ODIR} & ETS & IRM & \thead{IROvA-\\TS} & ReCal & Hoki{}\\
\midrule
MNIST & LeNet 5 & 0.0080 & 0.0044 & 0.0039 & 0.0037 & 0.0038 & 0.0032 & 0.0047 & \textbf{0.0026} & 0.0038 & \underline{0.0029}\\
\midrule
CIFAR10 & DenseNet 40 & 0.0524 & 0.0111 & 0.0116 & 0.0108 & \underline{0.0103} & 0.0135 & 0.0111 & 0.0113 & 0.0152 & \textbf{0.0064}\\
CIFAR10 & LeNet 5 & 0.0247 & 0.0182 & 0.0191 & 0.0194 & 0.0208 & 0.0184 & \underline{0.0179} & 0.0207 & 0.0312 & \textbf{0.0128}\\
CIFAR10 & ResNet 110 & 0.0462 & 0.0123 & 0.0113 & 0.0113 & 0.0107 & 0.0108 & \underline{0.0104} & 0.0105 & 0.0172 & \textbf{0.0071}\\
CIFAR10 & ResNet 110 SD & 0.0541 & 0.0146 & 0.0133 & 0.0142 & \textbf{0.0115} & 0.0161 & 0.0152 & \underline{0.0131} & 0.0181 & 0.0168\\
CIFAR10 & WRN 28-10 & 0.0255 & 0.0131 & 0.0120 & 0.0112 & 0.0109 & 0.0105 & 0.0169 & \underline{0.0054} & 0.0122 & \textbf{0.0026}\\
\midrule
CIFAR100 & DenseNet 40 & 0.1728 & 0.0244 & 0.0330 & 0.0339 & 0.0295 & 0.0240 & 0.0270 & \underline{0.0163} & 0.0435 & \textbf{0.0114}\\
CIFAR100 & LeNet 5 & 0.0241 & 0.0261 & \underline{0.0205} & 0.0230 & 0.0270 & 0.0281 & 0.0278 & \textbf{0.0148} & 0.0407 & 0.0455\\
CIFAR100 & ResNet 110 & 0.1423 & \textbf{0.0156} & 0.0322 & 0.0379 & 0.0256 & 0.0197 & 0.0234 & 0.0232 & 0.0515 & \underline{0.0192}\\
CIFAR100 & ResNet 110 SD & 0.1232 & \textbf{0.0164} & 0.0385 & 0.0398 & 0.0227 & 0.0233 & 0.0209 & 0.0221 & 0.0490 & \underline{0.0167}\\
CIFAR100 & WRN 28-10 & 0.0562 & 0.0465 & 0.0470 & 0.0387 & 0.0376 & 0.0413 & 0.0404 & \underline{0.0171} & 0.0387 & \textbf{0.0152}\\
\midrule
ImageNet & DenseNet 161 & 0.0567 & 0.0227 & 0.0257 & 0.0380 & 0.0478 & 0.0176 & 0.0147 & \underline{0.0131} & 0.0504 & \textbf{0.0044}\\
ImageNet & MobileNet V2 & 0.0299 & 0.0187 & 0.0204 & 0.0241 & 0.0287 & 0.0194 & 0.0145 & \underline{0.0094} & 0.0497 & \textbf{0.0016}\\
ImageNet & ResNet 152 & 0.0499 & 0.0239 & 0.0249 & 0.0365 & 0.0399 & 0.0177 & 0.0179 & \textbf{0.0107} & 0.0474 & \underline{0.0177}\\
ImageNet & WRN 101-2 & 0.0544 & 0.0341 & 0.0342 & 0.0428 & 0.0342 & 0.0271 & 0.0212 & \underline{0.0106} & 0.0438 & \textbf{0.0067}\\
\bottomrule
\end{tabular}
\end{sc}
}
\end{table*}
\begin{table*}[!hbt]
\caption{ECE using 100 bins}
\label{tab:ece_100_bins}
\centering
\resizebox{\textwidth}{!}{
\begin{sc}\begin{tabular}{llcccccccccc}
\toprule
Dataset & Model & Uncal. & TempS & VecS & \thead{MS-\\ODIR} & \thead{Dir-\\ODIR} & ETS & IRM & \thead{IROvA-\\TS} & ReCal & Hoki{}\\
\midrule
MNIST & LeNet 5 & 0.0086 & 0.0057 & 0.0053 & 0.0052 & 0.0057 & 0.0054 & 0.0062 & \textbf{0.0032} & \underline{0.0045} & 0.0053\\
\midrule
CIFAR10 & DenseNet 40 & 0.0538 & 0.0156 & 0.0166 & 0.0146 & 0.0158 & 0.0176 & 0.0162 & \underline{0.0125} & 0.0171 & \textbf{0.0069}\\
CIFAR10 & LeNet 5 & 0.0285 & 0.0268 & \underline{0.0243} & 0.0259 & 0.0273 & 0.0289 & 0.0279 & \textbf{0.0213} & 0.0353 & 0.0336\\
CIFAR10 & ResNet 110 & 0.0474 & 0.0150 & 0.0148 & 0.0148 & 0.0142 & 0.0141 & 0.0146 & \underline{0.0110} & 0.0208 & \textbf{0.0076}\\
CIFAR10 & ResNet 110 SD & 0.0551 & 0.0196 & 0.0179 & 0.0189 & \underline{0.0165} & 0.0195 & 0.0214 & \textbf{0.0150} & 0.0213 & 0.0282\\
CIFAR10 & WRN 28-10 & 0.0261 & 0.0146 & 0.0143 & 0.0148 & 0.0152 & 0.0121 & 0.0201 & \textbf{0.0061} & 0.0132 & \underline{0.0110}\\
\midrule
CIFAR100 & DenseNet 40 & 0.1731 & 0.0292 & 0.0369 & 0.0393 & 0.0333 & 0.0340 & 0.0329 & \textbf{0.0174} & 0.0518 & \underline{0.0189}\\
CIFAR100 & LeNet 5 & 0.0321 & 0.0346 & 0.0303 & \underline{0.0296} & 0.0297 & 0.0338 & 0.0340 & \textbf{0.0166} & 0.0497 & 0.0455\\
CIFAR100 & ResNet 110 & 0.1425 & \underline{0.0264} & 0.0356 & 0.0416 & 0.0307 & 0.0264 & 0.0304 & \textbf{0.0257} & 0.0562 & 0.0456\\
CIFAR100 & ResNet 110 SD & 0.1235 & 0.0248 & 0.0408 & 0.0424 & 0.0293 & 0.0318 & 0.0281 & \underline{0.0232} & 0.0537 & \textbf{0.0167}\\
CIFAR100 & WRN 28-10 & 0.0596 & 0.0486 & 0.0509 & 0.0455 & 0.0405 & 0.0449 & 0.0439 & \textbf{0.0173} & 0.0452 & \underline{0.0206}\\
\midrule
ImageNet & DenseNet 161 & 0.0575 & 0.0256 & 0.0292 & 0.0392 & 0.0483 & 0.0205 & 0.0195 & \underline{0.0136} & 0.0536 & \textbf{0.0050}\\
ImageNet & MobileNet V2 & 0.0316 & 0.0228 & 0.0242 & 0.0274 & 0.0323 & 0.0248 & 0.0194 & \underline{0.0113} & 0.0538 & \textbf{0.0096}\\
ImageNet & ResNet 152 & 0.0509 & 0.0264 & 0.0281 & 0.0384 & 0.0405 & 0.0214 & \underline{0.0201} & \textbf{0.0129} & 0.0508 & 0.0236\\
ImageNet & WRN 101-2 & 0.0554 & 0.0356 & 0.0373 & 0.0441 & 0.0362 & 0.0308 & 0.0267 & \underline{0.0127} & 0.0467 & \textbf{0.0085}\\
\bottomrule
\end{tabular}
\end{sc}
}
\end{table*}
\clearpage
\subsection{Initialization with Original Uncalibrated Confidence}
We perform an experiment to investigate the effect of different initializations.
In Algorithm \ref{alg:design_time} and \ref{alg:run_time}, we initialize the confidence with the validation set accuracy.
We can also use the original uncalibrated confidence from a classifier as the initial value, and we compare ECE values with those two different initialization.
Table \ref{tab:exp_init} shows that the initialization with the validation set accuracy is always better than the initialization with original uncalibrated confidence except two benchmarks, (CIFAR10, DenseNet40) and (CIFAR100, ResNet110 SD).
This difference illustrates the importance of the initialization of Hoki -- starting from a high-variance initial set of confidences may make it harder to converge to a good local optimum in terms of ECE.
\begin{table*}[!hbt]
\caption{{ECE by different initialization}}
\label{tab:exp_init}
\vskip 0.15in
\begin{sc}\begin{center}
\begin{tabular}{llcccccc}
\toprule
Dataset & Model & \thead{Val\\Accuracy} & \thead{Uncalibrated\\Confidence}\\
\midrule
\multirow{1}{*}{MNIST} & LeNet 5 & \textbf{0.0008} & 0.0018\\
\midrule
\multirow{5}{*}{CIFAR10} & DenseNet 40 & 0.0057 & \textbf{0.0038}\\
& LeNet 5 & \textbf{0.0110} & 0.0171\\
& ResNet 110 & \textbf{0.0071} & 0.0093\\
& ResNet 110 SD & \textbf{0.0044} & 0.0060\\
& WRN 28-10 & \textbf{0.0026} & 0.0042\\
\midrule
\multirow{5}{*}{CIFAR100} & DenseNet 40 & \textbf{0.0073} & 0.0178\\
& LeNet 5 & \textbf{0.0123} & 0.0189\\
& ResNet 110 & \textbf{0.0127} & 0.0157\\
& ResNet 110 SD & 0.0098 & \textbf{0.0090}\\
& WRN 28-10 & \textbf{0.0112} & 0.0117\\
\midrule
\multirow{4}{*}{ImageNet} & DenseNet 161 & \textbf{0.0043} & 0.0069\\
& MobileNet V2 & \textbf{0.0011} & 0.0061\\
& ResNet 152 & \textbf{0.0052} & 0.0081 \\
& WRN 101-2 & \textbf{0.0067} & 0.0077\\
\bottomrule
\end{tabular}
\end{center}
\end{sc}
\end{table*}
\clearpage
\section{Computing Environment}
All experiments were run on a server with the specifications described in Table~\ref{tab:computing_spec}.
\begin{table}[!hbt]
\centering
\caption{Computing Specification}
\label{tab:computing_spec}
\begin{tabular}{cc}
\toprule
Item & Specification \\
\midrule
CPU & Intel(R) Xeon(R) Gold 6248 CPU @ 2.50GHz \\
Memory & 768 GB\\
GPU & NVIDIA GeForce RTX 2080 Ti\\
\bottomrule
\end{tabular}
\end{table}
|
\section{Introduction}
\label{Sec:Intro}
In this article, we investigate several topological indices of a class of random graphs. The {\em topological index} of a graph is a graph-invariant descriptor that quantifies its structure or some kind of feature. Topological index has found a plethora of applications in chemical graph theory, mathematical chemistry and chemoinformatics. In practice, various kinds of topological indices, such as Zagreb index~\cite{Gutman1972} and Randi\'{c} index~\cite{Randic}, are used to compare molecular graphs of chemical compounds~\cite{Todeschini}, and to model quantitative structure-property relationship (QSPR) and quantitative structure-activity relationship (QSAR) between molecules~\cite{Devillers}. We refer interested readers to~\cite{Todeschini} for a text-style exposition of utilization of topological index in chemistry.
Specifically, we look into a class of random graphs that incorporate randomness into {\em caterpillar graphs}, i.e., {\em random caterpillars}. In mathematical chemistry and chemoinformatics, a caterpillar graph (or simply caterpillar) is an acyclic graph with the property that there remains a path (called spine) if all leaves are pruned, best known for modeling the structure and intrinsic properties of benzoid hydrocarbon molecules~\cite{Andrade, ElBasil, ElBasil1990, Gutman1985}.
Our motivation of incorporating randomness and caterpillar graphs is from a recent article~\cite{Kryven}, in which a random graph model was utilized to model the chemistry of a discrete polymerization process. More precisely, we consider random caterpillars that grow in a uniform manner. At time $0$, there is a {\em spine} consisting of $m \ge 2$ (fixed) nodes. At each subsequent point of time, a {\em leaf} is connected with one of the spine nodes by an edge, all spine nodes being equally likely to be selected. At time $n$, we denote the structure of a random caterpillar by $C_n$.
At first, we present some known results of $C_n$ as preliminaries, and give some notations that will be used throughout the manuscript. At time $n$, there is a total of $(n + m)$ nodes. Additionally, the total number of edges is fixed, i.e., $(n + 2m - 2)$. Enumerate $m$ spine nodes in a preserved order (e.g., from left to right) with distinct numbers in $\{1, 2, \ldots, m\}$. Let $X_{i, n}$ be the number of leaves attached to spine node $i$ for $i = 1, 2, \ldots, m$, and let $D_{i, n}$ be the degree of spine node $i$. According to the evolution of random caterpillars, we know that the joint distribution of $X_{i, n}$'s is multinomial with parameters $n$ and $\mathbb{I}_{m \times 1}/m$, where $\mathbb{I}_{m \times 1}$ is a column vector of all $1$'s. Additionally, there is an instantaneous relation between $X_{i, n}$ and $D_{i, n}$. That is $D_{i, n} = X_{i, n} + 1$ for $i = 1, m$; $D_{i, n} = X_{i, n} + 2$ for $i = 2, 3, \ldots, m - 1$.
In the remainder of the paper, we calculate several topological
indices of $C_n$. They are Gini index, Hoover index, Zagreb index,
Randi\'{c} index and Wiener index, respectively presented from
Section~\ref{Sec:Gini} to Section~\ref{Sec:Wiener}. We give the
definition of each index and some brief introductions about their
applications that initiate and promote the motivations of our
analysis in the sequel. In Section~\ref{Sec:numeric}, we carry out
numerical experiments to verify the theoretical results developed in
the preceding sections. Some concluding remarks are addressed at the
end of the article .
\section{Gini index}
\label{Sec:Gini}
The {\em Gini index} (or called Gini coefficient) is a widely used measure in economics~\cite{Gini}, mainly known for assessing statistical dispersion of income or wealth of a population. More precisely, the Gini index (of a target population) is a number measuring the degree of inequality in (income or wealth) distribution. Given a population of size $n$, the {\em Gini index} is given by
$$G = \frac{\sum_{i = 1}^{n} \sum_{j = 1}^{n} |w_i - w_j|}{2n\sum_{i = 1}^{n} w_i},$$
where $w_i$ refers to the wealth of person $i$.
Recently, different types of Gini index that are well defined for graphs were proposed. We present the results only in this section without proof, as random caterpillars were used as examples in all the relevant sources. A distance-based Gini index for rooted trees was given in~\cite{Balaji}, used as a measure of disparities of trees within random tree classes. Let $B_n$ be the distance-based Gini index of $C_n$. The exact and the asymptotic mean of $B_n$ were calculated in~\cite{Balaji}, and revisited by~\cite{Zhang}.
\begin{prop}[\cite{Balaji}]
The mean of the distance-based Gini index for a random caterpillar is
$$\E\left[G_n\right] = \frac{(2m^2 - 2)n^2 + (m^3 + 4m^2 - m + 2)n + 2m^4 - 2m^2}{(6m^2 + 6m)n^2 + 12m^3n + 6m^4 - 6m^3}.$$
As $n \to \infty$ and $m \to \infty$, the expectation of this index converges to $1/3$.
\end{prop}
In an independent work, a degree-based Gini index for general graphs was proposed by~\cite{Domicolo}. This particular Gini index, in general, is used to assess the regularity of classes of random graphs. A follow-up study that uncovers a duality theory is conducted in~\cite{Domicoloduality}. To avoid ambiguity, we denote the degree-based Gini index in ~\cite{Domicolo} by $\tilde{B}_n$.
\begin{prop}[\cite{Domicolo}]
The mean of the degree-based Gini index for the class of random caterpillars is
$$\tilde{B}_n \, \overset{L_1}{\longrightarrow} \, \frac{1}{2},$$
as $n \to \infty$.
\end{prop}
A slightly different degree-based Gini index for random caterpillars
was discussed in an independent source~\cite{Zhang}, where a
different target population was considered. It is evident that this
type of Gini index degenerates as $n$ goes to infinity, shown
in~\cite{Zhang} via numerical experiments and
in~\cite{Domicoloduality} via a probabilistic approach.
\section{Hoover index}
\label{Sec:Hoover}
Another important topological measure with applications in economics is the {\em Hoover index}~\cite{Hoover}, which is also known as the Robin Hood Index or the Schutz index in the literature. Like the Gini index, the Hoover index is another inequality metric used for measuring the deviation of the current (income or wealth) distribution from the perfectly even distribution. An alternative interpretation of this index is the portion of population income that would be taken from the richer half to the poorer half for the whole community to be perfectly equal. Mathematically, the {\em Hoover index} of a population with size $n$ is
\begin{equation}
\label{Eq:Hoover}
H=\frac{n}{2}\times \frac{\sum_{i = 1}^{n} |w_i - \bar{w}|}{\bar{w}},
\end{equation}
where $\bar{w}=1/n\sum_{i=1}^n w_i$ is the average of the entire population wealth. A graphical interpretation of the Hoover index is the longest vertical distance between the {\em Lorenz curve} and the $45$ degree line of a unit square. Thus, we immediately have $0 \le H < 1$.
A graph-friendly Hoover index is to replace all $w_i$'s in Equation~(\ref{Eq:Hoover}) with node degrees. Our intent is to consider the Hoover index for a class of graphs. Therefore, we propose a degree-based Hoover index for graphs analogous to the degree-based Gini index introduced in~\cite{Domicolo} as a competing measure for assessing graph regularity. Given an arbitrary graph $G = (V, E) \in \mathsf{G}$, where $V$ and $E$ respectively denotes the vertex set and the edge set of graph $G$ and $\mathsf{G}$ is the class to which $G$ belongs, the degree-based Hoover index is defined as follows:
$$H(G) = \frac{1}{2} \times \frac{\sum_{v \in V} \bigl|{\rm deg}(v) - \sum_{v \in V} {\rm deg}(v) / |V| \big|}{\E \bigl[|\mathcal{V}|\bigr] \times \E \bigl[{\rm deg}(\mathcal{U}) \bigr]},$$
where ${\rm deg}(v)$ represents the {\em degree} (i.e., the number of edges incident to $v$) of~$v$, $|V|$ is the cardinality of set $V$, $\E \bigl[|\mathcal{V}|\bigr]$ is the expected {\em order} (i.e., the number of nodes) of a randomly chosen graph in $\mathsf{G}$, and $\E \bigl[{\rm deg}(\mathcal{U}) \bigr]$ is the degree of a randomly chosen node in a randomly chosen graph in $\mathsf{G}$. The Hoover index of the class $\mathsf{G}$, $H(\mathsf{G})$, is the average of all $H(G)$ for $G$ in $\mathsf{G}$. An argument similar to~\cite[Theorem 4.1]{Domicolo} can be established to show that the Hoover index of a graph class takes values between $0$ and $1$ asymptotically, and a value closer to $0$ suggests that the graphs in the class tend to be more regular.
The order of an arbitrary $C_n$ is fixed, i.e., $(n + m)$. Let $\mathcal{U}_{C_n}$ be a randomly selected node of $C_n$ uniformly chosen from the class of random caterpillars. We have
$$
\E \bigl[{\rm deg}(\mathcal{U}_{C_n}) \bigr] = \frac{\sum_{i = 1}^{m} D_{i, n} + \sum_{i = 1}^{m} X_{i, n}}{n + m} = \frac{2n + 2m - 2}{n + m} = 2 - \frac{2}{n + m}.
$$
Next, we compute the numerator. Note that all leaves have degree $1$, less than the average. The probability that the (spine) nodes at the two ends of the spine are never selected in a long run is negligible. Hence, with high probability, the degree of each spine node is larger than the average. Thus, the expectation of the numerator is equivalent to
$$
\sum_{i = 1}^{m} \left(\E[D_{i, n}] - \frac{2n + 2m - 2}{n + m}\right) + n \, \left(\frac{2n + 2m - 2}{n + m} - 1\right) = \frac{2n(n + m - 2)}{n + m}.
$$
In what follows, we get the asymptotic mean of the Hoover index of the class of random caterpillars at time $n$, denoted by $H_n$, in the next proposition.
\begin{prop}
\label{Prop:meanH}
The mean of the Hoover index of the class of random caterpillars
$$H_n \, \to \, \frac{1}{2},$$
as $n \to \infty$.
\end{prop}
\begin{proof}
By the definition of the Hoover index, we have
$$
H_n = \E\bigl[H(C_n)\bigr] = \frac{2n(n + m - 2)/(n + m)}{2(n + m)\bigl(2 - 2/(n + m)\bigr)} \, \to \, \frac{1}{2},
$$
as $n \to \infty$.
\end{proof}
\section{Zagreb index}
\label{Sec:Zagreb}
In this section, we calculate the Zagreb index of a random
caterpillar at time $n$, denoted by $Z_n = \mbox{{\tt
Zagreb}}(C_n)$. The {\em Zagreb index} of a graph is defined as the
sum of the squared degrees of the nodes in the graph. Applications
of Zagreb index mostly appear in mathematical chemistry, used to
study molecular complexity~\cite{Nikolic},
chirality~\cite{Golbraikh}, ZE-isomerism~\cite{Golbraikh2002} and
heterosystems~\cite{Milicevic}. It is not even possible to list them
all, so we refer the interested readers to a survey
article~\cite{Nikolic2003}, in which the authors emphasized the
potential applicability of Zagreb index for deriving multilinear
regression models. In the literature of random graphs and
algorithms, the Zagreb indices of random recursive trees, random
$b$-ary tree, plain-oriented recursive trees and preferential
attachment caterpillars were respectively studied in~\cite{Feng,
Feng2015, Zhangarxiv, Zhangarxiv2019}.
By definition, the Zagreb index of random caterpillars is given by
$$Z_n = \sum_{i = 1}^{m} D^2_{i, n} + \sum_{i = 1}^{m} X_{i, n}.$$
Here, first, we give some additional useful notations. Let ${\bf 1}_n(i)$ denote the event that the spine node labeled with $i$ is selected at time $n$, and let $\mathbb{F}_n$ denote the $\sigma$-field generated by the history of the growth of a caterpillar in the first $n$ stages. We present the mean of $Z_n$ as well as a weak law in the next proposition.
\begin{prop}
\label{Prop:meanZ}
For $n \ge 0$, we have
$$\E[Z_n] = \frac{n^2}{m} + \frac{(6m - 5)n}{m} + 4m - 6.$$
As $n \to \infty$, we have
$$\frac{Z_n}{n^2} \, \overset{L_1}{\longrightarrow} \, \frac{1}{m}.$$
This convergence takes place in probability as well.
\end{prop}
\begin{proof}
We consider the following almost-sure recursive relation between $Z_{n - 1}$ and $Z_n$, conditional on ${\bf 1}_n(i)$ and $\mathbb{F}_{n - 1}$:
\begin{equation}
\label{Eq:recZ}
Z_n = Z_{n - 1} + (D_{i, n - 1} + 1)^2 - D^2_{i, n - 1} + 1 = Z_{n - 1} + 2 D_{i, n - 1} + 2.
\end{equation}
Averaging it out over $i$, we get
\begin{equation*}
\E[Z_n \, | \, \mathbb{F}_{n - 1}] = Z_{n - 1} + \frac{2}{m} \sum_{i = 1}^{m} D_{i, n - 1} + 2 = Z_{n - 1} + \frac{2(n + 2m - 3)}{m} + 2.
\end{equation*}
Taking another expectation with respect to $\mathbb{F}_{n - 1}$, we obtain the recurrence for $Z_n$. We solve it with initial condition $\E[Z_0] = Z_0 = 4m - 6$, and get the result stated in the proposition.
As $n \to \infty$, we have $Z_n / n^2$ converges to $1/m$ in $L_1$-space, suggesting that $Z_n / n^2$ converges to $1/m$ in probability as well.
\end{proof}
We continue to calculate the second moment of $Z_n$, and accordingly obtain the variance of $Z_n$.
\begin{prop}
\label{Prop:varZ}
For $n \ge 0$, we have
\begin{align*}
\E\left[Z^2_n\right] &= \frac{n^4}{m^2} + \frac{(12m - 10)n^3}{m^2} + \frac{(44m^2 - 70m + 23)n^2}{m^2}
\\ &\qquad{}+ \frac{(48m^3 - 112m^2 + 66m - 14)n}{m^2} + 16\left(m - \frac{3}{2}\right)^2,
\end{align*}
and
$$\mathbb{V}{\rm ar}[Z_n] = \frac{2n \bigl((m - 1)n + 3m - 7\bigr)}{m^2}.$$
\end{prop}
\begin{proof}
Recall the almost-sure recursive relation between $Z_{n - 1}$ and $Z_n$, conditional on ${\bf 1}_n(i)$ and $\mathbb{F}_n$, established in Proposition~\ref{Prop:meanZ} (c.f.\ Equation~(\ref{Eq:recZ})). We square it on both sides and get
\begin{align*}
Z_n^2 &= (Z_{n - 1} + 2 D_{i, n - 1} + 2)^2
\\ &= Z^2_{n - 1} + 4 D^2_{i, n - 1} + 4 + 4 Z_{n - 1} + 4 Z_{n - 1}D_{n - 1} + 8 D_{i, n - 1}
\end{align*}
Again, we average it out over $i$ to obtain
\begin{align*}
\E\left[Z^2_n \, | \, \mathbb{F}_{n - 1}\right] &= Z^2_{n - 1} + \frac{4}{m} \sum_{i = 1}^{m} D^2_{i, n - 1} + 4 + 4 Z_{n - 1} + \frac{4 Z_{n - 1}}{m} \sum_{i = 1}^{m} D_{i, n - 1}
\\ &\qquad{}+ \frac{8}{m} \sum_{i = 1}^{m} D_{i, n - 1}.
\\ &= Z^2_{n - 1} + \frac{4}{m} \bigl(Z_{n - 1} - (n - 1)\bigr) + 4 + 4Z_{n - 1}
\\ &\qquad{}+ \frac{4(n + 2m - 3)}{m} Z_{n - 1} + \frac{8(n +2 m - 3)}{m}.
\end{align*}
We then obtain the recurrence for the second moment of $Z_n$ by taking the expectation with respect to $\mathbb{F}_{n - 1}$ and by plugging the result of the first moment of $Z_{n - 1}$. Solve the recurrence with initial condition $\E\left[Z^2_0\right] = Z^2_0 = (4m - 6)^2$ to get the result stated in the proposition. In what follows, we obtained the exact expression of the variance of $Z_n$ by subtracting the square of the mean of $Z_n$ from its second moment.
\end{proof}
According to the expression of $\E\left[Z_n^2\right]$, we can simply conclude that
$$ \frac{Z^2_n}{n^4} \, \overset{L_1}{\longrightarrow} \, \frac{1}{m^2}, $$
done in a similar manner as the proof for $L_1$ convergence of $Z_n/n^2$. Thus, we also obtain a weak law for $Z^2_n$. Besides, we find a stronger $L_2$ convergence of $Z_n/n^2$, presented in the next corollary.
\begin{corollary}
As $n \to \infty$, we have
$$ \frac{Z_n}{n^2} \, \overset{L_2}{\longrightarrow} \, \frac{1}{m}.$$
\end{corollary}
\begin{proof}
By the $L_1$-convergence results for $Z^2_n$ and $Z_n$, we have
\begin{align*}
\lim_{n \to \infty}\E\left[\left|\frac{Z_n}{n^2} - \frac{1}{m}\right|^2\right] &= \lim_{n \to \infty}\E\left[\frac{Z^2}{n^4} - 2 \times \frac{Z_n}{n^2}\times\frac{1}{m} + \frac{1}{m^2}\right]
\\ &=\frac{1}{m^2} - 2 \times \frac{1}{m} \times \frac{1}{m} + \frac{1}{m^2}
\\ &= 0.
\end{align*}
\end{proof}
According to the variance of $Z_n$ given in Proposition~\ref{Prop:varZ}, we find that the order of its leading term is $n^2$, which is the same as that for the mean of $Z_n$. A sharp concentration of the variance often suggests asymptotic normality of the random variable. In what follows, we characterize the asymptotic behavior of $Z_n$ after properly scaled. Based on our investigation, the scale is $n$. Our strategy is to construct a {\em martingale} array based on a transformation of $Z_n$, and appeal to a {\em Martingale Central Limit Theorem} (MCLTs) for developing a Gaussian law of $Z_n/n$ as $n$ goes to infinity.
In the proof of Proposition~\ref{Prop:meanZ}, we find that
$$\E[Z_n \, | \, \mathbb{F}_{n - 1}] = Z_{n - 1} + \frac{2(n + 2m - 3)}{m} + 2,$$
suggesting that $\{Z_n\}_n$ is not a martingale. In the next lemma, we apply a transformation to $Z_n$ such that the new sequence is a martingale.
\begin{lemma}
\label{Lem:martingale}
The sequence of $\{M_n\}_n$ such that
$$M_n = Z_n - \frac{n(n + 6m - 5)}{m}$$
is a martingale.
\end{lemma}
\begin{proof}
Let us consider $M_n = Z_n + \beta_{n}$ such that the following fundamental property of martingale holds; namely,
\begin{align*}
\E[Z_n + \beta_n \, | \, \mathbb{F}_{n - 1}] = Z_{n - 1} + \frac{2(n + 2m - 3)}{m} + 2 + \beta_n = Z_{n - 1} + \beta_{n - 1}.
\end{align*}
This produces a recurrence for $\beta_n$, namely,
$$\beta_n = \beta_{n - 1} - \frac{2n + 6m - 6}{m}.$$
The solution is $\beta_n = -n(n + 6m - 5)/m$, obtained by taking an arbitrary choice of the initial condition; we choose $\beta_0 = 0$.
\end{proof}
The MCLT that we exploit to show asymptotic normality is based {\color{red} on} martingale differences, expressed in terms of a difference operator, i.e., $\nabla M_j := M_j - M_{j - 1}$. In fact, there are different versions of MCLT listed in~\cite{Hall}, requiring different sets of conditions. The MCLT that we use refers to~\cite[Corollary 3.2]{Hall}, requiring two conditions which are respectively known as the {\em conditional Lindeberg's condition} and the {\em conditional variance condition}. In the next two lemmas, we verify these two conditions one after another.
\begin{lemma}
\label{Lem:Lindeberg}
The Lindeberg's condition is given by
$$U_n := \sum_{j = 1}^{n} \E \left[ \left( \frac{\nabla M_j}{n} \right)^2 {\bf 1} \bigl(| \nabla M_j / n| > \varepsilon\bigr) \, \Bigg{|} \, \mathbb{F}_{j - 1}\right] \, \overset{P}{\longrightarrow} \, 0,$$
for arbitrary $\varepsilon > 0$.
\end{lemma}
\begin{proof}
We establish an absolutely uniform bound in all $j \le n$. By the construction of the martingale, we have
\begin{align*}
|\nabla M_j| &= |M_j - M_{j - 1}|
\\ &\le |Z_j - Z_{j - 1}| + \frac{2j + 6m - 6}{m}
\\ &\le 2 \max_{i} D_{i, j - 1} + 2 + \frac{2j + 6m - 6}{m}
\\ &\le 2\bigl(2 + (j - 1)\bigr) + 2 + \frac{2j + 6m - 6}{m}
\\ &= \left(\frac{2m + 2}{m}\right) j + \frac{10m - 6}{m},
\end{align*}
which is increasing in $j$ for any fixed integer $m \ge 2$. Thus, we conclude that $|\nabla M_j|$ is uniformly bounded by $n$. Hence, for any $\varepsilon > 0$, there exists $n_0(\varepsilon) > 0$ such that the sets $\{|M_j / n| > \varepsilon\}$ are all empty for $n > n_0(\varepsilon)$. In what follows, we conclude that $U_n$ converges to $0$ almost surely, which is stronger than the in-probability convergence required for the Lindeberg's condition.
\end{proof}
\begin{lemma}
\label{Lem:variance}
The conditional variance condition is given by
$$V_n := \sum_{j = 1}^{n} \E \left[ \left( \frac{\nabla M_j}{n} \right)^2 \, \Bigg{|} \, \mathbb{F}_{j - 1}\right] \, \overset{P}{\longrightarrow} \, \frac{2(m - 1)}{m^2}.$$
\end{lemma}
\begin{proof}
We rewrite $V_n$ as follows:
\begin{align*}
V_n &= \frac{1}{n^2} \sum_{j = 1}^{n} \E \left[\bigl(Z_j + \beta_j - (Z_{j - 1} + \beta_{j - 1})\bigr)^2 \, \Big{|} \, \mathbb{F}_{j - 1}\right]
\\ &= \frac{1}{n^2} \sum_{j = 1}^{n} \E \bigl[(Z_j - Z_{j - 1})^2 + 2(Z_j - Z_{j - 1})(\beta_j - \beta_{j - 1})
\\ &\qquad{}+ (\beta_j - \beta_{j - 1})^2 \, | \, \mathbb{F}_{j - 1}\bigr]
\end{align*}
We evaluate the three parts in the summand one after another, by considering the asymptotic equivalents of $Z^2_{j - 1}$ and $Z_{j - 1}$.
\begin{enumerate}
\item The first part is
\begin{align*}
\E[(Z_j - Z_{j - 1})^2 \, | \, \mathbb{F}_{j - 1}] &= \E[Z_j^2 \, | \, \mathbb{F}_{j - 1}] - 2 Z_{j - 1} \E[Z_j \, | \, \mathbb{F}_{j - 1}] + Z^2_{j - 1}
\\&= \frac{4j^2 + 28(m - 1)j + 6m^2 - 68m + 24}{m^2}.
\end{align*}
\item The second part is
\begin{align*}
\E[2(Z_j - Z_{j - 1})(\beta_j - \beta_{j - 1}) \, | \, \mathbb{F}_{j - 1}] &= 2(\beta_j - \beta_{j - 1}) \bigl(\E[Z_j \, | \, \mathbb{F}_{j - 1}] - Z_{j - 1}\bigr)
\\ &= -\frac{8(j + 3m + 3)(j + 3m - 3)}{m^2}.
\end{align*}
\item The third part it
$$\E[(\beta_j - \beta_{j - 1})^2 \, | \, \mathbb{F}_{j - 1}] = (\beta_j - \beta_{j - 1})^2 = \frac{(2j + 6m + 6)^2}{m^2}.$$
\end{enumerate}
Putting three parts together, we obtain the summand in $V_n$ for each $j$. Then, we sum these terms for $j = 1, 2, \ldots, n$, and let $n$ go to infinity, obtaining
$$
V_n = \frac{1}{n^2} \left(\frac{2(m - 1)}{m^2} n^2 + O(n)\right) \, \overset{L_1}{\longrightarrow} \, \frac{2(m - 1)}{m^2}.
$$
The $L_1$ convergence is stronger than the required in-probability convergence.
\end{proof}
\begin{theorem}
\label{Thm:zagreb}
As $n \to \infty$, $Z_n / n$ follows a Gaussian law, namely,
$$\frac{Z_n - n^2/m}{n} \, \overset{D}{\longrightarrow} \, \mathcal{N}\left(0, \frac{2(m - 1)}{m^2}\right).$$
\end{theorem}
The proof is easily verified by using the MCLT~\cite[Corollary 3.2]{Hall}.
\section{Randi\'{c} index}
\label{Sec:Randic}
In this section, the Randi\'{c} index of a random caterpillar, denoted by $R_n = \mbox{{\tt Randi\'{c}}}(C_n)$, is calculated. The {\em Randi\'{c} index} of a graph $G = (V, E)$ (with parameter $\alpha$) is the sum of the product of the degrees (raised to power $\alpha$) of all pairwise connected nodes. Mathematically, it is
$$R(G) = \sum_{\{u, v\} \in E} \bigl({\rm deg}(u){\rm deg}(v)\bigr)^{\alpha},$$
for all $u, v \in V$.
The classical choice of $\alpha$ is $-1/2$~\cite{Randic}. Under such choice, the Randi\'{c} index is also called {\em connectivity index}. The general definition given above was later proposed by~\cite{Bollobas1998}. Similar to Zagreb index, Randi\'{c} is also used for modeling QSAR and QSPR of chemical compounds (e.g., alkanes~\cite{Gutman1999}, saturated hydrocarbon~\cite{Gutman2000} and benzenoid systems~\cite{Rada}) in chemoinformatics. We refer interested readers to~\cite{Randic2001} for a concise survey, to~\cite{Randic2008} for a complete history review and to~\cite{Li} for a summary of mathematical properties. To the best of our knowledge, little work has been done for the Randi\'{c} index of random graph models. The only source that we find in the literature is the Randi\'{c} index of random binary trees~\cite{Feng2008}.
In $C_n$, note that no pair of leaves is connected. Each leaf, instead, is only connected to its parent spine node. Therefore, the contribution by each edge connecting a leaf and its corresponding spine node to $R_n$ is the degree of the spine node raised to power $\alpha$. There are $(m - 1)$ edges on the spine, and the contribution by each of them to $R_n$ is $(D_{i - 1, n} D_{i, n})^{\alpha}$ for $i = 2, 3, \ldots, m$. Hence, we arrive at
\begin{equation}
\label{Eq:Randic}
R_n = \sum_{i = 2}^{m} (D_{i - 1, n} D_{i, n})^{\alpha} + \sum_{i = 1}^{m} X_{i, n} D^{\alpha}_{i, n}.
\end{equation}
Specifically, we consider the Randi\'{c} index with parameter $\alpha = 1$. This particular type of Randi\'{c} index is also popular in mathematical chemistry, viewed as a molecular structure descriptor~\cite{Balaban}. Among all, this index is best known for measuring the branching of molecular carbon-atom skeleton~\cite{Gutman, Gutman2013}. In the rest of this section, all the Randi\'{c} indicies without specification all refer to the Randi\'{c} indicies with parameter $\alpha = 1$. Accordingly, we redefine $R_n$ in Equation~(\ref{Eq:Randic}) as
$$R_n = \sum_{i = 2}^{m} D_{i - 1, n} D_{i, n} + \sum_{i = 1}^{m} X_{i, n} D_{i, n}.$$
Many mathematical properties of this specific class of Randi\'{c} indices ($\alpha = 1$) were investigated by graph theorists and combinatorists recently~\cite{Gutman2015, Khalifeh, Mehdi}. In the graph theory community, the Randi\'{c} index is more often known by a different name: the {\em second-order Zagreb index}. In~\cite{Bollobas}, the Randi\'{c} index for extremal graphs was studied, where another different name ``extreme $1$-weight " was used. In the next proposition, we present the mean of $R_n$ as well as a weak law.
\begin{prop}
\label{Prop:meanR}
For $n \ge 0$, the mean of the Randi\'{c} index of a caterpillar is
$$\E[R_n] = \frac{(2m - 1)n^2 + (7m^2 - 10m + 1)n + 4m^2(m - 2)}{m^2}.$$
As $n \to \infty$, we have
$$\frac{R_n}{n^2} \, \overset{L_1}{\longrightarrow} \, \frac{2m - 1}{m^2}.$$
This convergence takes place in probability as well.
\end{prop}
\begin{proof}
Recall that the the distribution of $X_{i, n}$'s is multinomial with parameters $n$ and $\mathbb{I}_{m \times 1}/m$. Hence, we have $\E[X_{i, n}] = n/m$, $\mathbb{V}{\rm ar}[X_{i, n}] = (m - 1)n/m^2$, and $\E\left[X^2_{i, n}\right] = (n + m - 1)n/m^2$ for each $i$, and $\E[X_{i, n}X_{j, n}] = -n/m^2 + n^2/m^2 = n(n - 1)/m^2$ for all $i \neq j$. In what follows, we obtain
\begin{align*}
\E[R_n] &= \sum_{i = 2}^{m} \E[D_{i - 1, n} D_{i, n}] + \sum_{i = 1}^{m} \E[X_{i, n} D_{i, n}]
\\ &= 2 \, \E\bigl[(X_{1, n} + 1)(X_{2, n} + 2)\bigr] + (m - 3) \, \E\bigl[(X_{2, n} + 2)(X_{3, n} + 2)\bigr]
\\&\qquad{}+ 2 \, \E\bigl[X_{1, n} (X_{1, n} + 1)\bigr] + (m - 2) \, \E\bigl[X_{2, n} (X_{2, n} + 2)\bigr]
\\&= (m - 1) \, \E[X_{1, n}X_{2, n}] + m \, \E\left[X^2_{1, n}\right] + (6m - 8) \, \E[X_{1, n}] + (4m - 8)
\\&=\frac{(m - 1)n(n - 1)}{m^2} + \frac{m(n + m - 1)n}{m^2} + \frac{(6m - 8)n}{m} + 4m - 8
\\& = \frac{(2m - 1)n^2 + (7m^2 - 10m + 1)n + 4m^2(m - 2)}{m^2}.
\end{align*}
It is obvious that $R_n / n^2$ converges to $(2m - 1)/m^2$ in $L_1$-space, and this convergence is stronger than the in-probability convergence stated in the weak law.
\end{proof}
Another approach to calculating the mean of $R_n$ is to establish an almost-sure relation analogous to Equation~(\ref{Eq:recZ}) for $R_n$, to obtain an recurrence for $\E[R_n]$ by taking expectation twice, and lastly to solve the recurrence. We omit the details of the derivation, but just present an intermediate step that we need for the follow-up study:
\begin{align*}
\E[R_j \, | \, \mathbb{F}_{j - 1}] &= R_{j - 1} + \frac{1}{m}\bigl(2(2j + 3m - 5) - D_{1, j - 1} - D_{m, j - 1}\bigr) + 1
\\ &\ge R_{j - 1} + \frac{2j + 7m - 10}{m},
\end{align*}
suggesting that $\{R_j\}_{1 \le j \le n}$ is a super-martingale. Note that we use $j$ as subscript instead of $n$ to avoid potential confusion of notation. Consider a generic scale, $\xi_n = n^2$, free of $j$. It is obvious that $\{R_j / \xi_n \}_{1 \le j \le n}$ remains a super-martingale.
\begin{theorem}
\label{Thm:almostsureR}
There exists a random variable $\tilde{R}$ finite in its mean, such that
$$\frac{R_n}{n^2} \, \overset{a.s.}{\longrightarrow} \, \tilde{R},$$
as $n \to \infty$.
\end{theorem}
\begin{proof}
Notice that $R_j$ is increasing in $j$. So is $R_n / \xi_n$. We thus have
\begin{equation*}
\sup_j \E \left[\left|\frac{R_j}{n^2}\right|\right] = \E\left[\frac{R_n}{n^2}\right] = \frac{2m - 1}{m^2} + O\left(\frac{1}{n}\right) < + \infty.
\end{equation*}
We thus arrive at the conclusion by the {\em Doob's Convergence Theorem}.
\end{proof}
\section{Wiener index}
\label{Sec:Wiener}
In this section, we place focus on the Wiener index of a random caterpillar, i.e., $W_n = \mbox{{\tt Wiener}}(C_n)$. The Wiener index of a graph is defined as the sum of the lengths of the shortest paths (i.e., distances) of all pairs of nodes therein. Mathematically, given $G = (V, E)$, it is
$$W(G) = \sum_{u, v \in V} {\rm dist}(u, v),$$
where ${\rm dist}(u, v)$ is the distance between $u$ and $v$.
In chemistry, the Wiener index was first used to study carbon-carbon bonds between all pairs of carbon atoms in an alkane~\cite{Wiener}, and later was used to characterize QSPR for alkanes~\cite{Platt}. There is a variety of applications of the Wiener index in chemical graph theory and chemoinformatics, not limited to QSAR and QSPR modeling. For the sake of conciseness, we refer the interested readers to~\cite{Nikolic1995} and relevant references therein.
On the other hand, the Wiener index is very popular in the random graph community, and is extensively investigated for many random models, such as random binary search trees and random recursive trees~\cite{Neininger}, balanced binary trees~\cite{Bereg}, random digital trees~\cite{Fuchs}, random split trees~\cite{Munsonius2011}, random $b$-ary trees~\cite{Munsonius}, conditioned Galton-Watson trees~\cite{Fill} and more general rooted and unrooted trees~\cite{Wagner} and simply generated random trees~\cite{Janson}.
The Wiener index of a caterpillar is constituted by three classes of contributions:
\begin{enumerate}
\item The total distances between spine nodes and spine nodes;
\item The total distances between leaves and leaves;
\item The total distances between spine nodes and leaves.
\end{enumerate}
We calculate the three contributions to $W_n$ one after another.
The contribution purely among spine nodes is simple, as it is fixed. That is
$$\sum_{i = 1}^{m - 1} \sum_{j = i + 1}^{m} (j - i) = \frac{1}{2} \sum_{i = 1}^{m - 1} (m + 1 - i)(m - i) = \frac{1}{6}m(m^2 - 1).$$
To compute the contributions between leaves and leaves, we consider the following two scenarios. For all $1 \le i < j \le m$, the distance between a leaf attached to spine node $i$ and a leaf attached to spine node $j$ is $(j - i + 2)$, and the number of pairs is $X_{i, n}X_{j, n}$. On the other hand, for each $1 \le i \le m$, the distance between a leaf attached to spine node $i$ and another (different) leaf attached to the same spine node is $2$, and the number of pairs is $X_i(X_i - 1)/2$. Thus, the total contribution among leaves is
$$
\sum_{i = 1}^{m - 1} \sum_{j = i + 1}^{m} (j - i + 2) X_{i, n}X_{j, n} + \sum_{i = 1}^{m} X_{i, n}(X_{i, n} - 1).
$$
Lastly, we calculate the contributions between leaves and spine nodes. Note that the computation for this class is slightly different from the previous two. We consider the orders of indices $i$ and $j$ to avoid double counting, but it is unnecessary here. For all $1 \le i, j \le m$, the distance between a leaf attached to spine node $i$ and spine node $j$ is $(|i - j| + 1)$, and the number of leaves attached to spine node $i$ is $X_{i, n}$. Hence, the total contribution by this class is given by
$$
\sum_{i = 1}^{m} \sum_{j = 1}^{m} (|j - i| + 1) X_{i, n}.
$$
We are now ready to calculate the mean of the Wiener index. Likewise, we obtain a weak law.
\begin{prop}
\label{Prop:meanW}
For $n \ge 0$, the mean of the Wiener index of a caterpillar is
$$\E[W_n] = \frac{(m^2 +6m - 1)n^2 + (m - 1)(2m^2 + 7m - 1)n + m^2(m^2 - 1)}{6m}.$$
As $n \to \infty$, we have
$$\frac{W_n}{n^2} \, \overset{L_1}{\longrightarrow} \, \frac{m^2 + 6m - 1}{6m}.$$
The convergence takes place in probability as well.
\end{prop}
\begin{proof}
Putting the three types of contributions that we have calculated together, and taking the expectation for all $X_{i, n}$'s, we get
\begin{align*}
\E[W_n] &= \frac{m(m^2 - 1)}{6} + \frac{m(m + 7)(m - 1)}{6} \, \E[X_{1, n}X_{2, n}] + m \, \E\left[X^2_{1, n}\right]
\\ &\qquad{}- m \, \E\left[X_{1, n}\right] + \left(m^2 + 2 \times {m + 1 \choose 3} \right) \E[X_{1, n}]
\\ &= \frac{m(m^2 - 1)}{6} + \frac{(m + 7)(m - 1)n(n - 1)}{6m} + \frac{(n + m - 1)n}{m}
\\ &\qquad{} + \frac{(m + 4)(m - 1)n}{3}
\\ &= \frac{(m^2 +6m - 1)n^2 + (m - 1)(2m^2 + 7m - 1)n + m^2(m^2 - 1)}{6m}.
\end{align*}
The weak law stated in the proposition follows immediately.
\end{proof}
There exists a finite-mean random variable, to which $W_n$ converges, after properly scaled (by $n^2$). The proof is done mutandis mutatis, with an application of the Doob's Convergence Theorem. We thus omit the details, but only state the theorem.
\begin{theorem}
There exists a random variable $\tilde{W}$ finite in its mean, such that
$$\frac{W_n}{n^2} \, \overset{a.s.}{\longrightarrow} \, \tilde{W},$$
as $n \to \infty$.
\end{theorem}
An extension of the Wiener index is the {\em hyper Wiener index}, a relatively new topological index that characterizes molecular structure and feature of more complex chemical compounds. This index was proposed by~\cite{Randic1993} to analyze the structure of 2-Methylhexane, and later on was used in~\cite{Darafsheh} to investigate one-pentagonal carbon nanocone. The hyper Wiener index of $G = (V, E)$ is given by
$$W^{h}(G) = \sum_{u, v \in V} \bigl({\rm dist}(u, v) + {\rm dist}^2(u, v) \bigr).$$
Let $W^h_n$ be the hyper Wiener index of $C_n$. The computation of $W^h_n$ is indeed similar to that of $W_n$. We thus only list the key steps. We again decompose the index into three parts, where the first refers to spine-spine contributions:
$$\sum_{i = 1}^{m - 1} \sum_{j = i + 1}^{m} \bigl((j - i) + (j - i)^2\bigr) = \frac{m(m^3 + 2m^2 - m - 2)}{12},$$
which is deterministic. The second part is leaf-leaf contributions, the expectation of which is
\begin{align*}
&\frac{(m^2 + 11m + 46)(m - 1)m}{12} \, \E[X_{1, n}X_{2, n}] + 3m \left(\E\left[X^2_{1, n}\right] - \E[X_{1, n}]\right)
\\ &= \frac{(m^3 + 10m^2 +35m - 10)n^2 - (m + 10)(m + 1)(m - 1)n}{12m}.
\end{align*}
The last part is contributed by the distances between spine nodes and leaf nodes. Its expectation is
$$\frac{(m^2 + 7m + 18)(m - 1)m}{6} \, \E[X_{i, n}] = \frac{(m^2 + 7m + 18)(m - 1)n}{6}.$$
Putting three parts together, we get the expectation of the hyper Wiener index of $C_n$, presented in the next proposition.
\begin{prop}
For $n \ge 0$, the mean of the hyper Wiener index of a caterpillar is
\begin{align*}
\E\left[W^h_n\right] &= \frac{1}{12m} \bigl((m^3 + 10 m^2 + 35m - 10)n^2 + (2m^3 + 13m^2 + 25m - 10)
\\ &\qquad{}\times (m - 1) n + m^2(m + 2)(m + 1)(m - 1)\bigr).
\end{align*}
As $n \to \infty$, we have
$$\frac{W^h_n}{n^2} \, \overset{L_1}{\longrightarrow} \, \frac{m^3 + 10 m^2 + 35m - 10}{12m}.$$
The convergence takes place in probability as well.
\end{prop}
\section{Numerical experiments}
\label{Sec:numeric}
We conduct a series of numerical experiments to verify the results
developed from Section~\ref{Sec:Hoover} to~\ref{Sec:Wiener}. Given a
fixed $m =200$, we independently generate $R = 500$ replications of
random caterpillars after $n = 5000$ evolutionary steps. For each
simulated caterpillar, its Hoover, Zagreb, Randi\'{c}, Wiener and
hyper Wiener indices are computed and recorded. We evaluate each of
the simulation results one after another.
Our results from Section~\ref{Sec:Hoover} indicate that the Hoover
index of a random caterpillar is a deterministic function of $m$ and
$n$, and hence lacks randomness. From the simulation, we find that
the Hoover index of each generated caterpillar is around $0.4807$,
which is consistent with Proposition~\ref{Prop:meanH}.
Next, we compute the Zagreb index of each generated caterpillar, and
standardize the (random) sample of Zagreb indices according to
Propositions~\ref{Prop:meanZ} and~\ref{Prop:varZ}. The resulting
histogram is depicted in Figure~\ref{Fig:zagreb}. In addition, we
use the kernel method to estimate the density based on the sample of
the standardized Zagreb indices, presented in
Figure~\ref{Fig:zagreb} as well. The simulation results suggest that
the Zagreb index (after
proper scaling) of random caterpillars follows a Gaussian law, which
agrees with Theorem~\ref{Thm:zagreb}. We further confirm the
conclusion via the {\em Shapiro-Wilk} normality test, which yields
that the $p$-value equals $0.5442$.
\begin{figure}[tbp]
\begin{center}
\includegraphics[width = 0.75\textwidth]{zagreb.pdf}
\end{center}
\caption{Histogram of the standardized Zagreb indices of $500$
independently generated random caterpillars with $m = 200$ and
$n = 5000$; the thick blue curve is the estimated density of the
sample.}
\label{Fig:zagreb}
\end{figure}
For the Randi\'{c} index, the sample mean based on the simulated
caterpillars is $0.012$ after scaled by $n^2$. This simulation
result is reasonably close to the theoretical result $0.010$ from
Proposition~\ref{Prop:meanR}.
For the Wiener index, we use the {\tt igraph} package (from {\tt R}
program) to compute the distances among the vertices in the
generated caterpillars. The distance
calculation requires a great deal of computation powers, so we
adjust the
simulation parameters to $R^{\prime} = 500$, $m^{\prime} = 50$ and
$n^{\prime} = 2000$. The average of the Wiener indices (after
properly scaled) of the simulated caterpillars is $9.7732$, almost
identical to the theoretical result $9.7720$ from
Proposition~\ref{Prop:meanW}. Under the same setting of
$R^{\prime}$, $m^{\prime}$ and $n^{\prime}$, we also compute the
hyper Wiener indices. The simulation and
theoretical results (after properly scaled) are respectively given
by $264.7783$ and $264.6214$, which completes the verification.
\section{Concluding remarks}
In this section, we address some concluding remarks, and propose some potential future work as well. We investigate several popular statistical indicies for a class of random caterpillars, including Gini index, Hoover index, Zagreb index, Randi\'{c} index and Wiener index (and its extension, hyper Wiener index). The mean of each index is computed. Specifically, we show that the limit distribution of the Zagreb index of random caterpillars is Gaussian. Topological index of random graphs is a burgeoning research area in the applied probability community. The follow-up study of the present paper may be given to further investigations of the limit distribution of the Wiener index and the analysis of the Randi\'{c} index with a more general parameter $\alpha$. Brand new research in this area is three folded:
\begin{enumerate}
\item Propose novel indices according to practical needs; for instance, we can consider a global metric that captures the total weight of random caterpillars, where the weight can be added to nodes or edges or itself can be temporal.
\item Investigate other statistical indices that are not yet covered in the present paper; for instance, the {\em Hosoya's $Z$ index} that counts the number of matchings in a graph and the {\em Balaban's index} that interprets graph connectivity via the associated distance matrix.
\item Consider other types of random trees or more complex random networks that can be used for modeling molecular structures of chemical compounds, and investigate the relevant topological indices thereof. We will report our results elsewhere.
\end{enumerate}
|
\section{Introduction}
\label{sec:intro}
\DIFdelbegin
\DIFdelend The cosmic star formation history shows a rise of star formation activity in our Universe from high-redshift till redshift $z=2$, known as cosmic noon, and then a decrease towards $z=0$ \citep{Madau14}. Our current understanding is that star formation depends on the availability of $\mathrm{H_2}$ clouds, the raw material for stars \DIFaddbegin \DIFadd{\mbox
\citep{Kennicutt12}
}\DIFaddend . As a galaxy evolves, $\mathrm{H_2}$ clouds are consumed to form stars, and less material is available for the next star formation episode. However, infalling gas can provide a fresh supply which replenishes that consumed by star formation activity.
Morphology and colour often correlate with star formation in galaxies \DIFaddbegin \DIFadd{\mbox
\citep[e.g.][for a review]{RH94}
}\DIFaddend . Elliptical galaxies are dominated by an old stellar population and appear red \DIFaddbegin \DIFadd{\mbox
\citep{Thomas05, Kormendy09}
}\DIFaddend . These elliptical galaxies have run out their star-forming gas; therefore, they have no or low star formation\DIFdelbegin \DIFdel{rate}\DIFdelend . At the opposite end, spiral galaxies have bluer emission, showing signs for young stellar populations and active star formation \DIFaddbegin \DIFadd{\mbox
\citep[see also a review from][]{Kennicutt98}
}\DIFaddend .
Per ``conventional\DIFdelbegin \DIFdel{" }\DIFdelend \DIFaddbegin \DIFadd{'' }\DIFaddend galaxy evolution theory, galaxies evolve from star-forming spiral to quench elliptical. Dust follows this evolutionary sequence, and galaxies become dust-poor at late-times, either through dust destruction or ejection out of the ISM. \DIFdelbegin \DIFdel{As the }\DIFdelend \DIFaddbegin \DIFadd{Galaxy formation models assume that as }\DIFaddend star formation activity decreases\DIFdelbegin \DIFdel{over time}\DIFdelend , mergers and disk instabilities make galaxies more spheroidal \DIFdelbegin \DIFdel{with a decrease in the ISM dust mass}\DIFdelend \DIFaddbegin \DIFadd{\mbox
\citep{Somerville01, Baugh05, Croton06, Croton16}
}\DIFaddend . During such morphological transformations, a fraction of the ISM dust is ejected through SN and AGN feedback. \DIFaddbegin \DIFadd{Therefore, we see a decrease in the ISM dust mass in more elliptical galaxies.
}\DIFaddend
Since the dust content of a galaxy depends on its unique evolutionary history, the observed dust content in the local and high-redshift Universe provides insight into the physics of galaxy formation. Studies of high-redshift galaxies in the far-infrared and submillimetre regime reveal an abundance of star-forming galaxies with massive dust reservoirs \citep{Valiante09}. Dusty galaxies are even found in the very early Universe, including A1689-zD1 at $z=7.5$ with a dust mass of $4 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace \citep{Watson15}, and HFSL3 at $z=6.34$ with a dust mass of $1.3 \times 10^9$ \ensuremath{\mathrm{M}_{\odot}}\xspace \citep{Riechers13}. However, observations also see star-forming galaxies at early and late times with little dust and low metallicity \citep{Fisher14}. The dichotomy between dust-rich and dust-poor galaxies across redshift motivates this work.
Recent galaxy surveys have measured the dust mass function \citep{DEE03, VDE05, Dunne11, Eales09, Clemens13} and many scaling relations between dust and the fundamental properties of galaxies. These include the relationship between dust mass and stellar mass \DIFdelbegin \DIFdel{~\mbox
\citep{Santini14}
and }\DIFdelend \DIFaddbegin \DIFadd{\mbox
\citep{Santini14}
and the }\DIFaddend relationship between dust mass and star formation rate (SFR) \citep{daCunha10, Santini14}. These scaling relations provide constraints for galaxy evolution models that track the dust properties of galaxies.
\DIFaddbegin \DIFadd{Historically, dust in galaxies is commonly investigated using one zone analytical models \mbox
\citep[e.g.,][]{Dwek1998, ZGT08, Valiante09, Asano13, Zhukovska14}
. Such models are critical to assess the various processes of dust production and destruction, especially to reproduce the dust content of specific galaxy populations. However, they generally use a simplistic approach to model gas and stellar evolution in galaxies and do not provide predictions for galaxies within a cosmological volume. Zoom-simulations coupled with self-consistent dust tracking give a more realistic accounting of how dust, metals, gas and stellar populations interact within galaxy evolution framework \mbox
\citep{Bekki15, Aoyama17}
. But they are computationally expensive and tend to focus on reproducing the properties of individual galaxies.
}
\DIFadd{\mbox
\cite{McKinnon16}
was one of the first works to model dust and galaxy co-evolution within a cosmological simulation. They incorporated a detailed dust prescription to a hydrodynamical simulation, including stellar production, grain growth and grain destruction by supernovae shocks. To date, a few semi-analytic galaxy models have included self-consistent dust modelling. These models can reproduce a number of global trends between dust and various galaxy properties across cosmic time, with less detail but more efficient computing cost and time compared to the hydrodynamical simulations \mbox
\citep{Popping17, Vijayan19, Triani20}
.
}
\DIFaddend In this paper, we extend the analysis of \citet{Triani20} to investigate the different characteristics between dust-rich and dust-poor simulated galaxies using the \dustysage\DIFaddbegin \footnote{\url{https://github.com/dptriani/dusty-sage}} \DIFaddend semi-analytic model. \DIFaddbegin \DIFadd{Dust in the ISM forms in stellar ejecta \mbox
\citep{Dwek1998, ZGT08}
, grows further via grain accretion \mbox
\citep{Dwek1998, ZGT08, Draine09}
, and is destroyed by supernovae shocks \mbox
\citep{ZGT08, Slavin15}
. }\DIFaddend \dustysage \DIFdelbegin \DIFdel{includes analytical prescriptions for stellar dust production, dust grain growth in dense molecular clouds, dust destruction by SNe shocks , }\DIFdelend \DIFaddbegin \DIFadd{includes analytical prescriptions for these mechanisms, as well as }\DIFaddend dust locked in stars \DIFaddbegin \DIFadd{(astration)}\DIFaddend , and dust inflows \DIFdelbegin \DIFdel{and outflows. }\DIFdelend \DIFaddbegin \DIFadd{from and outflows to the halo and ejected reservoirs. In the model, dust undergoes further destruction via thermal sputtering in the halo and ejected reservoirs due to their high temperature. }\DIFaddend The primary constraints for dust modelled by \dustysage are the observed dust mass function and dust mass - stellar mass relation at $z=0$. \dustysage includes dust tracking in the ISM, halo, and ejected dust by feedback processes. It provides predictions for the dust mass function and the relation between dust mass and stellar mass at high redshift.
This paper is organised as follows: In Section \ref{sec:dustysage}, we provide a brief description of \dustysage. We study how the dust fraction in each reservoir relates to the stellar mass and morphology in Section \ref{sec:morphologies}. Then we categorise galaxies into groups based on their ISM and halo dust content in Section \ref{sec:classification}. In Section \ref{sec:relation}, we map these groups based on their morphology, star formation activity and stellar mass at redshift $z=0$. We extend the relations to higher redshift, up to $z=3$, in Section \ref{sec:evolution}. Section \ref{sec:discussion} and \ref{sec:conclusion} provide a discussion and conclusion, respectively. Throughout this paper, we assume $h = 0.73$ based on the cosmology used of the Millennium simulation.
\section{The semi-analytical model - Dusty SAGE}
\label{sec:dustysage}
Here we provide only a brief overview of the galaxy and dust model used in this work. \dustysage is built on the more generic galaxy model \texttt{SAGE}\xspace \citep{Croton06, Croton16}, but with a detailed prescription that tracks dust evolution. The summary of both galaxy and dust evolution mechanisms in \texttt{Dusty SAGE} is presented in Figure 1 in \citet{Triani20}. The model follows baryonic growth in dark matter halos taken from an N-body simulation. There are four baryonic reservoirs: \DIFaddbegin \DIFadd{the }\DIFaddend pristine gas, the hot halo, the ISM and an ejected reservoir. In this paper, we run \dustysage on the Millennium simulation \citep{Springel05} and select galaxies with stellar mass $\log M_* = 9 - 12$ \ensuremath{\mathrm{M}_{\odot}}\xspace.
Mass is exchanged among the different baryonic reservoirs across cosmic time. Pristine gas falls into the collapsed dark matter halo and is heated to the virial temperature -- the hot halo reservoir \DIFaddbegin \DIFadd{\mbox
\citep{RO77, WR78}
}\DIFaddend . A fraction of this hot gas then cools into the ISM \DIFaddbegin \DIFadd{\mbox
\citep{SD93}
}\DIFaddend . In the ISM, cold hydrogen differentiates into atomic and molecular hydrogen \DIFaddbegin \DIFadd{\mbox
\citep{HMK79}
}\DIFaddend . Stars then form from molecular clouds \DIFaddbegin \DIFadd{\mbox
\citep{Kennicutt12}
}\DIFaddend . During their evolution, stars create helium and elements heavier than helium in their core, known as `metals'. These metals are expelled in stellar ejecta and change the metallicity of the ISM (ratio of metals and gas mass). Massive stars end their life as supernovae and inject a large amount of energy into the cold ISM. This energy reheats cold ISM gas back to the hot halo reservoir, and potentially even expels it outside the halo -- the `ejected' reservoir \DIFaddbegin \DIFadd{\mbox
\citep[e.g.,][]{MW03}
}\DIFaddend . This process is known as SN feedback. The gas accretion onto supermassive \DIFdelbegin \DIFdel{blackholes }\DIFdelend \DIFaddbegin \DIFadd{black holes }\DIFaddend provides another source of feedback; this energetic phenomenon is known as active galactic nuclei (AGN) \DIFaddbegin \DIFadd{\mbox
\citep{Croton06}
}\DIFaddend . The ejected gas can later be reincorporated back to the halo and galaxy.
\DIFaddbegin \DIFadd{In both }\dustysage \DIFadd{and }\texttt{SAGE}\xspace \DIFadd{\mbox
\citep{Croton06, Croton16}
, the bulge-to-total stellar mass ratio represents the morphology of the galaxy, which evolves during disk instabilities and mergers. We adopt the disk stability criteria of \mbox
\citet{1998MMW}
and transfer sufficient stellar mass from the disk to the bulge to ensure stability every time an instability occurs. In a merger between a more massive central galaxy and a less massive satellite galaxy, bulge enrichment depends on the total stellar and gas mass ratio of both progenitors. If the mass ratio exceeds 0.3, a `major' merger has occurred: the disks of both galaxies are destroyed and all stars are placed in a bulge. Otherwise, the merger is `minor' and the satellite stellar content is added to the central bulge. Both instabilities and mergers can trigger a starburst. Unlike quiescent disk star formation, stars formed in a merger-induced starburst are placed in the bulge.
}
\DIFaddend Besides the usual galaxy evolution processes, \dustysage also incorporates a detailed dust evolution model. Analogous to the gas, we track dust in three reservoirs -- the ISM, the hot halo and an ejected reservoir. \DIFdelbegin \DIFdel{The following processes affect the mass in each of these dust reservoirs :
}\DIFdelend \DIFaddbegin \DIFadd{Dust is mainly processed in the ISM, then heated to the halo and ejected reservoirs via feedback mechanisms powered by supernovae and AGN. The total dust production rate in the ISM is described by:
}\begin{equation} \DIFadd{\label{eq:total_rate}
\dt{M}_\mathrm{d} = \dt{M}_\mathrm{d}^\mathrm{form} + \dt{M}_\mathrm{d}^\mathrm{growth} - \dt{M}_\mathrm{d}^\mathrm{dest} - \dt{M}_\mathrm{d}^\mathrm{SF} - \dt{M}_\mathrm{d}^\mathrm{outflow} + \dt{M}_\mathrm{d}^\mathrm{inflow},
}\end{equation}
\DIFadd{where:
}\DIFaddend \begin{itemize}
\item \DIFdelbegin \DIFdel{Dust formed via condensation of elements in the ejecta of SN II and AGB stars}\DIFdelend \DIFaddbegin \DIFadd{$\dt{M}_\mathrm{d}^\mathrm{form}$ is the stellar dust formation rate: In every star formation episode, }\dustysage \DIFadd{tracks the abundance of C, N and O in AGB stars and C, O, Mg, Si, S, Ca and Fe in SN II ejecta. The condensation of these elements to form dust is given in Table \ref{tab:params}}\DIFaddend .
\item \DIFdelbegin \DIFdel{In the ISM, dust is destroyed by SN blast waves}\DIFdelend \DIFaddbegin \DIFadd{$\dt{M}_\mathrm{d}^\mathrm{growth}$ is the grain growth rate in dense molecular clouds: Existing grains grow via metal accretion \mbox
\citep{Dwek1998, ZGT08}
where the timescale for this process depends on the metal abundance in the cold gas}\DIFaddend .
\item \DIFdelbegin \DIFdel{Dust growth by accretion in molecular clouds.
}\DIFdelend \DIFaddbegin \DIFadd{$\dt{M}_\mathrm{d}^\mathrm{dest}$ is the destruction rate via SN shocks: We follow the prescription from \mbox
\cite{DS80, Mckee89, Asano13}
to compute a destruction timescale from the total cold gas mass and the supernovae rate, efficiency and swept mass.
}\DIFaddend \item \DIFdelbegin \DIFdel{A fraction of dust in the ISM is locked up in stars during each star formation episode}\DIFdelend \DIFaddbegin \DIFadd{$\dt{M}_\mathrm{d}^\mathrm{SF}$ is the rate for dust locked in newly formed stars, which is proportional to the star formation rate and the dust-to-gas ratio}\DIFaddend .
\item \DIFdelbegin \DIFdel{Outflows transport dust out of the ISM to the hot haloand the }\DIFdelend \DIFaddbegin \DIFadd{$\dt{M}_\mathrm{d}^\mathrm{outflow}$ and $\dt{M}_\mathrm{d}^\mathrm{inflow}$ are the dust outflow and inflow rates. SN and AGN feedback can reheat cold ISM gas and expel it to the halo. The feedback energy, if large enough relative to the depth of the potential well, can even eject the gas to leave the galaxy and host halo. In an outflow, we assume that the dust-to-gas ratio of the ejected gas is equal to the ISM. Ejected gas can be reincorporated back to the halo, while maintaining the dust-to-gas ratio of the }\DIFaddend ejected reservoir. \DIFdelbegin
\ite
\DIFdel{Thermal sputtering destroys dust grains in the hot haloand the ejected reservoir}\DIFdelend \DIFaddbegin \DIFadd{Hot gas undergoes cooling process back to the disk, and we assume in this inflow the dust-to-gas ratio equals that of the halo}\DIFaddend .
\DIFdelbegin
\ite
\DIFdel{Dust is reincorporated back from the ejected reservoir to the hot halo.
\ite
\DIFdel{During halo cooling, dust cools down from the haloto }\DIFdelend \DIFaddbegin \end{itemize}
\DIFadd{Dust populates the halo and ejected reservoirs through outflows from the ISM. In both reservoirs, dust grains are destroyed via thermal sputtering on a short timescale that depends on both the density and temperature of each reservoir. In }\dustysage\DIFadd{, we assume that the temperature in both the halo and ejected reservoir reach the virial value. Although thermal sputtering is very efficient, we still find a significant abundance of dust in the halo. Figure 14 in \mbox
\citet{Triani20}
shows that at low-redshift, galaxies have more dust in their halo than in }\DIFaddend the ISM. \DIFdelbegin
\DIFdelend \DIFaddbegin \DIFadd{This outcome roughly agrees with the massive dust found in the circumgalactic medium \mbox
\citep[CGM,][]{Dunne11}
. \mbox
\citet{Popping17}
also predict a significant amount of dust in the halo.
}
\DIFaddend \dustysage provides a good agreement with the galaxy stellar mass function at redshift $z=0$. It also successfully reproduces the dust mass function and various dust scaling relations over a wide range of redshifts \citep{Triani20}. To achieve a more realistic distribution of the stellar mass in the bulge and disc, we make slight changes to a few of the parameters in \citet{Triani20}. These are listed in Table \ref{tab:params}.
\begin{table*}
\caption{Fiducial \dustysage parameters used throughout this work, also compared to those from \citet{Triani20}.}
\label{tab:params}
\begin{tabular}{lccc}
\hline
\hline
Parameter & Description & Value &\citet{Triani20} \\
\hline
\hline
$f_b^\mathrm{cosmic}$ & Cosmic baryon fraction & 0.17 & 0.17\\
\hline
$z_0$ & Redshift when $\mathrm{H_{II}}$ regions overlap & 8.0 & 8.0\\
$z_r$ & Redshift when the intergalactic medium is fully reionized & 7.0 & 7.0\\
\hline
$\alpha_\mathrm{SF}$ & Star formation efficiency from $\mathrm{H_2}$ [$\mathrm{Myr^{-1}}$] & 0.005 & 0.005\\
$R$ & Instanteneous recycling fraction & 0.43 & 0.43\\
\hline
$\epsilon_\mathrm{disc}$ & Mass-loading factor due to supernovae & 2.0 & 3.0\\
$\epsilon_\mathrm{halo}$ & Efficiency of supernovae to unbind gas from the hot halo & 0.2 & 0.3\\
$k_\mathrm{reinc}$ & Velocity scale for gas reincorporation & 0.15 & 0.15\\
\hline
$\kappa_\mathrm{R}$ & Radio mode feedback efficiency & 0.09 & 0.08\\
$\kappa_\mathrm{Q}$ & Quasar mode feedback \DIFdelbeginFL \DIFdelFL{efficency }\DIFdelendFL \DIFaddbeginFL \DIFaddFL{efficiency }\DIFaddendFL & 0.005 & 0.005\\
$f_\mathrm{BH}$ & Rate of black hole growth during quasar mode & 0.015 & 0.015\\
\hline
$f_\mathrm{friction}$ & Threshold subhalo-to-baryonic mass for satellite disruption or merging & 1.0 & 1.0\\
$f_\mathrm{major}$ & Threshold mass ratio for merger to be major & 0.15 & 0.3\\
\hline
$\alpha_\mathrm{burst}$ & Exponent for the powerlaw for starburst fraction in merger & 0.18* & 0.7\\
$\beta_\mathrm{burst}$ & Coefficient for starburst fraction in merger & 0.75* & 0.56\\
\hline
$\delta_\mathrm{C}^\mathrm{AGB}$ & Condensation efficiency for AGB stars & 0.2 & 0.2\\
$\delta_\mathrm{C}^\mathrm{SNII}$ & Condensation efficiency for SN II & 0.15 & 0.15\\
$\tau_\mathrm{acc,0}$ & Accretion timescale for grain growth [$\mathrm{yr}$] & $4.0 \times 10^5$ & $4.0 \times 10^5$\\
$f_\mathrm{SN}$ & Fraction of destroyed dust to the swept dust mass by SN & 0.1 & 0.1\\
\hline
\hline
\multicolumn{3}{l}{*value adopted from \citet{Somerville01}}\\
\end{tabular}
\end{table*}
\section{How morphology and stellar mass relate to dust fraction in the hot halo, ISM and ejected reservoir}
\label{sec:morphologies}
In this section, we investigate the relation between galaxy morphology, represented as the bulge-to-total (BTT) mass ratio, and the fraction of dust in the ISM, halo and ejected reservoirs relative to the total dust mass in all reservoirs. We use stellar mass in computing the BTT mass ratio and assume the gas mass to be negligible. We present the median dust fraction vs BTT mass ratio in three stellar mass bins in Figure \ref{fig:btt-fraction}. The grey histogram shows the distribution of BTT mass ratios across all galaxies. In all panels, galaxies with higher BTT mass ratio contain most of their dust in the halo while those with a lower BTT mass ratio tend to keep theirs in the ISM. For the low mass galaxies in the bottom panel, spiral galaxies contain a significant fraction of their dust in the ISM and ejected reservoir. These galaxies make up more than 50\% of the overall population. Conversely, the halo contains nearly 90\% of the total dust in elliptical galaxies with BTT > $0.8$.
For the galaxies with $\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10 - 11$ (central panel), the same trend occurs. Galaxies with a BTT mass ratio smaller than $\sim 0.2$ keep almost half of their dust in the ISM. For the most massive galaxies (top panel), all galaxies contain most of their dust in the halo. Massive galaxies are more likely to host AGNs compared to low-mass galaxies. The stronger AGN feedback is more likely to blow dust out into the halo (and ejected reservoir) compared to the weaker SN feedback commonly found in low-mass galaxies. Massive galaxies also have deeper potential wells than the low mass systems, allowing them to capture reheated dust in the hot halo. However, although the ISM dust fraction is smaller than in Milky Way sized galaxies, their actual dust mass in the ISM is significant and will likely make a noticeable contribution to the galaxy SED in the infrared.
Observations have found galaxies with a massive amount of dust in their circumgalactic medium (CGM). For a galaxy with stellar mass $\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10.1$, \citet{Peeples14} found a fiducial value of metals locked in CGM dust of $5 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace, almost twice the value for those found in the form of ISM dust, $2.6 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace. \citet{Menard10} derived a CGM dust mass of $5 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace for galaxies with stellar mass $\log \mathrm{M_*} = 10.4 \ensuremath{\mathrm{M}_{\odot}}\xspace$. Although the CGM is not exactly equivalent to our halo gas component, we will use CGM dust as a proxy to test our predictions of the dust bond to, but outside the galaxy.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{images/btt-fraction-mass-512files.png}
\caption{The median value for dust mass fraction in each baryon reservoir versus bulge-to-total stellar mass ratio (BTT). The purple, orange and green lines represents hot dust mass in the halo, cold dust mass in the ISM, and dust in the ejected reservoir, respectively. The grey histogram represents the number distribution of galaxies in 10 BTT bins.}
\label{fig:btt-fraction}
\end{figure*}
\section{Classifying galaxies based on their dust and fundamental properties}
\label{sec:classification}
To investigate how the dust content of galaxies correlates with their stellar mass, star formation activity, and morphology, we first establish our simulated galaxies into groups. In the \dustysage model, the dust in galaxies is distributed in three components - the ISM, the halo and the ejected reservoir. We exclude the dust in the ejected reservoir in our analysis since it is out of the system. Our four dust groups are described below.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/dust-stellarmass-512files.png}
\caption{The cold dust mass of simulated galaxies as a function of stellar mass at redshift $z=0$. The heatmap shows the density distribution with a brighter colour representing higher density, and the black line marks the median. The red line marks the threshold where the cold dust mass per stellar mass equals $10^{-4}$. Above this line, galaxies are classified as ISM dust-rich while below are classified as ISM dust-poor.}
\label{fig:dust-stellarmass}
\end{figure}
\subsection{ISM dust-poor}
\label{ssec:ISM-dust-poor}
When considering dust, observers typically report the properties of the ISM \DIFaddbegin \DIFadd{\mbox
\citep[e.g.,][]{Eales09, RR14, Santini14, Mancini15, daCunha15, Nersesian19}
}\DIFaddend . Although some works have extended dust measurements to include the CGM \DIFaddbegin \DIFadd{\mbox
\citep{Menard10, PMC15}
}\DIFaddend , such observations are rare. Therefore, as a first step, we have classified our model galaxies based on their cold dust content in the ISM. Figure~\ref{fig:dust-stellarmass} shows the relation between cold dust mass in the ISM and stellar mass of our model galaxies at redshift $z=0$. We find that the dust mass in the ISM increases with the host stellar mass, but with a significant scatter.
The median values of our predicted ISM dust mass versus stellar mass are in a good agreement with the observations of dust in local galaxies, such as from the DustPedia catalogue \citep{Nersesian19}. Figure \ref{fig:dust-stellarmass} shows a large scatter below the median values, marking galaxies with relatively less dust compared to the overall population.
We draw an arbitrary line at dust mass per stellar mass of $10^{-4}$, marked with the red line in Figure \ref{fig:dust-stellarmass} to divide the galaxy population into two categories: ISM dust-rich and ISM dust-poor. The ISM dust-poor group contains galaxies below the red line. Observations have found such ISM dust-poor galaxies in both the local and high-redshift Universe \citep{Fisher14}.
Further exploration reveals that \dustysage galaxies in this ISM dust-poor category vary in their halo dust content. Therefore, we additionally divide the ISM dust-poor group based on their fraction of halo dust per stellar mass described below.
\subsubsection{Halo dust-rich}
\label{sssec:halo-dust-rich}
Several observations \citep{PMC15, Peeples14, Menard10} extended their search for metals and dust to the CGM. \citet{Peeples14} find that only $25\%$ of metals created in stars stay in the ISM; a similar fate also occurs for dust. \citet{Menard10} found galaxies with massive CGM dust. These observations \DIFdelbegin \DIFdel{confirm }\DIFdelend \DIFaddbegin \DIFadd{discovered }\DIFaddend the existence of dust out of the ISM, which is \DIFdelbegin \DIFdel{predicted }\DIFdelend \DIFaddbegin \DIFadd{reproduced }\DIFaddend by our model.
In our model, a mix of halo and ejected dust might be a better representation of the CGM for some galaxies. However, we only use halo dust to represent the observed CGM dust. We use the same threshold as the ISM dust-mass to define the halo-dust rich galaxies: halo dust mass per stellar mass of $10^{-4}$. Halo dust-rich galaxies are defined as ISM dust-poor galaxies with halo dust mass above this threshold.
\subsubsection{Dust-poor}
\label{sssec:dust-poor}
The population of galaxies that lack dust in both their ISM and halo is classified as dust-poor. Due to the rarity of dust measurements outside the ISM, we found no counterpart of this category from the observations. Their existence in our model serves as a prediction for future surveys measuring the CGM dust.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/sSFR-hist-allmass-512files.png}
\caption{The distribution of the specific star-formation rate in our ISM dust-rich galaxies. The grey line at $2.71 \times 10^{-11} \mathrm{yr^{-1}}$ is the Milky Way value adopted from \citet{Licquia15} which we use as a threshold in defining quenched and star-forming galaxies in our model galaxies.}
\label{fig:sfr-hist}
\end{figure}
\subsection{ISM dust-rich}
\label{ssec:ISM-dust-rich}
We define the ISM dust-rich galaxies in our model as those with a fraction of ISM dust mass per stellar mass above $10^{-4}$. Dust obscures the intrinsic stellar spectra of such galaxies and re-emits the radiation in the infrared \DIFdelbegin \DIFdel{. Infrared }\DIFdelend \DIFaddbegin \DIFadd{\mbox
\citep{WTC92, WG00}
. Infrared and sub-millimeter }\DIFaddend surveys have found a substantial number of these galaxies, especially at low redshift \DIFaddbegin \DIFadd{\mbox
\citep[e.g.,][]{Eales09, DEE03, daCunha15, Clemens13, RR14}
}\DIFaddend .
Dust accounting in the galactic ISM is the main focus of many current and future infrared surveys. Forthcoming galaxy survey, like those using JWST, will measure dust in high redshift galaxies. Such surveys will provide additional constraints to our model predictions.
Besides the relation between dust mass and stellar mass shown in Figure \ref{fig:dust-stellarmass}, observations have also found a relation between dust mass and SFR \citep{daCunha10, Casey12, Santini14}. Based on this, we divide our ISM dust-rich galaxies further into star-forming and quenched subclasses.
\subsubsection{Star-forming ISM dust-rich}
\label{sssec:sf-dust-rich}
Figure \ref{fig:sfr-hist} shows the specific SFR (sSFR) distribution of the ISM dust-rich galaxies in our model. To separate star-forming from quenched galaxies, we use the Milky Way sSFR value as the threshold, $2.71 \times 10^{-11} \mathrm{yr^{-1}}$ \citep{Licquia15}. Galaxies with sSFR above the Milky Way value are defined as the star-forming ISM dust-rich.
\subsubsection{Quenched ISM dust-rich}
\label{sssec:q-dust-rich}
ISM dust-rich galaxies are commonly related to star-forming spirals since dust is initially formed in stellar ejecta and grows in molecular clouds \DIFaddbegin \DIFadd{\mbox
\citep{Valiante09}
}\DIFaddend . However, we also find a population of ISM dust-rich galaxies with relatively low star formation activity in our model (see Figure \ref{fig:sfr-hist}). The quenched ISM dust-rich galaxies are those with a sSFR below the Milky Way value of $2.71 \times 10^{-11} \mathrm{yr^{-1}}$.
\DIFdelbegin
\DIFdelend \DIFaddbegin \section{The relation between dust mass and fundamental galaxy properties}
\DIFaddend \label{sec:relation}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{images/btt-ssfr-Z-512files.png}
\caption{\textit{Top panel} shows the BTT ratio, sSFR and gas-phase metallicity of our model galaxies in the ISM dust-poor group. This group consists of two subgroups: halo dust-rich and dust-poor. Markers with small, medium and large size represents three mass bin: $\log \mathrm{M_*} = 9 - 10$ \ensuremath{\mathrm{M}_{\odot}}\xspace, $\log \mathrm{M_*} = 10 - 11$, and $\log \mathrm{M_*} = 11 - 12$ \ensuremath{\mathrm{M}_{\odot}}\xspace, respectively. The points mark the median sSFR and median BTT mass ratio of each group, with the errorbars spanning the $16^\mathrm{th}$ and $84^\mathrm{th}$ percentiles. The colour of each point marks the gas-phase metallicity, with brighter colours representing higher value. We assume a solar metallicity of 0.02 and $12 + \log[\mathrm{O/H}] - \log[\mathrm{Z/Z_\odot}] = 9.0$. The number inside parentheses in the legend represents the percentage of galaxies in each category with respect to the total number of galaxies with stellar mass $\log \mathrm{M_*} = 9 - 12$ at $z=0$. Open symbols indicate equivalent observational dataset taken from DustPedia \citep{Nersesian19} and KINGFISH \citep{Kennicutt11, RR14}.
\textit{Bottom panel} shows the same relation above but for the ISM dust-rich group, which is splitted into star-forming and quenched subgroups.}
\label{fig:btt-sfr}
\end{figure*}
In the previous sections, we categorized model galaxies into four classes: star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich and dust-poor. We now plot the BTT mass ratio vs sSFR of each group in three mass bin in Figure \ref{fig:btt-sfr} as indicated in the legend, with the colour distribution representing the median gas-phase metallicity of each group. The error bars denote the $16^\mathrm{th}$ and $84^\mathrm{th}$ percentile range.
Figure~\ref{fig:btt-sfr} shows a general trend where the ISM dust-rich galaxies in the bottom panel occupy the lower right space marking disky galaxies with higher star formation activity. Meanwhile, ISM dust-poor galaxies in the top panel generally have lower sSFR and a larger BTT mass ratio. Since galaxies typically change from star-forming spirals to quenched ellipticals over time, this plot indicates that the ISM also evolves from being dust-rich to dust-poor \DIFaddbegin \DIFadd{across h}\DIFaddend .
The gas-phase metallicity in Figure~\ref{fig:btt-sfr} increase with stellar mass, as one may expect from the galaxy mass-metallicity relation \citep{Tremonti04}. Both ISM dust-rich groups generally have higher metallicity than the ISM dust-poor groups.
Star-forming ISM dust-rich galaxies have the highest sSFR and lowest BTT mass ratio. \DIFdelbegin \DIFdel{This is because galaxies with active star formation allow dust to form via metal condensation in the stellar ejecta. The high sSFRalso }\DIFdelend \DIFaddbegin \DIFadd{Condensation in stellar ejecta is one of the main dust production channels in our model and is tightly correlated with star formation activity. Thus, we would expect ISM dust-rich galaxies to have a high sSFR. It is also possible that the high sSFR of such galaxies }\DIFaddend reflects the abundance of molecular gas \DIFdelbegin \DIFdel{supply }\DIFdelend in the ISM\DIFdelbegin \DIFdel{. Dense molecular clouds and high metallicity in these galaxies enable }\DIFdelend \DIFaddbegin \DIFadd{, which enables }\DIFaddend dust growth via accretion of metal grains\DIFdelbegin \DIFdel{, the dominant mechanism of dust production in
\DIFdel{\mbox
\citep{Triani20}
}\DIFdelend .
Quenched ISM dust-rich galaxies have \DIFdelbegin \DIFdel{the second-highest }\DIFdelend \DIFaddbegin \DIFadd{much lower }\DIFaddend sSFR (on average) and are \DIFdelbegin \DIFdel{still }\DIFdelend \DIFaddbegin \DIFadd{also located }\DIFaddend in the low BTT-ratio regime. Their gas-phase metallicity is higher than that of the star-forming ISM dust-rich galaxies. \DIFdelbegin \DIFdel{The quenched group consists of a more evolved population than }\DIFdelend \DIFaddbegin \DIFadd{As galaxies in this group have less sSFR than those in }\DIFaddend the star-forming group\DIFdelbegin \DIFdel{. These galaxies }\DIFdelend \DIFaddbegin \DIFadd{, their stellar dust production must be less efficient. However, their abundance of dust can be explained by their high availability of gas-phase metals, which can accrete onto existing dust grains in the dense environment. The low sSFR of these galaxies also suggest that they }\DIFaddend lack massive young stars that become SN\DIFaddbegin \DIFadd{, the primary destroyer of dust in the ISM}\DIFaddend . The scarcity of SN \DIFdelbegin \DIFdel{prevents }\DIFdelend \DIFaddbegin \DIFadd{can also prevent }\DIFaddend the dust and refractory elements from being expelled, allowing them to preserve more \DIFaddbegin \DIFadd{dust and }\DIFaddend metals in their ISM\DIFdelbegin \DIFdel{, hence the higher metallicity.
These are a few competing factors for dust. Based on the lower sSFR of the quenched ISM dust-rich group, their stellar dust production is less efficient than the star-forming ISM dust-rich group. However, the dust growth is more efficient because of the availability of refractory elements to accrete onto existing grains.
}\DIFdelend \DIFaddbegin \DIFadd{.
}\DIFaddend
In contrast to the ISM dust-rich galaxies, the ISM dust-poor group have higher BTT \DIFaddbegin \DIFadd{mass ratios }\DIFaddend (i.e., more elliptical)\DIFdelbegin \DIFdel{mass ratios}\DIFdelend , lower average sSFRs and lower gas-phase metallicities. The halo dust-rich group have the highest BTT mass ratio, incorporating the ``bulgiest'' galaxies for all mass bins. Galaxies in this group also have the lowest gas metallicities. To explore this further, \DIFdelbegin \DIFdel{Figure~\ref{fig:hot-cold-ratio} shows }\DIFdelend \DIFaddbegin \DIFadd{we plot }\DIFaddend the distribution of the ratio between the ISM cold gas mass and the hot halo gas mass \DIFaddbegin \DIFadd{in Figure~\ref{fig:hot-cold-ratio}}\DIFaddend . Here, the halo dust-rich group peaks lowest of all groups, showing that galaxies in this group either have less cold gas in their ISM or more gas in their halo. The fact that these galaxies have massive halo dust and hot gas, yet they lack ISM gas and dust, indicates an efficient mechanism to transport their ISM content out of the disk.
Figure~\ref{fig:outflowrate} shows the fraction of galaxies in each group versus outflow rate. \DIFdelbegin \DIFdel{Halo }\DIFdelend \DIFaddbegin \DIFadd{The halo }\DIFaddend dust-rich group dominates for outflow rates above $30\ \ensuremath{\mathrm{M}_{\odot}}\xspace \mathrm{yr^{-1}}$. \DIFdelbegin \DIFdel{The }\DIFdelend \DIFaddbegin \DIFadd{This suggests that the }\DIFaddend efficient outflow in these galaxies \DIFdelbegin \DIFdel{blows }\DIFdelend \DIFaddbegin \DIFadd{can blow }\DIFaddend their ISM content out of the disk, while their low sSFR does not allow them to replenish the metals and dust in the ISM.
Looking back to Figure \ref{fig:btt-sfr}, we see that the total percentage of galaxies in the most massive mass bin ($\log \mathrm{M_*} = 11 - 12$) is $1.47 \%$ and the halo dust-rich group makes up a quarter of them, a more substantial contribution compared to those in the lower mass bins. \DIFdelbegin \DIFdel{Massive }\DIFdelend \DIFaddbegin \DIFadd{In }\dustysage\DIFadd{, massive }\DIFaddend galaxies are likely to have deeper potential wells, \DIFdelbegin \DIFdel{allowing }\DIFdelend \DIFaddbegin \DIFadd{which allow }\DIFaddend them to capture most of the dust outflow from the ISM in the halo\DIFdelbegin \DIFdel{, so }\DIFdelend \DIFaddbegin \DIFadd{. Therefore, }\DIFaddend very little dust escapes into the ejected reservoir \DIFaddbegin \DIFadd{in feedback events}\DIFaddend . Figure~\ref{fig:btt-fraction} also shows that in the most massive galaxies, the majority of the dust mass is in the hot halo, and an only small fraction of their dust fraction is in the ejected reservoir at all BTT mass ratios.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/hot-cold-gas-512files.png}
\caption{The probability distribution of the cold to hot gas mass ratio of our simulated galaxy categories. Red, blue, green and purple lines represent the star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich and dust-poor groups, respectively.}
\label{fig:hot-cold-ratio}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/outflowrate-512files.png}
\caption{The gas outflow rate distribution due to feedback. We assume that the dust-to-gas mass ratio in the outflow is the same as the general ISM.}
\label{fig:outflowrate}
\end{figure}
In the dust-poor category, \DIFdelbegin \DIFdel{the }\DIFdelend \DIFaddbegin \DIFadd{we find an anti-correlation between stellar mass and sSFR. The }\DIFaddend median sSFR of dwarf galaxies is $\log \mathrm{sSFR}=-10.6$, more than two orders of magnitude larger than those in the most massive bin. Galaxies will \DIFdelbegin \DIFdel{eventually }\DIFdelend become dust-poor \DIFdelbegin \DIFdel{once }\DIFdelend \DIFaddbegin \DIFadd{if }\DIFaddend their dust depletion rate exceeds the dust production rate. Since \DIFdelbegin \DIFdel{massive galaxies }\DIFdelend \DIFaddbegin \DIFadd{Figure \ref{fig:dust-stellarmass} shows a positive correlation between dust mass and stellar mass, more massive galaxies are more likely to }\DIFaddend have more dust in their ISM\DIFdelbegin \DIFdel{(Figure \ref{fig:dust-stellarmass}), they become dust-poor at a later stage}\DIFdelend . Therefore, \DIFdelbegin \DIFdel{the dust-poor group in massive galaxies have a lower sSFR on average. }\DIFdelend \DIFaddbegin \DIFadd{only those galaxies with a very low production rate can become dust-poor. The anti-correlation reflects the link between dust and star formation rates.
}\DIFaddend
Figure \ref{fig:btt-sfr} also allows us to examine \DIFdelbegin \DIFdel{trends with mass }\DIFdelend \DIFaddbegin \DIFadd{the stellar mass evolution }\DIFaddend across the four groups. At the massive end, galaxies with mass $\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 11 - 12$ are comprised primarily of the star-forming ISM dust-rich category (0.6 \% of the total population), followed by the quenched ISM dust-rich and halo dust-rich groups each containing 0.3\% of the total galaxies. The dust-poor group is in the minority amongst massive galaxies. The dominance of the ISM dust-rich and halo dust-rich groups shows that the most massive galaxies rarely destroy or eject their dust out of the halo. Massive late-type and early-type galaxies contain the majority of their dust in the ISM and the halo, respectively.
In the middle bin ($\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10 - 11$), the biggest percentage is still contributed by the star-forming ISM dust-rich category (14 \%), followed by the quenched ISM dust-rich with 4.9 \%, halo dust-rich with 4.6 \%, and the dust-poor group with 1.1 \%. In the lowest mass bin ($\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 9 - 10$), the star-forming ISM dust-rich still dominates (28\%). However, the contributions of quenched ISM dust-rich and dust-poor groups in this mass bin are also the most significant (13.9 \% and 12.8 \%, respectively) compared to those in the higher mass bins. The halo dust-rich group shares the smallest percentage for dwarf galaxies with only 3.6\%.
\subsection{Comparison with observed datasets}
\DIFdelbegin \DIFdel{'}\DIFdelend \label{ssec:compared-with-data}
To add context to our theoretical predictions, we make comparisons with two observation datasets. DustPedia \citep{Nersesian19} is a compilation of local galaxies with dust measurements. To create this, the authors use the dust model \texttt{THEMIS} in the SED fitting code \texttt{CIGALE} to infer dust mass and galaxy properties from \textit{Herschel} multiwavelength photometry. The authors use numerical Hubble T as the morphology indicator, with lower value corresponds to a larger stellar composition in bulge relative to the disk. To convert their Hubble T value to BTT-ratio, we use the K-band bulge-to-disk flux ratio mapping from \citet{Graham08}. Note that the BTT ratio for our simulated galaxies is based on stellar mass, while the observations are based on flux.
We group their data based on our definition of the star-forming ISM dust-rich, quenched ISM dust-rich and ISM dust-poor galaxies. We can not divide their ISM dust-poor galaxies further because they do not provide a halo dust mass. We did not find galaxies classified as quenched ISM dust-rich in their dataset. Therefore, we plot the mean values of BTT mass ratio and sSFR of their star-forming ISM dust-rich group as a pink-filled circle and their ISM dust-poor group as a pink-filled square in Figure \ref{fig:btt-sfr}.
Our second observational dataset is taken from the KINGFISH survey \citep{Kennicutt11}. We group KINGFISH galaxies based on our definitions and compute the mean dust mass, BTT mass ratio, and SFR for each group. We plot the mean dust and galaxy properties with markers showing the group they belong. The dust measurement for KINGFISH galaxies is taken from \citet{RR14}.
Our prediction for star-forming ISM dust-rich and ISM dust-poor galaxies are in good agreement with the observed values. In both the KINGFISH and DustPedia data, dust-poor galaxies are those with the most elliptical morphology \DIFdelbegin \DIFdel{. Since the mapping of }\DIFdelend \DIFaddbegin \DIFadd{(E-type) with Hubble T below -2 \mbox
\citep{RR14, Nersesian19}
. However, the map of \mbox
\citet{Graham08}
that we use to convert the }\DIFaddend Hubble type to \DIFdelbegin \DIFdel{bulge-to-disk flux ratio in \mbox
\citet{Graham08}
}\DIFdelend \DIFaddbegin \DIFadd{BTT ratio }\DIFaddend only extends to \DIFdelbegin \DIFdel{S0 galaxies as the earliest type}\DIFdelend \DIFaddbegin \DIFadd{S0-type galaxies (Hubble T = -2). Therefore}\DIFaddend , we use \DIFdelbegin \DIFdel{this value to convert E-type galaxies as well. }\DIFdelend \DIFaddbegin \DIFadd{the Hubble T = -2 conversion for elliptical galaxies with Hubble T < -2. }\DIFaddend The converted BTT \DIFdelbegin \DIFdel{mass }\DIFdelend \DIFaddbegin \DIFadd{flux }\DIFaddend ratio of these galaxies are $0.15$ dex lower than our median \DIFdelbegin \DIFdel{value }\DIFdelend \DIFaddbegin \DIFadd{mass ratio }\DIFaddend but are still within our $16^\mathrm{th}$ percentile.
\section{Evolution of the relations between dust and fundamental properties}
\label{sec:evolution}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{images/btt-ssfr-Z-evol-512files.png}
\caption{The evolution of the relation shown in Figure \ref{fig:btt-sfr} between redshift $z=0$ and redshift $z=3$. For clarity, we focus on the middle mass bin ($\log M* = 10 - 11$ \ensuremath{\mathrm{M}_{\odot}}\xspace). Circle, inverted triangle, diamond and square represent the median value of the star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich and dust-poor groups, respectively.}
\label{fig:btt-sfr-evol}
\end{figure*}
Galaxy evolution \DIFdelbegin \DIFdel{theory predicts }\DIFdelend \DIFaddbegin \DIFadd{models predict }\DIFaddend that galaxies evolve from \DIFdelbegin \DIFdel{high sSFR, low BTT mass ratio to low sSFR, high BTT mass ratio }\DIFdelend \DIFaddbegin \DIFadd{star-forming spiral to quench elliptical }\DIFaddend (i.e., the bottom right to top left in Figure \ref{fig:btt-sfr}) \DIFaddbegin \DIFadd{\mbox
\citep{Somerville01, Baugh05, Croton06, Croton16}
}\DIFaddend . The dust content of the ISM \DIFdelbegin \DIFdel{follow }\DIFdelend \DIFaddbegin \DIFadd{follows }\DIFaddend this evolutionary timeline \DIFdelbegin \DIFdel{and }\DIFdelend \DIFaddbegin \DIFadd{to }\DIFaddend become dust-poor over time. There are two primary depletion \DIFdelbegin \DIFdel{mechanism }\DIFdelend \DIFaddbegin \DIFadd{mechanisms }\DIFaddend for ISM dust: destruction by SN shocks \DIFaddbegin \DIFadd{\mbox
\citep{Slavin15}
}\DIFaddend and ejection by SN and AGN feedback \DIFaddbegin \DIFadd{\mbox
\citep{Popping17}
}\DIFaddend . Ejected dust can be trapped in the halo if the potential well is sufficiently deep, causing the host to end up as \DIFaddbegin \DIFadd{a }\DIFaddend halo dust-rich \DIFdelbegin \DIFdel{galaxies}\DIFdelend \DIFaddbegin \DIFadd{galaxy}\DIFaddend . Otherwise, \DIFdelbegin \DIFdel{the }\DIFdelend \DIFaddbegin \DIFadd{such }\DIFaddend galaxies end up as dust-poor.
To see the evolution of \DIFaddbegin \DIFadd{the }\DIFaddend dust content in galaxies across time, we plot the relation between BTT mass ratio and sSFR of our dust groups from redshift $z=0$ to redshift $z=3$ in Figure \ref{fig:btt-sfr-evol}. Here, we focus on the middle mass bin ($\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10 - 11$) as representative of the Milky Way mass range. Circle, inverted triangle, diamond, and squares represent the star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich, and dust-poor categories, respectively. This plot does not show the evolution of the same group of galaxies. Instead, it shows us the map of galaxies belong to each group at different redshift in the BTT mass ratio and sSFR plane.
Figure~\ref{fig:btt-sfr-evol} shows how the percentage of galaxies in each group changes over time. The fraction of star-forming ISM dust-rich galaxies peaks at $z=1$ then remains at $\approx 25\%$ until $z=0$. The other three groups show lower fractions but continue to increase down to the present day. This evolutionary trend reveals how star-forming ISM dust-rich galaxies might convert into the other groups at lower redshift.
Across redshift, the behaviour of individual groups within the BTT mass ratio - sSFR space also changes. In the BTT mass ratio dimension, the quenched ISM dust-rich group gets diskier towards $z=0$ with a $0.3$ drop, while the other three groups do not show notable changes between $z=3$ and $z=0$. All groups except the quenched ISM dust-rich shows a significant decrease of the median sSFR from $z=3$ to $z=0$. The biggest sSFR decrease occurs in halo dust-rich galaxies, by more than 2 order of magnitude, while the star-forming ISM dust-rich and dust-poor galaxies each undergo a single dex decrease.
In the gas phase metallicity dimension, the halo dust-rich group have the lowest median value across redshift $z=3$ to $z=0$. The median metallicity of both dust-poor and star-forming ISM dust-rich galaxies increases slightly with decreasing redshift. Meanwhile, the median metallicity of the quenched ISM dust-rich group fluctuates with redshift; it increases $\approx 0.1$ dex between $z=3$ and $z=2$, does not change between $z=2$ and $z=1$, then decreases $\approx 0.3$ dex.
\section{Discussion: the behaviour of each group across redshift}
\label{sec:discussion}
From Figure~\ref{fig:btt-sfr} and Figure~\ref{fig:btt-sfr-evol}, we can see that although we divide galaxies into groups solely based on their dust content, the galaxy properties of each group are often distinct. Also, galaxy properties within each group evolve with redshift. Our predictions for how these groups change with time can help us understand the baryonic and dust physics occurring. In this section, we discussed the features of these groups between redshift $z=3$ and the present day.
\subsection{The star-forming ISM dust-rich group}
At all mass bins at redshift $z=0$, the star-forming ISM dust-rich group consistently stays at the lower right region in the BTT mass ratio - sSFR space (see Figure~\ref{fig:btt-sfr}). This indicates that these galaxies are still actively forming stars and have an abundance of molecular gas in their ISM, providing a dust production channel via condensation in stellar ejecta and grain growth in the molecular clouds. The gas-phase metallicity in this group is relatively high compared to those in the halo dust-rich and dust-poor groups. This group dominates the galaxy population for all mass bins at redshift $z=0$.
Across redshift, this group stays in the disky star-forming region (see Figure~\ref{fig:btt-sfr-evol}) and continues to dominate in the overall population. The median gas-phase metallicity is lower at higher redshift. This metallicity increase with decreasing redshift reflects the ability of such galaxies always to retain an abundance of metals in their ISM. Since galaxies in the star-forming ISM dust-rich group have high sSFR by \DIFaddbegin \DIFadd{the }\DIFaddend selection, we expect that the ISM will be both metal and dust-rich. Observations of the local and high redshift Universe find many galaxies that belong to this group. \DIFdelbegin \DIFdel{At }\DIFdelend \DIFaddbegin \DIFadd{Figure \ref{fig:btt-sfr} shows that at }\DIFaddend redshift $z=0$, both the KINGFISH~\citep{Kennicutt11, RR14} and DustPedia~\citep{Nersesian19} datasets have galaxies with dust-to-stellar mass and sSFR classifications similar to our star-forming ISM dust-rich group.
\DIFdelbegin \DIFdel{At high redshift }\DIFdelend \DIFaddbegin
\DIFadd{Figure \ref{fig:dtm-ssfr} shows the sSFR of our model galaxies as a function of dust-to-stellar mass ratio from redshift $z=0$ to $z=3$. The white line marks the divider between the ISM dust-rich and ISM dust-poor groups. Across redshift}\DIFaddend , many authors \DIFdelbegin \DIFdel{\mbox
\citep{Riechers13, Santini14, daCunha15, Watson15}
}\DIFdelend find star-forming galaxies with a \DIFdelbegin \DIFdel{dust-rich ISM. Unfortunately, most of them are observed in }\DIFdelend \DIFaddbegin \DIFadd{dust rich ISM. The general trend that the sSFR decreases with decreasing redshift is seen both for the model and in the observations. At $z=0$, our ISM dust-rich galaxies are in good agreement with the observational values from DustPedia \mbox
\citep{Nersesian19}
, \mbox
\citet{RR14}
and \mbox
\citet{Santini14}
. For the ISM dust poor galaxies below the grey line, the observations are lower than our median value but are still within the $16^\mathrm{th}$ percentile. At higher redshift, observations are yet to find ISM dust-poor galaxies. At $z=1$, our model roughly agrees with }\DIFaddend the \DIFdelbegin \DIFdel{infrared regime, }\DIFdelend \DIFaddbegin \DIFadd{dust-rich data from \mbox
\citet{Santini14}
, although our median value of sSFR lies $0.1 - 0.4$ dex below the observed values. At $z=2$ and $z=3$, however, our model fails to reproduce galaxies }\DIFaddend with \DIFdelbegin \DIFdel{only limited measurementa of morphology or BTT mass ratio. Therefore, we can not plot such high-redshift observed galaxies against our predictions in Figure~\ref{fig:btt-sfr-evol}. }\DIFdelend \DIFaddbegin \DIFadd{$\log (\mathrm{M_{dust}} / \mathrm{M_{star}}) > -2.5$ where most of the observational datasets lie. }\dustysage \DIFadd{and our base model }\texttt{SAGE}\xspace \DIFadd{\mbox
\citep{Croton16}
are constrained using various galaxy observations at $z=0$; therefore, we still find limitations in reproducing the high redshift population. We also note that the dataset from \mbox
\citet{daCunha15}
(red triangles) is obtained in the ALMA LESS survey which is biased towards the extremely bright galaxies.
}\DIFaddend
\DIFaddbegin \begin{figure*}
\centering
\includegraphics[width=1\textwidth]{images/dtm-ssfr-millennium.png}
\caption{\DIFaddFL{The relation between specific star formation rate (sSFR) and the dust-to-stellar mass ratio from redshift $z=0$ to $z=3$. The heat map shows the 2D density distribution of our model galaxies with brighter color representing higher density. The solid black lines mark the median while the dashed lines marked the $16^\mathrm{th}$ and $84^\mathrm{th}$ percentile. The white vertical lines mark a log dust-to-stellar mass ratio = -4, where we divide the ISM dust-rich and dust-poor categories. The red circles, blue diamonds, green squares and red triangles are observational values from DustPedia: \mbox
\citet{Nersesian19}
, \mbox
\citet{RR14}
, \mbox
\citet{Santini14}
and \mbox
\citet{daCunha15}
, respectively.}}
\label{fig:dtm-ssfr}
\end{figure*}
\DIFaddend \subsection{The quenched ISM dust-rich group}
The quenched ISM dust-rich group consists of galaxies that have an abundance of dust in their ISM despite their low star formation activity \DIFdelbegin \DIFdel{. These galaxies might still have stellar dust production in the ejecta of SN II and AGB stars from previous star formation episodes. However, they }\DIFdelend \DIFaddbegin \DIFadd{(defined as galaxies with sSFR below the Milky Way). One possible scenario for this is that due to the low sSFR, these galaxies }\DIFaddend no longer host supernovae \DIFaddbegin \DIFadd{events }\DIFaddend that destroy dust \DIFdelbegin \DIFdel{. Therefore, }\DIFdelend \DIFaddbegin \DIFadd{and expel metals and dust from the ISM. This also explains the higher gas-phase metallicity of }\DIFaddend galaxies in this group \DIFdelbegin \DIFdel{are more metal-rich than }\DIFdelend \DIFaddbegin \DIFadd{compared to }\DIFaddend those in the star-forming ISM dust-rich group. The abundance of metals in the ISM provides material for dust growth via grain accretion, allowing \DIFdelbegin \DIFdel{metal-rich }\DIFdelend \DIFaddbegin \DIFadd{such }\DIFaddend galaxies to accumulate \DIFdelbegin \DIFdel{more dust .
}\DIFdelend \DIFaddbegin \DIFadd{dust mass. Another scenario for their dust production is the condensation in the ejecta from previous star formation episodes. Although they do not have a fresh supply of stellar ejecta due to the low sSFR, it is possible that there are still refractory elements from previous SN and AGB winds in the ISM that condense into dust grains.
}\DIFaddend
Figure~\ref{fig:btt-sfr} shows that the majority of quenched ISM dust-rich galaxies can be found in the low mass population. Compared to the star-forming ISM dust-rich group, this group is rare at low and high redshift. Figure~\ref{fig:btt-sfr-evol} shows that the fraction of this group with stellar mass $\log \mathrm{M_*} = 10-11$ at $z > 0$ is always below $6\%$. Due to their low abundance, we have not found any observational evidence for this group.
\subsection{The halo dust-rich group}
Halo dust-rich galaxies have little or no dust in their ISM but contain a significant amount of dust in their halo. At redshift $z=0$, this group includes elliptical galaxies with the largest BTT mass ratios and low sSFR. The median gas-phase metallicity of this group is also the lowest compared to the other groups. The small sSFR does not allow for efficient instantaneous metal and stellar dust production in such galaxies. However, they should have accumulated metals and dust from previous star-formation episodes. The lack of metals and dust in their ISM, while their halo is dust-rich, indicates that the ISM content is transported into the halo (see Section~\ref{sec:relation}).
At higher redshift, however, the halo dust-rich group occupies an increasingly high sSFR region with a median $\log \mathrm{sSFR} >$ the Milky Way value. The high sSFR of these galaxies means that they are producing metals and dust in their stellar ejecta. However, unlike galaxies in the other groups, galaxies in this group show no increase in gas metallicity with decreasing redshift. The low abundance of metals and dust in their ISM is possible if \DIFaddbegin \DIFadd{the }\DIFaddend feedback is efficiently heating and blowing out the newly formed metals and dust. Such feedback might be provided by AGN activity \citep{Sarangi19}.
Galaxies in a deep potential well can retain the feedback-heated gas, dust and metals in their halo. Mass in the dark matter halo usually sets the depth of the potential well. Since stellar mass correlates with halo mass, halo dust-rich galaxies should be more common in the massive galaxies. Figure \ref{fig:btt-sfr} shows that in the massive bin, this group comprise a quarter of the population, a more significant fraction than the lower stellar mass bin.
\citet{Menard10} measured dust reddening effects of background quasars relative to the foreground SDSS galaxies. They observed the existence of diffuse dust in halos with an amount comparable to those in the galactic disk. For $0.5L^*$ galaxies, the authors estimated a dust mass of $5 \times 10^7 \ensuremath{\mathrm{M}_{\odot}}\xspace$ in the halo. \citet{PMC15}
confirmed this halo dust mass for $0.1L^* - 1L^*$ low redshift galaxies. Observations of CGM dust provides evidence for our prediction that dust is transported out of the ISM and trapped in the halo. However, such observations are still uncommon. Future CGM dust measurements will allow for better constraints for models such as \dustysage.
\subsection{The dust-poor group}
We define galaxies that lack dust in both the ISM and their halo as dust-poor galaxies. At redshift $z=0$ (Figure~\ref{fig:btt-sfr}), the dust-poor group has low median sSFR, low gas-phase metallicity and high BTT mass ratio. In the DustPedia~\citep{Nersesian19} and KINGFISH~\citep{Kennicutt11, RR14} datasets, galaxies with morphological type E-S0 have a median dust-to-stellar mass ratio below $1 \times 10^{-4}$ and are classified as ISM dust-poor as per our definition. However, both datasets do not provide measurement for dust in the halo. Therefore, we assume their halo is dust-poor, and we include them in the dust-poor group instead of the halo dust-rich group.
Dust depletion in the ISM occurs in two ways: destruction by SN shocks inside the galaxy disk and dust that is ejected out of the ISM by feedback. If the potential well of the system is enough to trap the dust, the ejected dust will end up in halo. However, galaxies with low stellar mass usually live in shallower potentials and thus are likely to lose their dust completely through ejection. This likely explains why Figure~\ref{fig:btt-sfr} shows that the dust-poor group contain the most significant fraction of their galaxies in the low mass bin.
In Figure \ref{fig:btt-sfr-evol}, the dust-poor group moves to the lower BTT mass ratio, lower sSFR region with decreasing redshift while the gas metallicity increases. At high redshift, galaxies in this group have relatively high sSFR, but their ISM still lacks both metals and dust. Dust formation is highly sensitive to the abundance of metals. In stellar ejecta, dust forms via the condensation of metals ejected from stars. In dense clouds, dust grains grow by accreting metals. Therefore, galaxies with low metallicity can not produce dust efficiently. If this low production rate can not keep up with the depletion rate, a galaxy will end up in the dust-poor group.
\citet{Fisher14} observed the dust emission from a local dwarf galaxy, I Zwicky 18, and found it lacks dust with a dust-to-stellar mass ratio of about $10^{-5}$ to $10^{-6}$. This galaxy has a SFR of $0.05$ \ensuremath{\mathrm{M}_{\odot}}\xspace/year and a very low \DIFdelbegin \DIFdel{stellar }\DIFdelend metallicity of $12+\log\mathrm{[O/H]} = 7.17$\DIFdelbegin \DIFdel{. The observed properties of }\DIFdelend \DIFaddbegin \DIFadd{, much lower than that seen in local galaxies \mbox
\citep{Fisher14}
. The reported stellar mass is $9 \times 10^7 \ensuremath{\mathrm{M}_{\odot}}\xspace$, which makes the sSFR $ = 5.5 \times 10^{-10}$, consistent with the median sSFR value of the dust-poor galaxies in the lower mass group from }\dustysage \DIFadd{(see Figure \ref{fig:btt-sfr}). Although the stellar mass of }\DIFaddend I Zwicky 18 \DIFdelbegin \DIFdel{agree well with our description of galaxies in the dust-poor galaxy group}\DIFdelend \DIFaddbegin \DIFadd{is well below our sample, we find a consistent feature where a lack of metals accompanies the lack of dust}\DIFaddend .
\section{Conclusion}
\label{sec:conclusion}
During galaxy's evolution, dust physics is heavily influenced by star formation, gas physics, outflows and the depth of the potential well. Therefore, the dust, stellar, and gas content of galaxies are tightly related to each other. In this paper, we use the \dustysage galaxy formation model \citep{Triani20} to explore the relation between dust content and the fundamental properties of galaxies. We have divided the galaxy population based on their dust mass and observable properties. At redshift $z=0$, the properties of our model galaxies are in good agreement with observations, especially where the statistics are sound.
Our investigation includes the comparison of morphology, specific star formation rate, metallicity and stellar mass for galaxies grouped by their dust content from redshift $z=3$ to $z=0$. Our main conclusions are the following:
\begin{itemize}
\item We find that the distribution of dust in the ISM, hot halo and ejected reservoir is affected by galaxies' morphology and stellar mass. The fraction of dust in the hot halo increases with BTT mass ratio and stellar mass (Figure \ref{fig:btt-fraction}).
\item At redshift $z=0$, ISM dust-rich galaxies have the highest gas-phase metallicities. Both star-forming and quenched ISM dust-rich groups shows a low BTT mass ratio and relatively high median sSFR. Galaxies in the halo dust-rich and dust-poor groups are metal poor. They are distributed in the high BTT mass ratio and low sSFR regime of Figure~\ref{fig:btt-sfr}.
\item We see an evolution in the behaviour of our grouped galaxies in Figure~\ref{fig:btt-sfr-evol}. From $z=3$ to $z=0$: (i) the BTT mass ratio of the quenched ISM dust-rich group drops $\approx 0.3$ dex while the other three groups show insignificant changes, (ii) the median sSFR of all groups except quenched ISM dust-rich trend towards lower region, and (iii) all groups show weak evolution of gas-phase metallicity with redshift except the quenched ISM dust-rich group where the metallicity fluctuates with redshift.
\item Across redshift, both star-forming ISM dust-rich and quenched ISM dust-rich groups maintain their high median sSFR and high gas-phase metallicity. This implies that such galaxies are still actively producing metals and dust via stellar production. The abundance of metals in the ISM also enables dust growth via metal accretion in dense clouds.
\item The halo dust-rich group consists of galaxies with a dust-poor ISM but dust-rich halo. They have a relatively high outflow rate of heated gas, dust, and metals out of the ISM. At present, they are elliptical with very low sSFR and metallicity. The gas-phase metallicity of these galaxies does not depend significantly on redshift. At high redshift, galaxies in this group have high sSFR that enables an efficient dust production mechanism. Their lack of ISM dust, therefore, implies effective feedback to reheat the newly formed dust. Because these massive galaxies have a deep potential well, they are more likely to retain the heated dust in their halo. Figure~\ref{fig:btt-sfr} and Figure \ref{fig:btt-sfr-evol} shows that this group is made up of a larger fraction of massive galaxies.
\item Galaxies in the dust-poor group have low dust mass both in their ISM and halo. From redshift $z=3$ to $z=0$, this group evolves from the high sSFR region to the lower sSFR region. However, they always have low metallicity. As metals are the main ingredients for dust, this low metallicity implies that their dust production is also small.
\end{itemize}
Our model provides predictions for future surveys with next-generation instruments and telescopes, such as ALMA and JWST, that will measure the dust and galaxy properties in extraordinary detail at high redshift. \DIFaddbegin \DIFadd{JWST will cover the IR spectrum in the 0.6 to 28.5 micron regime, which reveals the total galactic dust content up to $z=2$. At $z=3$, it will only cover the dust spectrum to $\approx 7$ micron, providing a measure for the mass of hot dust. As we have seen throughout this paper, current dust measurement at high-redshift are only available for the brightest galaxies with high dust mass. Complemented with data from ALMA and other IR and sub-mm surveys, JWST will provide a significantly improved view of high redshift dust across a wider range of masses. This will better constraint dust enrichment scenarios in models such as }\DIFaddend \dustysage\DIFdelbegin \DIFdel{is publicly available at
\DIFdel{. }\DIFdelend \DIFaddbegin \DIFadd{, leading to a better predictions and more detailed interpretation of the observations.
}\DIFaddend
\section*{Acknowledgements}
We would like to thank \DIFaddbegin \DIFadd{the anonymous referee whose valuable comments improved the quality of this paper, and }\DIFaddend Ned Taylor for helpful comments during the final stages of this work. This research \DIFdelbegin \DIFdel{were }\DIFdelend \DIFaddbegin \DIFadd{was }\DIFaddend supported by the Australian Research Council Centre of Excellence for All Sky Astro-physics in 3 Dimensions (ASTRO 3D), through project number CE170100013. The Semi-Analytic Galaxy Evolution (SAGE) model, on which \dustysage was built, is a publicly available codebase that runs on the dark matter halo trees of a cosmological N-body simulation. It is available for download at \url{https://github.com/darrencroton/sage}. This research has used \texttt{python} (\url{https://www.python.org/}), \texttt{numpy} \citep{Vanderwalt11} and \texttt{matplotlib} \citep{Hunter07}.
\section*{Data availability}
The data underlying this article are available in the article.
\bibliographystyle{mnras}
\section{Introduction}
\label{sec:intro}
The cosmic star formation history shows a rise of star formation activity in our Universe from high-redshift till redshift $z=2$, known as cosmic noon, and then a decrease towards $z=0$ \citep{Madau14}. Our current understanding is that star formation depends on the availability of $\mathrm{H_2}$ clouds, the raw material for stars \citep{Kennicutt12}. As a galaxy evolves, $\mathrm{H_2}$ clouds are consumed to form stars, and less material is available for the next star formation episode. However, infalling gas can provide a fresh supply which replenishes \DIFdelbegin \DIFdel{that }\DIFdelend \DIFaddbegin \DIFadd{gas that was }\DIFaddend consumed by star formation activity \DIFaddbegin \DIFadd{\mbox
\citep[e.g][]{Dekel06, Almeida14}
}\DIFaddend .
Morphology and colour often correlate with star formation in galaxies \citep[e.g.][for a review]{RH94}. Elliptical galaxies are dominated by an old stellar population and appear red \citep{Thomas05, Kormendy09}. These elliptical galaxies have run out their star-forming gas; therefore, they have no or low star formation. At the opposite end, spiral galaxies have bluer emission, showing signs for young stellar populations and active star formation \citep[see also a review from][]{Kennicutt98}.
Per ``conventional'' galaxy evolution theory, galaxies evolve from star-forming spiral \DIFdelbegin \DIFdel{to quench elliptical}\DIFdelend \DIFaddbegin \DIFadd{galaxies to quenched ellipticals \mbox
\citep{Bundy06, Skelton12, Tojeiro13}
}\DIFaddend . Dust follows this evolutionary sequence, and galaxies become dust-poor at late-times, either through dust destruction or ejection out of the ISM. Galaxy formation models assume that as star formation activity decreases, mergers and disk instabilities make galaxies more spheroidal \mbox{
\citep{Somerville01, Baugh05, Croton06, Croton16}} During such morphological transformations, a fraction of the ISM dust is ejected through SN and AGN feedback. Therefore, we see a decrease in the ISM dust mass in more elliptical galaxies.
Since the dust content of a galaxy depends on its unique evolutionary history, the observed dust content in the local and high-redshift Universe provides insight into the physics of galaxy formation. Studies of high-redshift galaxies in the far-infrared and submillimetre regime reveal an abundance of star-forming galaxies with massive dust reservoirs \citep{Valiante09}. Dusty galaxies are even found in the very early Universe, including A1689-zD1 at $z=7.5$ with a dust mass of $4 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace \citep{Watson15}, and HFSL3 at $z=6.34$ with a dust mass of $1.3 \times 10^9$ \ensuremath{\mathrm{M}_{\odot}}\xspace \citep{Riechers13}. However, observations also see star-forming galaxies at early and late times with little dust and low metallicity \citep{Fisher14}. The dichotomy between dust-rich and dust-poor galaxies across redshift motivates this work.
Recent galaxy surveys have measured the dust mass function \citep{DEE03, VDE05, Dunne11, Eales09, Clemens13} and many scaling relations between dust and the fundamental properties of galaxies. These include the relationship between dust mass and stellar mass \citep{Santini14} and the relationship between dust mass and star formation rate (SFR) \citep{daCunha10, Santini14}. These scaling relations provide constraints for galaxy evolution models that track the dust properties of galaxies.
Historically, dust in galaxies is commonly investigated using one zone analytical models \citep[e.g.,][]{Dwek1998, ZGT08, Valiante09, Asano13, Zhukovska14}. Such models are critical to assess the various processes of dust production and destruction, especially to reproduce the dust content of specific galaxy populations. However, they generally use a simplistic approach to model gas and stellar evolution in galaxies and do not provide predictions for galaxies within a cosmological volume. Zoom-simulations coupled with self-consistent dust tracking give a more realistic accounting of how dust, metals, gas and stellar populations interact within galaxy evolution framework \citep{Bekki15, Aoyama17}. But they are computationally expensive and tend to focus on reproducing the properties of individual galaxies.
\cite{McKinnon16} was one of the first works to model dust and galaxy co-evolution within a cosmological simulation. They incorporated a detailed dust prescription to a hydrodynamical simulation, including stellar production, grain growth and grain destruction by supernovae shocks. To date, a few semi-analytic galaxy models have included self-consistent dust modelling. These models can reproduce a number of global trends between dust and various galaxy properties across cosmic time, with less detail but more efficient computing cost and time compared to the hydrodynamical simulations \citep{Popping17, Vijayan19, Triani20}.
In this paper, we extend the analysis of \citet{Triani20} to investigate the different characteristics between dust-rich and dust-poor simulated galaxies using the \dustysage\footnote{\url{https://github.com/dptriani/dusty-sage}} semi-analytic model. Dust in the ISM forms in stellar ejecta \citep{Dwek1998, ZGT08}, grows further via grain accretion \DIFdelbegin \DIFdel{\mbox
\citep{Dwek1998, ZGT08, Draine09}
}\DIFdelend \DIFaddbegin \DIFadd{\mbox
\citep{Draine90, Dwek1998, ZGT08, Draine09}
}\DIFaddend , and is destroyed by supernovae shocks \DIFdelbegin \DIFdel{\mbox
\citep{ZGT08, Slavin15}
}\DIFdelend \DIFaddbegin \DIFadd{\mbox
\citep{DS80, Jones94, ZGT08, Slavin15}
}\DIFaddend . \dustysage includes analytical prescriptions for these mechanisms, as well as dust locked in stars (astration), and dust inflows from and outflows to the halo and ejected reservoirs. In the model, dust undergoes further destruction via thermal sputtering in the halo and ejected reservoirs due to their high temperature. The primary constraints for dust modelled by \dustysage are the observed dust mass function and dust mass - stellar mass relation at $z=0$. \dustysage includes dust tracking in the ISM, halo, and ejected dust by feedback processes. It provides predictions for the dust mass function and the relation between dust mass and stellar mass at high redshift.
This paper is organised as follows: In Section \ref{sec:dustysage}, we provide a brief description of \dustysage. We study how the dust fraction in each reservoir relates to the stellar mass and morphology in Section \ref{sec:morphologies}. Then we categorise galaxies into groups based on their ISM and halo dust content in Section \ref{sec:classification}. In Section \ref{sec:relation}, we map these groups based on their morphology, star formation activity and stellar mass at redshift $z=0$. We extend the relations to higher redshift, up to $z=3$, in Section \ref{sec:evolution}. Section \ref{sec:discussion} and \ref{sec:conclusion} provide a discussion and conclusion, respectively. Throughout this paper, we assume $h = 0.73$ based on the cosmology used of the Millennium simulation.
\section{The semi-analytical model - Dusty SAGE}
\label{sec:dustysage}
Here we provide only a brief overview of the galaxy and dust model used in this work. \dustysage is built on the more generic galaxy model \texttt{SAGE}\xspace \citep{Croton06, Croton16}, but with a detailed prescription that tracks dust evolution. The summary of both galaxy and dust evolution mechanisms in \texttt{Dusty SAGE} is presented in Figure 1 in \citet{Triani20}. The model follows baryonic growth in dark matter halos taken from an N-body simulation. There are four baryonic reservoirs: the pristine gas, the hot halo, the ISM and an ejected reservoir. In this paper, we run \dustysage on the Millennium simulation \citep{Springel05} and select galaxies with stellar mass $\log M_* = 9 - 12$ \ensuremath{\mathrm{M}_{\odot}}\xspace.
Mass is exchanged among the different baryonic reservoirs across cosmic time. Pristine gas falls into the collapsed dark matter halo and is heated to the virial temperature -- the hot halo reservoir \citep{RO77, WR78}. A fraction of this hot gas then cools into the ISM \citep{SD93}. In the ISM, cold hydrogen differentiates into atomic and molecular hydrogen \citep{HMK79}. Stars then form from molecular clouds \citep{Kennicutt12}. During their evolution, stars create helium and elements heavier than helium in their core, known as `metals'. These metals are expelled in stellar ejecta and change the metallicity of the ISM (ratio of metals and gas mass). Massive stars end their life as supernovae and inject a large amount of energy into the cold ISM. This energy reheats cold ISM gas back to the hot halo reservoir, and potentially even expels it outside the halo -- the `ejected' reservoir \citep[e.g.,][]{MW03}. This process is known as SN feedback. The gas accretion onto supermassive black holes provides another source of feedback; this energetic phenomenon is known as active galactic nuclei (AGN) \citep{Croton06}. The ejected gas can later be reincorporated back to the halo and galaxy.
In both \dustysage and \texttt{SAGE}\xspace \citep{Croton06, Croton16}, the bulge-to-total stellar mass ratio represents the morphology of the galaxy, which evolves during disk instabilities and mergers. We adopt the disk stability criteria of \citet{1998MMW} and transfer sufficient stellar mass from the disk to the bulge to ensure stability every time an instability occurs. In a merger between a more massive central galaxy and a less massive satellite galaxy, bulge enrichment depends on the total stellar and gas mass ratio of both progenitors. If the mass ratio exceeds 0.3, a `major' merger has occurred: the disks of both galaxies are destroyed and all stars are placed in a bulge. Otherwise, the merger is `minor' and the satellite stellar content is added to the central bulge. Both instabilities and mergers can trigger a starburst. Unlike quiescent disk star formation, stars formed in a merger-induced starburst are placed in the bulge.
Besides the usual galaxy evolution processes, \dustysage also incorporates a detailed dust evolution model. Analogous to the gas, we track dust in three reservoirs -- the ISM, the hot halo and an ejected reservoir. Dust is mainly processed in the ISM, then heated to the halo and ejected reservoirs via feedback mechanisms powered by supernovae and AGN. The total dust production rate in the ISM is described by:
\begin{equation} \label{eq:total_rate}
\dt{M}_\mathrm{d} = \dt{M}_\mathrm{d}^\mathrm{form} + \dt{M}_\mathrm{d}^\mathrm{growth} - \dt{M}_\mathrm{d}^\mathrm{dest} - \dt{M}_\mathrm{d}^\mathrm{SF} - \dt{M}_\mathrm{d}^\mathrm{outflow} + \dt{M}_\mathrm{d}^\mathrm{inflow},
\end{equation}
where:
\begin{itemize}
\item $\dt{M}_\mathrm{d}^\mathrm{form}$ is the stellar dust formation rate: In every star formation episode, \dustysage tracks the abundance of C, N and O in AGB stars and C, O, Mg, Si, S, Ca and Fe in SN II ejecta. The condensation of these elements to form dust is given in Table \ref{tab:params}.
\item $\dt{M}_\mathrm{d}^\mathrm{growth}$ is the grain growth rate in dense molecular clouds: Existing grains grow via metal accretion \citep{Dwek1998, ZGT08} where the timescale for this process depends on the metal abundance in the cold gas.
\item $\dt{M}_\mathrm{d}^\mathrm{dest}$ is the destruction rate via SN shocks: We follow the prescription from \cite{DS80, Mckee89, Asano13} to compute a destruction timescale from the total cold gas mass and the supernovae rate, efficiency and swept mass.
\item $\dt{M}_\mathrm{d}^\mathrm{SF}$ is the rate for dust locked in newly formed stars, which is proportional to the star formation rate and the dust-to-gas ratio.
\item $\dt{M}_\mathrm{d}^\mathrm{outflow}$ and $\dt{M}_\mathrm{d}^\mathrm{inflow}$ are the dust outflow and inflow rates. SN and AGN feedback can reheat cold ISM gas and expel it to the halo. The feedback energy, if large enough relative to the depth of the potential well, can even eject the gas to leave the galaxy and host halo. In an outflow, we assume that the dust-to-gas ratio of the ejected gas is equal to the ISM. Ejected gas can be reincorporated back to the halo, while maintaining the dust-to-gas ratio of the ejected reservoir. Hot gas undergoes cooling process back to the disk, and we assume in this inflow the dust-to-gas ratio equals that of the halo.
\end{itemize}
Dust populates the halo and ejected reservoirs through outflows from the ISM. In both reservoirs, dust grains are destroyed via thermal sputtering on a short timescale that depends on \DIFdelbegin \DIFdel{both the }\DIFdelend \DIFaddbegin \DIFadd{the gas }\DIFaddend density and temperature\DIFdelbegin \DIFdel{of each reservoir. In
\DIFdel{, we assume that the temperature in both the halo and ejected reservoir reach the virial value. }\DIFdelend \DIFaddbegin \DIFadd{. In both the halo and ejected reservoirs we assume the virial temperature, with an isothermal density profile for the hot gas extending to the virial radius, and a uniform density profile for the ejected. However, the computed timescale should be taken as the lower limit for the ejected reservoir. In reality, the ejected reservoir is a mix of the circumgalactic medium (CGM) and intergalactic medium (IGM). Gas in the IGM might extend beyond the virial radius, resulting in a lower density. The temperature might also be higher than in the model. Both factors would allow for a more efficient sputtering.
}
\DIFaddend Although thermal sputtering is very efficient, we still find a significant abundance of dust in the halo. \DIFaddbegin \DIFadd{Dust properties in the halo depend on the balance between the outflow of materials from the ISM and the depth of the galaxy's potential well. We assume that the gas outflow carries dust in the same proportion as the ISM and transfers it to the halo, with no dust destruction in the process. It may well be that a fraction of dust is destroyed when ejected out of the ISM, or the DTG ratio of the outflow differs to the ISM, which will alter our predictions. As we mentioned above, it is also possible that the sputtering rates are higher than our prediction. However, the nature of such processes are currently unclear. Future observations to quantify dust properties in the galactic wind and the IGM will provide additional constraints to these processes.
}
\DIFaddend Figure 14 in \citet{Triani20} shows that \DIFdelbegin \DIFdel{at low-redshift, }\DIFdelend galaxies have more dust in their halo than in the ISM \DIFaddbegin \DIFadd{at low-redshift}\DIFaddend . This outcome roughly agrees with the massive dust \DIFaddbegin \DIFadd{content }\DIFaddend found in the \DIFdelbegin \DIFdel{circumgalactic medium \mbox
\citep[CGM,][]{Dunne11}
. \mbox
\citet{Popping17}
also predict }\DIFdelend \DIFaddbegin \DIFadd{CGM \mbox
\citep{Dunne11, PMC15}
. The galaxy formation model of \mbox
\citet{Popping17}
also predicts }\DIFaddend a significant amount of dust in \DIFdelbegin \DIFdel{the halo.
}\DIFdelend \DIFaddbegin \DIFadd{both the halo and ejected reservoir. However, their dust mass density is notably higher than our results.
}\DIFaddend
\dustysage provides a good agreement with the galaxy stellar mass function at redshift $z=0$. It also successfully reproduces the dust mass function and various dust scaling relations over a wide range of redshifts \citep{Triani20}. To achieve a more realistic distribution of the stellar mass in the bulge and disc, we make slight changes to a few of the parameters in \citet{Triani20}. These are listed in Table \ref{tab:params}.
\begin{table*}
\caption{Fiducial \dustysage parameters used throughout this work, also compared to those from \citet{Triani20}.}
\label{tab:params}
\begin{tabular}{lccc}
\hline
\hline
Parameter & Description & Value &\citet{Triani20} \\
\hline
\hline
$f_b^\mathrm{cosmic}$ & Cosmic baryon fraction & 0.17 & 0.17\\
\hline
$z_0$ & Redshift when $\mathrm{H_{II}}$ regions overlap & 8.0 & 8.0\\
$z_r$ & Redshift when the intergalactic medium is fully reionized & 7.0 & 7.0\\
\hline
$\alpha_\mathrm{SF}$ & Star formation efficiency from $\mathrm{H_2}$ [$\mathrm{Myr^{-1}}$] & 0.005 & 0.005\\
$R$ & Instanteneous recycling fraction & 0.43 & 0.43\\
\hline
$\epsilon_\mathrm{disc}$ & Mass-loading factor due to supernovae & 2.0 & 3.0\\
$\epsilon_\mathrm{halo}$ & Efficiency of supernovae to unbind gas from the hot halo & 0.2 & 0.3\\
$k_\mathrm{reinc}$ & Velocity scale for gas reincorporation & 0.15 & 0.15\\
\hline
$\kappa_\mathrm{R}$ & Radio mode feedback efficiency & 0.09 & 0.08\\
$\kappa_\mathrm{Q}$ & Quasar mode feedback efficiency & 0.005 & 0.005\\
$f_\mathrm{BH}$ & Rate of black hole growth during quasar mode & 0.015 & 0.015\\
\hline
$f_\mathrm{friction}$ & Threshold subhalo-to-baryonic mass for satellite disruption or merging & 1.0 & 1.0\\
$f_\mathrm{major}$ & Threshold mass ratio for merger to be major & 0.15 & 0.3\\
\hline
$\alpha_\mathrm{burst}$ & Exponent for the powerlaw for starburst fraction in merger & 0.18* & 0.7\\
$\beta_\mathrm{burst}$ & Coefficient for starburst fraction in merger & 0.75* & 0.56\\
\hline
$\delta_\mathrm{C}^\mathrm{AGB}$ & Condensation efficiency for AGB stars & 0.2 & 0.2\\
$\delta_\mathrm{C}^\mathrm{SNII}$ & Condensation efficiency for SN II & 0.15 & 0.15\\
$\tau_\mathrm{acc,0}$ & Accretion timescale for grain growth [$\mathrm{yr}$] & $4.0 \times 10^5$ & $4.0 \times 10^5$\\
$f_\mathrm{SN}$ & Fraction of destroyed dust to the swept dust mass by SN & 0.1 & 0.1\\
\hline
\hline
\multicolumn{3}{l}{*value adopted from \citet{Somerville01}}\\
\end{tabular}
\end{table*}
\section{How morphology and stellar mass relate to dust fraction in the hot halo, ISM and ejected reservoir}
\label{sec:morphologies}
In this section, we investigate the relation between galaxy morphology, represented as the bulge-to-total (BTT) mass ratio, and the fraction of dust in the ISM, halo and ejected reservoirs relative to the total dust mass in all reservoirs. We use stellar mass in computing the BTT mass ratio and assume the gas mass to be negligible. We present the median dust fraction vs BTT mass ratio in three stellar mass bins in Figure \ref{fig:btt-fraction}. The grey histogram shows the distribution of BTT mass ratios across all galaxies. In all panels, galaxies with higher BTT mass ratio contain most of their dust in the halo while those with a lower BTT mass ratio tend to keep theirs in the ISM. For the low mass galaxies in the bottom panel, spiral galaxies contain a significant fraction of their dust in the ISM and ejected reservoir. These galaxies make up more than 50\% of the overall population. Conversely, the halo contains nearly 90\% of the total dust in elliptical galaxies with BTT > $0.8$.
For the galaxies with $\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10 - 11$ (central panel), the same trend occurs. Galaxies with a BTT mass ratio smaller than $\sim 0.2$ keep almost half of their dust in the ISM. For the most massive galaxies (top panel), all galaxies contain most of their dust in the halo. Massive galaxies are more likely to host AGNs compared to low-mass galaxies. The stronger AGN feedback is more likely to blow dust out into the halo (and ejected reservoir) compared to the weaker SN feedback commonly found in low-mass galaxies. Massive galaxies also have deeper potential wells than the low mass systems, allowing them to capture reheated dust in the hot halo. However, although the ISM dust fraction is smaller than in Milky Way sized galaxies, their actual dust mass in the ISM is significant and will likely make a noticeable contribution to the galaxy SED in the infrared.
Observations have found galaxies with a massive amount of dust in their \DIFdelbegin \DIFdel{circumgalactic medium (CGM)}\DIFdelend \DIFaddbegin \DIFadd{CGM}\DIFaddend . For a galaxy with stellar mass $\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10.1$, \citet{Peeples14} found a fiducial value of metals locked in CGM dust of $5 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace, almost twice the value for those found in the form of ISM dust, $2.6 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace. \citet{Menard10} derived a CGM dust mass of $5 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace for galaxies with stellar mass $\log \mathrm{M_*} = 10.4 \ensuremath{\mathrm{M}_{\odot}}\xspace$. Although the CGM is not exactly equivalent to our halo gas component, we will use CGM dust as a proxy to test our predictions of the dust bond to, but outside the galaxy.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{images/btt-fraction-mass-512files.png}
\caption{The median value for dust mass fraction in each baryon reservoir versus bulge-to-total stellar mass ratio (BTT). The purple, orange and green lines represents hot dust mass in the halo, cold dust mass in the ISM, and dust in the ejected reservoir, respectively. The grey histogram represents the number distribution of galaxies in 10 BTT bins.}
\label{fig:btt-fraction}
\end{figure*}
\section{Classifying galaxies based on their dust and fundamental properties}
\label{sec:classification}
To investigate how the dust content of galaxies correlates with their stellar mass, star formation activity, and morphology, we first establish our simulated galaxies into groups. In the \dustysage model, the dust in galaxies is distributed in three components - the ISM, the halo and the ejected reservoir. We exclude the dust in the ejected reservoir in our analysis since it is out of the system. Our four dust groups are described below.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/dust-stellarmass-512files.png}
\caption{The cold dust mass of simulated galaxies as a function of stellar mass at redshift $z=0$. The heatmap shows the density distribution with a brighter colour representing higher density, and the black line marks the median. The red line marks the threshold where the cold dust mass per stellar mass equals $10^{-4}$. Above this line, galaxies are classified as ISM dust-rich while below are classified as ISM dust-poor.}
\label{fig:dust-stellarmass}
\end{figure}
\subsection{ISM dust-poor}
\label{ssec:ISM-dust-poor}
When considering dust, observers typically report the properties of the ISM \citep[e.g.,][]{Eales09, RR14, Santini14, Mancini15, daCunha15, Nersesian19}. Although some works have extended dust measurements to include the CGM \citep{Menard10, PMC15}, such observations are rare. Therefore, as a first step, we have classified our model galaxies based on their cold dust content in the ISM. Figure~\ref{fig:dust-stellarmass} shows the relation between cold dust mass in the ISM and stellar mass of our model galaxies at redshift $z=0$. We find that the dust mass in the ISM increases with the host stellar mass, but with a significant scatter.
The median values of our predicted ISM dust mass versus stellar mass are in a good agreement with the observations of dust in local galaxies, such as from the DustPedia catalogue \citep{Nersesian19}. Figure \ref{fig:dust-stellarmass} shows a large scatter below the median values, marking galaxies with relatively less dust compared to the overall population.
We draw an arbitrary line at dust mass per stellar mass of $10^{-4}$, marked with the red line in Figure \ref{fig:dust-stellarmass} to divide the galaxy population into two categories: ISM dust-rich and ISM dust-poor. The ISM dust-poor group contains galaxies below the red line. Observations have found such ISM dust-poor galaxies in both the local and high-redshift Universe \citep{Fisher14}.
Further exploration reveals that \dustysage galaxies in this ISM dust-poor category vary in their halo dust content. Therefore, we additionally divide the ISM dust-poor group based on their fraction of halo dust per stellar mass described below.
\subsubsection{Halo dust-rich}
\label{sssec:halo-dust-rich}
Several observations \citep{PMC15, Peeples14, Menard10} extended their search for metals and dust to the CGM. \citet{Peeples14} find that only $25\%$ of metals created in stars stay in the ISM; a similar fate also occurs for dust. \citet{Menard10} found galaxies with massive CGM dust. These observations discovered the existence of dust out of the ISM, which is reproduced by our model.
In our model, a mix of halo and ejected dust might be a better representation of the CGM for some galaxies. However, we only use halo dust to represent the observed CGM dust. We use the same threshold as the ISM dust-mass to define the halo-dust rich galaxies: halo dust mass per stellar mass of $10^{-4}$. Halo dust-rich galaxies are defined as ISM dust-poor galaxies with halo dust mass above this threshold.
\subsubsection{Dust-poor}
\label{sssec:dust-poor}
The population of galaxies that lack dust in both their ISM and halo is classified as dust-poor. Due to the rarity of dust measurements outside the ISM, we found no counterpart of this category from the observations. Their existence in our model serves as a prediction for future surveys measuring the CGM dust.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/sSFR-hist-allmass-512files.png}
\caption{The distribution of the specific star-formation rate in our ISM dust-rich galaxies. The grey line at $2.71 \times 10^{-11} \mathrm{yr^{-1}}$ is the Milky Way value adopted from \citet{Licquia15} which we use as a threshold in defining quenched and star-forming galaxies in our model galaxies.}
\label{fig:sfr-hist}
\end{figure}
\subsection{ISM dust-rich}
\label{ssec:ISM-dust-rich}
We define the ISM dust-rich galaxies in our model as those with a fraction of ISM dust mass per stellar mass above $10^{-4}$. Dust obscures the intrinsic stellar spectra of such galaxies and re-emits the radiation in the infrared \citep{WTC92, WG00}. Infrared and sub-millimeter surveys have found a substantial number of these galaxies, especially at low redshift \citep[e.g.,][]{Eales09, DEE03, daCunha15, Clemens13, RR14}.
Dust accounting in the galactic ISM is the main focus of many current and future infrared surveys. Forthcoming galaxy survey, like those using JWST, will measure dust in high redshift galaxies. Such surveys will provide additional constraints to our model predictions.
Besides the relation between dust mass and stellar mass shown in Figure \ref{fig:dust-stellarmass}, observations have also found a relation between dust mass and SFR \citep{daCunha10, Casey12, Santini14}. Based on this, we divide our ISM dust-rich galaxies further into star-forming and quenched subclasses.
\subsubsection{Star-forming ISM dust-rich}
\label{sssec:sf-dust-rich}
Figure \ref{fig:sfr-hist} shows the specific SFR (sSFR) distribution of the ISM dust-rich galaxies in our model. To separate star-forming from quenched galaxies, we use the Milky Way sSFR value as the threshold, $2.71 \times 10^{-11} \mathrm{yr^{-1}}$ \citep{Licquia15}. Galaxies with sSFR above the Milky Way value are defined as the star-forming ISM dust-rich.
\subsubsection{Quenched ISM dust-rich}
\label{sssec:q-dust-rich}
ISM dust-rich galaxies are commonly related to star-forming spirals since dust is initially formed in stellar ejecta and grows in molecular clouds \citep{Valiante09}. However, we also find a population of ISM dust-rich galaxies with relatively low star formation activity in our model (see Figure \ref{fig:sfr-hist}). The quenched ISM dust-rich galaxies are those with a sSFR below the Milky Way value of $2.71 \times 10^{-11} \mathrm{yr^{-1}}$.
\section{The relation between dust mass and fundamental galaxy properties}
\label{sec:relation}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{images/btt-ssfr-Z-512files.png}
\caption{\textit{Top panel} shows the BTT ratio, sSFR and gas-phase metallicity of our model galaxies in the ISM dust-poor group. This group consists of two subgroups: halo dust-rich and dust-poor. Markers with small, medium and large size represents three mass bin: $\log \mathrm{M_*} = 9 - 10$ \ensuremath{\mathrm{M}_{\odot}}\xspace, $\log \mathrm{M_*} = 10 - 11$, and $\log \mathrm{M_*} = 11 - 12$ \ensuremath{\mathrm{M}_{\odot}}\xspace, respectively. The points mark the median sSFR and median BTT mass ratio of each group, with the errorbars spanning the $16^\mathrm{th}$ and $84^\mathrm{th}$ percentiles. The colour of each point marks the gas-phase metallicity, with brighter colours representing higher value. We assume a solar metallicity of 0.02 and $12 + \log[\mathrm{O/H}] - \log[\mathrm{Z/Z_\odot}] = 9.0$. The number inside parentheses in the legend represents the percentage of galaxies in each category with respect to the total number of galaxies with stellar mass $\log \mathrm{M_*} = 9 - 12$ at $z=0$. Open symbols indicate equivalent observational dataset taken from DustPedia \citep{Nersesian19} and KINGFISH \citep{Kennicutt11, RR14}.
\textit{Bottom panel} shows the same relation above but for the ISM dust-rich group, which is splitted into star-forming and quenched subgroups.}
\label{fig:btt-sfr}
\end{figure*}
In the previous sections, we categorized model galaxies into four classes: star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich and dust-poor. We now plot the BTT mass ratio vs sSFR of each group in three mass bin in Figure \ref{fig:btt-sfr} as indicated in the legend, with the colour distribution representing the median gas-phase metallicity of each group. The error bars denote the $16^\mathrm{th}$ and $84^\mathrm{th}$ percentile range.
Figure~\ref{fig:btt-sfr} shows a general trend where the ISM dust-rich galaxies in the bottom panel occupy the lower right space marking disky galaxies with higher star formation activity. Meanwhile, ISM dust-poor galaxies in the top panel generally have lower sSFR and a larger BTT mass ratio. Since galaxies typically change from star-forming spirals to quenched ellipticals over time, this plot indicates that the ISM also evolves from being dust-rich to dust-poor across h.
The gas-phase metallicity in Figure~\ref{fig:btt-sfr} increase with stellar mass, as one may expect from the galaxy mass-metallicity relation \citep{Tremonti04}. Both ISM dust-rich groups generally have higher metallicity than the ISM dust-poor groups.
Star-forming ISM dust-rich galaxies have the highest sSFR and lowest BTT mass ratio. Condensation in stellar ejecta is one of the main dust production channels in our model and is tightly correlated with star formation activity. Thus, we would expect ISM dust-rich galaxies to have a high sSFR. It is also possible that the high sSFR of such galaxies reflects the abundance of molecular gas in the ISM, which enables dust growth via accretion of metal grains.
Quenched ISM dust-rich galaxies have much lower sSFR (on average) and are also located in the low BTT-ratio regime. Their gas-phase metallicity is higher than that of the star-forming ISM dust-rich galaxies. As galaxies in this group have less sSFR than those in the star-forming group, their stellar dust production must be less efficient. However, their abundance of dust can be explained by their high availability of gas-phase metals, which can accrete onto existing dust grains in the dense environment. The low sSFR of these galaxies also suggest that they lack massive young stars that become SN, the primary destroyer of dust in the ISM. The scarcity of SN can also prevent the dust and refractory elements from being expelled, allowing them to preserve more dust and metals in their ISM.
In contrast to the ISM dust-rich galaxies, the ISM dust-poor group have higher BTT mass ratios (i.e., more elliptical), lower average sSFRs and lower gas-phase metallicities. The halo dust-rich group have the highest BTT mass ratio, incorporating the ``bulgiest'' galaxies for all mass bins. Galaxies in this group also have the lowest gas metallicities. To explore this further, we plot the distribution of the ratio between the ISM cold gas mass and the hot halo gas mass in Figure~\ref{fig:hot-cold-ratio}. Here, the halo dust-rich group peaks lowest of all groups, showing that galaxies in this group either have less cold gas in their ISM or more gas in their halo. The fact that these galaxies have massive halo dust and hot gas, yet they lack ISM gas and dust, indicates an efficient mechanism to transport their ISM content out of the disk.
Figure~\ref{fig:outflowrate} shows the fraction of galaxies in each group versus outflow rate. The halo dust-rich group dominates for outflow rates above $30\ \ensuremath{\mathrm{M}_{\odot}}\xspace \mathrm{yr^{-1}}$. This suggests that the efficient outflow in these galaxies can blow their ISM content out of the disk, while their low sSFR does not allow them to replenish the metals and dust in the ISM.
Looking back to Figure \ref{fig:btt-sfr}, we see that the total percentage of galaxies in the most massive mass bin ($\log \mathrm{M_*} = 11 - 12$) is $1.47 \%$ and the halo dust-rich group makes up a quarter of them, a more substantial contribution compared to those in the lower mass bins. In \dustysage, massive galaxies are likely to have deeper potential wells, which allow them to capture most of the dust outflow from the ISM in the halo. Therefore, very little dust escapes into the ejected reservoir in feedback events. Figure~\ref{fig:btt-fraction} also shows that in the most massive galaxies, the majority of the dust mass is in the hot halo, and an only small fraction of their dust fraction is in the ejected reservoir at all BTT mass ratios.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/hot-cold-gas-512files.png}
\caption{The probability distribution of the cold to hot gas mass ratio of our simulated galaxy categories. Red, blue, green and purple lines represent the star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich and dust-poor groups, respectively.}
\label{fig:hot-cold-ratio}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/outflowrate-512files.png}
\caption{The gas outflow rate distribution due to feedback. We assume that the dust-to-gas mass ratio in the outflow is the same as the general ISM.}
\label{fig:outflowrate}
\end{figure}
In the dust-poor category, we find an anti-correlation between stellar mass and sSFR. The median sSFR of dwarf galaxies is $\log \mathrm{sSFR}=-10.6$, more than two orders of magnitude larger than those in the most massive bin. Galaxies will become dust-poor if their dust depletion rate exceeds the dust production rate. Since Figure \ref{fig:dust-stellarmass} shows a positive correlation between dust mass and stellar mass, more massive galaxies are more likely to have more dust in their ISM. Therefore, only those galaxies with a very low production rate can become dust-poor. The anti-correlation reflects the link between dust and star formation rates.
Figure \ref{fig:btt-sfr} also allows us to examine the stellar mass evolution across the four groups. At the massive end, galaxies with mass $\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 11 - 12$ are comprised primarily of the star-forming ISM dust-rich category (0.6 \% of the total population), followed by the quenched ISM dust-rich and halo dust-rich groups each containing 0.3\% of the total galaxies. The dust-poor group is in the minority amongst massive galaxies. The dominance of the ISM dust-rich and halo dust-rich groups shows that the most massive galaxies rarely destroy or eject their dust out of the halo. Massive late-type and early-type galaxies contain the majority of their dust in the ISM and the halo, respectively.
In the middle bin ($\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10 - 11$), the biggest percentage is still contributed by the star-forming ISM dust-rich category (14 \%), followed by the quenched ISM dust-rich with 4.9 \%, halo dust-rich with 4.6 \%, and the dust-poor group with 1.1 \%. In the lowest mass bin ($\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 9 - 10$), the star-forming ISM dust-rich still dominates (28\%). However, the contributions of quenched ISM dust-rich and dust-poor groups in this mass bin are also the most significant (13.9 \% and 12.8 \%, respectively) compared to those in the higher mass bins. The halo dust-rich group shares the smallest percentage for dwarf galaxies with only 3.6\%.
\subsection{Comparison with observed datasets}
\label{ssec:compared-with-data}
To add context to our theoretical predictions, we make comparisons with two observation datasets. DustPedia \citep{Nersesian19} is a compilation of local galaxies with dust measurements. To create this, the authors use the dust model \texttt{THEMIS} in the SED fitting code \texttt{CIGALE} to infer dust mass and galaxy properties from \textit{Herschel} multiwavelength photometry. The authors use numerical Hubble T as the morphology indicator, with lower value corresponds to a larger stellar composition in bulge relative to the disk. To convert their Hubble T value to BTT-ratio, we use the K-band bulge-to-disk flux ratio mapping from \citet{Graham08}. Note that the BTT ratio for our simulated galaxies is based on stellar mass, while the observations are based on flux.
We group their data based on our definition of the star-forming ISM dust-rich, quenched ISM dust-rich and ISM dust-poor galaxies. We can not divide their ISM dust-poor galaxies further because they do not provide a halo dust mass. We did not find galaxies classified as quenched ISM dust-rich in their dataset. Therefore, we plot the mean values of BTT mass ratio and sSFR of their star-forming ISM dust-rich group as a pink-filled circle and their ISM dust-poor group as a pink-filled square in Figure \ref{fig:btt-sfr}.
Our second observational dataset is taken from the KINGFISH survey \citep{Kennicutt11}. We group KINGFISH galaxies based on our definitions and compute the mean dust mass, BTT mass ratio, and SFR for each group. We plot the mean dust and galaxy properties with markers showing the group they belong. The dust measurement for KINGFISH galaxies is taken from \citet{RR14}.
Our prediction for star-forming ISM dust-rich and ISM dust-poor galaxies are in good agreement with the observed values. In both the KINGFISH and DustPedia data, dust-poor galaxies are those with the most elliptical morphology (E-type) with Hubble T below -2 \citep{RR14, Nersesian19}. However, the map of \citet{Graham08} that we use to convert the Hubble type to BTT ratio only extends to S0-type galaxies (Hubble T = -2). Therefore, we use the Hubble T = -2 conversion for elliptical galaxies with Hubble T < -2. The converted BTT flux ratio of these galaxies are $0.15$ dex lower than our median mass ratio but are still within our $16^\mathrm{th}$ percentile.
\section{Evolution of the relations between dust and fundamental properties}
\label{sec:evolution}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{images/btt-ssfr-Z-evol-512files.png}
\caption{The evolution of the relation shown in Figure \ref{fig:btt-sfr} between redshift $z=0$ and redshift $z=3$. For clarity, we focus on the middle mass bin ($\log M* = 10 - 11$ \ensuremath{\mathrm{M}_{\odot}}\xspace). Circle, inverted triangle, diamond and square represent the median value of the star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich and dust-poor groups, respectively.}
\label{fig:btt-sfr-evol}
\end{figure*}
Galaxy evolution models predict that galaxies evolve from star-forming spiral to quench elliptical (i.e., the bottom right to top left in Figure \ref{fig:btt-sfr}) \mbox{\citep{Somerville01, Baugh05, Croton06, Croton16}}. The dust content of the ISM follows this evolutionary timeline to become dust-poor over time. There are two primary depletion mechanisms for ISM dust: destruction by SN shocks \citep{Slavin15} and ejection by SN and AGN feedback \DIFdelbegin \DIFdel{\mbox
\citep{Popping17}
}\DIFdelend \DIFaddbegin \DIFadd{\mbox
\citep{Feldmann15, Popping17}
}\DIFaddend . Ejected dust can be trapped in the halo if the potential well is sufficiently deep, causing the host to end up as a halo dust-rich galaxy. Otherwise, such galaxies end up as dust-poor.
To see the evolution of the dust content in galaxies across time, we plot the relation between BTT mass ratio and sSFR of our dust groups from redshift $z=0$ to redshift $z=3$ in Figure \ref{fig:btt-sfr-evol}. Here, we focus on the middle mass bin ($\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10 - 11$) as representative of the Milky Way mass range. Circle, inverted triangle, diamond, and squares represent the star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich, and dust-poor categories, respectively. This plot does not show the evolution of the same group of galaxies. Instead, it shows us the map of galaxies belong to each group at different redshift in the BTT mass ratio and sSFR plane.
Figure~\ref{fig:btt-sfr-evol} shows how the percentage of galaxies in each group changes over time. The fraction of star-forming ISM dust-rich galaxies peaks at $z=1$ then remains at $\approx 25\%$ until $z=0$. The other three groups show lower fractions but continue to increase down to the present day. This evolutionary trend reveals how star-forming ISM dust-rich galaxies might convert into the other groups at lower redshift.
Across redshift, the behaviour of individual groups within the BTT mass ratio - sSFR space also changes. In the BTT mass ratio dimension, the quenched ISM dust-rich group gets diskier towards $z=0$ with a $0.3$ drop, while the other three groups do not show notable changes between $z=3$ and $z=0$. All groups except the quenched ISM dust-rich shows a significant decrease of the median sSFR from $z=3$ to $z=0$. The biggest sSFR decrease occurs in halo dust-rich galaxies, by more than 2 order of magnitude, while the star-forming ISM dust-rich and dust-poor galaxies each undergo a single dex decrease.
In the gas phase metallicity dimension, the halo dust-rich group have the lowest median value across redshift $z=3$ to $z=0$. The median metallicity of both dust-poor and star-forming ISM dust-rich galaxies increases slightly with decreasing redshift. Meanwhile, the median metallicity of the quenched ISM dust-rich group fluctuates with redshift; it increases $\approx 0.1$ dex between $z=3$ and $z=2$, does not change between $z=2$ and $z=1$, then decreases $\approx 0.3$ dex.
\section{Discussion: the behaviour of each group across redshift}
\label{sec:discussion}
From Figure~\ref{fig:btt-sfr} and Figure~\ref{fig:btt-sfr-evol}, we can see that although we divide galaxies into groups solely based on their dust content, the galaxy properties of each group are often distinct. Also, galaxy properties within each group evolve with redshift. Our predictions for how these groups change with time can help us understand the baryonic and dust physics occurring. In this section, we discussed the features of these groups between redshift $z=3$ and the present day.
\subsection{The star-forming ISM dust-rich group}
At all mass bins at redshift $z=0$, the star-forming ISM dust-rich group consistently stays at the lower right region in the BTT mass ratio - sSFR space (see Figure~\ref{fig:btt-sfr}). This indicates that these galaxies are still actively forming stars and have an abundance of molecular gas in their ISM, providing a dust production channel via condensation in stellar ejecta and grain growth in the molecular clouds. The gas-phase metallicity in this group is relatively high compared to those in the halo dust-rich and dust-poor groups. This group dominates the galaxy population for all mass bins at redshift $z=0$.
Across redshift, this group stays in the disky star-forming region (see Figure~\ref{fig:btt-sfr-evol}) and continues to dominate in the overall population. The median gas-phase metallicity is lower at higher redshift. This metallicity increase with decreasing redshift reflects the ability of such galaxies always to retain an abundance of metals in their ISM. Since galaxies in the star-forming ISM dust-rich group have high sSFR by the selection, we expect that the ISM will be both metal and dust-rich. Observations of the local and high redshift Universe find many galaxies that belong to this group. Figure \ref{fig:btt-sfr} shows that at redshift $z=0$, both the KINGFISH~\citep{Kennicutt11, RR14} and DustPedia~\citep{Nersesian19} datasets have galaxies with dust-to-stellar mass and sSFR classifications similar to our star-forming ISM dust-rich group.
Figure \ref{fig:dtm-ssfr} shows the sSFR of our model galaxies as a function of dust-to-stellar mass ratio from redshift $z=0$ to $z=3$. The white line marks the divider between the ISM dust-rich and ISM dust-poor groups. Across redshift, many authors find star-forming galaxies with a dust rich ISM. The general trend that the sSFR decreases with decreasing redshift is seen both for the model and in the observations. At $z=0$, our ISM dust-rich galaxies are in good agreement with the observational values from DustPedia \citep{Nersesian19}, \citet{RR14} and \citet{Santini14}. For the ISM dust poor galaxies below the grey line, the observations are lower than our median value but are still within the $16^\mathrm{th}$ percentile. At higher redshift, observations are yet to find ISM dust-poor galaxies. At $z=1$, our model roughly agrees with the dust-rich data from \citet{Santini14}, although our median value of sSFR lies $0.1 - 0.4$ dex below the observed values. At $z=2$ and $z=3$, however, our model fails to reproduce galaxies with $\log (\mathrm{M_{dust}} / \mathrm{M_{star}}) > -2.5$ where most of the observational datasets lie. \dustysage and our base model \texttt{SAGE}\xspace \citep{Croton16} are constrained using various galaxy observations at $z=0$; therefore, we still find limitations in reproducing the high redshift population. We also note that the dataset from \citet{daCunha15} (red triangles) is obtained in the ALMA LESS survey which is biased towards the extremely bright galaxies.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{images/dtm-ssfr-millennium.png}
\caption{The relation between specific star formation rate (sSFR) and the dust-to-stellar mass ratio from redshift $z=0$ to $z=3$. The heat map shows the 2D density distribution of our model galaxies with brighter color representing higher density. The solid black lines mark the median while the dashed lines marked the $16^\mathrm{th}$ and $84^\mathrm{th}$ percentile. The white vertical lines mark a log dust-to-stellar mass ratio = -4, where we divide the ISM dust-rich and dust-poor categories. The red circles, blue diamonds, green squares and red triangles are observational values from DustPedia: \citet{Nersesian19}, \citet{RR14}, \citet{Santini14} and \citet{daCunha15}, respectively.}
\label{fig:dtm-ssfr}
\end{figure*}
\subsection{The quenched ISM dust-rich group}
The quenched ISM dust-rich group consists of galaxies that have an abundance of dust in their ISM despite their low star formation activity (defined as galaxies with sSFR below the Milky Way). One possible scenario for this is that due to the low sSFR, these galaxies no longer host supernovae events that destroy dust and expel metals and dust from the ISM. This also explains the higher gas-phase metallicity of galaxies in this group compared to those in the star-forming ISM dust-rich group. The abundance of metals in the ISM provides material for dust growth via grain accretion, allowing such galaxies to accumulate dust mass. Another scenario for their dust production is the condensation in the ejecta from previous star formation episodes. Although they do not have a fresh supply of stellar ejecta due to the low sSFR, it is possible that there are still refractory elements from previous SN and AGB winds in the ISM that condense into dust grains.
Figure~\ref{fig:btt-sfr} shows that the majority of quenched ISM dust-rich galaxies can be found in the low mass population. Compared to the star-forming ISM dust-rich group, this group is rare at low and high redshift. Figure~\ref{fig:btt-sfr-evol} shows that the fraction of this group with stellar mass $\log \mathrm{M_*} = 10-11$ at $z > 0$ is always below $6\%$. Due to their low abundance, we have not found any observational evidence for this group.
\subsection{The halo dust-rich group}
Halo dust-rich galaxies have little or no dust in their ISM but contain a significant amount of dust in their halo. At redshift $z=0$, this group includes elliptical galaxies with the largest BTT mass ratios and low sSFR. The median gas-phase metallicity of this group is also the lowest compared to the other groups. The small sSFR does not allow for efficient instantaneous metal and stellar dust production in such galaxies. However, they should have accumulated metals and dust from previous star-formation episodes. The lack of metals and dust in their ISM, while their halo is dust-rich, indicates that the ISM content is transported into the halo (see Section~\ref{sec:relation}).
At higher redshift, however, the halo dust-rich group occupies an increasingly high sSFR region with a median $\log \mathrm{sSFR} >$ the Milky Way value. The high sSFR of these galaxies means that they are producing metals and dust in their stellar ejecta. However, unlike galaxies in the other groups, galaxies in this group show no increase in gas metallicity with decreasing redshift. The low abundance of metals and dust in their ISM is possible if the feedback is efficiently heating and blowing out the newly formed metals and dust. Such feedback might be provided by AGN activity \citep{Sarangi19}.
Galaxies in a deep potential well can retain the feedback-heated gas, dust and metals in their halo. Mass in the dark matter halo usually sets the depth of the potential well. Since stellar mass correlates with halo mass, halo dust-rich galaxies should be more common in the massive galaxies. Figure \ref{fig:btt-sfr} shows that in the massive bin, this group comprise a quarter of the population, a more significant fraction than the lower stellar mass bin.
\citet{Menard10} measured dust reddening effects of background quasars relative to the foreground SDSS galaxies. They observed the existence of diffuse dust in halos with an amount comparable to those in the galactic disk. For $0.5L^*$ galaxies, the authors estimated a dust mass of $5 \times 10^7 \ensuremath{\mathrm{M}_{\odot}}\xspace$ in the halo. \citet{PMC15}
confirmed this halo dust mass for $0.1L^* - 1L^*$ low redshift galaxies. Observations of CGM dust provides evidence for our prediction that dust is transported out of the ISM and trapped in the halo. However, such observations are still uncommon. Future CGM dust measurements will allow for better constraints for models such as \dustysage.
\subsection{The dust-poor group}
We define galaxies that lack dust in both the ISM and their halo as dust-poor galaxies. At redshift $z=0$ (Figure~\ref{fig:btt-sfr}), the dust-poor group has low median sSFR, low gas-phase metallicity and high BTT mass ratio. In the DustPedia~\citep{Nersesian19} and KINGFISH~\citep{Kennicutt11, RR14} datasets, galaxies with morphological type E-S0 have a median dust-to-stellar mass ratio below $1 \times 10^{-4}$ and are classified as ISM dust-poor as per our definition. However, both datasets do not provide measurement for dust in the halo. Therefore, we assume their halo is dust-poor, and we include them in the dust-poor group instead of the halo dust-rich group.
Dust depletion in the ISM occurs in two ways: destruction by SN shocks inside the galaxy disk and dust that is ejected out of the ISM by feedback. If the potential well of the system is enough to trap the dust, the ejected dust will end up in halo. However, galaxies with low stellar mass usually live in shallower potentials and thus are likely to lose their dust completely through ejection. This likely explains why Figure~\ref{fig:btt-sfr} shows that the dust-poor group contain the most significant fraction of their galaxies in the low mass bin.
In Figure \ref{fig:btt-sfr-evol}, the dust-poor group moves to the lower BTT mass ratio, lower sSFR region with decreasing redshift while the gas metallicity increases. At high redshift, galaxies in this group have relatively high sSFR, but their ISM still lacks both metals and dust. Dust formation is highly sensitive to the abundance of metals. In stellar ejecta, dust forms via the condensation of metals ejected from stars. In dense clouds, dust grains grow by accreting metals. Therefore, galaxies with low metallicity can not produce dust efficiently. If this low production rate can not keep up with the depletion rate, a galaxy will end up in the dust-poor group.
\citet{Fisher14} observed the dust emission from a local dwarf galaxy, I Zwicky 18, and found it lacks dust with a dust-to-stellar mass ratio of about $10^{-5}$ to $10^{-6}$. This galaxy has a SFR of $0.05$ \ensuremath{\mathrm{M}_{\odot}}\xspace/year and a very low metallicity of $12+\log\mathrm{[O/H]} = 7.17$, much lower than that seen in local galaxies \citep{Fisher14}. The reported stellar mass is $9 \times 10^7 \ensuremath{\mathrm{M}_{\odot}}\xspace$, which makes the sSFR $ = 5.5 \times 10^{-10}$, consistent with the median sSFR value of the dust-poor galaxies in the lower mass group from \dustysage (see Figure \ref{fig:btt-sfr}). Although the stellar mass of I Zwicky 18 is well below our sample, we find a consistent feature where a lack of metals accompanies the lack of dust.
\section{Conclusion}
\label{sec:conclusion}
During galaxy's evolution, dust physics is heavily influenced by star formation, gas physics, outflows and the depth of the potential well. Therefore, the dust, stellar, and gas content of galaxies are tightly related to each other. In this paper, we use the \dustysage galaxy formation model \DIFdelbegin \DIFdel{\mbox
\citep{Triani20}
}\DIFdelend \DIFaddbegin \DIFadd{\mbox{\citep{Triani20}} }\DIFaddend to explore the relation between dust content and the fundamental properties of galaxies. We have divided the galaxy population based on their dust mass and observable properties. At redshift $z=0$, the properties of our model galaxies are in good agreement with observations, especially where the statistics are sound.
Our investigation includes the comparison of morphology, specific star formation rate, metallicity and stellar mass for galaxies grouped by their dust content from redshift $z=3$ to $z=0$. Our main conclusions are the following:
\begin{itemize}
\item We find that the distribution of dust in the ISM, hot halo and ejected reservoir is affected by galaxies' morphology and stellar mass. The fraction of dust in the hot halo increases with BTT mass ratio and stellar mass (Figure \ref{fig:btt-fraction}).
\item At redshift $z=0$, ISM dust-rich galaxies have the highest gas-phase metallicities. Both star-forming and quenched ISM dust-rich groups shows a low BTT mass ratio and relatively high median sSFR. Galaxies in the halo dust-rich and dust-poor groups are metal poor. They are distributed in the high BTT mass ratio and low sSFR regime of Figure~\ref{fig:btt-sfr}.
\item We see an evolution in the behaviour of our grouped galaxies in Figure~\ref{fig:btt-sfr-evol}. From $z=3$ to $z=0$: (i) the BTT mass ratio of the quenched ISM dust-rich group drops $\approx 0.3$ dex while the other three groups show insignificant changes, (ii) the median sSFR of all groups except quenched ISM dust-rich trend towards lower region, and (iii) all groups show weak evolution of gas-phase metallicity with redshift except the quenched ISM dust-rich group where the metallicity fluctuates with redshift.
\item Across redshift, both star-forming ISM dust-rich and quenched ISM dust-rich groups maintain their high median sSFR and high gas-phase metallicity. This implies that such galaxies are still actively producing metals and dust via stellar production. The abundance of metals in the ISM also enables dust growth via metal accretion in dense clouds.
\item The halo dust-rich group consists of galaxies with a dust-poor ISM but dust-rich halo. They have a relatively high outflow rate of heated gas, dust, and metals out of the ISM. At present, they are elliptical with very low sSFR and metallicity. The gas-phase metallicity of these galaxies does not depend significantly on redshift. At high redshift, galaxies in this group have high sSFR that enables an efficient dust production mechanism. Their lack of ISM dust, therefore, implies effective feedback to reheat the newly formed dust. Because these massive galaxies have a deep potential well, they are more likely to retain the heated dust in their halo. Figure~\ref{fig:btt-sfr} and Figure \ref{fig:btt-sfr-evol} shows that this group is made up of a larger fraction of massive galaxies.
\item Galaxies in the dust-poor group have low dust mass both in their ISM and halo. From redshift $z=3$ to $z=0$, this group evolves from the high sSFR region to the lower sSFR region. However, they always have low metallicity. As metals are the main ingredients for dust, this low metallicity implies that their dust production is also small.
\end{itemize}
Our model provides predictions for future surveys with next-generation instruments and telescopes, such as ALMA and JWST, that will measure the dust and galaxy properties in extraordinary detail at high redshift. JWST will cover the IR spectrum in the 0.6 to 28.5 micron regime, which reveals the total galactic dust content up to $z=2$. At $z=3$, it will only cover the dust spectrum to $\approx 7$ micron, providing a measure for the mass of hot dust. As we have seen throughout this paper, current dust measurement at high-redshift are only available for the brightest galaxies with high dust mass. Complemented with data from ALMA and other IR and sub-mm surveys, JWST will provide a significantly improved view of high redshift dust across a wider range of masses. This will better constraint dust enrichment scenarios in models such as \dustysage, leading to a better predictions and more detailed interpretation of the observations.
\section*{Acknowledgements}
We would like to thank the anonymous referee whose valuable comments improved the quality of this paper, and Ned Taylor for helpful comments during the final stages of this work. This research was supported by the Australian Research Council Centre of Excellence for All Sky Astro-physics in 3 Dimensions (ASTRO 3D), through project number CE170100013. The Semi-Analytic Galaxy Evolution (SAGE) model, on which \dustysage was built, is a publicly available codebase that runs on the dark matter halo trees of a cosmological N-body simulation. It is available for download at \url{https://github.com/darrencroton/sage}. This research has used \texttt{python} (\url{https://www.python.org/}), \texttt{numpy} \citep{Vanderwalt11} and \texttt{matplotlib} \citep{Hunter07}.
\section*{Data availability}
The data underlying this article are available in the article\DIFaddbegin \DIFadd{. The galaxy formation model used to generate the data is available at }\url{https://github.com/dptriani/dusty-sage}\DIFaddend .
\bibliographystyle{mnras}
\section{Introduction}
\label{sec:intro}
The cosmic star formation history shows a rise of star formation activity in our Universe from high-redshift till redshift $z=2$, known as cosmic noon, and then a decrease towards $z=0$ \citep{Madau14}. Our current understanding is that star formation depends on the availability of $\mathrm{H_2}$ clouds, the raw material for stars \citep{Kennicutt12}. As a galaxy evolves, $\mathrm{H_2}$ clouds are consumed to form stars, and less material is available for the next star formation episode. However, infalling gas can provide a fresh supply which replenishes gas that was consumed by star formation activity \citep[e.g][]{Dekel06, Almeida14}.
Morphology and colour often correlate with star formation in galaxies \citep[e.g.][for a review]{RH94}. Elliptical galaxies are dominated by an old stellar population and appear red \citep{Thomas05, Kormendy09}. These elliptical galaxies have run out their star-forming gas; therefore, they have no or low star formation. At the opposite end, spiral galaxies have bluer emission, showing signs for young stellar populations and active star formation \citep[see also a review from][]{Kennicutt98}.
Per ``conventional'' galaxy evolution theory, galaxies evolve from star-forming spiral galaxies to quenched ellipticals \citep{Bundy06, Skelton12, Tojeiro13}. Dust follows this evolutionary sequence, and galaxies become dust-poor at late-times, either through dust destruction or ejection out of the ISM. Galaxy formation models assume that as star formation activity decreases, mergers and disk instabilities make galaxies more spheroidal \citep{Somerville01, Baugh05, Croton06, Croton16}. During such morphological transformations, a fraction of the ISM dust is ejected through SN and AGN feedback. Therefore, we see a decrease in the ISM dust mass in more elliptical galaxies.
Since the dust content of a galaxy depends on its unique evolutionary history, the observed dust content in the local and high-redshift Universe provides insight into the physics of galaxy formation. Studies of high-redshift galaxies in the far-infrared and submillimetre regime reveal an abundance of star-forming galaxies with massive dust reservoirs \citep{Valiante09}. Dusty galaxies are even found in the very early Universe, including A1689-zD1 at $z=7.5$ with a dust mass of $4 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace \citep{Watson15}, and HFSL3 at $z=6.34$ with a dust mass of $1.3 \times 10^9$ \ensuremath{\mathrm{M}_{\odot}}\xspace \citep{Riechers13}. However, observations also see star-forming galaxies at early and late times with little dust and low metallicity \citep{Fisher14}. The dichotomy between dust-rich and dust-poor galaxies across redshift motivates this work.
Recent galaxy surveys have measured the dust mass function \citep{DEE03, VDE05, Dunne11, Eales09, Clemens13} and many scaling relations between dust and the fundamental properties of galaxies. These include the relationship between dust mass and stellar mass \citep{Santini14} and the relationship between dust mass and star formation rate (SFR) \citep{daCunha10, Santini14}. These scaling relations provide constraints for galaxy evolution models that track the dust properties of galaxies.
Historically, dust in galaxies is commonly investigated using one zone analytical models \citep[e.g.,][]{Dwek1998, ZGT08, Valiante09, Asano13, Zhukovska14}. Such models are critical to assess the various processes of dust production and destruction, especially to reproduce the dust content of specific galaxy populations. However, they generally use a simplistic approach to model gas and stellar evolution in galaxies and do not provide predictions for galaxies within a cosmological volume. Zoom-simulations coupled with self-consistent dust tracking give a more realistic accounting of how dust, metals, gas and stellar populations interact within galaxy evolution framework \citep{Bekki15, Aoyama17}. But they are computationally expensive and tend to focus on reproducing the properties of individual galaxies.
\cite{McKinnon16} was one of the first works to model dust and galaxy co-evolution within a cosmological simulation. They incorporated a detailed dust prescription to a hydrodynamical simulation, including stellar production, grain growth and grain destruction by supernovae shocks. To date, a few semi-analytic galaxy models have included self-consistent dust modelling. These models can reproduce a number of global trends between dust and various galaxy properties across cosmic time, with less detail but more efficient computing cost and time compared to the hydrodynamical simulations \citep{Popping17, Vijayan19, Triani20}.
In this paper, we extend the analysis of \citet{Triani20} to investigate the different characteristics between dust-rich and dust-poor simulated galaxies using the \dustysage\footnote{\url{https://github.com/dptriani/dusty-sage}} semi-analytic model. Dust in the ISM forms in stellar ejecta \citep{Dwek1998, ZGT08}, grows further via grain accretion \citep{Draine90, Dwek1998, ZGT08, Draine09}, and is destroyed by supernovae shocks \citep{DS80, Jones94, ZGT08, Slavin15}. \dustysage includes analytical prescriptions for these mechanisms, as well as dust locked in stars (astration), and dust inflows from and outflows to the halo and ejected reservoirs. In the model, dust undergoes further destruction via thermal sputtering in the halo and ejected reservoirs due to their high temperature. The primary constraints for dust modelled by \dustysage are the observed dust mass function and dust mass - stellar mass relation at $z=0$. \dustysage includes dust tracking in the ISM, halo, and ejected dust by feedback processes. It provides predictions for the dust mass function and the relation between dust mass and stellar mass at high redshift.
This paper is organised as follows: In Section \ref{sec:dustysage}, we provide a brief description of \dustysage. We study how the dust fraction in each reservoir relates to the stellar mass and morphology in Section \ref{sec:morphologies}. Then we categorise galaxies into groups based on their ISM and halo dust content in Section \ref{sec:classification}. In Section \ref{sec:relation}, we map these groups based on their morphology, star formation activity and stellar mass at redshift $z=0$. We extend the relations to higher redshift, up to $z=3$, in Section \ref{sec:evolution}. Section \ref{sec:discussion} and \ref{sec:conclusion} provide a discussion and conclusion, respectively. Throughout this paper, we assume $h = 0.73$ based on the cosmology used of the Millennium simulation.
\section{The semi-analytical model - Dusty SAGE}
\label{sec:dustysage}
Here we provide only a brief overview of the galaxy and dust model used in this work. \dustysage is built on the more generic galaxy model \texttt{SAGE}\xspace \citep{Croton06, Croton16}, but with a detailed prescription that tracks dust evolution. The summary of both galaxy and dust evolution mechanisms in \texttt{Dusty SAGE} is presented in Figure 1 in \citet{Triani20}. The model follows baryonic growth in dark matter halos taken from an N-body simulation. There are four baryonic reservoirs: the pristine gas, the hot halo, the ISM and an ejected reservoir. In this paper, we run \dustysage on the Millennium simulation \citep{Springel05} and select galaxies with stellar mass $\log M_* = 9 - 12$ \ensuremath{\mathrm{M}_{\odot}}\xspace.
Mass is exchanged among the different baryonic reservoirs across cosmic time. Pristine gas falls into the collapsed dark matter halo and is heated to the virial temperature -- the hot halo reservoir \citep{RO77, WR78}. A fraction of this hot gas then cools into the ISM \citep{SD93}. In the ISM, cold hydrogen differentiates into atomic and molecular hydrogen \citep{HMK79}. Stars then form from molecular clouds \citep{Kennicutt12}. During their evolution, stars create helium and elements heavier than helium in their core, known as `metals'. These metals are expelled in stellar ejecta and change the metallicity of the ISM (ratio of metals and gas mass). Massive stars end their life as supernovae and inject a large amount of energy into the cold ISM. This energy reheats cold ISM gas back to the hot halo reservoir, and potentially even expels it outside the halo -- the `ejected' reservoir \citep[e.g.,][]{MW03}. This process is known as SN feedback. The gas accretion onto supermassive black holes provides another source of feedback; this energetic phenomenon is known as active galactic nuclei (AGN) \citep{Croton06}. The ejected gas can later be reincorporated back to the halo and galaxy.
In both \dustysage and \texttt{SAGE}\xspace \citep{Croton06, Croton16}, the bulge-to-total stellar mass ratio represents the morphology of the galaxy, which evolves during disk instabilities and mergers. We adopt the disk stability criteria of \citet{1998MMW} and transfer sufficient stellar mass from the disk to the bulge to ensure stability every time an instability occurs. In a merger between a more massive central galaxy and a less massive satellite galaxy, bulge enrichment depends on the total stellar and gas mass ratio of both progenitors. If the mass ratio exceeds 0.3, a `major' merger has occurred: the disks of both galaxies are destroyed and all stars are placed in a bulge. Otherwise, the merger is `minor' and the satellite stellar content is added to the central bulge. Both instabilities and mergers can trigger a starburst. Unlike quiescent disk star formation, stars formed in a merger-induced starburst are placed in the bulge.
Besides the usual galaxy evolution processes, \dustysage also incorporates a detailed dust evolution model. Analogous to the gas, we track dust in three reservoirs -- the ISM, the hot halo and an ejected reservoir. Dust is mainly processed in the ISM, then heated to the halo and ejected reservoirs via feedback mechanisms powered by supernovae and AGN. The total dust production rate in the ISM is described by:
\begin{equation} \label{eq:total_rate}
\dt{M}_\mathrm{d} = \dt{M}_\mathrm{d}^\mathrm{form} + \dt{M}_\mathrm{d}^\mathrm{growth} - \dt{M}_\mathrm{d}^\mathrm{dest} - \dt{M}_\mathrm{d}^\mathrm{SF} - \dt{M}_\mathrm{d}^\mathrm{outflow} + \dt{M}_\mathrm{d}^\mathrm{inflow},
\end{equation}
where:
\begin{itemize}
\item $\dt{M}_\mathrm{d}^\mathrm{form}$ is the stellar dust formation rate: In every star formation episode, \dustysage tracks the abundance of C, N and O in AGB stars and C, O, Mg, Si, S, Ca and Fe in SN II ejecta. The condensation of these elements to form dust is given in Table \ref{tab:params}.
\item $\dt{M}_\mathrm{d}^\mathrm{growth}$ is the grain growth rate in dense molecular clouds: Existing grains grow via metal accretion \citep{Dwek1998, ZGT08} where the timescale for this process depends on the metal abundance in the cold gas.
\item $\dt{M}_\mathrm{d}^\mathrm{dest}$ is the destruction rate via SN shocks: We follow the prescription from \cite{DS80, Mckee89, Asano13} to compute a destruction timescale from the total cold gas mass and the supernovae rate, efficiency and swept mass.
\item $\dt{M}_\mathrm{d}^\mathrm{SF}$ is the rate for dust locked in newly formed stars, which is proportional to the star formation rate and the dust-to-gas ratio.
\item $\dt{M}_\mathrm{d}^\mathrm{outflow}$ and $\dt{M}_\mathrm{d}^\mathrm{inflow}$ are the dust outflow and inflow rates. SN and AGN feedback can reheat cold ISM gas and expel it to the halo. The feedback energy, if large enough relative to the depth of the potential well, can even eject the gas to leave the galaxy and host halo. In an outflow, we assume that the dust-to-gas ratio of the ejected gas is equal to the ISM. Ejected gas can be reincorporated back to the halo, while maintaining the dust-to-gas ratio of the ejected reservoir. Hot gas undergoes cooling process back to the disk, and we assume in this inflow the dust-to-gas ratio equals that of the halo.
\end{itemize}
Dust populates the halo and ejected reservoirs through outflows from the ISM. In both reservoirs, dust grains are destroyed via thermal sputtering on a short timescale that depends on the gas density and temperature. In both the halo and ejected reservoirs, we assume the virial temperature, with an isothermal density profile for the hot gas extending to the virial radius, and a uniform density profile for the ejected component. In every timestep, the ejected reservoir's density is evaluated by dividing the gas mass with the reservoir's volume, assuming the virial radius. However, the computed timescale should be taken as the upper limit for the ejected reservoir. In reality, the ejected reservoir is a mix of the circumgalactic medium (CGM) and intergalactic medium (IGM). Gas in the IGM might extend beyond the virial radius, resulting in a lower density. The temperature might also be higher than in the model, allowing for a more efficient sputtering.
Although thermal sputtering is very efficient, we still find a significant abundance of dust in the halo. Dust properties in the halo depend on the balance between the outflows and inflows from/to the ISM. The depth of the galaxy's potential well also affects the outflow mass trapped in the halo. Galaxies with a shallow potential are more likely to lose the outflowing materials to the ejected reservoir. We assume that the gas outflow carries dust in the same proportion as the ISM and transfers it to the halo, with no dust destruction in the process. It may well be that a fraction of dust is destroyed when ejected out of the ISM, or the DTG ratio of the outflow differs from the ISM, which will alter our predictions. As we mentioned above, it is also possible that the sputtering rates are higher than our prediction. However, the nature of such processes is currently unclear. Future observations to quantify dust properties in the galactic wind and the IGM will provide additional constraints to these processes.
Figure 14 in \citet{Triani20} shows that galaxies have more dust in their halo than in the ISM at low-redshift. This outcome roughly agrees with the massive dust content found in the CGM \citep{Dunne11, PMC15}. The galaxy formation model of \citet{Popping17} also predicts a significant amount of dust in both the halo and ejected reservoir. However, their dust mass density is notably higher than our results.
\dustysage provides a good agreement with the galaxy stellar mass function at redshift $z=0$. It also successfully reproduces the dust mass function and various dust scaling relations over a wide range of redshifts \citep{Triani20}. To achieve a more realistic distribution of the stellar mass in the bulge and disc, we make slight changes to a few of the parameters in \citet{Triani20}. These are listed in Table \ref{tab:params}.
\begin{table*}
\caption{Fiducial \dustysage parameters used throughout this work, also compared to those from \citet{Triani20}.}
\label{tab:params}
\begin{tabular}{lccc}
\hline
\hline
Parameter & Description & Value &\citet{Triani20} \\
\hline
\hline
$f_b^\mathrm{cosmic}$ & Cosmic baryon fraction & 0.17 & 0.17\\
\hline
$z_0$ & Redshift when $\mathrm{H_{II}}$ regions overlap & 8.0 & 8.0\\
$z_r$ & Redshift when the intergalactic medium is fully reionized & 7.0 & 7.0\\
\hline
$\alpha_\mathrm{SF}$ & Star formation efficiency from $\mathrm{H_2}$ [$\mathrm{Myr^{-1}}$] & 0.005 & 0.005\\
$R$ & Instanteneous recycling fraction & 0.43 & 0.43\\
\hline
$\epsilon_\mathrm{disc}$ & Mass-loading factor due to supernovae & 2.0 & 3.0\\
$\epsilon_\mathrm{halo}$ & Efficiency of supernovae to unbind gas from the hot halo & 0.2 & 0.3\\
$k_\mathrm{reinc}$ & Velocity scale for gas reincorporation & 0.15 & 0.15\\
\hline
$\kappa_\mathrm{R}$ & Radio mode feedback efficiency & 0.09 & 0.08\\
$\kappa_\mathrm{Q}$ & Quasar mode feedback efficiency & 0.005 & 0.005\\
$f_\mathrm{BH}$ & Rate of black hole growth during quasar mode & 0.015 & 0.015\\
\hline
$f_\mathrm{friction}$ & Threshold subhalo-to-baryonic mass for satellite disruption or merging & 1.0 & 1.0\\
$f_\mathrm{major}$ & Threshold mass ratio for merger to be major & 0.15 & 0.3\\
\hline
$\alpha_\mathrm{burst}$ & Exponent for the powerlaw for starburst fraction in merger & 0.18* & 0.7\\
$\beta_\mathrm{burst}$ & Coefficient for starburst fraction in merger & 0.75* & 0.56\\
\hline
$\delta_\mathrm{C}^\mathrm{AGB}$ & Condensation efficiency for AGB stars & 0.2 & 0.2\\
$\delta_\mathrm{C}^\mathrm{SNII}$ & Condensation efficiency for SN II & 0.15 & 0.15\\
$\tau_\mathrm{acc,0}$ & Accretion timescale for grain growth [$\mathrm{yr}$] & $4.0 \times 10^5$ & $4.0 \times 10^5$\\
$f_\mathrm{SN}$ & Fraction of destroyed dust to the swept dust mass by SN & 0.1 & 0.1\\
\hline
\hline
\multicolumn{3}{l}{*value adopted from \citet{Somerville01}}\\
\end{tabular}
\end{table*}
\section{How morphology and stellar mass relate to dust fraction in the hot halo, ISM and ejected reservoir}
\label{sec:morphologies}
In this section, we investigate the relation between galaxy morphology, represented as the bulge-to-total (BTT) mass ratio, and the fraction of dust in the ISM, halo and ejected reservoirs relative to the total dust mass in all reservoirs. We use stellar mass in computing the BTT mass ratio and assume the gas mass to be negligible. We present the median dust fraction vs BTT mass ratio in three stellar mass bins in Figure \ref{fig:btt-fraction}. The grey histogram shows the distribution of BTT mass ratios across all galaxies. In all panels, galaxies with higher BTT mass ratio contain most of their dust in the halo while those with a lower BTT mass ratio tend to keep theirs in the ISM. For the low mass galaxies in the bottom panel, spiral galaxies contain a significant fraction of their dust in the ISM and ejected reservoir. These galaxies make up more than 50\% of the overall population. Conversely, the halo contains nearly 90\% of the total dust in elliptical galaxies with BTT > $0.8$.
For the galaxies with $\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10 - 11$ (central panel), the same trend occurs. Galaxies with a BTT mass ratio smaller than $\sim 0.2$ keep almost half of their dust in the ISM. For the most massive galaxies (top panel), all galaxies contain most of their dust in the halo. Massive galaxies are more likely to host AGNs compared to low-mass galaxies. The stronger AGN feedback is more likely to blow dust out into the halo (and ejected reservoir) compared to the weaker SN feedback commonly found in low-mass galaxies. Massive galaxies also have deeper potential wells than the low mass systems, allowing them to capture reheated dust in the hot halo. However, although the ISM dust fraction is smaller than in Milky Way sized galaxies, their actual dust mass in the ISM is significant and will likely make a noticeable contribution to the galaxy SED in the infrared.
Observations have found galaxies with a massive amount of dust in their CGM. For a galaxy with stellar mass $\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10.1$, \citet{Peeples14} found a fiducial value of metals locked in CGM dust of $5 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace, almost twice the value for those found in the form of ISM dust, $2.6 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace. \citet{Menard10} derived a CGM dust mass of $5 \times 10^7$ \ensuremath{\mathrm{M}_{\odot}}\xspace for galaxies with stellar mass $\log \mathrm{M_*} = 10.4 \ensuremath{\mathrm{M}_{\odot}}\xspace$. Although the CGM is not exactly equivalent to our halo gas component, we will use CGM dust as a proxy to test our predictions of the dust bond to, but outside the galaxy.
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{images/btt-fraction-mass-512files.png}
\caption{The median value for dust mass fraction in each baryon reservoir versus bulge-to-total stellar mass ratio (BTT). The purple, orange and green lines represents hot dust mass in the halo, cold dust mass in the ISM, and dust in the ejected reservoir, respectively. The grey histogram represents the number distribution of galaxies in 10 BTT bins.}
\label{fig:btt-fraction}
\end{figure*}
\section{Classifying galaxies based on their dust and fundamental properties}
\label{sec:classification}
To investigate how the dust content of galaxies correlates with their stellar mass, star formation activity, and morphology, we first establish our simulated galaxies into groups. In the \dustysage model, the dust in galaxies is distributed in three components - the ISM, the halo and the ejected reservoir. We exclude the dust in the ejected reservoir in our analysis since it is out of the system. Our four dust groups are described below.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/dust-stellarmass-512files.png}
\caption{The cold dust mass of simulated galaxies as a function of stellar mass at redshift $z=0$. The heatmap shows the density distribution with a brighter colour representing higher density, and the black line marks the median. The red line marks the threshold where the cold dust mass per stellar mass equals $10^{-4}$. Above this line, galaxies are classified as ISM dust-rich while below are classified as ISM dust-poor.}
\label{fig:dust-stellarmass}
\end{figure}
\subsection{ISM dust-poor}
\label{ssec:ISM-dust-poor}
When considering dust, observers typically report the properties of the ISM \citep[e.g.,][]{Eales09, RR14, Santini14, Mancini15, daCunha15, Nersesian19}. Although some works have extended dust measurements to include the CGM \citep{Menard10, PMC15}, such observations are rare. Therefore, as a first step, we have classified our model galaxies based on their cold dust content in the ISM. Figure~\ref{fig:dust-stellarmass} shows the relation between cold dust mass in the ISM and stellar mass of our model galaxies at redshift $z=0$. We find that the dust mass in the ISM increases with the host stellar mass, but with a significant scatter.
The median values of our predicted ISM dust mass versus stellar mass are in a good agreement with the observations of dust in local galaxies, such as from the DustPedia catalogue \citep{Nersesian19}. Figure \ref{fig:dust-stellarmass} shows a large scatter below the median values, marking galaxies with relatively less dust compared to the overall population.
We draw an arbitrary line at dust mass per stellar mass of $10^{-4}$, marked with the red line in Figure \ref{fig:dust-stellarmass} to divide the galaxy population into two categories: ISM dust-rich and ISM dust-poor. The ISM dust-poor group contains galaxies below the red line. Observations have found such ISM dust-poor galaxies in both the local and high-redshift Universe \citep{Fisher14}.
Further exploration reveals that \dustysage galaxies in this ISM dust-poor category vary in their halo dust content. Therefore, we additionally divide the ISM dust-poor group based on their fraction of halo dust per stellar mass described below.
\subsubsection{Halo dust-rich}
\label{sssec:halo-dust-rich}
Several observations \citep{PMC15, Peeples14, Menard10} extended their search for metals and dust to the CGM. \citet{Peeples14} find that only $25\%$ of metals created in stars stay in the ISM; a similar fate also occurs for dust. \citet{Menard10} found galaxies with massive CGM dust. These observations discovered the existence of dust out of the ISM, which is reproduced by our model.
In our model, a mix of halo and ejected dust might be a better representation of the CGM for some galaxies. However, we only use halo dust to represent the observed CGM dust. We use the same threshold as the ISM dust-mass to define the halo-dust rich galaxies: halo dust mass per stellar mass of $10^{-4}$. Halo dust-rich galaxies are defined as ISM dust-poor galaxies with halo dust mass above this threshold.
\subsubsection{Dust-poor}
\label{sssec:dust-poor}
The population of galaxies that lack dust in both their ISM and halo is classified as dust-poor. Due to the rarity of dust measurements outside the ISM, we found no counterpart of this category from the observations. Their existence in our model serves as a prediction for future surveys measuring the CGM dust.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/sSFR-hist-allmass-512files.png}
\caption{The distribution of the specific star-formation rate in our ISM dust-rich galaxies. The grey line at $2.71 \times 10^{-11} \mathrm{yr^{-1}}$ is the Milky Way value adopted from \citet{Licquia15} which we use as a threshold in defining quenched and star-forming galaxies in our model galaxies.}
\label{fig:sfr-hist}
\end{figure}
\subsection{ISM dust-rich}
\label{ssec:ISM-dust-rich}
We define the ISM dust-rich galaxies in our model as those with a fraction of ISM dust mass per stellar mass above $10^{-4}$. Dust obscures the intrinsic stellar spectra of such galaxies and re-emits the radiation in the infrared \citep{WTC92, WG00}. Infrared and sub-millimeter surveys have found a substantial number of these galaxies, especially at low redshift \citep[e.g.,][]{Eales09, DEE03, daCunha15, Clemens13, RR14}.
Dust accounting in the galactic ISM is the main focus of many current and future infrared surveys. Forthcoming galaxy survey, like those using JWST, will measure dust in high redshift galaxies. Such surveys will provide additional constraints to our model predictions.
Besides the relation between dust mass and stellar mass shown in Figure \ref{fig:dust-stellarmass}, observations have also found a relation between dust mass and SFR \citep{daCunha10, Casey12, Santini14}. Based on this, we divide our ISM dust-rich galaxies further into star-forming and quenched subclasses.
\subsubsection{Star-forming ISM dust-rich}
\label{sssec:sf-dust-rich}
Figure \ref{fig:sfr-hist} shows the specific SFR (sSFR) distribution of the ISM dust-rich galaxies in our model. To separate star-forming from quenched galaxies, we use the Milky Way sSFR value as the threshold, $2.71 \times 10^{-11} \mathrm{yr^{-1}}$ \citep{Licquia15}. Galaxies with sSFR above the Milky Way value are defined as the star-forming ISM dust-rich.
\subsubsection{Quenched ISM dust-rich}
\label{sssec:q-dust-rich}
ISM dust-rich galaxies are commonly related to star-forming spirals since dust is initially formed in stellar ejecta and grows in molecular clouds \citep{Valiante09}. However, we also find a population of ISM dust-rich galaxies with relatively low star formation activity in our model (see Figure \ref{fig:sfr-hist}). The quenched ISM dust-rich galaxies are those with a sSFR below the Milky Way value of $2.71 \times 10^{-11} \mathrm{yr^{-1}}$.
\section{The relation between dust mass and fundamental galaxy properties}
\label{sec:relation}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{images/btt-ssfr-Z-512files.png}
\caption{\textit{Top panel} shows the BTT ratio, sSFR and gas-phase metallicity of our model galaxies in the ISM dust-poor group. This group consists of two subgroups: halo dust-rich and dust-poor. Markers with small, medium and large size represents three mass bin: $\log \mathrm{M_*} = 9 - 10$ \ensuremath{\mathrm{M}_{\odot}}\xspace, $\log \mathrm{M_*} = 10 - 11$, and $\log \mathrm{M_*} = 11 - 12$ \ensuremath{\mathrm{M}_{\odot}}\xspace, respectively. The points mark the median sSFR and median BTT mass ratio of each group, with the errorbars spanning the $16^\mathrm{th}$ and $84^\mathrm{th}$ percentiles. The colour of each point marks the gas-phase metallicity, with brighter colours representing higher value. We assume a solar metallicity of 0.02 and $12 + \log[\mathrm{O/H}] - \log[\mathrm{Z/Z_\odot}] = 9.0$. The number inside parentheses in the legend represents the percentage of galaxies in each category with respect to the total number of galaxies with stellar mass $\log \mathrm{M_*} = 9 - 12$ at $z=0$. Open symbols indicate equivalent observational dataset taken from DustPedia \citep{Nersesian19} and KINGFISH \citep{Kennicutt11, RR14}.
\textit{Bottom panel} shows the same relation above but for the ISM dust-rich group, which is splitted into star-forming and quenched subgroups.}
\label{fig:btt-sfr}
\end{figure*}
In the previous sections, we categorized model galaxies into four classes: star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich and dust-poor. We now plot the BTT mass ratio vs sSFR of each group in three mass bin in Figure \ref{fig:btt-sfr} as indicated in the legend, with the colour distribution representing the median gas-phase metallicity of each group. The error bars denote the $16^\mathrm{th}$ and $84^\mathrm{th}$ percentile range.
Figure~\ref{fig:btt-sfr} shows a general trend where the ISM dust-rich galaxies in the bottom panel occupy the lower right space marking disky galaxies with higher star formation activity. Meanwhile, ISM dust-poor galaxies in the top panel generally have lower sSFR and a larger BTT mass ratio. Since galaxies typically change from star-forming spirals to quenched ellipticals over time, this plot indicates that the ISM also evolves from being dust-rich to dust-poor across h.
The gas-phase metallicity in Figure~\ref{fig:btt-sfr} increase with stellar mass, as one may expect from the galaxy mass-metallicity relation \citep{Tremonti04}. Both ISM dust-rich groups generally have higher metallicity than the ISM dust-poor groups.
Star-forming ISM dust-rich galaxies have the highest sSFR and lowest BTT mass ratio. Condensation in stellar ejecta is one of the main dust production channels in our model and is tightly correlated with star formation activity. Thus, we would expect ISM dust-rich galaxies to have a high sSFR. It is also possible that the high sSFR of such galaxies reflects the abundance of molecular gas in the ISM, which enables dust growth via accretion of metal grains.
Quenched ISM dust-rich galaxies have much lower sSFR (on average) and are also located in the low BTT-ratio regime. Their gas-phase metallicity is higher than that of the star-forming ISM dust-rich galaxies. As galaxies in this group have less sSFR than those in the star-forming group, their stellar dust production must be less efficient. However, their abundance of dust can be explained by their high availability of gas-phase metals, which can accrete onto existing dust grains in the dense environment. The low sSFR of these galaxies also suggest that they lack massive young stars that become SN, the primary destroyer of dust in the ISM. The scarcity of SN can also prevent the dust and refractory elements from being expelled, allowing them to preserve more dust and metals in their ISM.
In contrast to the ISM dust-rich galaxies, the ISM dust-poor group have higher BTT mass ratios (i.e., more elliptical), lower average sSFRs and lower gas-phase metallicities. The halo dust-rich group have the highest BTT mass ratio, incorporating the ``bulgiest'' galaxies for all mass bins. Galaxies in this group also have the lowest gas metallicities. To explore this further, we plot the distribution of the ratio between the ISM cold gas mass and the hot halo gas mass in Figure~\ref{fig:hot-cold-ratio}. Here, the halo dust-rich group peaks lowest of all groups, showing that galaxies in this group either have less cold gas in their ISM or more gas in their halo. The fact that these galaxies have massive halo dust and hot gas, yet they lack ISM gas and dust, indicates an efficient mechanism to transport their ISM content out of the disk.
Figure~\ref{fig:outflowrate} shows the fraction of galaxies in each group versus outflow rate. The halo dust-rich group dominates for outflow rates above $30\ \ensuremath{\mathrm{M}_{\odot}}\xspace \mathrm{yr^{-1}}$. This suggests that the efficient outflow in these galaxies can blow their ISM content out of the disk, while their low sSFR does not allow them to replenish the metals and dust in the ISM.
Looking back to Figure \ref{fig:btt-sfr}, we see that the total percentage of galaxies in the most massive mass bin ($\log \mathrm{M_*} = 11 - 12$) is $1.47 \%$ and the halo dust-rich group makes up a quarter of them, a more substantial contribution compared to those in the lower mass bins. In \dustysage, massive galaxies are likely to have deeper potential wells, which allow them to capture most of the dust outflow from the ISM in the halo. Therefore, very little dust escapes into the ejected reservoir in feedback events. Figure~\ref{fig:btt-fraction} also shows that in the most massive galaxies, the majority of the dust mass is in the hot halo, and an only small fraction of their dust fraction is in the ejected reservoir at all BTT mass ratios.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/hot-cold-gas-512files.png}
\caption{The probability distribution of the cold to hot gas mass ratio of our simulated galaxy categories. Red, blue, green and purple lines represent the star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich and dust-poor groups, respectively.}
\label{fig:hot-cold-ratio}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{images/outflowrate-512files.png}
\caption{The gas outflow rate distribution due to feedback. We assume that the dust-to-gas mass ratio in the outflow is the same as the general ISM.}
\label{fig:outflowrate}
\end{figure}
In the dust-poor category, we find an anti-correlation between stellar mass and sSFR. The median sSFR of dwarf galaxies is $\log \mathrm{sSFR}=-10.6$, more than two orders of magnitude larger than those in the most massive bin. Galaxies will become dust-poor if their dust depletion rate exceeds the dust production rate. Since Figure \ref{fig:dust-stellarmass} shows a positive correlation between dust mass and stellar mass, more massive galaxies are more likely to have more dust in their ISM. Therefore, only those galaxies with a very low production rate can become dust-poor. The anti-correlation reflects the link between dust and star formation rates.
Figure \ref{fig:btt-sfr} also allows us to examine the stellar mass evolution across the four groups. At the massive end, galaxies with mass $\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 11 - 12$ are comprised primarily of the star-forming ISM dust-rich category (0.6 \% of the total population), followed by the quenched ISM dust-rich and halo dust-rich groups each containing 0.3\% of the total galaxies. The dust-poor group is in the minority amongst massive galaxies. The dominance of the ISM dust-rich and halo dust-rich groups shows that the most massive galaxies rarely destroy or eject their dust out of the halo. Massive late-type and early-type galaxies contain the majority of their dust in the ISM and the halo, respectively.
In the middle bin ($\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10 - 11$), the biggest percentage is still contributed by the star-forming ISM dust-rich category (14 \%), followed by the quenched ISM dust-rich with 4.9 \%, halo dust-rich with 4.6 \%, and the dust-poor group with 1.1 \%. In the lowest mass bin ($\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 9 - 10$), the star-forming ISM dust-rich still dominates (28\%). However, the contributions of quenched ISM dust-rich and dust-poor groups in this mass bin are also the most significant (13.9 \% and 12.8 \%, respectively) compared to those in the higher mass bins. The halo dust-rich group shares the smallest percentage for dwarf galaxies with only 3.6\%.
\subsection{Comparison with observed datasets}
\label{ssec:compared-with-data}
To add context to our theoretical predictions, we make comparisons with two observation datasets. DustPedia \citep{Nersesian19} is a compilation of local galaxies with dust measurements. To create this, the authors use the dust model \texttt{THEMIS} in the SED fitting code \texttt{CIGALE} to infer dust mass and galaxy properties from \textit{Herschel} multiwavelength photometry. The authors use numerical Hubble T as the morphology indicator, with lower value corresponds to a larger stellar composition in bulge relative to the disk. To convert their Hubble T value to BTT-ratio, we use the K-band bulge-to-disk flux ratio mapping from \citet{Graham08}. Note that the BTT ratio for our simulated galaxies is based on stellar mass, while the observations are based on flux.
We group their data based on our definition of the star-forming ISM dust-rich, quenched ISM dust-rich and ISM dust-poor galaxies. We can not divide their ISM dust-poor galaxies further because they do not provide a halo dust mass. We did not find galaxies classified as quenched ISM dust-rich in their dataset. Therefore, we plot the mean values of BTT mass ratio and sSFR of their star-forming ISM dust-rich group as a pink-filled circle and their ISM dust-poor group as a pink-filled square in Figure \ref{fig:btt-sfr}.
Our second observational dataset is taken from the KINGFISH survey \citep{Kennicutt11}. We group KINGFISH galaxies based on our definitions and compute the mean dust mass, BTT mass ratio, and SFR for each group. We plot the mean dust and galaxy properties with markers showing the group they belong. The dust measurement for KINGFISH galaxies is taken from \citet{RR14}.
Our prediction for star-forming ISM dust-rich and ISM dust-poor galaxies are in good agreement with the observed values. In both the KINGFISH and DustPedia data, dust-poor galaxies are those with the most elliptical morphology (E-type) with Hubble T below -2 \citep{RR14, Nersesian19}. However, the map of \citet{Graham08} that we use to convert the Hubble type to BTT ratio only extends to S0-type galaxies (Hubble T = -2). Therefore, we use the Hubble T = -2 conversion for elliptical galaxies with Hubble T < -2. The converted BTT flux ratio of these galaxies are $0.15$ dex lower than our median mass ratio but are still within our $16^\mathrm{th}$ percentile.
\section{Evolution of the relations between dust and fundamental properties}
\label{sec:evolution}
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{images/btt-ssfr-Z-evol-512files.png}
\caption{The evolution of the relation shown in Figure \ref{fig:btt-sfr} between redshift $z=0$ and redshift $z=3$. For clarity, we focus on the middle mass bin ($\log M* = 10 - 11$ \ensuremath{\mathrm{M}_{\odot}}\xspace). Circle, inverted triangle, diamond and square represent the median value of the star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich and dust-poor groups, respectively.}
\label{fig:btt-sfr-evol}
\end{figure*}
Galaxy evolution models predict that galaxies evolve from star-forming spiral to quench elliptical (i.e., the bottom right to top left in Figure \ref{fig:btt-sfr}) \citep{Somerville01, Baugh05, Croton06, Croton16}. The dust content of the ISM follows this evolutionary timeline to become dust-poor over time. There are two primary depletion mechanisms for ISM dust: destruction by SN shocks \citep{Slavin15} and ejection by SN and AGN feedback \citep{Feldmann15, Popping17}. Ejected dust can be trapped in the halo if the potential well is sufficiently deep, causing the host to end up as a halo dust-rich galaxy. Otherwise, such galaxies end up as dust-poor.
To see the evolution of the dust content in galaxies across time, we plot the relation between BTT mass ratio and sSFR of our dust groups from redshift $z=0$ to redshift $z=3$ in Figure \ref{fig:btt-sfr-evol}. Here, we focus on the middle mass bin ($\log \mathrm{M_*}\ (\ensuremath{\mathrm{M}_{\odot}}\xspace) = 10 - 11$) as representative of the Milky Way mass range. Circle, inverted triangle, diamond, and squares represent the star-forming ISM dust-rich, quenched ISM dust-rich, halo dust-rich, and dust-poor categories, respectively. This plot does not show the evolution of the same group of galaxies. Instead, it shows us the map of galaxies belong to each group at different redshift in the BTT mass ratio and sSFR plane.
Figure~\ref{fig:btt-sfr-evol} shows how the percentage of galaxies in each group changes over time. The fraction of star-forming ISM dust-rich galaxies peaks at $z=1$ then remains at $\approx 25\%$ until $z=0$. The other three groups show lower fractions but continue to increase down to the present day. This evolutionary trend reveals how star-forming ISM dust-rich galaxies might convert into the other groups at lower redshift.
Across redshift, the behaviour of individual groups within the BTT mass ratio - sSFR space also changes. In the BTT mass ratio dimension, the quenched ISM dust-rich group gets diskier towards $z=0$ with a $0.3$ drop, while the other three groups do not show notable changes between $z=3$ and $z=0$. All groups except the quenched ISM dust-rich shows a significant decrease of the median sSFR from $z=3$ to $z=0$. The biggest sSFR decrease occurs in halo dust-rich galaxies, by more than 2 order of magnitude, while the star-forming ISM dust-rich and dust-poor galaxies each undergo a single dex decrease.
In the gas phase metallicity dimension, the halo dust-rich group have the lowest median value across redshift $z=3$ to $z=0$. The median metallicity of both dust-poor and star-forming ISM dust-rich galaxies increases slightly with decreasing redshift. Meanwhile, the median metallicity of the quenched ISM dust-rich group fluctuates with redshift; it increases $\approx 0.1$ dex between $z=3$ and $z=2$, does not change between $z=2$ and $z=1$, then decreases $\approx 0.3$ dex.
\section{Discussion: the behaviour of each group across redshift}
\label{sec:discussion}
From Figure~\ref{fig:btt-sfr} and Figure~\ref{fig:btt-sfr-evol}, we can see that although we divide galaxies into groups solely based on their dust content, the galaxy properties of each group are often distinct. Also, galaxy properties within each group evolve with redshift. Our predictions for how these groups change with time can help us understand the baryonic and dust physics occurring. In this section, we discussed the features of these groups between redshift $z=3$ and the present day.
\subsection{The star-forming ISM dust-rich group}
At all mass bins at redshift $z=0$, the star-forming ISM dust-rich group consistently stays at the lower right region in the BTT mass ratio - sSFR space (see Figure~\ref{fig:btt-sfr}). This indicates that these galaxies are still actively forming stars and have an abundance of molecular gas in their ISM, providing a dust production channel via condensation in stellar ejecta and grain growth in the molecular clouds. The gas-phase metallicity in this group is relatively high compared to those in the halo dust-rich and dust-poor groups. This group dominates the galaxy population for all mass bins at redshift $z=0$.
Across redshift, this group stays in the disky star-forming region (see Figure~\ref{fig:btt-sfr-evol}) and continues to dominate in the overall population. The median gas-phase metallicity is lower at higher redshift. This metallicity increase with decreasing redshift reflects the ability of such galaxies always to retain an abundance of metals in their ISM. Since galaxies in the star-forming ISM dust-rich group have high sSFR by the selection, we expect that the ISM will be both metal and dust-rich. Observations of the local and high redshift Universe find many galaxies that belong to this group. Figure \ref{fig:btt-sfr} shows that at redshift $z=0$, both the KINGFISH~\citep{Kennicutt11, RR14} and DustPedia~\citep{Nersesian19} datasets have galaxies with dust-to-stellar mass and sSFR classifications similar to our star-forming ISM dust-rich group.
Figure \ref{fig:dtm-ssfr} shows the sSFR of our model galaxies as a function of dust-to-stellar mass ratio from redshift $z=0$ to $z=3$. The white line marks the divider between the ISM dust-rich and ISM dust-poor groups. Across redshift, many authors find star-forming galaxies with a dust rich ISM. The general trend that the sSFR decreases with decreasing redshift is seen both for the model and in the observations. At $z=0$, our ISM dust-rich galaxies are in good agreement with the observational values from DustPedia \citep{Nersesian19}, \citet{RR14} and \citet{Santini14}. For the ISM dust poor galaxies below the grey line, the observations are lower than our median value but are still within the $16^\mathrm{th}$ percentile. At higher redshift, observations are yet to find ISM dust-poor galaxies. At $z=1$, our model roughly agrees with the dust-rich data from \citet{Santini14}, although our median value of sSFR lies $0.1 - 0.4$ dex below the observed values. At $z=2$ and $z=3$, however, our model fails to reproduce galaxies with $\log (\mathrm{M_{dust}} / \mathrm{M_{star}}) > -2.5$ where most of the observational datasets lie. \dustysage and our base model \texttt{SAGE}\xspace \citep{Croton16} are constrained using various galaxy observations at $z=0$; therefore, we still find limitations in reproducing the high redshift population. We also note that the dataset from \citet{daCunha15} (red triangles) is obtained in the ALMA LESS survey which is biased towards the extremely bright galaxies.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{images/dtm-ssfr-millennium.png}
\caption{The relation between specific star formation rate (sSFR) and the dust-to-stellar mass ratio from redshift $z=0$ to $z=3$. The heat map shows the 2D density distribution of our model galaxies with brighter color representing higher density. The solid black lines mark the median while the dashed lines marked the $16^\mathrm{th}$ and $84^\mathrm{th}$ percentile. The white vertical lines mark a log dust-to-stellar mass ratio = -4, where we divide the ISM dust-rich and dust-poor categories. The red circles, blue diamonds, green squares and red triangles are observational values from DustPedia: \citet{Nersesian19}, \citet{RR14}, \citet{Santini14} and \citet{daCunha15}, respectively.}
\label{fig:dtm-ssfr}
\end{figure*}
\subsection{The quenched ISM dust-rich group}
The quenched ISM dust-rich group consists of galaxies that have an abundance of dust in their ISM despite their low star formation activity (defined as galaxies with sSFR below the Milky Way). One possible scenario for this is that due to the low sSFR, these galaxies no longer host supernovae events that destroy dust and expel metals and dust from the ISM. This also explains the higher gas-phase metallicity of galaxies in this group compared to those in the star-forming ISM dust-rich group. The abundance of metals in the ISM provides material for dust growth via grain accretion, allowing such galaxies to accumulate dust mass. Another scenario for their dust production is the condensation in the ejecta from previous star formation episodes. Although they do not have a fresh supply of stellar ejecta due to the low sSFR, it is possible that there are still refractory elements from previous SN and AGB winds in the ISM that condense into dust grains.
Figure~\ref{fig:btt-sfr} shows that the majority of quenched ISM dust-rich galaxies can be found in the low mass population. Compared to the star-forming ISM dust-rich group, this group is rare at low and high redshift. Figure~\ref{fig:btt-sfr-evol} shows that the fraction of this group with stellar mass $\log \mathrm{M_*} = 10-11$ at $z > 0$ is always below $6\%$. Due to their low abundance, we have not found any observational evidence for this group.
\subsection{The halo dust-rich group}
Halo dust-rich galaxies have little or no dust in their ISM but contain a significant amount of dust in their halo. At redshift $z=0$, this group includes elliptical galaxies with the largest BTT mass ratios and low sSFR. The median gas-phase metallicity of this group is also the lowest compared to the other groups. The small sSFR does not allow for efficient instantaneous metal and stellar dust production in such galaxies. However, they should have accumulated metals and dust from previous star-formation episodes. The lack of metals and dust in their ISM, while their halo is dust-rich, indicates that the ISM content is transported into the halo (see Section~\ref{sec:relation}).
At higher redshift, however, the halo dust-rich group occupies an increasingly high sSFR region with a median $\log \mathrm{sSFR} >$ the Milky Way value. The high sSFR of these galaxies means that they are producing metals and dust in their stellar ejecta. However, unlike galaxies in the other groups, galaxies in this group show no increase in gas metallicity with decreasing redshift. The low abundance of metals and dust in their ISM is possible if the feedback is efficiently heating and blowing out the newly formed metals and dust. Such feedback might be provided by AGN activity \citep{Sarangi19}.
Galaxies in a deep potential well can retain the feedback-heated gas, dust and metals in their halo. Mass in the dark matter halo usually sets the depth of the potential well. Since stellar mass correlates with halo mass, halo dust-rich galaxies should be more common in the massive galaxies. Figure \ref{fig:btt-sfr} shows that in the massive bin, this group comprise a quarter of the population, a more significant fraction than the lower stellar mass bin.
\citet{Menard10} measured dust reddening effects of background quasars relative to the foreground SDSS galaxies. They observed the existence of diffuse dust in halos with an amount comparable to those in the galactic disk. For $0.5L^*$ galaxies, the authors estimated a dust mass of $5 \times 10^7 \ensuremath{\mathrm{M}_{\odot}}\xspace$ in the halo. \citet{PMC15}
confirmed this halo dust mass for $0.1L^* - 1L^*$ low redshift galaxies. Observations of CGM dust provides evidence for our prediction that dust is transported out of the ISM and trapped in the halo. However, such observations are still uncommon. Future CGM dust measurements will allow for better constraints for models such as \dustysage.
\subsection{The dust-poor group}
We define galaxies that lack dust in both the ISM and their halo as dust-poor galaxies. At redshift $z=0$ (Figure~\ref{fig:btt-sfr}), the dust-poor group has low median sSFR, low gas-phase metallicity and high BTT mass ratio. In the DustPedia~\citep{Nersesian19} and KINGFISH~\citep{Kennicutt11, RR14} datasets, galaxies with morphological type E-S0 have a median dust-to-stellar mass ratio below $1 \times 10^{-4}$ and are classified as ISM dust-poor as per our definition. However, both datasets do not provide measurement for dust in the halo. Therefore, we assume their halo is dust-poor, and we include them in the dust-poor group instead of the halo dust-rich group.
Dust depletion in the ISM occurs in two ways: destruction by SN shocks inside the galaxy disk and dust that is ejected out of the ISM by feedback. If the potential well of the system is enough to trap the dust, the ejected dust will end up in halo. However, galaxies with low stellar mass usually live in shallower potentials and thus are likely to lose their dust completely through ejection. This likely explains why Figure~\ref{fig:btt-sfr} shows that the dust-poor group contain the most significant fraction of their galaxies in the low mass bin.
In Figure \ref{fig:btt-sfr-evol}, the dust-poor group moves to the lower BTT mass ratio, lower sSFR region with decreasing redshift while the gas metallicity increases. At high redshift, galaxies in this group have relatively high sSFR, but their ISM still lacks both metals and dust. Dust formation is highly sensitive to the abundance of metals. In stellar ejecta, dust forms via the condensation of metals ejected from stars. In dense clouds, dust grains grow by accreting metals. Therefore, galaxies with low metallicity can not produce dust efficiently. If this low production rate can not keep up with the depletion rate, a galaxy will end up in the dust-poor group.
\citet{Fisher14} observed the dust emission from a local dwarf galaxy, I Zwicky 18, and found it lacks dust with a dust-to-stellar mass ratio of about $10^{-5}$ to $10^{-6}$. This galaxy has a SFR of $0.05$ \ensuremath{\mathrm{M}_{\odot}}\xspace/year and a very low metallicity of $12+\log\mathrm{[O/H]} = 7.17$, much lower than that seen in local galaxies \citep{Fisher14}. The reported stellar mass is $9 \times 10^7 \ensuremath{\mathrm{M}_{\odot}}\xspace$, which makes the sSFR $ = 5.5 \times 10^{-10}$, consistent with the median sSFR value of the dust-poor galaxies in the lower mass group from \dustysage (see Figure \ref{fig:btt-sfr}). Although the stellar mass of I Zwicky 18 is well below our sample, we find a consistent feature where a lack of metals accompanies the lack of dust.
\section{Conclusion}
\label{sec:conclusion}
During galaxy's evolution, dust physics is heavily influenced by star formation, gas physics, outflows and the depth of the potential well. Therefore, the dust, stellar, and gas content of galaxies are tightly related to each other. In this paper, we use the \dustysage galaxy formation model \mbox{\citep{Triani20}} to explore the relation between dust content and the fundamental properties of galaxies. We have divided the galaxy population based on their dust mass and observable properties. At redshift $z=0$, the properties of our model galaxies are in good agreement with observations, especially where the statistics are sound.
Our investigation includes the comparison of morphology, specific star formation rate, metallicity and stellar mass for galaxies grouped by their dust content from redshift $z=3$ to $z=0$. Our main conclusions are the following:
\begin{itemize}
\item We find that the distribution of dust in the ISM, hot halo and ejected reservoir is affected by galaxies' morphology and stellar mass. The fraction of dust in the hot halo increases with BTT mass ratio and stellar mass (Figure \ref{fig:btt-fraction}).
\item At redshift $z=0$, ISM dust-rich galaxies have the highest gas-phase metallicities. Both star-forming and quenched ISM dust-rich groups shows a low BTT mass ratio and relatively high median sSFR. Galaxies in the halo dust-rich and dust-poor groups are metal poor. They are distributed in the high BTT mass ratio and low sSFR regime of Figure~\ref{fig:btt-sfr}.
\item We see an evolution in the behaviour of our grouped galaxies in Figure~\ref{fig:btt-sfr-evol}. From $z=3$ to $z=0$: (i) the BTT mass ratio of the quenched ISM dust-rich group drops $\approx 0.3$ dex while the other three groups show insignificant changes, (ii) the median sSFR of all groups except quenched ISM dust-rich trend towards lower region, and (iii) all groups show weak evolution of gas-phase metallicity with redshift except the quenched ISM dust-rich group where the metallicity fluctuates with redshift.
\item Across redshift, both star-forming ISM dust-rich and quenched ISM dust-rich groups maintain their high median sSFR and high gas-phase metallicity. This implies that such galaxies are still actively producing metals and dust via stellar production. The abundance of metals in the ISM also enables dust growth via metal accretion in dense clouds.
\item The halo dust-rich group consists of galaxies with a dust-poor ISM but dust-rich halo. They have a relatively high outflow rate of heated gas, dust, and metals out of the ISM. At present, they are elliptical with very low sSFR and metallicity. The gas-phase metallicity of these galaxies does not depend significantly on redshift. At high redshift, galaxies in this group have high sSFR that enables an efficient dust production mechanism. Their lack of ISM dust, therefore, implies effective feedback to reheat the newly formed dust. Because these massive galaxies have a deep potential well, they are more likely to retain the heated dust in their halo. Figure~\ref{fig:btt-sfr} and Figure \ref{fig:btt-sfr-evol} shows that this group is made up of a larger fraction of massive galaxies.
\item Galaxies in the dust-poor group have low dust mass both in their ISM and halo. From redshift $z=3$ to $z=0$, this group evolves from the high sSFR region to the lower sSFR region. However, they always have low metallicity. As metals are the main ingredients for dust, this low metallicity implies that their dust production is also small.
\end{itemize}
Our model provides predictions for future surveys with next-generation instruments and telescopes, such as ALMA and JWST, that will measure the dust and galaxy properties in extraordinary detail at high redshift. JWST will cover the IR spectrum in the 0.6 to 28.5 micron regime, which reveals the total galactic dust content up to $z=2$. At $z=3$, it will only cover the dust spectrum to $\approx 7$ micron, providing a measure for the mass of hot dust. As we have seen throughout this paper, current dust measurement at high-redshift are only available for the brightest galaxies with high dust mass. Complemented with data from ALMA and other IR and sub-mm surveys, JWST will provide a significantly improved view of high redshift dust across a wider range of masses. This will better constraint dust enrichment scenarios in models such as \dustysage, leading to a better predictions and more detailed interpretation of the observations.
\section*{Acknowledgements}
We would like to thank the anonymous referee whose valuable comments improved the quality of this paper, and Ned Taylor for helpful comments during the final stages of this work. This research was supported by the Australian Research Council Centre of Excellence for All Sky Astro-physics in 3 Dimensions (ASTRO 3D), through project number CE170100013. The Semi-Analytic Galaxy Evolution (SAGE) model, on which \dustysage was built, is a publicly available codebase that runs on the dark matter halo trees of a cosmological N-body simulation. It is available for download at \url{https://github.com/darrencroton/sage}. This research has used \texttt{python} (\url{https://www.python.org/}), \texttt{numpy} \citep{Vanderwalt11} and \texttt{matplotlib} \citep{Hunter07}.
\section*{Data availability}
The data underlying this article are available in the article. The galaxy formation model used to generate the data is available at \url{https://github.com/dptriani/dusty-sage}.
\bibliographystyle{mnras}
\section{Introduction}
The journal \textit{Monthly Notices of the Royal Astronomical Society} (MNRAS) encourages authors to prepare their papers using \LaTeX.
The style file \verb'mnras.cls' can be used to approximate the final appearance of the journal, and provides numerous features to simplify the preparation of papers.
This document, \verb'mnras_guide.tex', provides guidance on using that style file and the features it enables.
This is not a general guide on how to use \LaTeX, of which many excellent examples already exist.
We particularly recommend \textit{Wikibooks \LaTeX}\footnote{\url{https://en.wikibooks.org/wiki/LaTeX}}, a collaborative online textbook which is of use to both beginners and experts.
Alternatively there are several other online resources, and most academic libraries also hold suitable beginner's guides.
For guidance on the contents of papers, journal style, and how to submit a paper, see the MNRAS Instructions to Authors\footnote{\label{foot:itas}\url{http://www.oxfordjournals.org/our_journals/mnras/for_authors/}}.
Only technical issues with the \LaTeX\ class are considered here.
\section{Obtaining and installing the MNRAS package}
Some \LaTeX\ distributions come with the MNRAS package by default.
If yours does not, you can either install it using your distribution's package manager, or download it from the Comprehensive \TeX\ Archive Network\footnote{\url{http://www.ctan.org/tex-archive/macros/latex/contrib/mnras}} (CTAN).
The files can either be installed permanently by placing them in the appropriate directory (consult the documentation for your \LaTeX\ distribution), or used temporarily by placing them in the working directory for your paper.
To use the MNRAS package, simply specify \verb'mnras' as the document class at the start of a \verb'.tex' file:
\begin{verbatim}
\documentclass{mnras}
\end{verbatim}
Then compile \LaTeX\ (and if necessary \bibtex) in the usual way.
\section{Preparing and submitting a paper}
We recommend that you start with a copy of the \texttt{mnras\_template.tex} file.
Rename the file, update the information on the title page, and then work on the text of your paper.
Guidelines for content, style etc. are given in the instructions to authors on the journal's website$^{\ref{foot:itas}}$.
Note that this document does not follow all the aspects of MNRAS journal style (e.g. it has a table of contents).
If a paper is accepted, it is professionally typeset and copyedited by the publishers.
It is therefore likely that minor changes to presentation will occur.
For this reason, we ask authors to ignore minor details such as slightly long lines, extra blank spaces, or misplaced figures, because these details will be dealt with during the production process.
Papers must be submitted electronically via the online submission system; paper submissions are not permitted.
For full guidance on how to submit a paper, see the instructions to authors.
\section{Class options}
\label{sec:options}
There are several options which can be added to the document class line like this:
\begin{verbatim}
\documentclass[option1,option2]{mnras}
\end{verbatim}
The available options are:
\begin{itemize}
\item \verb'letters' -- used for papers in the journal's Letters section.
\item \verb'onecolumn' -- single column, instead of the default two columns. This should be used {\it only} if necessary for the display of numerous very long equations.
\item \verb'doublespacing' -- text has double line spacing. Please don't submit papers in this format.
\item \verb'referee' -- \textit{(deprecated)} single column, double spaced, larger text, bigger margins. Please don't submit papers in this format.
\item \verb'galley' -- \textit{(deprecated)} no running headers, no attempt to align the bottom of columns.
\item \verb'landscape' -- \textit{(deprecated)} sets the whole document on landscape paper.
\item \verb"usenatbib" -- \textit{(all papers should use this)} this uses Patrick Daly's \verb"natbib.sty" package for citations.
\item \verb"usegraphicx" -- \textit{(most papers will need this)} includes the \verb'graphicx' package, for inclusion of figures and images.
\item \verb'useAMS' -- adds support for upright Greek characters \verb'\upi', \verb'\umu' and \verb'\upartial' ($\upi$, $\umu$ and $\upartial$). Only these three are included, if you require other symbols you will need to include the \verb'amsmath' or \verb'amsymb' packages (see section~\ref{sec:packages}).
\item \verb"usedcolumn" -- includes the package \verb"dcolumn", which includes two new types of column alignment for use in tables.
\end{itemize}
Some of these options are deprecated and retained for backwards compatibility only.
Others are used in almost all papers, but again are retained as options to ensure that papers written decades ago will continue to compile without problems.
If you want to include any other packages, see section~\ref{sec:packages}.
\section{Title page}
If you are using \texttt{mnras\_template.tex} the necessary code for generating the title page, headers and footers is already present.
Simply edit the title, author list, institutions, abstract and keywords as described below.
\subsection{Title}
There are two forms of the title: the full version used on the first page, and a short version which is used in the header of other odd-numbered pages (the `running head').
Enter them with \verb'\title[]{}' like this:
\begin{verbatim}
\title[Running head]{Full title of the paper}
\end{verbatim}
The full title can be multiple lines (use \verb'\\' to start a new line) and may be as long as necessary, although we encourage authors to use concise titles. The running head must be $\le~45$ characters on a single line.
See appendix~\ref{sec:advanced} for more complicated examples.
\subsection{Authors and institutions}
Like the title, there are two forms of author list: the full version which appears on the title page, and a short form which appears in the header of the even-numbered pages. Enter them using the \verb'\author[]{}' command.
If the author list is more than one line long, start a new line using \verb'\newauthor'. Use \verb'\\' to start the institution list. Affiliations for each author should be indicated with a superscript number, and correspond to the list of institutions below the author list.
For example, if I were to write a paper with two coauthors at another institution, one of whom also works at a third location:
\begin{verbatim}
\author[K. T. Smith et al.]{
Keith T. Smith,$^{1}$
A. N. Other,$^{2}$
and Third Author$^{2,3}$
\\
$^{1}$Affiliation 1\\
$^{2}$Affiliation 2\\
$^{3}$Affiliation 3}
\end{verbatim}
Affiliations should be in the format `Department, Institution, Street Address, City and Postal Code, Country'.
Email addresses can be inserted with the \verb'\thanks{}' command which adds a title page footnote.
If you want to list more than one email, put them all in the same \verb'\thanks' and use \verb'\footnotemark[]' to refer to the same footnote multiple times.
Present addresses (if different to those where the work was performed) can also be added with a \verb'\thanks' command.
\subsection{Abstract and keywords}
The abstract is entered in an \verb'abstract' environment:
\begin{verbatim}
\begin{abstract}
The abstract of the paper.
\end{abstract}
\end{verbatim}
\noindent Note that there is a word limit on the length of abstracts.
For the current word limit, see the journal instructions to authors$^{\ref{foot:itas}}$.
Immediately following the abstract, a set of keywords is entered in a \verb'keywords' environment:
\begin{verbatim}
\begin{keywords}
keyword 1 -- keyword 2 -- keyword 3
\end{keywords}
\end{verbatim}
\noindent There is a list of permitted keywords, which is agreed between all the major astronomy journals and revised every few years.
Do \emph{not} make up new keywords!
For the current list of allowed keywords, see the journal's instructions to authors$^{\ref{foot:itas}}$.
\section{Sections and lists}
Sections and lists are generally the same as in the standard \LaTeX\ classes.
\subsection{Sections}
\label{sec:sections}
Sections are entered in the usual way, using \verb'\section{}' and its variants. It is possible to nest up to four section levels:
\begin{verbatim}
\section{Main section}
\subsection{Subsection}
\subsubsection{Subsubsection}
\paragraph{Lowest level section}
\end{verbatim}
\noindent The other \LaTeX\ sectioning commands \verb'\part', \verb'\chapter' and \verb'\subparagraph{}' are deprecated and should not be used.
Some sections are not numbered as part of journal style (e.g. the Acknowledgements).
To insert an unnumbered section use the `starred' version of the command: \verb'\section*{}'.
See appendix~\ref{sec:advanced} for more complicated examples.
\subsection{Lists}
Two forms of lists can be used in MNRAS -- numbered and unnumbered.
For a numbered list, use the \verb'enumerate' environment:
\begin{verbatim}
\begin{enumerate}
\item First item
\item Second item
\item etc.
\end{enumerate}
\end{verbatim}
\noindent which produces
\begin{enumerate}
\item First item
\item Second item
\item etc.
\end{enumerate}
Note that the list uses lowercase Roman numerals, rather than the \LaTeX\ default Arabic numerals.
For an unnumbered list, use the \verb'description' environment without the optional argument:
\begin{verbatim}
\begin{description}
\item First item
\item Second item
\item etc.
\end{description}
\end{verbatim}
\noindent which produces
\begin{description}
\item First item
\item Second item
\item etc.
\end{description}
Bulleted lists using the \verb'itemize' environment should not be used in MNRAS; it is retained for backwards compatibility only.
\section{Mathematics and symbols}
The MNRAS class mostly adopts standard \LaTeX\ handling of mathematics, which is briefly summarised here.
See also section~\ref{sec:packages} for packages that support more advanced mathematics.
Mathematics can be inserted into the running text using the syntax \verb'$1+1=2$', which produces $1+1=2$.
Use this only for short expressions or when referring to mathematical quantities; equations should be entered as described below.
\subsection{Equations}
Equations should be entered using the \verb'equation' environment, which automatically numbers them:
\begin{verbatim}
\begin{equation}
a^2=b^2+c^2
\end{equation}
\end{verbatim}
\noindent which produces
\begin{equation}
a^2=b^2+c^2
\end{equation}
By default, the equations are numbered sequentially throughout the whole paper. If a paper has a large number of equations, it may be better to number them by section (2.1, 2.2 etc.). To do this, add the command \verb'\numberwithin{equation}{section}' to the preamble.
It is also possible to produce un-numbered equations by using the \LaTeX\ built-in \verb'\['\textellipsis\verb'\]' and \verb'$$'\textellipsis\verb'$$' commands; however MNRAS requires that all equations are numbered, so these commands should be avoided.
\subsection{Special symbols}
\begin{table}
\caption{Additional commands for special symbols commonly used in astronomy. These can be used anywhere.}
\label{tab:anysymbols}
\begin{tabular}{lll}
\hline
Command & Output & Meaning\\
\hline
\verb'\sun' & \sun & Sun, solar\\[2pt]
\verb'\earth' & \earth & Earth, terrestrial\\[2pt]
\verb'\micron' & \micron & microns\\[2pt]
\verb'\degr' & \degr & degrees\\[2pt]
\verb'\arcmin' & \arcmin & arcminutes\\[2pt]
\verb'\arcsec' & \arcsec & arcseconds\\[2pt]
\verb'\fdg' & \fdg & fraction of a degree\\[2pt]
\verb'\farcm' & \farcm & fraction of an arcminute\\[2pt]
\verb'\farcs' & \farcs & fraction of an arcsecond\\[2pt]
\verb'\fd' & \fd & fraction of a day\\[2pt]
\verb'\fh' & \fh & fraction of an hour\\[2pt]
\verb'\fm' & \fm & fraction of a minute\\[2pt]
\verb'\fs' & \fs & fraction of a second\\[2pt]
\verb'\fp' & \fp & fraction of a period\\[2pt]
\verb'\diameter' & \diameter & diameter\\[2pt]
\verb'\sq' & \sq & square, Q.E.D.\\[2pt]
\hline
\end{tabular}
\end{table}
\begin{table}
\caption{Additional commands for mathematical symbols. These can only be used in maths mode.}
\label{tab:mathssymbols}
\begin{tabular}{lll}
\hline
Command & Output & Meaning\\
\hline
\verb'\upi' & $\upi$ & upright pi\\[2pt]
\verb'\umu' & $\umu$ & upright mu\\[2pt]
\verb'\upartial' & $\upartial$ & upright partial derivative\\[2pt]
\verb'\lid' & $\lid$ & less than or equal to\\[2pt]
\verb'\gid' & $\gid$ & greater than or equal to\\[2pt]
\verb'\la' & $\la$ & less than of order\\[2pt]
\verb'\ga' & $\ga$ & greater than of order\\[2pt]
\verb'\loa' & $\loa$ & less than approximately\\[2pt]
\verb'\goa' & $\goa$ & greater than approximately\\[2pt]
\verb'\cor' & $\cor$ & corresponds to\\[2pt]
\verb'\sol' & $\sol$ & similar to or less than\\[2pt]
\verb'\sog' & $\sog$ & similar to or greater than\\[2pt]
\verb'\lse' & $\lse$ & less than or homotopic to \\[2pt]
\verb'\gse' & $\gse$ & greater than or homotopic to\\[2pt]
\verb'\getsto' & $\getsto$ & from over to\\[2pt]
\verb'\grole' & $\grole$ & greater over less\\[2pt]
\verb'\leogr' & $\leogr$ & less over greater\\
\hline
\end{tabular}
\end{table}
Some additional symbols of common use in astronomy have been added in the MNRAS class. These are shown in tables~\ref{tab:anysymbols}--\ref{tab:mathssymbols}. The command names are -- as far as possible -- the same as those used in other major astronomy journals.
Many other mathematical symbols are also available, either built into \LaTeX\ or via additional packages. If you want to insert a specific symbol but don't know the \LaTeX\ command, we recommend using the Detexify website\footnote{\url{http://detexify.kirelabs.org}}.
Sometimes font or coding limitations mean a symbol may not get smaller when used in sub- or superscripts, and will therefore be displayed at the wrong size. There is no need to worry about this as it will be corrected by the typesetter during production.
To produce bold symbols in mathematics, use \verb'\bmath' for simple variables, and the \verb'bm' package for more complex symbols (see section~\ref{sec:packages}). Vectors are set in bold italic, using \verb'\mathbfit{}'.
For matrices, use \verb'\mathbfss{}' to produce a bold sans-serif font e.g. \mathbfss{H}; this works even outside maths mode, but not all symbols are available (e.g. Greek). For $\nabla$ (del, used in gradients, divergence etc.) use \verb'$\nabla$'.
\subsection{Ions}
A new \verb'\ion{}{}' command has been added to the class file, for the correct typesetting of ionisation states.
For example, to typeset singly ionised calcium use \verb'\ion{Ca}{ii}', which produces \ion{Ca}{ii}.
\section{Figures and tables}
\label{sec:fig_table}
Figures and tables (collectively called `floats') are mostly the same as built into \LaTeX.
\subsection{Basic examples}
\begin{figure}
\includegraphics[width=\columnwidth]{example}
\caption{An example figure.}
\label{fig:example}
\end{figure}
Figures are inserted in the usual way using a \verb'figure' environment and \verb'\includegraphics'. The example Figure~\ref{fig:example} was generated using the code:
\begin{verbatim}
\begin{figure}
\includegraphics[width=\columnwidth]{example}
\caption{An example figure.}
\label{fig:example}
\end{figure}
\end{verbatim}
\begin{table}
\caption{An example table.}
\label{tab:example}
\begin{tabular}{lcc}
\hline
Star & Mass & Luminosity\\
& $M_{\sun}$ & $L_{\sun}$\\
\hline
Sun & 1.00 & 1.00\\
$\alpha$~Cen~A & 1.10 & 1.52\\
$\epsilon$~Eri & 0.82 & 0.34\\
\hline
\end{tabular}
\end{table}
The example Table~\ref{tab:example} was generated using the code:
\begin{verbatim}
\begin{table}
\caption{An example table.}
\label{tab:example}
\begin{tabular}{lcc}
\hline
Star & Mass & Luminosity\\
& $M_{\sun}$ & $L_{\sun}$\\
\hline
Sun & 1.00 & 1.00\\
$\alpha$~Cen~A & 1.10 & 1.52\\
$\epsilon$~Eri & 0.82 & 0.34\\
\hline
\end{tabular}
\end{table}
\end{verbatim}
\subsection{Captions and placement}
Captions go \emph{above} tables but \emph{below} figures, as in the examples above.
The \LaTeX\ float placement commands \verb'[htbp]' are intentionally disabled.
Layout of figures and tables will be adjusted by the publisher during the production process, so authors should not concern themselves with placement to avoid disappointment and wasted effort.
Simply place the \LaTeX\ code close to where the figure or table is first mentioned in the text and leave exact placement to the publishers.
By default a figure or table will occupy one column of the page.
To produce a wider version which covers both columns, use the \verb'figure*' or \verb'table*' environment.
If a figure or table is too long to fit on a single page it can be split it into several parts.
Create an additional figure or table which uses \verb'\contcaption{}' instead of \verb'\caption{}'.
This will automatically correct the numbering and add `\emph{continued}' at the start of the caption.
\begin{table}
\contcaption{A table continued from the previous one.}
\label{tab:continued}
\begin{tabular}{lcc}
\hline
Star & Mass & Luminosity\\
& $M_{\sun}$ & $L_{\sun}$\\
\hline
$\tau$~Cet & 0.78 & 0.52\\
$\delta$~Pav & 0.99 & 1.22\\
$\sigma$~Dra & 0.87 & 0.43\\
\hline
\end{tabular}
\end{table}
Table~\ref{tab:continued} was generated using the code:
\begin{verbatim}
\begin{table}
\contcaption{A table continued from the previous one.}
\label{tab:continued}
\begin{tabular}{lcc}
\hline
Star & Mass & Luminosity\\
& $M_{\sun}$ & $L_{\sun}$\\
\hline
$\tau$~Cet & 0.78 & 0.52\\
$\delta$~Pav & 0.99 & 1.22\\
$\sigma$~Dra & 0.87 & 0.43\\
\hline
\end{tabular}
\end{table}
\end{verbatim}
To produce a landscape figure or table, use the \verb'pdflscape' package and the \verb'landscape' environment.
The landscape Table~\ref{tab:landscape} was produced using the code:
\begin{verbatim}
\begin{landscape}
\begin{table}
\caption{An example landscape table.}
\label{tab:landscape}
\begin{tabular}{cccccccccc}
\hline
Header & Header & ...\\
Unit & Unit & ...\\
\hline
Data & Data & ...\\
Data & Data & ...\\
...\\
\hline
\end{tabular}
\end{table}
\end{landscape}
\end{verbatim}
Unfortunately this method will force a page break before the table appears.
More complicated solutions are possible, but authors shouldn't worry about this.
\begin{landscape}
\begin{table}
\caption{An example landscape table.}
\label{tab:landscape}
\begin{tabular}{cccccccccc}
\hline
Header & Header & Header & Header & Header & Header & Header & Header & Header & Header\\
Unit & Unit & Unit & Unit & Unit & Unit & Unit & Unit & Unit & Unit \\
\hline
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
Data & Data & Data & Data & Data & Data & Data & Data & Data & Data\\
\hline
\end{tabular}
\end{table}
\end{landscape}
\section{References and citations}
\subsection{Cross-referencing}
The usual \LaTeX\ commands \verb'\label{}' and \verb'\ref{}' can be used for cross-referencing within the same paper.
We recommend that you use these whenever relevant, rather than writing out the section or figure numbers explicitly.
This ensures that cross-references are updated whenever the numbering changes (e.g. during revision) and provides clickable links (if available in your compiler).
It is best to give each section, figure and table a logical label.
For example, Table~\ref{tab:mathssymbols} has the label \verb'tab:mathssymbols', whilst section~\ref{sec:packages} has the label \verb'sec:packages'.
Add the label \emph{after} the section or caption command, as in the examples in sections~\ref{sec:sections} and \ref{sec:fig_table}.
Enter the cross-reference with a non-breaking space between the type of object and the number, like this: \verb'see Figure~\ref{fig:example}'.
The \verb'\autoref{}' command can be used to automatically fill out the type of object, saving on typing.
It also causes the link to cover the whole phrase rather than just the number, but for that reason is only suitable for single cross-references rather than ranges.
For example, \verb'\autoref{tab:journal_abbr}' produces \autoref{tab:journal_abbr}.
\subsection{Citations}
\label{sec:cite}
MNRAS uses the Harvard -- author (year) -- citation style, e.g. \citet{author2013}.
This is implemented in \LaTeX\ via the \verb'natbib' package, which in turn is included via the \verb'usenatbib' package option (see section~\ref{sec:options}), which should be used in all papers.
Each entry in the reference list has a `key' (see section~\ref{sec:ref_list}) which is used to generate citations.
There are two basic \verb'natbib' commands:
\begin{description}
\item \verb'\citet{key}' produces an in-text citation: \citet{author2013}
\item \verb'\citep{key}' produces a bracketed (parenthetical) citation: \citep{author2013}
\end{description}
Citations will include clickable links to the relevant entry in the reference list, if supported by your \LaTeX\ compiler.
\defcitealias{smith2014}{Paper~I}
\begin{table*}
\caption{Common citation commands, provided by the \texttt{natbib} package.}
\label{tab:natbib}
\begin{tabular}{lll}
\hline
Command & Ouput & Note\\
\hline
\verb'\citet{key}' & \citet{smith2014} & \\
\verb'\citep{key}' & \citep{smith2014} & \\
\verb'\citep{key,key2}' & \citep{smith2014,jones2015} & Multiple papers\\
\verb'\citet[table 4]{key}' & \citet[table 4]{smith2014} & \\
\verb'\citep[see][figure 7]{key}' & \citep[see][figure 7]{smith2014} & \\
\verb'\citealt{key}' & \citealt{smith2014} & For use with manual brackets\\
\verb'\citeauthor{key}' & \citeauthor{smith2014} & If already cited in close proximity\\
\verb'\defcitealias{key}{Paper~I}' & & Define an alias (doesn't work in floats)\\
\verb'\citetalias{key}' & \citetalias{smith2014} & \\
\verb'\citepalias{key}' & \citepalias{smith2014} & \\
\hline
\end{tabular}
\end{table*}
There are a number of other \verb'natbib' commands which can be used for more complicated citations.
The most commonly used ones are listed in Table~\ref{tab:natbib}.
For full guidance on their use, consult the \verb'natbib' documentation\footnote{\url{http://www.ctan.org/pkg/natbib}}.
If a reference has several authors, \verb'natbib' will automatically use `et al.' if there are more than two authors. However, if a paper has exactly three authors, MNRAS style is to list all three on the first citation and use `et al.' thereafter. If you are using \bibtex\ (see section~\ref{sec:ref_list}) then this is handled automatically. If not, the \verb'\citet*{}' and \verb'\citep*{}' commands can be used at the first citation to include all of the authors.
\subsection{The list of references}
\label{sec:ref_list}
It is possible to enter references manually using the usual \LaTeX\ commands, but we strongly encourage authors to use \bibtex\ instead.
\bibtex\ ensures that the reference list is updated automatically as references are added or removed from the paper, puts them in the correct format, saves on typing, and the same reference file can be used for many different papers -- saving time hunting down reference details.
An MNRAS \bibtex\ style file, \verb'mnras.bst', is distributed as part of this package.
The rest of this section will assume you are using \bibtex.
References are entered into a separate \verb'.bib' file in standard \bibtex\ formatting.
This can be done manually, or there are several software packages which make editing the \verb'.bib' file much easier.
We particularly recommend \textsc{JabRef}\footnote{\url{http://jabref.sourceforge.net/}}, which works on all major operating systems.
\bibtex\ entries can be obtained from the NASA Astrophysics Data System\footnote{\label{foot:ads}\url{http://adsabs.harvard.edu}} (ADS) by clicking on `Bibtex entry for this abstract' on any entry.
Simply copy this into your \verb'.bib' file or into the `BibTeX source' tab in \textsc{JabRef}.
Each entry in the \verb'.bib' file must specify a unique `key' to identify the paper, the format of which is up to the author.
Simply cite it in the usual way, as described in section~\ref{sec:cite}, using the specified key.
Compile the paper as usual, but add an extra step to run the \texttt{bibtex} command.
Consult the documentation for your compiler or latex distribution.
Correct formatting of the reference list will be handled by \bibtex\ in almost all cases, provided that the correct information was entered into the \verb'.bib' file.
Note that ADS entries are not always correct, particularly for older papers and conference proceedings, so may need to be edited.
If in doubt, or if you are producing the reference list manually, see the MNRAS instructions to authors$^{\ref{foot:itas}}$ for the current guidelines on how to format the list of references.
\section{Appendices and online material}
To start an appendix, simply place the \verb' |
\section*{Abstract}
}
\textit{Purpose}: We propose a general framework for quantifying predictive uncertainties of dose-related quantities and leveraging this information in a dose mimicking problem in the context of automated radiation therapy treatment planning.
\noindent \textit{Methods}: A three-step pipeline, comprising feature extraction, dose statistic prediction and dose mimicking, is employed. In particular, the features are produced by a convolutional variational autoencoder and used as inputs in a previously developed nonparametric Bayesian statistical method, estimating the multivariate predictive distribution of a collection of predefined dose statistics. Specially developed objective functions are then used to construct a probabilistic dose mimicking problem based on the produced distributions, creating deliverable treatment plans.
\noindent \textit{Results}: The numerical experiments are performed using a dataset of $94$ retrospective treatment plans of prostate cancer patients. We show that the features extracted by the variational autoencoder capture geometric information of substantial relevance to the dose statistic prediction problem and are related to dose statistics in a more regularized fashion than hand-crafted features. The estimated predictive distributions are reasonable and outperforms a non--input-dependent benchmark method, and the deliverable plans produced by the probabilistic dose mimicking agree better with their clinical counterparts than for a non-probabilistic formulation.
\noindent \textit{Conclusions}: We demonstrate that prediction of dose-related quantities may be extended to include uncertainty estimation and that such probabilistic information may be leveraged in a dose mimicking problem. The treatment plans produced by the proposed pipeline resemble their original counterparts well, illustrating the merits of a holistic approach to automated planning based on probabilistic modeling.
\newline
\begin{spacing}{0.9}
{\sffamily\small \noindent \textbf{Keywords:} Knowledge-based planning, uncertainty modeling, dose--volume histogram prediction, variational autoencoder, mixture-of-experts, dose mimicking.}
\end{spacing}
\end{quote}
\tolerance=1000
\section{Introduction}
In light of its rapid advancement of late, machine learning has been extensively applied to various areas within biomedical engineering, often with considerable success \citep{park, siddique}. One such area is that of automated radiation therapy treatment planning, where data-driven approaches may help in homogenizing the otherwise time-consuming process of creating clinically satisfactory treatment plans. Instead of the conventional approach of continually adjusting weights of objective functions and possibly reiterating multiple times with physicians, plan optimization based on machine learning often comprises predicting the achievable dose-related quantities---for example, spatial dose or dose--volume histograms (DVHs)---and solving the inverse problem of reconstructing them with respect to machine parameters. While much of previous literature focuses on pure prediction, the quantification of its associated uncertainty, which also contains valuable information for the subsequent optimization step, is often omitted. To address this, we present a general framework for, given any collection of dose-related quantities, estimating their multivariate predictive probability distribution and setting up a corresponding plan optimization problem taking this probabilistic information into account.
In general, automated treatment planning using machine learning, also known as knowledge-based planning, concerns the automatic plan generation using knowledge extracted from historically delivered clinical treatment plans \citep{ge, ng, wang, hussein}. One usually goes about this task by first assuming a parameterized optimization problem and then training a machine learning model to predict the unknown parameters, which is done in such a way that the solution to the resulting optimization problem will correspond to a clinically satisfactory plan. An example is the prediction of weights in a weighted-sum formulation \citep{boutilier}, which may even be sequentially adjusted by an agent trained using reinforcement learning methods \citep{shen}. More commonly, one predicts either a spatial dose distribution, a set of DVHs in some regions of interest, a collection of dose statistics or a combination thereof, and then minimizes the deviation between the quantities evaluated on the actual dose and their corresponding predicted values. This latter optimization is often referred to as dose mimicking. The prediction--mimicking division is prevalent in large parts of the current literature on knowledge-based planning, allowing for the parts to be separately investigated but also leading to a certain lack of causality between prediction accuracy and resulting plan quality.
The first part of predicting dose-related quantities has been abundantly addressed in previous literature. Regarding spatial dose prediction, two- or three-dimensional convolutional neural networks in various architectures, such as U-nets, have been widely used \citep{nguyen_unet, campbell, kearney, shiraishi, ma_isodose} and extended to generative models such as generative adversarial networks \citep{babier_gan, murakami}, although other methods such as random forests have also been studied \citep{mcintosh}. For prediction of DVHs or other dose statistics, while overlap volume histograms evaluated on the input image have been traditionally used for this purpose \citep{appenzoller, wall, jiao, skarpmanmunter, ma_features, zhu, wu, yuan}, more recent literature also includes the use of neural network--based methods to simultaneously predict spatial dose and DVHs directly from input images \citep{liu, nguyen_pareto}. Common for all of the aforementioned approaches, however, is that predictions are made deterministically with no associated predictive probability distribution. Various sources for inter-planner variations are investigated in \citet{nelms}, and as pointed out by \citet{babier_importance}, when designing and evaluating methods for automated planning, it is essential to maintain a holistic picture of both parts of a prediction--mimicking pipeline. Considering this, omitting the probabilistic information contained in predictive distributions when using purely predictive methods may be the cause of a substantial disconnect between the two parts. For example, the predictive uncertainty associated with the mean dose in some organ at risk may be orders of magnitude higher than that of the near-minimum dose in the planning target volume (PTV), which is crucial information for the dose mimicking problem; the same holds for individual voxel doses in spatial dose prediction. Without this information, finding the right weights in an optimization problem minimizing deviations between DVH statistics or voxelwise doses may be just as hard as tuning objective weights in conventional inverse planning.
While certain aspects of a general probabilistic approach have been previously addressed---for instance, in \citet{nilsson}, where a U-net--based mixture density network is trained to output voxelwise univariate Gaussian mixtures, or in \citet{nguyen_bagging}, where ensemble and Monte Carlo dropout techniques are used to provide uncertainties of spatial dose and dose statistics by sampling predictions---the estimation of concrete predictive distributions for DVHs or other dose statistics has yet to be given much attention in literature. In particular, given a collection of dose statistics, which may represent a discretized DVH, one would like to estimate their multivariate conditional probability distribution given the current patient geometry and the training dataset. For example, Eclipse RapidPlan (Varian Medical Systems, Palo Alto, CA, USA) provides DVH prediction uncertainties by forming confidence bands based on estimated standard deviations around the predictive mean for each point on the DVH \citep{fogliata}, although the actual underlying probability distribution is not available. As the raw inputs are typically contoured CT images (or, alternatively, only the contours), prediction of dose statistics or DVHs is a high-dimensional problem made especially hard by the scarcity of data often prevalent in similar settings. One way to remedy this problem is to employ dimensionality reduction using unsupervised learning methods to preprocess the input images before passing them to a machine learning model. Such dimensionality reduction methods have been previously explored in e.g. \citet{hanna}, where a variational autoencoder \citep{kingma} is used for the purpose, but also in other related areas, examples including deformable registration \citep{fu} and outcome prediction \citep{cui}.
In this paper, we propose a general probabilistic framework for quantifying predictive uncertainties of any collection of dose-related quantities and leveraging the information for creating a deliverable plan. This is done in a three-step pipeline comprising the extraction of features, the prediction of dose statistics and the solving of a dose mimicking problem, each of which enjoying the advantage of being substitutable by any equivalent method. In particular, we use a convolutional variational autoencoder for the dimensionality reduction of contoured CT images, producing features to be used as inputs in a recently developed nonparametric Bayesian method \citep{zhang_sbmoe} exploiting feature similarities and outputting predictive distributions as Gaussian mixtures. The estimation of such joint predictive distributions incorporating dependencies between dose statistics is novel in literature, as is the use of variational autoencoders for producing covariates to predict dose-related quantities. Furthermore, using specially developed objective functions designed to leverage the information contained in the estimated distributions, with gradients of dose statistics readily available using the method outlined in \citet{zhang_direct}, a probabilistic dose mimicking formulation is solved to produce complete treatment plans. The computational study shows that the proposed pipeline yields extracted features containing geometric information of substantial relevance for dose statistic prediction, reasonable estimations of multivariate predictive distributions over dose statistics and complete treatment plans agreeing well with their respective ground truths. More specifically, the variational autoencoder is shown to produce a less entangled feature space than that of hand-crafted features, the estimated predictive distributions outperform a non--input-dependent benchmark method, and the probabilistic dose mimicking formulation is shown to be more suited than a non-probabilistic analogue for handling planning tradeoffs. In conclusion, the experiments serve to demonstrate the feasibility of the framework for automated treatment planning.
\section{Materials and methods}
Let $\{(x^n, d^n)\}_{n} \subset \mathcal{X} \times \mathcal{D}$ be a dataset of clinical, historically delivered treatment plans consisting of pairs of contoured CT images $x^n$ and dose distributions $d^n$, which are represented as vectors $x^n = (x_i^n)_i$ and $d^n = (d_i^n)_i$, $i$ being the index over voxels. Here, $\mathcal{X}$ and $\mathcal{D}$ denote spaces of contoured images and dose distributions, respectively. For our purposes, we will only use the binary encodings of the regions of interest (ROIs) and not the radiodensities in the CT image. Let also $\{\psi_j\}_{j}$ be a collection of dose statistic functions $\psi_j : \mathcal{D} \to \mathbb{R}$---e.g., dose-at-volume $\operatorname{D}_v$ in different ROIs and at different volume levels $v$---and let $y^n$ be defined as the vector $y^n = (\psi_j(d^n))_{j}$ of evaluated dose statistic values on $d^n$ for each $n$. In the following, we will use the terms dose statistic and dose-related quantity interchangeably.
Given a new patient with input image $x^*$, the main task is to predict the corresponding dose statistic values $y^*$ and solve an accordingly constructed dose mimicking optimization problem. Our pipeline, shown in Figure \ref{flowchart}, is divided into the following three parts:
\begin{enumerate}
\item training a variational autoencoder to extract features $\phi(x^*) \in \mathcal{Z}$ for some relatively low-dimensional vector space $\mathcal{Z}$, where $\phi$ is the encoder part,
\item using $\{(\phi(x^n), y^n)\}_{n}$ as training set to train a similarity-based Bayesian mixture-of-experts model, developed in a previous paper \citep{zhang_sbmoe}, which outputs an estimate of the multivariate predictive density $p(y^* \mid x^*, \{(x^n, y^n)\}_n)$, and
\item solving a dose mimicking problem using specially developed objective functions incorporating probabilistic information in the predictive density $p(y^* \mid x^*, \{(x^n, y^n)\}_n)$, thus creating a deliverable plan.
\end{enumerate}
\begin{figure}[H]
\centering
\includegraphics[width=0.9\textwidth]{Flowchart.pdf}
\caption{The proposed automatic treatment planning pipeline.}
\label{flowchart}
\end{figure}
\subsection{Variational autoencoder}
\label{vae}
Using neural networks to extract features from high-dimensional images is a form of representation learning \citep{bengio}. A common method for this is using autoencoders \citep{goodfellow}, which is an unsupervised learning method using only the training inputs $x^n$. An autoencoder consists of an encoder $\phi : \mathcal{X} \to \mathcal{Z}$ and a decoder $\varphi : \mathcal{Z} \to \mathcal{X}$, and is trained in such a way that $\varphi(\phi(x^*))$ will resemble each new input $x^*$ as well as possible. If this is successful, one can deem the low-dimensional feature vector $\phi(x^*)$ to contain sufficient input to reconstruct $x^*$ reasonably well, thereby making $\phi$ qualify as a feature extractor.
A drawback with such plain autoencoders is the possible lack of regularity---that is, there is generally no guarantee that the decoded images $\varphi(z)$, $\varphi(z')$ will be close to each other whenever $z$, $z'$ are close. This may be problematic when using the produced features in another statistical model, particularly if the model is a nonparametric model relying on interpolations. While regularization methods for autoencoders have been extensively studied \citep{bengio}, another way of approaching this problem is to instead use a variational autoencoder, replacing the encoder and decoder by probabilistic counterparts. Modeling the latent representation $z \in \mathcal{Z}$ corresponding to each input $x \in \mathcal{X}$ as a random variable, in a variational autoencoder, one uses neural networks to approximate the conditional densities $p(z \mid x)$ and $p(x \mid z)$. The latter density, referred to as the likelihood, is often well-defined whereas the former, known as the posterior, is subsequently derived from Bayes' rule if one also defines a prior $p(z)$. However, the posterior $p(z \mid x)$ in this case will be intractable to compute---instead, it is approximated with some closed-form density $q(z)$, often called the variational posterior, in such a way that the Kullback--Leibler divergence $d_{\mathrm{KL}}(q \; \Vert \; p(\cdot \mid x))$ between $q(z)$ and $p(z \mid x)$ is minimized. Equivalently, one may find $q(z)$ by maximizing the evidence lower bound $\operatorname{ELBO}(q) = \operatorname{\mathbb{E}}^{q(z)} \log(p(x, z) / q(z))$. For a detailed review of variational inference in general, see \citet{blei}.
In our case, given a fixed set of ROIs $\{R_k\}_{k=1}^K$ as index sets over voxels (each voxel may belong to several ROIs), a binary encoding of each input image $x = (x_i)_i \in \mathcal{X}$ is used, where $x_i$ is the vector $x_i = (1_{i \; \in \; R_k})_k \in \mathbb{R}^{K}$---that is, entry $k$ of $x_i$ equals $1$ if voxel $i$ belongs to ROI $R_k$ and $0$ otherwise. For the model, we follow \citet{kingma} and define for each $n$ the likelihood and prior
\[
p(x^n \mid z^n) = \prod_i \prod_k \operatorname{Be}\!\left( x_{ik}^n \mid \varphi(z^n)_{ik} \right), \quad p(z^n) = \prod_j \operatorname{N}(z^n_j \mid 0, 1),
\]
where $z^n = (z^n_j)_j$, with the variational posterior being
\[
q(z^n) = \prod_j \operatorname{N}(z^n_j \mid \phi(x^n)_{j1}, \phi(x^n)_{j2}^2)
.\]
Here, $\operatorname{Be}$ and $\operatorname{N}$ denote the Bernoulli and the normal distribution, and $\phi$ and $\varphi$ are two neural networks with outputs of shape $\operatorname{dim} \mathcal{Z} \times 2$ and $\operatorname{dim} \mathcal{X} \times K$, respectively. The negative ELBO, which is the loss function to be minimized with respect to the network weights in $\phi$ and $\varphi$, is then obtained as
\begin{align*}
-\operatorname{ELBO}(q) = \sum_n \Bigg(& \frac{1}{2} \sum_j (1 + \log \phi(x^n)_{j2}^2 - \phi(x^n)_{j1}^2 - \phi(x^n)_{j2}^2) \\
\quad\quad &- \operatorname{\mathbb{E}}^{q(z^n)} \sum_i \sum_k \Big( x_{ik} \log \varphi(z^n)_{ik} + (1 - x_{ik}) \log(1 - \varphi(z^n)_{ik}) \Big) \Bigg)
\end{align*}
---here, the sum over $j$ regularizes the variational posterior by penalizing discrepancy from the prior, while the expected sum over $i$ and $k$ penalizes reconstruction error. In particular, the expectations over $q(z^n)$ and their gradients may be Monte Carlo--approximated using the reparameterization trick \citep{kingma}, which relies on rewriting $z^n$ as a location--scale transform $z^n = \phi(x^n)_{j1} + \epsilon \phi(x^n)_{j2}$ with $\epsilon \sim \operatorname{N}(0, 1)$ and then sampling $\epsilon$ at each objective evaluation.
\subsection{Dose statistic prediction}
Having trained the variational autoencoder, we may treat the encoder as fixed and use the set $\{(\phi(x^n), y^n)\}_n$, comprising pairs of feature vectors and evaluated dose statistics, as training dataset for a statistical model in which the multivariate predictive distribution $p(y^* \mid \phi(x^*), \{(\phi(x^n), y^n)\}_n)$ is obtained given a new input $x^*$. However, despite the embedding of the inputs as feature vectors, the dimensionalities involved are still high relative to typical dataset sizes. Due to the rather complex relationships between feature vectors and dose statistics, this will make the training of a parametric model of suitable size---e.g., a neural network outputting for each input a multivariate normal distribution predicting the corresponding output---hard to train to acceptable generalization accuracy without overfitting. Moreover, as distributions of dose statistics may be skewed or even multimodal \citep{nilsson}, such a model will often be inadequate even if it were possible to train using the available data.
Motivated by this, we settle on a so called similarity-based mixture-of-experts model \citep{zhang_sbmoe} for the purpose, which has predictive distributions in the form of multivariate Gaussian mixtures. This is a recently developed Bayesian nonparametric model in which predictions are made based on similarities between inputs rather than explicitly modeling input--output relationships. In particular, for each new input feature vector $\phi(x^*)$, mixture weights are calculated by first evaluating the similarity $\tau_n$ to each equivalent $\phi(x^n)$ in the training set, and then the probability $\sigma_{nc}$ of each $y^n$ belonging to each expert class $c$---the $\tau_n$ are referred to as first transitions and the $\sigma_{nc}$ as second transitions. The experts, in turn, are multivariate normal distributions $\{\operatorname{N}(\mu_c, \Sigma_c)\}_{c=1}^C$. Specifically, the distance metric $d_{\mathcal{Z}}$ used to determine similarities in the feature space $\mathcal{Z}$ is the Mahalanobis distance $d_{\mathcal{Z}}(z, z') = (z - z')^{\operatorname{T}} \Lambda (z - z')$, where $\Lambda$ is a precision matrix. This leads to a predictive likelihood on the form
\[
p(y^* \mid \phi(x^*), \{(\phi(x^n), y^n)\}_n, \theta) = \sum_c \sum_n \tau_n \sigma_{nc} \operatorname{N}(y^* \mid \mu_c, \Sigma_c)
,\]
where $\theta = (\Lambda, \{(\mu_c, \Sigma_c)\}_c)$ are the parameters and where the first and second transitions $\tau_n$, $\sigma_{nc}$ are evaluated as
\[
\tau_n = \frac{\operatorname{N}(\phi(x^*) \mid \phi(x^n), \Lambda^{-1})}{\sum_{n'} \operatorname{N}(\phi(x^*) \mid \phi(x^{n'}), \Lambda^{-1})}, \quad \sigma_{nc} = \frac{\operatorname{N}(y^n \mid \mu_c, \Sigma_c)}{\sum_{c'} \operatorname{N}(y^n \mid \mu_{c'}, \Sigma_{c'})}
.\]
The parameters $\theta$ are treated in a Bayesian fashion, fitted using updates according to a mean-field variational Bayes algorithm, and the resulting predictive distribution is obtained from Monte Carlo samples from the variational posterior. For further details, see \citet{zhang_sbmoe}.
\subsection{Dose mimicking}
\label{dosemimickingsec}
With a feature extraction method and dose statistic prediction model in place, consider a new patient with input image $x^*$ for which the predictive distribution $p(y^* \mid x^*, \{(x^n, y^n)\}_n)$ over a vector $y^* = (\psi_j(d^*))_j$ of dose statistic values has been estimated. In particular, let $F_j$ be the cumulative distribution function of the univariate marginal density $p(y^*_j \mid x^*, \{(x^n, y^n)\}_n)$ for each $j$. To utilize the probabilistic information captured in these distributions, we will base the objective function contribution from dose statistic $j$ for a given dose $d$ on the value of $F_j(\psi_j(d))$---loosely speaking, the contribution is based on the fraction of training patients the current patient is better than in terms of dose statistic $j$. While one could, in principle, use the joint cumulative distribution function over all dose statistics, the choice of using the marginal distributions separately is motivated by the advantage of increased control over individual dose statistics. We assume that each dose statistic is either ideally maximized or minimized, and assign binary labels $t_j = 1$ and $t_j = 0$ for the former and the latter case, respectively, to reflect this. Since the cumulative distribution function ranges from $0$ to $1$, we set up the loss on a cross-entropy form to obtain the contribution corresponding to $\psi_j$ as
\begin{equation}
\label{ce_objective}
\operatorname{CE}_j(d) = -t_j \log F_j(\psi_j(d)) - (1 - t_j) \log(1 - F_j(\psi_j(d))).
\end{equation}
Note that this means that the higher the certainty that the dose statistic is around some range of values, the more will the objective penalize deviations from those values. At the same time, since $F_j$ is continuous and strictly increasing in the case of a mixture-Gaussian predictive density, the optimizer will always have incentive to improve even when beyond the range of typical values. Figure \ref{tce} illustrates the behavior of the penalty for different probability distributions.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\textwidth]{a_lse.pdf}
\caption{Illustration of how the shapes of the predictive densities (dashed) affect their respective penalty contributions (solid). The blue and purple lines correspond to dose statistics to be minimized and the green lines to those to be maximized.}
\label{tce}
\end{figure}
As for the resulting optimization problem, let $\eta$ denote the optimization variables with feasible set $\mathcal{E}$, where the total dose $d$ is determined by some dose deposition mapping $d = d(\eta)$. For example, in the case of direct machine parameter optimization for volumetric modulated arc therapy (VMAT), the relation between the machine parameters $\eta$ and the dose $d$ may be modeled according to \citet{unkelbach}. The optimization problem may then be written as
\begin{equation}
\label{optformulation}
\underrel{\text{minimize}}{\eta \in \mathcal{E}} \quad \sum_j w_j \operatorname{CE}_j(d(\eta)),
\end{equation}
where $w_j$ is an importance weight for each dose statistic $\psi_j$. It is important to note that in contrast to conventional inverse planning or dose mimicking, the optimization problem above is relatively insensitive to the weights $w_j$ since a substantial part of the information regarding the relative importances of the dose statistics is already stored in the cumulative distribution functions $F_j$.
\subsection{Computational study}
To showcase the advantages and disadvantages of the proposed approach, we will use the following numerical experiments to demonstrate the full pipeline comprising feature extraction, dose statistic prediction and dose mimicking. The data used for this purpose consists of $94$ retrospective treatment plans of prostate cancer patients having undergone a prostatectomy prior to radiation therapy, originating from the Iridium Cancer Network in Antwerp, Belgium. The ROIs were contoured according to the RTOG \citep{gay} and ACROP \citep{salembier} guidelines, and the patients were treated with a prescribed dose of $7000 \; \mathrm{cGy}$ in the prostate bed and $5600 \; \mathrm{cGy}$ in the seminal vesicles and pelvic nodes, divided into $35$ fractions. All patients were treated using dual $360$-degree VMAT arcs, with final doses calculated using a collapsed cone algorithm. The DVHs of the training patients are shown in the backgrounds of Figure \ref{dvhband_1}. For the feature extraction and dose statistic prediction parts, the dataset was split into a training and a test set of $84$ and $10$ patients, respectively.
The variational autoencoder described in Section \ref{vae} was implemented in TensorFlow 2.3, using the architecture depicted in Figure \ref{vae_architecture}. Specifically, convolutional and transpose-convolutional layers were mostly used for the encoder and decoder parts, although complemented with dense layers, and layer normalization and dropout were applied after each rectified linear unit (ReLU) activation. Using the prostate PTV, seminal vesicles PTV, rectum, bladder, left femur, right femur and small bowel as the set of ROIs whose contours are inputted to the neural network---a total of seven ROIs---the input images were all preprocessed to binary arrays of size $64 \times 48 \times 83 \times 7$ with a $5 \; \mathrm{mm}$ voxel resolution, and the dimension of the latent space $\mathcal{Z}$ was set to $512$. In total, the network comprised around $6$ million parameters, which were fitted using a standard Adam optimizer \citep{kingma_adam} and a training and validation set of $75$ and $9$ patients, respectively. In particular, random data augmentations of small shifts and rotations were employed during training. The features produced by the variational autoencoder were qualitatively compared to hand-crafted features. In our case, we used distance transforms from the rectum and bladder regions to themselves and the two PTVs represented by normalized histograms with bin edges $(-3, -2, -1, 0, 1, 2, 3, 5, 7.5, 10) \; \mathrm{cm}$, amounting to $88$ dimensions in total.
\begin{figure}[h]
\hspace*{-1.8cm}
\includegraphics[width=1.3\textwidth]{vae.pdf}
\caption{Illustration of the architecture of the variational autoencoder described in Section \ref{vae}. The encoder consists of convolutional (yellow) and dense (purple) layers, and the decoder upsamples using transpose-convolutional (blue) layers. After each layer, a ReLU activation is applied if indicated with a darker color, followed by layer normalization and dropout. The green vector represents the stochastic source $\epsilon$ in the reparameterization trick described in Section \ref{vae}, and the last gray layer represents a softmax activation.}
\label{vae_architecture}
\end{figure}
The similarity-based mixture-of-experts model was subsequently trained using the implementation described in \citet{zhang_sbmoe}, in particular also using Tensorflow 2.3. For organs at risk, the set $\{\psi_j\}_j$ of dose statistics were dose-at-volume levels at volumes $10 \; \%, 20 \; \%, \dots, 90 \; \%$, all with $t_j = 0$; for targets, both dose-at-volume and mean-tail-dose \citep{romeijn} at volumes $2 \; \%, 5 \; \%, 10 \; \%, 90 \; \%, 95 \; \%, 98 \; \%$ were used, with $t_j = 1$ for volumes less than $50 \; \%$ and $t_j = 0$ otherwise. Here, mean-tail-dose was included to achieve better control of the tails of the target DVHs \citep{zhang_direct}. While one could in principle train one model to fit all dose statistics at once, in order to achieve better predictive accuracy given the small amount of available data, we trained one mixture-of-experts model per ROI at the cost of sacrificing inter-ROI dependencies. Each model comprised $C = 32$ experts with a posterior sample size of $10$ for each mean--covariance pair $(\mu_c, \Sigma_c)$, leading to a mixture model of $320$ classes.
The dose mimicking problem (\ref{optformulation}) was subsequently set up using the fitted dose statistic prediction model and the same beam configuration as used in the training dataset, with all optimizations and dose calculations performed in RayStation 10B (RaySearch Laboratories). In particular, the direct machine parameter optimization formulation for VMAT described in Section 2c of \citet{unkelbach} was used, leading to fully deliverable plans. To obtain increased control of the resulting plans, the clinical goals in Table \ref{tab:clinical_goals} were added to the set of dose statistics, with predictive distributions substituted by synthetically constructed normal distributions centered around the respective goal values. The importance weights $w_j$ were set to $0.1$ for the left and right femurs, $1$ for the rectum, bladder and small bowel regions, and $10$ for the PTVs. The objective functions, with gradients available due to the recent developments in \citet{zhang_direct}, were implemented in the RayStation source code. Each mimicking optimization comprised $3$ runs of $100$, $100$ and $40$ iterations with accurate dose computations in between---as per standard practice, approximate doses during optimization were calculated by a singular value decomposition algorithm and accurate doses by a collapsed cone algorithm. To illustrate the contribution of the uncertainty information captured in the cross-entropy objectives $\operatorname{CE}_j(d)$ from (\ref{ce_objective}), we also compared to a similar but non-probabilistic dose mimicking formulation. Specifically, each $\operatorname{CE}_j(d)$ in the proposed dose mimicking problem (\ref{optformulation}) was replaced by a quadratic penalty
\[
t_j (\psi_j(d) - \hat{\psi}_j)_{-}^2 + (1 - t_j) (\psi_j(d) - \hat{\psi}_j)_{+}^2
,\]
with $\hat{\psi_j}$ set to the mean of the distribution associated with $F_j$ and $(x)_-$, $(x)_+$ denoting the negative and positive parts of $x$, respectively. Otherwise, the quadratic penalty formulation used the same optimization settings and weights as the original dose mimicking procedure.
\begin{table}[h]
\caption{\label{tab:clinical_goals}Additional dose-at-volume ($\operatorname{D}_v$) and lower/upper mean-tail-dose ($\operatorname{MTD}_v^{\pm}$) clinical goals used in the dose mimicking problem.}
\centering
\begin{tabular}{ll}
\toprule
ROI & Goal \\
\midrule
PTV, prostate & $\operatorname{MTD}^-_{98 \, \%} \geq 6700 \; \mathrm{cGy}$ \\
PTV, prostate & $\operatorname{MTD}^+_{0.5 \, \%} \leq 7400 \; \mathrm{cGy}$ \\
PTV, seminal vesicles & $\operatorname{MTD}^-_{98 \, \%} \geq 5300 \; \mathrm{cGy}$ \\
PTV, seminal vesicles & $\operatorname{MTD}^+_{5 \, \%} \leq 5870 \; \mathrm{cGy}$ \\
Rectum & $\operatorname{D}_{50 \, \%} \leq 4500 \; \mathrm{cGy}$ \\
Bladder & $\operatorname{D}_{25 \, \%} \leq 6500 \; \mathrm{cGy}$ \\
Near-target & $\operatorname{MTD}^+_{5 \, \%} \leq 5500 \; \mathrm{cGy}$ \\
\bottomrule
\end{tabular}
\label{tableresults}
\end{table}
\begin{comment}
\begin{table}[h]
\centering
\caption{\label{tab:clinical_goals}Additional dose-at-volume ($\operatorname{D}_v$) and lower/upper mean-tail-dose ($\operatorname{MTD}_v^{\pm}$) clinical goals used in the dose mimicking problem.}
\begin{tabular}{l|l}
ROI & Goal \\\hline
PTV, prostate & $\operatorname{MTD}^-_{98 \, \%} \geq 6700 \; \mathrm{cGy}$ \\
PTV, prostate & $\operatorname{MTD}^+_{0.5 \, \%} \leq 7400 \; \mathrm{cGy}$ \\
PTV, seminal vesicles & $\operatorname{MTD}^-_{98 \, \%} \geq 5300 \; \mathrm{cGy}$ \\
PTV, seminal vesicles & $\operatorname{MTD}^+_{5 \, \%} \leq 5870 \; \mathrm{cGy}$ \\
Rectum & $\operatorname{D}_{50 \, \%} \leq 4500 \; \mathrm{cGy}$ \\
Bladder & $\operatorname{D}_{25 \, \%} \leq 6500 \; \mathrm{cGy}$ \\
Near-target & $\operatorname{MTD}^+_{5 \, \%} \leq 5500 \; \mathrm{cGy}$ \\
\end{tabular}
\end{table}
\end{comment}
\section{Results}
As the main purpose of the variational autoencoder is to provide input features for the dose statistic prediction models, we first ensure that the produced features contain sufficient predictive power. For this, we temporarily use a simplistic linear regression model fitted using the data $\{(\phi(x^n), y^n_{\mathrm{pc}})\}_n$, where $y^n_{\mathrm{pc}} = (y^n_{\mathrm{pc}, 1}, y^n_{\mathrm{pc}, 2})$ are the two first principal components of the dose statistic values over all ROIs. Then, for each patient $x^*$ in the test dataset, we compute the standard Euclidean distance between the predicted $y_{\mathrm{pc}}^*$ and each predicted $y^n_{\mathrm{pc}}$, which serves as a relevance-adapted distance metric in the feature space $\mathcal{Z}$. This distance metric is also compared to one constructed analogously but using the hand-crafted distance transform features. Figure \ref{vae_features_linear} depicts this distance in relation to the actual distance between the principal component vectors in scatterplots for three test patients, comparing between the variational autoencoder and the hand-crafted features. Despite the simplicity of the linear model, for the variational autoencoder features, the relevance-adapted distances agree well with the principal component distances, with points lying close to the ground truth also being close in feature space. As a start, this shows that the features produced by the variational autoencoder contain substantial information of relevance for predicting dose statistics. Moreover, the scatterplots for the hand-crafted features show a more irregular relation between the relevance-adapted distances and the true principal component distances. While this does not per se entail that these features contain less predictive power than those produced by the variational autoencoder, it is indicative of a more entangled feature space, which may aggravate the training of the subsequent dose statistic prediction model.
\begin{figure}[h]
\centering
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics[width=\textwidth]{vae_features_linear.pdf}
\caption{Variational autoencoder features.}
\end{subfigure}%
\vspace{0.5cm}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[width=\textwidth]{dt_features_linear.pdf}
\caption{Hand-crafted features.}
\end{subfigure}
\caption{Scatterplot of the first and second principal components (PCs) of the dose statistics in the training data (round) and one test patient (diamond), plotted for three patients in the test dataset. A darker purple (green) color corresponds to being closer (farther) in the feature space with respect to a relevance-adapted metric constructed from variational autoencoder (a) and hand-crafted (b) features.}
\label{vae_features_linear}
\end{figure}
Next, we assess the predictive performance of the seven ROI-specific dose statistic prediction models. Naturally, since the true predictive distributions are unknown, we cannot directly evaluate the accuracy of the estimated predictive distributions by computing e.g. Kullback--Leibler divergences to ground truths. However, we may use the training dataset as a reference sample and compare average errors between that and a sample drawn from the estimated predictive distribution. Table \ref{tableresults} shows the mean squared error (MSE) $\operatorname{\mathbb{E}} \Vert y - y^* \Vert^2_{\mathrm{std}}$ and the mean absolute error (MAE) $\operatorname{\mathbb{E}} \Vert y - y^* \Vert_{\mathrm{std}}$ for each model, averaged over the $10$ test patients---here, expectation is taken over $y \sim p(y \; | \; \phi(x^*), \{(\phi(x^n), y^n)\}_n)$ and $\Vert \cdot \Vert_{\mathrm{std}}$ denotes a standardized Euclidean norm using the sample standard deviations in the training dataset. Figure \ref{metric_boxplot} shows boxplots illustrating the spreads of the MSE and MAE errors over the test dataset beside corresponding benchmarks consisting of their empirical analogues over the training dataset. For the organs at risk, the dose statistic prediction outperforms the choice of a training data point at random, as would be expected---in particular, the interpretation of the figure is that the models mostly output predictions better than most other training data points, since the model boxes are centered around the lower half of their respective benchmark boxes. In this context, it is important to note that there is a theoretical limit on how low the boxplots may lie due to the inherent uncertainty of decisions made during treatment planning. For the targets, however, interestingly, the models perform slightly worse than the benchmark. There may be various reasons for this---for example, the predictive uncertainties may have been overestimated, or the target DVHs may not actually depend noticeably on the patient geometry. Finally, in Figure \ref{dvhband_1}, we show, for two patients in the test dataset, pointwise DVH confidence bands associated with the dose statistic prediction models for the seven ROIs.
\begin{table}[h]
\caption{Comparison of the standardized MSE and MAE values, averaged over the test dataset, between the proposed models and their respective benchmarks, with the benchmark value defined as the average error over the training dataset. Note that the associated spreads of the validation metric values correspond to the boxplots in Figure \ref{metric_boxplot}.}
\centering
\begin{tabular}{lrrrr}
\toprule
\multirow{2}{*}[-5pt]{ROI} & \multicolumn{2}{c}{MSE} & \multicolumn{2}{c}{MAE}\\
\addlinespace[0pt]
\cmidrule(lr){2-3} \cmidrule(lr){4-5} \\
\addlinespace[-10pt]
{} & Model & Benchmark & Model & Benchmark \\
\midrule
PTV, prostate & $77.51$ & $21.49$ & $22.13$ & $12.62$ \\
PTV, seminal vesicles & $42.60$ & $25.78$ & $17.10$ & $13.22$ \\
Rectum & $13.11$ & $14.84$ & $8.54$ & $9.04$ \\
Bladder & $13.15$ & $18.51$ & $8.37$ & $9.97$ \\
Left femur & $11.67$ & $16.77$ & $7.69$ & $9.39$ \\
Right femur & $12.11$ & $16.06$ & $8.10$ & $9.09$ \\
Small bowel & $11.61$ & $15.12$ & $8.26$ & $9.21$ \\
\bottomrule
\end{tabular}
\label{tableresults}
\end{table}
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{metric_boxplot.pdf}
\caption{Boxplots of the MSE and MAE for the different models in comparison to their benchmark models. The diamond-shaped markers show outliers.}
\label{metric_boxplot}
\end{figure}
\begin{figure}[h]
\centering
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics[width=\textwidth]{dvh_bands_4.pdf}
\caption{Patient 1.}
\end{subfigure}%
\vspace{0.5cm}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[width=\textwidth]{dvh_bands_6.pdf}
\caption{Patient 2.}
\end{subfigure}
\caption{Illustration of dose statistic predictions for two test patients, where the shaded bands correspond to $99 \; \%$, $95 \; \%$ and $70 \; \%$ prediction intervals for each dose statistic. The training equivalents are shown in gray and the ground truth in black.}
\label{dvhband_1}
\end{figure}
Finally, to demonstrate that the dose mimicking method outlined in Section \ref{dosemimickingsec} can leverage the information contained in the predictive distributions from the dose statistic prediction, we create deliverable plans for two patients in the test dataset. This is then compared to the non-probabilistic formulation using quadratic penalties instead of cross-entropy objectives. The resulting DVHs and spatial dose distributions of the mimicked dose distributions are shown in Figure \ref{mimic_dvhs} in comparison to their ground truth counterparts. Interestingly, we observe that for both patient cases, the probabilistically mimicked DVHs follow the ground truth mostly well, possibly with slight deviations for the targets. On the other hand, the quadratic penalty mimicking was overly aggressive in reducing the dose to the rectum and the bladder at the cost of sacrificing target coverage and homogeneity. The reason for this is that the quadratic penalties, containing no uncertainty insight, are overconfident that the predictive mean DVHs could be achieved---the predictions in the rectum and the bladder, while being colder than the clinical plan for both patient cases, are associated with a much higher uncertainty than for the targets, which the probabilistic dose mimicking formulation was able to take into account. The specification of preferences captured in the predictive distributions are thus being realized differently depending on the actual patient geometry. Hence, this showcases the advantage of articulating one's preferences by a probability distribution instead of a single reference value. On the other hand, we see that the spatial dose distributions for both mimicking formulations follow the ground truths mostly well but may differ significantly in parts not covered by any ROI, with the quadratic penalty formulation deviating slightly more, which is expected since the optimization problem does not penalize such deviations. While we may conclude that the comparison demonstrates the merits of a probabilistic approach to dose mimicking, it is emphasized that in a clinical setting, further post-processing is in general required to obtain plans of quality sufficient to be approved and delivered.
\begin{figure}[h]
\centering
\begin{subfigure}[t]{\textwidth}
\centering
\includegraphics[width=\textwidth]{mimic_comparison.pdf}
\caption{}
\end{subfigure}%
\vspace{0.5cm}
\begin{subfigure}[b]{\textwidth}
\centering
\includegraphics[width=\textwidth]{all_four.png}
\caption{}
\end{subfigure}
\caption{(a) DVH comparisons between the mimicked plans using the probabilistic (solid) and non-probabilistic (dash-dotted) formulations and their clinical counterparts (dashed), shown for the two test patients in Figure \ref{dvhband_1}. Note that the optimization problem is not set up to mimic the clinical plan but rather the predicted dose statistics. (b) Transversal cuts of the spatial dose distributions of the mimicked plans in comparison to their clinical counterparts, shown for the two test patients in Figure \ref{dvhband_1}. The colors of the ROI contours correspond to those in the legend in (a).}
\label{mimic_dvhs}
\end{figure}
\section{Discussion}
In the context of predicting dose-related quantities as part of an automated treatment planning pipeline, the importance of precise quantification of predictive uncertainties has been left largely unaddressed in literature. Motivated by this, in this work, we have presented a general framework in which the multivariate predictive distribution of any collection of dose-related quantities may be estimated and leveraged in a dose mimicking problem. We have shown that a variational autoencoder can be employed to extract from contoured CT images lower-dimensional features of substantial relevance for predicting dose statistics, and that a similarity-based mixture-of-experts model can be used to quantify predictive uncertainties of dose statistic values, including inter-statistic dependencies. Moreover, we have shown that by setting up a dose mimicking problem in which contributions from dose statistics are based on the marginal cumulative distribution functions in a cross-entropy--like form, with gradients of dose statistics directly available by virtue of recent developments, we may create deliverable plans agreeing well with their respective historical clinical plans with minimal need for weight tuning. In particular, the comparison between the proposed dose mimicking formulation and a non-probabilistic analogue has showcased the merits of including uncertainty information during optimization.
The nonparametric mixture-of-experts method used for dose statistic prediction is similar to previous work in terms of training requirements, portability and generalizability. Compared to the methods in early studies on knowledge-based planning \citep{zhu, wu, yuan}, instead of predicting principal component coefficients, our method outputs complete probability distributions capable of modeling inter-statistic dependencies, skewness and eventual multimodality. As argued by \citet{babier_importance}, it is crucial to maintain a holistic perspective when developing and assessing constituent methods in a prediction--mimicking pipeline---there is, for instance, no direct causality between the performance of a dose prediction model and the quality of the produced plan. In this sense, the proposed division of the pipeline into feature extraction, dose statistic prediction and dose mimicking has the advantages of being flexible, with each part being substitutable with any other algorithm, and general, with minimal loss of information between the steps. Moreover, the former two parts may naturally be combined into a semi-supervised learning algorithm, reducing the need for training data in the form of clinically delivered treatments. A disadvantage, however, is that the dose-related quantities to be predicted must be meaningful when evaluated on all training patients, which is not true for e.g. individual voxel doses of the current patient. Thus, a spatial dose prediction algorithm may complement the current dose mimicking problem setup by helping to shape the dose distribution outside the defined ROIs. If done also probabilistically, such as in \citet{nguyen_bagging}, the spatial dose prediction could be basis for a contribution to the dose mimicking objective in (\ref{optformulation}) penalizing spatial deviation, similarly to in \citet{nilsson}, as well as serve as a reference for our proposed dose statistic prediction method.
While the present work primarily focuses on conceptually demonstrating the merits of the proposed method, there are several interesting directions for future research. For example, the features extracted by the variational autoencoder may benefit from being augmented with handcrafted counterparts such as in \citet{cui}, and the estimated predictive distribution over dose statistics may be used for quality assurance purposes. Furthermore, with an estimated joint predictive distribution, one may use the inter-statistic dependencies to post-process the distribution to be more aggressive toward one or several dose statistics. This may, in turn, be used to form differently focused tradeoff objectives in a multicriteria optimization problem, leading to a semiautomated treatment planning workflow in which the user may optionally articulate their own preferences by navigating between Pareto optimal plans. In conclusion, we note that a rigorous probabilistic perspective of the automated treatment planning problem enables ample new opportunities to be further explored in future work.
\section{Conclusions}
In this work, we have presented a general framework for automated radiation therapy treatment planning based on a three-step pipeline comprising feature extraction, dose statistic prediction and dose mimicking. By using latent representations produced by a variational autoencoder and a nonparametric Bayesian dose statistic prediction method, we obtain reasonable predictive distributions as multivariate Gaussian mixtures. Furthermore, using objective functions specially designed to leverage such predictive information, we perform dose mimicking to create deliverable plans resembling their original counterparts well. Thus, we have demonstrated the advantages of a holistic approach based on probabilistic modeling, with applications to many related tasks.
\printbibliography
\end{document} |
\section{Introduction}\label{sec:intro}
Causality is important for designing interpretable and robust methods in artificial intelligence research~\citep{miller2019explanation}, and has been used in many fields of artificial intelligence, such as causal transfer learning \citep{zhang2020domain,Bengio2020A} and causality-based algorithmic fairness~\citep{counter, ijcai2019-199}. One of the main problems in many of these studies is to infer whether a treatment variable is a cause of a target variable, or to further identify the causes/non-causes of a specified target variable or the effects/non-effects of a given treatment. For example, in a telecommunication network, a single fault (or alarm) in the network can trigger a flood of alarms, and conversely, a recovery of a single fault may clear many alarms. Therefore, knowing the causal relations among the alarms (or faults) is helpful to localize the key failure points for fault recovery in practice.
Directed acyclic graphs (DAGs) can be used to represent causal relationships among variables~\citep{pearl2009causality}. Following Pearl's definition of \emph{inferred causation}~\citep[Definition~2.3.1]{pearl2009causality}, we call $X$ a \emph{cause} of $Y$ and $Y$ an effect of $X$ if $X$ has a directed path to $Y$ in the true DAG. From observational data, however, instead of an exact causal DAG,\footnote{We note that, the recent progresses in identifying the causal relation between two variables indeed provide an opportunity to learn an exact DAG. However, such methods need to pose additional distributional conditions~\citep{Shimizu2006lingam, Zhang2009identifiability, Shimizu2011directlingam, Peters2013identifiability,Peters2014causal}} we generally learn a Markov equivalence class of DAGs represented by a completed partially directed acyclic graph (CPDAG). The undirected edges in a CPDAG imply that some causal relations among variables can not be read from the graph directly. Therefore, given a Markov equivalence class of DAGs, a variable $X$ is a \emph{definite cause} of a target $Y$ if $X$ is always a cause of $Y$ in every equivalent DAG, and a variable $X$ is a \emph{definite non-cause} of $Y$ if $X$ is never a cause of $Y$ in any DAG in the class. If $X$ is neither a definite cause nor a definite non-cause of $Y$, $X$ is called a \emph{possible cause} of $Y$.
Some approaches can be used to identify the type of causal relation between a treatment and a target. An intuitive approach is first to learn a Markov equivalence class from observational data, and then enumerate all DAGs in the class to check whether the treatment is definitely or definitely not a cause of the target in all of these equivalent DAGs. However, the intuitive approach is inefficient when the number of DAGs in the learned Markov equivalence class is large~\citep{he2015counting}.
Another way is to check the paths from the treatment to the target in a CPDAG. It has been shown that a treatment is a definite non-cause of a target if and only if there is no partially directed path from the treatment to the target \citep[see, e.g.][]{zhang2006, perkovic2017interpreting}. Given a CPDAG, \citet{Roumpelaki} also introduced a sufficient condition for identifying definite causes. However, the necessity of this condition remains a conjecture ~\citep{zhang2006, Mooij2020constraint} and the corresponding approach could be inefficient since it needs to learn an entire CPDAG first.
The third approach is to estimate all possible causal effects of the treatment on the target~\citep{maathuis2009estimating, perkovic2017interpreting, nandy2017estimating, fang2020bgida, liu2020cida, liu2020local, Witte2020efficient, Guo2020minimal}. This approach, which we call the causal-effect-based method, determines whether a treatment is a cause of a target by judging whether all possible causal effects are zeros/non-zeros based on a certain criterion or method, such as hypothesis testing. However, the causal-effect-based method requires additional assumptions,\footnote{We remark that,
$X$ has a zero-valued causal effect on $Y$ does not necessarily mean that there is no directed path from $X$ to $Y$ (See \ref{app:ce} for an example). Nevertheless, with the causal faithfulness assumption as well as some model assumptions such as linear-Gaussianity, the former implies the latter.} and it could be time-consuming as the number of possible effects grows exponentially in the worst case.
In this paper, we study the problem of locally identifying causal relations under Markov equivalence with the assumption that there is no hidden variable or selection bias. That is, given a pair of treatment and target variables, we intend to decide whether the treatment is a definite cause, a possible cause or a definite non-cause of the target only based on a local induced subgraph and a few independence tests related to the treatment without learning an entire CPDAG. This local approach is usually more efficient than the global ones that need an entire CPDAG, especially when the underlying causal graph is large.
To this end, we first discuss the existence of a causal path from one variable to another given a CPDAG, and prove the necessity of the condition in \citet{Roumpelaki} for CPDAGs in Section \ref{sec:anatomy}. This yields a sufficient and necessary graphical condition to check the existence of a causal path.
Next, in Section \ref{sec:characterization} we propose local identification criteria for definite causes, possible causes and definite non-causes separately. These criteria depend only on the induced subgraph of the true CPDAG over the adjacent variables of the treatment as well as some queries about d-separations, thus directly lead to a local learning algorithm given in Section \ref{sec:learn}. For the completeness of the paper, a global algorithm and several causal-effect-based methods for learning types of causal relations are also provided in Section \ref{sec:learn}. In Section \ref{sec:simulation}, we compare experimentally the proposed local learning method with the global and the causal-effect-based methods, and show the efficiency and efficacy of the proposed method. Finally, we discuss some applications and possible extensions of our work in Section \ref{sec: conclusion}, and give some graph terminology, additional algorithms, proofs and additional experimental results in Appendix A, B, C and D, respectively.
\section{Preliminaries and Related Work}\label{sec:background}
In this paper, we use $pa(\textbf{S}, \cal G)$, $ch(\textbf{S}, \cal G)$, $sib(\textbf{S}, \cal G)$, $adj(\textbf{S}, \cal G)$, $an(\textbf{S}, \cal G)$ and $de(\textbf{S}, \cal G)$ to denote the union of the parents, children, siblings (or undirected neighbors), adjacent vertices, ancestors, and descendants of each variable in set $\textbf{S}$ in $\mathcal{G}$, respectively, where ${\cal G}=(\mathbf{V}, \mathbf{E})$ can be a directed, an undirected, or a partially directed graph. The basic graph terminology can be found in \ref{app:graph}. As a convention, we regard a vertex as an ancestor and a descendant of itself. If $\textbf{S}=\{X\}$ is a singleton set, we will replace $\textbf{S}$ by $X$ for ease of presentation. Let $\cal G$ be a causal \emph{directed acyclic graph} (causal DAG) and $X$ be a vertex in $\cal G$, the vertices in $an(X, \mathcal{G})\setminus X$ are \emph{causes} of $X$, and the vertices in $pa(X, \mathcal{G})$ are \emph{direct causes} of $X$. If $X$ is a cause of $Y$, then the directed paths from $X$ to $Y$ are called \emph{causal paths}.
\subsection{Causal DAG Models}\label{sec:sec:causalDAG}
The notion of \emph{d-separation} induces a set of conditional independence relations encoded in a DAG \citep{pearl1988probabilistic}. Let $\mathcal{G}$ be a DAG and $\pi = (X = X_0, X_1, ... , X_n = Y)$ be a path from $X$ to $Y$ in $\mathcal{G}$. An intermediate vertex $X_i$ is a \emph{collider} on $\pi$ if $X_{i-1} \rightarrow X_i$ and $X_i \leftarrow X_{i+1}$, otherwise, $X_i$ is a \emph{non-collider} on $\pi$. For three distinct vertices $X_i, X_j$ and $X_k$, if $X_i\rightarrow X_j\leftarrow X_k$ and $X_i$ is not adjacent to $X_k$ in $\cal G$, then the triple $(X_i, X_j, X_k)$ is called a \emph{v-structure} collided on $X_j$ in $\cal G$. Given $\textbf{Z} \subseteq \textbf{V}$, we say $\pi$ is \emph{d-connected (or active)} given $\textbf{Z}$ if $\textbf{Z}$ does not contain any endpoint or non-collider on the path and every collider on the path has a descendant in $\textbf{Z}$. If $\pi$ is not d-connected given $\textbf{Z}$, then $\pi$ is \emph{blocked} by $\textbf{Z}$. For pairwise disjoint sets $\textbf{X}, \textbf{Y}, \textbf{Z} \subseteq \textbf{V}$, $\textbf{X}$ and $\textbf{Y}$ are d-separated by $\textbf{Z}$ (denoted by $\textbf{X} \indep \textbf{Y} \mid \textbf{Z}$) if and only if every path between some $X\in \textbf{X}$ and $Y\in \textbf{Y}$ is blocked by $\textbf{Z}$.
Let $\mathcal{J}_{\mathcal{G}}$ be the set of d-separation relations read off from a DAG $\cal G$. Two DAGs $\mathcal{G}_1$ and $\mathcal{G}_2$ are Markov \emph{equivalent} if $\mathcal{J}_{\mathcal{G}_1}=\mathcal{J}_{\mathcal{G}_2}$. \citet{pearl1989conditional} have shown that two DAGs are equivalent if and only if they have the same skeleton and the same v-structures. A \emph{Markov equivalence class} or simply \emph{equivalence class}, denoted by $[\mathcal{G}]$, contains all DAGs equivalent to $\mathcal{G}$. A Markov equivalence class $[\mathcal{G}]$ can be uniquely represented by a partially directed graph called \emph{completed partially directed acyclic graph} (CPDAG) $\mathcal{G}^*$. Two vertices are adjacent in $\mathcal{G}^*$ if and only if they are adjacent in $\mathcal{G}$ and a directed edge occurs in $\mathcal{G}^*$ if and only if it appears in every DAG in $[\mathcal{G}]$~\citep{pearl1989conditional}. For the ease of presentation, we will also use $[\mathcal{G}^*]$ to represent the Markov equivalence class represented by $\mathcal{G}^*$. Given a CPDAG $\mathcal{G}^*$, we use $\mathcal{G}^*_u$ and $\mathcal{G}^*_d$, which consist of all undirected edges and all directed edges in ${\cal G}^*$, to denote the \emph{undirected subgraph} and the \emph{directed subgraph} of ${\cal G}^*$, respectively.
\citet{andersson1997characterization} proved that (1) the undirected subgraph $\mathcal{G}^*_u$ of $\mathcal{G}^*$ is the union of disjoint connected chordal graphs (the definition of chordal graph is provided in \ref{app:graph}), and (2) every partially directed cycle in $\mathcal{G}^*$ is an undirected cycle, that is, none of the partially directed cycles in $\mathcal{G}^*$ contains a directed edge. Each isolated connected undirected subgraph of $\mathcal{G}^*_u$ is called a \emph{chain component} of ${\cal G}^*$~\citep{andersson1997characterization, lauritzen2002chain}.
For a given distribution $P$, we use $\textbf{X} \indep_{P} \textbf{Y} \mid \textbf{Z}$ to denote that $\textbf{X}$ is independent of $\textbf{Y}$ given $\textbf{Z}$ with respect to $P$, where $\textbf{X}, \textbf{Y}, \textbf{Z} \subseteq \textbf{V}$ are pairwise disjoint.
If both $\textbf{X}=\{X\}$ and $\textbf{Y}=\{Y\}$ are singleton sets, we allow that $X$ or $Y \in \textbf{Z}$, and assume that $X\indep Y\mid \textbf{Z}$ trivially holds in this case.
Let $\mathcal{J}_{P}$ be the set of all (conditional) independencies that hold with respect to $P$. The main results of this paper are based on the following assumptions: the \emph{causal Markov assumption}, which states that $\textbf{X} \indep \textbf{Y} \mid \textbf{Z} $ in $\mathcal{J}_{\mathcal{G}} $ implies $\textbf{X} \indep_P \textbf{Y} \mid \textbf{Z} $ in $\mathcal{J}_{P}$; the \emph{causal faithfulness assumption}, which states that $\textbf{X} \indep_P \textbf{Y} \mid \textbf{Z} $ in $ \mathcal{J}_{P}$ implies $\textbf{X} \indep \textbf{Y} \mid \textbf{Z} $ in $ \mathcal{J}_{\mathcal{G}} $; and the assumption that there is no hidden variable or selection bias. A distribution $P$ is called \emph{Markovian} and \emph{faithful} to a DAG $\mathcal{G}$ if $P$ and ${\cal G}$ satisfy the causal Markov assumption and the causal faithfulness assumption. A causal DAG model consists of a DAG $\cal G$ and a joint distribution $P$ over a common vertex set $\mathbf{V}$ such that $P$ satisfies the causal Markov assumption with respect to $\cal G$. $\cal G$ is called the \emph{causal structure} of the model and $P$ is called the \emph{observational distribution} (or simply \emph{distribution})~\citep{hauser2012characterization}.
\subsection{Global and Local Causal Structure Learning}\label{sec:sec:structure_learning}
Causal structure learning methods try to recover the causal structure from data.
Global causal structure learning focuses on learning an entire causal structure over all variables while local causal structure learning aims to recover only a part of the underlying causal structure.
Existing approaches for learning global causal structures roughly fall into two classes: constraint-based and score-based methods. Constraint-based methods, such as the PC algorithm~\citep{spirtes1991algorithm} and the stable PC algorithm~\citep{colombo2014order}, use conditional independence tests to find causal skeleton and then determine the edge directions according to a series of orientation rules~\citep{meek1995causal}. Under the causal Markov and causal faithfulness assumptions, constraint-based methods can identify causal graphs up to a Markov equivalence class. On the other hand, score-based methods, such as exact search algorithms like dynamic programming \citep{Koivisto2004exact,Singh2005finding} and A* \citep{Yuan2011learning,Xiang2013lasso}, greedy search algorithms like GES \citep{chickeringo2002optimal}, and gradient-based methods like NOTEARS~\citep{zheng2018dags}, evaluate candidate graphs with a predefined score function and search for the optimal DAGs or CPDAGs.
Local learning algorithms usually learn the Markov blanket~\citep[see, e.g.][]{tsamardinos2003algorithms, Tsamardinos2003towards, fu2010} or the parent and child set of a given target~\citep[see, e.g.][]{wang2014discovering, gao2015local, liu2019}. Recently, \citet[Algorithm~3]{liu2020local} extended the MB-by-MB algorithm \citep{wang2014discovering} to learn the chain component containing a given target and the directed edges surrounding the chain component. This variant of MB-by-MB can thus learn the induced subgraph of the true CPDAG over the target and its neighbors, that is, the parents, siblings and children of the target in the CPDAG.
\subsection{Related Work}
As discussed in Section \ref{sec:intro}, when a learned CPDAG is provided, one can either enumerate all equivalent DAGs, or check the paths in the CPDAG~\citep{zhang2006, Roumpelaki, perkovic2017interpreting}, or use the causal-effect-based method to identify types of causal relations~\citep{maathuis2009estimating, perkovic2017interpreting}.
Many sufficient conditions are also available to identify some of causal relations without estimating a global causal structure \citep{cooper97, spirtes2000causation, mani06y, pearl2009causality, Claassen11alogical, colombo2014order, Magliacane2016ancestral}.
For example, if $X\nindep Y\mid \mathbf{W}\cup Z$ while $X\indep Y\mid \mathbf{W}$, then $Z$ is a definite non-cause of every variable in $X\cup Y \cup \mathbf{W}$~\citep{Claassen11alogical}. Since these rules are sound but not complete, they may fail to identify the causal relation of a given pair of treatment and target.
Recently, a related work from \cite{entner13a} proposed sound and complete rules for inferring whether a given variable $X$ has a causal effect on another variable $Y$.
Compared with our work, their criteria allow the existence of unmeasured confounders, but also require two additional assumptions: $Y$ is not a cause of $X$, and neither $X$ nor $Y$ is a cause of other observed variables.
\section{An Anatomy of Causal Relations}\label{sec:anatomy}
In this section, we provide a sufficient and necessary condition to identify definite causal relations, and show that definite causal relations can be divided into two subtypes: explicit and implicit causal relations.
\subsection{Graphical Criteria for Identifying Types of Causal Relations}\label{sec:sec:graphical}
As mentioned in Section \ref{sec:intro},
given a CPDAG, a variable $X$ is a definite non-cause of another variable $Y$ if and only if there is no partially directed path from $X$ to $Y$~\citep{zhang2006, perkovic2017interpreting}.
\citet[Theorem~3.1]{Roumpelaki} proved that a treatment is a definite cause of a target if there is a directed path from the treatment to the target or the treatment has two chordless partially directed paths to the target on which two vertices adjacent to the treatment are distinct and non-adjacent. In the section, we will show that this condition is also necessary, and before that, a concept of \emph{critical set} is introduced as follows.
\begin{definition}[Critical Set]\label{def:critical-set}
{\rm {\citep[Definition~2]{fang2020bgida}}}
Let $\mathcal{G}^*$ be a CPDAG, and $X$ and $Y$ be two distinct vertices in $\mathcal{G}^*$. The critical set of $X$ with respect to $Y$ in $\mathcal{G}^*$ consists of all adjacent vertices of $X$ lying on at least one chordless partially directed path from $X$ to $Y$.
\end{definition}
The definition of chordless partially directed path can be found in \ref{app:graph}. With Definition \ref{def:critical-set}, we have the following lemma.
\begin{lemma}\label{lem:child_critical}
Let $\mathcal{G}^*$ be a CPDAG. For any two distinct vertices $X$ and $Y$ in $\mathcal{G}^*$, $X$ is a definite cause of $Y$ in the underlying DAG if and only if the critical set of $X$ with respect to $Y$ in $\mathcal{G}^*$ contains a child of $X$ in every DAG ${\cal G}\in[\mathcal{G}^*]$.
\end{lemma}
Lemma \ref{lem:child_critical} follows from Lemma~2 in \citet{fang2020bgida}. It gives a sufficient and necessary condition to decide whether $X$ is a definite cause of $Y$. However, checking the condition given in Lemma \ref{lem:child_critical} also requires to enumerate all equivalent DAGs. To mitigate this problem, we discuss a graphical characteristic of critical set in the corresponding CPDAG.
\begin{lemma}\label{lem:critical_has_a_child}
Let $\mathcal{G}^*$ be a CPDAG and $X, Y$ be two distinct vertices in $\mathcal{G}^*$. Denote by $\mathbf{C}$ the critical set of $X$ with respect to $Y$ in $\mathcal{G}^*$, then $\mathbf{C}\cap ch(X, {\cal G})=\emptyset$ for some ${\cal G} \in [\mathcal{G}^*]$ if and only if $\mathbf{C} = \emptyset$, or $\mathbf{C}$ induces a complete subgraph of $\mathcal{G}^*$ but $\mathbf{C}\cap ch(X, \mathcal{G}^*)= \emptyset$.
\end{lemma}
Based on Lemmas \ref{lem:child_critical} and \ref{lem:critical_has_a_child}, we have the desired sufficient and necessary graphical criterion.
\begin{theorem}\label{thm:graphical_definite_cause}
Suppose that $\mathcal{G}^*$ is a CPDAG, $X, Y$ are two distinct vertices in $\mathcal{G}^*$, and $\mathbf{C}$ is the critical set of $X$ with respect to $Y$ in $\mathcal{G}^*$. Then, $X$ is a definite cause of $Y$ if and only if $\mathbf{C}\cap ch(X, \mathcal{G}^*)\neq \emptyset$, or $\mathbf{C}$ is non-empty and induces an incomplete subgraph of $\mathcal{G}^*$.
\end{theorem}
The sufficiency of the condition in Theorem \ref{thm:graphical_definite_cause} has been extended to other types of causal graphs by \citet{Roumpelaki} and \citet{Mooij2020constraint}.\footnote{We note that, although ~\citet[Theorem~3.1]{Roumpelaki} also claimed that they have proved the necessity, their proof is flawed. As mentioned by \citet{Mooij2020constraint}, the last part of the proof appears to be incomplete. How to prove the necessity for more general types of causal graphs remains an open problem~\citep{zhang2006}.} With the help of Theorem \ref{thm:graphical_definite_cause}, we can identify the type of causal relation based on a learned CPDAG by enumerating paths and finding critical sets. Below, we give an example to illustrate this idea
\begin{example}\label{exa:difinitecause} Consider the respiratory disease network shown in Figure \ref{fig:motivation_1}. The meanings of the node labels are given in the caption. Let smoking be the treatment and dyspnoea be the target. From Figure \ref{fig1_2} we can see that the partially directed paths from smoking to dyspnoea are ${\rm Smok}-{\rm Lung}\rightarrow{\rm Either}\rightarrow{\rm Dysp}$ and ${\rm Smok}-{\rm Bronc}\rightarrow{\rm Dysp}$. Therefore, the critical set of smoking with respect to dyspnoea is $\{{\rm Lung}, {\rm Bronc}\}$. As ${\rm Lung}$ and ${\rm Bronc}$ are not adjacent, by Theorem \ref{thm:graphical_definite_cause} smoking is a definite cause of dyspnoea. Similarly, the critical set of lung cancer with respect to dyspnoea is $\{{\rm Smok}, {\rm Either}\}$. Since ${\rm Either}$ is a child of ${\rm Lung}$, lung cancer is also a definite cause of dyspnoea.
\begin{figure}[!t]
\centering
\subfigure[$\mathcal{G}_{t}$]{
\begin{minipage}[t]{0.31\linewidth}
\centering
\includegraphics[width=1.3in]{figs/fig1new_1.png}
\vspace{0.35cm}
\label{fig1_1}
\end{minipage}%
}%
\subfigure[$\mathcal{G}^*$]{
\begin{minipage}[t]{0.31\linewidth}
\centering
\includegraphics[width=1.3in]{figs/fig1new_2.png}
\vspace{0.35cm}
\label{fig1_2}
\end{minipage}%
}%
\subfigure[Markov eqivelence class]{
\begin{minipage}[t]{0.35\linewidth}
\centering
\includegraphics[width=1.25in]{figs/fig1new_3.png}
\vspace{0.35cm}
\label{fig1_3}
\end{minipage}
}%
\centering
\caption{This example is adapted from the ASIA network. The original network structure and related parameters can be found in \citet{Lauritzen1988Local}. Figure \ref{fig1_1} shows the true underlying causal DAG, and Figure \ref{fig1_2} shows the corresponding CPDAG. Figure \ref{fig1_3} enumerates all equivalent DAGs in the Markov equivalence class. The meanings of the node labels are: recently have been to the \textbf{Hosp}ital, test positive for \textbf{Tub}erculosis, \textbf{Smok}ing, test positive for \textbf{Lung} cancer, \textbf{Bronc}hitis, \textbf{Either} have lung cancer or have tuberculosis, test positive for \textbf{X-ray}, and test positive for \textbf{Dysp}noea.}
\label{fig:motivation_1}
\end{figure}
\end{example}
\subsection{Explicit and Implicit Causal Relations}\label{sec:sec:concepts}
We now study the properties of definite causal relations, and show that definite causal relations can be divided into two subtypes based on the existence of causal paths in a CPDAG. The results in this section are of key importance to build local characterizations in Section \ref{sec:characterization}, and are also useful for developing an efficient global learning algorithm.
\begin{proposition}\label{prop:definite-cause}
For two distinct vertices $X$ and $Y$, if $X$ is a definite cause of $Y$, then $X$ and $Y$ are not in the same chain component.
\end{proposition}
Given a target variable $Y$, Proposition \ref{prop:definite-cause} shows that $Y$ and its definite causes do not appear in the same chain component. Thus, if a treatment $X$ is a definite cause of a target $Y$, then in ${\cal G}^*$ there must be a partially directed path from $X$ to $Y$ which contains a directed edge. On the other hand, for two distinct vertices lying in the same chain component, we have,
\begin{proposition}\label{prop:chain-comp}
Two distinct vertices $X$ and $Y$ are possible causes of each other if and only if they are in the same chain component.
\end{proposition}
Recall that in Figure \ref{fig1_2}, both ${\rm Smok}$ and ${\rm Lung}$ are definite causes of ${\rm Dysp}$. However, in the CPDAG there exists a directed path from ${\rm Lung}$ to ${\rm Dysp}$ while no directed path exists from ${\rm Smok}$ to ${\rm Dysp}$. That is, the cause ${\rm Lung}$ of ${\rm Dysp}$ is explicit and the cause ${\rm Smok}$ of ${\rm Dysp}$ is implicit in the CPDAG. This difference motivates the following two concepts.
\begin{definition}[Explicit Cause]\label{def:explicit}
A variable $X$ is an explicit cause of $Y$ if
there is a common causal path from $X$ to $Y$ in every DAG in the Markov equivalence class represented by a CPDAG $\mathcal{G}^*$.
\end{definition}
As there is a common causal path from an explicit cause $X$ to the target $Y$ in every DAG in the Markov equivalence class represented by $\mathcal{G}^*$, there is a directed path from $X$ to $Y$ in $\mathcal{G}^*$, and thus $X$ is a definite cause of $Y$.
\begin{definition}[Implicit Cause]\label{def:implicit}
A variable $X$ is an implicit cause of $Y$ if $X$ is a definite cause of $Y$ and there is no common causal path from $X$ to $Y$ in all DAGs in the Markov equivalence class represented by a CPDAG $\mathcal{G}^*$.
\end{definition}
We notice that $X$ is a
definite cause of $Y$ if only if it satisfies one of the two conditions given in Theorem \ref{thm:graphical_definite_cause}. The first condition, $\mathbf{C}\cap ch(X, \mathcal{G}^*)\neq \emptyset$, is the sufficient and necessary condition for identifying explicit causes, while the second condition corresponds to implicit causes. In Section \ref{sec:characterization}, we will exploit this difference between explicit and implicit causes to develop local characterizations for both of them. Below, we give an illustrative example.
\begin{figure}[!t]
\centering
\includegraphics[width=0.35\linewidth]{figs/fig4.png}
\caption{ An example for identifying the types of causal relations}
\label{fig4}
\end{figure}
\begin{example}\label{e1}
Consider the causes of the target variable $Y$ based on the CPDAG $\mathcal{G}^*$ in Figure \ref{fig4}. It is clear that all the variables other than $Y$ are definite or possible causes of $Y$. Obviously, $\{E, D, F\}$ are explicit causes of $Y$. For $B$, since $B-E\to Y$, $B-D\to Y$ and $B-G-F\to Y$ are chordless partially directed paths, the critical set of $B$ with respect to $Y$ is $\{E, D, G\}$. As the induced subgraph of $\mathcal{G}^*$ over $\{E, D, G\}$ is not complete, $B$ is a definite cause of $Y$, and $B$ is also implicit. Similarly, $G$ is another implicit cause of $Y$. For $X$ and $A$, the critical set of $X$ and $A$ with respect to $Y$ are $\{B, D, G\}$ and $\{X, G\}$, respectively. Since the corresponding induced subgraphs are complete, by Theorem \ref{thm:graphical_definite_cause}, $X$ and $A$ are not implicit causes of $Y$. Thus, they are possible causes of $Y$.
\end{example}
Despite the difference, explicit and implicit causes also have some interesting connections. The following Proposition \ref{prop:nec-implicit} proves that the existence of an implicit cause implies the existence of at least two explicit causes.
\begin{proposition}\label{prop:nec-implicit}
Let $\mathcal{G}^*$ be a CPDAG and $X$ and $Y$ be two vertices of it in different chain components. If $X$ is the only explicit cause of $Y$ in the chain component to which $X$ belongs, then every vertex in this chain component, except $X$, is a possible cause of $Y$.
\end{proposition}
\section{Local Characterizations of Types of Causal Relations}\label{sec:characterization}
In this section, we introduce the theoretical results on locally characterizing different types of causal relations. Our local characterizations depend on the induced subgraph of the true CPDAG over the treatment's neighbors as well as some queries about d-separation relations. The first result is about definite non-causal relations, as given in Theorem~\ref{thm:non-cause-no-bk}.
\begin{theorem}\label{thm:non-cause-no-bk}
Let $\mathcal{G}^*$ be a CPDAG. For any two distinct vertices $X$ and $Y$ in $\mathcal{G}^*$, $X$ is a definite non-cause of $Y$ if and only if $X \indep Y \mid pa(X, \mathcal{G}^*)$ holds.
\end{theorem}
Theorem \ref{thm:non-cause-no-bk} introduces a local characterization for definite non-causal relations, which is based on the local structure around the treatment $X$ and a single d-separation claim. The d-separation claim $X \indep Y \mid pa(X, \mathcal{G}^*)$ is similar to the following well-known result called \emph{local Markov property} of a causal DAG model: any variable is d-separated from its non-descendants given its parents. The difference is that in our local characterization, only the parents of $X$ in the CPDAG are included in the separation set, and we rule out the siblings of $X$ even if they may be the parents of $X$ in the true causal DAG. Since in a causal DAG, the non-descendants of a variable are those which are definitely not caused by the variable, Theorem \ref{thm:non-cause-no-bk} can be regarded as an extension of the local Markov property to CPDAGs.
Following Theorem \ref{thm:non-cause-no-bk}, we can distinguish definite and possible causes from definite non-causes with a local causal structure query and a d-separation query. Next, we characterize explicit and implicit causal relations locally in Theorem \ref{thm:explicit-cause-no-bk} and Theorem \ref{thm:implicit-cause-no-bk}, respectively, which together characterize definite causal relations.
\begin{theorem}\label{thm:explicit-cause-no-bk}
Let $\mathcal{G}^*$ be a CPDAG. For any two distinct vertices $X$ and $Y$ in $\mathcal{G}^*$, $X$ is an explicit cause of $Y$ if and only if $X \nindep Y \mid pa(X, \mathcal{G}^*)\cup sib(X, \mathcal{G}^*)$ holds.
\end{theorem}
The local characterization in Theorem \ref{thm:explicit-cause-no-bk} includes a single d-separation claim, $X \nindep Y \mid pa(X, \mathcal{G}^*)\cup sib(X, \mathcal{G}^*)$, which means the set $pa(X, \mathcal{G}^*)\cup sib(X, \mathcal{G}^*)$ cannot block all paths from $X$ to $Y$. In the proof of this theorem, we show that this claim is equivalent to that there exists at least one path from $X$ to $Y$ in $\mathcal{G}^*$ on which the node adjacent to $X$ is a child of $X$. Based on \citet[Lemma~7.2]{maathuis2015generalized} and \citet[Lemma~B.1]{perkovic2017interpreting}, this implies that there is a directed path from $X$ to $Y$ in $\mathcal{G}^*$.
We remark that the sufficiency of Theorem~\ref{thm:explicit-cause-no-bk} is related to the LWF local Markov property~\citep{Frydenberg1990}. Given a chain graph $\cal C$ over a vertex set $\mathbf{V}$ and a distribution $P$ over the same vertex set, $P$ is called LWF local Markovian (or local G-Markovian) to $\cal C$ if $X\indep_P \mathbf{V}\setminus (de(X, {\cal C})\cup pa(X, {\cal C})\cup sib(X, {\cal C})) \mid pa(X, {\cal C})\cup sib(X, {\cal C})$ for any $X\in \mathbf{V}$~\citep{Frydenberg1990}. Since \cite{andersson1997characterization} proved that a CPDAG ${\cal G}^*$ is a chain graph, if a distribution $P$ is LWF local Markovian to ${\cal G}^*$, then $X\nindep_P Y \mid pa(X, {\cal G}^*)\cup sib(X, {\cal G}^*)$ for $Y\notin pa(X, {\cal G}^*)\cup sib(X, {\cal G}^*)$ implies that $Y\in de(X, {\cal G}^*)$. That is, $X$ is an explicit cause of $Y$. In \ref{proof:explict}, we show the sketch of proving the sufficiency of Theorem~\ref{thm:explicit-cause-no-bk} based on the theories of chain graph models.
\begin{theorem}\label{thm:implicit-cause-no-bk}
Suppose that $\mathcal{G}^*$ is a CPDAG and $\mathcal{M}$ is the set of maximal cliques of the induced subgraph of $\mathcal{G}^*$ over $sib(X, \mathcal{G}^*)$. Then, $X$ is an implicit cause of $Y$ if and only if $X \indep Y \mid pa(X, \mathcal{G}^*)\cup sib(X, \mathcal{G}^*)$ and $X \nindep Y \mid pa(X, \mathcal{G}^*)\cup \textbf{M}$ for any $\textbf{M}\in \mathcal{M}$.
\end{theorem}
The definition of maximal clique is given in \ref{app:graph}. In Theorem \ref{thm:implicit-cause-no-bk}, the first condition $X \indep Y \mid pa(X, \mathcal{G}^*)\cup sib(X, \mathcal{G}^*)$ makes sure that $X$ is not an explicit cause of $Y$ and the second condition, which is $X \nindep Y \mid pa(X, \mathcal{G}^*)\cup \textbf{M}$ for any $\textbf{M}\in \mathcal{M}$, guarantees that $X$ is not a possible cause of $Y$. These two conditions in Theorem \ref{thm:implicit-cause-no-bk} are local in the sense that both $sib(X, \mathcal{G}^*)$ and $pa(X, \mathcal{G}^*)$ are subsets of $X$'s neighbors in $\mathcal{G}^*$, and a maximal clique $\textbf{M}$ is also a subset of $sib(X, \mathcal{G}^*)$. Once we obtain the induced subgraph of $\mathcal{G}^*$ over $adj(X, \mathcal{G}^*)$, we can know $sib(X, \mathcal{G}^*)$ and $\mathcal{M}$, and thus the conditional independence queries can be answered accordingly if we have the oracles.
As mentioned in Section \ref{sec:sec:concepts}, definite causes include both explicit and implicit causes. Therefore, Theorems \ref{thm:explicit-cause-no-bk} and \ref{thm:implicit-cause-no-bk} give a sound and complete local characterization of definite causal relations as follows.
\begin{corollary}\label{cor:definite}
Suppose that $\mathcal{G}^*$ is a CPDAG and $\mathcal{M}$ is the set of maximal cliques of the induced subgraph of $\mathcal{G}^*$ over $sib(X, \mathcal{G}^*)$. Then, $X$ is a definite cause of $Y$ if and only if $X \nindep Y \mid pa(X, \mathcal{G}^*)\cup sib(X, \mathcal{G}^*)$ or $X \nindep Y \mid pa(X, \mathcal{G}^*)\cup \textbf{M}$ for any $\textbf{M}\in \mathcal{M}$.
\end{corollary}
Together with Theorem \ref{thm:non-cause-no-bk}, Corollary \ref{cor:definite} can be used to identify definite causal relations and definite non-causal relations. This result is local in the sense that it only depends on the local structure around the treatment $X$ and a limited number of d-separation queries. When data is available in practice, d-separation queries can be answered by performing statistical independence tests. Thus, local characterizations are particularly meaningful for identifying types of causal relations from observational data.
\section{Algorithms}\label{sec:learn}
In this section, we discuss how to learn the types of causal relations from observational data. A local algorithm, which exploits the local characterizations in Section \ref{sec:characterization} directly, is provided in Section~\ref{sec:sec:local}. For the completeness of the paper, we also provide an efficient global learning method in Section~\ref{sec:sec:global}, and causal-effect-based methods in Section~\ref{sec:sec:ce}.
\subsection{A Local Learning Algorithm}\label{sec:sec:local}
\begin{algorithm}[!t]
\caption{A local algorithm for identifying the type of causal relation (local ITC) }
\label{algo:local}
\begin{algorithmic}[1]
\REQUIRE
A treatment $X$, a target $Y$, $pa(X, {\cal G}^*)$, the induced subgraph of ${\cal G}^*$ over $sib(X, {\cal G}^*)$, and independence oracles.
\ENSURE
The type of causal relation between X and Y.
\IF {$X \indep Y \mid pa(X, {\cal G}^*)$}
\RETURN {$X$ is a \textbf{definite non-cause} of $Y$,}
\ENDIF
\IF {$X \nindep Y \mid pa(X, {\cal G}^*)\cup sib(X, {\cal G}^*)$}
\RETURN {$X$ is an \textbf{explicit cause} of $Y$,}
\ENDIF
\STATE {$\mathcal{M}=$ the set of maximal cliques of $sib(X, {\cal G}^*)$,}
\IF {exists $\textbf{M}\in \mathcal{M}$ such that $X \indep Y \mid pa(X, {\cal G}^*)\cup \textbf{M}$,}
\RETURN {$X$ is a \textbf{possible cause} of $Y$,}
\ENDIF
\RETURN {$X$ is an \textbf{implicit cause} of $Y$.}
\end{algorithmic}
\end{algorithm}
The main procedure of our local algorithm is summarized in Algorithm \ref{algo:local}.
The input of Algorithm \ref{algo:local} consists of $pa(X, {\cal G}^*)$, the induced subgraph of ${\cal G}^*$ over $sib(X, {\cal G}^*)$, and some independence oracles. The first two arguments, $pa(X, {\cal G}^*)$ and the induced subgraph over $sib(X, {\cal G}^*)$, can be learned locally by using the variant of the MB-by-MB algorithm proposed by \citet[Algorithm~3]{liu2020local}, which is designed for learning the chain component containing a given target variable and the directed edges connected to the variables in the chain component.
The third argument (the independence oracles), as discussed in Section \ref{sec:characterization}, can be replaced by statistical independence tests in practice. Overall, the procedure given in Algorithm \ref{algo:local} is a direct application of the local characterizations in Theorems \ref{thm:non-cause-no-bk}, \ref{thm:explicit-cause-no-bk} and \ref{thm:implicit-cause-no-bk}, and thus we have,
\begin{theorem}\label{thm:algo-local}
Given a CPDAG ${\cal G}^*$ over $\mathbf{V}$ and the independence oracles faithful to a DAG in $[{\cal G}^*]$, the local ITC (Algorithm \ref{algo:local}) is sound and complete for identifying explicit causes, implicit causes, possible causes and definite non-causes of any variable $Y$ in ${\cal G}^*$.
\end{theorem}
Here, the soundness and completeness mean that the identified causes of each type are all and only those variables satisfying the definition of the corresponding type of cause. For example, the learned explicit causes of $Y$ are all and only the variables in $\mathbf{V}\setminus\{Y\}$ each of which has at least a common directed path to $Y$ in all equivalent DAGs.
The complexity of Algorithm \ref{algo:local} can be measured by the maximum number of conditional independence tests (or d-separation queries). Clearly, the maximum number of conditional independence tests performed by Algorithm \ref{algo:local} is $m+2$, where $m$ is the number of maximal cliques of $sib(X, {\cal G}^*)$. Fortunately, there are only linearly many maximal cliques (with respect to the number of vertices) in a chordal graph \citep{Rose1975elimination, blair1993chordal}, so the number of conditional independence tests needed in Algorithm \ref{algo:local} is at most $O(|sib(X, {\cal G}^*)|)$.
\subsection{A Global Learning Algorithm}\label{sec:sec:global}
Given a CPDAG, identifying definite non-causal and explicit causal relations is straightforward. To discriminate implicit causal relations from possible causal relations, we need an approach to find critical sets. The next proposition is particularly useful.
\begin{proposition}\label{prop:equivalent-critical-set}
For any two distinct vertices $X, Y$ in a CPDAG $\mathcal{G}^*$ such that $X$ is not an explicit cause of $Y$, it holds that
$\mathbf{C}_{XY}=\cup_{Z\in \mathbf{Z}}\mathbf{C}_{XZ}$,
where $\mathbf{C}_{UV}$ denotes the critical set of $U$ with respect to $V$, and $\mathbf{Z}$ is the set of ancestors of $Y$ in $\mathcal{G}^*$ which are also in the chain component containing $X$.
\end{proposition}
\begin{algorithm}[!t]
\caption{Finding the critical set of a given $X$ with respect to a set $\textbf{Z}$}
\label{algo:critical}
\begin{algorithmic}[1]
\REQUIRE
A chordal graph $\mathcal{G}^*_u$, a variable $X$ in $\mathcal{G}^*_u$, and a variable set $\textbf{Z}\neq \emptyset$ such that $X\notin \textbf{Z}$.
\ENSURE
$\textbf{C}$, which is the critical set of $X$ with respect to $\textbf{Z}$ in $\mathcal{G}^*_u$.
\STATE {Initialize $\textbf{C}=\emptyset$, a waiting queue $\mathcal{S}=[\;]$, and a set $\mathcal{H}=\emptyset$,}
\FOR {$\alpha \in adj(X)$}
\STATE {add $(\alpha, X, \alpha)$ to the end of $\cal S$,}
\ENDFOR
\WHILE {$\mathcal{S}$ is not empty}
\STATE {take the first element $( \alpha, \psi,\tau)$ out of $\cal S$ and add it to $\mathcal{H}$,}
\IF {$\tau\in \textbf{Z}$}
\STATE {add $\alpha$ to $\textbf{C}$, remove from $\mathcal{S}$ all triples where the first element is $\alpha$,}
\ELSE
\FOR {$\beta \in adj(\tau)$ and $\beta\notin adj(\psi)\cup\{\psi\}$ }
\IF { $(\alpha, \tau,\beta) \notin \mathcal{H}$ and $ ( \alpha, \tau,\beta) \notin \mathcal{S}$}
\STATE {add $( \alpha, \tau,\beta)$ to the end of $\cal S$,}
\ENDIF
\ENDFOR
\ENDIF
\ENDWHILE
\RETURN {$\textbf{C}$}
\end{algorithmic}
\end{algorithm}
Proposition \ref{prop:equivalent-critical-set} provides a factorization of the critical set of $X$ with respect to $Y$. For simplicity, we call $\cup_{Z\in \mathbf{Z}}\mathbf{C}_{XZ}$ the critical set of $X$ with respect to $\mathbf{Z}$. Algorithm \ref{algo:critical} shows how to find $\cup_{Z\in \mathbf{Z}}\mathbf{C}_{XZ}$ efficiently. Algorithm \ref{algo:critical} runs a breadth-first-search and returns the critical set of $X$ with respect to $\textbf{Z}$ in $\mathcal{G}^*_u$.
In Algorithm \ref{algo:critical}, we start from the siblings of $X$, then search chordless paths from the siblings until reaching some $Z_i\in \textbf{Z}$. Every chordless path starting from a sibling of $X$ is recorded in a queue $\cal S$ as a triple like $(\alpha, \psi, \tau)$, where $\alpha$ and $\tau$ are the start and the end points of the path, respectively, and $\psi$ is the sibling of $\tau$ on the path. If $\tau$ is a member of $\textbf{Z}$, we add $\alpha$ to the critical set $\textbf{C}$ and remove from $\mathcal{S}$ all triples where the first element is $\alpha$, that is, we stop enumerating chordless paths starting with $\alpha$. Otherwise, we extend the chordless path to the siblings of $\tau$ that are neither $\psi$ nor siblings of $\psi$ and add the corresponding triples to the queue $\cal S$. In this algorithm, a set of visited triples, $\cal H$, is introduced to speed up the search by avoiding visiting the same triple twice.
\begin{algorithm}[!t]
\caption{A global algorithm for identifying the type of causal relation (global ITC). }
\label{algo:global}
\begin{algorithmic}[1]
\REQUIRE
A CPDAG $\mathcal{G}^*$, a variable $X$ and a target $Y$ in $\mathcal{G}^*$.
\ENSURE
The type of causal relation between X and Y.
\IF {$X$ and $Y$ are connected by a path in $\mathcal{G}^*_u$}
\RETURN {$X$ is a \textbf{possible cause} of $Y$,}
\ENDIF
\STATE {let $\textbf{Z}=an(Y, \mathcal{G}^*)$,}
\IF {$X\in \textbf{Z}$}
\RETURN {$X$ is an \textbf{explicit cause} of $Y$,}
\ENDIF
\STATE {use Algorithm \ref{algo:critical} to find the critical set $\textbf{C}$ of $X$ with respect to $\textbf{Z}$ in $\mathcal{G}^*_u$,}
\IF {$|\textbf{C}|=0$}
\RETURN {$X$ is a \textbf{definite non-cause} of $Y$,}
\ENDIF
\IF {$\textbf{C}$ induces a complete subgraph of $\mathcal{G}^*_u$}
\RETURN {$X$ is a \textbf{possible cause} of $Y$,}
\ENDIF
\RETURN {$X$ is an \textbf{implicit cause} of $Y$.}
\end{algorithmic}
\end{algorithm}
Finally, we present a global learning approach for identifying types of causal relations in Algorithm \ref{algo:global}. Algorithm \ref{algo:global} is global in the sense that it takes an entire CPDAG as input. In Algorithm \ref{algo:global}, we first check whether $X$ and $Y$ are in the same chain component. If they are, $X$ is a possible cause of $Y$ based on Proposition \ref{prop:chain-comp}. Otherwise, we find the set of explicit causes of $Y$ and denote it by $\textbf{Z}$. This can be done by searching for the vertices that are connected to $Y$ in the directed subgraph of $\mathcal{G}^*$. If $X \in \textbf{Z}$, $X$ is an explicit cause of $Y$, otherwise, we find the critical set $\textbf{C}$ of $X$ with respect to $\textbf{Z}$. When $\textbf{C}=\emptyset$, we have that there are no explicit causes of $Y$ in the chain component containing $X$, so $X$ is not a cause of $Y$. Finally, using Theorem \ref{thm:implicit-cause-no-bk}, Algorithm \ref{algo:global} distinguishes between possible causes and implicit causes.
Since Algorithm \ref{algo:critical} does not visit the same triple like $(\alpha, \psi,\tau)$ twice, where $\alpha$ is a sibling of $X$ and $\tau$ is a sibling of $\psi$ in $\mathcal{G}^*_u$, the complexity of Algorithm \ref{algo:critical} in the worst case is $O(|sib(X, \mathcal{G}^*)| \cdot |\mathbf{E}(\mathcal{G}^*_u)|)$, where $|\mathbf{E}(\mathcal{G}^*_u)|$ is the number of edges in $\mathcal{G}^*_u$. Now we consider the computational complexity of global ITC (Algorithm \ref{algo:global}). We know that the complexity to check the undirected connectivity of $X$ and $Y$ or to find $an(Y, \mathcal{G}^*)$ is $O(|\mathbf{E}(\mathcal{G}^*)|^2)$, where $|\mathbf{E}(\mathcal{G}^*)|$ is the number of vertices in ${\cal G}^*$. Consequently, the complexity of global ITC is $O(|\mathbf{E}(\mathcal{G}^*)|^2+|sib(X, \mathcal{G}^*)| \cdot |\mathbf{E}(\mathcal{G}^*_u)|)$. Clearly, the worst case is $O(|\mathbf{E}(\mathcal{G}^*)|^3)$.
\subsection{Causal-Effect-Based Methods}\label{sec:sec:ce}
\begin{algorithm}[!t]
\caption{The framework of causal effect testing based algorithms.}
\label{algo:ce-based}
\begin{algorithmic}[1]
\REQUIRE
A treatment $X$, a target $Y$, a CPDAG $\mathcal{G}^*$ over a vertex set $\bf V$, and a significance level $\alpha$.
\ENSURE
The type of causal relation between X and Y.
\STATE{set $\Theta= [\;]$ and ${\rm P}_{\rm val}= [\;]$,}
\FOR{each $\mathbf{S}\subset \mathbf{V}$ such that $\mathbf{S}$ is an adjustment set for $(X, Y)$ in some DAG in $[\mathcal{G}^*]$}
\STATE{estimate the causal effect $\theta$ of $X$ on $Y$ by adjusting for $\mathbf{S}$, and add the causal effect to $\Theta$,}
\STATE{test the null hypothesis $\theta=0$ and add the corresponding p-value to ${\rm P}_{\rm val}$,}
\ENDFOR
\IF {every p-value in ${\rm P}_{\rm val}$ is less than or equal to $\alpha$}
\RETURN {$X$ is a \textbf{definite cause} of $Y$,}
\ENDIF
\IF {every p-value in ${\rm P}_{\rm val}$ is greater than $\alpha$}
\RETURN {$X$ is a \textbf{definite non-cause} of $Y$,}
\ENDIF
\RETURN {$X$ is a \textbf{possible cause} of $Y$.}
\end{algorithmic}
\end{algorithm}
We now discuss the causal-effect-based methods, which are modifications of the IDA-type algorithms. For simplicity, we assume that the observed variables follow a linear-Gaussian structural equation model and that the observational distribution is faithful to the underlying DAG. With these assumptions, $X$ has a non-zero total causal effect on $Y$ if and only if there is a directed path from $X$ to $Y$ in the underlying DAG. Following the work of~\citet{maathuis2009estimating}, we use
\[ACE(Y\mid do(X=x)) \coloneqq \frac{\partial E(Y\mid do(X=x))}{\partial x},\]
to measure the (average) total causal effect of $X$ on $Y$.\footnote{Here, $do(X=x)$ is the \emph{do-operator} proposed by \citet{pearl2009causality} to denote the intervention on $X$ by forcing $X$ to be $x$. \citet{pearl2009causality} defined that $X$ has a causal effect on $Y$ if there exists an $x\neq x'$ such that $P(Y\mid do(X=x))\neq P(Y\mid do(X=x'))$, where $P(Y\mid do(X=x))$ is the post-intervention distribution of $Y$. On the other hand, it is common to summarize $P(Y\mid do(X=x))$ by its mean \citep{pearl2009causality, maathuis2009estimating}, i.e., the mean of $Y$ w.r.t. $P(Y\mid do(X=x))$, which is denoted by $E(Y\mid do(X=x))$.} As mentioned in the introduction, the idea of the causal-effect-based methods is to estimate all possible causal effects of the treatment on the target first, and then check whether the possible effects are all zeros or non-zeros. If all of the possible effects are evaluated as zeros (non-zeros), then the treatment is a definite non-cause (definite cause) of the target.
However, due to estimation error, an estimated effect may not be exactly zero. In the work of~\citet{maathuis2010nature}, the authors first estimated all possible effects for all pairs of treatment and target, and then summarized each set of possible causal effects by its minimum absolute value. Finally, the minimum values were sorted in descending order. The top ones were evaluated as relatively strong effects. Though this method has been widely applied to real-world problems, it requires to estimate all possible effects for all pairs of variables, which may bring unnecessary costs if someone only wants to know the causal relation between one pair of treatment and target. Moreover, in this method, the order of an effect depends on the other effects. Thus, this method is suitable for comparing the magnitude of causal effects, rather than identifying the causal relation of a given pair.
In this paper, we focus on a testing-based solution, whose framework is summarized by Algorithm \ref{algo:ce-based}. After initializing two sequences $\Theta$ and ${\rm P}_{\rm val}$, Algorithm~\ref{algo:ce-based} enumerates all possible causal effects of $X$ on $Y$ and tests the null hypothesis $\theta=0$ for each estimated effect $\theta$. Different modifications adopt different enumeration and testing strategies. We introduce four modifications below.
\begin{enumerate}
\item [(M1)] IDA {\rm +} testing all enumerated effects. Following the original IDA framework~\citep{maathuis2009estimating}, this modification enumerates all possible effects by listing all possible parental sets of the treatment $X$. Thus, line 2 of Algorithm~\ref{algo:ce-based} is replaced by
{\vspace{2mm} \it ``for each $\mathbf{Q}\subset sib(X, {\cal G}^*)$ such that orienting $\mathbf{Q}\to X$ and $X\rightarrow sib(X, \mathcal{G}^{*})\setminus \textbf{Q}$ does not introduce any v-structure collided on $X$, let $\mathbf{S}=\mathbf{Q}\cup pa(X, {\cal G}^*)$ and do ..." \vspace{2mm}}
All enumerated effects are then tested according to line 4 of Algorithm~\ref{algo:ce-based}. In the linear-Gaussian case, estimating the causal effect of $X$ on $Y$ by adjusting for $\mathbf{S}$ is equivalent to estimating the coefficient of $X$ in the linear regression of $Y$ on $X$ and $\mathbf{S}$. Hence, a t-test for the coefficient of $X$ is used to test the significance of the causal effect of $X$ on $Y$.
\item [(M2)] IDA {\rm +} testing the minimum and maximum absolute enumerated effects. This modification is inspired by the work of~\citet{maathuis2010nature}. It first enumerates all possible effects by listing all possible parental sets of the treatment $X$. Then, it tests the effects of $X$ on $Y$ with the minimum and maximum absolute values to obtain two p-values, $p_{\rm min}$ and $p_{\rm max}$, respectively. Consequently, if $p_{\rm min} \leq \alpha$, it returns that X is a definite cause of Y, if $p_{\rm max} > \alpha$, it returns that X is a definite non-cause of Y, and otherwise returns that X is a possible cause of Y. The details of this modification is provided in \ref{app:ce}.
\item [(M3)] IDA {\rm +} utilizing non-ancestral relations {\rm +} testing all enumerated effects. This modification first lists all possible parental sets of the treatment $X$. Then, to reduce the number of estimations and significance tests, it checks whether $X$ is a non-ancestor of $Y$ before estimating the causal effects of $X$ on $Y$. More formally, we insert a step between lines 2 and 3 of Algorithm~\ref{algo:ce-based} as follows.
{\vspace{2mm} \it ``Orient $\mathbf{Q}\to X$ and $X\rightarrow sib(X, \mathcal{G}^{*})\setminus \textbf{Q}$, complete the orientations using Meek's rules~\citep{meek1995causal}, and use Lemma 3.2 in~\cite{perkovic2017interpreting} to check whether $X$ is a b-possible ancestor of $Y$." \vspace{2mm}}
The definition of b-possible ancestor can be found in~\citet[Definition~3.3]{perkovic2017interpreting}. We directly set $\theta=0$ and the p-value $p=1$ if $X$ is not a b-possible ancestor of $Y$. Otherwise, we estimate the causal effect of $X$ on $Y$ by adjusting for $\mathbf{S}=\mathbf{Q}\cup pa(X, {\cal G}^*)$, and test the significance of the estimated effect. The details are provided in~\ref{app:ce}.
\item [(M4)] IDA {\rm +} utilizing non-ancestral relations {\rm +} testing the minimum and maximum absolute enumerated effects. This modification takes the same enumeration strategy used by the third modification, and uses the same testing strategy as the second method does.
\end{enumerate}
Except for the second and forth modifications, the other two modifications usually compute a list of p-values. These p-values are compared with a given significance level $\alpha$, as shown in lines $6$-$12$ of Algorithm~\ref{algo:ce-based}. The adjustment methods for multiple p-values, such as the Bonferroni correction, may be used to control the false discovery rate.
The input CPDAG of Algorithm \ref{algo:ce-based} can be replaced by the induced subgraph over $pa(X, {\cal G}^*) \cup sib(X, {\cal G}^*)$ for the first two modifications.
Since the induced subgraph over $pa(X, {\cal G}^*) \cup sib(X, {\cal G}^*)$ can be learned locally using the variant of MB-by-MB \citep[Algorithm~3]{liu2020local}, we can combine the first two modifications with the variant of MB-by-MB to make them fully local. However, for the last two modifications, the input CPDAG cannot be replaced. The reason is that these two modifications need to run Meek's rules to extend the local orientations $\mathbf{Q}\to X$ and $X\rightarrow sib(X, \mathcal{G}^{*})\setminus \textbf{Q}$, and Meek's rules require an entire CPDAG.
We remark that, the causal-effect-based methods are not restricted to the aforementioned four modifications. For example, in the linear-Gaussian case, one can use the optimal IDA~\citep{Witte2020efficient} to replace the original IDA. For each $\mathbf{Q}\subset sib(X, {\cal G}^*)$ such that orienting $\mathbf{Q}\to X$ and $X\rightarrow sib(X, \mathcal{G}^{*})\setminus \textbf{Q}$ does not introduce any v-structure collided on $X$, the optimal IDA first runs Meek's rules to extend the local orientations $\mathbf{Q}\to X$ and $X\rightarrow sib(X, \mathcal{G}^{*})\setminus \textbf{Q}$, and then finds the optimal adjustment set so that the estimation of the causal effect has the smallest asymptotic variance. Another modification, which we call the hybrid method, is to infer whether $X$ is a definite non-cause of $Y$ by checking whether $X$ has a partially directed path to $Y$ in the input CPDAG~\citep{zhang2006, perkovic2017interpreting} first, and then call a causal-effect-based method if $X$ is not a definite non-cause of $Y$. Compared with the third and forth modifications, this hybrid method utilizes non-ancestral relations before listing all possible parental sets of $X$, and therefore, is
generally more efficient if $X$ is graphically identified as a definite non-cause of $Y$.
To end this section, we theoretically compare the proposed local algorithm to the causal-effect-based methods in terms of computational complexity. In the worst case, the number of causal effect estimations required by a causal-effect-based method is $2^{|sib(X, {\cal G}^*)|}$, since every causal-effect-based method needs to enumerate the possible causal effects of $X$ on $Y$. At the same time, at most $2^{|sib(X, {\cal G}^*)|}$ tests on these estimated causal effects are required in the worst case.
Besides, the modifications that utilize non-ancestral relations or the optimal IDA have to run $2^{|sib(X, {\cal G}^*)|}$ times Meek's rules, while the complexity of Meek's rules is polynomial to the number of vertices in the graph. Therefore, the proposed local method (Algorithm \ref{algo:local}) is more efficient than the current causal-effect-based methods as the former only needs linearly many hypothesis tests.
\section{ Experiments}\label{sec:simulation}
In this section, we illustrate and evaluate the proposed methods experimentally using synthetic data sets generated from linear structural equation models with Erd\"os-R\'enyi random DAGs and the DREAM4 data sets. We compare the local ITC with the global one as well as the four modifications of causal-effect-based methods (CE-based for short).
The details of the CE-based methods are provided in Section~\ref{sec:sec:ce} and \ref{app:ce}. In this section, the four modifications of CE-based methods from M1 to M4 are denoted by ``IDA + test (all)", ``IDA + test (min/max)", ``IDA + an + test (all)", and `IDA + an + test (min/max)", respectively. We use the Bonferroni correction to adjust p-values for multiple comparisons, and the corresponding methods are denoted by ``multi". In addition, the CE-based methods with the optimal IDA and the hybrid method mentioned in Section~\ref{sec:sec:ce} are also studied experimentally in~\ref{app:app:opt} and \ref{app:app:hybrid} respectively.
In Section \ref{sec:sec:true}, we assume that the true CPDAG or its local structure of interest is available. In this case, the synthetic data sets are only used by the CE-based methods to estimate causal effects, and by the local ITC method to perform conditional independence tests.
In Section \ref{sec:sec:est}, we further evaluate the methods based on the structures learned from data. Three global structure learning algorithms, including the PC algorithm~\citep{spirtes1991algorithm}, the stable PC algorithm~\citep{colombo2014order} and the GES algorithm~\citep{chickering2002learning}, are used to learn CPDAGs, and the variant of MB-by-MB~\citep{liu2020local} is used to learn parents and siblings of the vertices of interest. In all of these experiments, algorithms like PC, stable PC, GES and IDA are called from \texttt{R-package pcalg}~\citep{Kalisch2012pcalg}, and the Bonferroni correction is called from \texttt{R-package stats}. The significance level $\alpha$ of statistical independence tests is $0.001$.\footnote{Experiments show different significance levels give similar results.} All codes were run on a computer with an Intel 2.5GHz CPU and 8 GB of memory.
Let $\text{ER}(n,d)$ denote a random DAG with $n$ vertices and average in-and-out degree $d$. In our experiments, $n$ is chosen from $\{50, 100\}$ and $d$ is chosen from $\{1.5, 2.0, 2.5, 3.0, 3.5, 4.0\}$. For a sampled $\text{ER}(n,d)$ graph, we drew an edge weight $\beta_{ij}$ from a Uniform($[0.8, 1.6]$) or a Uniform($[-1.6, -0.8]\cup[0.8, 1.6]$) distribution for each directed edge $X_i\rightarrow X_j$ in the DAG. Then, we constructed a linear structural equation model as follows,
\begin{align}
X_j = \sum_{X_i \in pa(X_j)}\beta_{ij}X_i+\epsilon_j\;, \quad j = 1,...,n\;,
\end{align}
where $\epsilon_1, ..., \epsilon_n$ are independent $\mathcal{N}(0,1)$ noises.
For each combination of $n$, $d$ and the distribution of edge weights, we generated $5,000$ weighted DAGs. Finally, in Section~\ref{sec:sec:true} we drew $N_{\rm effect}\in \{50, 100, 150\}$ samples from this linear model to estimate causal effects and perform conditional independence tests, and in Section~\ref{sec:sec:est} we drew additional $N_{\rm graph}\in \{100, 200, 500\}$ samples to learn the required causal structures. In summary, there were totally $2\times 6 \times 2=24$ graph settings and $2\times 6\times 2\times 3\times 3=216$ experiment parameter settings, and for each experiment parameter setting, we repeated the experiment $5,000$ times.
Given a sampled DAG, we randomly drew a treatment variable and a target variable, and compared their causal relation (definite cause, definite non-cause, or possible cause) learned from data with the true one read from the corresponding CPDAG of the sampled DAG. The Kappa coefficient~\citep{Jacob1960kappa} as well as the true positive rate (TPR) and the false positive rate (FPR) were used to measure the performance of each method. The Kappa coefficient is an adjustment of accuracy rate. It is the proportion of agreement after chance agreement is removed from consideration~\citep{Jacob1960kappa}. The formal definition of the Kappa coefficient is given as follows. For an experiment parameter setting, let $M_{ij}$ be the number of experiments in which the $i$th causal relation is identified as the $j$th causal relation. The Kappa coefficient $\kappa$ is defined as
\[\kappa = \frac{p - q}{1-q},\]
where
\[p = \frac{\sum_{i=1}^{3} M_{ii}}{\sum_{i,j=1}^{3} M_{ij}}, \qquad q = \frac{\sum_{i=1}^{3} (\sum_{j=1}^3 M_{ij})\cdot(\sum_{j=1}^3 M_{ji}) }{(\sum_{i,j=1}^{3} M_{ij})^2}.\]
The Kappa coefficient ranges from $-1$ to $+1$, and the higher the value of Kappa, the better the evaluated method. For ease of visualization, we focus on the kappa coefficients in this section, and the TPRs and FPRs are reported in~\ref{app:app:detailed}.
\subsection{Learning with True Graphs}\label{sec:sec:true}
\begin{figure}[t!]
\centering
\subfigure{
\begin{minipage}[t]{0.9\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/legend.pdf}
\end{minipage}%
}%
\vspace{-1em}
\addtocounter{subfigure}{-1}
\subfigure[$n=50$, $N_{\rm effect}=50$ \label{fig:true:kappa_50_50}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/kappa_nodes_50_wt_positive_0_50.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=50$, $N_{\rm effect}=100$ \label{fig:true:kappa_50_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/kappa_nodes_50_wt_positive_0_100.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=50$, $N_{\rm effect}=150$ \label{fig:true:kappa_50_150}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/kappa_nodes_50_wt_positive_0_150.pdf}
\end{minipage}%
}%
\subfigure[$n=100$, $N_{\rm effect}=50$ \label{fig:true:kappa_100_50}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/kappa_nodes_100_wt_positive_0_50.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=100$, $N_{\rm effect}=100$ \label{fig:true:kappa_100_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/kappa_nodes_100_wt_positive_0_100.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=100$, $N_{\rm effect}=150$ \label{fig:true:kappa_100_150}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/kappa_nodes_100_wt_positive_0_150.pdf}
\end{minipage}%
}%
\caption{The Kappa coefficients of different methods on random graphs with positive weights. The true graph structures are provided. }
\label{fig:true:kappa}
\end{figure}
\begin{figure}[t!]
\centering
\subfigure{
\begin{minipage}[t]{1\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision/legend.pdf}
\end{minipage}%
}%
\vspace{-1em}
\addtocounter{subfigure}{-1}
\subfigure[$n=50$, $N_{\rm effect}=50$ \label{fig:true:time_50_50}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision/total_time_nodes_50_wt_positive_0_50revised.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=50$, $N_{\rm effect}=100$ \label{fig:true:time_50_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision/total_time_nodes_50_wt_positive_0_100revised.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=50$, $N_{\rm effect}=150$ \label{fig:true:time_50_150}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision/total_time_nodes_50_wt_positive_0_150revised.pdf}
\end{minipage}%
}%
\subfigure[$n=100$, $N_{\rm effect}=50$ \label{fig:true:time_100_50}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision/total_time_nodes_100_wt_positive_0_50revised.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=100$, $N_{\rm effect}=100$ \label{fig:true:time_100_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision/total_time_nodes_100_wt_positive_0_100revised.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=100$, $N_{\rm effect}=150$ \label{fig:true:time_100_150}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision/total_time_nodes_100_wt_positive_0_150revised.pdf}
\end{minipage}%
}%
\caption{The CPU time (in seconds) of different methods on random graphs with positive weights. The true graph structures are provided. The CPU time of ``IDA + an + test (all)" and ``IDA + an + test (min/max)" is not shown, as they are more than 50-100 times slower than the other methods.}
\label{fig:true:time}
\end{figure}
In this section, the true CPDAGs or their local structures are provided to exclude estimation biases caused by graph structure learning from data. In this case, the global ITC shown in Algorithm \ref{algo:global} can identify all causal relations correctly since the input CPDAG is true. Except for the global ITC, the local ITC and the CE-based methods need to perform hypothesis tests, which may introduce errors. To assess these methods, we run experiments on data with positive weights (Uniform($[0.8, 1.6]$)) as well as a mixture of negative and positive weights (Uniform($[-1.6, -0.8]\cup[0.8, 1.6]$)). The results on the graphs with positive weights are shown in this section and the rests are in~\ref{app:app:mix}.
\begin{figure}[!h]
\centering
\subfigure{
\begin{minipage}[t]{1\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision/legend.pdf}
\end{minipage}%
}%
\vspace{-1em}
\addtocounter{subfigure}{-1}
\subfigure[$n=50$, $N_{\rm effect}=50$ \label{fig:true:time_zoom_50_50}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/total_time_zoom_nodes_50_wt_positive_0_50.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=50$, $N_{\rm effect}=100$ \label{fig:true:time_zoom_50_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/total_time_zoom_nodes_50_wt_positive_0_100.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=50$, $N_{\rm effect}=150$ \label{fig:true:time_zoom_50_150}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/total_time_zoom_nodes_50_wt_positive_0_150.pdf}
\end{minipage}%
}%
\subfigure[$n=100$, $N_{\rm effect}=50$ \label{fig:true:time_zoom_100_50}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/total_time_zoom_nodes_100_wt_positive_0_50.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=100$, $N_{\rm effect}=100$ \label{fig:true:time_zoom_100_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/total_time_zoom_nodes_100_wt_positive_0_100.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=100$, $N_{\rm effect}=150$ \label{fig:true:time_zoom_100_150}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/true_graph_res/total_time_zoom_nodes_100_wt_positive_0_150.pdf}
\end{minipage}%
}%
\caption{The ratio of the CPU time of different methods to that of the local ITC on random graphs with positive weights. The true graph structures are provided. The results of ``IDA + an + test (all)" and ``IDA + an + test (min/max)" are not shown, as they are more than 50-100 times slower than the other methods. }
\label{fig:true:zoom}
\end{figure}
Figure~\ref{fig:true:kappa} shows the Kappa coefficients of the proposed local method and five CE-based methods on random graphs with positive weights. We can see that the proposed local ITC is significantly better than the CE-based methods, especially when the sample size is small and the average degree is high. Increasing the sample size can improve the performance of all methods, and reduce the difference between the local ITC and the CE-based methods.
In these experiments, probably because the total number of hypothesis tests is not large, adjusting p-values for multiple comparisons does not bring much improvement. Besides, testing all enumerated effects usually performs better than testing the minimum and maximum absolute effects. Moreover, utilizing non-ancestral information can improve the performance. Consequently, the CE-based-method denoted by ``IDA + an + test (all)" performs the best among the five CE-based-methods.
However, utilizing non-ancestral information will significantly increase the computational time because of the use of Meek's rules. In our experiments, the CE-based methods which utilize non-ancestral relations are 50-100 times slower than the others. To compare the other methods, Figure~\ref{fig:true:time} reports the CPU time (in seconds) and Figure~\ref{fig:true:zoom} further shows the ratio of the time used by each CE-based method to the local ITC. As one can see from the figures, the CE-based methods without using non-ancestral relations are 2-4 times slower than the local ITC.
Benefiting from fewer hypothesis tests, the local ITC algorithm is more stable, more accurate, and more efficient than the CE-based methods. Additional evidence also comes from the experiments on models with mixed edge weights. In these experiments, although all Kappa coefficients drop, the Kappa coefficients of the local ITC drop less than those of the CE-based methods. The details can be found in~\ref{app:app:mix}.
\subsection{Learning with Estimated Graphs}\label{sec:sec:est}
\begin{figure}[t!]
\centering
\subfigure{
\begin{minipage}[t]{1\textwidth}
\centering
\includegraphics[width=1\textwidth]{figs/revision_learn/legend.pdf}
\end{minipage}%
}%
\vspace{-1.2em}
\addtocounter{subfigure}{-1}
\subfigure[$n=50$, $N=(100,100)$ \label{fig:learned:kappa_50_100_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision_learn/kappa_nodes_50_wt_positive_100_100revised.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=50$, $N=(200,100)$ \label{fig:learned:kappa_50_200_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision_learn/kappa_nodes_50_wt_positive_200_100revised.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=50$, $N=(500,150)$ \label{fig:learned:kappa_50_500_150}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision_learn/kappa_nodes_50_wt_positive_500_150revised.pdf}
\end{minipage}%
}%
\subfigure[$n=100, N=(100,100)$ \label{fig:learned:kappa_100_50}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision_learn/kappa_nodes_100_wt_positive_100_100revised.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=100,N=(200,100)$ \label{fig:learned:kappa_100_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision_learn/kappa_nodes_100_wt_positive_200_100revised.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n = 100$, $N=(500,150)$ \label{fig:learned:kappa_100_150}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/revision_learn/kappa_nodes_100_wt_positive_500_150revised.pdf}
\end{minipage}%
}%
\caption{The Kappa coefficients of different methods on random graphs with positive weights. The graph structures are learned from data. $N=(N_{\rm graph}, N_{\rm effect})$ denotes the sample sizes for learning graphs and estimating causal effects.
}
\label{fig:learn:kappa}
\end{figure}
In this section, we further study experimentally our proposed methods when the true causal structures are not available. We used the variant of MB-by-MB~\citep{liu2020local} to learn the parents and siblings of the vertices of interest (denoted by ``local +"), and used the PC algorithm, the stable PC algorithm (denoted by ``PCS +") and GES to learn entire CPDAGs. The learned structures are then passed to the local ITC, the global ITC and the CE-based methods. For ease of presentation, we mainly report twelve methods in this section, but the conclusions obtained coincide with all the experiments.
\begin{figure}[t!]
\centering
\subfigure{
\begin{minipage}[t]{1\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/learned_graph_res/legend.pdf}
\end{minipage}%
}%
\vspace{-1em}
\addtocounter{subfigure}{-1}
\subfigure[$n=50$, $N=(100,100)$ \label{fig:learned:time_50_100_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/learned_graph_res/total_time_nodes_50_wt_positive_100_100.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=50$, $N=(200,100)$ \label{fig:learned:time_50_200_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/learned_graph_res/total_time_nodes_50_wt_positive_200_100.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=50$, $N=(500,150)$ \label{fig:learned:time_50_500_150}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/learned_graph_res/total_time_nodes_50_wt_positive_500_150.pdf}
\end{minipage}%
}%
\subfigure[$n=100,N=(100,100)$ \label{fig:learned:time_100_50}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/learned_graph_res/total_time_nodes_100_wt_positive_100_100.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=100$, $N=(200,100)$ \label{fig:learned:time_100_100}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/learned_graph_res/total_time_nodes_100_wt_positive_200_100.pdf}
\end{minipage}%
}%
\hspace{0.01\textwidth}
\subfigure[$n=100$, $N=(500,150)$ \label{fig:learned:time_100_150}]{
\begin{minipage}[t]{0.3\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/learned_graph_res/total_time_nodes_100_wt_positive_500_150.pdf}
\end{minipage}%
}%
\caption{The total CPU time (in seconds) of different methods on random graphs with positive weights. The graph structures are learned from data. $N=(N_{\rm graph}, N_{\rm effect})$ denotes the sample sizes for learning graphs and estimating causal effects.}
\label{fig:learn:time}
\end{figure}
Figure \ref{fig:learn:kappa} shows the Kappa coefficients based on 50- and 100-node graphs. For ease of presentation, we omit the results of the global ITC combined with PCS and GES as well as the results of the CE-based methods combined with PCS and GES, since the PCS-based methods perform similarly to PC and the GES-based methods do not perform well (see~\ref{app:app:detailed} for the detailed TPRs and FPRs).
As one can see, the proposed local ITC outperforms the other methods in almost all settings, especially when the sample size is small. The CE-based methods combined with the variant of MB-by-MB perform slightly worse than the local ITC. The global ITC combined with PC is also competitive when the sample size is large. When $N_{\rm graph}\geq 200$, the global ITC combined with PC outperforms the corresponding CE-based methods. Besides, the CE-based methods that use non-ancestral relations are usually better than the other CE-based methods, as they take the advantage of the correctly learned global graphical structure. Moreover, testing all enumerated effects performs similarly to testing the minimum and maximum absolute effects, rather than outperforms the latter as shown in Section~\ref{sec:sec:true}.
We next compare the total computational time of different methods.
As shown in Figure \ref{fig:learn:time},
the local ITC and the local versions of the CE-based methods are more efficient than the global ones. For the methods using the same structure learning algorithm, the ITC methods are more efficient than the CE-based methods that take much more computational time to identify non-ancestral relations, and are slightly more efficient than the other CE-based methods because the structure learning generally dominates the computational time of these methods.
We also compare the time spend in identifying types of causal relations (excluding the time for structure learning from the total time), and the results are similar to those shown in Figure~\ref{fig:true:zoom}. Using non-ancestral information in CE-based methods usually makes them 40-80 times slower than the local ITC,
while the other CE-based methods without using non-ancestral relations are almost $2$ times slower than the local ITC.
The experiments of the CE-based methods using the optimal IDA and the hybrid method can be found in \ref{app:app:opt} and in~\ref{app:app:hybrid} respectively. Briefly, these experiments show that the CE-based methods with the optimal IDA are usually better than the methods with the original IDA, but do not outperform the global ITC, and the hybrid method is slightly better than the non-hybrid CE-based methods that use non-ancestral relations
\subsection{An Application to the DREAM4 Data Sets}\label{sec:sec:d4}
In this section, we apply our method to the synthetic gene expression data sets from the
DREAM4 \emph{in silico} challenge, to show the potential of our method for supporting causal inference. A detailed description of the data sets can be found at \texttt{https://dreamchallenges.org/dream-4-in-silico-\\network-challenge/}. In this study, we focus on $5$ data sets provided by the DREAM4 challenge, each of which contains a gene regulatory network (possibly cyclic) with 100 genes, observational gene expression data with 310 observations and interventional gene expression data.
The used $5$ data sets, including the true network structures, can be obtained from the \texttt{R-package DREAM4}.\footnote{The $5$ data sets are named by ``dream4\_100\_01" to ``dream4\_100\_05" in the \texttt{R-package DREAM4}.}
We normalized each data set such that each gene has a sample mean 0 and a sample variance 1. The marginal distributions of the variables are approximately Gaussian. Following ~\citet{maathuis2010nature}, we assume that the multivariate Gaussianity holds for all variables.
To evaluate the ``true" relationship for each pair of $X$ and $Y$, we first estimate the causal effect of $X$ on $Y$ using the back-door adjustment, based on the observational data and the true network structure. Then, we use the t-test to decide the significance of the estimated causal effect at the significance level $\alpha=0.001$. All pairs of $X$ and $Y$ whose corresponding p-values of the t-tests are less than or equal to $\alpha$ are regarded as ``true" causal pairs and constitute the target set.
Assuming that the true gene network is unknown, we next use
the following three methods to identify the type of causal relation for each treatment-target pair $(X, Y)$.
\begin{itemize}
\item {Method 1.} Using the PC algorithm to estimate a CPDAG first, and then calling the global ITC (Algorithm~\ref{algo:global}) to identify the type of causal relation for each $(X, Y)$.
\item Method 2. Using the PC algorithm to estimate a CPDAG first, and then calling the local ITC (Algorithm~\ref{algo:local}) to identify the type of causal relation for each $(X, Y)$. The required local structures, i.e. $pa(X, {\cal G}^*)$ and the induced subgraph over $sib(X, {\cal G}^*)$, are read from the learned CPDAG.
\item Method 3. Using the variant of MB-by-MB to estimate $pa(X, {\cal G}^*)$ and the induced subgraph over $sib(X, {\cal G}^*)$ for each $X$, and then identifying the type of causal relation for each $(X, Y)$ by using the local ITC (Algorithm~\ref{algo:local}).
\end{itemize}
For each variable pair $(X, Y)$,
we further use IDA to estimate all possible causal effects of $X$ on $Y$. To build a sequence of variable pairs based on the magnitude of the causal effects, we first rank (in descending order) the pairs of treatment and target whose corresponding causal relations are definite causal according to their minimum absolute effects. Then, we rank (in descending order) the pairs whose corresponding causal relations are definite non-causal and possible causal according to their maximum absolute effects. Finally, we append the ordered sequence of definite non-causal and possible causal relation pairs to the ordered sequence of definite causal relation pairs, and select top $q$ pairs as the predicted pairs.
Note that compared to the work of~\citet{maathuis2010nature}, this method has two differences. First, we rank definite non-causal and possible causal pairs by their maximum absolute effects, while \citet{maathuis2010nature} rank all pairs by their minimum absolute effects. This is because that the minimum absolute effect of a definite non-causal or possible causal pair should be zero, while the maximum absolute effect is more informative since it measures the upper bound on the true causal effect. Second, we rank the definite causal pairs before the other pairs regardless of their possible estimated causal effects.
We compare Methods 1 and 2 to the IDA algorithm combined with PC, and compare Method 3 to the IDA algorithm combined with the variant of MB-by-MB. All significance levels used in these methods are set to be 0.001, aligned with those used in the simulation
studies. For each method, we compare the predicted pairs to the target pairs for different $q$'s and compute the area under the receiver operating characteristic curve (AUC). Tables~\ref{tab:d4-0.001} shows the results. For each data set and each method, we also perform DeLong's test to test whether the AUC of the method is different from the AUC of the IDA algorithm~\citep{delong1988}, using the \texttt{R function roc.test} implemented in \texttt{R-package pROC}. The null hypothesis is that the difference in AUC is equal to 0. The p-values of the tests are reported in parentheses.
\begin{table}[!t]
\centering
\scalebox{0.72}{%
\begin{tabular}{@{}lccccc@{}}
\toprule
Methods & dream4\_100\_01 & dream4\_100\_02 & dream4\_100\_03 & dream4\_100\_04 & dream4\_100\_05 \\ \midrule
\multirow{2}{*}{PC + IDA} & \multirow{2}{*}{0.6578} & \multirow{2}{*}{0.6870} & \multirow{2}{*}{0.6894} & \multirow{2}{*}{0.6921} & \multirow{2}{*}{0.6823} \\
& & & & & \\
\multirow{2}{*}{Method 1 + IDA} & 0.6439 & \textbf{0.6905} & \textbf{0.6930} & 0.6808 & \textbf{0.6898} \\
& (0.0000) & (0.2809) & (0.1049) & (0.0000) & (0.0050) \\
\multirow{2}{*}{Method 2 + IDA} & 0.6481 & \textbf{0.6972} & \textbf{0.6956} & \textbf{0.6945} & \textbf{0.6916} \\
& (0.0000) & (0.0000) & (0.0000) & (0.0151) & (0.0000) \\ \midrule
\multirow{2}{*}{local + IDA} & \multirow{2}{*}{0.6624} & \multirow{2}{*}{0.6354} & \multirow{2}{*}{0.7000} & \multirow{2}{*}{0.6750} & \multirow{2}{*}{0.6696} \\
& & & & & \\
\multirow{2}{*}{Method 3 + IDA} & \textbf{0.6672} & 0.6349 & \textbf{0.7019} & \textbf{0.6769} & \textbf{0.6776} \\
& (0.0002) & (0.2000) & (0.0890) & (0.0203) & (0.0000) \\
\bottomrule
\end{tabular}%
}
\caption{ AUC of different methods on DREAM4 data sets. The p-values of DeLong's tests are reported in parentheses, which test whether the AUC of a proposed method is significantly different from the AUC of the ``PC + IDA" or ``local + IDA" algorithm.}
\label{tab:d4-0.001}
\end{table}
Table~\ref{tab:d4-0.001} displays that the modified versions of IDA with Methods 1 to 3 outperform the original IDA in 3, 4 and 4 data sets, respectively, and 9 of them are significant at the level $0.1$. As a result, our proposed methods, especially the local ITC (Methods 2 and 3), can improve the performance of the IDA algorithm when predicting the magnitude of a causal effect.
We remark that the above results on the DREAM4 data sets are proof-of-concept and show that identifying types of causal relations do have the potential to support causal inference. Apart from the listed Methods 1, 2 and 3, the practitioners may develop their own specific modifications. Of course, as discussed in~\citet{maathuis2010nature}, great care should be taken in real applications when the underlying assumptions, such as the multivariate Gaussianity and the faithfulness, are violated. Nevertheless, we hope that the example given in this section could motivate more studies on the use of the local and global ITC in observational studies.
\section{Concluding Remarks}\label{sec: conclusion}
In this paper, we present a local method for identifying types of causal relations without evaluating causal effects and learning a global causal structure. A sufficient and necessary graphical condition is provided to check the existence of a causal path from a treatment to a target based on a CPDAG. We also study the graphical properties of each type of causal relation. Inspired by these properties, we further propose a local identification criterion for each type of causal relation, which depends only on the induced subgraph of the true CPDAG over the adjacent variables of the treatment as well as some queries about d-separation relations. The local criteria naturally lead to a local learning algorithm for identifying types of causal relations if one assumes that the faithfulness condition holds. Experimental studies empirically prove that the proposed local algorithm performs well.
Our work introduces the local characterizations of types of causal relations, which are helpful for understanding causal relations hidden behind observational data. Except for the theoretical contributions, our results have many potential applications as well. Firstly, as mentioned in the introduction, some real-world problems, such as fault analysis in telecommunication networks and online product recommendation, qualitative analysis is enough for making decisions.
Secondly, even in quantitative analysis, when the causal effect is not uniquely identifiable due to Markov equivalence, we may also use the proposed methods to check whether the bounds on a causal effect cover zero. For example, as shown in the experiments, our methods can be used to modify the current IDA-type algorithms to predict which interventions are likely to have a strong effect, as mentioned by~\citet{maathuis2010nature}. Thirdly, the proposed local method can be combined with the IDA algorithm to reduce the computational costs. For instance, if a treatment is a non-cause of a target, then without any computation we can conclude that all possible effects are zeros \citep{maathuis2009estimating}. Compared to the existing global method that depends on the global structure of an input CPDAG, the proposed local methods are more effective and efficient, especially when we are only interested in the causal effect of one treatment on one target, not the causal effects of all treatments on all targets.
{Finally, \citet{Shi2021mediation} provided a method to find all ``mediators" lying on at least one directed path from a given treatment to a given target, assuming that the underlying DAG is identifiable from data. When the underlying DAG is not identifiable but a CPDAG is identifiable, our proposed methods are potentially useful for finding ``definite mediators", which are not only the definite effects of the treatment but also the definite causes of the target. These variables must lie on at least one directed path from the treatment to the target, no matter which equivalent DAG is the true one.}
Our results can be easily extended to interventional essential graphs \citep{he2008active,hauser2012characterization}, which can be used to represent Markov equivalence classes where some variables are intervened. Basically, interventional essential graphs are also chain graphs and can be learned from the mixture of observational and interventional data. Extending our proposed concepts, theorems, and algorithms to interventional essential graphs is straightforward. A possible future work is to extend the global characterization for definite causal relations to maximal PDAGs. Maximal PDAGs are generalizations of CPDAGs, and have been frequently used for representing causal background knowledge \citep{perkovic2017interpreting, fang2020bgida, Perkovic2020mpdag,Witte2020efficient, Guo2020minimal}. Another interesting direction is to take hidden variables and selection biases into account. For example, one may extend the results to partially ancestral graphs~\citep{richardson2002ancestral,ali2012towards,zhang2008completeness}.
\section*{Acknowledgements}
We would like to thank the editor and the three referees for their helpful comments and suggestions that greatly improved the previous versions of this paper. This work was supported by National Key R\&D Program of China (2018YFB1004300), NSFC (11671020,11971040,11771028,12071015).
|
\section{Introduction}
We consider the familiar setting of inverse problems where the goal is to recover an $n$-dimensional signal $x^*$ that is indirectly observed via a linear measurement operation $y=Ax^*$. The measurement vector can be noisy, and its dimension $m$ may be less than $n$. Several practical applications fit this setting, including super-resolution \citep{srcnn}, in-painting, denoising \cite{vincent2010stacked}, and compressed sensing \cite{cs,onenet}.
Since such an inverse problem is ill-posed in general, the recovery of $x^*$ from $y$ often requires assuming a low-dimensional structure or \emph{prior} on $x^*$. Choices of good priors have been extensively explored in the past three decades, including sparsity \cite{bpdn,cosamp}, structured sparsity \cite{modelcs}, end-to-end training via convolutional neural networks \cite{onenet, mousavi2017learning}, pre-trained generative priors \cite{CSGAN}, as well as untrained deep image priors \cite{DIP,netgd}.
In this paper, we focus on a powerful class of priors based on deep generative models. The setup is the following: the unknown signal $x^*$ is assumed to lie in the range of some pre-trained generator network, obtained from (say) a generative adversarial network (GAN) or a variational autoencoder (VAE). That is, $x^* = G(z^*)$ for some $z^*$ in the latent space. The task is again to recover $x^*$ from (noisy) linear measurements.
Such generative priors have been shown to achieve high empirical success \cite{onenet,CSGAN,deepcs}. However, progress on the theoretical side for inverse problems with generative priors has been much more modest. On the one hand, the seminal work of~\cite{bora2017CS} established the first \emph{statistical} upper bounds (in terms of measurement complexity) for compressed sensing for fairly general generative priors, which was later shown in~\cite{scarlett2020} to be nearly optimal. On the other hand, provable \emph{algorithmic guarantees} for recovery using generative priors are only available in very restrictive cases. The paper~\cite{hand2018global} proves the convergence of (a variant of) gradient descent for shallow generative priors whose weights obey a distributional assumption. The paper~\cite{shah2018solving} proves the convergence of projected gradient descent (PGD) under the assumption that the range of the (possibly deep) generative model $G$ admits a polynomial-time oracle projection. To our knowledge, the most general algorithmic result in this line of work is by~\citet{latorre2019fast}. Here, the authors show that under rather mild and intuitive assumptions on $G$, a linearized alternating direction method of multipliers (ADMM) applied to a regularized mean-squared error loss converges to a (potentially large) neighborhood of $x^*$.
The main barrier for obtaining guarantees for recovery algorithms based on gradient descent is the \emph{non-convexity} of the recovery problem induced by the generator network. Therefore, in this paper we sidestep traditional gradient descent-style optimization methods, and instead show that a very good estimate of $x^*$ can also be obtained by performing stochastic gradient Langevin Dynamics (SGLD) \cite{welling2011sgld, raginsky2017, zhang2017hitting, zou2020faster}. We show that this dynamics amounts to \emph{sampling} from a Gibbs distribution whose energy function is precisely the reconstruction loss \footnote{While preparing this manuscript, we became aware of concurrent work by \citet{jalaldimakis2020} which also pursues a similar Langevin-style approach for solving compressed sensing problems; however, they do not theoretically analyze its dynamics.}.
As a stochastic version of gradient descent, SGLD is simple to implement. However, care must be taken in constructing the additive stochastic perturbation to each gradient update step. Nevertheless, the sampling viewpoint enables us to achieve finite-time convergence guarantees for compressed sensing recovery. To the best of our knowledge, this is the first such result for solving compressed sensing problems with generative neural network priors. Moreover, our analysis succeeds under (slightly) weaker assumptions on the generator network than those made in~\cite{latorre2019fast}. Our specific contributions are as follows:
\begin{enumerate}[]
\item We propose a provable compressed sensing recovery algorithm for generative priors based on stochastic gradient Langevin dynamics (SGLD).
\item We prove polynomial-time convergence of our proposed recovery algorithm to the true underlying solution, under assumptions of smoothness and near-isometry of $G$. These are technically weaker than the mild assumptions made in~\cite{latorre2019fast}.
We emphasize that these conditions are valid for a wide range of generator networks. Section~\ref{sec:results} describes them in greater details.
\item We provide several empirical results and demonstrate that our approach is competitive with existing (heuristic) methods based on gradient descent.
\end{enumerate}
\section{Prior work}
We briefly review the literature on compressed sensing with deep generative models. For a thorough survey on deep learning for inverse problems, see~\cite{ongie2020deep}.
In \cite{CSGAN}, the authors provide sufficient conditions under which the solution of the inverse problem is a minimizer of the (possibly non-convex) program:
\begin{align}\label{eq:csprior}
\min_{x = G(z)}\|Ax-y\|_2^2 \, .
\end{align}
Specifically, they show that if $A$ satisfies the so-called set-Restricted Eigenvalue Condition (REC), then the solution to \eqref{eq:csprior} equals the unknown vector $x^*$. They also show that if the generator $G$ has a latent dimension $k$ and is $L$-Lipschitz, then a matrix $A \in \mathbb{R}^{m \times n}$ populated with i.i.d. Gaussian entries satisfies the REC, provided $m = O(k \log L)$. However, they propose gradient descent as a heuristic to solve \eqref{eq:csprior}, but do not analyze its convergence. In \cite{shah2018solving}, the authors show that projected gradient descent (PGD) for \eqref{eq:csprior} converges at a linear rate under the REC, but only if there exists a tractable \emph{projection} oracle that can compute $\operatornamewithlimits{arg\,min}_z \|x-G(z)\|$ for any $x$. The recent work~\cite{lei2019inverting} provides sufficient conditions under which such a projection can be approximately computed. In \cite{latorre2019fast}, a provable recovery scheme based on ADMM is established, but guarantees recovery only up to a neighborhood around $x^*$.
Note that all the above works assume mild conditions on the weights of the generator, use variations of gradient descent to update the estimate for $x$, and require the forward matrix $A$ to satisfy the REC over the range of $G$. \citet{hand2018global,hand2019global} showed \emph{global} convergence for gradient descent, but under the (strong) assumption that the weights of the trained generator are Gaussian distributed.
Generator networks trained with GANs are most commonly studied. However, \citet{whang2020compressed,asim2019invertible} have recently advocated using \textit{invertible} generative models, which use real-valued non-volume preserving (NVP) transformations \cite{dinh2016density}. An alternate strategy for sampling images consistent with linear forward models was proposed in \cite{lindgren2020conditional} where the authors assume an invertible generative mapping and sample the latent vector $z$ from a second generative invertible prior.
Our proposed approach also traces its roots to Bayesian compressed sensing \cite{ji2007bayesian}, where instead of modeling the problem as estimating a (deterministic) sparse vector, one models the signal $x$ to be sampled from a sparsity promoting {distribution}, such as a Laplace prior. One can then derive the maximum \emph{a posteriori} (MAP) estimate of $x$ under the constraint that the measurements $y=Ax$ are consistent. Our motivation is similar, except that we model the distribution of $x$ as being supported on the range of a generative prior.
\section{Recovery via Langevin dynamics}
\label{sec:results}
In the rest of the paper, $x\wedge y$ denotes $\min\{x,y\}$ and $x\vee y$ for $\max\{x, y\}$. Given a distribution $\mu$ and set $\cA$, we denote $\mu(\cA)$ the probability measure of $\cA$ with respect to $\mu$. $\|\mu-\nu\|_{TV}$ is the total variation distance between two distributions $\mu$ and $\nu$. Finally, we use standard big-O notation in our analysis.
\subsection{Preliminaries}
We focus on the problem of recovering a signal $x^* \in \mathbb{R}^n$ from a set of linear measurements $y \in \mathbb{R}^m$ where $$y= Ax^* + \varepsilon.$$
To keep our analysis and results simple, we consider zero measurement noise, i.e., $\varepsilon = 0$\footnote{We not in passing that our analysis techniques succeed for any vector $\varepsilon$ with bounded $\ell_2$ norm.}. Here, $A \in \mathbb{R}^{m \times n}$ is a matrix populated with i.i.d. Gaussian entries with mean 0 and variance $1/m$. We assume that $x^*$ belongs to the range of a known generative model $G: \mathcal{D} \subset \mathbb{R}^d \rightarrow \mathbb{R}^n$; that is,
$$x^* = G(z^*)~~\text{for some}~~z^* \in \mathcal{D}.$$
Following~\cite{bora2017CS}, we restrict $z$ to belong to a $d$-dimensional Euclidean ball, i.e., $\mathcal{D} = \mathcal{B}(0, R)$. Then, given the measurements $y$, our goal is to recover $x^*$. Again following~\cite{bora2017CS}, we do so by solving the usual optimization problem:
\begin{equation}
\label{eqnReconLoss}
\min_{z \in \mathcal{D}} F(z) \triangleq \| y - AG(z) \|^2.
\end{equation}
Hereon and otherwise stated, $\|\cdot\|$ denotes the $\ell_2$-norm. The most popular approach to solving~\eqref{eqnReconLoss} is to use gradient descent \cite{bora2017CS}. For generative models $G(z)$ defined by deep neural networks, the function $F(z)$ is highly non-convex, and as such, it is impossible to guarantee global signal recovery using regular (projected) gradient descent.
We adopt a slightly more refined approach. Starting from an initial point $z_0 \sim \mu_0$,
our algorithm computes stochastic gradient updates of the form:
\begin{equation}
\label{eqnSGLD}
z_{k+1} = z_{k} - \eta \nabla_z F(z) + \sqrt{2\eta \beta^{-1}} \xi_k, \quad k = 0, 1, 2, \dots
\end{equation}
where $\xi_k$ is a unit Gaussian random vector in $\mathbb{R}^d$, $\eta$ is the step size and $\beta$ is an inverse temperature parameter. This update rule is known as \emph{stochastic gradient Langevin dynamics} (SGLD) \cite{welling2011sgld} and has been widely studied both in theory and practice \cite{raginsky2017, zhang2017hitting}. Intuitively,~\eqref{eqnSGLD} is an Euler discretization of the continuous-time {diffusion equation}:
\begin{equation}
\label{eqnConDiffusion}
\mathrm{d} Z(t) = - \nabla_z F(Z(t)) \mathrm{d}t + \sqrt{2\beta^{-1}} \mathrm{d}B(t), \quad t \geq 0,
\end{equation}
where $Z(0) \sim \mu_0$. Under standard regularity conditions on $F(z)$, one can show that the above diffusion has a unique invariant Gibbs measure.
We refine the standard SGLD to account for the boundedness of $z$. Specifically, we require an additional Metropolis-like accept/reject step to ensure that $z_{k+1}$ always belongs to the support $\mathcal{D}$, and also is not too far from $z_k$ of the previous iteration. We study this variant for theoretical analysis; in practice we have found that this is not necessary. Algorithm \ref{alg:sgld} (CS-SGLD) shows the detailed algorithm. Note that we can use stochastic (mini-batch) gradient instead of the full gradient $\nabla_z F(z)$.
We wish to derive sufficient conditions on the convergence (in distribution) of the random process in Algorithm \ref{alg:sgld} to the target distribution $\pi$, denoted by:
\begin{equation}
\pi(\mathrm{d}z) \propto \exp(-\beta F(z)){\mathbf 1}(z \in \mathcal{D}), \label{eq:gibbs}
\end{equation}
and study its consequence in recovering the true signal $x^*$. This leads to the first guarantees of a stochastic gradient-like method for compressed sensing with generative priors. In order to do so, we make the following three assumptions on the generator network $G(z)$.
\begin{algorithm}[!t]
\caption{CS-SGLD}
\label{alg:sgld}
\begin{algorithmic}
\STATE \textbf{Input:} step size $\eta$; inverse temperature parameter $\beta$, radius $r$ and Lipschitz constant $L$ of $F(z)$.
\STATE Draw $z_0$ from $\mu_0 = \mathcal{N}(0, \frac{1}{2L\beta} \mathbb{I})$ truncated on $\mathcal{D}$.
\FOR {$k = 0,1,\ldots, $}
\STATE Randomly sample $\xi_k \sim \mathcal{N}(0, \mathbb{I})$.
\STATE $z_{k+1}=z_k-\eta \nabla_z F(z_k) +\sqrt{2\eta/\beta}\xi_k$
\IF{$z_{k+1} \not \in \mathcal{B}(z_k, r)\cap K$}
\STATE $z_{k+1}= z_k$
\ENDIF
\ENDFOR
\STATE \textbf{Output: $\widehat{z} = z_k$}.
\end{algorithmic}
\end{algorithm}
\begin{enumerate}[leftmargin=3em]
\item[{\bf (A.1)}] {\bf Boundedness.} For all $z \in \mathcal{D}$, we have that $\| G(z) \| \leq B$ for some $B > 0$.
\item[{\bf (A.2)}] {\bf Near-isometry.} $G(z)$ is a near-isometric mapping if there exist $0 < \iota_G \le \kappa_G$ such that the following holds for any $z, z' \in \mathcal{D}$:
\begin{align*}
\iota_G \| z - z'\| \le \| G(z) - G(z') \| \le \kappa_G \| z - z' \|.
\end{align*}
\item[{\bf (A.3)}] {\bf Lipschitz gradients.} The Jacobian of $G(z)$ is $M$-Lipschitz, i.e., for any $z, z' \in \mathcal{D}$, we have
\begin{equation*}
\| \nabla_z G(z) - \nabla_z G(z') \| \le M \| z - z' \|,
\end{equation*}
where $\nabla_z G(z) = \frac{\partial G(z)}{\partial z}$ is the Jacobian of the mapping $G(\cdot)$ with respect to $z$.
\end{enumerate}
All three assumptions are justifiable. Assumption {\bf (A.1)} is reasonable due to the bounded domain $K$ and for well-trained generative models $G(z)$ whose target data distribution is normalized. Assumption {\bf (A.2)} is reminiscent of the ubiquitous restricted isometry property (RIP) used for compressed sensing analysis \citep{candes2005RIP} and is recently adopted in \citep{latorre2019fast}. Finally, Assumption {\bf (A.3)} is needed so that the loss function $F(z)$ is smooth, following typical analyses of Markov processes.
Next, we introduce a new concept of smoothness for generative networks. This concept is a weaker version of a condition on $G(\cdot)$ introduced in \cite{latorre2019fast}.
\begin{definition}[Strong smoothness]
\label{defStrongSmoothness}
The generator network $G(z)$ is $(\alpha,\gamma)$-strongly smooth if there exist $\alpha >0$ and $\gamma \ge 0$ such that for any $z, z' \in \mathcal{D} $, we have
\begin{equation}
\label{eqnStrongSmooth}
\langle G(z) - G(z'), \nabla_z G(z) (z - z') \rangle \ge \alpha \|z - z' \| ^2 - \gamma.
\end{equation}
\end{definition}
Following~\cite{latorre2019fast} (Assumption 2), we call this property {``strong smoothness''}. However, our definition of strong smoothness requires two parameters instead of one, and is weaker since we allow for an additive slack parameter $\gamma \ge 0$.
Definition~\ref{defStrongSmoothness} can be closely linked to the following property of the {loss function} $F(z)$ that turns out to be crucial in establishing convergence results for CS-SGLD.
\begin{definition}[Dissipativity \citep{hale1990asymptotic}]
\label{defDissipative}
A differentiable function $F(z)$ on $\mathcal{D}$ is $(\alpha, \gamma)$-dissipative around $z^*$ if for constants $\alpha >0$ and $\gamma \ge 0$, we have
\begin{equation}
\label{eqnDissipative}
\langle z - z^*, \nabla_z F(z) \rangle \ge \alpha \|z - z^* \| ^2 - \gamma.
\end{equation}
\end{definition}
It is straightforward to see that \eqref{eqnDissipative} essentially recovers the strong smoothness condition \eqref{eqnStrongSmooth} if the measurement matrix $A$ is assumed to be the identity matrix. In compressed sensing, it is often the case that $A$ is a (sub)Gaussian matrix and that given a sufficient number of measurements as well as Assumptions {\bf (A.1)}, {\bf (A.2)} and {\bf (A.3)}, the dissipativity of $F(z)$ for such an $A$ can still be established.
Once $F$ is shown to be dissipative, the machinery of~\cite{raginsky2017, zhang2017hitting, zou2020faster} can be adapted to show that the convergence of CS-SGLD. The majority of the remainder of the paper is devoted to proving this series of technical claims.
\subsection{Main results}
We first show that a very broad class of generator networks satisfies the assumptions made above. The following proposition is an extension of a result in~\cite{latorre2019fast}.
\begin{proposition}
\label{propAssumptionHold}
Suppose $G(z) : \mathcal{D} \subset \mathbb{R}^d \rightarrow \mathbb{R}^n$ is a feed-forward neural network with layers of non-decreasing sizes and compact input domain $\mathcal{D}$. Assume that the non-linear activation is a continuously differentiable, strictly increasing function. Then, $G(z)$ satisfies Assumptions \textbf{(A.2)} \& \textbf{(A.3)} with constants $\iota_G, \kappa_G, M$, and if $2\iota_G^2 > M \kappa_G$, the strong smoothness in Definition \ref{defStrongSmoothness} also holds almost surely with respect to the Lebesgue measure.
\end{proposition}
This proposition merits a thorough discussion. First, architectures with increasing layer sizes are common; many generative models (such as GANs) assume architectures of this sort. Observe that the non-decreasing layer size condition is much milder than the expansivity ratios of successive layers assumed in related work~\cite{hand2018global,asim2019invertible}.
Second, the compactness assumption of the domain of $G$ is mild, and traces its provenance to earlier related works \cite{bora2017CS, latorre2019fast}. Moreover, common empirical techniques for training generative models (such as GANs) indeed assume that the latent vectors $z$ lie on the surface of a sphere~\cite{white2016sampling}.
Third, common activation functions such as the sigmoid, or the Exponential Linear Unit (ELU) are continuously differentiable and monotonic. Note that the standard Rectified Linear Unit (ReLU) activation does \emph{not} satisfy these conditions, and establishing similar results for ReLU networks is deferred to future work.
The key for our theoretical analysis, as discussed above, is Definition~\ref{defStrongSmoothness}, and establishing this requires Proposition~\ref{propAssumptionHold}. Interestingly however, in Section~\ref{sec:exp} below we provide \emph{empirical} evidence that strong smoothness holds for generative adversarial networks with ReLU activation trained on the MNIST and CIFAR-10 image datasets.
We now obtain a measurement complexity result by deriving a bound on the number of measurements required for $F$ to be dissipative.
\begin{lemma}
\label{lmSampleComplexity}
Let $G(z) : \mathcal{D} \subset \mathbb{R}^d \rightarrow \mathbb{R}^n$ be a feed-forward neural network that satisfies the conditions in Proposition~\ref{propAssumptionHold}. Let $\kappa_G$ be its Lipschitz constant. Suppose the number of measurements $m$ satisfies:
\[
m = \Omega\left(\frac{d}{\delta^2} \log (\kappa_G/\gamma) \right) \, ,
\]
for some small constant $\delta > 0$. If the elements of $A$ are drawn according to $\mathcal{N}(0, \frac{1}{m})$, then the loss function $F(z)$ is $(1-\delta, \gamma)$-dissipative with probability at least $1 - \exp(-\Omega(m\delta^2))$.
\end{lemma}
The above result can be derived using covering number arguments, similar to the treatment in~\cite{bora2017CS}. Observe that the number of measurements scales linearly with the dimension of the \emph{latent} vector $z$ instead of the \emph{ambient} dimension, keeping in line with the flavor of results in standard compressed sensing. Recent lower bounds reported~\cite{scarlett2020} also have shown that the scaling of $m$ with respect to $d$ and $\log L$ might be \emph{tight} for compressed sensing recovery in several natural parameter regimes.
We need two more quantities to readily state our convergence guarantee. Both definitions are widely used in the convergence analysis of MCMC methods.
The first quantity defines the goodness of an initial distribution $\mu_0$ with respect to the target distribution $\pi$.
\begin{definition}[$\lambda$-warm start, \cite{zou2020faster}]\label{eq:def_warm_start}
Let $\nu$ be a distribution on $\mathcal{D}$. An initial distribution $\mu_0$ is $\lambda$-warm start with respect to $\nu$ if
\begin{align*}
\sup_{\mathcal{A}: \mathcal{A} \subseteq \mathcal{D} } \frac{\mu_0(\mathcal{A})}{\nu(\mathcal{A})}\le \lambda.
\end{align*}
\end{definition}
The next quantity is the Cheeger constant that connects the geometry of the objective function and the hitting time of SGLD to a particular set in the domain \cite{zhang2017hitting}.
\begin{definition}[Cheeger constant]\label{def:cheeger} Let $\mu$ be a probability measure on $\mathcal{D}$. We say $\mu$ satisfies the isoperimetric inequality with Cheeger constant $\rho$ if for any $\cA \subset \mathcal{D}$,
\begin{align*}
\liminf_{h\rightarrow 0^+} \frac{\mu(\cA_h)-\mu(\cA)}{h}\ge \rho\min\big\{\mu(\cA), 1-\mu(\cA)\big\},
\end{align*}
where $\cA_h = \{u \in K: \exists v \in\cA, \|u - v\|_2\le h\}$.
\end{definition}
Putting all the above ingredients together, our main theoretical result describing the convergence of Algorithm~\ref{alg:sgld} (CS-SGLD) for compressed sensing recovery is given as follows.
\begin{theorem}[Convergence of CS-SGLD]
\label{thmMain}
Assume that the generative network $G$ satisfies Assumptions {\bf (A.1)} -- {\bf (A.3)} as well as the strong smoothness condition.
Consider a signal $x^* = G(z^*)$, and assume that it is measured with $m$ (sub)Gaussian measurements such that $m = \Omega(d \log \kappa_G/\gamma)$. Choose an inverse temperature $\beta$
and precision parameter $\epsilon > 0$. Then, after $k$ iterations of SGLD in Algorithm \ref{alg:sgld}, we obtain a latent vector $z_k$ such that
\begin{equation}\label{eq:conv}
\mathbb{E}\left[F(z_k)\right] \leq \epsilon + O\left(\frac{d}{\beta}\log \left( \frac{\beta}{d}\right)\right),
\end{equation}
provided the step size $\eta$ and the number of iterations $k$ are chosen such that:
\[
\eta = \widetilde{O}\left(\frac{\rho^2 \epsilon^2}{d^2\beta} \right),~ \text{and} \quad k = \widetilde{O}\left(\frac{d^3\beta^2}{\rho^4\epsilon^2}\right).
\]
\end{theorem}
In words, if we choose a high enough inverse temperature and appropriate step size, CS-SGLD converges (in expectation) to a signal estimate with very low loss within a polynomial number of iterations.
Let us parse the above result further. First, observe that the right hand side of \eqref{eq:conv} consists of two terms. The first term can be made arbitrarily small (at the cost of greater computational cost since $\eta$ decreases ). The second term represents the irreducible expected error of the exact sampling algorithm on the Gibbs measure $\pi({\mathrm d} z)$, which is worse than the optimal loss obtained at $z=z^*$.
Second, suppose the right hand side of \eqref{eq:conv} is upper bounded by $\epsilon'$. Once SGLD finds an $\epsilon'$-approximate minimizer of the loss, in the regime of sufficient compressed sensing measurements (as specified by Lemma~\ref{lmSampleComplexity}), we can invoke Theorem 1.1 in \cite{bora2017CS} along with Jensen's inequality to immediately obtain a recovery guarantee, i.e.,
$$
\mathbb{E}\left[\norm{x^* - G(z_k)}\right] \leq \sqrt{\epsilon'} .
$$
Third, the convergence rate of CS-SGLD can be slow. In particular, SGLD may require a polynomial number of iterations to recover the true signal, while linearized ADMM~\cite{latorre2019fast} converges within a logarithmic number of iterations up to a \emph{neighborhood} of the true signal. Obtaining an improved characterization of CS-SGLD convergence (or perhaps devising a new linearly convergent algorithm) is an important direction for future work.
Fourth, the above result is for noiseless measurements. A rather similar result can be derived with noisy measurements of bounded noise (says, $\|\varepsilon\| \le \sigma)$. This quantity (times a constant depending on $A$) will affect~\eqref{eq:conv} up to an additive term that scales with $\sigma$. This is precisely in line with most compressed sensing recovery results and for simplicity we omit such a
derivation.
\section{Proof outline}
\label{sec:proof}
In this section, we provide a brief proof sketch of Theorem~\ref{thmMain}, while relegating details to the appendix. At a high level, our analysis is an adaptation of the framework of \cite{zhang2017hitting, zou2020faster} specialized to the problem of compressed sensing recovery using generative priors. The basic ingredient in the proof is the use of conductance analysis to show the convergence of CS-SGLD to the target distribution in total variation distance.
Let $\mu_{k}$ denote the probability measure of $Z_k$ generated by Algorithm \ref{alg:sgld} and $\pi$ denote the target distribution in \ref{eq:gibbs}. The proof of Theorem \ref{thmMain} consists of three main steps:
\begin{enumerate
\item First, we construct an auxiliary Metropolis-Hasting Markov process to show that $\mu_k$ converges to $\pi$ in total variation for a sufficiently large $k$ and a ``good'' initial distribution $\mu_0$.
\item
Next, we construct an initial distribution $\mu_0$ that serves as a $\lambda$-warm start with respect to $\pi$.
\item Finally, we show that a random draw from $\pi$ is a near-minimizer of $F(z)$, proving that CS-SGLD recovers the signal to high fidelity
\end{enumerate}
We proceed with a characterization of the evolution of the distribution of $z_k$ in Algorithm \ref{alg:sgld}, which basically follows \citep{zou2020faster}.
\subsection{Construction of Metropolis-Hasting SGLD}\label{sec:metropolized_SGLD}
Let $g(z) = \nabla_z F(z)$, $u$ and $w$ respectively be the points before and after one iteration of Algorithm \ref{alg:sgld}; the Markov chain is written as $u \rightarrow v \rightarrow w$, where $v \sim \mathcal{N}(u - \eta g(u), \frac{2\eta}{\beta} I)$ with the following density:
\begin{align}\label{eq:trans_SGLD}
\begin{split}
&P(v|u)= \bigg[\frac{1}{(4\pi\eta/\beta)^{d/2}}\exp\bigg(-\frac{\|v - u + \eta g(u) \|_2^2}{4\eta/\beta}\bigg)\bigg|u\bigg].
\end{split}
\end{align}%
Without the correction step, $P(v|u)$ is exactly the transition probability of the standard Langevin dynamics. Note also that one can construct a similar density with a stochastic (mini-batch) gradient. The process of $v \rightarrow w$ is
\begin{align}\label{eq:markov_chain_v2w}
w =
\begin{cases}
v & v\in\mathcal{B}(u,r)\cap \mathcal{D}; \\
u & \mbox{otherwise}.
\end{cases}
\end{align}
Let $p(u) = \mathbb{P}_{v\sim P(\cdot|u)}[v\in\mathcal{B}(u,r)\cap \mathcal{D}]$ be the probability of accepting $v$. The conditional density $Q(w|u)$ is
\begin{align*}
Q(w|u) = (1-p(u))&\delta_{u}(w)
+ P(w|u)\cdot{\mathbf 1}\big[w\in\mathcal{B}(u,r)\cap \mathcal{D} \big],
\end{align*}
where $\delta_u(\cdot)$ is the Dirac-delta function at $u$. Similar to \cite{zou2020faster, zhang2017hitting}, we consider the $1/2$-lazy version of the above Markov process, with the transition distribution
\begin{align}\label{eq:def_trans_lz_sgld}
\mathcal{T}_{u}(w) = \frac{1}{2}\delta_{u}(w) + \frac{1}{2}Q(w|u),
\end{align}
and construct an auxiliary Markov process by adding an extra Metropolis accept/reject step. While proving the ergodicity of the Markov process with transition distribution $\mathcal{T}_u(w)$ is difficult, the auxiliary chain does indeed converge to a unique stationary distribution $\pi \propto e^{-\beta F(z)}\cdot{\mathbf 1}(z\in \mathcal{D})$ due to the Metropolis-Hastings correction step.
The auxiliary Markov chain is given as follows: starting from $u$, let $w$ be the state generated from $\mathcal{T}_u(\cdot)$. The Metropolis-Hasting SGLD accepts $w$ with probability,
\begin{align*}
\alpha_{u}(w) = \min\bigg\{1, \frac{\mathcal{T}_{w}(u)}{\mathcal{T}_{u}(w)}\cdot\exp\big[-\beta\big(F(w) - F(u)\big)\big]\bigg\}.
\end{align*}
Let $\mathcal{T}^{\star}_u(\cdot)$ denote the transition distribution of the auxiliary Markov process, such that
\begin{align*}
\mathcal{T}^{\star}_u(w) = (1-\alpha_u(w))\delta(u) + \alpha_u(w)\mathcal{T}_u(w).
\end{align*}
Below, we establish the connection between $\mathcal{T}_u(\cdot)$ and $\mathcal{T}^{\star}_u(\cdot)$, as well as the convergence of the original chain in Algorithm \ref{alg:sgld} through a conductance analysis on $\mathcal{T}^{\star}_u(\cdot)$.
\begin{lemma
\label{lmAuxiliaryCloseness}
Under Assumptions, $F(z)$ is $L$-smooth and satisfies $\| \nabla_z F(z) \| \le D$ for $z \in \mathcal{D}$. For $r=\sqrt{10\eta d/\beta}$, the transition distribution of the chain in Algorithm \ref{alg:sgld} is $\delta$-close the auxiliary chain, i.e., for any set $\mathcal{A} \subseteq \mathcal{D}$
\[
(1-\delta)\mathcal{T}_u^{\star}(\mathcal{A}) \le \mathcal{T}_u(\mathcal{A}) \le (1+\delta)\mathcal{T}_u^{\star}(\mathcal{A}).
\]
where $\delta = 10Ld\eta +10LD d^{1/2}\beta^{1/2}\eta^{3/2}$.
\end{lemma}
In Appendix \ref{app:keyLemmas}, we show that $F(z)$ is $L$-smooth with $L=(MB + \kappa_G^2)$ and its gradient is bounded by $D=\kappa_G^2\|A^\top A\|$.
One can verify that $\mathcal{T}^{\star}_u(\cdot)$ is time-reversible \citep{zhang2017hitting}. Moreover, following \cite{lovasz1993random, vempala2007geometric}, the convergence of a time-reversible Markov chain to its stationary distribution depends on its conductance, which is defined as follows:
\begin{definition}[Restricted conductance]\label{def:s-conductance}
The conductance of a time-reversible Markov chain with transition distribution $\mathcal{T}^{\star}_u(\cdot)$ and stationary distribution $\pi$ is defined by,
\begin{align*}
\phi \triangleq \inf_{\cA: \cA\subseteq \mathcal{D}, \pi(\cA)\in(0,1)}\frac{\int_{\cA}\mathcal{T}_u(\mathcal{D} \backslash \cA) \pi({\mathrm d} u)}{\min\{\pi(\cA), \pi(\mathcal{D} \backslash\cA)\}}.
\end{align*}
\end{definition}
Using the conductance parameter $\phi$ and the closeness $\delta$ between $\mathcal{T}_u(\cdot)$ and $\mathcal{T}^{\star}_u(\cdot)$, we can derive the convergence of $\mathcal{T}_u(\cdot)$ in total variation distance.
\begin{lemma}[\citet{zou2020faster}]
\label{lmApproximateConvergence}
Assume the conditions of Lemma \ref{lmAuxiliaryCloseness} hold. If $\mathcal{T}_u(\cdot)$ is $\delta$-close to $\mathcal{T}^{\star}_u(\cdot)$ with $\delta\le\min\{1-\sqrt{2}/2, \phi/16\}$, and the initial distribution $\mu_0$ serves as a $\lambda$-warm start with respect to $\pi$, then
\begin{align*}
\|\mu_k -\pi\|_{TV}\le \lambda\big(1 - \phi^2/8\big)^k + 16\delta/\phi.
\end{align*}
\end{lemma}
We will further give a lower bound on $\delta$ in order to establish an explicit convergence rate.
\begin{lemma}[\citet{zou2020faster}]
\label{lmConductanceLB}
Under the same conditions of Lemma \ref{lmAuxiliaryCloseness} and the step size $\eta \le \frac{1}{30Ld} \wedge \frac{d}{25\beta D^2}$, there exists a constant $c_0$ such that
\[
\phi \ge c_0\rho\sqrt{\eta/\beta}.
\]
\end{lemma}
\subsection{Convergence of $\mu_k$ to the target distribution $\pi$}
Armed with these tools, we formally establish the first step of the proof.
\begin{theorem}\label{thmConvergenceInTV}
Suppose that the generative network $G$ satisfies Assumptions {\bf (A.1)} -- {\bf (A.3)} as well as the strong smoothness condition. Set $\eta = O\big(d^{-1} \wedge\rho^2 \beta^{-1} d^{-2}\big)$ and $r = \sqrt{10\eta d/\beta}$, then for any $\lambda$-warm start with respect to $\pi$, the output of Algorithm \ref{alg:sgld} satisfies
\begin{align*}
\|\mu_k- \pi\|_{TV}\le \lambda(1-C_0\eta)^{k} + C_1\eta^{1/2},
\end{align*}
where $\rho$ is the Cheeger constant of $\pi$, $C_0 = \widetilde{O}\big(\rho^2\beta^{-1}\big)$, and $C_2 = \widetilde{O}\big(d\beta^{1/2}\rho^{-1}\big)$. In particular, if the step size and the number of iterations satisfy:
\[
\eta = \widetilde{O}\left(\frac{\rho^2 \epsilon^2}{d^2\beta} \right),~ \text{and} \quad k = \widetilde{O}\left(\frac{d^2\beta^2\log(\lambda)}{\rho^4\epsilon^2}\right),
\]
then $\|\mu_k- \pi\|_{TV} \le \epsilon$ for $\epsilon > 0$.
\end{theorem}
The convergence rate is polynomial in the Cheeger constant $\rho$ whose lower bound is difficult to obtain generally. A rough bound $\rho = e^{-\tilde O(d)}$ can be derived using the Poincar\'e constant of the distribution $\pi$, under the smoothness assumption. See \citep{bakry2008simple} for details.
\begin{proof}[Proof outline of Theorem \ref{thmConvergenceInTV}] To prove the result, we find a sufficient condition for $\eta$ that fulfills the requirements of Lemmas \ref{lmAuxiliaryCloseness}, \ref{lmApproximateConvergence} and \ref{lmConductanceLB} hold. For $\eta \le \frac{d}{25\beta D^2}$, we have
\[
\delta = 10Ld\eta + 10LD d^{1/2}\beta^{1/2}\eta^{3/2} \le 12Ld\eta.
\]
Moreover, Lemma \ref{lmApproximateConvergence} requires $\delta \le \min\{1-\sqrt{2}/2, \phi/16\}$, while $\phi \ge c_0\rho\sqrt{\eta/\beta}$ by Lemma \ref{lmConductanceLB}, so we can set
\[
\eta = \min \biggl\{\frac{1}{30Ld}, \frac{d}{25\beta D^2} , \frac{c_0^2\rho^2}{(156Ld)^2\beta} \biggr\}
\]
for these conditions to hold. Putting all together, we obtain
\begin{align*}
\|\mu_k - \pi\|_{TV} &\le \lambda\big(1 - \phi^2/8\big)^k + \frac{16\delta}{\phi} \notag\\
&\le \lambda(1- C_0 \eta)^{k} + C_1\eta^{1/2},
\end{align*}
where $C_0 = c_0^2\rho^2/8\beta$, $C_1 = 156Ld\beta^{1/2}\rho^{-1}/c_0$. Therefore, we have proved the first part.
For the second part, to achieve $\epsilon$-sampling error, it suffices to choose $\eta$ and $k$ such that
\begin{align*}
\lambda(1-C_0\eta)^k \le \frac{\epsilon}{2}, ~\text{and} \quad C_1\eta^{1/2} \le \frac{\epsilon}{2}.
\end{align*}
Plugging in $C_0, C_1$ above, we can choose
\begin{align*}
\eta = O\bigg(\frac{\rho^2\epsilon^2}{d^2\beta}\bigg)\ \text{and} \ k = O\bigg(\frac{\log(\lambda/\epsilon)}{C_0\eta}\bigg) = \widetilde{O}\bigg(\frac{d^2\beta^2\log(\lambda)}{\rho^4\epsilon^2}\bigg)
\end{align*}
such that $\|\mu_k - \pi\|_{TV} \le \epsilon$, which completes the proof.
\end{proof}
\subsection{Existence of $\lambda$-warm start initial distribution $\mu_0$}
Apart from the step size and the number of iterations, the convergence depends on $\lambda$, the goodness of the initial distribution $\mu_0$. In this part, we specify a particular choice of $\mu_0$ in establish this.
\begin{definition}[Set-Restricted Eigenvalue Condition, \citep{bora2017CS}]
For some parameters $\tau >0$ and $o \ge 0$, $A \in \mathbb{R}^{m\times n}$ is called $\text{S-REC}(\tau, o)$ if for all $z, z' \in \mathcal{D}$,
\[
\| A(G(z) - G(z'))\| \ge \tau \| G(z) - G(z') \| - o.
\]
\end{definition}
\begin{lemma}
\label{lmGaussInit}
Suppose that $G(z)$ satisfies the near-isometry property in Assumption \textbf{A.2}, and $F(z)$ is $L$-smooth. If $A$ is $\text{S-REC}(\tau, 0)$, then the Gaussian distribution $\mathcal{N}(0, \frac{1}{2\beta L}\mathbb{I})$ supported on $\mathcal{D}$ is a $\lambda$-warm start with respect to $\pi$ with $\lambda = e^{O(d)}$.
\end{lemma}
\begin{proof}
Let $\mu_0$ denote the truncated Gaussian distribution $\mathcal{N}(0, \frac{1}{2\beta L}\mathbb{I})$ on $\mathcal{D}$ whose measure is
$$\mu_0({\mathrm d} z) = {e^{-\beta L\|z\|_2^2} {\mathbf 1}(z\in \mathcal{D}) {\mathrm d} z}/{\Gamma}$$
where $ \Gamma = \int_{\mathcal{D}}e^{-\beta L\|z \|_2^2}{\mathrm d} z $ is the normalization constant.
Along with the target measure $\pi$, we can easily verify that
\begin{align*}
\frac{\mu_0({\mathrm d} z)}{\pi({\mathrm d} z)} \le \frac{\int_{\mathcal{D}} e^{-\beta F(z)}{\mathrm d} z}{\Gamma} \cdot e^{-\beta L\|z\|_2^2+\beta F(z)}.
\end{align*}
Our goal is to bound the right hand side. Using the smoothness and the simple fact $F(z^*) = 0$, we have
\begin{align*}
F(z)\le \frac{L}{2}\|z - z^*\|_2^2 \le L\|z^*\|_2^2 + L\|z\|_2^2,
\end{align*}
which implies that $e^{-\beta L\|z\|_2^2 + \beta F(z)} \le e^{\beta L\|z^*\|_2^2}$.
To bound $\int_{\mathcal{D}} e^{-\beta F(z)}{\mathrm d} z$, we use the S-REC property of $A$ as well as the near-isometry of $G(z)$. Recall the objective function:
\begin{align*}
F(z) &= \| y - AG(z) \|^2 = \| A(G(z) - G(z^*) \|^2
\ge \tau^2 \|G(z) - G(z^*)\|^2 - o \ge \tau^2 \iota_G^2 \|z - z^*\|^2,
\end{align*}
where we have dropped $o$ for simplicity. Therefore,
\begin{align*}
\int_{\mathcal{D}} e^{-\beta F(z)}{\mathrm d} z &\le \int_{\mathcal{D}} e^{-\beta \tau^2 \iota_G^2 \|z - z^*\|^2}{\mathrm d} z
\le \left(\frac{\pi}{\beta \tau^2 \iota_G^2}\right)^{d/2}.
\end{align*}
Putting the above results together, we can get
\begin{align*}
\lambda \le \max_{z\in K}\frac{\mu_0({\mathrm d} z)}{\pi({\mathrm d} z)} \le \bigg(\frac{\pi}{\beta \tau^2 \iota_G^2}\bigg)^{d/2} \frac{e^{\beta L\|z^*\|_2^2}}{\Gamma} = e^{O(d)},
\end{align*}
and conclude the proof.
\end{proof}
\subsection{Completing the proof}
\label{sec:completeproof}
\begin{proof}[Proof of Theorem \ref{thmMain}]
Consider a random draw $\wh{Z}$ from $\mu_{k}$ and another $\wh{Z}^*$ from $\pi$. We hav
\begin{align*}
\mathbb{E}[F(\wh{Z})]
& = \left( \mathbb{E} [F(\wh{Z})] - \mathbb{E} [F(\wh{Z}^*)] \right) + \mathbb{E} [F(\wh{Z}^*)].
\end{align*}
We will first give a crude bound for the second term $\mathbb{E} [F(\wh{Z}^*)]$ following the idea from \citep{raginsky2017}:
\begin{align*}
\mathbb{E} [F(\wh{Z}^*)] = \int_{\mathcal{D}} F(z) \pi({\mathrm d} z)
&\le \cO\left(\frac{d}{\beta}\log \frac{\beta}{d}\right) .
\end{align*}
The detailed proof is given in Appendix \ref{app:MarkovSemi}. The first term is related to the convergence of $\mu_k$ to $\pi$ in total variation shown in Theorem \ref{thmConvergenceInTV}. Notice that $F(z) \le 2R\|A\|\kappa_G $ for all $z \in \mathcal{D}$ due the Lipschitz property of the generative network $G$. Moreover, by Theorem \ref{thmConvergenceInTV}, we have $\|\mu_k - \pi\|_{TV} \le \epsilon'$ for any $\epsilon' >0$ and a sufficiently large $k$. Hence, the first term is upper bounded by
\begin{align*}
&\left\vert \int_{\mathcal{D}} F(z) \mu_{k}({\mathrm d} z) - \int_{\mathcal{D}} F(z) \pi({\mathrm d} z) \right\vert
\le 2R\|A\|\kappa_G \left\vert \int_{\mathcal{D}} \mu_{k}({\mathrm d} z) - \int_{\mathcal{D}} \pi({\mathrm d} z) \right\vert
\le 2R\|A\|\kappa_G \epsilon'.
\end{align*}
Given the target error $\epsilon$, choose $\epsilon' = {\epsilon}/{(2R\|A\|\kappa_G)}$. By Lemma \ref{lmGaussInit}, we have $\lambda = e^{O(d)}$.
Then, for
\[
\eta = \widetilde{O}\left(\frac{\rho^2 \epsilon^2}{d^2\beta} \right),~ \text{and} \quad k = \widetilde{O}\left(\frac{d^3\beta^2}{\rho^4\epsilon^2}\right), ~\text{we have}
\]
\[
\mathbb{E}[F(\wh{Z})] \le \epsilon + \cO\left(\frac{d}{\beta}\log \frac{\beta}{d}\right).
\]
Therefore, we complete the proof of our main result.
\end{proof}
\section{Experimental results}
\label{sec:exp}
While we emphasize that the primary focus of our paper is theoretical, we corroborate our theory with representative experimental results on MNIST and CIFAR-10. Even though we require requires a bounded domain within $d$-dimensional Euclidean ball in our analysis, empirical results demonstrate that our approach works without the restriction.
\begin{figure}[h!]
\centering
\begin{tabular}{ccc}
\raisebox{1.3\height}{\includegraphics[width=0.1\linewidth]{assets/digit.png}} &
\includegraphics[width=0.42\linewidth]{assets/mnist2.png} &
\includegraphics[width=0.42\linewidth]{assets/mnist4.png} \\
& (a) & (b) \\
\raisebox{1.3\height}{\includegraphics[width=0.1\linewidth]{assets/grass.png}} &
\includegraphics[width=0.42\linewidth]{assets/cifar2.png}
&
\includegraphics[width=0.42\linewidth]{assets/cifar4.png}
\\
& (c) & (d)
\end{tabular}
\caption{[MNIST] selected base digit $G(z^*)$, evaluating (a) dissipativity (b) \eqref{eqnDissipative4-app}, [CIFAR] selected base image $G(z^*)$, evaluating (c) dissipativity (d) \eqref{eqnDissipative4-app}.}
\label{fig:bounds-app}
\end{figure}
\subsection{Validation of strong smoothness}
As mentioned above, our theory relies on the assumption that the following condition holds for some constants $\alpha > 0, \gamma \ge 0$ and $\forall z, z' \in \mathcal{D}$ for a domain $\mathcal{D}$. Here, we take $\mathcal{D} = \mathbb{R}^d$.
\begin{align*}
\langle G(z)-G(z'), \nabla_z G(z)(z-z') \rangle \geq \alpha \|z-z'\|^2 - \gamma.
\end{align*}
To estimate these constants, we generate samples $z$ and $z'$ from $\mathcal{N}(0,\mathbb{I})$. To establish $\alpha$ and $\gamma$, we perform experiments on two different datasets (i) MNIST (Net1) and (ii) CIFAR10 (Net2). For both datasets, we compute the terms $u(z,z') = \nabla_{z} G(z)^{\top}(G(z)-G(z')),z-z' \rangle$ and $v(z,z') = \|z-z'\|^2$ for 500 different instantiations of $z$ and $z'$. We then plot these pairs of $(\alpha v - \gamma,u)$ samples for different $z$'s and $z'$'s and compute the values of $\alpha$ and $\gamma$ by a simple linear program. We do this experiment for two DCGAN generators trained on MNIST (Figure \ref{fig:bounds-app} (a)) as well as on CIFAR10 (Figure \ref{fig:bounds-app} (c)).
Similarly for the compressed sensing case, we also derive values $\alpha_A$ and $\gamma_A$, where a compressive matrix $A$ acts on the output of the generator $G$. Here, we have picked the number of measurements $m = 0.1n$ where $n$ is the signal dimension. This is encapsulated in the following equation:
\begin{align}
\langle \nabla_z (AG(z))^{\top}(AG(z)-AG(z')),z-z' \rangle \geq \alpha_A \|z-z'\|^2 - \gamma_A
\label{eqnDissipative4-app}
\end{align}
\begin{comment}
\begin{align}
\quad \langle \nabla_z (AG(z))^{\top}(AG(z)-AG(z^*)),z-z^* \rangle \geq \alpha_A \|z-z^*\|^2 - \gamma_A
\label{eqnDissipative3-app}
\end{align}
\end{comment}
for all possible Gaussian matrices $A$ and different instantiations of $z$ and $z'$. Here, we capture the left side of the inequality in $u(z,z') = \langle \nabla_{z} (AG(z))^{\top}(AG(z)-AG(z')),z-z' \rangle$.
We similarly plot points $(\alpha_A v(z, z') -\gamma_A, u(z, z'))$ for all pairs $(z, z')$. The scatter plot is generated for 500 different instantiations of $z$ and $z'$ and $5$ different instantiations of $A$. We do this experiment for two DCGAN generators, one trained on MNIST (Figure \ref{fig:bounds-app} (b)) and the other trained on CIFAR10 (Figure \ref{fig:bounds-app} (d)). These experiments indicate that the dissipativity constant $\alpha$ is positive in all cases.
\subsection{Comparison of SGLD against GD}
\begin{figure*}[!ht]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.2\linewidth]{assets/recons/mnist_true.png} & \includegraphics[width=0.2\linewidth]{assets/recons/mnist_init.png} & \includegraphics[width=0.2\linewidth]{assets/recons/mnist_gd.png} & \includegraphics[width=0.2\linewidth]{assets/recons/mnist_lgd.png} \\
(a) Ground truth & (b) Initial & (c) GD & (d) SGLD \\
& & MSE = 0.0447 & MSE = 0.0275
\end{tabular}
\caption{Comparing the recovery performance of SGLD and GD at $m=0.2n$ measurements. \label{fig:reconstructions}}
\end{figure*}
We test the SGLD reconstruction by using the update rule in \eqref{eqnSGLD} and compare against the optimizing the updates of $z$ using standard gradient descent as in \cite{bora2017CS}. For all experiments, we use a pre-trained DCGAN generator, with network configuration described as follows: the generator consists of four different layers consisting of transposed convolutions, batch normalization and RELU activation; this is followed by a final layer with a transposed convolution and $\tanh$ activation \cite{DCGAN}.
We display the reconstructions on MNIST in Figure \ref{fig:reconstructions}. Note that the implementation in \cite{bora2017CS} requires 10 random restarts for CS reconstruction and they report the results corresponding to the best reconstruction. This likely suggests that the standard implementation is likely to get stuck in bad local minima or saddle points. For the sake of fair comparison, we fix the same random initialization of latent vector $z$ for both GD and SGLD with no restarts. We select $m=0.2n$. In Figure \ref{fig:reconstructions} we show reconstructions for the 16 different examples, which were all reconstructed at once using same $k=2000$ steps, learning rate of $\eta=0.02$ and the inverse temperature $\beta=1$ for both approaches. The only difference is the additional noise term in SGLD (Figure \ref{fig:reconstructions} part (d)). Notice that this additional noise component helps achieve better reconstruction performance overall as compared to simple gradient descent.
Phase transition plots scanning a range of compression ratios $m/n$ as well as example reconstructions on CIFAR-10 images can be found in the supplement. More thorough empirical comparisons with PGD-based approaches~\cite{shah2018solving,raj2019gan} are deferred to future work.
\subsection{Reconstructions for CIFAR10}
We display the reconstructions on CIFAR10 in Figure \ref{fig:reconstructions2}. As with the implementation for MNIST, for the sake of fair comparison, we fix the same random initialization of latent vector $z$ for both GD and SGLD with no restarts. We select $m=0.3n$. In Figure \ref{fig:reconstructions2} we show reconstructions for the 16 different examples from MNIST, which were all reconstructed at once using same $k=2000$ steps, learning rate of $\eta=0.05$ and the inverse temperature $\beta=1$ for both approaches. The only difference is the additional noise term in SGLD (Figure \ref{fig:reconstructions} part (d)). Similar to our experiments on MNIST we notice that this additional noise component helps achieve better reconstruction performance overall as compared to simple gradient descent.
\begin{figure}[!h]
\centering
\begin{tabular}{cccc}
\includegraphics[width=0.23\linewidth]{assets/recons/cifar_true.png} & \includegraphics[width=0.23\linewidth]{assets/recons/cifar_init.png} & \includegraphics[width=0.23\linewidth]{assets/recons/cifar_gd.png} & \includegraphics[width=0.23\linewidth]{assets/recons/cifar_lgd.png} \\
(a) Ground truth & (b) Initial & (c) GD & (d) SGLD \\
& & MSE = 0.0248 & MSE = 0.0246
\end{tabular}
\caption{ [CIFAR10] Comparing the recovery performance of SGLD and GD at $m=0.3n$ measurements.} \label{fig:reconstructions2}
\end{figure}
Next, we plot phase transition diagrams by scanning the compression ratio $f=m/n = [0.2,0.4,0.6,0.8,1.0]$ for the MNIST dataset in Figure \ref{fig:phasetrans}. For this experiment, we have chosen 5 different instantiations of the sampling matrix $A$ for each compression ratio $f$. In Figure \ref{fig:phasetrans} we report the average Mean Square Error (MSE) of reconstruction $\|\hat{x}-x\|^2$ over 5 different instances of $A$.
\begin{figure}[!h]
\centering
\begin{tikzpicture}[scale=0.5]
\begin{semilogyaxis
[width=0.75\textwidth,
xlabel= Compression ratio $f$,
ylabel= $\log$ MSE,
grid style = dashed,
grid=both,
y label style={at={(axis description cs:(-0.1,0.5)},anchor=north},
legend style=
{at={(0.7,0.95)},
anchor=south west,
anchor= north ,
} ,
]
\addplot[color=magenta, solid, mark=*, mark size = 2, line width =2] plot coordinates {
(0.2, 0.1464)
(0.4, 0.0732)
(0.6, 0.0461)
(0.8, 0.0429)
(1.0, 0.0436)
};
\addplot[color=blue, solid, mark=square, line width=2] plot coordinates {
(0.2, 0.1434)
(0.4, 0.0679)
(0.6, 0.0451)
(0.8, 0.0395)
(1.0, 0.0404)
};
\legend{GD\\SGLD\\}
\end{semilogyaxis}
\end{tikzpicture}
\caption{Phase transition plots representing average MSE of reconstructed image using gradient descent and stochastic gradient Langevin dynamics.}
\label{fig:phasetrans}
\end{figure}
We conclude that SGLD gives improved reconstruction quality as compared to GD.
\clearpage
\bibliographystyle{icml2021}
\section{Conductance Analysis}
In this section, we provide the proofs of Lemma \ref{lmAuxiliaryCloseness} and \ref{lmConductanceLB} based on the conductance analysis laid out in \citep{zhang2017hitting} and similarly in \citep{zou2020faster}. The proof of \ref{lmApproximateConvergence} directly follows from Lemma 6.3, \citep{zou2020faster}.
\begin{proof}[Proof of Lemma \ref{lmAuxiliaryCloseness}]
We use the same idea in Lemma 3 from \citep{zhang2017hitting} and similarly that in Lemma 6.1 from \citep{zou2020faster}.
The main difference of our proof is that we use full gradient $\nabla_z F(z)$ in Algorithm \ref{alg:sgld}, instead of stochastic mini-batch gradient, which simplifies the proof of this lemma a little.
We consider two cases for each $u$: $u \not\in \cA$ and $u\in\cA$. As long as we can prove the first case, the second case easily follows, by splitting $\cA$ into $\{u\}$ and $\cA \backslash \{u\}$ and using the result of the first case. For a detailed treatment of the latter case, we refer the reader to the proof of Lemma 6.1 in \citep{zou2020faster}.
Now that $u \notin \cA$, we have
\begin{align}\label{eq:relation_Q2P_case1}
\mathcal{T}^{\star}_{u}(\cA)=\int_{\cA \cap \mathcal{B}(u, r)} \mathcal{T}^{\star}_{u}(w){\mathrm d} w = \int_{\cA \cap \mathcal{B}(u, r) } \alpha_u(w) \mathcal{T}_{u}(w){\mathrm d} w.
\end{align}
where $\alpha_u(w)$ is the acceptance ratio of the Metropolis-Hasting. If suffices to show that $\alpha_u(w)\ge 1-\delta/2$ for all $w\in K \cap \cB(u,r)$, which implies
\begin{align*}
(1-\delta/2)\mathcal{T}_u(\cA)\le \mathcal{T}^{\star}_u(\cA)\le \mathcal{T}_u(\cA).
\end{align*}
The right hand side is obvious by the definition of $\alpha_u(w)$ while we can ensure $\delta \le 1/2$ with a sufficiently small $\eta$. What remains is to show that
\begin{align}\label{eq:lowerbound_accept}
\frac{\mathcal{T}_{w}(u)}{\mathcal{T}_u(w)}\cdot\exp(-\beta(F(w) - F(u)))\ge 1-\delta/2.
\end{align}
The left hand side is simplified by definition of $\mathcal{T}_u(w)$ as
\begin{align*}
\exp\bigg(\frac{\|w - u + \eta g(u)\|_2^2}{4\eta/\beta} - \frac{\|u - w + \eta g(w)\|_2^2}{4\eta/\beta}\bigg) \exp(-\beta(F(w) - F(u))) \ge 1- \delta/2.
\end{align*}
Note that $g(z) = \nabla_z F(z)$. Simplify the first exponent and combine with the second one gives the following form:
\begin{align}
\label{eq:simplified_exponent}
-\beta \left(F(w) - F(u) - \frac{1}{2}\langle w-u, \nabla_z F(w) + \nabla_z F(u) \rangle \right) + \frac{\eta\beta}{4} (\|\nabla_z F(u)\|^2 - \|\nabla_z F(w)\|^2).
\end{align}
To lower bound the left hand side, we appeal to the smoothness of $F(z)$. Specifically, by Lemmas \ref{lmGradientBound} and \ref{lmSmoothness}, we have $F$ is $L$-smooth and $\| \nabla_z F(z) \| \le D$ with $L=(MB + \kappa_G^2)$ and $D=\kappa_G^2\|A^\top A\|$. Then,
\begin{align*}
F(w) &\leq F(u) +\langle w - u, \nabla F(u) \rangle+\frac{L\|w -u\|_2^2}{2},\\
F(u) &\geq F(w) +\langle u - w, \nabla F(w)\rangle-\frac{L\|w - u\|_2^2}{2}.
\end{align*}
This directly implies that
\begin{align}
\big|F(w) - F(u) -\langle w - u, \frac{1}{2}\nabla F(w) + F(u)\rangle\big|\le \frac{L\|w - u\|_2^2}{2}.\label{eq:Nestorov}
\end{align}
Moreover,
\begin{align}\label{eq:upperbound_variance_difference}
\big|\|\nabla_z F(u)\|_2^2 - \|\nabla_z F(w)\|_2^2\big|&\leq \|\nabla F(u)-\nabla F(w)\|_2\cdot\|\nabla F(u)+\nabla F(w)\|_2\notag\\
&\le 2LD\|w-u\|_2.
\end{align}
Combining \eqref{eq:Nestorov} and \eqref{eq:upperbound_variance_difference} in \eqref{eq:simplified_exponent}, and together with $w \in \mathcal{B}(u, r)$ with $r = \sqrt{10\eta d/\beta}$,
\begin{align*}
\text{LHS of \eqref{eq:simplified_exponent}} &\ge -\frac{L\beta \| w -u\|^2}{2} - \frac{\eta\beta LD \|w - u\|}{2} \\
&\ge -5Ld\eta - 5LG d^{1/2}\beta^{-1/2}\eta^{3/2}.
\end{align*}
Pick $\delta/2 = 5Ld\eta + 5LD d^{1/2}\beta^{-1/2}\eta^{3/2}$, and use the fact $e^{-x} \ge 1 - x$ for $x \ge 0$, then we have proved the result.
\end{proof}
Next, we lower bound the conductance $\phi$ of $\mathcal{T}^{\star}_u(\cdot)$ using the idea in \citep{lee2018convergence, zou2020faster}, by first restating the following lemma:
\begin{lemma}[Lemma 13 in \citet{lee2018convergence}]\label{lemma:lowerbound_transitionprob}
Let $\mathcal{T}^{\star}_u(\cdot)$ be a time-reversible Markov chain on $\mathcal{D}$ with stationary distribution $\pi$. Suppose for any $u,v\in \mathcal{D}$ and a fixed $\Delta> 0$ such that $\|u-v\|_2\le \Delta$, we have $\|\mathcal{T}^{\star}_u(\cdot) - \mathcal{T}^{\star}_v(\cdot)\|_{TV}\le 0.99$, then the conductance of $\mathcal{T}^{\star}_u(\cdot)$ satisfies $\phi\ge C\rho\Delta$ for some constant $C > 0$ and $\rho$ is the Cheeger constant of $\pi$.
\end{lemma}
\begin{proof}[Proof of Lemma \ref{lmConductanceLB}] To apply Lemma \ref{lemma:lowerbound_transitionprob}, we follow the same idea of \citet{zou2020faster} and reuse some of their results without proof. To this end, we prove that for some $\Delta$, any pair of $u, v \in \mathcal{D}$ such that $\|u - v\|_2\le \Delta$, we have $\|\mathcal{T}^{\star}_u(\cdot) - \mathcal{T}^{\star}_v(\cdot)\|_{TV}\le 0.99$. Recall the distribution of the iterate $z$ after one-step standard SGLD without the accept/reject step in \eqref{eq:trans_SGLD} is
\begin{align*}
P(z|u) = \frac{1}{(4\pi\eta/\beta)^{d/2}}\exp\bigg(-\frac{\|z - u + \eta g(u)\|_2^2}{4\eta/\beta}\bigg)
\end{align*}
Since Algorithm \ref{alg:sgld} accepts the candidate only if it falls in the region $\mathcal{D} \cap \cB(u,r)$, the acceptance probability is
\begin{align*}
p(u) = \mathbb{P}_{z\sim P(\cdot|u)}\big[z\in \mathcal{D} \cap \cB(u,r)\big].
\end{align*}
Therefore, the transition probability $\mathcal{T}^{\star}_{u}(z)$ for $z\in \mathcal{D} \cap \cB(u,r)$ is given by
\begin{align*}
\mathcal{T}^{\star}_u(z) = \frac{2 - p(u) + p(u)(1-\alpha_u(z))}{2} \delta_u(z) + \frac{\alpha_u(z)}{2}P(z|u)\cdot{\mathbf 1}[z\in \mathcal{D} \cap \cB(u,r)].
\end{align*}
Take $u, v \in \mathcal{D}$ and let $\cS_u = \mathcal{D} \cap\cB(u,r)$ and $\cS_v = \mathcal{D} \cap\cB(v,r)$. By the definition of the total variation, there exists a set $\cA\in \mathcal{D} $ such that
\begin{align*}
\|\mathcal{T}^{\star}_u(\cdot) - \mathcal{T}^{\star}_v(\cdot)\|_{TV} &= |\mathcal{T}^{\star}_u(\cA) - \mathcal{T}^{\star}_v(\cA)|\notag\\
&\le \underbrace{\max_{u,z} \bigg[\frac{2 - p(u) + p(u)(1-\alpha_u(z))}{2}\bigg]}_{I_1} \notag\\
\qquad &+ \frac{1}{2}\underbrace{\bigg|\int_{z\in\cA} \alpha_u(z)P(z|u){\mathbf 1}(z\in\cS_u) - \alpha_v(z)P(z|v){\mathbf 1}(z\in\cS_v){\mathrm d} z\bigg|}_{I_2}.
\end{align*}
Using the mini-batch size that is exactly the same as the number of samples, we can reuse the bounds of $I_1$ and $I_2$ in Lemmas C.4 and C.5 of \citep{zou2020faster}. Consequently,
\begin{align*}
\|\mathcal{T}^{\star}_u(\cdot) - \mathcal{T}^{\star}_v(\cdot)\|_{TV}\le I_1 + I_2/2 \le 0.85+0.1\delta + \frac{\sqrt{\beta}\|u-v\|_2}{\sqrt{2\eta}}.
\end{align*}
By Lemma \ref{lmAuxiliaryCloseness}, we have $\delta =10Ld\eta +10LDd^{1/2}\beta^{1/2}\eta^{3/2} \le 12Ld\eta$ if $\eta \le \frac{d}{25\beta D^2}$. Thus if
\begin{align*}
\eta \le \frac{1}{25\beta D^2}\wedge \frac{1}{30Ld\eta} \quad \mbox{and} \quad \|u-v\|_2\le \frac{\sqrt{2\eta}}{10\sqrt{\beta}}\le 0.1 r,
\end{align*}
we have $\|\mathcal{T}^{\star}_u(\cdot) - \mathcal{T}^{\star}_v(\cdot)\|_{TV}\le 0.99$. As the result of Lemma \ref{lemma:lowerbound_transitionprob}, we prove a lower bound on the conductance $\phi$ of $\mathcal{T}^{\star}_u(\cdot)$
\begin{align*}
\phi \ge c_0\rho\sqrt{\eta/\beta},
\end{align*}
and finish the proof.
\end{proof}
|
\section{Introduction and main results}
\subsection{Introduction}
Many examples of Schrödinger operators with ``strange" spectral properties involve sparse potentials. In his seminal work \cite{MR484145} Pearson constructed examples of real-valued potentials (on the half-line) leading to singular continuous spectrum.
The potentials consists of an infinite sequence of ``bumps" of identical profile, and the separation between these bumps increases rapidly. The physical interpretation is that a quantum mechanical particle will ultimately be reflected from a bump. These ideas were further developed in several directions, see e.g.\ \cite{MR1463044} \cite{MR1342046}, \cite{MR1628290}, \cite{MR1666767}, \cite{MR1866165}, \cite{MR2027640}, and the references therein. Scattering from sparse potentials in higher dimensions was studied by Molchanov and Vainberg \cite{MR1690097}, \cite{MR1756700}; see also \cite{MR1606716}, \cite{MR1803388},\cite{MR2724615}, \cite{MR3390425}, \cite{MR2073594}. The discrete spectrum for multidimensional lattice Schrödinger operators was investigated by Rozenblum and Solomyak \cite{MR2544038}. They constructed examples of sparse potentials whose number of negative eigenvalues grows like an arbitrary given polynomial power in the large coupling limit.
In the recent work \cite{MR3627408} Boegli constructed a complex-valued sparse potential with arbitrary small $L^q$ norm ($q>d$) that has infinitely many non-real eigenvalues accumulating at every point of the essential spectrum.
Since the proof is based on compactness arguments, there is no quantitative bound on the rate of separation between the bumps, and hence no estimate on the pointwise decay of the potential or the accumulation rate of the eigenvalues is possible.
\subsection{A quantitative version of Boegli's example}
Our first result provides quantitative decay bounds for the example in \cite{MR3627408}. Perhaps more importantly, the construction can be used to produce a potential together with an infinite number of eigenvalues (possibly not all of them) satisfying given upper and lower bounds on their accumulation rate. We formulate our result for the most interesting spectral region
\begin{align}\label{def. Sigma_epsilon_0}
\Sigma_{0}=\set{z\in{\mathbb C}}{|\im z|\leq \epsilon_0 \re z},
\end{align}
where $\epsilon_0>0$ is small but fixed.
\begin{theorem}\label{thm quant Boegli intro}
Let $d\geq 1$, $q>d$, $\epsilon_1,\epsilon_2\in (0,1],\gamma>0$, and let $(\zeta_n)_n\subset \Sigma_{0}$ be a sequence satisfying
\begin{align}\label{condition on sequence zetan intro}
\big(\sum_{n\in{\mathbb N}}|\zeta_n|^{\frac{d}{2}}|\im\zeta_n|^{q-d}|\log^{d}|\im\zeta_n/\zeta_n||\big)^{\frac{1}{q}}\asymp \epsilon_1.
\end{align}
Then there exists a complex sparse potential $V$ such that the following hold:
\begin{enumerate}
\item[a)] For each $n\in{\mathbb N}$ there exists a discrete eigenvalue $z_n$ of $H_V=-\Delta+V$ which is exponentially close to $\zeta_n$, in the sense $|z_n-\zeta_n|\leq \exp(-|\im\zeta_n|^{-\gamma})$.
\item[b)] The potential satisfies $\|V\|_{L^q({\mathbb R}^d)}\lesssim \epsilon_1$.
\item[c)] The potential decays polynomially, i.e.\ there exists a positive constant $\beta=\beta(\gamma,d,q)$ such that $|V(x)|\lesssim \epsilon_2\langle x\rangle^{-\beta}$.
\end{enumerate}
\end{theorem}
\begin{remark}
\noindent (i) In particular, for any $\lambda\in (0,\infty)$ there exists a sequence $(\zeta_n)_n\subset \Sigma_{0}$ satisfying \eqref{condition on sequence zetan intro} such that $\lim_{n\to\infty}\zeta_n=\lambda$. In this way one can find a sequence accumulating to every point of the essential spectrum. This yields a constructive proof of the result of Boegli \cite{MR3627408}.
\noindent (ii) One can remove the logarithm in \eqref{condition on sequence zetan intro} at the expense of replacing the $L^q$ norm of $V$ by the ``Davies--Nath norm" (see \eqref{def. Davies--Nath norm}).
\noindent (iii) We will give explicit bounds on the polynomial decay $\beta$.
\noindent (iv) Substituting the trivial lower bound $|\zeta_n|\geq |\im\zeta_n|$ into \eqref{condition on sequence zetan intro} shows $\im\zeta_n\to 0$. This is the reason why we say that $z_n$ is exponentially close to $\zeta_n$.
\end{remark}
We believe that the pointwise condition c) is more natural than the $L^q$ condition~b) for the phenomenon that takes place in Theorem \ref{thm quant Boegli intro}. This is because complex analogues of classical phase space bounds, which motivate the consideration on $L^q$ norms in the first place, lack many of the features that make them so useful for real potentials (more on that in Subsection \ref{subsection Weyl} below). Put simply, the $L^q$ norm does not see the separation between the bumps, while the pointwise bound does. We will nevertheless work with $L^q$ norms since we allow the bumps to have singularities. In the case where they are bounded the pointwise decay of the whole potential can easily be estimated by comparing the $L^{\infty}$ norms of the bumps to their spatial separation from the origin. In his fundamental work on non-selfadjoint Schrödinger operators, Pavlov \cite{MR0203530}, \cite{MR0234319} showed that the number of eigenvalues in one dimension is finite if $|V(x)|\lesssim \exp(-c|x|^{1/2})$, and that this exponential rate is best possible. This means that the potential in Theorem \ref{thm quant Boegli intro} cannot decay too fast. The $L^q$ bound imposes no decay whatsoever, but we can at least establish polynomial decay. For recent quantitative improvements of Pavlov's bound we refer to Borichev--Frank--Volberg~\cite{borichev2019counting} and Sodin \cite{MR4088346}.
The proof of the example in \cite{MR3627408} is based on ``soft" methods like weak convergence, compact embedding and the notion of the limiting essential spectrum. In contrast, our proof uses ``hard" estimates for the resolvent and the Birman-Schwinger operator, combined with tools from complex analysis such as Rouch\'e's theorem, Jensen's formula and Cartan type estimates. This allows us to obtain more precise results thatn those in \cite{MR3627408}. Rouch\'e's theorem and Jensen's formula are among the most ubiquitous albeit simple tools in non-selfadjoint spectral theory, where such machinery as the variational principle or the spectral theorem is not available. In the present paper Cartan type estimates are crucial to bound a certain Fredholm determinant from below and get upper upper bounds on the norm of the resolvent. This opens the way to proving existence of eigenvalues by means of quasimode construction. We are then in a setting similar to the selfadjoint case where a quasimode of size $\epsilon$ guarantees the existence of a spectral point in an $\epsilon$-neighborhood of the quasi-eigenvalue. This follows from the inequality $\|(H_V-z)^{-1}\|\leq 1/{\rm d}(z,\sigma(H_V))$, where $\sigma(H_V)$ is the spectrum. In the non-selfadjoint case the inequality may fail dramatically. This phenomenon gives rise to the notion of pseudospectrum, which we will not discuss here (see e.g.\ the monograph \cite{MR2359869}). The upper bounds obtained by Cartan type estimates generally grow exponentially in $1/{\rm d}(z,\sigma(H_V))$. In order to beat this, we are forced to construct exponentially small quasimodes, a challenging task in all but the simplest models. The strategy is reminiscent of the proof of existence of resonances close to the real axis due to Tang--Zworski \cite{MR1637824} and Stefanov \cite{MR1700740} (see also the recent book by Dyatlov--Zworski \cite{MR3969938}). Our method is perhaps closest to that of Dencker--Sjöstrand--Zworski \cite[Section 6]{MR2020109} for non-selfajoint dissipative Schrödinger operators. The difference is that we consider decaying potentials and do not assume, as these authors do, that the quasi-eigenvalue is real (see \cite[Proposition 6.4]{MR2020109}). This means that the amplification of the exponential upper bound through the maximum principle (see \cite[Proposition 6.2]{MR2020109}) is in general not possible in our case.
Another crucial difference is that we need a more quantitative version of the Cartan type estimate (Lemma \ref{lemma Levin}) as well as of the conformal transformations between the spectral region and the model domain (the unit disk). The Riemann mapping theorem is notoriously non-quantitative. Instead, we use Cayley and Schwarz--Christoffel transformations, which have previously been used in other contexts related to non-selfadjoint spectral theory, especially in connection with Lieb--Thirring type inequalities. The combination with Rouch\'e's theorem and the Cartan type bounds is new and leads to results with an inverse problem flavor, as in Theorem \ref{thm quant Boegli intro}.
\subsection{Magnitude bounds}
The second result gives precise bounds on the magnitude of eigenvalues of Schrödinger operators with complex sparse potentials, or more generally, potentials of the form $V=\sum_{j=1}^NV_j$, where the $V_j$ have disjoint support and separate rapidly from each other. We will call these ``separating" potentials. The Schrödinger operator $H_V=-\Delta+V$ behaves like an almost orthogonal sum, due to the rapid decoupling between the $N$ ``channels". This enables us to improve upon the bounds for general complex potentials due to Frank \cite{MR2820160}, \cite{MR3717979}. For simplicity we state the result here for $d\geq 3$. The general case along with further refinements can be found in Subsection \ref{subsection Magnitude bounds}. We refer to Section \ref{section Definitions} for a more in-depth explanation of the terminology.
\begin{theorem}\label{theorem sparse magnitude bound intro}
Assume that $d\geq 3$ and $d/2\leq q\leq (d+1)/2$. If $V$ is separating at scale $\eta^{-1}$, then every eigenvalue $z$ of $H_V$ with $\im\sqrt{z}\geq (d+1)\eta$ satisfies
\begin{align}\label{sparse magnitude bound}
|z|^{q-\frac{d}{2}}\lesssim \sup_{j\in[N]}\|V_j\|_{L^q({\mathbb R}^d)}^q.
\end{align}
If $q>(d+1)/2$, then
\begin{align}\label{sparse magnitude bound q>(d+1)/2}
|z|^{\frac{1}{2}}{\rm d}(z,{\mathbb R}_+)^{q-\frac{d+1}{2}}\lesssim \sup_{j\in[N]}\|V_j\|_{L^q({\mathbb R}^d)}^q.
\end{align}
\end{theorem}
The bound \eqref{sparse magnitude bound} follows from \eqref{BS bound} by a Birman--Schwinger argument. It could also be proved by using the eigenvalue bounds of the author \cite{MR4104544}, which are inspired by a method of Davies and Nath \cite{MR1946184} in one dimension. For $N=1$ the estimates \eqref{sparse magnitude bound}, \eqref{sparse magnitude bound q>(d+1)/2} coincide with those of Frank \cite{MR2820160}, \cite{MR3717979}, respectively. The difference is that here $V$ might decay very slowly or not at all. Nevertheless, on the $\eta^{2}$ energy (spectral) scale the estimate is of the same quality as for $N=1$.
We make a short remark about the connection with the Laptev--Safronov conjecture \cite{MR2540070}, which stipulates that
\begin{align}\label{LS conjecture}
\sup_{V\in L^q({\mathbb R}^d)}\sup_{z\in\sigma(-\Delta+V)\setminus{\mathbb R}_+}\frac{|z|^{q-\frac{d}{2}}}{\|V\|_q^q}<\infty\quad\mbox{for all}\quad q\in[d/2,d].
\end{align}
For the range $q\in[d/2,(d+1)/2]$ the conjecture was proven by Frank \cite{MR2820160}. The question whether \eqref{LS conjecture} is true for $q\in((d+1)/2,d]$ is still open.
The expectation, based on intuition from counterexamples to Fourier restriction (see e.g.\ \cite{MR4107520}, \cite{MR4104544} for more explanations) is that the conjecture is false in this range. Incidentally, Theorem \ref{thm quant Boegli intro} clearly implies the necessity of $q>d$ in the conjecture, but this already follows from B\"ogli's result (without the pointwise bound). In fact, a single bump of the sparse potential used in Boegli's construction (and in Theorem \ref{thm quant Boegli intro}) already provides a counterexample. Since there seems to be some confusion about this issue we use the results of Section \ref{section Complex square wells} to show necessity of the condition $q>d$. Indeed, Lemma \ref{lemma single square well} implies that for small $\epsilon>0$ there is a potential $V(\epsilon)$ and an eigenvalue $z(\epsilon)$ such that
$|z(\epsilon)|^{q-\frac{d}{2}}/\|V(\epsilon)\|_q^q\gtrsim \epsilon^{q-d}\log^q(1/\epsilon)$.
The example (a complex step potential) is simple but, quite amazingly, generic enough to show optimality of several estimates in the literature (see \cite{MR4104544}). In one dimension, the step potential can be tuned to essentially saturate any of the known magnitude bounds. For example, the last inequality also shows that the Davies--Nath bound \cite{MR1946184} is sharp and, since $\im z(\epsilon)\asymp\epsilon$, that Frank's bound \cite[Theorem 1.1]{MR3717979}
is sharp up to a logarithm. In Subsection \ref{subsection Weyl} we will show how the complex step potential also implies sharpness of another bound in \cite{MR3717979}.
\subsection{A generalization of Klaus' theorem}
The following is a generalization of a result due to Klaus \cite{MR700696} on the characterization of the essential spectrum. The generalization is twofold: First, we admit complex potentials and second, we prove it for any dimension (whereas Klaus only proved the one-dimensional case). In this introduction we again focus on the case $d\geq 3$, but the statement is valid in $d=1,2$ for $q$ in the range \eqref{range q}.
\begin{theorem}[Klaus' theorem \cite{MR700696} for complex potentials]\label{thm. Klaus' theorem intro}
Assume that $d\geq 3$ and $d/2\leq q\leq (d+1)/2$. If $V$ is a separating potential and $\sup_{j\in[N]}\|V_j\|_{L^q({\mathbb R}^d)}<\infty$, then
\begin{align*}
\sigma_{\rm e}(H)=[0,\infty)\cup S,
\end{align*}
where $S$ is the set of all $z\in{\mathbb C}\setminus [0,\infty)$ such that there exist infinite sequences $(i_n)_n,(z_n)_n$ with $z_n\in \sigma(H_{V_{i_n}})$, $i_n\to\infty$ and $z_n\to z$ as $n\to\infty$.
\end{theorem}
An alternative proof (also in one dimension) of Klaus' theorem can be found in~\cite{MR883643}. The role of Theorem \ref{thm. Klaus' theorem intro} in this paper will be an auxiliary one, and we will only use it to argue that the essential spectrum is invariant under the perturbations we consider in Theorem \ref{thm quant Boegli intro}. Although our proof follows the general strategy of \cite{MR700696} it is still worth emphasizing that some parts of it require somewhat novel techniques.
\subsection{Weyl's law and locality}\label{subsection Weyl}
Recently, Boegli and Štampach \cite{boegli2020liebthirring} disproved a conjecture by Demuth, Hansmann and Katriel \cite{MR3016473} for one-dimensional Schrödinger operators with complex potentials by establishing a lower bound on a certain Riesz means of eigenvalues.
More precisely, consider $H_{\alpha V}$,
where $V=\I\mathbf{1}_{[-1,1]}$ is a purely imaginary step potential and $\alpha$ is a large semiclassical parameter. Boegli and Štampach prove that, for any $p\geq 1$,
\begin{align}\label{main result SF}
\alpha^{-p}\sum_{z\in\sigma_d(H_{\alpha V})}\frac{(\im z)^p}{|z|^{1/2}}\geq C_p\log\alpha.
\end{align}
The interesting feature of this bound is that it shows a logarithmic violation of Weyl's law. To recall Weyl's law, consider a \emph{self-adjoint} operator, with a smooth real-valued potential. Note that if we set $h=1/\sqrt{\alpha}$, then $\alpha^{-1}H_{\alpha V}$ takes the form of a semiclassical Schrödinger operator, $-h^2\partial_x^2+V(x)$. Semiclassical asymptotics (Weyl's law) yield, for a suitable class of functions $f$,
\begin{align}\label{Weyl law f}
\Tr f(H_{\alpha V})=\frac{\sqrt{\alpha}}{2\pi}\big(\int f(\alpha(\xi^2+V(x))){\rm d} x{\rm d} \xi+o(1)\big)
\end{align}
as $\alpha\to \infty$. In particular, if $f$ is homogeneous of degree $\gamma$, then
\begin{align}\label{Weyl law}
\lim_{\alpha\to \infty}\alpha^{-1/2-\gamma}\Tr f(H_{\alpha V})=(2\pi)^{-1}\int f(\xi^2+V(x)){\rm d} x{\rm d} \xi.
\end{align}
For $f(\lambda):=\lambda_-^{\gamma}$ and $\gamma\geq 1/2$ (since we are considering $d=1$) the Lieb-Thirring inequality
\begin{align}\label{LT}
\sum_{\lambda\in\sigma_{\rm d}(H_{\alpha V})}\lambda_-^{\gamma}\leq C_{\gamma}\alpha^{1/2+\gamma}\int V_-(x)^{1/2+\gamma}{\rm d} x
\end{align}
captures the semiclassical behavior \eqref{Weyl law}, but is valid for any $\alpha>0$, not only asymptotically. Returning to the complex potential $V=\I\mathbf{1}_{[-1,1]}$ and noticing that $f(z):=\frac{(\im z)^p}{|z|^{1/2}}$ is homogeneous of degree $\gamma=p-1/2$, we observe that \eqref{main result SF} implies that the formal analogue of \eqref{Weyl law} cannot hold, i.e.\ that
\begin{align*}
\liminf_{\alpha\to\infty}\alpha^{-1/2-\gamma}\sum_{z\in\sigma_d(H_{\alpha V})}\frac{(\im z)^{1/2+\gamma}}{|z|^{1/2}}=\infty,
\end{align*}
hence violating Weyl's law \eqref{Weyl law}. The comparison with Weyl's law is formal because $f(H_{\alpha})$ does not make sense in general for a non-normal operator. However, \eqref{main result SF} also shows that the complex analogue of the Lieb-Thirring inequality \eqref{LT},
\begin{align}
\sum_{z\in\sigma_d(H_{\alpha V})}\frac{(\im z)^{1/2+\gamma}}{|z|^{1/2}}\leq C_{\gamma}'\alpha^{1/2+\gamma}\int V_-(x)^{1/2+\gamma}{\rm d} x,
\end{align}
cannot be true, thus disproving the conjectured bound in \cite{MR3016473}.
For a non-selfadjoint (pseudo)-differential operator with analytic symbol and in one dimension the eigenvalues typically lie on a complex curve, hence violating Weyl's law (in terms of complex phase space). Small random perturbations typically restore the Weyl law (in real phase space). The literature on the subject is vast and we merely refer the interested reader to the book of Sjöstrand \cite{MR3931708} for an overview of recent developments. In contrast, a classical result of Markus and Macaev \cite{MR545380} implies that Weyl's law holds for the real part of the eigenvalues, provided the non-selfadjoint perturbation is small.
A slightly different view on the same phenomenon (violation of Weyl's law) is connected with the notion of ``locality". In the semiclassical limit, in the self-adjoint case, each state occupies a volume of $(2\pi/\sqrt{\alpha})$ in phase space $T^*{\mathbb R}$.
Indeed, with $f(\lambda):=\mathbf{1}_{(-\infty,-E]}(H_{\alpha V})$, \eqref{Weyl law f} yields
\begin{align*}
N(H_{\alpha V};(-\infty,-E])=\sqrt{\alpha}\,\mathrm{Vol}(S_{E,V})(1+o(1)),
\end{align*}
where $N(H;\Sigma)$ denotes the number of eigenvalues of an operator $H$ in a set $\Sigma$ and $S_{E,V}:=\set{(x,\xi)\in T^*{\mathbb R}}{\xi^2+V(x)\leq -E}$ is the relevant part of phase space. If we consider a sum of disjoint bumps $V=\sum_{j=1}^N V_j$, say with $V_j(x)=W(x-x_j)$, then $S_{E,V}=N\,S_{E,W}$, and hence
\begin{align*}
\frac{N(H_{\alpha V};(-\infty,-E])}{N(H_{\alpha W};(-\infty,-E])}=N(1+o(1)).
\end{align*}
This means that in the semiclassical limit each bump is responsible for an equal number of eigenvalues. In particular, two distinct realizations of $V$ as a sum of bumps have the same number of eigenvalues, as long as the bumps are disjoint. This feature of locality is also captured by the Lieb-Thirring inequalities since the bound involves the integral linearly.
Our third result shows that this kind of locality can be violated in the non-selfadjoint case. We adopt the same notation $N(H;\Sigma)$ for the number of eigenvalues of $H$ in $\Sigma$ as in the self-adjoint case, but emphasize that these are counted according to their \emph{algebraic} multiplicity.
We consider one sparse one non-sparse (or non-separating) realization of $V$ and denote these by $V_{\rm s}$ and $V_{\rm n}$, respectively. For simplicity, we will consider the same potential as in \cite{boegli2020liebthirring}, i.e.\ the purely imaginary step potential $W=\I|W_0|\mathbf{1}_{[-R_0,R_0]}$ of size $|W_0|$ and width $R_0$. For simplicity we fix these scales to be of order one.
Then $V_{\rm n}$ is the single well of width $R=NR_0$, while $V_{\rm s}$ is a sum of $N$ disjoint wells $W(x-x_j)$ of width $R_0$. We will fix the coupling strength $\alpha$ and consider the limit $N\to\infty$. For the non-sparse operator this is in fact still a semiclassical limit, as can easily be seen by rescaling.
\begin{theorem}\label{thm. violation of locality into}
Let $d=1$, $N\gg 1$ and consider the rectangular set
\begin{align*}
\Sigma:=\set{z\in{\mathbb C}}{C^{-1}\frac{N^2}{\log^2 N}\leq \re z\leq C\frac{N^2}{\log^2 N},\, C^{-1}\leq \im z\leq C},
\end{align*}
where $C$ is a large constant.
Then we have
\begin{align}\label{Vnsp}
N(H_{V_{\rm n}};\Sigma)\gtrsim \frac{N^2}{\log N}.
\end{align}
Moreover, there exists a sequence $(x_j)_j$ such that
\begin{align}\label{Vsp}
N(H_{V_{\rm s}};\Sigma)\lesssim N.
\end{align}
\end{theorem}
The estimates in \cite{boegli2020liebthirring} would be sufficient to prove a lower bound of size $N^{2-\epsilon}$ in \eqref{Vnsp}. Their argument proceeds by approximating the characteristic eigenvalue equation and finding the roots of this equation in an asymptotic regime.
They did not prove that the original equation has nearby roots. This can be done e.g.\ by a contraction mapping argument \cite{stepanenko}. We will give a proof using Rouch\'e's theorem. Note that, by power counting (dimensional analysis), the constants in \eqref{Vnsp}, \eqref{Vsp} only depend on the dimensionless quantity $|W_0|^{1/2}R_0$.
In the selfadjoint case, i.e.\ when $W$ is replaced by $|W_0|\mathbf{1}_{[-R_0,R_0]}$, the number of negative eigenvalues of $V_{\rm n}$ is of order~$N$, in agreement with semiclassics. Also note that, since in one dimension each $H_{V_j}$ has at least one negative eigenvalue, we have $N(H_{V_{\rm s}};{\mathbb R}_-)\asymp N$ in this case. Hence the left hand sides of \eqref{Vnsp} and \eqref{Vsp} are equal in magnitude, which may be seen as a manifestation of locality. However, this locality is violated if one takes into account not only eigenvalues but also resonances. Zworski \cite{MR899652} proved that, for a compactly supported, bounded, complex potential, the number $n(r)$ of resonances $\lambda_j^2$ in a disk $|\lambda_j|\leq r$ asymptotically satisfies
\begin{align}\label{Zworski}
n(r)=\frac{2\,\mathrm{|ch\, supp(V)}|}{\pi}\,r(1+o_V(1))
\end{align}
as $r\to\infty$. Moreover, for any $\epsilon_0>0$, the number of resonances in $|\lambda_j|\leq r$ but outside the sector $\Sigma_{0}$ (see \eqref{def. Sigma_epsilon_0}) is $o(r)$. This and the fact that eigenvalue bounds outside $\Sigma_{0}$ are ``trivial" (in the sense that they can be proved by the same standard estimates as for real potentials, with the provisio that the constants blow up as the implicit constant in \eqref{def. Sigma_epsilon_0} becomes small) motivates us to often restrict attention to the spectral set $\Sigma_{0}$.
The result \eqref{Zworski} was obtained earlier by Regge \cite{MR143532} in some special cases. Different proofs were given by Froese \cite{MR1456597} and Simon \cite{MR1802901}. Froese's proof also works for complex potentials.
Note that eigenvalues are included in the definition of resonances. Formula \eqref{Zworski} can be seen as a Weyl law for resonances but is \emph{nonlocal} as it includes the convex hull of the support of the potential. An obvious corollary of Zworski's formula is that the right hand side of \eqref{Zworski} is an upper bound for the number of eigenvalues $\lambda_j^2$ in the disk $|\lambda_j|\leq r$ (resonances in the upper half plane). For the potential $V_{\rm n}$, taking $r\asymp N/\log N$ and observing that $\mathrm{ch\, supp(V_{\rm n})}\asymp N$, we find that the leading term in \eqref{Zworski} is of order $N^2/\log N$, in agreement with the lower bound \eqref{Vnsp}. Note, however, that the asymptotics are not uniform, i.e.\ the error may depend on $N$ (recall that $V_{\rm n}$ depends on $N$). Korotyaev \cite[Theorem 1.1]{MR3574651} proved an upper bound which yields
\begin{align*}
n(r)\leq \frac{8N}{\pi\log 2}r+\mathcal{O}(N\log N)
\end{align*}
for $r\asymp N/\log N$. Hence, \eqref{Vnsp} shows that, at the scale considered here, \textit{a substantial fraction of resonances are actual eigenvalues}. In the special case of a step potential Stepanenko \cite{stepanenko2020bounds} proved that the total number of eigenvalues is bounded by $N^2/\log N$. The lower bound \eqref{Vnsp} shows that this is sharp up to constants; this was observed independently by Stepanenko \cite{stepanenko}.
In dimension $d\geq 3$ it may of course happen that all the $H_{V_j}$, and thus also~$H_{V_{\rm s}}$, have no eigenvalues at all if $|W_0|^{1/2}R_0$ is small (by the CLR bound), while~$H_{V_{\rm n}}$ has of the order $N^d$ eigenvalues. This is not the kind of phenomenon that takes place in Theorem \ref{thm. violation of locality into}. Indeed, there we allow $|W_0|^{1/2}R_0$ to be of unit size. We also note that in higher dimensions the results on the asymptotics of the resonance counting function are weaker, and there are example of complex potentials with no resonances at all (see Christiansen \cite{MR2225694}). On the other hand, Christiansen \cite{MR2174422} and Christiansen--Hislop \cite{MR2189242}, \cite{MR2648080} proved that $n(r)$ has maximal growth rate $r^d$ for ``most" potentials in certain families.
A final comment regarding the implications of Theorem \ref{thm. violation of locality into}, also related to locality (or the lack thereof), concerns a general observation on Lieb--Thirring type inequalities for complex potentials. It is a fact that all known upper bounds for eigenvalue sums have a superlinear dependence on $N$. For example, in $d=1$, \cite[Theorem 1.3]{MR3717979} yields the bound
\begin{align}\label{Frank d=1 sum}
\sum_j{\rm d}(z_j,{\mathbb R}_+)^a\lesssim \left(\int_{{\mathbb R}}|V|^b{\rm d} x\right)^{c}
\end{align}
with $(a,b,c)=(1,1,2)$. The lower bound \eqref{Vnsp} shows that the power $c$ cannot be decreased, while preserving the homogeneity condition $-2a=(-2b+1)c$. Indeed,
\begin{align*}
\sum_j{\rm d}(z_j,{\mathbb R}_+)^a\geq \sum_{z_j\in\Sigma}{\rm d}(z_j,{\mathbb R}_+)^a\gtrsim \frac{|W_0|^aN^2}{\log N},
\end{align*}
while
\begin{align*}
\left(\int_{{\mathbb R}}|V_{\rm n}|^{b}{\rm d} x\right)^{c}=(2NR_0|W_0|^{b})^{c},
\end{align*}
so that the ratio of the first to the second is bounded below by $(|W_0|^{1/2}R_0)^{-c}\frac{N^{2-c}}{\log N}$, which tends to infinity as $N\to\infty$, for every $c<2$. On the other hand, for sufficient rapid separation of the bumps in $V_{\rm s}$, we can prove Frank's bound \eqref{Frank d=1 sum} with a linear dependence on $N$, at least locally in the spectrum.
\begin{theorem}\label{thm. Frank's 1d bound for sparse intro} Let $V\in \ell^2L^1({\mathbb R})$ and $N\gg 1$.
For any $\eta>0$ there exists a sequence $(x_j)_{j=1}^N$, $x_j=x_j(\eta,\|V\|_{\ell^2L^1})$, such that
\begin{align}\label{N times Frank's bound}
\sum_{\im\sqrt{z_j}>(d+1)\eta}{\rm d}(z_j,{\mathbb R}_+)\lesssim N\sum_{j=1}^N\left(\int_{{\mathbb R}}|V_{j}|{\rm d} x\right)^{2}
\end{align}
\end{theorem}
Our methods could be used to prove similar generalizations of the higher-dimensional results of \cite{MR3717979}, but we will not pursue this.
\subsection*{Notation}
We write $\sigma(H)$, $\rho(H)$ for the spectrum, respectively the resolvent set of a closed linear operator $H$. The free and the perturbed resolvent operators are denoted by $R_0(z)=(-\Delta-z)^{-1}$ and $R_V(z)=(H_V-z)^{-1}$, respectively. We denote by $\mathfrak{S}^p$ the Schatten spaces of order $p$ over the Hilbert space $L^2({\mathbb R}^d)$ and by $\|\cdot\|_p$ the corresponding Schatten norms. We also write $\|\cdot\|=\|\cdot\|_{\infty}$ for the operator norm. To distinguish it from $L^p$ norms of functions we denote the latter by $\|\cdot\|_{L^p}$. We will use the notation $V^{1/2}=V/|V|^{1/2}$ and $\langle x\rangle:=2+|x|$. The statement $a\lesssim b$ means that $|a|\leq C |b|$ for some absolute constant $C$. We write $a\asymp b$ if $a\lesssim b\lesssim a$. If the estimate depends on a list of parameters $\tau$, we indicate this by writing $a\lesssim_{\tau} b$. The dependence on the dimension $d$ and the Lebesgue exponents $p,q$ is usually suppressed. We write $a\ll b$ if $|a|\leq c|b|$ with a small absolute constant $c$, independent of any parameters. By an absolute constant we always understand a dimensionless constant $C=C(d,p,q)$. Here we choose units of length $l$ such that position, momentum and energy have dimensions $l$, $l^{-1}$ and $l^{-2}$, respectively. We chose the branch of the square root $\sqrt{\cdot}$ on ${\mathbb C}\setminus[0,\infty)$ such that $\sqrt{z}\in\mathbb{H}$, where $\mathbb{H}=\set{\kappa\in{\mathbb C}}{\im\kappa>0}$ denotes the upper half plane. The open unit disk in ${\mathbb C}$ is denoted by $\mathbb{D}$.
\subsection*{Acknowledgements}
The author gratefully acknowledges correspondence with Sabine Boegli and comments of Rupert Frank, who pointed out the failure of Weyl's law and the connection with nonlocality. Many thanks also go to St\'ephane Nonnenmacher for useful discussions on resonances and to Alexei Stepanenko for explaining his recent preprint. Special thanks go to Tanya Christiansen for many helpful remarks on a preliminary version of the introduction.
\section{Definitions and preliminaries}\label{section Definitions}
\subsection{Separating and sparse potentials}\label{subsection Separating and sparse potentials}
We consider sparse potentials of the form
\begin{align}\label{separating potential}
V(x)=V_j(x),\quad x\in \Omega_j,
\end{align}
where $j\in [N]=\{1,2,\ldots,N\}$, $N\in {\mathbb N}\cup\{\infty\}$, and $\Omega_j\subset{\mathbb R}^d$ are mutually disjoint (not necessarily bounded) sets. We then set
\begin{align}
L_j:={\rm d}(\Omega_j,\cup_{i\in[N]}\Omega_i\setminus\Omega_j).
\end{align}
We assume that $V_j\in \ell^{p}L^q$, where the norms are defined by
\begin{align*}
\|V\|_{\ell^{p}L^q}:=\big(\sum_{j\in[N]}\|V_j\|_{L^{q}({\mathbb R}^d)}^p\big)^{1/p}
\end{align*}
for $p\in [1,\infty)$ and $\|V\|_{\ell^{\infty}L^q}:=\sup_{j\in [N]}\|V_j\|_{L^{q}({\mathbb R}^d)}$. Here $q$ will be in the range
\begin{align}\label{range q}
q\in [1,\infty] \quad \mbox{if } d=1,\quad q\in (1,\infty]\quad \mbox{if } d=2,\quad q\in[\tfrac{d}{2},\infty] \quad \mbox{if } d\geq 3.
\end{align}
In particular, we have $\|V\|_{L^q}=\|V\|_{\ell^q L^q}$. We sometimes write $V=V(L)$ or $V=V(L,\Omega)$ to emphasize the dependence of $V$ on the sequences $L=(L_j)_{j=1}^N$ or $\Omega=(\Omega)_{j=1}^N$.
\begin{definition}\label{def. separating}
We say that $V=V(L)$ is \emph{separating} if
\begin{align*}
\mathrm{sep}(L,\eta):=\sum_{j\in [N]}\exp(-\eta L_j)<\infty
\end{align*}
for every $\eta>0$. We call $V$ \emph{separating at scale} $\eta^{-1}$ if $\mathrm{sep}(L,\eta)\leq 1$. We say that $V$ is \emph{strongly separating} if $\mathrm{sep}(L,\delta\eta)\lesssim_{\delta}\mathrm{sep}(L,\eta)$ for every $\delta,\eta>0$.
\end{definition}
The constant $\mathrm{sep}(L,\eta)$ only depends on $L$ and not on $V$ itself. We will sometimes abuse terminology and call the sequence $L$ separating. Note that since $\eta L_j$ is dimensionless, $\eta$ has the dimension of inverse length.
We shall always assume that the sequence $L$ is increasing.
\begin{definition}\label{def. sparse}
We say that $V=V(L,\Omega)$ is \emph{sparse} if it is separating, the supports $\Omega_n$ are bounded,
and $\lim_{n\to\infty}\mathrm{diam}(\Omega_n)/L_n=0$.
\end{definition}
Most of our results hold for separating potentials. The strong separation condition is convenient and facilitates some of the proofs.
Typical examples of strongly separating sequences are ($\eta\ll\eta_0$):
\begin{itemize}
\item[a)] It $\eta_0L_k\gtrsim k^{\alpha}$ for $\alpha>0$, then $\mathrm{sep}(L,\eta)\lesssim (\eta/\eta_0)^{-1/\alpha}$.
\item[b)] If $\eta_0L_k\gtrsim\exp(k)$, then $\mathrm{sep}(L,\eta)\lesssim \log(\eta_0/\eta)$.
\item[c)] If $\eta_0L_k\gtrsim\exp(\exp(k))$, then $\mathrm{sep}(L,\eta)\lesssim\log\log(\eta_0/\eta)$.
\end{itemize}
See Subsection \ref{subsection Distribution function} for a proof.
The explicit example used to prove Theorem \ref{thm quant Boegli intro} turns out to be sparse. Note that, by the disjoint supports, the definition \eqref{separating potential} is equivalent to $V=\sum_{j=1}^NV_j$.
\subsection{Comparison with a direct sum}
In Section \ref{Section Comparison between Hdiag and HV} we will compare the two operators
\begin{align}\label{compare the two operators}
H_V=-\Delta+V,\quad
H_{\rm diag}=\bigoplus_{j\in[N]} (-\Delta+V_j).
\end{align}
Note that the point spectrum (eigenvalues) of $H_{\rm diag}$,
\begin{align}\label{spectrum direct sum}
\sigma_{\rm p} (H_{\rm diag})=\bigcup_{j=1}^N\sigma_{\rm p} (H_{V_j})=\bigcup_{j=1}^N\sigma_{\rm p} (H_{W_j}),
\end{align}
is independent of the sequence $L$. We will consider $\sigma_{\rm p} (H_{W_j})$ as part of the data and seek to prove lower bounds on $L_j$ or $|x_i-x_j|$ that imply that $\sigma_{\rm p} (H_{V})$ is close to $\sigma_{\rm p} (H_{\rm diag})$. In fact, we will consider a subset of the point spectrum, the discrete spectrum. We will consider the $V_j$ as given only \emph{up to translations}, i.e.\ we stipulate that
\begin{align}\label{def. Wj}
V_j(x)=W_j(x-x_j),
\end{align}
where $W_j\in \ell^{p}L^q$ contains the origin in its support.
Using the triangle inequality, it is easy to see that $|x_i-x_j|\geq L_j$ for $i\neq j$, and therefore
\begin{align*}
\sup_{i\in[N]}\sum_{j\in [N]\setminus\{i\}}\exp(-\eta |x_i-x_j|)\leq \mathrm{sep}(L,\eta).
\end{align*}
Straightforward arguments also show that
\begin{align}\label{lower bound Li in terms of xi, Omegai}
L_i\geq |x_i-x_j|-\mathrm{diam}(\Omega_i)-\mathrm{diam}(\Omega_j)
\end{align}
for all $j\in [N]\setminus\{i\}$. In particular, for sparse potentials,
\begin{align}\label{lower bound Li in terms of xi, Omegai sparse}
L_i(1+o(1))\geq \sup_{j<i}|x_i-x_j|.
\end{align}
Note that $\mathrm{diam}(\Omega_j)=\mathrm{diam}(\supp W_j)$ is part of the data. We can thus obtain a lower bound for $|x_i-x_j|$ in terms of $L_j$ and vice versa. For this reason we restrict attention to $L_j$ here.
\subsection{Truncations}
For technical reaons, it will turn out to be convenient to consider finite truncations. For $n\in[N]$ we define
\begin{align}
V^{(n)}&:=\sum_{j\in[n]} V_j,\quad
H^{(n)}:=-\Delta+V^{(n)}.\label{def. Vuppern}
\end{align}
\subsection{Abstract Birman-Schwinger principle}
We mostly disregard operator theoretic discussions here and refer e.g.\ to \cite{MR3717979} for the (standard) definition of $H_{V_j}$ as $m$-sectorial operators. The rigorous definition of $H_V$ is a bit more subtle since $V$ need not be decaying. However, a classical construction of Kato \cite{MR0190801} produces a closed extension $H_V$ of $-\Delta+V$ via a Birman-Schwinger type argument. This approach works as soon as one can find point $z_0$ in the resolvent set of $H_0=-\Delta$ at which the Birman-Schwinger operator
\begin{align}\label{def. BS_V(z)}
BS_V(z):=|V|^{1/2}(-\Delta^2-z)^{-1}V^{1/2}
\end{align}
has norm less than one. Such bounds are provided by Lemma \ref{lemma BS bounds}, but to avoid technicalities it is useful to think of the potential as being bounded by a large cutoff (and all of the bounds will be independent of that cutoff).
By iterating the second resolvent identity,
\begin{align*}
R_V(z)=R_0(z)-R_0(z)VR(z),
\end{align*}
it is then easy to see that
\begin{align}\label{resolvent identity}
R_V(z)-R_0(z)=-R_0(z)V^{1/2}(I+BS_V(z))^{-1}|V|^{1/2}R_0(z).
\end{align}
For more background on the abstract Birman-Schwinger principle in a nonselfadjoint setting we refer to \cite{MR2201310}, \cite{MR3717979}, \cite{behrndt2020generalized}, \cite{hansmann2020abstract}.
\subsection{The essential and discrete spectrum}
We briefly recall some facts about the essential and discrete spectrum of a closed operator $H$. There are several inequivalent definitions of essential spectrum for non-selfadjoint operators (but these all coincide for Schrödinger operators with decaying potentials \cite[Appendix B]{MR3717979}). We use the following standard definition.
\begin{align*}
\sigma_{\rm e}(H):=\set{z\in{\mathbb C}}{H-z \mbox{ is not a Fredholm operator}}.
\end{align*}
The discrete spectrum is defined as
\[
\sigma_{\rm d}(H):=\set{z\in{\mathbb C}}{z\mbox{ is an isolated eigenvalue of $H$ of finite multiplicity}}.
\]
Note that, if $H$ is not selfadjoint, then, in general, $\sigma(H)$ is not the disjoint union of $\sigma_{\rm e}(H)$ and $\sigma_{\rm d}(H)$. However, by \cite[Theorem XVII.2.1]{GGK1}, if every connected component of ${\mathbb C}\setminus\sigma_{\rm e}(H)$ contains points of $\rho(H)$, then
\begin{align}\label{sigmae,sigmad}
\sigma(H)\setminus\sigma_{\rm e}(H)=\sigma_{\rm d}(H).
\end{align}
In the situations we consider here \eqref{sigmae,sigmad} will always be true for $H=H_V$ and $H_{\rm diag}$. In fact, Corollary \ref{cor. essential spectrum} tells us that
$\sigma_{\rm e}(H)=[0,\infty)$, just as for decaying potentials.
\section{Universal bounds for separating potentials}
\label{section Quantitative bounds for separating potentials}
In this section we consider $H_V$ as a perturbation of $-\Delta$. We will thus only make assumptions about $V$ and not about $H_{\rm diag}$.
\subsection{Birman--Schwinger analysis}
Since the $V_j$ have mutually disjoint supports, we can write the Birman-Schwinger operator \eqref{def. BS_V(z)} as
\begin{align*}
BS_V(z)=\sum_{i=1}^NBS_{ii}(z)+\sum_{i\neq j}BS_{ij}(z),
\end{align*}
where
\begin{align*}
BS_{ij}(z)=|V_i|^{1/2}R_0(z)V_j^{1/2}.
\end{align*}
The first term is unitarily equivalent to the orthogonal sum
\begin{align*}
BS_{\rm diag}(z):=\bigoplus_{i=1}^NBS_{ii}(z).
\end{align*}
on the Hilbert space $\mathcal{H}\simeq\bigoplus_{i=1}^N \mathcal{H}_i$, where $\mathcal{H}=L^2({\mathbb R}^d)$, $\mathcal{H}_i=L^2(\Omega_i)$. By abuse of notation we will always identify these two Hilbert spaces and the corresponding operators. The off-diagonal contribution is
\begin{align*}
BS_{\rm off}(z):=\sum_{i\neq j}BS_{ij}(z).
\end{align*}
In the following we will use the notation
\begin{equation}\label{def omegap(z)}
\begin{split}
\omega_q(z):=\begin{cases}
|z|^{\frac{d}{2q}-1}\quad&\mbox{if } q\leq q_d,\\
|z|^{-\frac{1}{2q}}{\rm d}(z,{\mathbb R}_+)^{\frac{q_d}{q}-1}\quad&\mbox{if } q\geq q_d,
\end{cases}
\end{split}
\end{equation}
where $q_d=(d+1)/2$. Note that for $z\in\Sigma_{0}$ (see \eqref{def. Sigma_epsilon_0}) we have ${\rm d}(z,{\mathbb R}_+)=|\im z|$. We use the abbreviation
\begin{align}\label{def s(L,z)}
s(L,z):=\mathrm{sep}(L,\im\sqrt{z}/(d+1))
\end{align}
and set $\alpha(q):=2\max(q,q_d)$.
\begin{lemma}\label{lemma BS bounds}
Assume $q$ is in the range \eqref{range q} and $p\in[\alpha(q),\infty]$. Then the following hold.
\begin{itemize}
\item[(i)] For any $i\in [N]$,
\begin{align}\label{BSii bound}
\|BS_{ii}(z)\|_{\alpha(q)}\lesssim \omega_q(z)\|V_i\|_{L^q}.
\end{align}
\item[(ii)] For any $i,j\in [N]$, $i\neq j$,
\begin{align}\label{BSij bound}
\|BS_{ij}(z)\|_{\alpha(q)}\lesssim \exp(-\tfrac{2}{d+1}\im\sqrt{z}\,{\rm d}(\Omega_i,\Omega_j)) \omega_q(z)\|V_i\|_{L^q}^{1/2}\|V_j\|_{L^q}^{1/2}.
\end{align}
\item[(iii)] The diagonal part satisfies
\begin{align}\label{BSdiag bound}
\|BS_{\rm diag}(z)\|_{p}\lesssim \omega_q(z)\|V\|_{\ell^{p} L^q}.
\end{align}
\item[(iv)] The off-diagonal part satisfies
\begin{align}\label{BSoff bound}
\|BS_{\rm off}(z)\|_{\alpha(q)}\lesssim s(L,z)^2\omega_q(z)\|V\|_{\ell^{\infty} L^q},
\end{align}
\item[(v)] The full Birman--Schwinger operator satisfies
\begin{align}\label{BS bound}
\|BS_V(z)\|_{p}\lesssim \omega_q(z)\langle s(L,z)\rangle^2\|V\|_{\ell^{p} L^q}.
\end{align}
\end{itemize}
\end{lemma}
\begin{proof}
It follows from
\begin{align*}
\|BS_{\rm diag}(z)\|_{p}=\big(\sum_{i\in[N]}\|BS_{ii}(z)\|_{p}^p\big)^{1/p}
\end{align*}
that (iii) is a consequence of (i). In view of the embeddings
$\mathfrak{S}^{p_1}\subset \mathfrak{S}^{p_2}$, $\ell^{p_1}\subset \ell^{p_2}$ for $p_1\leq p_2$, (v) follows from (iii) and (iv). Moreover, (iv) follows from (ii) and the triangle inequality, using the estimate ${\rm d}(\Omega_i,\Omega_j)\geq \frac{1}{2}L_i+\frac{1}{2}L_j$ to sum the double series.
The estimate (i) is the same as for the $N=1$ case and follows from known results in the literature: For $q\leq q_d$, from \cite[Theorem 12]{MR3730931} for $d\geq 3$, from \cite[Theorem 4.1]{MR3608659} for $d=2$ and from \cite[Theorem 4]{AAD01} for $d=1$, for $q\geq q_d$ and $d\geq 1$ from \cite[Proposition 2.1]{MR3717979}. The bound (ii) is proved by complex interpolation as in \cite[Theorem 12]{MR3730931} in the case $q\leq q_d$ and in \cite[Proposition 2.1]{MR3717979} for $q\geq q_d$. The only difference is that we include the (second) exponential in the pointwise bound
\begin{align}\label{pointwise bound on complex power of the resolvent}
|(-\Delta-z)^{-(a+\I t)}(x-y)|\leq C_1{\rm e}^{C_2t^2}{\rm e}^{-\im\sqrt{z}|x-y|}|x-y|^{-\frac{d+1}{2}+a}
\end{align}
for $a\in [(d-1)/2,(d+1)/2]$ and $d\geq 2$, see e.g.\ \cite[(2.5)]{MR3865141}. For $d=1$ one can use the explicit formula for the resolvent kernel.
\end{proof}
\begin{remark}\label{remark BS bounds with Davies-Nath}
Using the results of \cite{MR4104544} (or \cite{MR1946184} in one dimension) in the proof of Lemma \ref{lemma BS bounds} we could replace the bounds \eqref{BSii bound}, \eqref{BSij bound} by the following. For $q\leq q_d$ and $i,j\in [N]$,
\begin{align}
\|BS_{ij}(z)\|_{\alpha(q)}\lesssim \exp(-\tfrac{2}{d+1}\im\sqrt{z}\,{\rm d}(\Omega_i,\Omega_j)) |z|^{\frac{d}{2q}-1}F_{V_i,q}^{1/2}(\im\sqrt{z})F_{V_j,q}^{1/2}(\im\sqrt{z}),
\end{align}
where $F_{V,q}(s)$ is the ``Davies--Nath norm"
\begin{align}\label{def. Davies--Nath norm}
F_{V,q}(s):=\left(\sup_{y\in{\mathbb R}^d}\int_{{\mathbb R}^d}|V(x)|^{q}\exp(-s|x-y|){\rm d} x\right)^{\frac{1}{q}}.
\end{align}
This implies the bounds
\begin{align}
\|BS_{\rm diag}(z)\|_{p}&\lesssim |z|^{\frac{d}{2q}-1}\sup_{i\in[N]}F_{V_i,q},\\
\|BS_{\rm off}(z)\|_{\alpha(q)}&\lesssim s(L,z)^2|z|^{\frac{d}{2q}-1}\sup_{i\in[N]}F_{V_i,q},\\
\|BS_{V}(z)\|_{\alpha(q)}&\lesssim \langle s(L,z)\rangle^2|z|^{\frac{d}{2q}-1}\sup_{i\in[N]}F_{V_i,q}.
\end{align}
\end{remark}
\subsection{Norm resolvent convergence}
\begin{lemma}\label{lemma BS bound for half the potential}
Under the assumptions of Lemma \ref{lemma BS bounds} we have
\begin{align}
\||V|^{1/2}R_0(z)\|_{2p}\lesssim |\im z|^{-1} \omega_q(z)\langle s(L,z)\rangle^2\|V\|_{\ell^{p} L^q}^{1/2}.
\end{align}
\end{lemma}
\begin{proof}
The claim readily follows from \eqref{BS bound},
the identity
\begin{align*}
R_0(z)R_0(\overline{z})=\frac{1}{2\im z}(R_0(\overline{z})-R_0(z))
\end{align*}
and a $TT^*$ argument.
\end{proof}
\begin{proposition}\label{prop. nrc}
Under the assumptions of Lemma \ref{lemma BS bounds} the Schrödinger operators $H_{V^{(n)}}$ with truncated potentials converge in norm resolvent sense to $H_V$.
\end{proposition}
\begin{proof}
We first note that all the bounds of Lemma \ref{lemma BS bounds} also hold for $H_{V^{(n)}}$, uniformly in $n$. Since $V^{(n)}$ converges to $V$ in $\ell^pL^q$, it follows that the Birman-Schwinger operator associated to $H_{V^{(n)}}$ converges to $BS_V(z)$. Moreover, by Lemma \ref{lemma BS bound for half the potential}, the operators $|V^{(n)}|^{1/2}R_0(z)$ converge in $\mathfrak{S}^{2p}$-norm. Note that the square root of $V$ is trivial to compute, owing to the disjointness of supports of the $V_j$. We choose $z=\I t$, $t\gg 1$. For such $z$ the norm of the Birman-Schwinger operator is $<1$, hence a Neumann series argument and the resolvent identity \eqref{resolvent identity} yield the claim.
\end{proof}
\begin{proposition}\label{prop. nrc2}
As $s(L,z)\to 0$ for fixed $z$, the Schrödinger operators $H_{V}$ (depending on $L$) converge in norm resolvent sense to $H_{\rm diag}$.
\end{proposition}
\begin{proof}
By \eqref{BSoff bound}, it follows that $BS_{\rm off}(z)\to 0$ in norm as $s(L,z)\to 0$. The remainder of the proof is similar to that of Proposition \ref{prop. nrc}.
\end{proof}
\subsection{Proof of Klaus' theorem}
The proof is an adaptation of Klaus' original argument, which is based on the Birman-Schwinger principle. In the non-selfadjoint setting one may use e.g.\ \cite[Lemma 11.2.1]{MR2359869} as a substitute for Weyl sequences. This yields an easy proof of the inclusion $\sigma_{\rm e}(H_V)\supset [0,\infty)\cup S$. For the converse inclusion we prove an analogue of \cite[Proposition 2.3]{MR700696}.
\begin{proposition}\label{prop. 1 Klaus}
For $V\in \ell^{\infty}L^q$ and $z\in{\mathbb C}\setminus[0,\infty)$ we have
\begin{align*}
\sigma(BS_{\rm diag}(z))=\overline{\bigcup_{i}\sigma(BS_{ii}(z))}.
\end{align*}
\end{proposition}
\begin{proof}
Only the inclusion $\subset$ is nontrivial. The resolvent set of a direct sum of bounded operators $A=\bigoplus_i A_i$ is known to be
\begin{align*}
\rho(A)=\set{\lambda\in \bigcap_i\rho(A_i)}{\sup_i\|(A_i-\lambda)^{-1}\|<\infty}.
\end{align*}
We set $A_i=BS_{ii}(z)$, whence $A=BS_{\rm diag}(z)$. Assume that $\lambda \in {\mathbb C}\setminus \overline{\bigcup_{i}\sigma(BS_{ii}(z))}$. This clearly implies $\lambda\in \bigcap_i\rho(A_i)$. It remains to prove that $\sup_i\|(A_i-\lambda)^{-1}\|<\infty$. By assumption, there exists $\delta>0$ such that ${\rm d}(\lambda,\sigma(A_i))\geq \delta$. By \cite[Theorem 4.1]{MR2047381},
\begin{align*}
\|(A_i-\lambda)^{-1}\|\leq \delta^{-1}\exp(a\|A_i\|_{{\alpha}}^{{\alpha}}\delta^{-{\alpha}}+b),
\end{align*}
where $\alpha=\alpha(q)$ and $a=a(q)$, $b=b(q)$ are constants. Since, by \eqref{BSdiag bound}, $\|A_i\|_{\alpha}\lesssim \|V\|_{\ell^{\infty}L^q}$,
it follows that $\sup_i\|(A_i-\lambda)^{-1}\|<\infty$.
\end{proof}
With Proposition \ref{prop. 1 Klaus} in hand is immediate that \cite[Lemma 2.4]{MR700696} ($K_i=A_i$ in our notation) holds in the generality needed here. The Schatten bound \eqref{BSoff bound} provides a substitute for the compactness arguments \cite[Proposition 2.1-2.2]{MR700696}. For the remainder of the proof one can follow the arguments in \cite{MR700696} verbatim.
\begin{corollary}\label{cor. essential spectrum}
If $V\in \ell^pL^q$ with $q$ in the range \eqref{range q} and $p<\infty$, then we have $S=\emptyset$, i.e.\ $\sigma_{\rm e}(H_V)=[0,\infty)$. The same holds for $H_{\rm diag}$.
\end{corollary}
\begin{proof}
By \eqref{BS bound}, we have that $\lim_{t\to\infty}\|BS_V(\I t)\|=0$, which means that the inverse exists as a bounded operator for $z=\I t$ and $t\gg 1$. Corollary \ref{lemma BS bound for half the potential} implies that $R_0(z)V^{1/2}$ is compact, whence the resolvent difference \eqref{resolvent identity} is compact. The claim now follows by Weyl's theorem \cite[Theorem IX.2.4]{MR929030}.
\end{proof}
\subsection{Magnitude bounds}\label{subsection Magnitude bounds}
The following universal bounds generalize those of Theorem \ref{theorem sparse magnitude bound intro} in the introduction. They are an immediate consequence of \eqref{BS bound}, Remark \ref{remark BS bounds with Davies-Nath} and the Birman--Schwinger principle.
\begin{theorem}
Let $q$ be in the range \eqref{range q}. If $V$ is separating, then every eigenvalue $z$ of $H_V$ satisfies
\begin{align}\label{magnitude bounds 1}
\omega_q(z)^{-1} \langle s(L,z)\rangle^{-2}\lesssim \|V\|_{\ell^{\infty}L^q}
\end{align}
as well as
\begin{align}\label{magnitude bounds 2}
|z|^{q-\frac{d}{2}}\langle s(L,z)\rangle^{-2}\lesssim \sup_{j\in[N]}F_{V_j,q}(\im\sqrt{z}).
\end{align}
\end{theorem}
\begin{remark}
$(i)$ In the case of a single ``bump" ($N=1$) the bound \eqref{magnitude bounds 1} was proved by Frank in \cite{MR3556444} for $q\leq q_d$ and in \cite{MR3717979} for $q>q_d$. In the latter case it was observed that the inequality implies $\im z\to 0$ as $\re z\to +\infty$ for eigenvalues $z$ of $H_V$. More precisely, if we fix the norm of the potential, then
\begin{align*}
|\im z|^{1-\frac{q_d}{q}}\lesssim (\re z)^{-\frac{1}{2q}}.
\end{align*}
In the case $N\neq 1$ \eqref{magnitude bounds 1} implies that the above holds with an additional factor $\langle s(L,z)\rangle^{2}$ on the right. If $L$ grows at least polynomially, $\eta_0 L_k\gtrsim k^{\alpha}$, then we obtain
\begin{align*}
|\im z|^{1-\frac{q_d}{q}+\frac{2}{\alpha}}\lesssim (\re z)^{-\frac{1}{2q}+\frac{1}{\alpha}}\eta_0^{-\frac{2}{\alpha}},
\end{align*}
see Example a) after Definition \ref{def. sparse}. Hence, for sufficiently large $\alpha$ (depending on $q$ and $d$) the exponent of $|\im z|$ remains positive, while that of $\re z$ remains negative, and we still get the conclusion that $\im z\to 0$ as $\re z\to +\infty$.
\noindent $(ii)$ The $N=1$ case of \eqref{magnitude bounds 2} was proved in one dimension by Davies and Nath \cite{MR1946184} and in higher dimensions by the author \cite{MR4104544}. The inequality is similar to \eqref{magnitude bounds 1} for $q> q_d$. Both are relevant for ``long-range" potentials. In the case of the step potential \eqref{magnitude bounds 2} is sharp, while \eqref{magnitude bounds 1} (both for $N=1$) loses a logarithm (see \eqref{single square well norm V} and \eqref{single square well Davies Nath}).
\end{remark}
\section{Determinant bounds}
\begin{assumption}\label{assumption 1}
Let $q$ be in the range \eqref{range q}, $p\in[2\max(q,q_d),\infty)$, $V=V(L)$ strongly separating and $\|V\|_{\ell^{p} L^q}\lesssim 1$.
\end{assumption}
The assumption $\|V\|_{\ell^{p} L^q}\lesssim 1$ is for convenience only and can easily be removed by power counting arguments (Since the estimates are scale-invariant).
\subsection{Upper bounds}
We collect some useful estimates that we will repeatedly use (these follows from \cite[Theorem 9.2]{MR2154153}):
\begin{align}\label{Gamma}
|f(z)|\leq \exp(\mathcal{O}(1)\langle\|BS(z)\|_p\rangle^p),
\end{align}
\begin{align}\label{continuity of det}
|f_{\rm diag}(z)-f_V(z)|\leq \|BS_{\rm off}(z)\|_p\exp(\mathcal{O}(1) \langle\|BS_{\rm diag}(z)\|_p+\|BS_V(z)\|_p\rangle^p),
\end{align}
where $f=f_{\rm diag}$ or $f_V$ and $BS(z)=BS_{\rm diag}(z)$ or $BS_V(z)$.
Formulas \eqref{Gamma}, \eqref{continuity of det}, together with the bounds of Lemma \ref{lemma BS bounds} motivates the following definitions,
\begin{align}\label{def. phi_p, psi_p}
\psi_p(t):=\exp(\mathcal{O}_p(1)\langle t\rangle^p),\quad\varphi_p(t):=t\exp(\mathcal{O}_p(1)\langle t\rangle^p),\quad t\geq 0,
\end{align}
and for $z\in {\mathbb C}\setminus[0,\infty)$,
\begin{equation}\label{def. Phi_p, Psi_p operator}
\begin{split}
\Psi_{p,q}(z)&:=\psi_p(\omega_q(z)\|V\|_{\ell^{p} L^q}),\\
\Phi_{p,q}(z)&:=\varphi_p(\omega_q(z)\|V\|_{\ell^{p} L^q}),\\
\Phi_{p,q}(L,z)&:=\varphi_p(s(L,z)^2\omega_q(z)\|V\|_{\ell^{p} L^q}),
\end{split}
\end{equation}
where we suppressed the dependence on $V$. We recall that $\omega_q(z)$ and $s(L,z)$ were defined in \eqref{def omegap(z)} and \eqref{def s(L,z)}.
The constant $O_p(1)$ is allowed to vary from one occurence to another. Thus, for example, the inequality $\varphi_p(t)\psi_p(t)\lesssim \varphi_p(t)$ holds, but $\psi_p(t)\lesssim 1$ need not be true.
\subsection{Lower bounds away from zeros}
The following lemma can be considered as one of the key technical results. To state it we introduce the notation
\begin{align}
M_{p,q}(z)&:=\frac{\langle z\rangle}{|\im z|}\left(\frac{\langle z\rangle}{|z|}\right)^{5p(\frac{q_d}{q}-1)_-+8}\langle\omega_q(z)\rangle^p,\label{def. Mzeta}\\
M_{p,q}(L,z)&:=M_{p,q}(z)\langle s(L,\left(\frac{|z|}{\langle z\rangle}\right)^{5}z)\rangle^{2p},\label{def. MLzeta}
\end{align}
Moreover, we set
\begin{align}\label{def delta(z)}
\delta_H(z):=\min(\tfrac{1}{2},{\rm d}(z,\sigma(H))).
\end{align}
In the following $H$ denotes either $H_{\rm diag}$ or $H_V$. We then write $\delta_{\rm diag}(z):=\delta_{H_{\rm diag}}(z)$ and $\delta_V(z):=\delta_{H_V}(z)$.
\begin{lemma}\label{lemma Lower bound on regularized determinants}
Suppose Assumption \ref{assumption 1} holds. Then for all $z\in \Sigma_{0}$,
\begin{align}
|f_{\rm diag}(z)|^{-1}&\leq \exp(\mathcal{O}(1) M_{p,q}(z)\log\frac{1}{\delta_{\rm diag}(z)}),\label{eq. Lower bound on regularized determinants fdiag}\\
|f_{V}(z)|^{-1}&\leq \exp(\mathcal{O}(1) M_{p,q}(L,z)\log\frac{1}{\delta_V(z)}).\label{eq. Lower bound on regularized determinants fV}
\end{align}
\end{lemma}
\begin{proof}
In the following, $f$ denotes either $f_{\rm diag}$ or $f_V$. We are going to apply Lemma~\ref{lemma wedge} with parameters (in the notation of that lemma)
\begin{align}
&r_1^2=c| z|,\quad r_2^2=c^2| z|,\quad r_3^2=c^3| z|,\\
&R_1^2=c^{-1}| z|,\quad R_2^2=c^{-2}\langle z\rangle,\\
&2\varphi_1=\epsilon|\arg( z)|,\quad 2\varphi_2=\epsilon^2|\arg( z)|,\quad 2\varphi_3=c\epsilon^2|\arg( z)|,\\
&\theta_1=\pi-\varphi_1,\quad \theta_2=\pi-\varphi_2,\quad \theta_3=\pi-\varphi_3,
\end{align}
and with $U_j$, $j=1,2,3$, the wedges defined in \eqref{def. wedge}. Here $c\ll 1$ is a small absolute constant and $\epsilon=\epsilon(z)$ is a small parameter that will be chosen momentarily.
Note that $|\arg( z)|\ll 1$ since $ z\in\Sigma_{0}$ and thus $\sin(2\varphi_j)\asymp \varphi_j\asymp \tan(2\varphi_j)$, which will be used repeatedly in the proof. We will now verify the assumptions of Lemma~\ref{lemma wedge}. The first condition in \eqref{condition lemma wedge} is satisfied by definition. Using that, for $j=1,2$,
\begin{align*}
{\rm d}(\partial U_j,\partial U_3)
=\sin(2\varphi_j-2\varphi_3)r_j^2\asymp (c\epsilon)^{j}|\im z|,
\end{align*}
we find that the second and third condition in \eqref{condition lemma wedge} become
\begin{align*}
c\epsilon\ll (c^2(|z|/\langle z\rangle)^{1/2})^{\frac{2\pi}{\theta_3-\varphi_3}+2},\quad c^3\epsilon|\im z|/\langle z\rangle (c^2(|z|/\langle z\rangle)^{1/2})^{\frac{2\pi}{\theta_3-\varphi_3}}\ll 1,
\end{align*}
respectively, which means that the third condition is trivially satisfied. Since $c^2(|z|/\langle z\rangle)^{1/2}\ll 1$ and $\frac{2\pi}{\theta_3-\varphi_3}\leq 3$, the second condition is satisfied if we choose e.g.\ $\epsilon= c^{10}(|z|/\langle z\rangle)^{5/2}$, which we do.
We will next show that, for a suitable choice of $z_2\in U_2$,
\begin{equation}\label{max less than omegazeta}
\max_{w\in U_3}\log|f(w)|-\log|f( z_2)|\lesssim_c (\langle s(L,\epsilon^2 z)\rangle^{2\nu}\epsilon^{-2(\frac{q_d}{q}-1)_-}\langle\omega_q( z)\rangle)^p,
\end{equation}
where $\nu=0$ if $f=f_{\rm diag}$ and $\nu=1$ if $f=f_{V}$.
Since in the present case
\begin{align*}
\frac{R_2^2}{{\rm d}(\partial U_2,\partial U_3)} \left(\frac{R_2}{r_2}\right)^{\frac{2\pi}{\theta_3-\varphi_3}}\lesssim c^{-28}\frac{\langle z\rangle}{|\im z|}\left(\frac{\langle z\rangle}{| z|}\right)^{8},
\end{align*}
we see that \eqref{eq. lemma wedge} holds.
Lemma~\ref{lemma wedge} thus implies that \eqref{eq. Lower bound on regularized determinants fdiag}, \eqref{eq. Lower bound on regularized determinants fV} hold. To prove~\eqref{max less than omegazeta} we first observe that, by the maximum principle, $|f|$ attains its maximum on the boundary of the wedge $U_3$. We estimate this on the boundary component corresponding to the ray $w\rho^2{\rm e}^{2\I\varphi_3}$, $\rho>r_3$, the estimates for the other two boundary components being similar. By \eqref{Gamma} and the bounds of Lemma \ref{lemma BS bounds}, this maximum is bounded by the right hand side of \eqref{max less than omegazeta}, where we used that $\sup_{\rho>r_3}\omega_q(\rho^2{\rm e}^{2\I\varphi_3})\lesssim_c \epsilon^{-2(\frac{q_d}{q}-1)_-}\omega_q(z)$ and that $L$ is strongly separating. This proves \eqref{max less than omegazeta} for the first term on the left. The other part follows by selecting, for instance, $z_2=\tfrac{\I}{4}R_2^2$ and using the estimates (similar to \eqref{continuity of det}, see \cite[Theorem 9.2]{MR2154153})
\begin{align*}
|f_{\rm diag}(z_2)-1|\lesssim \Phi_{p,q}(z_2),\quad |f_{V}(z_2)-1|\lesssim \Phi_{p,q}(L,z_2),
\end{align*}
where we once again used Lemma \ref{lemma BS bounds} and $\omega_q(z_2)\lesssim \omega_q(z)$. Note that in the case of $f_V$ we can absorb the factor $\langle s(L,z_2)\rangle^2$ in the definition of $\Phi_{p,q}(L,z_2)$ into the $O(1)$ term since $\im\sqrt{z_2}\gtrsim 1$ and $L$ is strongly separating. Since $\omega_q(z_2)\ll 1$ for $R_2\gg 1$ (which is true whenever $c\ll 1$), it follows that $|f(z_2)|\geq 1/2$ for $c$ sufficiently small. This finishes the proof of \eqref{max less than omegazeta}.
\end{proof}
\subsection{Upper bound on the resolvent away from eigenvalues}
As a consequence of Lemma \ref{lemma Lower bound on regularized determinants} we also obtain an upper bound for the resolvent of $H_V$ away from the spectrum. The idea is to use the following infinite-dimensional analogue of Caramer's rule (see \cite[(7.10)]{MR482328}),
\begin{align}\label{Cramers rule}
\|(I+BS(z))^{-1}\|\leq \frac{\exp(\mathcal{O}(1)\|BS(z)\|_{p}^p)}{|f(z)|}.
\end{align}
\begin{proposition}\label{prop. upper bound for the resolvent away from the spectrum}
Suppose Assumption \ref{assumption 1} holds. Then for all $z\in \Sigma_{0}$,
\begin{align}
\|(H_{\rm diag}-z)^{-1}\|\leq \exp(\mathcal{O}(1) M_{p,q}(z)\log\frac{1}{\delta_{\rm diag}(z)}),\label{upper bound for the resolvent Hdiag}\\
\|(H_V-z)^{-1}\|\leq \exp(\mathcal{O}(1) M_{p,q}(L,z)\log\frac{1}{\delta_V(z)}),\label{upper bound for the resolvent HV}
\end{align}
where $M_{q}(z)$, $M_{q}(L,z)$ are given by \eqref{def. Mzeta}.
\end{proposition}
\begin{proof}
In view of the trivial bound
\begin{align*}
\|(-\Delta-z)^{-1}\|\leq |\im z|^{-1},
\end{align*}
the claim follows from Lemma \ref{lemma upper bound for the BS away from the spectrum}, the resolvent identity \eqref{resolvent identity}, Corollary \ref{lemma BS bound for half the potential} and the fact that $|\im z|^{-1}\lesssim M_{p,q}(z)$.
\end{proof}
\begin{lemma}\label{lemma upper bound for the BS away from the spectrum}
The operator norms of $(I+BS_{\rm diag}(z))^{-1}$ and $(I+BS_{V}(z))^{-1}$ are bounded by the right hand side of \eqref{upper bound for the resolvent Hdiag} and \eqref{upper bound for the resolvent HV}, respectively.
\end{lemma}
\begin{proof}
We only prove the claim for $BS_{\rm diag}(z)$; the other part is similar. By \eqref{Cramers rule}, \eqref{BS bound} and \eqref{def. Phi_p, Psi_p operator}, we have $\|(I+BS_{\rm diag}(z))^{-1}\|\lesssim|f_{\rm diag}(z)|^{-1}\Psi_{p,q}(z)$. Since $\Psi_{p,q}(z)\lesssim M_{p,q}(z)$, Lemma \ref{lemma Lower bound on regularized determinants} implies that the latter is bounded by \eqref{upper bound for the resolvent Hdiag}.
\end{proof}
\section{Comparison between $H_{\rm diag}$ and $H_V$}\label{Section Comparison between Hdiag and HV}
\subsection{Ghershgorin type upper bounds}
We record the following Ghershgorin type bound.
We temporarily restore the norm of the potential and define
\begin{align}
\omega_{q,i}(z):=\omega_q(z)\|V_i\|_{L^q},
\end{align}
and $M_{p,q,i}(z)$ is defined by \eqref{def. Mzeta} with $\omega_{q}(z)$ replaced by $\omega_{q,i}(z)$.
\begin{proposition
Under Assumption \ref{assumption 1} the discrete spectrum of $\sigma(H_{V})$ in $\Sigma_0$ is contained in the set
\begin{align}\label{Ghershgorin set}
\bigcup_{i\in[N]}\set{z\in{\mathbb C}}{\im\sqrt{z}L_i-\log \langle s(L,z)\rangle\lesssim - M_{p,q,i}(z)\log\delta_{H_{V_i}}(z)+\log \omega_{q,i}(z)}.
\end{align}
\end{proposition}
\begin{proof}
Assume first that $N<\infty$, and consider the Hilbert spaces $\mathcal{H}_n=L^2(\Omega_n)$, $\mathcal{H}=\bigoplus_{n\in[N]}\mathcal{H}_n$, with operators $A_{ij}=\delta_{ij}I_{\mathcal{H}}+BS_{ij}(z)$ and $A=(A_{ij})_{i,j=1}^{N}$. Applying the Gershgorin theorem for bounded block operator matrices due to Salas \cite{MR1685637} (see also \cite[Theorem 1.13.1]{MR2463978}) yields
\begin{align*}
\sigma(A)\subset\bigcup_{i=1}^N\set{\lambda\in{\mathbb C}}{\|(A_{ii}-\lambda)^{-1}\|^{-1}\leq \sum_{j\in[N]\setminus\{i\}}\|A_{ij}\|}.
\end{align*}
Note that here we are using the convention that $\|(A_{ii}-\lambda)^{-1}\|=\infty$ if $\lambda\in \sigma(A_{ii})$.
By the Birman--Schwinger principle, this implies that
\begin{align*}
\sigma(H_V)\subset\bigcup_{i=1}^N\set{z\in{\mathbb C}}{\|(I_{\mathcal{H}}+BS_{ii}(z))^{-1}\|^{-1}\leq \sum_{j\in[N]\setminus\{i\}}\|BS_{ij}(z)\|}.
\end{align*}
Again, we include the spectrum of $A_{ii}$ in the set on the right.
By Lemma \ref{lemma upper bound for the BS away from the spectrum}, we have
\begin{align*}
\|(I_{\mathcal{H}}+BS_{ii}(z))^{-1}\|\leq \exp(\mathcal{O}(1) M_{p,q,i}(z)\log\frac{1}{\delta_{H_{V_i}}(z)}).
\end{align*}
On the other hand, by \eqref{BSij bound} and the strong separation property,
\begin{align*}
\sum_{j\in[N]\setminus\{i\}}\|BS_{ij}(z)\|\lesssim s(L,z)\exp(-\tfrac{1}{d+1}\im\sqrt{z}L_i)\omega_q(z)
\end{align*}
The claim for $N<\infty$ follows. Similarly, it follows for the truncated operators $H^{(n)}$. Since the set \eqref{Ghershgorin set} is independent of $n$ the claim for the case $N=\infty$ then follows from the norm resolvent convergence of the truncated operators (Proposition \ref{prop. nrc}).
\end{proof}
\subsection{Lower bounds: Qualitative results}
In the following we establish criteria on the sequence $(L_j)_j$ that guarantee proximity of $\sigma_{\rm d} (H_{V})$ to $\sigma_{\rm d} (H_{\rm diag})$ in various regions of the spectral plane.
Let us fist discuss some standard facts from perturbation theory (\cite[Chaper IV]{Ka}). It is well knows that the spectrum of a closed operator $H$ is upper semicontinuous under perturbations, and the same is true for each separated part of the spectrum \cite[Theorem 3.16]{Ka}. Moreover, a \emph{finite system} of eigenvalues $\{\zeta_1,\ldots,\zeta_n\}$ changes continuously, just as in the finite-dimensional case. This follows from the fact that the Riesz projection
\begin{align}\label{Riesz projection}
\frac{1}{2\pi\I}\int_{\Gamma}(\zeta-H)^{-1}{\rm d}\zeta
\end{align}
is continuous in $H$ in the uniform topology (in the sense of generalized convergence of operators). Here $\Gamma$ is a closed contour (a piecewise smooth curve) in the resolvent set of $H$ and encircling the eigenvalues $\zeta_1,\ldots,\zeta_n$ (and no other point of the spectrum) once in the counterclockwise direction. Then
\begin{align}\label{trace Riesz proj.}
\frac{1}{2\pi\I}\Tr \int_{\Gamma}(\zeta-H)^{-1}{\rm d}\zeta=n.
\end{align}
Here we consider a finite system of eigenvalues of $H_{\rm diag}$ in some compact subset $\Sigma\Subset{\mathbb C}\setminus[0,\infty)$. By Corollary \ref{cor. essential spectrum}, Assumption \ref{assumption 1} implies that each point in $\Sigma$ is either in the resolvent set or a discrete eigenvalue of $H_{\rm diag}$. By compactness, $\Sigma\cap \sigma(H_{\rm diag})$ is a finite set.
We then have
\begin{align}\label{def delta0(U)}
\delta_0(\Sigma):=\min\set{{\rm d}(\zeta,\sigma(H_{\rm diag})\setminus\{\zeta\})}{\zeta \in \Sigma\cap \sigma(H_{\rm diag})}>0.
\end{align}
For $\delta\in (0,\tfrac{1}{3}\min(1,\delta_0(\Sigma)))$ we set
\begin{align}\label{def Udelta Gammadelta}
U_{\delta}:=\bigcup_{\zeta \in \Sigma\cap \sigma(H_{\rm diag})}D(\zeta,\delta),\quad \Gamma_{\delta}:=\partial U_{\delta}.
\end{align}
In general it is hard to determine $\delta_0(\Sigma)$, but we still have $\Gamma_{\delta}\subset\rho(H_{\rm diag})$ for generic~$\delta$. This is all that is needed for a lower bound on the number of eigenvalues in $U_{\delta}$.
The norm resolvent convergence (Proposition \ref{prop. nrc2}) implies the following proposition. In Subsection \ref{subsection Proof of Prop. comparison principle first version} we will give an alternative proof using the argument principle.
Let us first state our assumptions for the remainder of this section.
\begin{assumption}\label{assumption 2}
Let $\Sigma \subset\Sigma_0\cap{\mathbb C}\setminus[0,\infty)$ be a compact subset, let $\delta_0(\Sigma),U_{\delta},\Gamma_{\delta}$ be defined by \eqref{def delta0(U)}, \eqref{def Udelta Gammadelta} and let $\delta\in(0,\delta_0(\Sigma))$.
\end{assumption}
\begin{proposition}\label{Prop. comparison principle first version}
Suppose Assumptions \ref{assumption 1}, \ref{assumption 2} hold. Then for any $\delta\in (0,\delta_0(\Sigma))$ there exists a constant $C=C(\delta,\Sigma)$ such that, if $s(L,\zeta)\leq 1/C$, then
\begin{align*}
N(H_V;U_{\delta})=N(H_{\rm diag};U_{\delta}).
\end{align*}
\end{proposition}
\subsection{Argument principle}\label{subsection argument principle}
The argument in the previous subsection involved compactness and continuity and is obviously non-quantitative. The issue is of course the need for a quantitative estimate of the resolvent on the curve $\Gamma_{\delta}$. We will prove such estimates in Proposition~\ref{prop. upper bound for the resolvent away from the spectrum}. Here we argue in a slightly different (albeit closely related) manner. We will use the regularized Fredholm determinants (see for instance \cite[IV.2]{MR0246142}, \cite[Chapter 9]{MR2154153} or \cite[XI.9.21]{MR1009163})
\begin{align}
f_{\rm diag}(z)&:=\Det_{p}(I+BS_{\rm diag}(z)),\label{f}\\
f_V(z)&:=\Det_{p}(I+BS_V(z)),\label{fdiag}
\end{align}
where $p\in [2\max(q,q_d),\infty)$ is assumed to be an integer. The main property that we will use is that the $f_{\rm diag},f_V$ are analytic functions in ${\mathbb C}\setminus[0,\infty)$ and have zeros (counted with multiplicity) exactly at the eigenvalues of $H_{\rm diag}$, $H_V$, respectively. Moreover, by the generalized argument principle (see e.g.\ \cite[Theorem 3.2]{MR3717979} or \cite[Theorem 6.7]{behrndt2020generalized}),
\begin{align}\label{argument principle}
N(H;U_{\delta})=\frac{1}{2\pi\I}\int_{\Gamma_{\delta}}\frac{{\rm d}}{{\rm d} z}\log f(\zeta){\rm d} \zeta,
\end{align}
where $H=H_{\rm diag}$ or $H_V$ and $f=f_{\rm diag}$ or $f_V$.
This suggests a comparison between $H_{\rm diag}$ and $H_{V}$ via Rouch\'e's theorem (see e.g.\ \cite{MR1001877} for related ideas).
We set
\begin{align}\label{def r}
r_{\delta}:=\sup_{z\in \Gamma_{\delta}}\frac{|f_{\rm diag}(z)-f_V(z)|}{|f_{\rm diag}(z)|}.
\end{align}
We will show that $r_{\delta}<1$ if $\max_{j\in[n]}s(L,\zeta_j)$ is sufficiently small. Rouch\'e's theorem then asserts that $f_{\rm diag}$ and $f$ have the same number of zeros in $U_{\delta}$.
\subsection{Alternative proof of Proposition \ref{Prop. comparison principle first version}}\label{subsection Proof of Prop. comparison principle first version}
Without loss of generality we may assume that $\Sigma$ contains exactly one eigenvalue $\zeta$ of $H_{\rm diag}$.
We are going to use Lemma \ref{lemma appendix simplest version}. For this purpose we set $U_1=U_{\delta}$ and let $U_2\subset{\mathbb C}\setminus[0,\infty)$ be a precompact simply connected neighborhood of $U_1$ containing a point $\zeta_2\notin \sigma(H_{\rm diag})$. This is possible by \eqref{sparse magnitude bound} applied to $V_j$ (i.e.\ with $N=1$) since we can take $\zeta_2=-A$, where $A\gg 1$. By \eqref{continuity of det} we find that
\begin{align}\label{fdiag-f}
\sup_{z\in \Gamma_{\delta}}|f_{\rm diag}(z)-f_V(z)|
\leq C_1\,s(L,\zeta)^2,
\end{align}
where $C_1=C_1(\delta,\Sigma)$. We take $A$ so large that
\begin{align}\label{Upsilon}
\Phi_{p,q}(z_0)\leq\tfrac{1}{2}.
\end{align}
This is possible since $\lim_{A\to\infty}\omega_q(-A)=0$ by \eqref{BSdiag bound}. By Lemma \ref{lemma appendix simplest version} there exists a constant $C_2=C_2(\delta,\Sigma)$ such that
\begin{align}\label{lower bound fdiag simplest}
\log|f_{\rm diag}(z)|\geq -C_2\quad \mbox{for all}\quad z\in \Gamma_{\delta}.
\end{align}
Here we used that $\log|f_{\rm diag}(z_0)|\geq -\log 2$, which follows from \eqref{Upsilon}, \eqref{BSdiag bound} and \eqref{Gamma}.
Combining \eqref{fdiag-f} and \eqref{lower bound fdiag simplest}, we infer that $r_{\delta}<1$ if $s(L,\zeta)$ is sufficiently small.
\subsection{Lower bounds: Quantitative results}
In the following we establish quantitative versions of Proposition \ref{Prop. comparison principle first version}.
We return to estimating the quantity $r_{\delta}$ in~\eqref{def r} featuring in Rouch\'e's theorem.
\begin{lemma}
Suppose Assumptions \ref{assumption 1}, \ref{assumption 2} hold. Then
\begin{align}\label{upper bound formula for r improved}
r_{\delta}\leq \max_{\zeta\in\Sigma} s(L,\zeta)^2\exp\big(\mathcal{O}(1) \langle s(L,\zeta)\rangle^2 M_{p,q}(\zeta)\log\frac{1}{\delta}\big),
\end{align}
where $\delta_{H_{\rm diag}}(\zeta)$, $s(L,\zeta)$, $M_{p,q}(\zeta)$ are given by \eqref{def delta(z)}, \eqref{def s(L,z)}, \eqref{def. Mzeta}, respectively.
\end{lemma}
\begin{proof}
Again we may assume that $\Sigma$ contains exactly one eigenvalue $\zeta$ of $H_{\rm diag}$, so that $U_{\delta}=D(\zeta,\delta)$ and $\Gamma_{\delta}=\partial D(\zeta,\delta)$.
We clearly have $\delta(z)=\delta$ and $\omega_q(z)\asymp \omega_q(\zeta)$ for $z\in\Gamma_{\delta}$.
It is easy to see that \eqref{continuity of det} and Lemma \ref{lemma BS bounds} imply
\begin{align}
\sup_{z\in\Gamma_{\delta}}|f_{\rm diag}(z)-f_V(z)|\lesssim \Phi_{p,q}(L,\zeta)\Psi_{p,q}(\zeta).
\end{align}
We have also used that $L$ is strongly separating, hence $s(L,z)\lesssim s(L,\zeta)$.
In order to estimate $r_{\delta}$ it remains to bound $|f_{\rm diag}(z)|$ from below using \eqref{eq. Lower bound on regularized determinants fdiag}.
\end{proof}
As an immediate corollary we obtain an improvement of Proposition
\ref{Prop. comparison principle first version}.
\begin{proposition}\label{Prop. comparison principle improvement}
Suppose Assumptions \ref{assumption 1}, \ref{assumption 2} hold. Then
\begin{align*}
N(H_V;U_{\delta})=N(H_{\rm diag};U_{\delta}),
\end{align*}
provided that $L$ is so large that $r_{\delta}<1$ in \eqref{upper bound formula for r improved}.
\end{proposition}
\section{From quasimodes to eigenvalues}\label{section From quasimodes to eigenvalues}
\subsection{Existence of a single eigenvalue}
We record a useful corollary of Proposition \ref{prop. upper bound for the resolvent away from the spectrum}. The proof is obvious.
\begin{corollary}\label{corollary from quasimodes to eigenvalues}
Suppose Assumption \ref{assumption 1} holds.
Assume that there is a normalized $\psi\in L^2({\mathbb R}^d)$ such that $$\|(H_V-\zeta)\psi\|\leq \epsilon,$$ where $\epsilon^{-1}$ is larger than the right hand side of \eqref{upper bound for the resolvent HV} at $z=\zeta$.
Then $\sigma_{\rm d}(H_V)\cap D(\zeta,\delta))\neq\emptyset$.
\end{corollary}
\subsection{Existence of a sequence of eigenvalues}
If instead of a single quasi-eigenvalue we consider a sequence $(\zeta_j)_j$ with $\lim_{j\to\infty}\im\sqrt{\zeta_j}=0$, an across-the-board assumption like the one in Corollary \ref{corollary from quasimodes to eigenvalues} is not feasible since the right hand side of \eqref{upper bound for the resolvent HV} at $z=\zeta_n$ tends to infinity as $n\to\infty$. One possible solution would be to modify the previous arguments and select $L$ as a function of the sequence $(\zeta_j)_j$. We will follow a similar, albeit slightly different strategy which we find more intuitive. It is also closer in spirit to the inductive argument in \cite{MR3627408}, which is based on strong resolvent convergence. Once more, the approach we will outline can be viewed as a quantitative version of that method.
The strategy will be to first construct quasimodes of $H_V$ in a direct way (Lemma \ref{lemma quasimodes Hdiag to HV}) and then use Corollary \ref{corollary from quasimodes to eigenvalues} to obtain existence of eigenvalues. The quasi-eigenvalues and quasimodes will be actual eigenvalues and eigenfunctions of $H_{\rm diag}$. We introduce a sequence of scales $\varepsilon_j$ and $a_j$, where $\varepsilon_j$ has dimension of energy and $a_j$ has dimension of length, and assume that the eigenfunctions $\psi_j$ corresponding to $\zeta_j$ decay exponentially away from $\Omega_j$ in such a way that
\begin{align}\label{assumption exponential decay}
\|V_i\psi_j\|\leq C_q a_j^{-d/\widetilde{q}}\exp(-c_0\,\im\sqrt{\zeta_j}\,{\rm d}\,(\Omega_i,\Omega_j))\|V_i\|_{L^{\widetilde{q}}},
\end{align}
where $\widetilde{q}\geq 2$ and $c_0>0$. In the following applications we can take $c_0=1$. We will then choose $L$ such that
\begin{align}\label{choice of Lj}
\im\sqrt{\zeta_n} L_n\geq C\log\left(n\log^2\langle n\rangle\sup_{j\in[n]}\varepsilon_j^{-1}a_j^{-d/\widetilde{q}}\sup_{i\in[n]}\|V_i\|_{L^{\widetilde{q}}}\right),
\end{align}
where $C=C(d,\widetilde{q})$ is a large constant.
\begin{lemma}\label{lemma quasimodes Hdiag to HV}
Assume that $V\in \ell^{\infty}L^{\widetilde{q}}$ for some $\widetilde{q}\geq 2$ and that
$\zeta_j$ are eigenvalues of $H_{V_j}$ with normalized eigenfunctions $\psi_j$ satisfying \eqref{assumption exponential decay}. Then there exists an absolute constant $C=C(d,q)$ such that if $V(L)$ is separating and satisfies \eqref{choice of Lj}, then $H_V$ has a sequence of normalized quasimodes $\psi_j$,
\begin{align*}
\|(H_{V}-\zeta_j)\psi_j\|\leq \varepsilon_j.
\end{align*}
\end{lemma}
\begin{remark}
Lemma \ref{lemma quasimodes Hdiag to HV} could be seen as a quantitative version of Lemma 2 in \cite{MR3627408}.
\end{remark}
\begin{proof}
Let $\psi_j$ be the eigenfunctions of $H_{V_j}$ corresponding to $\zeta_j$, i.e.\ $(H_{V_j}-\zeta_j)\psi_j=0$. For $n\in[N]$ we make the following (stronger) induction hypothesis $P(n)$:
\begin{align}\label{P(n)}
\|(H^{(n)}-\zeta_j)\psi_j\|\leq \varepsilon_j\big(1-\frac{1}{\log\langle n\rangle}\big),\quad j\in [n]
\end{align}
(recall \eqref{def. Vuppern} for the definition of $H^{(n)}$).
The base case $n=1$ is true by assumption.
Assume now that $P(n-1)$ holds. By the exponential decay \eqref{assumption exponential decay},
\begin{align*}
\|(H^{(n)}-\zeta_{n})\psi_{n}\|&\leq \sum_{j=1}^{n-1}\|V_j\psi_{n}\|\leq nC_q a_n^{-d/\widetilde{q}}\exp(-c_0\eta_n\,L_n)\sup_{j\in[n]}\|V_j\|_{L^{\widetilde{q}}},
\end{align*}
where we have set $\eta_n:=\im\sqrt{\zeta_n}$.
Moreover, by induction hypothesis, for $j\in [n-1]$,
\begin{align*}
\|(H^{(n)}-\zeta_{j})\psi_{j}\|&\leq \|(H^{(n-1)}-\zeta_{j})\psi_{j}\|+\|V_{n}\psi_{j}\|\\
&\leq \varepsilon_j\big(1-\frac{1}{\log\langle n-1\rangle}\big)+C_q a_j^{-d/\widetilde{q}}\exp(-c_0\eta_j\,L_n)\|V_n\|_{L^{\widetilde{q}}}.
\end{align*}
Hence $P(n)$ would hold if $L_n$ satisfied the estimates
\begin{align}
&nC_q a_n^{-d/\widetilde{q}}\exp(-c_0\eta_n\,L_n)\sup_{j\in[n]}\|V_j\|_{L^{\widetilde{q}}}\leq \varepsilon_n,\label{accomplished our goal 1}\\
&C_q a_j^{-d/\widetilde{q}}\exp(-c_0\eta_j\,L_n)\|V_n\|_{L^{\widetilde{q}}}\leq \varepsilon_j\big(\frac{1}{\log\langle n-1\rangle}-\frac{1}{\log\langle n\rangle}\big),\label{accomplished our goal 2}
\end{align}
for $j\in [n]$.
By the mean value theorem
\begin{align*}
\big(\frac{1}{\log\langle n-1\rangle}-\frac{1}{\log\langle n\rangle}\big)\gtrsim \frac{1}{n\log^2\langle n\rangle},
\end{align*}
and it is easy to check that \eqref{accomplished our goal 1}--\eqref{accomplished our goal 2} are satisfied for the choice \eqref{choice of Lj}. This completes the induction step. For $N<\infty$ the claim now follows from \eqref{P(n)} with $n=N$. Now consider the case $N=\infty$.
Since $L$ is separating,
\begin{align*}
\lim_{n\to\infty}\|(H_V-H^{(n)})\psi_j\|^2&=\lim_{n\to\infty}\|(V-V^{(n)})\psi_j\|^2=\lim_{n\to\infty}\sum_{k=n+1}^{\infty}\|V_k\psi_j\|^2\\
&\leq a_j^{-d/\widetilde{q}}\|V\|_{\ell^{\infty}L^{\widetilde{q}}}^2\lim_{n\to\infty}\sum_{k=n+1}^{\infty}\exp(-2\eta_j L_k)=0.
\end{align*}
Together with \eqref{P(n)} this yields the claim for $N=\infty$.
\end{proof}
\begin{remark}\label{remark nlog2n}
The factor $n\log^2 n$ in \eqref{choice of Lj} comes from the induction hypothesis and should not be taken too seriously. One could of course replace $\log\langle n\rangle$ by any other slowly varying sequence tending to infinity. However, this would not change the bound \eqref{choice of Lj} significantly.
\end{remark}
\subsection{Quasimode construction}
We now construct the potential $W_j$ having $\zeta_j$ as an eigenvalue.
\begin{lemma}\label{lemma single square well}
Given $\zeta\in\Sigma_{0}$ and $x_0\in{\mathbb R}^d$ there exists a potential $W=W(\zeta,x_0)\in L^{\infty}_{\rm comp}({\mathbb R}^d)$ such that the following hold.
\begin{enumerate}
\item $H_W$ has eigenvalue $\zeta$;
\item $\supp W\subset B(x_0,R)$, where
\begin{align}\label{single square well R}
R=R(\zeta)\asymp |\zeta|^{\frac{1}{2}}|\im\zeta|^{-1}|\log|\im\zeta|/|\zeta||.
\end{align}
\item For any $1\leq q\leq \infty$
\begin{align}\label{single square well norm V}
\|W\|_{L^q({\mathbb R}^d)}\asymp |\zeta|^{\frac{d}{2q}}|\im\zeta|^{1-\frac{d}{q}}|\log^{\frac{d}{q}}|\im\zeta/\zeta||.
\end{align}
\item For $1\leq q\leq q_d$,
\begin{align}\label{single square well Davies Nath}
F_{W,q}(\im\sqrt{\zeta})\lesssim |\zeta|^{\frac{d}{2q}}|\im\zeta|^{1-\frac{d}{q}}
\end{align}
\item The normalized eigenfunction $\psi=\psi(\zeta,x_0)$ of $H_W$ corresponding to $\zeta$ satisfies the exponential decay estimate
\begin{align}\label{exponential decay single square well}
|\psi(x)|\leq C|\zeta|^{1/4} |x|^{-\frac{d-1}{2}}\exp(-\im\sqrt{\zeta}\,{\rm d}\,(x,\supp W)).
\end{align}
\end{enumerate}
\end{lemma}
\begin{proof}
By scaling it suffices to prove this for $|\zeta|\asymp 1$.
In view of the results of Section \ref{section Complex square wells} (and $\zeta=E$ in the notation of that section) we can then simply choose $W$ as a shifted step potential. The shift of course does not affect the eigenvalues nor the $L^q$ norms. The latter are trivial to compute using the size bound $|W|=\mathcal{O}(\epsilon)$ and the formula \eqref{single square well R} for the width of the step. The estimate \eqref{single square well Davies Nath} follows from a direct computation. The exponential decay follows from Lemma \ref{lemma exponential decay} or the explicit form of the wavefunction for the step potential.
\end{proof}
\begin{remark}
Similar results involving complex step potentials are contained in \cite{MR3627408}, \cite{MR4104544}, \cite{MR4157680}, albeit in a less quantitative form. A technical detail that distinguishes our proof from these is that we first pick the eigenvalue, then find the potential. This avoids the use of Rouch\'e's theorem in \cite{MR4104544}, \cite{MR4157680}.
\end{remark}
\subsection{Exponential decay}
We prove that the exponential decay bound \eqref{assumption exponential decay} holds for a class of compactly supported potentials that will be relevant in the next section. The important point here is that the constant $C$ in \eqref{exponential decay} is independent of $W$.
\begin{lemma}\label{lemma exponential decay}
Assume that $\supp W\subset B(0,R)$ and $\zeta\in [0,\infty)$, $\im\sqrt{\zeta}\leq \tfrac{1}{2}R^{-1}\log R$, $|\zeta|^{1/2}\geq K R^{-1}$ for a large absolute constant $K$. Assume that $\psi$ is a normalized eigenfunction of $H_W$ with eigenvalue $\zeta$. Then there exists an absolute constant $C=C(d)$ such that for $|x|>R$,
\begin{align}\label{exponential decay}
|\psi(x)|\leq C|\zeta|^{1/4} |x|^{-\frac{d-1}{2}}\exp(-\im\sqrt{\zeta}|x|).
\end{align}
\end{lemma}
\begin{proof}
Since $\psi$ is normalized in $L^2$ it has units $l^{-d/2}$. By homogeneity, we may thus assume that $|\zeta|=1$. Since $\psi$ solves the Helmholtz equation
\begin{align*}
-\Delta\psi(x)=\kappa^2\psi(x)
\end{align*}
for $|x|>R$ and $\kappa^2=\zeta$, we have (see e.g.\ \cite[Chapter 1, Section 2]{MR2598115})
\begin{align*}
\psi(x)=A|x|^{-\nu}H^{(1)}_{\nu}(\kappa|x|)
\end{align*}
in this region, where $H^{(1)}_{\nu}$ is the Hankel function, $\nu=(d-2)/2$ and $A=A(d,W)$ is a normalization constant. By the well-known asymptotics of the Hankel function at infinity,
\begin{align*}
\psi(x)=Ac_d|x|^{-\frac{d-1}{2}}\exp(-\im\kappa|x|)(1+\mathcal{O}(|x|^{-1})).
\end{align*}
This would imply \eqref{exponential decay} if we could show that $A$ has an upper bound independent of $W$. Since $\psi$ is normalized,
\begin{align*}
A^2c_d^2\int_{|x|>R}|x|^{-(d-1)}\exp(-2\im\kappa|x|)(1+\mathcal{O}(|x|^{-1})){\rm d} x\leq \|\psi\|^2=1.
\end{align*}
For sufficiently large $K$ we estimate the integral from below by
\begin{align*}
(1-\mathcal{O}(K^{-1}))\int_R^{2R}\exp(-2\im\kappa r){\rm d} r\geq (1-\mathcal{O}(K^{-1}))R\exp(-\log R)\geq \frac{1}{4},
\end{align*}
which proves that $A\leq 2/c_d$.
\end{proof}
\begin{corollary}\label{cor exp decay}
Assume that $V_j(x)=W_j(x-x_j)$ and that the assumptions of Lemma \ref{lemma exponential decay} are satisfied for $W_j,\zeta,\psi_j,R_j$. Then \eqref{assumption exponential decay} holds for any $\widetilde{q}\geq 2$ and with
\begin{align}\label{def. aj}
a_j^{-d/\widetilde{q}}\lesssim|\zeta_j|^{1/4}\Big(\frac{L_j}{\im\sqrt{\zeta_j}}\Big)^{-\frac{d-1}{2}(\frac{1}{2}-\frac{1}{\widetilde{q}})}.
\end{align}
\end{corollary}
\begin{proof}
Let $\tfrac{1}{\widetilde{q}}+\tfrac{1}{r}=\tfrac{1}{2}$. By Hölder,
\begin{align*}
\|V_i\psi_j\|\leq \|V_i\|_{L^{\widetilde{q}}}\|\psi_j\|_{L^r(\Omega_i)}
\end{align*}
and by \eqref{exponential decay},
\begin{align*}
\|\psi_j\|_{L^r(\Omega_i)}\lesssim |\zeta_j|^{1/4}\big(\int_{\Omega_i}(|x-x_j|^{-\frac{d-1}{2}}\exp(-\im\sqrt{\zeta_j}\,|x-x_j|))^r{\rm d} x\big)^{1/r}.
\end{align*}
Since $|x-x_j|\geq {\rm d}(x,\Omega_j)$,
\begin{align*}
\|\psi_j\|_{L^r(\Omega_i)}\lesssim |\zeta_j|^{1/4}\Big(\frac{L_j}{\im\sqrt{\zeta_j}}\Big)^{-\frac{d-1}{2r}}.
\end{align*}
The claim follows.
\end{proof}
\subsection{A quantitative version of Boegli's example}\label{subsection A quantitative version of Boegli's result}
In view of Corollary \ref{corollary from quasimodes to eigenvalues}, given $\zeta_j\in \sigma_{H_{V_j}}$ and $\delta_j>0$ we would like to choose $\varepsilon_j=\epsilon_j(\zeta_j,\delta_j,L)$ as
\begin{align}\label{def. epsilon_n}
\varepsilon_j^{-1}=\exp(\mathcal{O}(1) M_{p,q}(L,\zeta_j)\log\frac{1}{\delta_j})
\end{align}
and require that \eqref{choice of Lj} holds with $a_j$ as in \eqref{def. aj}. This gives a sufficient condition on the sequence $L$ ensuring that ${\rm d}(\zeta_j,\sigma(H_V))\leq \delta_j$. The following proposition follows immediately from Corollary \ref{corollary from quasimodes to eigenvalues}, Lemma \ref{lemma quasimodes Hdiag to HV} and Corollary \ref{cor exp decay}.
\begin{proposition}\label{prop. inverse result}
Suppose Assumption \ref{assumption 1} holds, $V_j\in \ell^{\infty}L^{\widetilde{q}}$ for some $\widetilde{q}\geq 2$ and that $\supp V_j(\cdot+x_j)\subset B(0,R_j)$ for some positive $R_j$. Let $\zeta_j,\delta_j$ be sequences satisfying $\im\sqrt{\zeta_j}\leq \tfrac{1}{2}R_j^{-1}\log R_j$, $|\zeta_j|^{1/2}\geq K R_j^{-1}$ for some large absolute constant $K$ and $\delta_j\in(0,1/2)$. Assume that $L$ satisfies \eqref{choice of Lj} with $a_j$ as in \eqref{def. aj} and $\epsilon_j$ as in \eqref{def. epsilon_n}.
If $\zeta_j\in \Sigma_{0}$ is an eigenvalue of $H_{V_j}$ of multiplicity $m_j$, then $D(\zeta_j,\delta_j)$ contains at least $m_j$ eigenvalues of $H_V$, counted with multiplicity.
\end{proposition}
In the following we apply Proposition \ref{prop. inverse result} with $V_j=W(\zeta_j,x_j)$, where $W$ is the complex step potential in Lemma \ref{lemma quasimodes Hdiag to HV}. Clearly, $V_j\in L^q({\mathbb R}^d)$ for every $q\in[1,\infty]$, with
\begin{align}
\|V\|_{\ell^pL^q}&\lesssim \big(\sum_n\big(|\zeta_n|^{\frac{d}{2q}}|\im\zeta_n|^{1-\frac{d}{q}}|\log^{\frac{d}{q}}|\im\zeta_n/\zeta_n||\big)^p\big)^{\frac{1}{p}},\label{VellpLq}\\
\sup_{j\in [n]}\|V_j\|_{L^{\infty}}&\lesssim \sup_{j\in [n]}|\im\zeta_n|.\label{Vjellinfty}
\end{align}
We will also take $\widetilde{q}=\infty$, so that $a_j^{-d/\widetilde{q}}=1$. For the remainder of this section we assume the following.
\begin{assumption}\label{assumption 3}
Let $q>d$, and assume that \eqref{condition on sequence zetan intro} holds. Without loss of generality we may and will also assume that $\im\zeta_n$ is monotonically decreasing.
\end{assumption}
\begin{lemma}\label{lemma condition implies}
Under Assumption \eqref{assumption 3} the following hold.
\begin{enumerate}
\item[(i)] $\|V\|_{\ell^pL^q}\leq\epsilon_1$, $\|V\|_{\ell^{\infty}L^{\infty}}\lesssim 1$.
\item[(ii)] $|\im\zeta_n|\lesssim 1$.
\item[(iii)] $\langle\zeta_n\rangle\lesssim |\im\zeta_n|^{-\frac{2(q-d)}{d}}$.
\item[(iv)] $M_{p,q}(\zeta_n)\gtrsim |\im\zeta_n|^{p(\frac{(q-d)}{dq}+\frac{q_d}{q}-1)}$.
\item[(v)] $M_q(\zeta_n)\lesssim |\im\zeta_n|^{-1-\frac{2(q-d)}{d}-5p(\frac{q_d}{d}-1)-8+p(\frac{(q-d)}{dq}+\frac{q_d}{q}-1)}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Condition \eqref{condition on sequence zetan intro} states that the right hand side of \eqref{VellpLq} with $p=q$ is bounded by $\epsilon_1$. Since $p>q$ and the embedding $\ell^{q}\subset\ell^{p}$ is contractive, the first claim in (i) follows. Since $|\im\zeta_n|\leq |\zeta_n|$ and $|\log|\im\zeta_n/\zeta||\geq 1$ for $\zeta_n \in\Sigma_0$, Condition \eqref{condition on sequence zetan intro} also implies
\begin{align}\label{firstboud secondbound}
|\im\zeta_n|^{1-\frac{d}{2q}}\leq\epsilon_1,\quad |\zeta_n|^{\frac{d}{2}}\leq \epsilon_1|\im\zeta_n|^{q-d}.
\end{align}
Since $q>d$ the first bound implies (ii) and thus the second claim in (i) follows from \eqref{Vjellinfty}. The claim (iii) follows from the second bound in \eqref{firstboud secondbound} and (ii). Using~(iii), we find
\begin{align}
\omega_q(\zeta_n)=|\zeta_n|^{-\frac{1}{2q}}|\im\zeta_n|^{\frac{q_d}{q}-1}\gtrsim |\im\zeta_n|^{\frac{(q-d)}{dq}+\frac{q_d}{q}-1}.\label{estimate omegaqzetan}
\end{align}
This yields (iv) since $M_{p,q}(\zeta_n)\geq \omega_q(\zeta_n)^p$. It also follows from (ii) that
\begin{align}\label{2 also yields}
\frac{|\zeta_n|}{\langle \zeta_n\rangle}\gtrsim |\im\zeta_n|.
\end{align}
Combining (iii), \eqref{estimate omegaqzetan} and \eqref{2 also yields} with the trivial lower bound $|\zeta_n|\geq |\im\zeta_n|$ in~\eqref{def. Mzeta} yields~(v).
\end{proof}
\begin{remark}
From the first equality in \eqref{estimate omegaqzetan} and the definition of $M_{p,q}(z)$ in \eqref{def. Mzeta} it is easy to see that for $|\zeta_n|\asymp 1$, we have better bounds
\begin{align}\label{better bounds for Mpq if zeta order 1}
|\im\zeta_n|^{p(\frac{q_d}{q}-1)}\lesssim M_{p,q}(\zeta_n)\lesssim |\im\zeta_n|^{p(\frac{q_d}{q}-1)-1}.
\end{align}
\end{remark}
\begin{lemma}\label{lemma Lk kalpha}
Suppose that $L_k\gtrsim k^{\alpha}$. Then, under Assumption \ref{assumption 3},
\begin{align}
s(L,\left(\frac{|\zeta_n|}{\langle \zeta_n\rangle}\right)^{5}\zeta_n)\rangle\lesssim |\im\zeta_n|^{-\frac{1}{\alpha}(\frac{7}{2}+\frac{(q-d)}{d})},\quad
M_{p,q}(L,\zeta_n)\lesssim |\im\zeta_n|^{-\kappa_{\alpha}},
\end{align}
where $\kappa_{\alpha}:=1+\frac{2(q-d)}{d}+5p(\frac{q_d}{d}-1)+8-p(\frac{(q-d)}{dq}+\frac{q_d}{q}-1)+\frac{2p}{\alpha}(\frac{7}{2}+\frac{(q-d)}{d})$.
\end{lemma}
\begin{proof}
Combining \eqref{2 also yields} with the estimate
\begin{align}\label{estimate imsqrzetan}
\im\sqrt{\zeta_n}\asymp\frac{|\im\zeta_n|}{|\zeta_n|^{1/2}}\gtrsim |\im\zeta_n|^{1+\frac{(q-d)}{d}},
\end{align}
where the first bound holds since $\zeta_n\in\Sigma_0$ and the second bound follows from Lemma \ref{lemma condition implies} (iii), we obtain
\begin{align}
s(L,\left(\frac{|\zeta_n|}{\langle \zeta_n\rangle}\right)^{5}\zeta_n)\rangle\lesssim \mathrm{sep}(L,|\im\zeta_n|^{\frac{7}{2}+\frac{(q-d)}{d}}).
\end{align}
The claim thus follows from Proposition \ref{lemma distribution function} and Example a) following it.
\end{proof}
\begin{remark}
For $|\zeta_n|\asymp 1$, we again have better bounds
\begin{align}\label{better bounds sLzeta if zeta order 1}
s(L,\left(\frac{|\zeta_n|}{\langle \zeta_n\rangle}\right)^{5}\zeta_n)\rangle&\lesssim |\im\zeta_n|^{-\frac{1}{\alpha}},\\
M_{p,q}(L,\zeta_n)&\lesssim |\im\zeta_n|^{p(\frac{q_d}{q}-1)-1-\frac{2p}{\alpha}}.
\end{align}
\end{remark}
We assume now that
\begin{align}\label{assumption deltan}
\delta_n\geq \exp(-|\im\zeta_n|^{-\gamma})
\end{align}
for some $\gamma>0$. This lower bound is motivated from the corresponding upper bound that results from the Ghershgorin estimate \eqref{Ghershgorin set} and a posteriori by \eqref{choice of Lj 3}.
\begin{lemma}
Fix a compact set $\Sigma\subset\Sigma_0\cap {\mathbb C}\setminus[0,\infty)$. The there exists $c=c(\Sigma)$ such that
\begin{align}
\sigma(H_V)\cap\Sigma\subset\set{z}{\delta_{H_{V_n}}(z)\leq \exp(-c L_n)}.
\end{align}
\end{lemma}
The following lemma is obvious.
\begin{lemma}\label{lemma logepsilonn}
If $\epsilon_n$ is defined by \eqref{def. epsilon_n}, $L_k\gtrsim k^{\alpha}$ and $\delta_n$ satisfies \eqref{assumption deltan}, then under Assumption \ref{assumption 3},
\begin{align*}
\log\epsilon_n^{-1}\lesssim |\im\zeta_n|^{-\kappa_{\alpha}-\gamma}.
\end{align*}
\end{lemma}
Lemma \eqref{lemma condition implies} (i) and Lemma \ref{lemma logepsilonn} imply that the right hand side of \eqref{choice of Lj} (with $\widetilde{q}=\infty$) is bounded by $|\im\zeta_n|^{-\kappa_{\alpha}-\gamma}\log\langle n\rangle$. We will show that
\begin{align}\label{n<imzetanegpower}
\langle n\rangle\leq |\im\zeta_n|^{-\frac{d}{2}-q+1},
\end{align}
for all but finitely many $n\in{\mathbb N}$, which will then give a sufficient condition for the choice of $L$ in Proposition \ref{prop. inverse result}, namely
\begin{align}\label{choice of Lj 2}
L_n\geq C|\im\zeta_n|^{-\kappa_{\alpha}-\gamma-1-\frac{(q-d)}{d}}\log(|\im\zeta_n|^{-1}).
\end{align}
Here we have used \eqref{estimate imsqrzetan} to estimate $\im\sqrt{\zeta_n}$ from below. In order to be consistent with our assumption $L_k\gtrsim k^{\alpha}$ we actually choose
\begin{align}\label{choice of Lj 3}
L_n= C|\im\zeta_n|^{-\widetilde{\kappa}},
\end{align}
where, in view of \eqref{n<imzetanegpower}, it suffices to take
\begin{align}\label{tildekappa}
\widetilde{\kappa}:=\max(\kappa_{\alpha}+\gamma+2+\frac{(q-d)}{d},\alpha(\frac{d}{2}+q-1)).
\end{align}
The exact choice of $\alpha$ is not important for us and we choose $\alpha=1$ for convenience.
\begin{lemma}
Under Assumption \ref{assumption 3} we have \eqref{n<imzetanegpower} for all but finitely many $n\in{\mathbb N}$.
\end{lemma}
\begin{proof}
Suppose the claim is false. Then there exists a subsequence, again denoted by $(\zeta_n)_n$, such that $\langle n\rangle>|\im\zeta_n|^{-\frac{d}{2}-q+1}$. Since $|\zeta_n|\geq |\im\zeta_n|$, \eqref{condition on sequence zetan intro} implies
\begin{align*}
\sum_n\langle n\rangle^{-1}<\sum_n|\im\zeta_n|^{\frac{d}{2}+q-1}\lesssim 1,
\end{align*}
a contradiction.
\end{proof}
\subsection{Proof of Theorem \ref{thm quant Boegli intro}}
We now specialize Proposition \ref{prop. inverse result} to the step potential $V_j=W(\zeta_j,x_j)$ and the explicit choice \eqref{choice of Lj 3}, which will prove Theorem \ref{thm quant Boegli intro}. Since we already know that the exponential decay bound is true for these potentials (see \eqref{exponential decay single square well}) we do not need to check the conditions $\im\sqrt{\zeta_j}\leq \tfrac{1}{2}R_j^{-1}\log R_j$, $|\zeta_j|^{1/2}\geq K R_j^{-1}$, but it is easy to see from \eqref{single square well R} that they do hold.
\begin{proposition}\label{prop. inverse result step potential}
Suppose Assumption \ref{assumption 3} holds, $\delta_n>0$ satisfies \eqref{assumption deltan} for some $\gamma>0$, and let $V=V(L)$ be the potential whose bumps $V_n=W(\zeta_n,x_n)$ are separated by $L_n$ in \eqref{choice of Lj 3}. Then $D(\zeta_n,\delta_n)$ contains an eigenvalues of $H_V$. Moreover, $\|V\|_{L^q}\leq \epsilon_1$ and $V$ decays polynomially,
\begin{align}
|V(x)|\lesssim \langle x\rangle^{-\frac{1}{\widetilde{\kappa}}},
\end{align}
where $\widetilde{\kappa}$ is given by \eqref{tildekappa} for some arbitrary $\alpha>0$.
\end{proposition}
\begin{proof}
By \eqref{single square well norm V}, we have the bound $|V(x)|\lesssim |\im\zeta_n|$ for $|x-x_n|\leq R(\zeta_n)$ and zero elsewhere. Since $|\im\zeta_n|\lesssim 1$ by Lemma \ref{lemma condition implies} (i), it follows that $V$ is bounded. Since $\widetilde{\kappa}\geq 2$, a comparison between $L_n$ and $|\Omega_n|=R(\zeta_n)$ in \eqref{single square well R} shows that $V$ is sparse. Therefore, by \eqref{lower bound Li in terms of xi, Omegai sparse}, we have $L_n\gtrsim |x_n|$. Hence, \eqref{choice of Lj 3} yields
\begin{align*}
|V(x_n)|\lesssim |\im\zeta_n|\lesssim L_n^{-\frac{1}{\widetilde{\kappa}}}\lesssim x_n^{-\frac{1}{\widetilde{\kappa}}},
\end{align*}
from which the decay bound follows.
\end{proof}
\section{Complex step potential}\label{section Complex square wells}
In this section we will establish precise estimates for eigenvalues of the sperically symmetric complex step potential $V=V_0\mathbf{1}_{B(0,R)}$, where $V_0\in {\mathbb C}$ and $R>0$. The bound state problem for $V_0<0$ and $d=1,3$ is treated in virtually any quantum mechanics textbook (see e.g.\ Problem 25 and Problem 63 in \cite{MR1746199}).
We adopt the notation
\begin{align}\label{step potential notation}
\chi=\sqrt{E},\quad \kappa=\sqrt{\chi^2-V_0}.
\end{align}
Here $E\in {\mathbb C}$ is the eigenvalue parameter, i.e.\ we consider the stationary Schrödinger equation
\begin{align}\label{Schrödinger equation square well}
-\Delta \psi+(V-E)\psi=0,
\end{align}
which becomes $-\Delta \psi-\kappa^2\psi=0$
inside the step and
$-\Delta \psi-\chi^2\psi=0$
outside the step.
\subsection{One dimension}\label{subsection one dimension step potential}
We start with one-dimensional case. The solution space to~\eqref{Schrödinger equation square well} then splits into even and odd functions, while in higher dimensions it splits into functions with definite angular momentum $\ell$. We consider odd functions as these also provide a solution for the case $d=3$ and $\ell=0$ ($s$-waves). The standard procedure to solving the square well problem reduces the task to finding zeros of the nonlinear scalar function $F(V_0,\kappa):=\I\chi-\kappa\cot(\kappa R)$, where $\chi=\sqrt{\kappa^2+V_0}$ by~\eqref{step potential notation}. A complete study of all the complex poles of this equation was initiated by Nussenzveig \cite{Nussenzveig} for $V_0\in {\mathbb R}\setminus\{0\}$. Subsequent articles in the physics literature \cite{JOFFILY1973301}, \cite{PhysRevC.53.2004}, \cite{Dabrowski_1997}, \cite{PhysRevA.61.032716} investigated the case of complex potentials. The solution $\kappa=\kappa(V_0)$ is not single-valued as there are branch points where $\partial F/\partial\chi=0$. The viewpoint endorsed by \cite{PhysRevA.61.032716} is to regard the equation $F(V_0,\kappa)=0$ as the definition of a Riemann surface. This approach treats the complex variables $\kappa$ and $V_0$ on equal footing. In fact, it is easy to see that one can always use $\kappa$ as a coordinate, i.e.\ one can solve for $V_0$,
\begin{align}\label{solve for V0 1d}
V_0=-\kappa^2\sec^2(\kappa R).
\end{align}
For the purpose of the construction of the sparse potential in Subsection \ref{subsection A quantitative version of Boegli's result} we do not need to solve for $\kappa$. Instead, we pick $\kappa$ first and then define $V_0$ by \eqref{solve for V0 1d}. To get an eigenvalue (i.e.\ a resonance on the physical sheet) we simply need to take care of the condition $\im\chi>0$, i.e.
\begin{align}\label{take care of im chi}
\re (\kappa R\cot(\kappa R))>0.
\end{align}
We are only interested in complex eigenvalues $E$ with $|E|\asymp 1$ (the general case can be obtained by scaling). We will try to make $V_0$ in \eqref{solve for V0 1d} small, i.e.\ we postulate that $V_0=\epsilon \widetilde{V}_0$, where $\epsilon>0$ is a small parameter and $\widetilde{V}_0\in {\mathbb C}$ is of unit size. By~\eqref{step potential notation} this implies that $|\kappa|\asymp 1$, and \eqref{solve for V0 1d} then reveals that $|\sin^2(\kappa R)|\asymp \epsilon^{-1}$, which means that
\begin{align}\label{dominant balance exponentials}
{\rm e}^{2\im\kappa R}\asymp \epsilon^{-1},\quad {\rm e}^{-2\im\kappa R}\asymp \epsilon.
\end{align}
Since we are free to choose $\kappa$,
set $\kappa=\pm 1+\I\epsilon\sigma$ with $\sigma> 0$, which will yield an eigenvalue with $\re E=1+\mathcal{O}(\epsilon)$ and $|\im E|=\mathcal{O}(\epsilon)$.
Going back to \eqref{dominant balance exponentials} we see that we must have
\begin{align}\label{R consistent with Davies bound}
R=\frac{1}{2\sigma\epsilon}\log\frac{C}{\epsilon}
\end{align}
for some constant $C$. It is quickly checked that this is consistent with the bound of Abramov et al.\ \cite{AAD01} since $\|V\|_{L^1}\gtrsim \log \frac{1}{\epsilon}$ and $|E|\asymp 1$. In view of \eqref{dominant balance exponentials} we may write ${\rm e}^{2\I\kappa R}=\epsilon u$, where $u=C{\rm e}^{2\I\re \kappa R}$. Using the Taylor approximation
\begin{align}\label{sec2 Taylor}
\sec^2(\kappa R)=-4\epsilon u(1+2\epsilon u +\mathcal{O}(\epsilon^2))
\end{align}
we obtain from \eqref{solve for V0 1d} that
\begin{align}\label{V0 result 1d}
V_0=4\epsilon u(\re\kappa)^2+\mathcal{O}(\epsilon^2),
\end{align}
which provides the desired smallness $|V_0|=\mathcal{O}(\epsilon)$. As already mentioned, we need to make sure that \eqref{take care of im chi} holds. Using the Taylor approximation
\begin{align*}
\cot(\kappa R)=\I(1+2\epsilon u+\mathcal{O}(\epsilon^2)),
\end{align*}
we find that \eqref{take care of im chi} holds if
\begin{align}\label{condition u}
-(\re\kappa)(\im u)-(\im\kappa)(\re u)>0
\end{align}
In particular, for $u\in \I{\mathbb R}_+$, we find that \eqref{condition u} forces us to choose $\re\kappa=-1$. Adopting this choice for $u$, it is then easy to check that we get an eigenvalue with $\re E=1+\mathcal{O}(\epsilon)$ and $\im E=\epsilon(4|u|-2\sigma)+\mathcal{O}(\epsilon^2)$ as desired. By simple scaling arguments this proves the one-dimensional case of Lemma \ref{lemma single square well}. We observe that the result is consistent with the trivial numerical range bound $\im E\leq \im V_0$; in fact, by choosing $\sigma$ small, $E$ can be taken arbitrarily close to the boundary of the numerical range $\im z=4\epsilon|u|$, up to errors of order $\epsilon^2$.
Before we conclude the one-dimensional case we note that the same result could have been obtained with an even wavefunction, in which case $\sec$ replaced by $\csc$ in \eqref{solve for V0 1d} and $\cot$ is replaced by $\tan$ in \eqref{take care of im chi}. The Taylor approximations
\begin{align}
\csc^2(\kappa R)=4\epsilon u(1+\mathcal{O}(\epsilon)),\quad
\tan(\kappa R)=\I(1-2\epsilon u+\mathcal{O}(\epsilon^2)),
\end{align}
and the freedom to choose the signs and the imaginary part of $u$ yields a proof of Lemma \ref{lemma single square well} using odd solutions.
\subsection{Higher dimensions}
By symmetry reductions we are led to consider the radial Schrödinger equation
\begin{align*}
(-\partial_r^2-\frac{d-1}{r}\partial_r+\frac{\ell(\ell+d-2)}{r^2}+V(r)-E)\psi_{\ell}(r)=0.
\end{align*}
It can be shown (see e.g.\ \cite[(5.12)]{MR2648080}) that an eigenvalue $E$ corresponds to a zero of the function (Wronskian)
\begin{align}\label{Wronskian}
F(V_0,\kappa):=\kappa J_{\nu}'(\kappa R)H^{(1)}_{\nu}(\chi R)-\chi J_{\nu}(\kappa R)H^{(1)'}_{\nu}(\chi R),
\end{align}
where $\nu=\ell+\frac{d-2}{2}$. We recall that $\chi,\kappa,E,V_0$ are related by \eqref{step potential notation}. Computations of resonances for spherically symmetric potentials can be
found in \cite{MR115692}, \cite{MR987299}, \cite{MR1953522}, \cite{MR2648080}. The last three papers use uniform asymptotic expansion of Bessel functions for large order. Here we only consider $s$-waves, i.e.\ $\ell=0$. Then we have the asymptotics
\begin{align}
J_{\nu}(z)&=\big(\frac{2}{\pi z}\big)^{1/2}\cos(z-\frac{\pi\nu}{2}-\frac{\pi}{4})(1+\mathcal{O}(|z|^{-1})),\\
H^{(1)}_{\nu}(z)&=\big(\frac{2}{\pi z}\big)^{1/2}\exp(\I z-\I\frac{\pi\nu}{2}-\I\frac{\pi}{4})(1+\mathcal{O}(|z|^{-1})),\\
J_{\nu}'(z)&=-\big(\frac{2}{\pi z}\big)^{1/2}\sin(z-\frac{\pi\nu}{2}-\frac{\pi}{4})(1+\mathcal{O}(|z|^{-1})),\\
H^{(1)'}_{\nu}(z)&=\I\big(\frac{2}{\pi z}\big)^{1/2}\exp(\I z-\I\frac{\pi\nu}{2}-\I\frac{\pi}{4})(1+\mathcal{O}(|z|^{-1}).
\end{align}
With the same choice of $\kappa$ as in the one-dimensional case and with $u\in{\mathbb C}$ such that ${\rm e}^{2\I(\kappa R-\frac{\pi\nu}{2}-\frac{\pi}{4})}=\epsilon u$, we then obtain that the zeros of $F(V_0,\kappa)$ coincide with the zeros of a function
\begin{align}
\kappa\sin(\omega(\kappa R))-\I\chi\cos(\omega(\kappa R))+\mathcal{O}(R^{-1}),
\end{align}
where $\omega(\kappa R)=\kappa R-\frac{\pi\nu}{2}-\frac{\pi}{4}$ and the $\kappa$-derivative of the error term is $\mathcal{O}(1)$. Recall that $\chi=\chi(V_0,\kappa)$ is given by \eqref{step potential notation}. The zeros of the function without the error term are found exactly as in the one-dimensional case and can be parametrized by $\kappa$, v.i.z.
\begin{align}\label{V_0 result higher d}
V_0=-\kappa^2\csc^2(\omega(\kappa R)).
\end{align}
This follows by dividing the above expression by $\cos(\omega(\kappa R))$ which has no zeros since $\im\kappa>0$. Since $|\cos(\omega(\kappa R))^{-1}|=\mathcal{O}(\epsilon^{1/2})$ we get from \eqref{R consistent with Davies bound} that the error after dividing is $\mathcal{O}(\epsilon^{3/2})$, i.e.\ we are looking for the zeros of a function
\begin{align}\label{Ftilde}
\widetilde{F}(V_0,\kappa)=V_0+\kappa^2\csc^2(\omega(\kappa R))+\mathcal{O}(\epsilon^{3/2}),
\end{align}
where the derivative of the error is $\mathcal{O}(\epsilon^{1/2})$.
The implicit function theorem thus yields $\partial \widetilde{F}(V_0,\kappa)/\partial V_0=1+\mathcal{O}(\epsilon^{1/2})$, which means that we can solve $\widetilde{F}(V_0,\kappa)$ for $V_0$, and the solution satisfies \eqref{V_0 result higher d} up to errors $\mathcal{O}(\epsilon^{3/2})$. Hence we obtain that $|V_0|=\mathcal{O}(\epsilon)$ as before.
\subsection{Proof of Theorem \ref{thm. violation of locality into}} We return to one dimension. We first prove the upper bound \eqref{Vsp}. Since $\sqrt{|V_0|}R_0$ is of order one, the bound in \cite{MR3556444} yields that the total number of eigenvalues of $H_{V_j}$ is also of order one. By Proposition \ref{Prop. comparison principle first version} (it is clear that the assumption on the norm of $V$ can be dropped), given $N\gg 1$, we can find $L=L(N)$ such that $H_{\rm s}$ has the same number of eigenvalues in $\Sigma=\Sigma(N)$ as $H_{\rm diag}$, which is just the $N$-fold orthogonal sum of the $H_{V_j}$ and hence has less than $\mathcal{O}(N)$ eigenvalues by the first part of the argument.
To prove the lower bound \eqref{Vnsp} we return to the formulas \eqref{solve for V0 1d}, \eqref{take care of im chi}, but we now fix $V_0=\I$. We also set $R=NR_0$ and $R_0\asymp 1$, so that the dimensionless parameter $\sqrt{|V_0|}R$ is of size $N$. We first solve an approximate equation and then use Rouch\'e's theorem to show that the exact equation \eqref{solve for V0 1d} has solutions close to the approximate ones. Finally, we use \eqref{take care of im chi} to check that we have found a pole on the physical plane (i.e.\ an eigenvalue). The approximate equation is $G_1(\kappa)=0$, where
\begin{align*}
G_1(\kappa):=V_0-4\kappa^2{\rm e}^{2\I\kappa R},
\end{align*}
and the approximation will be valid in the regime $\im\kappa R\gg 1$. Since $G_1$ can be factored,
\begin{align*}
G_1(\kappa)=(\sqrt{V_0}-2\kappa{\rm e}^{\I\kappa R})(\sqrt{V_0}+2\kappa{\rm e}^{\I\kappa R}),
\end{align*}
we only look for zeros of the first factor.
These zeros $\kappa_n$ are expressed by means of the Lambert $W$ function,
\begin{align*}
\kappa_n R=-\I W_n(\I \sqrt{V_0}R/2),
\end{align*}
where $n\in{\mathbb Z}$ and $W_n$ are the branches of the Lambert $W$ function. According to in \cite[(4.19)]{MR1414285} the asymptotic expansion of $W_n(z)$ as $|z|\to\infty$ is
\begin{align}\label{asymptotics Lambert W}
W_n(z)=\log(z)+2\pi\I n-\log(\log(z)+2\pi\I n)+\mathcal{O}(\frac{\log(\log(z)+2\pi\I n)}{\log(z)+2\pi\I n}),
\end{align}
where $\log$ is the principal branch of the logarithm on the slit plane with the negative real axis as branch cut. For $z=\I \sqrt{V_0}R/2$ this gives
\begin{align}\label{kappan formula1}
\kappa_nR=2\pi n-\I\log(\I \sqrt{V_0}R/2)+\I\log(\log(\I \sqrt{V_0}R/2)+2\pi\I n)+E_n(V_0,R),
\end{align}
where, for $N\gg 1$, the error satisfies $|E_n(V_0,R)|\lesssim \log(\log N+|n|)/(\log N+|n|)$, where we recalled that $\sqrt{|V_0|}R\asymp N$. For the assumption $\im\kappa R\gg 1$ made before to be consistent with the formula for $\kappa_n$ we require
\begin{align*}
\re\log\left(\frac{\log(\I \sqrt{V_0}R/2)+2\pi\I n}{\I \sqrt{V_0}R/2}\right)\gg 1\iff \left|\frac{\log(\I \sqrt{V_0}R/2)+2\pi\I n}{\I \sqrt{V_0}R/2}\right|\gg 1.
\end{align*}
Since $N\gg 1$ we can neglect the logarithm in the second expression and deduce the condition $|n|\gg N$, which we will assume henceforth. This gives us the error bound $|E_n(V_0,R)|\lesssim \log |n|/|n|$, which implies that
\begin{align}\label{imkappa=logn}
\im\kappa_n R\gtrsim \log \frac{|n|}{N},
\end{align}
in agreement with the assumption $\im\kappa R\gg 1$. We also obtain the more precise formulas
\begin{align}\label{kappan}
\re \kappa_nR=2\pi n+\frac{5\pi}{4}+\mathcal{O}(\log |n|/|n|),\quad
\im\kappa_n R=\log \frac{|n|}{N}+\mathcal{O}(1),
\end{align}
where we assumed that $n<0$. To justify this assumption, we recall from the discussion at the end of Subsection \ref{subsection one dimension step potential} that \eqref{V0 result 1d} and \eqref{condition u}, together with \eqref{kappan} and the assumption $V_0=\I$ made at the beginning of this subsection, imply that $n$ must be negative.
Having found the large zeros of $G_1(\kappa)$ we proceed to find those of
\begin{align*}
G_2(\kappa):=V_0+\kappa^2\sec^2(\kappa R),
\end{align*}
which determines the eigenvalues of the step potential (see the beginning of Subsection \ref{subsection one dimension step potential}). We define $\epsilon_n:=\exp(-2\im\kappa_n R)$, so that ${\rm e}^{2\I\kappa_n R}=\epsilon_n u_n$ for some $u_n$ on the unit circle. Note that, by \eqref{kappan}, $\epsilon_n=\mathcal{O}(1)\big(\frac{N}{|n|}\big)^2$. Using~\eqref{sec2 Taylor} with $\epsilon=\epsilon_n$, $u=u_n$, we estimate, for $\widetilde{\epsilon}_n\ll 1$,
\begin{align}\label{G1-G2}
\sup_{|\kappa-\kappa_n|=\widetilde{\epsilon}_n}|G_2(\kappa)-G_1(\kappa)|=\mathcal{O}(n^2\epsilon_n^2).
\end{align}
Moreover, for $|\kappa-\kappa_n|=\widetilde{\epsilon}_n$ we have
\begin{align*}
|G_1'(\kappa)|&\gtrsim |\kappa|^2R{\rm e}^{-2\im\kappa R}\gtrsim Nn^{2}\epsilon_n,\\
|G_1''(\kappa)|&\lesssim |\kappa|^2R^2{\rm e}^{-2\im\kappa R}\lesssim N^2n^{2}\epsilon_n.
\end{align*}
Using $G_1(\kappa_n)=0$ and Taylor expanding, it follows that
\begin{align}
|G_1(\kappa)|\gtrsim Nn^{2}\epsilon_n\widetilde{\epsilon}_n+\mathcal{O}(N^2n^{2}\epsilon_n\widetilde{\epsilon}_n^2).
\end{align}
For this to be meaningful we must of course assume $\widetilde{\epsilon}_n\ll 1/N$, which we do. Then we have $|G_1(\kappa)|\gtrsim Nn^{2}\epsilon_n\widetilde{\epsilon}_n$. Comparing this with \eqref{G1-G2} we see that
\begin{align*}
\sup_{|\kappa-\kappa_n|=\widetilde{\epsilon}_n}|G_1(\kappa)|^{-1}|G_2(\kappa)-G_1(\kappa)|<1,
\end{align*}
provided $\widetilde{\epsilon}_n\gg \epsilon_n/N$. Adopting the choice $\widetilde{\epsilon}_n=C\frac{N}{n^2}$, where $C$ is a large constant, we see that there exists a zero $\widetilde{\kappa}_n\in D(\kappa_n,C\frac{N}{n^2})$ of $G_2$. By the smallness of $N/n^2$, it follows that $\widetilde{\kappa}_n$ also satisfies \eqref{kappan}. We drop the tilde, i.e.\ we now denote the zeros of $G_2$ by $\kappa_n$. Summarizing what we have done so far, we have found infinitely many resonances $\kappa_n$, $|n|\gg N$, of of the step potential satisfying \eqref{kappan}. The last step is to check which of the resonances lie on the physical sheet, i.e.\ are actual eigenvalues. For this we need to check condition \eqref{take care of im chi}. By \eqref{kappan},
\begin{align*}
&\cot(\kappa_n R)=-\I\big(1+\I\big(\frac{N}{|n|}\big)^2{\rm e}^{\mathcal{O}(1)+\I\mathcal{O}(\log|n|/|n|)}+\mathcal{O}(\big(\frac{N}{|n|}\big)^4)\big),\\
&\re (\kappa_n R\cot(\kappa_n R))=2\pi\frac{N^2}{n}{\rm e}^{\mathcal{O}(1)}(1+\mathcal{O}(\log^2|n|/|n|))+\log\frac{|n|}{N}(1+\mathcal{O}(\log|n|/|n|)).
\end{align*}
Hence, recalling that $n<0$, the condition \eqref{take care of im chi} is nonvoid and is satisfied whenever $|n|\log\frac{|n|}{N}\ll N^2$; this holds for $|n|\ll N^2/\log N$. Recalling \eqref{step potential notation} we obtain the complex energies $E=E_n$,
\begin{align*}
\re E_n\asymp \frac{n^2}{N^2},\quad \im E_n\asymp \frac{|n|}{N^2}\log\frac{|n|}{N},
\end{align*}
and those energies with $c\frac{N^2}{\log N}\leq |n|\leq C\frac{N^2}{\log N}$ lie in the rectangle $\Sigma$ (see Theorem~\ref{thm. violation of locality into}).
This completes the proof of the lower bound \eqref{Vnsp}.
\section{Technical tools}\label{section Technical tools}
\subsection{Lower bounds on moduli of holomorphic functions}\label{subsection holomorphic functions}
We collect some well known results about the modulus of holomorphic functions away from zeros, based on Cartan's bound for polynomials (see e.g.\ \cite{MR589888}).
Let $U_1\Subset U_2\Subset {\mathbb C}$, where $U_2$ is simply connected. Assume that $f$ is holomorphic in a neighborhood of $U_2$ and $\zeta_2\in U_2$. Let $z_1, z_2\ldots, z_n$, be the zeros of $f$ in $U_2$. Define
\begin{align*}
Z_{f,\delta,U_2}:=\bigcup_{j=1}^n D(z_j,\delta).
\end{align*}
The following version can be found in \cite[Appendix D]{MR3969938}.
\begin{lemma}\label{lemma appendix simplest version}
There exists a constant $C=C(U_1,U_2,z_0)$ such that for any sufficiently small $\delta>0$,
\begin{align*}
\log|f(z)|\geq -C\log\frac{1}{\delta}\big(\log|f(z)|-\log\max_{z\in U_2}|f(\zeta_2)|\big)
\end{align*}
for all $z\in U_1\setminus Z_{f,\delta,U_2}$.
\end{lemma}
We need also use a more precise version, where $f$ is holomorphic in a neighborhood of $U_3$,
\begin{align}\label{Uj=D(0,rj)}
U_j=D(0,r_j),\quad j=1,2,3,
\end{align}
with $r_1<r_2<r_3$ and $\zeta_2=0$.
\begin{lemma}\label{lemma Levin}
Assume \eqref{Uj=D(0,rj)}. Then there exists an absolute constant $C$ such that for any sufficiently small $\delta>0$,
\begin{align*}
\log|f(z)|\geq -C\log\frac{1}{\delta}\big((r_2-r_1)^{-1}+\log^{-1}\big(\frac{r_3}{r_2}\big)\big)\big(\log\max_{|z|=r_3}|f(z)|-\log|f(\zeta_2)|\big)
\end{align*}
for all $z\in U_1\setminus Z_{f,\delta r_2,U_2}$.
\end{lemma}
\begin{proof}
The proof is a straightforward adaptation of \cite[Chapter 1, Theorem 11]{MR589888}, but we include it for completeness.
In the following, $r_1,r_2,r_3$ play the roles of $R,2R,2{\rm e} R$ in \cite{MR589888}. Without loss of generality we may and will assume that $\zeta_2=0$ and that $f(0)=1$; otherwise we could replace $f(z)$ by $\frac{f(z+\zeta_2)}{f(\zeta_2)}$. Consider the function
\begin{align*}
\varphi(z):=\frac{(-r_2)^n}{z_1 z_2\ldots z_n}\prod_{k=1}^n\frac{r_2(z-z_k)}{r_2^2-\overline{z_k}z}.
\end{align*}
We recall that $z_1,z_2,\ldots, z_n$ are the zeros of $f$ in $D(0,r_2)$. Observe that $\varphi(0)=1$ and
\begin{align*}
\varphi(r_2{\rm e}^{\I\theta})=\frac{r_2^n}{|z_1 z_2\ldots z_n|}
\end{align*}
for $\theta\in{\mathbb R}$. The function
\begin{align*}
\Psi(z):=\frac{f(z)}{\varphi(z)}
\end{align*}
has no zeros in $D(0,r_2)$ and satisfies $\psi(0)=1$; therefore, by Carath\'eodory's theorem \cite[Theorem 9]{MR589888}, for $|z|\leq r_1$,
\begin{align*}
\log|\psi(z)|&\geq -\frac{2r_1}{r_2-r_1}\big(\log\max_{|z|=r_2}|f(z)|+\log\frac{r_2^n}{|z_1 z_2\ldots z_n|}\big)\\
&\geq -\frac{2r_1}{r_2-r_1}\log\max_{|z|=r_2}|f(z)|.
\end{align*}
To estimate $\varphi$ from below for $|z|\leq r_1$ we use
\begin{align*}
\prod_{k=1}^n|r_2^2-\overline{z_k}z|&<(2r_2^2)^n,\\
\prod_{k=1}^n|r_2(z-z_k)|&>\left(\frac{\delta r_2}{{\rm e}}\right)^nr_2^n,\quad z\notin Z_{f,2\delta r_2,U_2}.
\end{align*}
The second inequality follows from Cartan's estimate \cite[Theorem 10]{MR589888}. We thus obtain the lower bound
\begin{align*}
|\varphi(z)|>(2r_2^2)^{-n}\left(\frac{\delta r_2}{{\rm e}}\right)^n\frac{r_2^{2n}}{|z_1 z_2\ldots z_n|}>\left(\frac{\delta}{2{\rm e}}\right)^n,\quad z\notin Z_{f,2\delta r_2,U_2}.
\end{align*}
By Jensen's formula \cite[Lemma 4]{MR589888}, since $f(0)=1$,
\begin{align*}
n\leq \log^{-1}\big(\frac{r_3}{r_2}\big)\log\max_{|z|=r_3}|f(z)|,
\end{align*}
and consequently
\begin{align*}
\log|\varphi(z)|>\log^{-1}\big(\frac{r_3}{r_2}\big)\log\max_{|z|=r_3}|f(z)|\log \left(\frac{\delta}{2{\rm e}}\right),\quad z\notin Z_{f,2\delta r_2,U_2}.
\end{align*}
Together with the lower bound for $\log|\psi|$ this leads to the claimed estimate, upon redefining $\delta$ and absorbing an error into the constant $C$.
\end{proof}
Next we state a version of Lemma \ref{lemma Levin} for ``wedges" of the form
\begin{align}\label{def. wedge}
W(\varphi,\theta;r,R):=\set{z\in{\mathbb C}\setminus[0,\infty)}{\arg(z)\in(2\varphi,2\theta),\, |z|\in (r^2,R^2)}
\end{align}
and $W(\varphi,\theta;R):=W(\varphi,\theta;R,\infty)$, where $0\leq\varphi<\theta\leq\pi$. In the following we fix $0\leq \varphi_3<\varphi_2<\varphi_1<\theta_1<\theta_2<\theta_3\leq\pi$ and $0<r_3<r_2<r_1<R_1<R_2$, and define
\begin{align*}
U_1:=W(\varphi_1,\theta_1;r_1,R_1),\quad
U_2:=W(\varphi_2,\theta_2;r_2,R_2),\quad
U_3:=W(\varphi_3,\theta_3;r_3).
\end{align*}
\begin{lemma}\label{lemma wedge}
Assume $f$ is a bounded holomorphic function on $U_3$ and that
\begin{align}\label{condition lemma wedge}
r_3\ll r_2,\!\!\!\!\quad\frac{{\rm d}(\partial U_2,\partial U_3)}{{\rm d}(\partial U_1,\partial U_3)}\ll \left(\frac{r_2}{R_2}\right)^{\frac{2\pi}{\theta_3-\varphi_3}+2},\!\!\!\!\quad \frac{{\rm d}(\partial U_1,\partial U_3)}{(\theta_3-\varphi_3)R_2^2} \left(\frac{r_2}{R_2}\right)^{\frac{2\pi}{\theta_3-\varphi_3}}\ll 1.
\end{align}
Then there exists an absolute constant $C$ such that for any $\zeta_2\in U_2$ and any sufficiently small $\delta>0$,
\begin{align}\label{eq. lemma wedge}
\log|f(z)|\geq -C\frac{R_2^2}{{\rm d}(\partial U_2,\partial U_3)} \left(\frac{R_2}{r_2}\right)^{\frac{2\pi}{\theta_3-\varphi_3}}\log\frac{1}{\delta}\big(\log\max_{z\in U_3}|f(z)|-\log|f(\zeta_2)|\big)
\end{align}
for all $z\in U_1\setminus Z_{f,\delta,U_2}$.
\end{lemma}
\begin{remark}
The constant $C$ only depends on the implicit constants in~\eqref{condition lemma wedge}. Hence, one can optimize the inequality with respect to $\zeta_2$, subject to the conditions above.
\end{remark}
\begin{proof}
We map $U_3$ conformally onto the unit disk, using a composition of the following conformal maps (where, by abuse of notation, we denote the variable and the map by the same letter):
\begin{itemize}
\item[i)] $U_3\to\kappa(U_3)\subset \mathbb{H}$, $\kappa(z):=\sqrt{z}$;
\item[ii)] $\kappa(U_3)\to S:=\set{\sigma\in {\mathbb C}}{0<\im\sigma<\pi,\,\re\sigma>0}$,
\begin{align*}
\sigma(\kappa):=\log({\rm e}^{-\frac{\I\pi\varphi_3}{\theta_3-\varphi_3}}(\kappa/r_3)^{\frac{\pi}{\theta_3-\varphi_3}}),
\end{align*}
where we select the principal branch of the logarithm on ${\mathbb C}\setminus[0,-\I\,\infty)$;
\item[iii)] The Schwarz-Christoffel transformation $S\to \mathbb{H}$, $\tau(\sigma):=\cosh(\sigma)$.
\item[iv)] The Möbius transformation $\mathbb{H}\to D(0,1)$, $w(\tau):=\frac{\tau-\tau_2}{\tau+\tau_2}$, where $\tau_2:=\tau(\sigma(\sqrt{\zeta_2}))$.
\end{itemize}
The choice of $\tau_2$ has been made in such a way that $w(z)=0$ if $z=\zeta_2$. Here we again abuse notation and write $w(z)=w(\tau(\sigma(\sqrt{z})))$. Note that
\begin{align}
\tau(\sigma(\kappa))=\frac{1}{2}(\alpha(\kappa)+\alpha(\kappa)^{-1}),\quad
\alpha(\kappa):={\rm e}^{-\frac{\I\pi\varphi_3}{\theta_3-\varphi_3}}(\kappa/r_3)^{\frac{\pi}{\theta_3-\varphi_3}}.
\end{align}
By distortion bounds \cite[Cor. 1.4]{MR1217706},
\begin{align}\label{Koebe distortion}
\left|\frac{{\rm d} w(z)}{{\rm d} z}\right|{\rm d}(z,U_3)\leq 1-|w(z)|^2\leq 4 \left|\frac{{\rm d} w(z)}{{\rm d} z}\right|{\rm d}(z,U_3).
\end{align}
We compute the differential of $w$ at $z\in U_2$ by the chain rule,
\begin{align*}
&\left|\frac{{\rm d} w}{{\rm d} z}\right|=\left|\frac{{\rm d} w}{{\rm d}\tau}\frac{{\rm d} \tau}{{\rm d}\sigma}\frac{{\rm d} \sigma}{{\rm d}\kappa}\frac{{\rm d}\kappa}{{\rm d} z}\right|=\frac{\pi|\tau_2||\sinh(\sigma)|}{(\theta_3-\varphi_3)|z|(\tau+\tau_2)^2}\\
&=\frac{\pi}{(\theta_3-\varphi_3)|z|}\frac{|\alpha(\kappa_2)+\alpha(\kappa_2)^{-1}||\alpha(\kappa)+\alpha(\kappa)^{-1}|}{|\alpha(\kappa_2)+\alpha(\kappa_2)^{-1}+\alpha(\kappa)+\alpha(\kappa)^{-1}|^2}.
\end{align*}
Since $|\kappa|\geq r_2\gg r_3$, we have $|\alpha(\kappa)|\gg 1$, whence
\begin{align*}
\frac{|\alpha(\kappa_2)+\alpha(\kappa_2)^{-1}||\alpha(\kappa)+\alpha(\kappa)^{-1}|}{|\alpha(\kappa_2)+\alpha(\kappa_2)^{-1}+\alpha(\kappa)+\alpha(\kappa)^{-1}|^2}
\asymp \frac{|\alpha(\kappa_2)||\alpha(\kappa)|}{(|\alpha(\kappa_2)|+|\alpha(\kappa)|)^2},
\end{align*}
which, in view of $|\alpha(\kappa)|=(|\kappa|/r_3)^{\frac{\pi}{\theta_3-\varphi_3}}$, leads to
\begin{align*}
\frac{1}{(\theta_3-\varphi_3)R_2^2} \left(\frac{r_2}{R_2}\right)^{\frac{2\pi}{\theta_3-\varphi_3}}\lesssim\left|\frac{{\rm d} w}{{\rm d} z}\right|\lesssim
\frac{1}{(\theta_3-\varphi_3)r_2^2}
\end{align*}
for $z\in U_2$. Denoting the numbers $r_j$ in Lemma \ref{lemma Levin} by $\rho_j$ instead (with $\rho_3=1$), we then find, using \eqref{Koebe distortion},
\begin{align*}
1-\rho_2&\gtrsim \frac{{\rm d}(\partial U_2,\partial U_3)}{(\theta_3-\varphi_3)R_2^2} \left(\frac{r_2}{R_2}\right)^{\frac{2\pi}{\theta_3-\varphi_3}},\\
\rho_2-\rho_1&\gtrsim \frac{{\rm d}(\partial U_1,\partial U_3)}{(\theta_3-\varphi_3)R_2^2} \left(\frac{r_2}{R_2}\right)^{\frac{2\pi}{\theta_3-\varphi_3}}-\frac{{\rm d}(\partial U_2,\partial U_3)}{(\theta_3-\varphi_3)r_2^2}
\gtrsim \frac{{\rm d}(\partial U_1,\partial U_3)}{(\theta_3-\varphi_3)R_2^2} \left(\frac{r_2}{R_2}\right)^{\frac{2\pi}{\theta_3-\varphi_3}},
\end{align*}
where in the second line we used the triangle inequality and the second inequality in \eqref{condition lemma wedge}. By the third inequality in \eqref{condition lemma wedge} we can Taylor expand
\begin{align*}
\log(\frac{1}{\rho_2})=-\log(1-(1-\rho_2))\asymp 1-\rho_2.
\end{align*}
Lemma \ref{lemma Levin} now yields the claim.
\end{proof}
\subsection{Distribution function}\label{subsection Distribution function}
For $s>0$, we define
\begin{align*}
h_L(s)=|\set{k\in [N]}{\eta_0 L_k\leq 1/s}|\in {\mathbb Z}_+,
\end{align*}
where $\eta_0^{-1}$ is an arbitrary length scale. Note that $h_L$ is decreasing and tends to infinity as $s\to 0$.
In fact, $h_L$ is the distribution function of the sequence $(\eta_0L_k)^{-1}$. Since we assume that $L_k$ is increasing, we also have
\begin{align*}
h_L(s)=\min\set{k\in{\mathbb Z}_+}{\eta_0 L_{k+1}> 1/s}.
\end{align*}
We will show that, under the assumption
\begin{align}\label{assumption on h_L}
\exists\lambda\in (0,1)\quad \mbox{such that}\quad\limsup_{s\to 0+}\frac{h_L(\lambda s)}{{\rm e}\, h_L(s)}<1,
\end{align}
the potential $V(L)$ is strongly separating in the sense of Definition \ref{def. sparse}.
\begin{proposition}\label{lemma distribution function}
Assume \eqref{assumption on h_L}. Then
\begin{align}\label{bound of lemma distribution function}
\mathrm{sep}(L,\eta)\lesssim\exp(-\eta L_1)\langle h_{L}(\eta/\eta_0)\rangle.
\end{align}
\end{proposition}
In particular, this implies that the examples in Subsection \ref{subsection Separating and sparse potentials} are strongly separating:
\begin{itemize}
\item[a)] If $\eta_0L_k\gtrsim k^{\alpha}$ for $\alpha>0$, then $h_L(s)\lesssim s^{-1/\alpha}$.
\item[b)] If $\eta_0L_k\gtrsim\exp(k)$, then $h_L(s)\lesssim\log(1/s)$.
\item[c)] If $\eta_0L_k\gtrsim\exp(\exp(k))$, then $h_L(s)\lesssim\log\log(1/s)$.
\end{itemize}
\begin{lemma}\label{lemma s vs delta s}
Assume \eqref{assumption on h_L}. Then for any $\delta>0$ and for all $s>0$,
\begin{align*}
\langle h_L(\delta s)\rangle\lesssim_{\delta} \langle h_L(s)\rangle.
\end{align*}
\end{lemma}
\begin{proof}
We may restrict our attention to the case $\delta<1$ as the case $\delta\geq 1$ is trivial. By \eqref{assumption on h_L} there exist $\lambda\in (0,1)$ and $s_0>0$ such that
\begin{align}\label{h_L(s) vs h_L(delta s)}
h_L(\lambda s)<{\rm e}\, h_L(s)
\end{align}
holds for all $s\in (0,s_0]$. Now let $n$ be the smallest integer such that $\lambda^n\leq \delta$. Iterating \eqref{h_L(s) vs h_L(delta s)} $n$ times, we get
\begin{align*}
h_L(\delta s)<{\rm e}^n h_L(s)
\end{align*}
for all $s\in (0,s_0]$. For $s>s_0$, the inequality holds trivially.
\end{proof}
\begin{proof}[Proof of Proposition \ref{lemma distribution function}] Without loss of generality we may assume that $\eta=\eta_0$.
We first consider the case $\eta L_1\leq 1$, and hence $h_L(1)\geq 1$.
Then
\begin{align*}
\sum_{k=1}^{\infty}\exp(-\eta L_k)\leq {h_L(1)}\exp(-\eta L_1)+\sum_{k=h_L(1)}^{\infty}\exp(-\eta L_k).
\end{align*}
It remains to show that the second term is bounded by the right hand side of \eqref{bound of lemma distribution function}. To this end, we decompose the sum into dyadic intervals $I_j=[h_L(2^{-j}),h_L(2^{-j-1})]$, $j\in {\mathbb Z}_+$. Then
\begin{align*}
\sum_{k\in I_j}\exp(-\eta L_k)\leq \exp(-2^j)h_L(2^{-j-1}).
\end{align*}
Summing over $j$ and using Cauchy's condensation test yields
\begin{align}\label{Cauchy condensation}
\sum_{k=h_L(1)}^{\infty}\exp(-\eta L_k)\lesssim \sum_{n=1}^{\infty}\exp(-n)h_L(\tfrac{1}{2n}).
\end{align}
By the quotient test, the series converges provided that
\begin{align}\label{quotient test}
\limsup_{n\to\infty}\frac{h_L(\lambda_n\tfrac{1}{2n})}{{\rm e}\,h_L(\tfrac{1}{2n})}<1,
\end{align}
where $\lambda_n=\tfrac{n}{n+1}$.
But this follows from assumption \eqref{assumption on h_L}. Indeed, since $\lambda_n\to 1$, we have $\lambda< \lambda_n$ for large $n$ and hence $h_L(\lambda_n\tfrac{1}{2n})\leq h_L(\lambda\tfrac{1}{2n})$. The series~\eqref{Cauchy condensation} is thus bounded by $\langle h_L(1)\rangle$, where we have used Lemma \ref{lemma s vs delta s} with $\delta=1/2$. This proves \eqref{bound of lemma distribution function} in the case $\eta L_1\leq 1$. The case $\eta L_1>1$ is similar, but \eqref{Cauchy condensation} is bounded by $\exp(-n_0)h_L(\tfrac{1}{2n_0})$, where $n_0$ is the least integer such that $h_L(\tfrac{1}{2n_0})\geq 1$, i.e.\ $n_0=\lceil\tfrac{\eta L_1}{2}\rceil$. Another application of Lemma \ref{lemma s vs delta s} completes the proof.
\end{proof}
\bibliographystyle{abbrv}
|
\section{Introduction}\label{section:introduction}
Credit rating systems have been widely employed in all kinds of financial institutions such as banks and agency securities, to help them to mitigate credit risk. However, the credit assessment process, which is expensive and complicated, often takes months with experts involved to analyze. Therefore, it is necessary to design a model to predict credit level automatically.
The banking industry has developed some credit risk models since the middle of the twentieth century. Initially, logistic regression and hidden Markov model were applied in credit rating \cite{gogas2014forecasting,petropoulos2016novel}. With the progress of science and technology, machine learning, deep learning and hybrid models have shown their power in this field \cite{wu2012credit,yeh2012hybrid,pai2015credit}. With the advent of graph neural networks, some graph-based models \cite{bruss2019deeptrax,cheng2019dynamic,cheng2019risk,cheng2020contagious,cheng2020spatio}, were built based on the loan guarantee network to utilize relations between corporations.
Although these models have achieved promising results, they all ignore a fatal problem, the class imbalance problem, which prevents these models from improving continuously. We find out that credit rating agencies, like Standard \& Poor's, Moody's and CCXI, often give AA, A, BBB to most corporations, and few of them are CC, C. This will incur the class-imbalanced problem, which poses a great challenge to corporate credit rating model.
When the supervised model is limited, these models may fail due to the class-imbalanced problem, self-supervised manners will show their power. Recently, contrastive self-supervised learning brings a new chance to this problem. In the work \cite{yang2020rethinking}, they systematically investigate and demonstrate, theoretically and empirically, the class-imbalanced learning can significantly benefit in contrastive self-supervised manners.
Contrastive self-supervised learning attracted the interest of many researchers due to its ability to avoid the cost of annotating large-scale datasets. Mikolov et al. \cite{mikolov2013distributed} firstly proposed contrastive learning for natural language processing in 2013. And it started to prevail on several NLP tasks in recent years \cite{chi2020infoxlm,fang2020cert,giorgi2020declutr}. On the one hand, contrastive self-supervised learning had shown its power by four main pretext tasks: color transformation, geometric transformation, context-based tasks, and cross-modal tasks \cite{jaiswal2020survey}. On the other hand, in the field of recommendation systems, Yao et al. \cite{Yao2020SelfsupervisedLF} proposed a multi-task self-supervised learning framework for large-scale recommendations. Xie et al. \cite{xie2020contrastive} developed CP4Rec to utilize the contrastive pre-training framework to extract meaningful sequential information. Besides, it can also alleviate generalization error, spurious correlations and adversarial attacks. That's why this kind of technology is applied in many fields.
Due to the limitations mentioned before, the application of contrastive self-supervised learning in the corporate credit rating is less well study. Different from previous works, we aim to tackle the class-imbalanced problem and improve the model performance of the class with few samples by leverage self-supervised signals constructed from corporate profile data. In specific, we propose a novel framework named Contrastive Pre-training for Corporate Credit Rating, CP4CCR for brevity. In the pre-training phase, we propose two self-supervised tasks, (i) Feature Masking and (ii) Feature Swapping (FS), to perform self-supervised pre-training. Then in the following phase, any standard corporate credit rating model can perform based on the pre-trained data. To sum up, the main contributions of this work are summarized as follows:
\begin{itemize}
\item[$\bullet$]Two novel unsupervised tasks of corporate credit rating are proposed to perform contrastive pre-training: (i) Feature Masking and (ii) Feature Swapping (FS).
\item[$\bullet$]We proposed a new framework named \textbf{C}ontrastive \textbf{P}re-training for \textbf{C}orporate \textbf{C}redit \textbf{R}ating (\textbf{CP4CCR}) to tackle the class-imbalanced problem.
\item[$\bullet$]Comprehensive experiments on the real public-listed corporate rating dataset demonstrate that our model can improve the performance of common credit rating models experimentally.
\end{itemize}
\section{The Proposed Method: CP4CCR}\label{section:TheProposedMethod}
In this section, we introduce the proposed framework: CP4CCR. It includes two phases. The architecture is illustrated in Figure \ref{figure1}.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.21]{figure/1.pdf}
\caption{The architecture of the CP4CCR framework.}
\label{figure1}
\end{figure}
In the first phase, the original dataset without label information is performed contrastive self-supervised pre-training to learn better initialization. Then to make full use of the original and pre-trained dataset, vector space stacking is proposed. We get the better-transformed dataset at the end of phase 1. Afterward, any standard corporate credit rating model can be applied based on the transformed dataset to predict the credit level in phase 2.
\subsection{Notations and Problem Statement}
In this paper, we represent column vectors and matrices by italic lower case letters (e.g., $c$) and bold upper case letters (e.g., \bm{$X$}), respectively. And we use calligraphic letters to represent sets (e.g., $\mathcal{C}$).
Let $\mathcal{C}=\{c_1,c_2,\cdots,c_n\}$ denotes the set of corporations, $n$ is the number of corporations. $c\in \mathbb{R}^d$ includes the profile of corporation, which has $d$ dimensions feature. Every corporation has a corresponding label that represents its credit level. Let $\mathcal{Y}=\{y_1,y_2,\cdots,y_m\}$ denotes the set of labels, and $m$ is the number of all unique labels. Corporate credit rating models aim to predict the credit level $\hat{y}$ according the profile of corporation $c$.
\subsection{Contrastive Self-supervised Pre-training}
The architecture of the Contrastive Self-supervised Pre-training is described as Figure \ref{figure2}. Positive samples and negative samples are generated by self-supervised learning tasks and negative sampling, respectively. They are all mapped into another vector space through an encoder function $Encoder(\cdot)$ to get better initialization. Finally, the encoder model is trained by constructing the contrastive loss between positive samples and negative samples.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.35]{figure/2.pdf}
\caption{The architecture of the Contrastive Self-superivised Pre-training.}
\label{figure2}
\end{figure}
\noindent
\textit{\textbf{Self-supervised Learning Tasks.}} Inspired by masked-LM used in pre-training for NLP and unordered nature of corporation profile, we propose two methods for generating positive samples: Feature Masking (FM), Feature Swapping (FS), and design three self-supervised learning tasks: FS, FM, FS+FM. During the FS task, we can randomly swap the location of features for $@k$ times due to the unordered nature of the corporation profile to build the self-supervised task. In addition, as demonstrated in figure \ref{figure2}, FM is designed to mask a subset of input features, therefore, they can only train by partial features $@k$. FS+FM is a more complicated task by using FS and FM simultaneously.
\\
\noindent
\textit{\textbf{Encoder Function.}} Theoretically, any encoder function which could map a vector space into another vector space can be applied to our framework. In our experiments, we use MLP and Transformer. MLP is easy and simple while Transformer is more powerful but more complicated. It is defined as follows:
\begin{equation}
c^{\prime} = Encoder(c)
\end{equation}
where $c^{\prime}\in \mathbb{R}^{d^{\prime}}$, $c^{\prime}$ is the representation of corporation in new vector space, $d^{\prime}$ is the dimension of new vector space.
\\
\noindent
\textit{\textbf{Negative Samples and Contrastive Loss.}} For negative samples, we can randomly select $@L$ sample which different from corresponding positive samples. Here we use cosine similarity to measure the distance between the embeddings of two samples. And InfoNce loss is applied to build contrastive loss. The \textit{self-supervised pre-training }(SSP) loss can be represented as follows:
\begin{equation}
sim(u,v)=\frac{u\cdot v}{ \Vert u\Vert \Vert v\Vert}
\end{equation}
\begin{equation}
\mathcal{L}_{ssp} = -log\frac{\exp \left(\frac{sim(c^{\prime},pos^{\prime})}{\tau}\right)}{\exp \left(\frac{sim(c^{\prime},pos^{\prime})}{\tau}\right) + \sum_{l=0}^L\exp \left(\frac{sim(c^{\prime},neg_l^{\prime})}{\tau}\right)}
\end{equation}
\subsection{Vector Space Stacking}
To fully use the information of data, we propose two methods of vector space stacking: Space Concatenation (SC) and Space Fusion (SF). They can be formulated as follows:
\begin{equation}
x_{sc} = Concatenation(c,c^{\prime})
\end{equation}
\begin{equation}
x_{sf} = \alpha c +(1-\alpha)c^{\prime}
\end{equation}
where $x_{sc}\in \mathbb{R}^{d+d^{\prime}}$, and $x_{fc}\in \mathbb{R}^{d}$ due to $d=d^{\prime}$.
When the new vector space and origin vector space are in different space, in other words, $d^{\prime}\neq d$, we can concatenate them together to form a large space. In this space, both origin data and pre-trained data can be used to predict the credit level. What's more, if $d^{\prime} = d$, not only Space Concatenation but also Space Fusion can be used to mix two spaces. When $\alpha=1$ CP4CCR will degenerate into the common credit rating model. While $\alpha=1$, only pre-trained data will be used. The reason why we design this structure will be explained in Section \ref{section:Experiment} (Experiment).
\subsection{Phase 2}
After the pre-training in phase 1, the better initialization of data will be got. In this vector space, similar samples will be closer, and dissimilar samples will be far from each other. The border between different classes will be wider. Therefore, this distribution of data will be easy for standard credit rating models to learn, especially for the class with few samples. The result of experiments will demonstrate this point. The training process can be formulated as follows.
\begin{equation}
z =CreditRatingModel(x)
\end{equation}
\begin{equation}
\hat{y} = log\_softmax(z)
\end{equation}
\begin{equation}
\mathcal{L}(\hat{y})=-\sum\limits_{i=1}^{n}{{{y}_{i}}\log ({{{\hat{y}}}_{i}})+(1-{{y}_{i}})\log (1-{{{\hat{y}}}_{i}})}+\lambda {{\left\| \Delta \right\|}^{2}}
\end{equation}
where $y$ denotes the one-hot encoding vector of ground truth item, $\lambda$ is parameter-specific regularization hyperparameters to prevent overfitting, and the model parameters are $\Delta$. The Back-Propagation algorithm is performed to train the model.
\section{Experiment}\label{section:Experiment}
In this section, we aim to answer the following three questions:
\textbf{RQ1.} Could the proposed CP4CCR improve the performance of standard credit rating models?
\textbf{RQ2.} How the proposed CP4CCR infect the data distribution of class with few samples?
\textbf{RQ3.} How do different types of vector space stacking affect the model performance?
\subsection{Experimental Configurations}
\textit{\textbf{Dataset.}} We evaluate the proposed CP4CCR on the corporate credit dataset. It has been built based on the annual financial statements of Chinese listed companies and China Stock Market \& Accounting Research Database (CSMRA). The results of credit ratings are conducted by the famous credit rating agencies, including CCXI, China Lianhe Credit Rating (CLCR), etc. The financial data of the corporation includes six aspects: profit capability, operation capability, growth capability, repayment capability, cash flow capability, Dupont identity. After the same preprocess as work \cite{feng2020corporation},we get 39 features and 9 rating labels: AAA, AA, A, BBB, BB, B, CCC, CC, C.
\\
\noindent
\textit{\textbf{Baselines.}} To evaluate the performance of the proposed CP4CCR, we select several models including KNN, Logistic Regression (LR), Random Forest (RF), Decision Tree (DT), GBDT, AdaBoost, GaussianNB, LDA, SVM (linear), SVM (rbf), MLP and Xgboost due to the particularity of the financial field. But we believe our framework will have more or less improvement based on other corporate credit rating models.
\\
\noindent
\textit{\textbf{Evaluation Metrics.}} We adopt three commonly-used metrics for evaluation, including Recall, Accuracy, and F1-score. Let $\mathcal{R}$ denote the set of credit rating models, and $\mathcal{M}$ is the set of metrics. For comparing overall improvement on all kinds of rating models, we propose a metric, \underline{O}ver\underline{a}ll \underline{P}erformance (OaP), which can be computed as follows:
\begin{equation}
OaP = \sum_{r \in \{\mathcal{R}\}} \sum_{t \in \{\mathcal{M}\}} {SSP(r,t)-Origin(r,t)}
\end{equation}
where $SSP(r,t)$ means the metric $t$ of model $r$ under self-supervised pre-training, while $Origin(r,t)$ uses the origin dataset. On the whole, $OaP>0$ means CP4CCR has a positive gain on all models, negative gain vice versa.
\\
\noindent
\textit{\textbf{Hyperparameter Setup.}} During the data preprocessing, we perform the Feature Masking, Feature Swapping and FS+FM respectively. $k$ is set to 4, and negative samples $L$ is set to 8. In the experiment, we use the MLP or Transformer as the encoder function. The initial learning rate for Adam is set to 0.001 and will decay by 0.00001.
\subsection{Improvement with Baseline Methods (RQ1)}
\begin{table}[!hbtp]
\centering
\caption{The Overall Performance of Experiments.}
\label{table2}
\begin{tabular}{ccc}
\toprule[1.5pt]
& { \textbf{MLP}} & {\textbf{Transformer}} \\ \hline
{ \textbf{FS}} & 0.60619 & 1.01328\\
{ \textbf{FM}} & 0.81989 & 1.03799\\
{ \textbf{FS+FM}} & 0.86575 & 1.04080\\
\toprule[1.5pt]
\end{tabular}
\end{table}
To demonstrate the improvement of our proposed framework CP4CCR, we conduct experiments based on the baseline methods. The origin denotes the data without pre-training. Data are pre-trained by three different self-supervised tasks:@FS, @FM, @FS+FM. We find that most baselines are improved by pre-training. Due to the space limit, we omit the sophisticated original result. For clearly verifying, the overall performance (OaP) is displayed in Table \ref{table2}. Both two encoder functions have a positive gain for data pre-training while the transformer encoder is better. FM is superior to FS, and perform them simultaneously will bring better results.
Besides we find another interesting result, in terms of the powerful encoder (Transformer), different types of pretext tasks bring almost the same positive gain for the model. However, a normal encoder (MLP) has a relative gap between different pretext tasks. It is might that a more powerful encoder is more robust to the pretext tasks. Therefore, we should pay more attention to the design of pretext tasks for the normal encoder, while for the powerful encoder, concentrating on improving the ability of it is a better way.
\subsection{Influence on the Class Imbalance(RQ2)}
To clearly demonstrate the influence on the class-imbalanced problem, we perform the t-SNE visualization on the original and pre-trained data.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.25]{figure/4.pdf}
\caption{t-SNE visualization of origin data and pretrained data by FS, FM and FS+FM}
\label{figure3}
\end{figure}
From Figure \ref{figure3}, the original data is usually chaotic. But after contrastive self-supervised learning, the imbalanced problem is mitigated, which results in clearer boundary and better initialization. For example, we can find that there are clear boundaries with feature swapping self-supervised tasks from Figure \ref{figure3} (b). And in Figure \ref{figure3} (c) yellow and brown samples get better initialization compared with original data to easily perform downstream tasks (Corporate Credit Rating).
\subsection{Different Vector Space Stackings (RQ3)}
To answer the \textbf{RQ3}, in this section, we analyze how hyper-parameter $\alpha$ impacts the metric OvP when the dimensions of origin and new vector space are the same, in other words, $d=d^{\prime}$. The following Figure \ref{figure4} shows the result.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.28]{figure/3.pdf}
\caption{The influence on the Space Fusion}
\label{figure4}
\end{figure}
We can find that when $\alpha$ is too small, space fusion even results in a negative gain. It gets better performance when $\alpha$ is about 0.9. And in the whole, the MLP encoder is stable, while the transformer encoder performs better but bigger variance.
What's more, we analyze the influence on space concatenation when $d\neq d^{\prime}$, the variation of OaP follows the changes of encoding dimension $d^{\prime}$. From Figure \ref{figure5}, we can draw some conclusions. When $d^{\prime}$ is less than 64, the Transformer encoder performs better than MLP. As $d^{\prime}$ increasing, larger vector space brings more positive gain in MLP than Transformer. We believe that increasing $d^{\prime}$ leads to more parameters for Transformer encoder which converges difficultly. In addition to considering the time of cost, we believe that a relatively satisfied result is got when $d=d^{\prime}$ with space concatenation.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.2]{figure/test.pdf}
\caption{The influence on the Space Concatenation}
\label{figure5}
\end{figure}
All in all, both FS and FM have positive gain to get better initialization and clearer boundary. Then Transformer is better than MLP to be the encoder function in the majority of cases. Finally, in terms of different space stackings, space concatenation exactly a great way to mix the original and new data information.
\section{Conclusion}\label{section:Conclusion}
In this work, we propose a novel framework named Contrastive Pre-training for Imbalanced Corporate Credit Rating (CP4CCR), which can get better initialization and clearer boundary by contrastive self-supervised learning. In specific, we proposed two self-supervised tasks: Feature Masking and Feature Swapping, and two types of vector space stacking: Space Concatenation and Space Fusion. Extensive experiments demonstrate that our CP4CCR can improve the performance of baseline methods by pre-trained data. What's more, Transformer is a better encoder function for contrastive self-supervised pre-training. In terms of space stackings, space concatenation is still a better way. We believe this work will bring a new framework to tackle imbalanced corporate credit rating in reality.
\bibliographystyle{ACM-Reference-Format}
|
\subsection{Motivating Examples}
\input{motivation}
\subsection{Online Worker Selection in Crowdsourcing Platforms}
Consider a crowdsourcing platform that provides data labelling service for a payment. The platform interact with a group of customers and a pool of workers. The customers request tasks to the platform for labeling and the platform is then responsible of selecting workers from the worker pool to complete the tasks. This process repeats, during which the platform learns the best deployment. We maintain a mild assumption that the payments from customers are non-decreasing with the quality of labels. To maximize its profit, it is desired for the platform to select workers that provides most reliable labels. The workflow of the platform can be described in the diagram below.
While the platform and the customers desire quality labels, it may not entirely be in the worker's interest to exert highest effort and thus report quality labels each time, which factors into a range of reasons. Workers may adapt strategic behaviors to maximum their own utility instead. Thus, it becomes crucial to identify the reliable workers to prevent strategic behaviors that jeopardize profits.
This naturally translate to a combinatorial multi-armed bandits problem under strategic manipulations where the workers are the strategic arm and the payments acts as the rewards.
\begin{figure}[H]
\centering
\label{crowdsys}
\includegraphics[width=7.5cm]{figs/crowdsys.png}
\caption{Online crowdsourcing system}
\end{figure}
\subsection{Online Influence Maximization}
For a social network that can be represented by the below directed graph where white dots are people and black dots labeled with ``SI'' are social influencers. The temporal nature of social network results in each directed edges exists with a probability at each time step. Formally, a node is said to be reached by a seed node if there it has a directed path from the seed node. Note that this definition considers a diffusion process where a node can be reached by a seed node through multiple edges. Given a set of seed nodes, we define a spread across social network at a given time $t$ by the expected number of nodes that can be reached by the nodes in the seed node set at that time. Intuitively, this models information spread among social networks and reflects scenarios like viral marketing.
\begin{figure}[H]
\centering
\includegraphics[scale = 0.35]{figs/example.png}
\caption{Probabilistic social network}
\end{figure}
In the case of product marketing, a company may need to select the most influential social influencers to collaborate with to maximize the expected spread across time. This can be expressed as an online influence maximization problem as the company seek to maximize the size of population reached by promotions across number of promotions. However, in many cases social influencers may report inaccurate information regarding their social influential ability, which may corresponds to a number of ways in real applications, such as faking follower numbers on social media platforms.
\subsection{Suboptimal pulls maximization}
\begin{claim}
Under a strategic stochastic multi-armed bandit problem, when the strategic suboptimal arms collude to maximize the number of suboptimal pulls of the UCB algorithm with confidence parameter $\sqrt{\frac{3\log t_i}{K_{i,t-1}}}$, the maximum number of suboptimal pulls is the solution of
\begin{equation*}
\begin{aligned}
& \underset{Y_{i},t_i,\ i\in [m]}{\mathrm{maximize}}
& & Y_{1}+\dots+Y_{m} \\
& \mathrm{subject}\text{ }\mathrm{to}
& & \frac{B_i}{Y_{i}} + \sqrt{\frac{3\log t_i}{Y_{i}}} \ge \delta_i + \sqrt{3\log t_i},\ i \in [m]\,, \\
& & & t_{i_j} \ge \sum_{l=1}^j Y_{i_l}, \ j\in [m]\,,\\
& & & \textup{for some permutation}\ i_1,\dots\ i_m\ \textup{of}\ [m]\,.\\
\end{aligned}
\end{equation*}
\end{claim}
In the program, $Y_i$ represents the total number of times arm $i$ is pulled, which corresponds to $K_{i,T-1}$ used in previous sections. The permutation $i_1,...,i_m$ then describes, before the optimal arm is first pulled, the chronological order of the suboptimal arms to complete their last pull.
Observe that the confidence parameter of the optimal arm
increases with respect to $t$, the inequality holds for $t_i$ if and only if it holds for all $t\le t_i$.
By the same observation that the advantage of the optimal arm increases with time,
strategic arms who optimally spend their budgets for the maximum amount of pulls will also maximize their number of pulls till the first optimal arm pull.
The inequality condition of the combinatorial program is thus $\frac{B_i}{K_{i,t-1}} + \sqrt{\frac{3\log t_i}{K_{i,t-1}}} \ge \delta_i + \sqrt{3\log t_i}$.
We term the class of approximately optimal strategy as `sequential pull strategy', in which the strategic arms deceit the learning algorithm to pull suboptimal arms until the depletion of the budgets and this class includes the LSI strategy.
\section{Introduction}
\label{sec:intro}
Sequential learning methods feature prominently in a range of real applications such as online recommendation systems, crowdsourcing systems, and online influence maximization problems. Among those methods, the multi-armed bandits problem serves as a fundamental framework. Its simple yet powerful model characterizes the dilemma of exploration and exploitation which is critical to the understanding of online sequential learning problems and their applications \citep{10.2307/2332286, Robbins:1952,auer2002finite,lattimore_szepesvari_2020}.
The model describes an iterative game constituted by a bandit algorithm and many arms. The bandit algorithm is required to, through a horizon $T$, choose an arm to pull at each time step. As the objective is to maximize the cumulative reward over time, the algorithm balances between exploiting immediate rewards based on the information collected or pulling less explored arms to gain more information about arms \citep{anantharam_asymptotically_1987,auer_nonstochastic_2002,cesa-bianchi_prediction_2006}.
Out of the real applications, many motivate the extension of MAB towards combinatorial multi-armed bandits (CMAB), where multiple arms can be selected in each round \citep{pmlr-v28-chen13a,DBLP:conf/nips/CombesSPL15,pmlr-v48-lif16,pmlr-v97-zimmert19a,pmlr-v117-rejwan20a}. CMAB demonstrates its effectiveness on problems like online social influence maximization, viral marketing, and advertisement placement, within which many offline variants are NP-hard. However, existing MAB and CMAB algorithms are often developed either under benign assumptions on the arms \citep{pmlr-v28-chen13a,pmlr-v38-kveton15,pmlr-v75-wei18a} or purely adversarial arms \citep{auer2002nonstochastic}. In the former setting, the arms are commonly assumed to report their reward signals truthfully without any strategic behavior, under which the drawbacks are apparent. In the latter setting, arms can attack any deployed algorithm to regret of $O(T)$ with this capability of reward manipulations, which is catastrophic for a bandit algorithm.
This assumption is stringent and rarely realistic.
In this paper, we adapt the combinatorial UCB (CUCB) algorithm with a carefully designed UCB-based exploration term.
A major difficulty stems from not knowing the manipulation term made by the arms, while our algorithm overcomes this by depending only on the knowledge of the maximum possible strategic budget.
Previous results by \citep{pmlr-v119-feng20c} only implies robustness of UCB style algorithm under stochastic multi armed bandits setting under only full knowledge of step wise deployment of strategic budget.
New tail bounds over the proposed exploration term and a new trade-off parameter that balances exploration and exploitation are utilized to facilitate the analysis of our algorithm.
We further establish results on the robustness of our UCB variant under strategic arms with an algorithm-dependent budget lower bound.
Our proposed algorithms are also evaluated empirically through an extensive set of synthetic environments and real datasets.
The real applications include reliable workers selection in online crowdsourcing systems, where workers might misrepresent a result for a better chance to be selected in the future; online information maximization, where nodes modify the spread to include itself into the seed set; and online recommendation systems, which characterizes the ``click framing'' behavior. Through a wide range of tasks and parameters, the experiment results corroborate our theoretical findings and demonstrate the effectiveness of our algorithms.
\subsection{Motivating Examples}
The setting of strategic manipulations describes the strategic behavior found in a variety of real applications.
Consider a crowdsourcing platform that provides a data labeling service for payment. The platform interacts with a group of customers and a pool of workers. The customers request tasks to the platform for labeling and the platform is then responsible for selecting workers from the worker pool to complete the tasks. This process repeats, during which the platform learns the best deployment. We maintain a mild assumption that the payments from customers are non-decreasing with the quality of labels. To maximize its profit, it is desired for the platform to select workers that provides the most reliable labels. The workflow of the platform can be described in the diagram below.
While the platform and the customers desire quality labels, it may not entirely be in the worker's interest to exert the highest effort and thus report quality labels each time, which factors into a range of reasons. Workers may adapt strategic behaviors to maximize their own utility instead. Thus, it becomes crucial to identify reliable workers to prevent strategic behaviors that jeopardize profits.
This naturally translates to a multi-armed bandits problem under strategic manipulations where the workers are the strategic arm and the payments act as the rewards. Under ideal assumptions it had shown the effectiveness of bandits algorithms on such problems \citep{jain2014quality, tran2014efficient,rangi2018multi}.
\begin{figure}[H]
\includegraphics[width=7.5cm]{figs/crowdsys.png}
\caption{Online crowdsourcing system}\label{crowdsys}
\end{figure}
\section{Related Work}
The problem of multi-armed bandits (MAB) was first investigated back in 1952 while some techniques utilized were developed back in 1933 \citep{10.2307/2332286, Robbins:1952, berry,auer2002finite,lattimore_szepesvari_2020}. Since then it has been extensively explored and serves as the foundation of many modern areas, including reinforcement learning, recommendation systems, graph algorithms, etc. \citep{DBLP:conf/iconip/BouneffoufBG12,pmlr-v70-vaswani17a, pmlr-v85-durand18a,10.5555/3367243.3367445,li2020stochastic}. With the need to model the selection of multiple arms in one round, MAB is then extended to combinatorial multi-armed bandits (CMAB), which see many deployments in real applications like news/goods recommendation, medical trials, routing, and so forth \citep{Wang_Ouyang_Wang_Chen_Asamov_Chang_2017,8022964,pmlr-v84-sankararaman18a, li2020online}.
Real applications motivate the community to derive algorithms in more adaptive environments. The first of which considers adversarial bandits with the classic EXP3 algorithm. The arms are assumed to be non-stationary but non-adaptive (which means that algorithms will adapt to the adversarial) \cite{auer2002nonstochastic}. Despite that adversarial bandits do not fall into the scale of this paper's related work, it leads tremendous effort to the following topics in adaptive arms.
\paragraph{Adversarial corruptions}
The adversary is given the ability to replace the observations of the bandit algorithm with arbitrary values within some constraints.
\citet{pmlr-v99-gupta19a} discuss the constraint that rewards are modified for at most $T_0$ rounds out of the total $T$ rounds. The asymptotic regret bound $O(mT_0)$ achieved in the work is shown to match the negative result, where $m$ is the number of arms. \citet{lykouris_stochastic_2018} discussed a different constraint where the corruption $z_t$ at time $t$ is cumulative up to some constant $B$.
The regret achieved under this constrained corruption is bounded by $O(mB\log(\frac{mT}{\delta}))$ with probability at least $1 - \delta$. \citet{JMLR:v20:18-395} restricts corruptions as independent Bernoulli events with probability $b$, while if corruption happens the reward becomes arbitrary and adversarial. This problem is addressed with median statistics with gap-dependent matching regret upper and lower bounds of $O(\sum_{i \neq i^{\ast}} \frac{1}{\Delta_i}\log(\frac{K}{\Delta_i}))$, where $i^{\ast}$ denotes the optimal arm and $\Delta_i$ is the suboptimality of arm $i$.
Several lines of research also discuss corruption topics in frequentist inference and partial monitoring \citep{pmlr-v83-gajane18a}, best arm identification \citep{pmlr-v99-gupta19a}, and adversarial bandits \citep{NEURIPS2020_e655c771}.
\paragraph{Adversarial attacks}
Different from arms in adversarial corruptions who intend to maximize the regret of the bandit algorithm, arms in the adversarial attack setting have the goal to maximize their number of pulls. The setting is first considered by \citet{NEURIPS2018_85f007f8}, where it shows that the attacker may spend a budget of $O(\log T)$ to deceit UCB and $\epsilon$-greedy algorithms to pull the target arms for at least $T - o(T)$ times. Stochastic and contextual MAB algorithms also suffer from undesired loss from this adversarial attack, under a variety of attack strategies \citep{pmlr-v97-liu19e}. Linear MAB algorithms, however, are shown to be near-optimal up to some linear or quadratic terms on the attack budget $B$ \citep{garcelon_adversarial_2020,bogunovic_stochastic_2020}.
\paragraph{Strategic manipulations}
The setting of strategic manipulations further weakens the capability of the adaptive arms and prevents the bandit algorithms from being overcautious.
The objective of an arm is still utility maximization under strategic manipulations, while each arm works on its utility individually.
Instead of having a global coordinator for the adversarial attack strategy, the strategic arms seek the best response via the dominant Nash equilibrium.
The strategic manipulation setting is first studied by \citet{pmlr-v99-braverman19b} where each arm is assumed to keep a portion of the reward in each round and arms maximize the cumulative reward it keeps.
The performance of bandit algorithms will then be catastrophic under the dominant Nash equilibrium, receiving an expected cumulative reward of $0$ almost surely.
With the utility of maximizing the number of pulls for each strategic arm, common bandits algorithms are proved to be robust with guaranteed $O(\log T)$ regret bound, but only under constant strategic budgets \citep{pmlr-v119-feng20c}.
For $\omega(\log T)$ budgets, bandits under strategic manipulations remain an open problem and will be discussed in this paper.
\section{Problem Formulation}
We consider the problem of combinatorial multi-armed bandits (CMAB) under the setting of strategic arms. In strategic CMAB, each arm is given a budget and the arm can strategically increase its emitted reward signals cumulatively up to this budget \emph{for its own interest}. This problem is a Stackelberg game that involves two parties. The bandit learning algorithm is deployed first to maximize its cumulative reward under the best response of the followers. The $m$ strategic bandits arms then deplete their budget where each of them aims to maximize its expected number of pulls. Knowing the principal's strategy and other followers' budget information, the followers are assumed to place their strategies according to the dominant Nash equilibrium between them.
Formally, given a time horizon $T$, the the bandit learning algorithm is asked to pull a subset of arms, denoted by an arm subset $S_t \in \mathcal{S} $ at each time $t\in [T]$, where $\mathcal{S} = { \{0,1\}^{m}} $ is the set of all possible arm subsets. At the time $t$, based on the principal's choice of arm subset, stochastic rewards $x_{i,t}$ are generated by arm $i$ from the underlying $1$-sub-Gaussian distribution with mean $\mu_i$, independent of other arms in $S_t$. The principal does not observe these rewards. Instead, each arm can increase the emitted reward signal by an arbitrary amount $z_{i,t}\ge 0$, as long as the cumulative manipulation over the horizon does not exceed a given budget $B_i$. The principal then receives the semi-bandit feedback $\{x_{i,t}+z_{i,t}\}_{i\in S_t}$.
Let $\bm{\mu} = (\mu_1, \mu_2, ..., \mu_i) $ be the vector of expectations of all arms. The expected reward of playing any arm subset S in
any round $r_{\bm{\mu}}(S) = \mathbb{E}[R_t(S)]$,which is a function of arm subset $S$ and $\bm{\mu}$. The reward function $r_{\bm{\mu}}(S)$ is assumed to satisfy two axioms:
\begin{itemize}
\item \textbf{Monotonicity.} The expected reward of playing any arm subset $S \in \mathcal{S}$ is monotonically non-decreasing with respect to the expected reward vector. That is, if for all $i \in [m], \mu_i \leq \mu_{i}'$, then $r_{\bm{\mu}}(S) \leq r_{\bm{\mu}'}(S)$.
\item \textbf{Bounded smoothness.} There exists a strictly increasing function $f(\cdot)$, termed the bounded smoothness function, such that for any two expected reward vectors $\bm{\mu}$ and $\bm{\mu'}$ satisfying $\|\bm{\mu}-\bm{\mu'}\|_{\infty}\le \Lambda $, we have $|r_{\bm{\mu}} - r_{\bm{\mu'}}| \leq f({\Lambda})$.
\end{itemize}
These axioms cover a rich set of reward functions and the explicit forms of $R_t(S)$ and $r_{\bm{\mu}}(S)$ are not needed to be specified \citep{pmlr-v28-chen13a}.
Without loss of generality, assume that $S^\ast = \argmax_{S\subseteq \mathcal{S}} r_{\bm{\mu}}(S)$ is the unique optimal subset of arms. When placing reward manipulations, each strategic arm has access to its own history $h_{i,t} = \{I_{i,t'}, x_{i,t'}, z_{i,t'}\}_{t'\ge 1}$, where $I_{i,t'}$ is the indicator of whether arm $i$ is pulled at time $t'$ and $t' < t$. The strategy of arm $i$ is determined by a function that maps this history to a manipulation $z_{i,t}$, as $Z_{i,t} \colon h_{i,t-1} \to \mathbb{R}$. Without loss of generality, we assume that arms in the optimal arm subset $i\in S^\ast$ have strategic budgets of $0$, which restricts their $z_{i,t}$ to be $0$.
In the combinatorial setting, even with the exact reward vector $\bm\mu$ provided, it can be hard to exactly compute the optimized $r_{\bm{\mu}}(S)$. In view of this, many have studied probabilistic approximation algorithms in combinatorial problems, which indicates that an ($\alpha,\beta$)-approximation oracle defined below can be usually available.
\begin{definition}[Approximation oracle]
Let $0\le \alpha,\beta \leq 1$ and define ${OPT}_{\bm{\mu}} = \max_{S\in \mathcal{S}} r_{\bm{\mu}}(S)$.
An oracle is called an ($\alpha,\beta$)-approximation oracle if it takes an expected reward vector $\bm{\mu}$ as input and outputs an arm subset $S \in \mathcal{S}$ such that $\mathbb{P}(r_{\bm{\mu}}(S) \geq \alpha \cdot {OPT}_{\bm{\mu}}) \geq \beta$. That is, the oracle gives an arm subset $S$ that is at least as good as $\alpha$ times the reward of an optimal arm subset with probability at least $\beta$.
\end{definition}
Denote ${S_B} = \{ r_{\bm{\mu}}(S) < \alpha \cdot {OPT}_{\bm{\mu}} | S \in \mathcal{S}\}$ to be the set of suboptimal arm subsets under the approximation oracle.
Note that a suboptimal arm subset can be given by the oracle for two reasons. The ($\alpha,\beta$)-approximation oracle can fail, which happens with probability at most $1-\beta$. The estimation of $\bm{\mu}$ can deviate from the true value by a significant amount, resulting in accurate input to the oracle.
The objective of the principal is to maximize the expected cumulative reward before manipulation over the time horizon $T$. Equivalently, the principal minimizes the regret, the cumulative difference between the scaled optimal reward and expected actual reward, as defined below.
\begin{definition}[Regret]
With access to an ($\alpha,\beta$)-\textit{approximation oracle}, the regret of a combinatorial bandit algorithm for $T$ rounds is
\[
Regret_{\bm{\mu},\alpha,\beta}(T) = T \cdot \alpha \cdot \beta \cdot {OPT}_{\bm{\mu}} - \mathbb{E}\big[\sum^T_{t=1}r_{\bm{\mu}}(S_t)\big]\,,
\]
where the randomness in $\mathbb{E}[\sum^T_{t=1}r_{\bm{\mu}}(S_t)]$ involves the stochasticity of the bandit algorithm and the oracle.
\end{definition}
The objective of each strategic arm $i\in [m]\setminus S^{\ast}$, however, is to maximize the number $\sum_{t=1}^T I_{i,t}$ of times it is pulled over the time horizon. To achieve this, the arm needs to confuse the principal by deviating the emitted reward signals up to the possessed budget.
\section{Strategic Combinatorial UCB}
We now propose a variant of combinatorial upper confidence bound algorithm that is robust to strategic manipulations of rewards in Algorithm \ref{algo}. The only mild assumption we maintain is that the learning algorithm has the knowledge of the largest budget possessed among all bandits arms, i.e, $B_{max}$. This is a relaxation of the assumptions made by \citet{pmlr-v119-feng20c} on the strategic UCB algorithm, in which the learning algorithm has access to the cumulative use of budget at every time step. We start with a detailed description of the algorithm and then analyze the theoretical upper bound of regret, which enjoys $O(m \log{T} + m B_{max})$.
For each arm $i$, our algorithm maintains a counter $K_{i,t-1}$ as the total number of times arm $i$ has been pulled up to time $t-1$ and $\widetilde{\mu}_{i,t} = \frac{\sum_{j=1}^{t-1} (x_{i,j}+ z_{i,j})}{K_{i,t-1}}$ as the empirical mean estimation based on the observations. At each time step, the algorithm computes the UCB estimation $\Bar{\mu}_{i,t}= \widetilde{\mu}_{i,t} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}} + \frac{B_{max}}{K_{i,t-1}}$ for $i \in [m]$. With $\Bar{\mu}_{i,t}$, the $(\alpha, \beta)$-approximation oracle then outputs an approximately optimal arm subset $S_t \in \mathcal{S}$. The algorithm plays the return arm subset and update the counter $K_{i,t}$ and the estimation $\widetilde{\mu}_{i,t}$ accordingly.
\begin{algorithm}[htb]
\DontPrintSemicolon
{\bf Input:}
Horizon $T$, number $m$ of arms, maximum budget $B_{max}$\;
{\bf Output:} Arm subset $S_t$\;
Initialize $K_{i,0} = 0$ and $\widetilde{\mu}_{i,0} = 0$ for $i \in [m]$\;
\For{t = 1 $\to$ m}{
Play an arbitrary arm subset $S_t \in \mathcal{S}$ such that $t \in S_t$\;
$K_{i,t} = K_{i,t-1} + 1$ for $i \in S_t$\;
$\widetilde{\mu}_j = x_{i,t}+z_{i,t}$ for $i \in S_t$ \;}
\For{t = m+1 $\to$ T}{
For $i \in [m]$, compute $\Bar{\mu}_{i,t}= \widetilde{\mu}_{i,t-1} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}} + \frac{B_{max}}{K_{i,t-1}} $\;
$S_t$ = {\textbf{oracle}}$(\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$\;
Play arm subset $S_t$ and update
$K_{i,t} = K_{i,t-1} + 1$ and $ \widetilde{\mu}_{i,t} = \frac{\sum_{j=1}^{t} (x_{i,j} + z_{i,j})}{K_{i,t}}$ for $i \in S_t$ \;
}
\caption{SCUCB}
\label{algo}
\end{algorithm}
We first introduce a few notations that are used in our results. Define, for any arm $i\in[m]$, the suboptimality gaps as
\begin{align*}
&\Delta^i_{min} = \alpha \cdot {OPT}_{\bm{\mu}} - \max\{r_{\bm{\mu}}(S) \mid S \in S_B, i \in S\}\,, \notag \\
&\Delta^i_{max} = \alpha \cdot {OPT}_{\bm{\mu}} - \min\{r_{\bm{\mu}}(S) \mid S \in S_B, i \in S\}\,.
\end{align*}
We then denote the maximum and minimum of the suboptimality gaps as $\Delta_{max} = \max_{i \in [m]} \Delta^i_{max}\notag $ and $\Delta_{min} = \min_{i \in [m]} \Delta^i_{min}$.
The following lemma re-establish the canonical tail bound inequality in UCB under the setting of strategic manipulations.
\begin{lemma}\label{lambda}
Let ${\Lambda}_{i,t} = \sqrt{\frac{3 \log{t} }{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}$, where $\rho_{i,t-1}$ is the total strategic budget spent by arm $i$ up to time $t-1$ and $K_{i,t-1}$ be the total number of pulls of arm $i$ up to time $t-1$. Define the event $E_t = \{ |\ \widetilde{\mu}_{i,t-1} - \mu_{i}| \leq \Lambda_{i,t}, \forall i \in [m]\}$, where $\mu_i$ is the true mean of arm $i$'s underlying distribution.
Then, $\mathbb{P}(\lnot E_t)\leq 2m \cdot t^{-2}$.
\end{lemma}
Armed with Lemma \ref{lambda}, we present one of our main theorems, Theorem \ref{appendixA}, which gives the regret bound of $O(\log T)$ of SCUCB. The outline of the proof follows that of CUCB by \citet{pmlr-v28-chen13a}. To complete the proof, we carefully choose $\psi_t$, which controls the trade-off between the exploration and exploitation periods.
\begin{theorem}
\label{appendixA}
The regret of the SCUCB algorithm with $m$ strategic arms in time horizon $T$ using an $(\alpha,\beta)$-approximation oracle is at most
\begin{align}
m\cdot\Delta_{max}\left( \frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}+1 \right)\,, \notag \nonumber
\end{align}
where $f^{-1}(\cdot)$ is the inverse bounded smoothness function.
\end{theorem}
\begin{proof}[Proof sketch]
We introduce a counter $N_i$ for each arm $i\in[m]$ after the $m$-round initialization and let $N_{i,t}$ be the value of $N_i$ at time $t$. We initialize $N_{i,m} = 1$. By definition, $\sum_{i \in [m]} N_{i,m} = m$. For $t > m$, the counter $N_{i,t}$ is updated as follows:
\begin{itemize}
\item If $S_t \in S_B$, then $N_{{i'},t} = N_{{i'},t-1} + 1$ where $i' =$ $\argmin_{i \in S_t}$ $N_{i,t-1}$. In the case that $i'$ is not unique, we break ties arbitrarily;
\item If $S_t \notin S_B$, then no counters will be updated.
\end{itemize}
As such, the total number of pulls of suboptimal arm subsets is less than or equal to $\sum_{i=1}^m N_{i,T}$.
Define $\psi_t = \frac{8B_{max}f^{-1}(\Delta_{min}) + 6\log{t}} {\left(f^{-1}(\Delta_{min})\right)^2} > c $, where $c$ is the larger solution of
\begin{align*}
\label{square}
&(f^{-1}(\Delta_{min}))^2 c^2 + 16B_{max}^2 - ((8B_{max} f^{-1}(\Delta_{min})- 6\log{t}) c = 0 \,.
\end{align*}
Then, we decompose the total number $\sum_{i=1}^m N_{i,T}$ of pulls of suboptimal arm subsets as
\begin{align}
\sum_{i=1}^m N_{i,T} = \ & m +\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B\} \notag \\
=\ &m +\sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} \leq \psi_t\} \notag \\
&+ \sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} > \psi_t\}\notag \\
\leq \ & m + m\psi_T+\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}\,. \notag
\end{align}
The inequality follows as $\sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} \leq \psi_t\}$ can be trivially bounded by $m\psi_T$. Thus the key to bound the total number of pulls of suboptimal arm subset is to upper bound $\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}$. Let $F_t$ denotes the event that the oracle fails to provide an $\alpha$-approximate arm subset with respect to the input vector $\Bar{\bm{\mu}} = (\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$.
Then we can decompose $\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}$ as
\begin{align}
&\quad \sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}\notag\\
&\leq \sum_{t=m+1}^T (\mathbb{I}\{F_t\} + \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\})\notag \\
&\leq (T-m)(1-\beta) + \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\}\notag \,. \notag
\end{align}
By leveraging the monotonicity and smoothness assumptions of the reward function, we show that $\mathbb{P}(\{\lnot F_t, E_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\}) = 0$. Therefore, by the inclusion-exclusion principle,
\begin{align}
\mathbb{P}(\{\lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\})
&\leq \mathbb{P}(\lnot E_t) \leq 2m \cdot t^{-2} \,.\notag
\end{align}
Leveraging the upper bound of total number of pulls of suboptimal arm subsets, the regret is bounded by
\begin{align}
& \quad Regret_{\bm{\mu},\alpha,\beta}(T)
\notag \\
& \leq T\alpha \beta \cdot \text{OPT}_{\bm{\mu}} - \left( T\alpha\cdot \text{OPT}_{\bm{\mu}} - \mathbb{E} \left[ \sum_{i=1}^m N_{i,T} \right]\cdot \Delta_{max} \right)
\notag \\
&\leq m\cdot \Delta_{max}\left(\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}+1\right)\,. \notag \qedhere
\end{align}
\end{proof}
It remains in question whether a bandit algorithm can achieve a regret upper bound sublinear in $B_{max}$.
Our conjecture is negative.
In fact, under strategic manipulations of rewards, the design of robust bandit algorithms, e.g. UCB and $\epsilon$-greedy, is analogous to the design of outlier-robust mean estimation algorithms. Existing works on robust mean estimation, such as \citet{steinhardt_et_al:LIPIcs:2018:8368}, argue that from an information theoretical point of view mean estimation error must be depending on the variance of the data. Casting this argument to bandit with the strategic manipulation setting, we believe that a tight regret bound is unlikely to be independent of the strategic budget $B_{max}$. Moreover, Proposition 4 in \citet{steinhardt_et_al:LIPIcs:2018:8368} implies that the dependency on $B_{max}$ is linear. This corresponds to our linear dependency of regret in Theorem \ref{appendixA}.
\section{Lower Bounds for strategic budget}
To fully understand the effects of strategic manipulations, we investigate the relationship between the strategic budget and the performance of UCB-based algorithms. Our results provide the dependency between an arm's strategic budget and the number of times it is pulled, which influence the regret of the algorithm incurred by the arm. Our analysis gains some insight from \citet{zuo_near_2020}, which limits the discussion to the 2-armed bandit setting and cannot be applied to the general MAB setting directly.
We first define some notations used in the theorem.
Without loss of generality, let arm $i^{\ast}$ be the optimal arm and arm $i, i \neq i^{\ast}$ be an arbitrary strategic suboptimal arm. Assume that arm $i$ has no access to the information regarding other arms. Let $K_{i,t}$ and $K_{i^{\ast},t}$ denote the number of times arm $i$ and arm $i^{\ast}$ have been pulled up to time $t$, respectively. Denote $\hat{\mu}_{i,t}$ as the empirical estimate of the underlying mean $\mu$ without manipulations, i.e., $\hat{\mu}_{i,t} = \frac{\sum^{t-1}_{j=1}x_{i,j}}{K_{i,t-1}}$. Define the suboptimality gap for each strategic arm $i$ to be $\delta_i = \mu_{i^{\ast}} - \mu_i$. We use a slightly revised UCB algorithm as the basic algorithm, where the UCB estimation term for arm $i$ is $\hat{\mu}_{i,t} + \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}}$ and $\eta$ is a confidence parameter chosen by the algorithm.
\begin{theorem}\label{lower}
In stochastic multi-armed bandit problems, for a strategic suboptimal arm $i$ without access to other arms' information, to be pulled for $\omega(k)$ in $T$ steps under the UCB algorithm where $k\ge O(\log{T})$, the minimum strategic budget is $\omega(k)$.
\end{theorem}
This dependency of $k$ can be extended to CMAB and CUCB straightforwardly when arms within an arm subset collude. Counter-intuitively, the dependency between the number of pulls of a strategic arm and its strategic budget is linear and subsequently, this infers that for a strategic arm to manipulate the algorithm to suffer an undesired regret of $\omega(k)$, where $k\ge O(\log T)$, the strategic budget must be at least $\omega(k)$.
\section{Experiments}
In this section, we evaluate our SCUCB algorithm empirically on synthetic data and three real applications, namely online worker selection in crowdsourcing, online recommendation, and online influence maximization. We highlight the best performance among all algorithms with bold text.
\subsection{Baseline Algorithms}
We compare our proposed SCUCB algorithm with both stochastic and adversarial bandits algorithms that achieves optimal asymptotic regret in their settings.
\begin{enumerate}
\allowdisplaybreaks
\item CUCB \cite{pmlr-v28-chen13a}. CUCB is the naive counterpart of our algorithm. The algorithm calculates an upper confidence interval for each arm and picks the best arm subset with the highest upper confidence interval.
\item TSCB \cite{wang2018thompson}. TSCB is the combinatorial version of the classical Thompson sampling algorithm. The algorithm maintains a prior beta distribution estimation for each arm and updates according to the received reward. At each time, the algorithm samples from the estimated distributions and pick actions according to the highest sample.
\item Combinatorial variant of EXP3 \cite{auer_nonstochastic_2002}. The algorithm maintains a weight for each arm and draws actions according to the normalized weight distribution. Upon receiving rewards, the algorithm update weight according to the classical EXP3 update rule.
\end{enumerate}
\subsection{Synthetic Experiments}
We conduct experiments presented in this section with synthetic data and compare our proposed algorithm with its naive counterpart. The approximation oracle is designed to succeed with probability $1$. Each bandit arm is modeled to follow a Bernoulli distribution with randomly populated $\mu \in [0,1]$ and all arms adapt LSI strategy.
The arms in the optimal arm subset have a strategic budget of $0$ since an additional strategic budget for optimal arms would only boost the performance of our algorithm. All other arms are equipped with a randomly allocated budget $B$, $ 0 \leq B \leq B_{max}$ by definition.
As is typical in the bandit literature, for example in \citet{auer2002finite}, we evaluate both the naive and strategic CUCB algorithms on their tuned versions, where the UCB exploration parameter is scaled by a constant factor $\gamma \in [0,1]$. To ensure reproducible results, each experiment is repeated for $10$ random seeds and the averaged result is presented.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 171.74 & 187.51 & 259.04 & 256.66 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{143.85} & \textbf{172.57} & \textbf{208.53} & \textbf{233.57} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 434.82 & 492.02 & 520.97 & 549.57 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{301.57} & \textbf{365.94} & \textbf{450.10} & \textbf{505.453} \\ \bottomrule
\end{tabularx}
\caption{Cumulative regret achieved by CUCB and SCUCB with synthetic data with various action size.}\label{table:1}
\end{table}
The first set of experiment is conducted with $K = \{10, 20\}$ arms through a time horizon of $T=5000$ time steps with maximum possible budget of $B_{max} = \{70, 90, 110, 130\}$. The algorithms are asked to select $\text{Action size} = 2$ arms as an arm subset at each time step. The cumulative regret incurred by CUCB and SCUCB algorithm is presented in the table \ref{table:1} where the best performance is highlighted in bold text. Clearly, SCUCB demonstrated its effectiveness as it achieves significantly smaller regrets in various possible maximum strategic budgets.
The next two tables reveals the advantage of SCUCB algorithm with various size of action set. The experiments are conducted with time horizon of $T=5000$ time steps with $K = \{10, 20\}$ arms and maximum possible budget of $B_{max} = \{50\}$. The algorithms are asked to select $\text{Action size} = \{2, 4, 6, 8\}$ arms as an arm subset at each time step. Once again, our SCUCB outperforms its naive counterpart by achieving much smaller cumulative regret,
the best numerical results across algorithms are highlighted in bold text.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 140.21 & 150.88 & 158.22 & 123.91 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{105.58} & \textbf{113.63} & \textbf{103.42} & \textbf{88.72} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 302.10 & 316.09 & 340.52 & 328.80 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{232.23} & \textbf{247.32} & \textbf{268.27} & \textbf{306.77} \\ \bottomrule
\end{tabularx}
\caption{Cumulative regret achieved by UCB and SCUCB with synthetic data with various $B_{max}$.}
\end{table}
To compare the two algorithms in detail, we plot the cumulative regret incurred by algorithms against the time steps. It becomes apparent that the SCUCB algorithm features a cumulative regret versus time steps line that is much smaller than the one caused by CUCB even under a higher maximum possible strategic budget. We compare the two algorithm under settings where the maximize possible budget is $B_{max} = 70, 90$ and $B_{max} = 110, 130$.
\begin{figure}[H]
\centering
\subfigure[]{
\label{Fig1}
\includegraphics[width=3.8cm]{figs/B7090_K20_N10000.png}}
\subfigure[]{
\label{Fig2}
\includegraphics[width=3.8cm]{figs/B110130_K20_N10000.png}}
\caption{\ref{Fig1} Comparison of CUCB and SCUCB with $B_{max} = 70, 90$. \ref{Fig2} Comparison of CUCB and SCUCB with $B_{max} = 110, 130$.}
\end{figure}
\subsection{Online Worker Selection in Crowdsourcing Systems}
We simulated an online crowdsourcing system that has a workflow resembled by figure \ref{crowdsys} and model it as a combinatorial bandits problem. We performed an extensive empirical analysis of our SCUCB algorithm against CUCB, the combinatorial version of Thompson sampling, and EXP3. The experiments were conducted with the Amazon sentiment dataset, where Amazon item reviews are labeled 'is/is not book' or 'is/is not negative \cite{Krivosheev:2018:CCM:3290265.3274366}. We split the dataset into two, where 'is book' contains data that only has labels 'is/is not book' and 'is negative' contains data that only has labels 'is/is not negative'. Both datasets consists of 7803 reviews, 284 workers, and 1011 tasks.
The correctness of workers and tasks can be visualized by the figures as below. Notice that most of the workers attain an accuracy of $60$ percent and above. This highlights that most of the workers we are interacting with has the ability to label most of the tasks correctly.
\begin{figure}[H]
\centering
\subfigure[]{
\label{worker}
\includegraphics[width=3.8cm]{figs/worker_visual.png}}
\subfigure[]{
\label{task}
\includegraphics[scale = 0.28]{figs/task_visual.png}}
\caption{\ref{worker} Visualization of correctness of worker's labels. \ref{task}Visualization of correctness of task's labels.}
\end{figure}
For each task, the data contains responses from 5 workers and the crowdsourcing platform is asked to select 2, 3, or 4 workers at each time step. To measure the performance of the algorithms, we made the assumption that the reward is monotonically correlated to the accuracy of the label compared to the ground truth label. Thus, we define the reward to be 1, if the worker's label is the same as ground truth label and 0 otherwise. To model the strategic behavior of the workers, we model each worker with a randomly allocated strategic level $s \in [0,1]$ such that the worker provides its honest response with probability $s$.
For the first set of experiments, we choose the maximum possible strategic budget to be $B_{max} = 500$. For actions size of $2, 3, 4$. We obtain the following cumulative rewards and cumulative regret results where the best performance is highlighted in bold text. To ensure reproducible results, each experiment is repeated 5 times and the averaged result is presented.
To further investigate the effect of the maximum possible strategic budget on the performance of algorithms, we fix the size of action size to be 2 and plot the cumulative reward/cumulative regret incurred by each algorithm again $B_{max} \in [100, 700]$. Regardless of the $B_{max}$ value, our proposed SCUCB algorithm consistently outperforms other methods. For better visualization, we omit the line for combinatorial Thompson sampling due to its relatively weaker performance.
\begin{figure}[H]
\centering
\subfigure[]{
\label{culre}
\includegraphics[width=3.8cm]{figs/culre.png}}
\subfigure[]{
\label{culreg}
\includegraphics[width=3.8cm]{figs/reg.png}}
\caption{\ref{culre} Cumulative rewards attained by algorithms with various level of $B_{max}$. \ref{culreg} Cumulative regrets attained by algorithms with various level of $B_{max}$. }
\end{figure}
\begin{table}[H]
\label{first}
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31672.6 & 46181.2 & 68244.6 \\
\multicolumn{1}{l|}{TSCB} & 19853.0 & 34226.2 & 60621.2 \\
\multicolumn{1}{l|}{EXP3} & 31569.6 & 46570.6 & 68482.2 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{32082.6} & \textbf{46623.6} & \textbf{68524.6} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 27787.4 & 42198.8 & 48125.4 \\
\multicolumn{1}{l|}{TSCB} & 39607.0 & 54153.8 & 55748.8 \\
\multicolumn{1}{l|}{EXP3} & 27890.4 & 41809.4 & 47887.8 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{27377.4} & \textbf{41756.4} & \textbf{47845.4} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31108.8 & 44757.6 & 59424.8 \\
\multicolumn{1}{l|}{TSCB} & 17240.6 & 32010.4 & 51109.0 \\
\multicolumn{1}{l|}{EXP3} & 30879.0 & 44767.2 & 58207.8 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{31759.0} & \textbf{44843.4} & \textbf{59441.4} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 28411.2 & 43566.2 & 56465.2 \\
\multicolumn{1}{l|}{TSCB} & 42279.4 & 56279.6 & 64781.0 \\
\multicolumn{1}{l|}{EXP3} & 28641.0 & 43523.4 & 57682.2 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{27761.0} & \textbf{43446.6} & \textbf{56448.6} \\ \bottomrule
\end{tabularx}
\caption{Cumulative rewards and cumulative regret attained by SCUCB and baseline algorithms with various action sizes on isbook dataset and isnegative dataset.}
\end{table}
The results of our crowdsourcing experiments reveal the robustness of our SCUCB algorithm under strategic manipulations. It also indicates the performance of adversarial and stochastic bandits algorithms under strategic manipulations. The combinatorial variant of Thompson sampling is vulnerable under manipulations, which agrees with our expectation given that the algorithm is a Bayesian algorithm in nature and is heavily relying on estimating the underlying distribution of arms. Its estimation can be easily perturbed under strategic manipulation, and thus leads to undesired performance results. It is also expected that the SCUCB algorithm far outperforms its naive counterpart, CUCB algorithm. The combinatorial version of EXP3 algorithm is the most competitive algorithm with our SCUCB. As the EXP3 algorithm was originally designed for a pure adversarial setting, the algorithm has some robustness under strategic manipulations. Shown in the Figure \ref{culre} and Figure \ref{culreg}, the combinatorial EXP3 algorithm consistently shows robustness across various levels of maximum possible strategic budget. However, as our SCUCB algorithm considers the maximum possible strategic budget and thus more adaptive, SCUCB far outperforms EXP3CB as $B_{max}$ increases.
\subsection{Online Recommendation System}
The online recommendation is a classic example of combinatorial bandits in applications. We evaluated algorithms in the latest MovieLens dataset which contains 9742 movies and 100837 ratings of the movies with each rating being $1 - 5$ \cite{10.1145/2827872}. We compare the algorithms based on the total values of ratings received, i.e. recommending a movie and receiving a rating of 5 is more desired than recommending a movie and receiving a rating of 1.
As the dataset may consist of unbalanced data for each movie, we adapted collaborative filtering and $k$-means clustering into our evaluation system. We used the collaborative filtered for training and testing and instead of recommending movies, we clustered the movies by $k$-means clustering and asked the bandits algorithm to choose one cluster at each time. The movie with the highest rating in the chosen cluster is recommended. The workflow of the experiment setup is summarized in the above diagram.
\begin{figure}[H]
\centering
\includegraphics[width=7cm]{figs/arch.png}
\caption{Experiment set up of online recommendation.}
\end{figure}
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 10 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2297.60 & 2494.36 & 2631.02 \\
\multicolumn{1}{l|}{TS} & 2160.14 & 2314.76 & 2468.26 \\
\multicolumn{1}{l|}{EXP3} & 2181.11 & 2449.28 & 2430.96 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2305.75} & \textbf{2314.76} & \textbf{2636.07} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 20 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2469.93 & 2336.92 & 2159.25 \\
\multicolumn{1}{l|}{TS} & 2260.20 & 1850.55 & 1905.42 \\
\multicolumn{1}{l|}{EXP3} & 2380.74 & 2317.82 & 2025.39 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2474.69} & \textbf{2341.19} & \textbf{2177.70} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 30 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\\midrule
\multicolumn{1}{l|}{UCB} & 2443.57 & 2393.88 & 2472.31 \\
\multicolumn{1}{l|}{TS} & 2132.16 & 2248.56 & 1855.14 \\
\multicolumn{1}{l|}{EXP3} & 2368.42 & 2265.32 & 2339.49 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2436.43} & \textbf{2397.72} & \textbf{2476.52} \\ \bottomrule
\end{tabularx}
\caption{Cumulative rewards attained by algorithms with different number of clusters $10, 20, 30$ and different level of $\{B_{max} = 30, 50, 70\}$ across 500 time steps. }
\end{table}
We performed the evaluation with various number of clusters (10, 20 or 30) and $B_{max} = 30, 50, 70$ through a time horizon of $T = 500$. The results presented below are averages over 5 runs to ensure reproducibility. The best performance across algorithms is highlighted in bold text. From the tables below, we conclude the effectiveness of SCUCB algorithm.
\subsection{Online Influence Maximization}
We implemented the online influence maximization with an offline influence maximization algorithm TIM as our oracle \cite{10.1145/2588555.2593670}. TIM is one of the offline influence maximization algorithms that achieve asymptotic optimality. We perform the experiments on two datasets. One is a simulation dataset with 16 nodes and 44 edges, the other is a 100 nodes subset from Digg dataset, where each node represents a user from Digg website \cite{nr}. We simulate the connectivity of edges at each time step by randomly assigning each edge a connectivity probability at the start of the experiment.
\begin{figure}[hbt!]
\centering
\subfigure[]{
\label{test_graph}
\includegraphics[width=3.6cm]{figs/test_graph.png}}
\subfigure[]{
\label{16c2}
\includegraphics[width=3.6cm]{figs/test.png}}
\caption{\ref{test_graph} Visualization of synthetic graph for evaluation. \ref{16c2} Mean rewards attained by algorithms.}
\end{figure}
Figure \ref{test_graph} visualizes the synthetic graph we created for evaluation. We then tested the algorithms over $T = 800$ time steps with action size of 2 and maximum possible strategic budget of $B_{max} = 300$. Figure \ref{16c2} shows the effectiveness of algorithms by evaluating based on the averaged nodes influenced. To ensure reproducible results, we calculate the mean of influence spread over 100 trials. For better visualization, we kept a running mean of over 50 time steps to smooth the lines.
We then investigate the effect of action size and the results are summarized in the following tables \ref{table:5}.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYYY@{}}
\toprule
\multicolumn{5}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Synthetic, 16 nodes, 44 edges} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 4 & 6 & 8 \\ \midrule
\multicolumn{1}{c|}{CUCB} & 5.46 & 6.54 & 10.4 & 12.5 \\
\multicolumn{1}{c|}{TSCB} & 6.7 & 8.50 & 10.44 & 12.7 \\
\multicolumn{1}{c|}{EXP3CB} & 5.86 & 8.14 & 10.34 & 12.46 \\
\multicolumn{1}{c|}{SCUCB} & \textbf{7.44} & \textbf{9.32} & \textbf{11.16} & \textbf{13.04} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{digg, 100 nodes} \\ \midrule
\multicolumn{1}{l|}{Action size} & 10 & 15 & 20 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 15.64 & 20.64 & 25.32 \\
\multicolumn{1}{l|}{TSCB} & 16.30 & 21.62 & 26.08 \\
\multicolumn{1}{l|}{EXP3} & 12.16 & 17.76 & 21.72 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{16.56} & \textbf{22.16} & \textbf{26.14} \\ \bottomrule
\end{tabularx}
\caption{Averaged final rewards attained by algorithms with synthetic and Digg dataset with action size of 2, 800 time steps and $B_{max} = 200$.}\label{table:5}
\end{table}
\section{Conclusion}
We investigate the problem of combinatorial MAB under strategic manipulations of rewards. We propose a variant of the UCB algorithm, SCUCB, which attains a regret at most $O(m \log T + m B_{max})$, with a lower bounds on the strategic budget for a malicious arm to incur a $\omega(\log T)$ regret of the bandit algorithm.
Compared to previous studies on the bandit problems under strategic manipulations, we relax on the assumption that the algorithm has access to the cumulative strategic budget spent by each arm at each time step.
For the robustness of bandit algorithms, we present lower bounds on the strategic budget for a malicious arm to incur a $\omega(\log T)$ regret of the bandit algorithm.
We provide extensive empirical results on both synthetic and real datasets with a range of applications to verify the effectiveness of the proposed algorithm. Our algorithm consistently outperforms baseline algorithms that were designed for stochastic and adversarial settings.
\subsection{Baseline Algorithms}
We compare our proposed SCUCB algorithm with both stochastic and adversarial bandits algorithms that achieves optimal asymptotic regret in their settings.
\begin{enumerate}
\item CUCB \cite{pmlr-v28-chen13a}. CUCB is the naive counterpart of our algorithm. The algorithm calculates a upper confidence interval for each arm and pick the best arm subset with the highest upper confidence interval.
\item TSCB \cite{wang2018thompson}. TSCB is the combinatorial version of the classical Thompson sampling algorithm. The algorithm maintain a prior beta distribution estimation for each arm and updates according to the received reward. At each time, the algorithm samples from the estimated distributions and pick actions according to the highest sample.
\item Combinatorial variant of EXP3 \cite{auer_nonstochastic_2002}. The algorithm maintains a weight for each arm and draw actions according to the normalized weight distribution. Upon receiving rewards, the algorithm update weight according to the classical EXP3 update rule.
\end{enumerate}
\subsection{Synthetic Experiments}
We conduct experiments presented in this section with synthetic data and compare our proposed algorithm with its naive counter part. The approximation oracle is designed to succeed with probability $1$. Each bandit arm is modeled to follow a Bernoulli distribution with randomly populated $\mu \in [0,1]$ and all arms adapt LSI strategy.
The arms in the optimal arm subset have a strategic budget of $0$ since additional strategic budget for optimal arms would only boost the performance of our algorithm. All other arms are equipped with a randomly allocated budget $B$,$ 0 \leq B \leq B_{max}$ by definition.
As is typical in the bandit literature, for example in \citet{auer2002finite}, we evaluate both the naive and strategic CUCB algorithms on their tuned versions, where the UCB exploration parameter is scaled by a constant factor $\gamma \in [0,1]$. To ensure reproducible result, each experiment is repeated for $10$ random seeds and the averaged result is presented.
\begin{table}[H]
\centering
\label{table:1}
\begin{tabular}{@{}lcccc@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 171.74 & 187.51 & 259.04 & 256.66 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{143.85} & \textbf{172.57} & \textbf{208.53} & \textbf{233.57} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lcccc@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 434.82 & 492.02 & 520.97 & 549.57 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{301.57} & \textbf{365.94} & \textbf{450.10} & \textbf{505.453} \\ \bottomrule
\end{tabular}
\caption{Cumulative regret achieved by UCB and SCUCB with synthetic data with various action size.}
\end{table}
The first set of experiment is conducted with $K = \{10, 20\}$ arms through a time horizon of $T=5000$ time steps with maximum possible budget of $B_{max} = \{70, 90, 110, 130\}$. The algorithms are asked to select $\text{Action size} = 2$ arms as an arm subset at each time step. The cumulative regret incurred by CUCB and SCUCB algorithm are presented in the table \ref{table:1} where the performance of our proposed SCUCB algorithm is highlighted in bold text. Clearly, SCUCB demonstrated its effectiveness as it achieves significantly smaller regrets in various possible maximum strategic budget.
The next two tables reveals the advantage of SCUCB algorithm with various size of action set. The experiments are conducted with time horizon of $T=5000$ time steps with $K = \{10, 20\}$ arms and maximum possible budget of $B_{max} = \{50\}$. The algorithms are asked to select $\text{Action size} = \{2, 4, 6, 8\}$ arms as an arm subset at each time step. Once again, our SCUCB outperforms its naive counterpart by achieving much smaller cumulative regret, for which numerical results are highlighted in bold text.
\begin{table}[H]
\centering
\begin{tabular}{@{}lcccc@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 140.21 & 150.88 & 158.22 & 123.91 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{105.58} & \textbf{113.63} & \textbf{103.42} & \textbf{88.72} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lcccc@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 302.10 & 316.09 & 340.52 & 328.80 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{232.23} & \textbf{247.32} & \textbf{268.27} & \textbf{306.77} \\ \bottomrule
\end{tabular}
\caption{Cumulative regret achieved by UCB and SCUCB with synthetic data with various $B_{max}$.}
\end{table}
To compare the two algorithms in details, we plot the cumulative regret incurred by algorithms against the time steps. It becomes apparent that the SCUCB algorithm features a cumulative regret versus time steps line that is much smaller than the one caused by CUCB even under a higher maximum possible strategic budget. We compare the two algorithm under settings where the maximize possible budget is $B_{max} = 70, 90$ and $B_{max} = 110, 130$.
\begin{figure}[H]
\centering
\subfigure[]{
\label{Fig1}
\includegraphics[width=3.8cm]{figs/B7090_K20_N10000.png}}
\subfigure[]{
\label{Fig2}
\includegraphics[width=3.8cm]{figs/B110130_K20_N10000.png}}
\caption{\ref{Fig1} Comparison of CUCB and SCUCB with $B_{max} = 70, 90$. \ref{Fig2} Comparison of CUCB and SCUCB with $B_{max} = 110, 130$.}
\end{figure}
\subsection{Online Worker Selection in Crowdsourcing Systems}
We simulated a online crowdsourcing system that has a workflow resembled by \ref{crowdsys} and model it as a combinatorial bandits problem. We performed extensive empirical analysis of our SCUCB algorithm against CUCB, the combinatorial version of Thompson sampling and EXP3. The experiments were conducted with the Amazon sentiment dataset, where Amazon items reviews are labeled 'is/is not book' or 'is/is not negative \cite{Krivosheev:2018:CCM:3290265.3274366}. We split the dataset into two, where 'is book' contains data that only has labels 'is/is not book' and 'is negative' contains data that only has labels 'is/is not negative'. Both dataset consists of 7803 reviews, 284 workers, and 1011 tasks.
The correctness of workers and tasks can be visualized by the figures as below.
\begin{figure}[H]
\centering
\subfigure[]{
\label{worker}
\includegraphics[width=3.8cm]{figs/worker_visual.png}}
\subfigure[]{
\label{task}
\includegraphics[scale = 0.28]{figs/task_visual.png}}
\caption{\ref{worker} Visualization of correctness of worker's labels. \ref{task}Visualization of correctness of task's labels.}
\end{figure}
For each task, the data contains responses from 5 workers and the crowdsourcing platform is asked to select 2, 3, or 4 workers at each time step. To measure the performance of the algorithms, we made the assumption that the reward is monotonically correlated to the accuracy of label compared to ground truth label. Thus, we define the reward to be 1, if the worker's label is the same with ground truth label and 0 otherwise. To model the strategic behavior of the workers, we model each worker with a randomly allocated strategic level $s \in [0,1]$ such that the worker provides its honest response with probability $s$.
\begin{table}[H]
\label{first}
\centering
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31672.6 & 46181.2 & 68244.6 \\
\multicolumn{1}{l|}{TSCB} & 19853.0 & 34226.2 & 60621.2 \\
\multicolumn{1}{l|}{EXP3} & 31569.6 & 46570.6 & 68482.2 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{32082.6} & \textbf{46623.6} & \textbf{68524.6} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 27787.4 & 42198.8 & 48125.4 \\
\multicolumn{1}{l|}{TSCB} & 39607.0 & 54153.8 & 55748.8 \\
\multicolumn{1}{l|}{EXP3} & 27890.4 & 41809.4 & 47887.8 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{27377.4} & \textbf{41756.4} & \textbf{47845.4} \\ \bottomrule
\end{tabular}
\end{table}
\begin{table}[H]\centering
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31108.8 & 44757.6 & 59424.8 \\
\multicolumn{1}{l|}{TSCB} & 17240.6 & 32010.4 & 51109.0 \\
\multicolumn{1}{l|}{EXP3} & 30879.0 & 44767.2 & 58207.8 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{31759.0} & \textbf{44843.4} & \textbf{59441.4} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 28411.2 & 43566.2 & 56465.2 \\
\multicolumn{1}{l|}{TSCB} & 42279.4 & 56279.6 & 64781.0 \\
\multicolumn{1}{l|}{EXP3} & 28641.0 & 43523.4 & 57682.2 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{27761.0} & \textbf{43446.6} & \textbf{56448.6} \\ \bottomrule
\end{tabular}
\caption{Cumulative rewards and cumulative regret attained by SCUCB and baseline algorithms with various action size on isbook dataset and isnegative datase.}
\end{table}
For the first set of experiments, we choose the maximum possible strategic budget to be $B_{max} = 500$. For actions size of $2, 3, 4$. We obtain the following cumulative rewards and cumulative regret results where SCUCB's performance is highlight in bold text. To ensure reproducible results, each experiment is repeated for 5 times and the averaged result is presented.
To further investigate the effect of maximum possible strategic budget on the performance of algorithms, we fix the size of action size to be 2 and plot the cumulative reward/cumulative regret incurred by each algorithm again $B_{max} \in [100, 700]$. Regardless of the $B_{max}$ value, our proposed SCUCB algorithm consistently outperforms other methods. For better visualization, we omit the line for combinatorial Thompson sampling due to its relatively weaker performance.
\begin{figure}[H]
\centering
\subfigure[]{
\label{culre}
\includegraphics[width=3.8cm]{figs/culre.png}}
\subfigure[]{
\label{culreg}
\includegraphics[width=3.8cm]{figs/reg.png}}
\caption{\ref{culre} Cumulative rewards attained by algorithms with various level of $B_{max}$. \ref{culreg} Cumulative regrets attained by algorithms with various level of $B_{max}$. }
\end{figure}
The results of our crowdsourcing experiments reveals the robustness of our SCUCB algorithm under strategic manipulations. It also indicates the performance of adversarial and stochastic bandits algorithms under strategic manipulations. The combinatorial variant of Thompson sampling is vulnerable under manipulations, which agrees with our expectation given that the algorithm is a Bayesian algorithm in nature and is heavily relying on estimating the underlying distribution of arms. Its estimation can be easily perturbed under strategic manipulation, and thus leads to undesired performance results. It is also expected that the SCUCB algorithm far outperform its naive counterpart, CUCB algorithm. The combinatorial version of EXP3 algorithm is the most competitive algorithm with our SCUCB. As the EXP3 algorithm was originally designed for pure adversarial setting, the algorithm has some robustness under strategic manipulations. Shown in the Figure \ref{culrereg}, the combinatorial EXP3 algorithm consistently show robustness across various level of maximum possible strategic budget. However, as our SCUCB algorithm considers the maximum possible strategic budget and thus more adaptive, SCUCB far outperforms EXP3CB as $B_{max}$ increases.
\subsection{Online Recommendation System}
Online recommendation is a classic example of combinatorial bandits in applications. We evaluated algorithms in the latest MovieLens dataset which contains 9742 movies and 100837 rating of the movies with each rating being $1 - 5$ \cite{10.1145/2827872}. We compare the algorithms based on the total values of ratings received, i.e. recommending a movie and receiving a rating of 5 is more desired than recommending a movie and receiving a rating of 1.
\begin{figure}[H]
\centering
\includegraphics[width=7cm]{figs/arch.png}
\caption{Experiment set up of online recommendation.}
\end{figure}
As the dataset may consists of unbalanced data for each movie, we adapted collaborative filtering and $k$-means clustering into our evaluation system. We used the collaborative filtered for training and testing and instead of recommending movies, we clustered the movies by $k$-means clustering and asked bandits algorithm to choose one cluster at each time. The movie with the highest rating in the chosen cluster is recommended. The workflow of the experiment set up is summarized in above diagram.
We performed the evaluation with various number of clusters (10, 20 or 30) and $B_{max} = 30, 50, 70$ through a time horizon of $T = 500$. The results presented below are averages over 5 runs to ensure reproducibility. From the tables below, we concludes the effectiveness of SCUCB algorithm, which performance is highlighted in bold text.
\begin{table}[H]
\centering
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 10 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2297.60 & 2494.36 & 2631.02 \\
\multicolumn{1}{l|}{TS} & 2160.14 & 2314.76 & 2468.26 \\
\multicolumn{1}{l|}{EXP3} & 2181.11 & 2449.28 & 2430.96 \\
\multicolumn{1}{l|}{\textbf{SUCB}} & \textbf{2305.75} & \textbf{2314.76} & \textbf{2636.07} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 20 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2469.93 & 2336.92 & 2159.25 \\
\multicolumn{1}{l|}{TS} & 2260.20 & 1850.55 & 1905.42 \\
\multicolumn{1}{l|}{EXP3} & 2380.74 & 2317.82 & 2025.39 \\
\multicolumn{1}{l|}{\textbf{SUCB}} & \textbf{2474.69} & \textbf{2341.19} & \textbf{2177.70} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 30 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\\midrule
\multicolumn{1}{l|}{UCB} & 2443.57 & 2393.88 & 2472.31 \\
\multicolumn{1}{l|}{TS} & 2132.16 & 2248.56 & 1855.14 \\
\multicolumn{1}{l|}{EXP3} & 2368.42 & 2265.32 & 2339.49 \\
\multicolumn{1}{l|}{\textbf{SUCB}} & \textbf{2436.43} & \textbf{2397.72} & \textbf{2476.52} \\ \bottomrule
\end{tabular}
\caption{Cumulative rewards attained by algorithms with different number of clusters $10, 20, 30$ and different level of $\{B_{max} = 30, 50, 70\}$ across 500 time steps. }
\end{table}
\subsection{Online Influence Maximization}
We implemented the online influence maximization with an offline influence maximization algorithm TIM as our oracle \cite{10.1145/2588555.2593670}. TIM is one of the offline influence maximization algorithm that achieves asymptotic optimality. We perform the experiments on two datasets. One is a simulation dataset with 16 nodes and 44 edges, the other is a 100 nodes subset from Digg dataset, where each nodes represents a user from Digg website \cite{nr}. We simulate the connectivity of edges at each time step by randomly assigning each edge a connectivity probability at the start of the experiment.
\begin{figure}[hbt!]
\centering
\subfigure[]{
\label{test_graph}
\includegraphics[width=3.6cm]{figs/test_graph.png}}
\subfigure[]{
\label{16c2}
\includegraphics[width=3.6cm]{figs/test.png}}
\caption{\ref{test_graph} Visualization of synthetic graph for evaluation. \ref{16c2} Mean rewards attained by algorithms.}
\end{figure}
Figure \ref{test_graph} visualizes the synthetic graph we created for evaluation. We then tested the algorithms over $T = 800$ time steps with action size of 2 and maximum possible strategic budget of $B_{max} = 300$. Figure \ref{16c2} show the effectiveness of algorithms by evaluating based on the averaged nodes influenced. To ensure reproducible results, we calculate the mean of nodes influence over 100 trials. For better visualization, we kept a running mean over 50 time steps to smooth the lines.
We then investigate the effect of action size and the results are summarized in the following two tables.
\begin{table}[hbt!]
\centering
\begin{tabular}{@{}ccccc@{}}
\toprule
\multicolumn{5}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Synthetic, 16 nodes, 44 edges} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 4 & 6 & 8 \\ \midrule
\multicolumn{1}{c|}{CUCB} & 5.46 & 6.54 & 10.4 & 12.5 \\
\multicolumn{1}{c|}{TSCB} & 6.7 & 8.50 & 10.44 & 12.7 \\
\multicolumn{1}{c|}{EXP3CB} & 5.86 & 8.14 & 10.34 & 12.46 \\
\multicolumn{1}{c|}{\textbf{SCUCB}} & \textbf{7.44} & \textbf{9.32} & \textbf{11.16} & \textbf{13.04} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{digg, 100 nodes} \\ \midrule
\multicolumn{1}{l|}{Action size} & 10 & 15 & 20 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 15.64 & 20.64 & 25.32 \\
\multicolumn{1}{l|}{TSCB} & 16.30 & 21.62 & 26.08 \\
\multicolumn{1}{l|}{EXP3} & 12.16 & 17.76 & 21.72 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{16.56} & \textbf{22.16} & \textbf{26.14} \\ \bottomrule
\end{tabular}
\caption{Averaged final rewards attained by algorithms with synthetic and Digg dataset with action size of 2, 800 time steps and $B_{max} = 200$.}
\end{table}
\iffalse
\paragraph{Experiment setting}
We conduct the experiments with 10 arms where each arm subset constitutes $2$ arms unless otherwise specified. The time horizon is set to be $T = 5\times 10^3$ and the approximation oracle is designed to succeed with probability $1$. Each bandit arm follows a Bernoulli distribution with randomly populated $\mu \in [0,1]$ and all arms adapt LSI except in the discussion of arm strategies.
The arms in the optimal arm subset have a strategic budget of $0$ and all other arms are equipped with the same strategic budget, which is $B_{max}$ by definition. The setting of heterogeneous budgets is discussed in the appendix.
As is typical in the bandit literature, for example in \citet{auer2002finite}, we evaluate both naive and strategic CUCB algorithms on their tuned versions, where the UCB exploration parameter is scaled by a constant factor $\gamma \in [0,1]$. Each experiment is repeated $10$ times and the averaged result is presented.
\begin{figure*}[htbp]
\centering
\subfigure[]{
\label{Fig1}
\includegraphics[scale=0.3]{figs/diff_B.png}}
\subfigure[]{
\label{Fig2}
\includegraphics[scale=0.24]{figs/linearB.png}}
\subfigure[]{
\label{Fig3}
\includegraphics[scale=0.3]{figs/LSI2.png}}
\subfigure[]{
\label{Fig4}
\includegraphics[scale=0.3]{figs/50arm.png}}
\caption{\ref{Fig1} Cumulative regret v.s. time horizon under different maximum budgets; \ref{Fig2} Cumulative regret v.s. strategic budget; \ref{Fig3} Cumulative regret under the LSI strategy and random strategy; \ref{Fig4} Cumulative regret v.s. time horizon under different number of arms in each arm subset ($B_{max} = 50$).}
\end{figure*}
\paragraph{Cumulative regret v.s. maximum budgets} We experiment with $B_{max} \in \{0, 10, 50, 100\}$ and $\gamma = 0.2$. Figure \ref{Fig1} highlights the failure of naive CUCB as the regret grows linearly with respect to $T$. In comparison, the regret grows in a logarithmic fashion with strategic CUCB, which agrees with our theoretical finding of the $O(\log T)$ regret bound.
\paragraph{Cumulative regret v.s strategic budget} We conduct our empirical study with and the strategic budget ranging from $1$ to $100$ under $\gamma = 0.2$. Figure \ref{Fig2} shows that the regret is linear with the largest strategic budget $B_{max}$ under both strategic CUCB and naive CUCB, which agrees with our linear term on the budget in Theorem \ref{appendixA}.
We observe that, though at a small value of $B_{max}$ CUCB display unstable performance, strategic CUCB can achieve much lower regret than that of naive CUCB in general. We, therefore, conclude the superiority of strategic CUCB in small-budget and large-budget regimes.
\paragraph{Cumulative regret under LSI strategy and random strategy} We simulate the random strategy for arms as arms randomly spend a budget that is generated from a standard uniform distribution in each round.
It is illustrated in Figure \ref{Fig3} that regardless of the choice of the bandit learning algorithm, the regret under LSI is higher than the regret under a random strategy. As is discussed previously, LSI strategy is one of the most powerful strategies that strategic arms can perform to hamper regret. The effect of strategies with collusion is demonstrated in the appendix.
\paragraph{Cumulative regret v.s. different arm subset size} We relax the number of arms in each arm subset from $2$ to $4$ and $6$ to observe the sensitivity with respect to this parameter. We choose naive CUCB with $4$ arms in each arm subset as a baseline and the maximum budget is set to be $50$. As shown in Figure \ref{Fig4}, our algorithm outperforms naive CUCB regardless of the number of arms in each arm subset, which indicates the robustness of strategic CUCB on the arm subset size.
\fi
\section{Regret analysis of strategic UCB}
\begin{namedtheorem}[Lemma \ref{lambda} (Re-statement)]
Let $\Lambda_{i,t} = \sqrt{\frac{3 \log{t} }{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}
, where $\rho_{i,t-1}$ is the total strategic budget spent by arm $i$ up to time $t-1$ and $K_{i,t-1}$ be the total number of pulls of arm $i$ up to time $t-1$. Define the event $E_t = \{ |\ \widetilde{\mu}_{i,t-1} - \mu_{i}| \leq \Lambda_{i,t}, \forall i \in [m]\}$, where $\mu_i$ is the true mean of arm $i$'s underlying distribution.
Then, $\mathbb{P}(\lnot E_t)\leq 2m \cdot t^{-2}$.
\end{namedtheorem}
\begin{proof}
Denote $\hat{\mu}_{i,t-1} = \frac{\sum_{j=1}^{t-1} x_{i,j}}{K_{i,t-1}}$ as the estimation of the expected reward without the manipulated variable $z_{i,t}$. By the definition of $\hat{\mu}_{i,t}$ and $\widetilde{\mu}_{i,t}$,
\begin{align}
& \mathbb{P}\left(|\widetilde{\mu}_{i,t-1} - \mu_i| > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}\right) \notag \\
=\ & \mathbb{P}\left(\left|\hat{\mu}_{i,t-1} + \frac{\rho_{i,t-1}}{K_{i,t-1}}- \mu_i\right| > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}} \ \right) \notag \\
=\ & \mathbb{P}\left(\hat{\mu}_{i,t-1} - \mu_i > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} \ \right) + \mathbb{P}\left(\hat{\mu}_{i,t-1} - \mu_i < -\sqrt{\frac{3\log{t}}{2K_{i,t-1}}} -2 \frac{\rho_{i,t-1}}{K_{i,t-1}} \right) \notag \\
\leq\ & \mathbb{P}\left(\hat{\mu}_{i,t-1} - \mu_i > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} \ \right) + \mathbb{P}\left(\hat{\mu}_{i,t-1} - \mu_i < -\sqrt{\frac{3\log{t}}{2K_{i,t-1}}}\right) \notag \\
=\ & \sum_{s=1}^{t-1} \mathbb{P}\left(|\hat{\mu}_{i,t-1} - \mu_i|> \sqrt{\frac{3\log{t}}{2s}}, K_{i,t-1} = s \ \right) \notag \\
\leq\ & \sum_{s=1}^{t-1} \mathbb{P}\left(|\hat{\mu}_{i,{t-1}} - \mu_i| > \sqrt{\frac{3\log{t}}{2s}}\ \right)\notag\\
\leq\ & 2t\cdot \exp{(-3\log{t})} = \frac{2}{t^2} \,, \nonumber
\end{align}
where the last inequality follows the Chernoff-Hoeffding bound. By the union bound,
\[
\quad \mathbb{P}(\lnot E_t) = \mathbb{P}\left( \{ |\widetilde{\mu}_{i,t-1} - \mu_{i}| > \Lambda_{i,t}, \forall i \in [m]\} \right)\leq 2m \cdot t^{-2} \,. \qedhere
\]
\end{proof}
\begin{namedtheorem}[Theorem \ref{appendixA} (Re-statement)]
The regret of the strategic CUCB algorithm with m strategic arms in time horizon $T$ using an $(\alpha,\beta)$-approximation oracle is at most
\begin{align}
m\cdot\Delta_{max}\left( \frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}+1 \right)\,, \notag \nonumber
\end{align}
where $f^{-1}(\cdot)$ is the inverse bounded smoothness function.
\end{namedtheorem}
\begin{proof}
We start by introducing a few notations. Let $F_t$ to be the event where the $(\alpha,\beta)$-approximation oracle fails to produce an $\alpha$-approximation answer with respect to the input $\bm{\Bar{\mu}} = (\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$ at time $t$. By definition of a $(\alpha,\beta)$-approximation oracle, we have $\mathbb{P}(F_t) \leq 1 - \beta$. Observe that at an arbitrary time $t$, a suboptimal arm subset may be selected due to two reasons, i) the $(\alpha, \beta)$-approximation oracle fails to provide an $\alpha$-approximate arm subset with respect to the input vector $\bm{\Bar{\mu}} = (\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$ and ii) the estimated mean vector $\bm{\bar{\mu}} = (\bar{\mu}_{1,t}, \bar{\mu}_{2,t},......,\bar{\mu}_{m,t})$ deviates from true values by a significant amount.
To account for $\sum^T_{t=1} \mathbb{I}(S_t \in S_B)$,where $\mathbb{I}(S_t \in S_B)$ is 1 when algorithm choose a suboptimal arm subset, $S_t \in S_B$, at time $t$, we introduce a counter $N_i$ for each arm $i\in[m]$ after the $m$-round initialization and let $N_{i,t}$ be the value of $N_i$ at time $t$. We initialize $N_{i,m} = 1$. By definition, $\sum_{i \in [m]} N_{i,m} = m$. For $t > m$, the counter $N_{i,t}$ is updated as follows.
\begin{itemize}
\item If $S_t \in S_B$, then $N_{{i'},t} = N_{{i'},t-1} + 1$ where $i' = \argmin_{i \in S_t} N_{i,t-1}$. In the case that $i'$ is not unique, we break ties arbitrarily.
\item If $S_t \not \in S_B$, no counters will be updated then.
\end{itemize}
As such, the total number of pulls of suboptimal arm subsets is less than or equal to $\sum_{i=1}^m N_{i,T}$.
Define $\psi_t = \frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{t}} {\left(f^{-1}(\Delta_{min})\right)^2} > c\,, $
where $c$ is the larger solution of the following equation,
\begin{align}
&\frac{\left(f^{-1}(\Delta_{min})\right)^2}{4} c^2
- \left(2B_{max} f^{-1}(\Delta_{min}) + \frac{3\log{t}}{2}\right)c + 4B_{max}^2 = 0\,. \label{square}
\end{align}
By solving Equation \eqref{square}, we have
\begin{align}
c &= \frac{2\left(2B_{max} f^{-1}(\Delta_{min}) + \frac{3\log{t}}{2}\right)} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{2\sqrt{\left(2B_{max} f^{-1}(\Delta_{min}) + \frac{3\log{t}}{2}\right)^2 -4(f^{-1}(\Delta_{min})) B_{max}}} {\left(f^{-1}(\Delta_{min})\right)^2}\,. \label{psi}
\end{align}
We then decompose the total number $\sum_{i=1}^m N_{i,T}$ of pulls of suboptimal arm subsets as
\begin{align}
\quad\sum_{i=1}^m N_{i,T} =\ & m +\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B\} \notag \\
=\ & m +\sum_{t=m+1}^T \sum_{i\in[m]}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} \leq \psi_t\} \notag \\
& + \sum_{t=m+1}^T \sum_{i\in[m]}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} > \psi_t\}\notag \\
\leq\ & m + m\psi_T +\sum_{t=m+1}^T \sum_{i\in[m]}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} > \psi_t\}\notag \\
=\ & m + m\psi_T+\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B , \forall i \in S_t, N_{i,t-1} > \psi_t\} \,.\label{8}
\end{align}
The first inequality follows as $\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B, \forall i \in S_t, N_{i,t-1} \leq \psi_t\}$ can be trivially upper bounded by $\psi_t$ and the second equality holds by our rule of updating the counters.
The third term of Equation \eqref{8} can be further decomposed according to whether the oracle fails,
\begin{align}
&\quad \sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B, N_{i,t-1} > \psi_t, \forall i \in S_t\}\notag\\
&\leq \sum_{t=m+1}^T (\mathbb{I}\{F_t\} + \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\})\label{ke} \\
&= (T-m)(1-\beta)+\sum_{t=m+1}^T \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t , \forall i \in S_t\}\,. \notag
\end{align}
Let $\Lambda_{i,t} = \sqrt{\frac{3 \log{t} }{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}$ where $\rho_{i,t-1}$ is the total strategic budget spent by arm $i$ up to time $t-1$. Define event $E_t = \{ |\ \widetilde{\mu}_{i,t-1} - \mu_{i}| \leq \Lambda_{i,t}, \forall i \in [m]\}$, where $\mu_i$ is the true mean of arm $i$'s underlying distribution without manipulation. We continue the proof under $\mathbb{P}(\{ E_t, \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t \}) = 0$ and prove it afterwards.
Since $\mathbb{P}\left(\{ E_t, \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t , \forall i \in S_t\} \right) = 0$, by inclusion-exclusion principle,
$ \mathbb{P}(\{\lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\}) \leq \mathbb{P}(\lnot E_t)\,.$ Denote $\hat{\mu}_{i,t-1} = \frac{\sum_{j=1}^{t-1} x_{i,j}}{K_{i,t-1}}$ as the estimation of the expected reward without the manipulated variable $z_{i,t}$. By the definition of $\hat{\mu}_{i,t}$ and $\widetilde{\mu}_{i,t-1}$,
\begin{align}
&\quad \mathbb{P}\left(|\widetilde{\mu}_{i,t-1} - \mu_i| > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}} + \frac{B_{max}}{K_{i,t-1}} \ \right) \notag \notag \\
&= \mathbb{P}\left(\left|\hat{\mu}_{i,t-1} + \frac{\rho_{i,t-1}}{K_{i,t-1}}- \mu_i\right| > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}+\frac{B_{max}}{K_{i,t-1}} \right) \notag \\
&\leq \mathbb{P}\left(\left|\hat{\mu}_{i,t-1} + \frac{\rho_{i,t-1}}{K_{i,t-1}}- \mu_i\right| > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}} \ \right) \notag \\
&\leq 2t^{-2}\,,
\end{align}
where the last inequality holds due to Lemma \ref{lambda}.
By the union bound,
\begin{align}
&\ \quad \mathbb{P}(\lnot E_t) = \mathbb{P}( \{ |\widetilde{\mu}_{i,t-1} - \mu_{i}| > \Lambda_{i,t}, \forall i \in [m] \} \ ) \leq 2m \cdot t^{-2}\,. \notag
\end{align}
Hence,
\begin{align}
\mathbb{P}(\{\lnot F_t, S_t \in S_B, \forall i \in S_t, K_{i,t-1} > \psi_t, \forall i \in S_t\}) \leq \mathbb{P}(\lnot E_t) \leq 2m \cdot t^{-2}\,. \notag
\end{align}
We now show that $\mathbb{P}(\{ E_t, \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t \}) = 0$. Let $\Lambda = \sqrt{\frac{3 \log{t}}{2\psi_t}}+ \frac{2B_{max}}{\psi_t} $, which is not a random variable, and $B_{max} = max_{i \in [m]} B_i$, where $B_i$ is the strategic budget for arm $i$. For variable $\Lambda_{i,t}$, let $\Lambda_{i,t}^{\ast} = \max \{\Lambda_{i,t}\}$. Since $K_{i,t-1} > \psi_t$ and $B_{max} \geq B_i \geq \rho_i$, we have $\Lambda > \Lambda_{i,t}^{\ast}$. According to line 7 of Algorithm 1, we have $\Bar{\mu}_{i,t}= \widetilde{\mu}_{i,t-1} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}} + \frac{B_{max}}{K_{i,t-1}}$ and $\bar{\mu}_{i,t} = \hat{\mu}_{i,t-1} + \frac{\rho_{i,t-1}}{K_{i,t-1}} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}}+ \frac{B_{max}}{K_{i,t-1}}$. Thus, $|\hat{\mu}_{i,t-1} - \mu_i| \leq \Lambda_{i,t}$ implies that $0 < \bar{\mu}_{i,t-1} - \mu_i \leq 2\Lambda_{i,t} \leq 2\Lambda_{i,t}^{\ast} \leq 2\Lambda$, for $ i \in S_t$.
Recall $S^{\ast} = \argmax_{S \in \mathcal{S}} r_{\bm{\mu}}(S)$ and ${OPT}_{\bm{\mu}} = \max_{S\in \mathcal{S}} r_{\bm{\mu}}(S)$. Suppose $\{E_t, \lnot F_t, S_t \in S_B, \forall i \in S_t, K_{i,t-1} > \psi_t \} $ happens at time $t$, the following holds
\begin{align}
\quad r_{\bm{\mu}}(S_t) + f(2\Lambda) \notag & \geq r_{\bm{\mu}}(S_t) + f(2\Lambda_{i,t}^{\ast}) \geq r_{\bm{\bar{\mu}}}(S_t) \\ \notag
& \geq \alpha \cdot {OPT}_{\bar{\bm{\mu}}}\geq \alpha \cdot r_{\bar{\bm{\mu}}}(S^{\ast}_{\bm{\mu}}) \geq \alpha \cdot r_{{\bm{\mu}}} (S^{\ast}_{\bm{\mu}}) \\ \notag
& = \alpha \cdot {OPT}_{\bm{\mu}}\,. \notag
\end{align}
The first inequality is due to the strict monotonicity of $f(\cdot)$ and $\Lambda > \Lambda_{i,t}^{\ast}$. The second inequality is due to the bounded smoothness property and $|\bar{\mu}_{i,t-1} - \mu_i | \leq 2\Lambda_{i,t}$. The third inequality is due to the fact that $\lnot F_t$ implies $S_t \geq \alpha \cdot opt_{\bm{\mu}}$. The forth inequality is by the definition of $opt_{\bm{\mu}}$. The last inequality is is due to the monotonicity of $r_{\bm{\mu}}(S)$ and $0 < \bar{\mu}_{i,t-1} - \mu_i$.
Let $\kappa = \sqrt{\frac{3 \log{t}}{2c}} + \frac{2B_{max}}{c}$ where $c$ takes the value defined in Equation \eqref{square}. Given $\psi_t > c$, we have $\Lambda = \sqrt{\frac{3 \log{t}}{2\psi_t}} + \frac{2B_{max}}{\psi_t} < \kappa$. By Equation \eqref{psi}, we have $f(2\Lambda) < f(2\kappa) = \Delta_{min}$ and $\Delta_{min} > \alpha \cdot opt_{\bm{\mu}} - r_{\bm{\mu}}(S_t)$, which contradicts the definition of $\Delta_{min}$ and the fact that $S_t \in S_B$. Therefore,
\begin{equation}
\mathbb{P}(E_t, \lnot F_t, S_t \in S_B, \forall i \in S_t, K_{i,t-1} > \psi_t )= 0\,. \notag
\end{equation}
Based on the above analysis, we can upper bound the total number of suboptimal pulls as
\begin{align}
\mathbb{E}\left[\sum_{i=1}^m N_{i,T} \right]
&\leq m(1+\psi_T) + (T-m)(1-\beta) + \sum_{t=m+1}^T \frac{2m}{t^2}\notag \\
&= m\left(1+\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2}\right) + (T-m)(1-\beta) + \sum_{t=m+1}^T \frac{2m}{t^2}\,.
\end{align}
Since the cumulative regret relate closely to the total number of suboptimal pulls $\mathbb{E}[\sum_{i=1}^m N_{i,T}]$, the upper bound of cumulative regret is thus
\begin{align}
\quad Regret_{\bm{\mu},\alpha,\beta}(T)
\leq\ & T \cdot \alpha \beta \text{OPT}_{\bm{\mu}} - \left( T \cdot \alpha \text{OPT}_{\bm{\mu}} - \mathbb{E} \left[ \sum_{i=1}^m N_{i,T} \right] \Delta_{max} \right)
\notag \\
=\ & (\beta-1) T \cdot \alpha \text{OPT}_{\bm{\mu}} + \Delta_{max}\left( m \left(1+\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2}\right)\notag \right) \notag \\
&+ \Delta_{max}\left((T-m) (1-\beta) +\sum_{t=m+1}^T \frac{2m}{t^2}\right)\notag\\
\leq\ & ((T-m) \cdot \Delta_{max}- T \cdot \alpha \text{OPT}_{\bm{\mu}})(1-\beta)
\notag \\
& + m \cdot \Delta_{max}\left(1+\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}\right)\,.\notag
\end{align}
Each time the algorithm pull a suboptimal arm subset $S_t \in S_B$ at time $t$, the algorithm incur an additional regret of at most $\Delta_{max}$, which is less than or equal to $\alpha \cdot opt_{\bm{\mu}}$. Thus,
\begin{align}
&\quad (T-m) \Delta_{max}-T\alpha \cdot \text{OPT}_{\bm{\mu}} \notag\\
&\leq (T-m) \alpha \cdot \text{OPT}_{\bm{\mu}}- T\alpha\cdot \text{OPT}_{\bm{\mu}}\notag\\
&=-m \alpha\cdot \text{OPT}_{\bm{\mu}} < 0\,.\notag
\end{align}
As a result, the regret of the strategic CUCB algorithm under strategic manipulations of reward is at most
\begin{align}
&Regret_{\bm{\mu},\alpha,\beta}(T)\leq m \cdot \Delta_{max} \left(\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3} + 1\right)\notag \,. \qedhere
\end{align}
\end{proof}
\section{Lower bound on the strategic budget}
\begin{namedtheorem}[Theorem \ref{lower}]
In stochastic multi-armed bandit problems, for a strategic suboptimal arm $i$, under time horizon $T$ and without access to other arms' information, the minimum strategic budget needed for it to be pulled $\omega(\log{T})$ is $\omega(\log{T})$. The regret incurred for any bandits learning algorithm is thus $\omega(\log{T})$.
\end{namedtheorem}
\begin{proof}
Let time $t \in [1,T]$ be the time step arm $i$ is last pulled under UCB algorithm and $\eta$ is a parameter chosen by the algorithm. The following inequality must stands at time $t$,
\begin{align*}
& \hat{\mu}_{i,t} + \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}} + \frac{\rho_i}{K_{i,t}}\geq \hat{\mu}_{i^{\ast},t} + \sqrt{\frac{2\log(K_{i^{\ast},t}^2/\eta^2)}{K_{i^{\ast},t}}}\,.
\end{align*}
By the Chernoff-Hoeffding bound and the union bound,
\begin{align*}
&\quad \mathbb{P} \left(\hat{\mu}_{i,t} - \mu_i \geq \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}}\right)\\
&\leq \sum^{t}_{s=1} \mathbb{P} \left( \hat{\mu}_{i,t} - \mu_i \geq \sqrt{\frac{2\log(s^2/\eta^2)}{s}}, K_{i,t} = s \right) \\
&\leq\sum^{t}_{s=1} \mathbb{P} \left( \hat{\mu}_{i,t} - \mu_i \geq \sqrt{\frac{2\log(s^2/\eta^2)}{s}}\right) \\
& \leq \sum^{t}_{s=1} \frac{\eta^2}{s^2} = \eta^2\sum^{t}_{s=1} \frac{1}{s^2} \leq \frac{\pi^2}{6} \eta^2 \,.
\end{align*}
Thus $\hat{\mu}_{i,t} - \mu_i \leq \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}}$ and similarly $\mu_{i^{\ast},t} - \hat{\mu}_{i^{\ast}} \leq \sqrt{\frac{2\log(K_{i^{\ast},t}^2/\eta^2)}{K_{i^{\ast},t}}} $, each with probability of at least $1 - \frac{\pi^2}{6}\eta^2$. Hence with probability $1 - 2 \frac{\pi^2}{6}\eta^2$, we have
\begin{align}
\mu_i + \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}} + \frac{\rho_i}{K_{i,t}} \geq \mu_{i^{\ast}}\,. \notag
\end{align}
When arm $i$ is pulled, arm $i$ wants to ensure the following holds
\begin{align}
\sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}} + \frac{\rho_i}{K_{i,t}} \geq \delta_i\,,\notag
\end{align}
where $\delta_i = \mu_{i^{\ast}} - \mu_i$. Then,
\begin{equation*}
B_i \geq \rho_i \geq \left(\delta_i - \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}}\right) \cdot K_{i,t}\,. \qedhere
\end{equation*}
\end{proof}
\section{Collusion strategy with the objective of regret maximization}
Stemming from the strong correlation between the cumulative regret and the number of suboptimal pulls, the optimal collusion strategy under the regret maximization objective is a variant of the combinatorial optimization problem defined earlier. Specifically, the maximization objective of the optimization problem takes a priority weight of each $Y_i$ based on the $\delta_i = \mu_i - \mu_{i^\ast}$, as
\begin{equation*}
\begin{aligned}
& \underset{Y_{i},t_i,\ i\in [m]}{\mathrm{maximize}}
& & \delta_1 \cdot Y_{1}+\dots+ \delta_m \cdot Y_{m} \\
& \mathrm{subject}\text{ }\mathrm{to}
& & \frac{B_i}{Y_{i}} + \sqrt{\frac{3\log t_i}{Y_{i}}} \ge \delta_i + \sqrt{3\log t_i},\ i \in [m]\,, \\
& & & t_{i_j} \ge \sum_{l=1}^j Y_{i_l}, \ j\in [m]\,,\\
& & & \textup{for some permutation}\ i_1,\dots\ i_m\ \textup{of}\ [m]\,.\\
\end{aligned}
\end{equation*}
\section{Additional simulation results}
\begin{figure*}[htbp]
\centering
\subfigure[]{
\label{Fig7}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/diff_B_card4.png}}
\subfigure[]{
\label{Fig8}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/diff_scale.png}}
\caption{\ref{Fig7} Cumulative regret v.s. time horizon with different level of strategic budget; \ref{Fig8} Effects of tuning UCB exploration parameter.}
\end{figure*}
\begin{figure*}[htbp]
\centering
\subfigure[]{
\label{Fig9}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/diff_arm_1010.png}}
\subfigure[]{
\label{Fig10}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/diff_arm_100.png}}
\caption{\ref{Fig9} Cumulative regret v.s. time horizon with different number of arms in each arm subset under $B_{max} = 10$; \ref{Fig10} Cumulative regret v.s. time horizon with different number of arms in each arm subset under $B_{max} = 100$.}
\end{figure*}
\begin{figure*}[htbp]
\centering
\subfigure[]{
\label{Fig5}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/pulls_final_final.png}}
\subfigure[]{
\label{Fig6}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/regret_final_final.png}}
\caption{\ref{Fig5} The number of suboptimal pulls v.s. collusion strategies; \ref{Fig6} Regret v.s. collusion strategies}
\end{figure*}
\bigbreak
\paragraph{Experiment setting} All of the experiment settings remain the same as the experiments section unless otherwise indicated.
\paragraph{Cumulative regret v.s. maximum budget} We experiment with $B_{max} = \{0, 10, 50, 100\}$ and $\gamma$ = $0.2$. Previously we have empirically evaluated the performance of strategic CUCB when each suboptimal arm process the same amount of strategic budget in Figure \ref{Fig1}. We now simulate the situation when suboptimal arms have strategic budgets of different levels by uniformly randomly assigning budgets from $[0,B_{max}]$ to each suboptimal arm. As shown in Figure \ref{Fig7}, the performance of strategic CUCB is still robust when there is a diversity of strategic budget owned by arms. Meanwhile, our strategic CUCB algorithm still outperforms naive CUCB under this scenario.
\paragraph{Effects of tuning UCB exploration parameter} As shown in Figure \ref{Fig8}, our strategic CUCB achieves the best performance with $\gamma = 0.2$. This determines the discount parameter used in the experiment section.
\paragraph{More experiments on arm subset size} Similar to the experiment section, we relax the number of arms in each arm subset from $2$ to $4$ and $6$. We choose naive CUCB with $4$ arms in each arm subset as a baseline. We experiment on $B_{max} \in \{10, 50, 100\}$, respectively. Figure \ref{Fig9} and Figure \ref{Fig10} showcase that our algorithm outperforms naive CUCB regardless of the number of arms in each arm subset under all budget values.
\paragraph{Performance of different collusion strategies}
We experiment with three collusion strategies that are variants of the lump sum investing (LSI) strategy, which is the dominant Nash equilibrium strategy for the subgame between strategic arms. The maximum budget is set to be $50$. In prioritized budget LSI (PB-LSI) strategy, each strategic arm $i$ spends $\delta_i = \mu_{i^{\ast}} - \mu_i$ in the initialization rounds and then spends the rest of their budget all at once by the order of the value of their strategic budget. The prioritized delta LSI (PD-LSI) and prioritized budget-delta LSI (PBD-LSI) operate in similar fashions, but strategic arms spend budget according to the increasing order of their suboptimality gap $\delta_i = \mu_{i^{\ast}} - \mu_i$ and the increasing order of $B_i - \delta_i$, respectively. Figure \ref{Fig5} and Figure \ref{Fig6} showcase scenarios where the LSI strategy is suboptimal under the common objective of maximizing suboptimal pulls and maximizing regret. The figures show that considering the budget or suboptimal gap might not be sufficient for the decision of the collusion strategy, under both the objectives. It is clear that PBD-LSI outperforms the other three strategies.
However, as the suboptimality gaps are constants, the effects of different collusion strategies on cumulative regret is limited and is fully observed in a short horizon.
\subsubsection*{References}}
\usepackage{mathtools}
\usepackage{tikz}
\DeclareSymbolFont{extraup}{U}{zavm}{m}{n}
\DeclareMathSymbol{\varheart}{\mathalpha}{extraup}{86}
\DeclareMathSymbol{\vardiamond}{\mathalpha}{extraup}{87}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\mathbb{P}}{\mathbb{P}}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{\mathbb{I}}{\mathbb{I}}
\newcommand{\mathbb{N}}{\mathbb{N}}
\newcommand{\mathbb{B}}{\mathbb{B}}
\newcommand{\mathbb{Q}}{\mathbb{Q}}
\newcommand{\mathbb{Z}}{\mathbb{Z}}
\newcommand{{O}}{{O}}
\newcommand{{o}}{{o}}
\newcommand{\mathbf{e}}{\mathbf{e}}
\newcommand{\mathcal{S}}{\mathcal{S}}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{\mathcal{D}}{\mathcal{D}}
\newcommand{\mathcal{M}}{\mathcal{M}}
\newcommand{\mathcal{N}}{\mathcal{N}}
\newcommand{\mathcal{I}}{\mathcal{I}}
\newcommand{\mathcal{Y}}{\mathcal{Y}}
\newcommand{\mathcal{T}}{\mathcal{T}}
\newcommand{\mathcal{R}}{\mathcal{R}}
\newcommand{\overline{R}}{\overline{R}}
\newcommand{\boldsymbol{a}}{\boldsymbol{a}}
\newcommand{\mathcal{S}}{\mathcal{S}}
\newcommand{\mathbb{A}}{\mathbb{A}}
\newcommand{\mathbbm{1}}{\mathbbm{1}}
\def\mathcal{A}{\mathcal{A}}
\def\mathcal{B}{\mathcal{B}}
\def\mathcal{D}{\mathcal{D}}
\def\bm{V}{\bm{V}}
\def\bm{X}{\bm{X}}
\def\bm{\overline{X}}{\bm{\overline{X}}}
\def\bm{Y}{\bm{Y}}
\def\bm{Z}{\bm{Z}}
\def\bm{W}{\bm{W}}
\def\bm{A}{\bm{A}}
\def\bm{\hat{A}}{\bm{\hat{A}}}
\def\bm{B}{\bm{B}}
\def\bm{\hat{B}}{\bm{\hat{B}}}
\def\bm{x}{\bm{x}}
\def\bm{y}{\bm{y}}
\def\bm{z}{\bm{z}}
\def\bProofm{u}{\bProofm{u}}
\def\underline{X}{\underline{X}}
\def\underline{Y}{\underline{Y}}
\def\underline{Z}{\underline{Z}}
\def\underline{x}{\underline{x}}
\def\underline{y}{\underline{y}}
\def\underline{z}{\underline{z}}
\def{\varepsilon}{{\varepsilon}}
\def\{0,1\}{\{0,1\}}
\def\mathrm{NP}{\mathrm{NP}}
\def\mathrm{co}{\mathrm{co}}
\def\varepsilon{\varepsilon}
\def\oplus{\oplus}
\def\mid{\mid}
\def\hat{\ell}{\hat{\ell}}
\providecommand\floor[1]{\lfloor#1\rfloor}
\providecommand\ceil[1]{\lceil#1\rceil}
\providecommand\blog[1]{\log_2\ceil{#1}}
\providecommand\abs[1]{\lvert#1\rvert}
\providecommand\bigabs[1]{\bigl\lvert#1\bigr\rvert}
\providecommand\ip[1]{\langle#1\rangle}
\providecommand\absip[1]{\abs{\ip{#1}}}
\DeclareMathOperator\Var{\mathrm{Var}}
\DeclareMathOperator\Cov{\mathrm{Cov}}
\DeclareMathOperator\poly{\mathrm{poly}}
\DeclareMathOperator{\pr}{\mathrm{Pr}}
\newcommand{\red}[1]{\textcolor{red}{#1}}
\newcommand{\blue}[1]{\textcolor{blue}{#1}}
\newcommand{\orange}[1]{\textcolor{orange}{#1}}
\newcommand{\violet}[1]{\textcolor{violet}{#1}}
\def\mathrm{P}{\mathrm{P}}
\providecommand\norm[1]{\|#1\|}
\newcommand{\bracket}[1]{\left(#1\right)}
\newcommand{\sbracket}[1]{\left[#1\right]}
\makeatletter
\newcommand*{\rom}[1]{\expandafter\@slowromancap\romannumeral #1@}
\usepackage{booktabs}
\section{Introduction}
\label{sec:intro}
Sequential learning methods feature prominently in a range of real applications such as online recommendation systems, crowdsourcing systems, and online influence maximization problems. Among those methods, the multi-armed bandits problem serves as a fundamental framework. Its simple yet powerful model characterizes the dilemma of exploration and exploitation which is critical to the understanding of online sequential learning problems and their applications \citep{10.2307/2332286, Robbins:1952,auer2002finite,lattimore_szepesvari_2020}.
The model describes an iterative game constituted by a bandit algorithm and many arms. The bandit algorithm is required to, through a horizon $T$, choose an arm to pull at each time step. As the objective is to maximize the cumulative reward over time, the algorithm balances between exploiting immediate rewards based on the information collected or pulling less explored arms to gain more information about arms \citep{anantharam_asymptotically_1987,auer_nonstochastic_2002,cesa-bianchi_prediction_2006}.
Out of the real applications, many motivate the extension of MAB towards combinatorial multi-armed bandits (CMAB), where multiple arms can be selected in each round \citep{pmlr-v28-chen13a,DBLP:conf/nips/CombesSPL15,pmlr-v48-lif16,pmlr-v97-zimmert19a,pmlr-v117-rejwan20a}. CMAB demonstrates its effectiveness on problems like online social influence maximization, viral marketing, and advertisement placement, within which many offline variants are NP-hard. However, existing MAB and CMAB algorithms are often developed either under benign assumptions on the arms \citep{pmlr-v28-chen13a,pmlr-v38-kveton15,pmlr-v75-wei18a} or purely adversarial arms \citep{auer2002nonstochastic}. In the former setting, the arms are commonly assumed to report their reward signals truthfully without any strategic behavior, under which the drawbacks are apparent. In the latter setting, arms can attack any deployed algorithm to regret of $O(T)$ with this capability of reward manipulations, which is catastrophic for a bandit algorithm.
This assumption is stringent and rarely realistic.
In this paper, we adapt the combinatorial UCB (CUCB) algorithm with a carefully designed UCB-based exploration term.
A major difficulty stems from not knowing the manipulation term made by the arms, while our algorithm overcomes this by depending only on the knowledge of the maximum possible strategic budget.
Previous results by \citep{pmlr-v119-feng20c} only implies robustness of UCB style algorithm under stochastic multi armed bandits setting under only full knowledge of step wise deployment of strategic budget.
New tail bounds over the proposed exploration term and a new trade-off parameter that balances exploration and exploitation are utilized to facilitate the analysis of our algorithm.
We further establish results on the robustness of our UCB variant under strategic arms with an algorithm-dependent budget lower bound.
Our proposed algorithms are also evaluated empirically through an extensive set of synthetic environments and real datasets.
The real applications include reliable workers selection in online crowdsourcing systems, where workers might misrepresent a result for a better chance to be selected in the future; online information maximization, where nodes modify the spread to include itself into the seed set; and online recommendation systems, which characterizes the ``click framing'' behavior. Through a wide range of tasks and parameters, the experiment results corroborate our theoretical findings and demonstrate the effectiveness of our algorithms.
\subsection{Motivating Examples}
The setting of strategic manipulations describes the strategic behavior found in a variety of real applications.
Consider a crowdsourcing platform that provides a data labeling service for payment. The platform interacts with a group of customers and a pool of workers. The customers request tasks to the platform for labeling and the platform is then responsible for selecting workers from the worker pool to complete the tasks. This process repeats, during which the platform learns the best deployment. We maintain a mild assumption that the payments from customers are non-decreasing with the quality of labels. To maximize its profit, it is desired for the platform to select workers that provides the most reliable labels. The workflow of the platform can be described in the diagram below.
While the platform and the customers desire quality labels, it may not entirely be in the worker's interest to exert the highest effort and thus report quality labels each time, which factors into a range of reasons. Workers may adapt strategic behaviors to maximize their own utility instead. Thus, it becomes crucial to identify reliable workers to prevent strategic behaviors that jeopardize profits.
This naturally translates to a multi-armed bandits problem under strategic manipulations where the workers are the strategic arm and the payments act as the rewards. Under ideal assumptions it had shown the effectiveness of bandits algorithms on such problems \citep{jain2014quality, tran2014efficient,rangi2018multi}.
\begin{figure}[H]
\includegraphics[width=7.5cm]{figs/crowdsys.png}
\caption{Online crowdsourcing system}\label{crowdsys}
\end{figure}
\section{Related Work}
The problem of multi-armed bandits (MAB) was first investigated back in 1952 while some techniques utilized were developed back in 1933 \citep{10.2307/2332286, Robbins:1952, berry,auer2002finite,lattimore_szepesvari_2020}. Since then it has been extensively explored and serves as the foundation of many modern areas, including reinforcement learning, recommendation systems, graph algorithms, etc. \citep{DBLP:conf/iconip/BouneffoufBG12,pmlr-v70-vaswani17a, pmlr-v85-durand18a,10.5555/3367243.3367445,li2020stochastic}. With the need to model the selection of multiple arms in one round, MAB is then extended to combinatorial multi-armed bandits (CMAB), which see many deployments in real applications like news/goods recommendation, medical trials, routing, and so forth \citep{Wang_Ouyang_Wang_Chen_Asamov_Chang_2017,8022964,pmlr-v84-sankararaman18a, li2020online}.
Real applications motivate the community to derive algorithms in more adaptive environments. The first of which considers adversarial bandits with the classic EXP3 algorithm. The arms are assumed to be non-stationary but non-adaptive (which means that algorithms will adapt to the adversarial) \cite{auer2002nonstochastic}. Despite that adversarial bandits do not fall into the scale of this paper's related work, it leads tremendous effort to the following topics in adaptive arms.
\paragraph{Adversarial corruptions}
The adversary is given the ability to replace the observations of the bandit algorithm with arbitrary values within some constraints.
\citet{pmlr-v99-gupta19a} discuss the constraint that rewards are modified for at most $T_0$ rounds out of the total $T$ rounds. The asymptotic regret bound $O(mT_0)$ achieved in the work is shown to match the negative result, where $m$ is the number of arms. \citet{lykouris_stochastic_2018} discussed a different constraint where the corruption $z_t$ at time $t$ is cumulative up to some constant $B$.
The regret achieved under this constrained corruption is bounded by $O(mB\log(\frac{mT}{\delta}))$ with probability at least $1 - \delta$. \citet{JMLR:v20:18-395} restricts corruptions as independent Bernoulli events with probability $b$, while if corruption happens the reward becomes arbitrary and adversarial. This problem is addressed with median statistics with gap-dependent matching regret upper and lower bounds of $O(\sum_{i \neq i^{\ast}} \frac{1}{\Delta_i}\log(\frac{K}{\Delta_i}))$, where $i^{\ast}$ denotes the optimal arm and $\Delta_i$ is the suboptimality of arm $i$.
Several lines of research also discuss corruption topics in frequentist inference and partial monitoring \citep{pmlr-v83-gajane18a}, best arm identification \citep{pmlr-v99-gupta19a}, and adversarial bandits \citep{NEURIPS2020_e655c771}.
\paragraph{Adversarial attacks}
Different from arms in adversarial corruptions who intend to maximize the regret of the bandit algorithm, arms in the adversarial attack setting have the goal to maximize their number of pulls. The setting is first considered by \citet{NEURIPS2018_85f007f8}, where it shows that the attacker may spend a budget of $O(\log T)$ to deceit UCB and $\epsilon$-greedy algorithms to pull the target arms for at least $T - o(T)$ times. Stochastic and contextual MAB algorithms also suffer from undesired loss from this adversarial attack, under a variety of attack strategies \citep{pmlr-v97-liu19e}. Linear MAB algorithms, however, are shown to be near-optimal up to some linear or quadratic terms on the attack budget $B$ \citep{garcelon_adversarial_2020,bogunovic_stochastic_2020}.
\paragraph{Strategic manipulations}
The setting of strategic manipulations further weakens the capability of the adaptive arms and prevents the bandit algorithms from being overcautious.
The objective of an arm is still utility maximization under strategic manipulations, while each arm works on its utility individually.
Instead of having a global coordinator for the adversarial attack strategy, the strategic arms seek the best response via the dominant Nash equilibrium.
The strategic manipulation setting is first studied by \citet{pmlr-v99-braverman19b} where each arm is assumed to keep a portion of the reward in each round and arms maximize the cumulative reward it keeps.
The performance of bandit algorithms will then be catastrophic under the dominant Nash equilibrium, receiving an expected cumulative reward of $0$ almost surely.
With the utility of maximizing the number of pulls for each strategic arm, common bandits algorithms are proved to be robust with guaranteed $O(\log T)$ regret bound, but only under constant strategic budgets \citep{pmlr-v119-feng20c}.
For $\omega(\log T)$ budgets, bandits under strategic manipulations remain an open problem and will be discussed in this paper.
\section{Problem Formulation}
We consider the problem of combinatorial multi-armed bandits (CMAB) under the setting of strategic arms. In strategic CMAB, each arm is given a budget and the arm can strategically increase its emitted reward signals cumulatively up to this budget \emph{for its own interest}. This problem is a Stackelberg game that involves two parties. The bandit learning algorithm is deployed first to maximize its cumulative reward under the best response of the followers. The $m$ strategic bandits arms then deplete their budget where each of them aims to maximize its expected number of pulls. Knowing the principal's strategy and other followers' budget information, the followers are assumed to place their strategies according to the dominant Nash equilibrium between them.
Formally, given a time horizon $T$, the the bandit learning algorithm is asked to pull a subset of arms, denoted by an arm subset $S_t \in \mathcal{S} $ at each time $t\in [T]$, where $\mathcal{S} = { \{0,1\}^{m}} $ is the set of all possible arm subsets. At the time $t$, based on the principal's choice of arm subset, stochastic rewards $x_{i,t}$ are generated by arm $i$ from the underlying $1$-sub-Gaussian distribution with mean $\mu_i$, independent of other arms in $S_t$. The principal does not observe these rewards. Instead, each arm can increase the emitted reward signal by an arbitrary amount $z_{i,t}\ge 0$, as long as the cumulative manipulation over the horizon does not exceed a given budget $B_i$. The principal then receives the semi-bandit feedback $\{x_{i,t}+z_{i,t}\}_{i\in S_t}$.
Let $\bm{\mu} = (\mu_1, \mu_2, ..., \mu_i) $ be the vector of expectations of all arms. The expected reward of playing any arm subset S in
any round $r_{\bm{\mu}}(S) = \mathbb{E}[R_t(S)]$,which is a function of arm subset $S$ and $\bm{\mu}$. The reward function $r_{\bm{\mu}}(S)$ is assumed to satisfy two axioms:
\begin{itemize}
\item \textbf{Monotonicity.} The expected reward of playing any arm subset $S \in \mathcal{S}$ is monotonically non-decreasing with respect to the expected reward vector. That is, if for all $i \in [m], \mu_i \leq \mu_{i}'$, then $r_{\bm{\mu}}(S) \leq r_{\bm{\mu}'}(S)$.
\item \textbf{Bounded smoothness.} There exists a strictly increasing function $f(\cdot)$, termed the bounded smoothness function, such that for any two expected reward vectors $\bm{\mu}$ and $\bm{\mu'}$ satisfying $\|\bm{\mu}-\bm{\mu'}\|_{\infty}\le \Lambda $, we have $|r_{\bm{\mu}} - r_{\bm{\mu'}}| \leq f({\Lambda})$.
\end{itemize}
These axioms cover a rich set of reward functions and the explicit forms of $R_t(S)$ and $r_{\bm{\mu}}(S)$ are not needed to be specified \citep{pmlr-v28-chen13a}.
Without loss of generality, assume that $S^\ast = \argmax_{S\subseteq \mathcal{S}} r_{\bm{\mu}}(S)$ is the unique optimal subset of arms. When placing reward manipulations, each strategic arm has access to its own history $h_{i,t} = \{I_{i,t'}, x_{i,t'}, z_{i,t'}\}_{t'\ge 1}$, where $I_{i,t'}$ is the indicator of whether arm $i$ is pulled at time $t'$ and $t' < t$. The strategy of arm $i$ is determined by a function that maps this history to a manipulation $z_{i,t}$, as $Z_{i,t} \colon h_{i,t-1} \to \mathbb{R}$. Without loss of generality, we assume that arms in the optimal arm subset $i\in S^\ast$ have strategic budgets of $0$, which restricts their $z_{i,t}$ to be $0$.
In the combinatorial setting, even with the exact reward vector $\bm\mu$ provided, it can be hard to exactly compute the optimized $r_{\bm{\mu}}(S)$. In view of this, many have studied probabilistic approximation algorithms in combinatorial problems, which indicates that an ($\alpha,\beta$)-approximation oracle defined below can be usually available.
\begin{definition}[Approximation oracle]
Let $0\le \alpha,\beta \leq 1$ and define ${OPT}_{\bm{\mu}} = \max_{S\in \mathcal{S}} r_{\bm{\mu}}(S)$.
An oracle is called an ($\alpha,\beta$)-approximation oracle if it takes an expected reward vector $\bm{\mu}$ as input and outputs an arm subset $S \in \mathcal{S}$ such that $\mathbb{P}(r_{\bm{\mu}}(S) \geq \alpha \cdot {OPT}_{\bm{\mu}}) \geq \beta$. That is, the oracle gives an arm subset $S$ that is at least as good as $\alpha$ times the reward of an optimal arm subset with probability at least $\beta$.
\end{definition}
Denote ${S_B} = \{ r_{\bm{\mu}}(S) < \alpha \cdot {OPT}_{\bm{\mu}} | S \in \mathcal{S}\}$ to be the set of suboptimal arm subsets under the approximation oracle.
Note that a suboptimal arm subset can be given by the oracle for two reasons. The ($\alpha,\beta$)-approximation oracle can fail, which happens with probability at most $1-\beta$. The estimation of $\bm{\mu}$ can deviate from the true value by a significant amount, resulting in accurate input to the oracle.
The objective of the principal is to maximize the expected cumulative reward before manipulation over the time horizon $T$. Equivalently, the principal minimizes the regret, the cumulative difference between the scaled optimal reward and expected actual reward, as defined below.
\begin{definition}[Regret]
With access to an ($\alpha,\beta$)-\textit{approximation oracle}, the regret of a combinatorial bandit algorithm for $T$ rounds is
\[
Regret_{\bm{\mu},\alpha,\beta}(T) = T \cdot \alpha \cdot \beta \cdot {OPT}_{\bm{\mu}} - \mathbb{E}\big[\sum^T_{t=1}r_{\bm{\mu}}(S_t)\big]\,,
\]
where the randomness in $\mathbb{E}[\sum^T_{t=1}r_{\bm{\mu}}(S_t)]$ involves the stochasticity of the bandit algorithm and the oracle.
\end{definition}
The objective of each strategic arm $i\in [m]\setminus S^{\ast}$, however, is to maximize the number $\sum_{t=1}^T I_{i,t}$ of times it is pulled over the time horizon. To achieve this, the arm needs to confuse the principal by deviating the emitted reward signals up to the possessed budget.
\section{Strategic Combinatorial UCB}
We now propose a variant of combinatorial upper confidence bound algorithm that is robust to strategic manipulations of rewards in Algorithm \ref{algo}. The only mild assumption we maintain is that the learning algorithm has the knowledge of the largest budget possessed among all bandits arms, i.e, $B_{max}$. This is a relaxation of the assumptions made by \citet{pmlr-v119-feng20c} on the strategic UCB algorithm, in which the learning algorithm has access to the cumulative use of budget at every time step. We start with a detailed description of the algorithm and then analyze the theoretical upper bound of regret, which enjoys $O(m \log{T} + m B_{max})$.
For each arm $i$, our algorithm maintains a counter $K_{i,t-1}$ as the total number of times arm $i$ has been pulled up to time $t-1$ and $\widetilde{\mu}_{i,t} = \frac{\sum_{j=1}^{t-1} (x_{i,j}+ z_{i,j})}{K_{i,t-1}}$ as the empirical mean estimation based on the observations. At each time step, the algorithm computes the UCB estimation $\Bar{\mu}_{i,t}= \widetilde{\mu}_{i,t} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}} + \frac{B_{max}}{K_{i,t-1}}$ for $i \in [m]$. With $\Bar{\mu}_{i,t}$, the $(\alpha, \beta)$-approximation oracle then outputs an approximately optimal arm subset $S_t \in \mathcal{S}$. The algorithm plays the return arm subset and update the counter $K_{i,t}$ and the estimation $\widetilde{\mu}_{i,t}$ accordingly.
\begin{algorithm}[htb]
\DontPrintSemicolon
{\bf Input:}
Horizon $T$, number $m$ of arms, maximum budget $B_{max}$\;
{\bf Output:} Arm subset $S_t$\;
Initialize $K_{i,0} = 0$ and $\widetilde{\mu}_{i,0} = 0$ for $i \in [m]$\;
\For{t = 1 $\to$ m}{
Play an arbitrary arm subset $S_t \in \mathcal{S}$ such that $t \in S_t$\;
$K_{i,t} = K_{i,t-1} + 1$ for $i \in S_t$\;
$\widetilde{\mu}_j = x_{i,t}+z_{i,t}$ for $i \in S_t$ \;}
\For{t = m+1 $\to$ T}{
For $i \in [m]$, compute $\Bar{\mu}_{i,t}= \widetilde{\mu}_{i,t-1} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}} + \frac{B_{max}}{K_{i,t-1}} $\;
$S_t$ = {\textbf{oracle}}$(\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$\;
Play arm subset $S_t$ and update
$K_{i,t} = K_{i,t-1} + 1$ and $ \widetilde{\mu}_{i,t} = \frac{\sum_{j=1}^{t} (x_{i,j} + z_{i,j})}{K_{i,t}}$ for $i \in S_t$ \;
}
\caption{SCUCB}
\label{algo}
\end{algorithm}
We first introduce a few notations that are used in our results. Define, for any arm $i\in[m]$, the suboptimality gaps as
\begin{align*}
&\Delta^i_{min} = \alpha \cdot {OPT}_{\bm{\mu}} - \max\{r_{\bm{\mu}}(S) \mid S \in S_B, i \in S\}\,, \notag \\
&\Delta^i_{max} = \alpha \cdot {OPT}_{\bm{\mu}} - \min\{r_{\bm{\mu}}(S) \mid S \in S_B, i \in S\}\,.
\end{align*}
We then denote the maximum and minimum of the suboptimality gaps as $\Delta_{max} = \max_{i \in [m]} \Delta^i_{max}\notag $ and $\Delta_{min} = \min_{i \in [m]} \Delta^i_{min}$.
The following lemma re-establish the canonical tail bound inequality in UCB under the setting of strategic manipulations.
\begin{lemma}\label{lambda}
Let ${\Lambda}_{i,t} = \sqrt{\frac{3 \log{t} }{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}$, where $\rho_{i,t-1}$ is the total strategic budget spent by arm $i$ up to time $t-1$ and $K_{i,t-1}$ be the total number of pulls of arm $i$ up to time $t-1$. Define the event $E_t = \{ |\ \widetilde{\mu}_{i,t-1} - \mu_{i}| \leq \Lambda_{i,t}, \forall i \in [m]\}$, where $\mu_i$ is the true mean of arm $i$'s underlying distribution.
Then, $\mathbb{P}(\lnot E_t)\leq 2m \cdot t^{-2}$.
\end{lemma}
Armed with Lemma \ref{lambda}, we present one of our main theorems, Theorem \ref{appendixA}, which gives the regret bound of $O(\log T)$ of SCUCB. The outline of the proof follows that of CUCB by \citet{pmlr-v28-chen13a}. To complete the proof, we carefully choose $\psi_t$, which controls the trade-off between the exploration and exploitation periods.
\begin{theorem}
\label{appendixA}
The regret of the SCUCB algorithm with $m$ strategic arms in time horizon $T$ using an $(\alpha,\beta)$-approximation oracle is at most
\begin{align}
m\cdot\Delta_{max}\left( \frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}+1 \right)\,, \notag \nonumber
\end{align}
where $f^{-1}(\cdot)$ is the inverse bounded smoothness function.
\end{theorem}
\begin{proof}[Proof sketch]
We introduce a counter $N_i$ for each arm $i\in[m]$ after the $m$-round initialization and let $N_{i,t}$ be the value of $N_i$ at time $t$. We initialize $N_{i,m} = 1$. By definition, $\sum_{i \in [m]} N_{i,m} = m$. For $t > m$, the counter $N_{i,t}$ is updated as follows:
\begin{itemize}
\item If $S_t \in S_B$, then $N_{{i'},t} = N_{{i'},t-1} + 1$ where $i' =$ $\argmin_{i \in S_t}$ $N_{i,t-1}$. In the case that $i'$ is not unique, we break ties arbitrarily;
\item If $S_t \notin S_B$, then no counters will be updated.
\end{itemize}
As such, the total number of pulls of suboptimal arm subsets is less than or equal to $\sum_{i=1}^m N_{i,T}$.
Define $\psi_t = \frac{8B_{max}f^{-1}(\Delta_{min}) + 6\log{t}} {\left(f^{-1}(\Delta_{min})\right)^2} > c $, where $c$ is the larger solution of
\begin{align*}
\label{square}
&(f^{-1}(\Delta_{min}))^2 c^2 + 16B_{max}^2 - ((8B_{max} f^{-1}(\Delta_{min})- 6\log{t}) c = 0 \,.
\end{align*}
Then, we decompose the total number $\sum_{i=1}^m N_{i,T}$ of pulls of suboptimal arm subsets as
\begin{align}
\sum_{i=1}^m N_{i,T} = \ & m +\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B\} \notag \\
=\ &m +\sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} \leq \psi_t\} \notag \\
&+ \sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} > \psi_t\}\notag \\
\leq \ & m + m\psi_T+\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}\,. \notag
\end{align}
The inequality follows as $\sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} \leq \psi_t\}$ can be trivially bounded by $m\psi_T$. Thus the key to bound the total number of pulls of suboptimal arm subset is to upper bound $\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}$. Let $F_t$ denotes the event that the oracle fails to provide an $\alpha$-approximate arm subset with respect to the input vector $\Bar{\bm{\mu}} = (\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$.
Then we can decompose $\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}$ as
\begin{align}
&\quad \sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}\notag\\
&\leq \sum_{t=m+1}^T (\mathbb{I}\{F_t\} + \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\})\notag \\
&\leq (T-m)(1-\beta) + \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\}\notag \,. \notag
\end{align}
By leveraging the monotonicity and smoothness assumptions of the reward function, we show that $\mathbb{P}(\{\lnot F_t, E_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\}) = 0$. Therefore, by the inclusion-exclusion principle,
\begin{align}
\mathbb{P}(\{\lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\})
&\leq \mathbb{P}(\lnot E_t) \leq 2m \cdot t^{-2} \,.\notag
\end{align}
Leveraging the upper bound of total number of pulls of suboptimal arm subsets, the regret is bounded by
\begin{align}
& \quad Regret_{\bm{\mu},\alpha,\beta}(T)
\notag \\
& \leq T\alpha \beta \cdot \text{OPT}_{\bm{\mu}} - \left( T\alpha\cdot \text{OPT}_{\bm{\mu}} - \mathbb{E} \left[ \sum_{i=1}^m N_{i,T} \right]\cdot \Delta_{max} \right)
\notag \\
&\leq m\cdot \Delta_{max}\left(\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}+1\right)\,. \notag \qedhere
\end{align}
\end{proof}
It remains in question whether a bandit algorithm can achieve a regret upper bound sublinear in $B_{max}$.
Our conjecture is negative.
In fact, under strategic manipulations of rewards, the design of robust bandit algorithms, e.g. UCB and $\epsilon$-greedy, is analogous to the design of outlier-robust mean estimation algorithms. Existing works on robust mean estimation, such as \citet{steinhardt_et_al:LIPIcs:2018:8368}, argue that from an information theoretical point of view mean estimation error must be depending on the variance of the data. Casting this argument to bandit with the strategic manipulation setting, we believe that a tight regret bound is unlikely to be independent of the strategic budget $B_{max}$. Moreover, Proposition 4 in \citet{steinhardt_et_al:LIPIcs:2018:8368} implies that the dependency on $B_{max}$ is linear. This corresponds to our linear dependency of regret in Theorem \ref{appendixA}.
\section{Lower Bounds for strategic budget}
To fully understand the effects of strategic manipulations, we investigate the relationship between the strategic budget and the performance of UCB-based algorithms. Our results provide the dependency between an arm's strategic budget and the number of times it is pulled, which influence the regret of the algorithm incurred by the arm. Our analysis gains some insight from \citet{zuo_near_2020}, which limits the discussion to the 2-armed bandit setting and cannot be applied to the general MAB setting directly.
We first define some notations used in the theorem.
Without loss of generality, let arm $i^{\ast}$ be the optimal arm and arm $i, i \neq i^{\ast}$ be an arbitrary strategic suboptimal arm. Assume that arm $i$ has no access to the information regarding other arms. Let $K_{i,t}$ and $K_{i^{\ast},t}$ denote the number of times arm $i$ and arm $i^{\ast}$ have been pulled up to time $t$, respectively. Denote $\hat{\mu}_{i,t}$ as the empirical estimate of the underlying mean $\mu$ without manipulations, i.e., $\hat{\mu}_{i,t} = \frac{\sum^{t-1}_{j=1}x_{i,j}}{K_{i,t-1}}$. Define the suboptimality gap for each strategic arm $i$ to be $\delta_i = \mu_{i^{\ast}} - \mu_i$. We use a slightly revised UCB algorithm as the basic algorithm, where the UCB estimation term for arm $i$ is $\hat{\mu}_{i,t} + \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}}$ and $\eta$ is a confidence parameter chosen by the algorithm.
\begin{theorem}\label{lower}
In stochastic multi-armed bandit problems, for a strategic suboptimal arm $i$ without access to other arms' information, to be pulled for $\omega(k)$ in $T$ steps under the UCB algorithm where $k\ge O(\log{T})$, the minimum strategic budget is $\omega(k)$.
\end{theorem}
This dependency of $k$ can be extended to CMAB and CUCB straightforwardly when arms within an arm subset collude. Counter-intuitively, the dependency between the number of pulls of a strategic arm and its strategic budget is linear and subsequently, this infers that for a strategic arm to manipulate the algorithm to suffer an undesired regret of $\omega(k)$, where $k\ge O(\log T)$, the strategic budget must be at least $\omega(k)$.
\section{Experiments}
In this section, we evaluate our SCUCB algorithm empirically on synthetic data and three real applications, namely online worker selection in crowdsourcing, online recommendation, and online influence maximization. We highlight the best performance among all algorithms with bold text.
\subsection{Baseline Algorithms}
We compare our proposed SCUCB algorithm with both stochastic and adversarial bandits algorithms that achieves optimal asymptotic regret in their settings.
\begin{enumerate}
\allowdisplaybreaks
\item CUCB \cite{pmlr-v28-chen13a}. CUCB is the naive counterpart of our algorithm. The algorithm calculates an upper confidence interval for each arm and picks the best arm subset with the highest upper confidence interval.
\item TSCB \cite{wang2018thompson}. TSCB is the combinatorial version of the classical Thompson sampling algorithm. The algorithm maintains a prior beta distribution estimation for each arm and updates according to the received reward. At each time, the algorithm samples from the estimated distributions and pick actions according to the highest sample.
\item Combinatorial variant of EXP3 \cite{auer_nonstochastic_2002}. The algorithm maintains a weight for each arm and draws actions according to the normalized weight distribution. Upon receiving rewards, the algorithm update weight according to the classical EXP3 update rule.
\end{enumerate}
\subsection{Synthetic Experiments}
We conduct experiments presented in this section with synthetic data and compare our proposed algorithm with its naive counterpart. The approximation oracle is designed to succeed with probability $1$. Each bandit arm is modeled to follow a Bernoulli distribution with randomly populated $\mu \in [0,1]$ and all arms adapt LSI strategy.
The arms in the optimal arm subset have a strategic budget of $0$ since an additional strategic budget for optimal arms would only boost the performance of our algorithm. All other arms are equipped with a randomly allocated budget $B$, $ 0 \leq B \leq B_{max}$ by definition.
As is typical in the bandit literature, for example in \citet{auer2002finite}, we evaluate both the naive and strategic CUCB algorithms on their tuned versions, where the UCB exploration parameter is scaled by a constant factor $\gamma \in [0,1]$. To ensure reproducible results, each experiment is repeated for $10$ random seeds and the averaged result is presented.
The first set of experiment is conducted with $K = \{10, 20\}$ arms through a time horizon of $T=5000$ time steps with maximum possible budget of $B_{max} = \{70, 90, 110, 130\}$. The algorithms are asked to select $\text{Action size} = 2$ arms as an arm subset at each time step. The cumulative regret incurred by CUCB and SCUCB algorithm is presented in the table \ref{table:1} where the best performance is highlighted in bold text. Clearly, SCUCB demonstrated its effectiveness as it achieves significantly smaller regrets in various possible maximum strategic budgets.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 171.74 & 187.51 & 259.04 & 256.66 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{143.85} & \textbf{172.57} & \textbf{208.53} & \textbf{233.57} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 434.82 & 492.02 & 520.97 & 549.57 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{301.57} & \textbf{365.94} & \textbf{450.10} & \textbf{505.453} \\ \bottomrule
\end{tabularx}
\caption{Cumulative regret achieved by CUCB and SCUCB with synthetic data with various action size.}\label{table:1}
\end{table}
The next two tables reveals the advantage of SCUCB algorithm with various size of action set. The experiments are conducted with time horizon of $T=5000$ time steps with $K = \{10, 20\}$ arms and maximum possible budget of $B_{max} = \{50\}$. The algorithms are asked to select $\text{Action size} = \{2, 4, 6, 8\}$ arms as an arm subset at each time step. Once again, our SCUCB outperforms its naive counterpart by achieving much smaller cumulative regret,
the best numerical results across algorithms are highlighted in bold text.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 140.21 & 150.88 & 158.22 & 123.91 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{105.58} & \textbf{113.63} & \textbf{103.42} & \textbf{88.72} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 302.10 & 316.09 & 340.52 & 328.80 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{232.23} & \textbf{247.32} & \textbf{268.27} & \textbf{306.77} \\ \bottomrule
\end{tabularx}
\caption{Cumulative regret achieved by UCB and SCUCB with synthetic data with various $B_{max}$.}
\end{table}
To compare the two algorithms in detail, we plot the cumulative regret incurred by algorithms against the time steps. It becomes apparent that the SCUCB algorithm features a cumulative regret versus time steps line that is much smaller than the one caused by CUCB even under a higher maximum possible strategic budget. We compare the two algorithm under settings where the maximize possible budget is $B_{max} = 70, 90$ and $B_{max} = 110, 130$.
\begin{figure}[H]
\centering
\subfigure[]{
\label{Fig1}
\includegraphics[width=3.8cm]{figs/B7090_K20_N10000.png}}
\subfigure[]{
\label{Fig2}
\includegraphics[width=3.8cm]{figs/B110130_K20_N10000.png}}
\caption{\ref{Fig1} Comparison of CUCB and SCUCB with $B_{max} = 70, 90$. \ref{Fig2} Comparison of CUCB and SCUCB with $B_{max} = 110, 130$.}
\end{figure}
\subsection{Online Worker Selection in Crowdsourcing Systems}
We simulated an online crowdsourcing system that has a workflow resembled by figure \ref{crowdsys} and model it as a combinatorial bandits problem. We performed an extensive empirical analysis of our SCUCB algorithm against CUCB, the combinatorial version of Thompson sampling, and EXP3. The experiments were conducted with the Amazon sentiment dataset, where Amazon item reviews are labeled 'is/is not book' or 'is/is not negative \cite{Krivosheev:2018:CCM:3290265.3274366}. We split the dataset into two, where 'is book' contains data that only has labels 'is/is not book' and 'is negative' contains data that only has labels 'is/is not negative'. Both datasets consists of 7803 reviews, 284 workers, and 1011 tasks.
The correctness of workers and tasks can be visualized by the figures as below. Notice that most of the workers attain an accuracy of $60$ percent and above. This highlights that most of the workers we are interacting with has the ability to label most of the tasks correctly.
\begin{figure}[H]
\centering
\subfigure[]{
\label{worker}
\includegraphics[width=3.8cm]{figs/worker_visual.png}}
\subfigure[]{
\label{task}
\includegraphics[scale = 0.28]{figs/task_visual.png}}
\caption{\ref{worker} Visualization of correctness of worker's labels. \ref{task}Visualization of correctness of task's labels.}
\end{figure}
For each task, the data contains responses from 5 workers and the crowdsourcing platform is asked to select 2, 3, or 4 workers at each time step. To measure the performance of the algorithms, we made the assumption that the reward is monotonically correlated to the accuracy of the label compared to the ground truth label. Thus, we define the reward to be 1, if the worker's label is the same as ground truth label and 0 otherwise. To model the strategic behavior of the workers, we model each worker with a randomly allocated strategic level $s \in [0,1]$ such that the worker provides its honest response with probability $s$.
For the first set of experiments, we choose the maximum possible strategic budget to be $B_{max} = 500$. For actions size of $2, 3, 4$. We obtain the following cumulative rewards and cumulative regret results where the best performance is highlighted in bold text. To ensure reproducible results, each experiment is repeated 5 times and the averaged result is presented.
To further investigate the effect of the maximum possible strategic budget on the performance of algorithms, we fix the size of action size to be 2 and plot the cumulative reward/cumulative regret incurred by each algorithm again $B_{max} \in [100, 700]$. Regardless of the $B_{max}$ value, our proposed SCUCB algorithm consistently outperforms other methods. For better visualization, we omit the line for combinatorial Thompson sampling due to its relatively weaker performance.
\begin{figure}[H]
\centering
\subfigure[]{
\label{culre}
\includegraphics[width=3.8cm]{figs/culre.png}}
\subfigure[]{
\label{culreg}
\includegraphics[width=3.8cm]{figs/reg.png}}
\caption{\ref{culre} Cumulative rewards attained by algorithms with various level of $B_{max}$. \ref{culreg} Cumulative regrets attained by algorithms with various level of $B_{max}$. }
\end{figure}
The results of our crowdsourcing experiments reveal the robustness of our SCUCB algorithm under strategic manipulations. It also indicates the performance of adversarial and stochastic bandits algorithms under strategic manipulations. The combinatorial variant of Thompson sampling is vulnerable under manipulations, which agrees with our expectation given that the algorithm is a Bayesian algorithm in nature and is heavily relying on estimating the underlying distribution of arms. Its estimation can be easily perturbed under strategic manipulation, and thus leads to undesired performance results. It is also expected that the SCUCB algorithm far outperforms its naive counterpart, CUCB algorithm. The combinatorial version of EXP3 algorithm is the most competitive algorithm with our SCUCB. As the EXP3 algorithm was originally designed for a pure adversarial setting, the algorithm has some robustness under strategic manipulations. Shown in the Figure \ref{culre} and Figure \ref{culreg}, the combinatorial EXP3 algorithm consistently shows robustness across various levels of maximum possible strategic budget. However, as our SCUCB algorithm considers the maximum possible strategic budget and thus more adaptive, SCUCB far outperforms EXP3CB as $B_{max}$ increases.
\begin{table}[H]
\label{first}
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31672.6 & 46181.2 & 68244.6 \\
\multicolumn{1}{l|}{TSCB} & 19853.0 & 34226.2 & 60621.2 \\
\multicolumn{1}{l|}{EXP3} & 31569.6 & 46570.6 & 68482.2 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{32082.6} & \textbf{46623.6} & \textbf{68524.6} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 27787.4 & 42198.8 & 48125.4 \\
\multicolumn{1}{l|}{TSCB} & 39607.0 & 54153.8 & 55748.8 \\
\multicolumn{1}{l|}{EXP3} & 27890.4 & 41809.4 & 47887.8 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{27377.4} & \textbf{41756.4} & \textbf{47845.4} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31108.8 & 44757.6 & 59424.8 \\
\multicolumn{1}{l|}{TSCB} & 17240.6 & 32010.4 & 51109.0 \\
\multicolumn{1}{l|}{EXP3} & 30879.0 & 44767.2 & 58207.8 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{31759.0} & \textbf{44843.4} & \textbf{59441.4} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 28411.2 & 43566.2 & 56465.2 \\
\multicolumn{1}{l|}{TSCB} & 42279.4 & 56279.6 & 64781.0 \\
\multicolumn{1}{l|}{EXP3} & 28641.0 & 43523.4 & 57682.2 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{27761.0} & \textbf{43446.6} & \textbf{56448.6} \\ \bottomrule
\end{tabularx}
\caption{Cumulative rewards and cumulative regret attained by SCUCB and baseline algorithms with various action sizes on isbook dataset and isnegative dataset.}
\end{table}
\subsection{Online Recommendation System}
The online recommendation is a classic example of combinatorial bandits in applications. We evaluated algorithms in the latest MovieLens dataset which contains 9742 movies and 100837 ratings of the movies with each rating being $1 - 5$ \cite{10.1145/2827872}. We compare the algorithms based on the total values of ratings received, i.e. recommending a movie and receiving a rating of 5 is more desired than recommending a movie and receiving a rating of 1.
As the dataset may consist of unbalanced data for each movie, we adapted collaborative filtering and $k$-means clustering into our evaluation system. We used the collaborative filtered for training and testing and instead of recommending movies, we clustered the movies by $k$-means clustering and asked the bandits algorithm to choose one cluster at each time. The movie with the highest rating in the chosen cluster is recommended. The workflow of the experiment setup is summarized in the above diagram.
We performed the evaluation with various number of clusters (10, 20 or 30) and $B_{max} = 30, 50, 70$ through a time horizon of $T = 500$. The results presented below are averages over 5 runs to ensure reproducibility. The best performance across algorithms is highlighted in bold text. From the tables below, we conclude the effectiveness of SCUCB algorithm.
\begin{table}[tbh]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 10 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2297.60 & 2494.36 & 2631.02 \\
\multicolumn{1}{l|}{TS} & 2160.14 & 2314.76 & 2468.26 \\
\multicolumn{1}{l|}{EXP3} & 2181.11 & 2449.28 & 2430.96 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2305.75} & \textbf{2314.76} & \textbf{2636.07} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 20 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2469.93 & 2336.92 & 2159.25 \\
\multicolumn{1}{l|}{TS} & 2260.20 & 1850.55 & 1905.42 \\
\multicolumn{1}{l|}{EXP3} & 2380.74 & 2317.82 & 2025.39 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2474.69} & \textbf{2341.19} & \textbf{2177.70} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 30 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\\midrule
\multicolumn{1}{l|}{UCB} & 2443.57 & 2393.88 & 2472.31 \\
\multicolumn{1}{l|}{TS} & 2132.16 & 2248.56 & 1855.14 \\
\multicolumn{1}{l|}{EXP3} & 2368.42 & 2265.32 & 2339.49 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2436.43} & \textbf{2397.72} & \textbf{2476.52} \\ \bottomrule
\end{tabularx}
\caption{Cumulative rewards attained by algorithms with different number of clusters $10, 20, 30$ and different level of $\{B_{max} = 30, 50, 70\}$ across 500 time steps. }
\end{table}
\begin{figure}[tbh]
\centering
\includegraphics[width=7cm]{figs/arch.png}
\caption{Experiment set up of online recommendation.}
\end{figure}
\subsection{Online Influence Maximization}
We implemented the online influence maximization with an offline influence maximization algorithm TIM as our oracle \cite{10.1145/2588555.2593670}. TIM is one of the offline influence maximization algorithms that achieve asymptotic optimality. We perform the experiments on two datasets. One is a simulation dataset with 16 nodes and 44 edges, the other is a 100 nodes subset from Digg dataset, where each node represents a user from Digg website \cite{nr}. We simulate the connectivity of edges at each time step by randomly assigning each edge a connectivity probability at the start of the experiment.
\begin{figure}[hbt!]
\centering
\includegraphics[width=0.3\textwidth]{figs/test_graph.png}
\caption{Visualization of synthetic graph for evaluation.}
\label{test_graph}
\end{figure}
\begin{figure}[hbt!]
\centering
\includegraphics[width=0.35\textwidth]{figs/test.png}
\caption{ Mean rewards attained by algorithms.}\label{16c2}
\end{figure}
Figure \ref{test_graph} visualizes the synthetic graph we created for evaluation. We then tested the algorithms over $T = 800$ time steps with action size of 2 and maximum possible strategic budget of $B_{max} = 300$. Figure \ref{16c2} shows the effectiveness of algorithms by evaluating based on the averaged nodes influenced. To ensure reproducible results, we calculate the mean of influence spread over 100 trials. For better visualization, we kept a running mean of over 50 time steps to smooth the lines.
We then investigate the effect of action size and the results are summarized in the following tables \ref{table:5}.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYYY@{}}
\toprule
\multicolumn{5}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Synthetic, 16 nodes, 44 edges} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 4 & 6 & 8 \\ \midrule
\multicolumn{1}{c|}{CUCB} & 5.46 & 6.54 & 10.4 & 12.5 \\
\multicolumn{1}{c|}{TSCB} & 6.7 & 8.50 & 10.44 & 12.7 \\
\multicolumn{1}{c|}{EXP3CB} & 5.86 & 8.14 & 10.34 & 12.46 \\
\multicolumn{1}{c|}{SCUCB} & \textbf{7.44} & \textbf{9.32} & \textbf{11.16} & \textbf{13.04} \\ \bottomrule
\end{tabularx}
\end{table}
\begin{table}[H]\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{digg, 100 nodes} \\ \midrule
\multicolumn{1}{l|}{Action size} & 10 & 15 & 20 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 15.64 & 20.64 & 25.32 \\
\multicolumn{1}{l|}{TSCB} & 16.30 & 21.62 & 26.08 \\
\multicolumn{1}{l|}{EXP3} & 12.16 & 17.76 & 21.72 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{16.56} & \textbf{22.16} & \textbf{26.14} \\ \bottomrule
\end{tabularx}
\caption{Averaged final rewards attained by algorithms with synthetic and Digg dataset with action size of 2, 800 time steps and $B_{max} = 200$.}\label{table:5}
\end{table}
\section{Conclusion}
We investigate the problem of combinatorial MAB under strategic manipulations of rewards. We propose a variant of the UCB algorithm, SCUCB, which attains a regret at most $O(m \log T + m B_{max})$.
Compared to previous studies on the bandit problems under strategic manipulations, we relax on the assumption that the algorithm has access to the cumulative strategic budget spent by each arm at each time step.
For the robustness of bandit algorithms, we present lower bounds on the strategic budget for a malicious arm to incur a $\omega(\log T)$ regret of the bandit algorithm.
We provide extensive empirical results on both synthetic and real datasets with a range of applications to verify the effectiveness of the proposed algorithm. Our algorithm consistently outperforms baseline algorithms that were designed for stochastic and adversarial settings.
\subsection{Motivating Examples}
\input{motivation}
\subsection{Online Worker Selection in Crowdsourcing Platforms}
Consider a crowdsourcing platform that provides data labelling service for a payment. The platform interact with a group of customers and a pool of workers. The customers request tasks to the platform for labeling and the platform is then responsible of selecting workers from the worker pool to complete the tasks. This process repeats, during which the platform learns the best deployment. We maintain a mild assumption that the payments from customers are non-decreasing with the quality of labels. To maximize its profit, it is desired for the platform to select workers that provides most reliable labels. The workflow of the platform can be described in the diagram below.
While the platform and the customers desire quality labels, it may not entirely be in the worker's interest to exert highest effort and thus report quality labels each time, which factors into a range of reasons. Workers may adapt strategic behaviors to maximum their own utility instead. Thus, it becomes crucial to identify the reliable workers to prevent strategic behaviors that jeopardize profits.
This naturally translate to a combinatorial multi-armed bandits problem under strategic manipulations where the workers are the strategic arm and the payments acts as the rewards.
\begin{figure}[H]
\centering
\label{crowdsys}
\includegraphics[width=7.5cm]{figs/crowdsys.png}
\caption{Online crowdsourcing system}
\end{figure}
\subsection{Online Influence Maximization}
For a social network that can be represented by the below directed graph where white dots are people and black dots labeled with ``SI'' are social influencers. The temporal nature of social network results in each directed edges exists with a probability at each time step. Formally, a node is said to be reached by a seed node if there it has a directed path from the seed node. Note that this definition considers a diffusion process where a node can be reached by a seed node through multiple edges. Given a set of seed nodes, we define a spread across social network at a given time $t$ by the expected number of nodes that can be reached by the nodes in the seed node set at that time. Intuitively, this models information spread among social networks and reflects scenarios like viral marketing.
\begin{figure}[H]
\centering
\includegraphics[scale = 0.35]{figs/example.png}
\caption{Probabilistic social network}
\end{figure}
In the case of product marketing, a company may need to select the most influential social influencers to collaborate with to maximize the expected spread across time. This can be expressed as an online influence maximization problem as the company seek to maximize the size of population reached by promotions across number of promotions. However, in many cases social influencers may report inaccurate information regarding their social influential ability, which may corresponds to a number of ways in real applications, such as faking follower numbers on social media platforms.
\subsection{Suboptimal pulls maximization}
\begin{claim}
Under a strategic stochastic multi-armed bandit problem, when the strategic suboptimal arms collude to maximize the number of suboptimal pulls of the UCB algorithm with confidence parameter $\sqrt{\frac{3\log t_i}{K_{i,t-1}}}$, the maximum number of suboptimal pulls is the solution of
\begin{equation*}
\begin{aligned}
& \underset{Y_{i},t_i,\ i\in [m]}{\mathrm{maximize}}
& & Y_{1}+\dots+Y_{m} \\
& \mathrm{subject}\text{ }\mathrm{to}
& & \frac{B_i}{Y_{i}} + \sqrt{\frac{3\log t_i}{Y_{i}}} \ge \delta_i + \sqrt{3\log t_i},\ i \in [m]\,, \\
& & & t_{i_j} \ge \sum_{l=1}^j Y_{i_l}, \ j\in [m]\,,\\
& & & \textup{for some permutation}\ i_1,\dots\ i_m\ \textup{of}\ [m]\,.\\
\end{aligned}
\end{equation*}
\end{claim}
In the program, $Y_i$ represents the total number of times arm $i$ is pulled, which corresponds to $K_{i,T-1}$ used in previous sections. The permutation $i_1,...,i_m$ then describes, before the optimal arm is first pulled, the chronological order of the suboptimal arms to complete their last pull.
Observe that the confidence parameter of the optimal arm
increases with respect to $t$, the inequality holds for $t_i$ if and only if it holds for all $t\le t_i$.
By the same observation that the advantage of the optimal arm increases with time,
strategic arms who optimally spend their budgets for the maximum amount of pulls will also maximize their number of pulls till the first optimal arm pull.
The inequality condition of the combinatorial program is thus $\frac{B_i}{K_{i,t-1}} + \sqrt{\frac{3\log t_i}{K_{i,t-1}}} \ge \delta_i + \sqrt{3\log t_i}$.
We term the class of approximately optimal strategy as `sequential pull strategy', in which the strategic arms deceit the learning algorithm to pull suboptimal arms until the depletion of the budgets and this class includes the LSI strategy.
\section{Introduction}
\label{sec:intro}
Sequential learning methods feature prominently in a range of real applications such as online recommendation systems, crowdsourcing systems, and online influence maximization problems. Among those methods, the multi-armed bandits problem serves as a fundamental framework. Its simple yet powerful model characterizes the dilemma of exploration and exploitation which is critical to the understanding of online sequential learning problems and their applications \citep{10.2307/2332286, Robbins:1952,auer2002finite,lattimore_szepesvari_2020}.
The model describes an iterative game constituted by a bandit algorithm and many arms. The bandit algorithm is required to, through a horizon $T$, choose an arm to pull at each time step. As the objective is to maximize the cumulative reward over time, the algorithm balances between exploiting immediate rewards based on the information collected or pulling less explored arms to gain more information about arms \citep{anantharam_asymptotically_1987,auer_nonstochastic_2002,cesa-bianchi_prediction_2006}.
Out of the real applications, many motivate the extension of MAB towards combinatorial multi-armed bandits (CMAB), where multiple arms can be selected in each round \citep{pmlr-v28-chen13a,DBLP:conf/nips/CombesSPL15,pmlr-v48-lif16,pmlr-v97-zimmert19a,pmlr-v117-rejwan20a}. CMAB demonstrates its effectiveness on problems like online social influence maximization, viral marketing, and advertisement placement, within which many offline variants are NP-hard. However, existing MAB and CMAB algorithms are often developed either under benign assumptions on the arms \citep{pmlr-v28-chen13a,pmlr-v38-kveton15,pmlr-v75-wei18a} or purely adversarial arms \citep{auer2002nonstochastic}. In the former setting, the arms are commonly assumed to report their reward signals truthfully without any strategic behavior, under which the drawbacks are apparent. In the latter setting, arms can attack any deployed algorithm to regret of $O(T)$ with this capability of reward manipulations, which is catastrophic for a bandit algorithm.
This assumption is stringent and rarely realistic.
In this paper, we adapt the combinatorial UCB (CUCB) algorithm with a carefully designed UCB-based exploration term.
A major difficulty stems from not knowing the manipulation term made by the arms, while our algorithm overcomes this by depending only on the knowledge of the maximum possible strategic budget.
Previous results by \citep{pmlr-v119-feng20c} only implies robustness of UCB style algorithm under stochastic multi armed bandits setting under only full knowledge of step wise deployment of strategic budget.
New tail bounds over the proposed exploration term and a new trade-off parameter that balances exploration and exploitation are utilized to facilitate the analysis of our algorithm.
We further establish results on the robustness of our UCB variant under strategic arms with an algorithm-dependent budget lower bound.
Our proposed algorithms are also evaluated empirically through an extensive set of synthetic environments and real datasets.
The real applications include reliable workers selection in online crowdsourcing systems, where workers might misrepresent a result for a better chance to be selected in the future; online information maximization, where nodes modify the spread to include itself into the seed set; and online recommendation systems, which characterizes the ``click framing'' behavior. Through a wide range of tasks and parameters, the experiment results corroborate our theoretical findings and demonstrate the effectiveness of our algorithms.
\subsection{Motivating Examples}
The setting of strategic manipulations describes the strategic behavior found in a variety of real applications.
Consider a crowdsourcing platform that provides a data labeling service for payment. The platform interacts with a group of customers and a pool of workers. The customers request tasks to the platform for labeling and the platform is then responsible for selecting workers from the worker pool to complete the tasks. This process repeats, during which the platform learns the best deployment. We maintain a mild assumption that the payments from customers are non-decreasing with the quality of labels. To maximize its profit, it is desired for the platform to select workers that provides the most reliable labels. The workflow of the platform can be described in the diagram below.
While the platform and the customers desire quality labels, it may not entirely be in the worker's interest to exert the highest effort and thus report quality labels each time, which factors into a range of reasons. Workers may adapt strategic behaviors to maximize their own utility instead. Thus, it becomes crucial to identify reliable workers to prevent strategic behaviors that jeopardize profits.
This naturally translates to a multi-armed bandits problem under strategic manipulations where the workers are the strategic arm and the payments act as the rewards. Under ideal assumptions it had shown the effectiveness of bandits algorithms on such problems \citep{jain2014quality, tran2014efficient,rangi2018multi}.
\begin{figure}[H]
\includegraphics[width=7.5cm]{figs/crowdsys.png}
\caption{Online crowdsourcing system}\label{crowdsys}
\end{figure}
\section{Related Work}
The problem of multi-armed bandits (MAB) was first investigated back in 1952 while some techniques utilized were developed back in 1933 \citep{10.2307/2332286, Robbins:1952, berry,auer2002finite,lattimore_szepesvari_2020}. Since then it has been extensively explored and serves as the foundation of many modern areas, including reinforcement learning, recommendation systems, graph algorithms, etc. \citep{DBLP:conf/iconip/BouneffoufBG12,pmlr-v70-vaswani17a, pmlr-v85-durand18a,10.5555/3367243.3367445,li2020stochastic}. With the need to model the selection of multiple arms in one round, MAB is then extended to combinatorial multi-armed bandits (CMAB), which see many deployments in real applications like news/goods recommendation, medical trials, routing, and so forth \citep{Wang_Ouyang_Wang_Chen_Asamov_Chang_2017,8022964,pmlr-v84-sankararaman18a, li2020online}.
Real applications motivate the community to derive algorithms in more adaptive environments. The first of which considers adversarial bandits with the classic EXP3 algorithm. The arms are assumed to be non-stationary but non-adaptive (which means that algorithms will adapt to the adversarial) \cite{auer2002nonstochastic}. Despite that adversarial bandits do not fall into the scale of this paper's related work, it leads tremendous effort to the following topics in adaptive arms.
\paragraph{Adversarial corruptions}
The adversary is given the ability to replace the observations of the bandit algorithm with arbitrary values within some constraints.
\citet{pmlr-v99-gupta19a} discuss the constraint that rewards are modified for at most $T_0$ rounds out of the total $T$ rounds. The asymptotic regret bound $O(mT_0)$ achieved in the work is shown to match the negative result, where $m$ is the number of arms. \citet{lykouris_stochastic_2018} discussed a different constraint where the corruption $z_t$ at time $t$ is cumulative up to some constant $B$.
The regret achieved under this constrained corruption is bounded by $O(mB\log(\frac{mT}{\delta}))$ with probability at least $1 - \delta$. \citet{JMLR:v20:18-395} restricts corruptions as independent Bernoulli events with probability $b$, while if corruption happens the reward becomes arbitrary and adversarial. This problem is addressed with median statistics with gap-dependent matching regret upper and lower bounds of $O(\sum_{i \neq i^{\ast}} \frac{1}{\Delta_i}\log(\frac{K}{\Delta_i}))$, where $i^{\ast}$ denotes the optimal arm and $\Delta_i$ is the suboptimality of arm $i$.
Several lines of research also discuss corruption topics in frequentist inference and partial monitoring \citep{pmlr-v83-gajane18a}, best arm identification \citep{pmlr-v99-gupta19a}, and adversarial bandits \citep{NEURIPS2020_e655c771}.
\paragraph{Adversarial attacks}
Different from arms in adversarial corruptions who intend to maximize the regret of the bandit algorithm, arms in the adversarial attack setting have the goal to maximize their number of pulls. The setting is first considered by \citet{NEURIPS2018_85f007f8}, where it shows that the attacker may spend a budget of $O(\log T)$ to deceit UCB and $\epsilon$-greedy algorithms to pull the target arms for at least $T - o(T)$ times. Stochastic and contextual MAB algorithms also suffer from undesired loss from this adversarial attack, under a variety of attack strategies \citep{pmlr-v97-liu19e}. Linear MAB algorithms, however, are shown to be near-optimal up to some linear or quadratic terms on the attack budget $B$ \citep{garcelon_adversarial_2020,bogunovic_stochastic_2020}.
\paragraph{Strategic manipulations}
The setting of strategic manipulations further weakens the capability of the adaptive arms and prevents the bandit algorithms from being overcautious.
The objective of an arm is still utility maximization under strategic manipulations, while each arm works on its utility individually.
Instead of having a global coordinator for the adversarial attack strategy, the strategic arms seek the best response via the dominant Nash equilibrium.
The strategic manipulation setting is first studied by \citet{pmlr-v99-braverman19b} where each arm is assumed to keep a portion of the reward in each round and arms maximize the cumulative reward it keeps.
The performance of bandit algorithms will then be catastrophic under the dominant Nash equilibrium, receiving an expected cumulative reward of $0$ almost surely.
With the utility of maximizing the number of pulls for each strategic arm, common bandits algorithms are proved to be robust with guaranteed $O(\log T)$ regret bound, but only under constant strategic budgets \citep{pmlr-v119-feng20c}.
For $\omega(\log T)$ budgets, bandits under strategic manipulations remain an open problem and will be discussed in this paper.
\section{Problem Formulation}
We consider the problem of combinatorial multi-armed bandits (CMAB) under the setting of strategic arms. In strategic CMAB, each arm is given a budget and the arm can strategically increase its emitted reward signals cumulatively up to this budget \emph{for its own interest}. This problem is a Stackelberg game that involves two parties. The bandit learning algorithm is deployed first to maximize its cumulative reward under the best response of the followers. The $m$ strategic bandits arms then deplete their budget where each of them aims to maximize its expected number of pulls. Knowing the principal's strategy and other followers' budget information, the followers are assumed to place their strategies according to the dominant Nash equilibrium between them.
Formally, given a time horizon $T$, the the bandit learning algorithm is asked to pull a subset of arms, denoted by an arm subset $S_t \in \mathcal{S} $ at each time $t\in [T]$, where $\mathcal{S} = { \{0,1\}^{m}} $ is the set of all possible arm subsets. At the time $t$, based on the principal's choice of arm subset, stochastic rewards $x_{i,t}$ are generated by arm $i$ from the underlying $1$-sub-Gaussian distribution with mean $\mu_i$, independent of other arms in $S_t$. The principal does not observe these rewards. Instead, each arm can increase the emitted reward signal by an arbitrary amount $z_{i,t}\ge 0$, as long as the cumulative manipulation over the horizon does not exceed a given budget $B_i$. The principal then receives the semi-bandit feedback $\{x_{i,t}+z_{i,t}\}_{i\in S_t}$.
Let $\bm{\mu} = (\mu_1, \mu_2, ..., \mu_i) $ be the vector of expectations of all arms. The expected reward of playing any arm subset S in
any round $r_{\bm{\mu}}(S) = \mathbb{E}[R_t(S)]$,which is a function of arm subset $S$ and $\bm{\mu}$. The reward function $r_{\bm{\mu}}(S)$ is assumed to satisfy two axioms:
\begin{itemize}
\item \textbf{Monotonicity.} The expected reward of playing any arm subset $S \in \mathcal{S}$ is monotonically non-decreasing with respect to the expected reward vector. That is, if for all $i \in [m], \mu_i \leq \mu_{i}'$, then $r_{\bm{\mu}}(S) \leq r_{\bm{\mu}'}(S)$.
\item \textbf{Bounded smoothness.} There exists a strictly increasing function $f(\cdot)$, termed the bounded smoothness function, such that for any two expected reward vectors $\bm{\mu}$ and $\bm{\mu'}$ satisfying $\|\bm{\mu}-\bm{\mu'}\|_{\infty}\le \Lambda $, we have $|r_{\bm{\mu}} - r_{\bm{\mu'}}| \leq f({\Lambda})$.
\end{itemize}
These axioms cover a rich set of reward functions and the explicit forms of $R_t(S)$ and $r_{\bm{\mu}}(S)$ are not needed to be specified \citep{pmlr-v28-chen13a}.
Without loss of generality, assume that $S^\ast = \argmax_{S\subseteq \mathcal{S}} r_{\bm{\mu}}(S)$ is the unique optimal subset of arms. When placing reward manipulations, each strategic arm has access to its own history $h_{i,t} = \{I_{i,t'}, x_{i,t'}, z_{i,t'}\}_{t'\ge 1}$, where $I_{i,t'}$ is the indicator of whether arm $i$ is pulled at time $t'$ and $t' < t$. The strategy of arm $i$ is determined by a function that maps this history to a manipulation $z_{i,t}$, as $Z_{i,t} \colon h_{i,t-1} \to \mathbb{R}$. Without loss of generality, we assume that arms in the optimal arm subset $i\in S^\ast$ have strategic budgets of $0$, which restricts their $z_{i,t}$ to be $0$.
In the combinatorial setting, even with the exact reward vector $\bm\mu$ provided, it can be hard to exactly compute the optimized $r_{\bm{\mu}}(S)$. In view of this, many have studied probabilistic approximation algorithms in combinatorial problems, which indicates that an ($\alpha,\beta$)-approximation oracle defined below can be usually available.
\begin{definition}[Approximation oracle]
Let $0\le \alpha,\beta \leq 1$ and define ${OPT}_{\bm{\mu}} = \max_{S\in \mathcal{S}} r_{\bm{\mu}}(S)$.
An oracle is called an ($\alpha,\beta$)-approximation oracle if it takes an expected reward vector $\bm{\mu}$ as input and outputs an arm subset $S \in \mathcal{S}$ such that $\mathbb{P}(r_{\bm{\mu}}(S) \geq \alpha \cdot {OPT}_{\bm{\mu}}) \geq \beta$. That is, the oracle gives an arm subset $S$ that is at least as good as $\alpha$ times the reward of an optimal arm subset with probability at least $\beta$.
\end{definition}
Denote ${S_B} = \{ r_{\bm{\mu}}(S) < \alpha \cdot {OPT}_{\bm{\mu}} | S \in \mathcal{S}\}$ to be the set of suboptimal arm subsets under the approximation oracle.
Note that a suboptimal arm subset can be given by the oracle for two reasons. The ($\alpha,\beta$)-approximation oracle can fail, which happens with probability at most $1-\beta$. The estimation of $\bm{\mu}$ can deviate from the true value by a significant amount, resulting in accurate input to the oracle.
The objective of the principal is to maximize the expected cumulative reward before manipulation over the time horizon $T$. Equivalently, the principal minimizes the regret, the cumulative difference between the scaled optimal reward and expected actual reward, as defined below.
\begin{definition}[Regret]
With access to an ($\alpha,\beta$)-\textit{approximation oracle}, the regret of a combinatorial bandit algorithm for $T$ rounds is
\[
Regret_{\bm{\mu},\alpha,\beta}(T) = T \cdot \alpha \cdot \beta \cdot {OPT}_{\bm{\mu}} - \mathbb{E}\big[\sum^T_{t=1}r_{\bm{\mu}}(S_t)\big]\,,
\]
where the randomness in $\mathbb{E}[\sum^T_{t=1}r_{\bm{\mu}}(S_t)]$ involves the stochasticity of the bandit algorithm and the oracle.
\end{definition}
The objective of each strategic arm $i\in [m]\setminus S^{\ast}$, however, is to maximize the number $\sum_{t=1}^T I_{i,t}$ of times it is pulled over the time horizon. To achieve this, the arm needs to confuse the principal by deviating the emitted reward signals up to the possessed budget.
\section{Strategic Combinatorial UCB}
We now propose a variant of combinatorial upper confidence bound algorithm that is robust to strategic manipulations of rewards in Algorithm \ref{algo}. The only mild assumption we maintain is that the learning algorithm has the knowledge of the largest budget possessed among all bandits arms, i.e, $B_{max}$. This is a relaxation of the assumptions made by \citet{pmlr-v119-feng20c} on the strategic UCB algorithm, in which the learning algorithm has access to the cumulative use of budget at every time step. We start with a detailed description of the algorithm and then analyze the theoretical upper bound of regret, which enjoys $O(m \log{T} + m B_{max})$.
For each arm $i$, our algorithm maintains a counter $K_{i,t-1}$ as the total number of times arm $i$ has been pulled up to time $t-1$ and $\widetilde{\mu}_{i,t} = \frac{\sum_{j=1}^{t-1} (x_{i,j}+ z_{i,j})}{K_{i,t-1}}$ as the empirical mean estimation based on the observations. At each time step, the algorithm computes the UCB estimation $\Bar{\mu}_{i,t}= \widetilde{\mu}_{i,t} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}} + \frac{B_{max}}{K_{i,t-1}}$ for $i \in [m]$. With $\Bar{\mu}_{i,t}$, the $(\alpha, \beta)$-approximation oracle then outputs an approximately optimal arm subset $S_t \in \mathcal{S}$. The algorithm plays the return arm subset and update the counter $K_{i,t}$ and the estimation $\widetilde{\mu}_{i,t}$ accordingly.
\begin{algorithm}[htb]
\DontPrintSemicolon
{\bf Input:}
Horizon $T$, number $m$ of arms, maximum budget $B_{max}$\;
{\bf Output:} Arm subset $S_t$\;
Initialize $K_{i,0} = 0$ and $\widetilde{\mu}_{i,0} = 0$ for $i \in [m]$\;
\For{t = 1 $\to$ m}{
Play an arbitrary arm subset $S_t \in \mathcal{S}$ such that $t \in S_t$\;
$K_{i,t} = K_{i,t-1} + 1$ for $i \in S_t$\;
$\widetilde{\mu}_j = x_{i,t}+z_{i,t}$ for $i \in S_t$ \;}
\For{t = m+1 $\to$ T}{
For $i \in [m]$, compute $\Bar{\mu}_{i,t}= \widetilde{\mu}_{i,t-1} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}} + \frac{B_{max}}{K_{i,t-1}} $\;
$S_t$ = {\textbf{oracle}}$(\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$\;
Play arm subset $S_t$ and update
$K_{i,t} = K_{i,t-1} + 1$ and $ \widetilde{\mu}_{i,t} = \frac{\sum_{j=1}^{t} (x_{i,j} + z_{i,j})}{K_{i,t}}$ for $i \in S_t$ \;
}
\caption{SCUCB}
\label{algo}
\end{algorithm}
We first introduce a few notations that are used in our results. Define, for any arm $i\in[m]$, the suboptimality gaps as
\begin{align*}
&\Delta^i_{min} = \alpha \cdot {OPT}_{\bm{\mu}} - \max\{r_{\bm{\mu}}(S) \mid S \in S_B, i \in S\}\,, \notag \\
&\Delta^i_{max} = \alpha \cdot {OPT}_{\bm{\mu}} - \min\{r_{\bm{\mu}}(S) \mid S \in S_B, i \in S\}\,.
\end{align*}
We then denote the maximum and minimum of the suboptimality gaps as $\Delta_{max} = \max_{i \in [m]} \Delta^i_{max}\notag $ and $\Delta_{min} = \min_{i \in [m]} \Delta^i_{min}$.
The following lemma re-establish the canonical tail bound inequality in UCB under the setting of strategic manipulations.
\begin{lemma}\label{lambda}
Let ${\Lambda}_{i,t} = \sqrt{\frac{3 \log{t} }{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}$, where $\rho_{i,t-1}$ is the total strategic budget spent by arm $i$ up to time $t-1$ and $K_{i,t-1}$ be the total number of pulls of arm $i$ up to time $t-1$. Define the event $E_t = \{ |\ \widetilde{\mu}_{i,t-1} - \mu_{i}| \leq \Lambda_{i,t}, \forall i \in [m]\}$, where $\mu_i$ is the true mean of arm $i$'s underlying distribution.
Then, $\mathbb{P}(\lnot E_t)\leq 2m \cdot t^{-2}$.
\end{lemma}
Armed with Lemma \ref{lambda}, we present one of our main theorems, Theorem \ref{appendixA}, which gives the regret bound of $O(\log T)$ of SCUCB. The outline of the proof follows that of CUCB by \citet{pmlr-v28-chen13a}. To complete the proof, we carefully choose $\psi_t$, which controls the trade-off between the exploration and exploitation periods.
\begin{theorem}
\label{appendixA}
The regret of the SCUCB algorithm with $m$ strategic arms in time horizon $T$ using an $(\alpha,\beta)$-approximation oracle is at most
\begin{align}
m\cdot\Delta_{max}\left( \frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}+1 \right)\,, \notag \nonumber
\end{align}
where $f^{-1}(\cdot)$ is the inverse bounded smoothness function.
\end{theorem}
\begin{proof}[Proof sketch]
We introduce a counter $N_i$ for each arm $i\in[m]$ after the $m$-round initialization and let $N_{i,t}$ be the value of $N_i$ at time $t$. We initialize $N_{i,m} = 1$. By definition, $\sum_{i \in [m]} N_{i,m} = m$. For $t > m$, the counter $N_{i,t}$ is updated as follows:
\begin{itemize}
\item If $S_t \in S_B$, then $N_{{i'},t} = N_{{i'},t-1} + 1$ where $i' =$ $\argmin_{i \in S_t}$ $N_{i,t-1}$. In the case that $i'$ is not unique, we break ties arbitrarily;
\item If $S_t \notin S_B$, then no counters will be updated.
\end{itemize}
As such, the total number of pulls of suboptimal arm subsets is less than or equal to $\sum_{i=1}^m N_{i,T}$.
Define $\psi_t = \frac{8B_{max}f^{-1}(\Delta_{min}) + 6\log{t}} {\left(f^{-1}(\Delta_{min})\right)^2} > c $, where $c$ is the larger solution of
\begin{align*}
\label{square}
&(f^{-1}(\Delta_{min}))^2 c^2 + 16B_{max}^2 - ((8B_{max} f^{-1}(\Delta_{min})- 6\log{t}) c = 0 \,.
\end{align*}
Then, we decompose the total number $\sum_{i=1}^m N_{i,T}$ of pulls of suboptimal arm subsets as
\begin{align}
\sum_{i=1}^m N_{i,T} = \ & m +\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B\} \notag \\
=\ &m +\sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} \leq \psi_t\} \notag \\
&+ \sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} > \psi_t\}\notag \\
\leq \ & m + m\psi_T+\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}\,. \notag
\end{align}
The inequality follows as $\sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} \leq \psi_t\}$ can be trivially bounded by $m\psi_T$. Thus the key to bound the total number of pulls of suboptimal arm subset is to upper bound $\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}$. Let $F_t$ denotes the event that the oracle fails to provide an $\alpha$-approximate arm subset with respect to the input vector $\Bar{\bm{\mu}} = (\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$.
Then we can decompose $\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}$ as
\begin{align}
&\quad \sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}\notag\\
&\leq \sum_{t=m+1}^T (\mathbb{I}\{F_t\} + \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\})\notag \\
&\leq (T-m)(1-\beta) + \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\}\notag \,. \notag
\end{align}
By leveraging the monotonicity and smoothness assumptions of the reward function, we show that $\mathbb{P}(\{\lnot F_t, E_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\}) = 0$. Therefore, by the inclusion-exclusion principle,
\begin{align}
\mathbb{P}(\{\lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\})
&\leq \mathbb{P}(\lnot E_t) \leq 2m \cdot t^{-2} \,.\notag
\end{align}
Leveraging the upper bound of total number of pulls of suboptimal arm subsets, the regret is bounded by
\begin{align}
& \quad Regret_{\bm{\mu},\alpha,\beta}(T)
\notag \\
& \leq T\alpha \beta \cdot \text{OPT}_{\bm{\mu}} - \left( T\alpha\cdot \text{OPT}_{\bm{\mu}} - \mathbb{E} \left[ \sum_{i=1}^m N_{i,T} \right]\cdot \Delta_{max} \right)
\notag \\
&\leq m\cdot \Delta_{max}\left(\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}+1\right)\,. \notag \qedhere
\end{align}
\end{proof}
It remains in question whether a bandit algorithm can achieve a regret upper bound sublinear in $B_{max}$.
Our conjecture is negative.
In fact, under strategic manipulations of rewards, the design of robust bandit algorithms, e.g. UCB and $\epsilon$-greedy, is analogous to the design of outlier-robust mean estimation algorithms. Existing works on robust mean estimation, such as \citet{steinhardt_et_al:LIPIcs:2018:8368}, argue that from an information theoretical point of view mean estimation error must be depending on the variance of the data. Casting this argument to bandit with the strategic manipulation setting, we believe that a tight regret bound is unlikely to be independent of the strategic budget $B_{max}$. Moreover, Proposition 4 in \citet{steinhardt_et_al:LIPIcs:2018:8368} implies that the dependency on $B_{max}$ is linear. This corresponds to our linear dependency of regret in Theorem \ref{appendixA}.
\section{Lower Bounds for strategic budget}
To fully understand the effects of strategic manipulations, we investigate the relationship between the strategic budget and the performance of UCB-based algorithms. Our results provide the dependency between an arm's strategic budget and the number of times it is pulled, which influence the regret of the algorithm incurred by the arm. Our analysis gains some insight from \citet{zuo_near_2020}, which limits the discussion to the 2-armed bandit setting and cannot be applied to the general MAB setting directly.
We first define some notations used in the theorem.
Without loss of generality, let arm $i^{\ast}$ be the optimal arm and arm $i, i \neq i^{\ast}$ be an arbitrary strategic suboptimal arm. Assume that arm $i$ has no access to the information regarding other arms. Let $K_{i,t}$ and $K_{i^{\ast},t}$ denote the number of times arm $i$ and arm $i^{\ast}$ have been pulled up to time $t$, respectively. Denote $\hat{\mu}_{i,t}$ as the empirical estimate of the underlying mean $\mu$ without manipulations, i.e., $\hat{\mu}_{i,t} = \frac{\sum^{t-1}_{j=1}x_{i,j}}{K_{i,t-1}}$. Define the suboptimality gap for each strategic arm $i$ to be $\delta_i = \mu_{i^{\ast}} - \mu_i$. We use a slightly revised UCB algorithm as the basic algorithm, where the UCB estimation term for arm $i$ is $\hat{\mu}_{i,t} + \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}}$ and $\eta$ is a confidence parameter chosen by the algorithm.
\begin{theorem}\label{lower}
In stochastic multi-armed bandit problems, for a strategic suboptimal arm $i$ without access to other arms' information, to be pulled for $\omega(k)$ in $T$ steps under the UCB algorithm where $k\ge O(\log{T})$, the minimum strategic budget is $\omega(k)$.
\end{theorem}
This dependency of $k$ can be extended to CMAB and CUCB straightforwardly when arms within an arm subset collude. Counter-intuitively, the dependency between the number of pulls of a strategic arm and its strategic budget is linear and subsequently, this infers that for a strategic arm to manipulate the algorithm to suffer an undesired regret of $\omega(k)$, where $k\ge O(\log T)$, the strategic budget must be at least $\omega(k)$.
\section{Experiments}
In this section, we evaluate our SCUCB algorithm empirically on synthetic data and three real applications, namely online worker selection in crowdsourcing, online recommendation, and online influence maximization. We highlight the best performance among all algorithms with bold text.
\subsection{Baseline Algorithms}
We compare our proposed SCUCB algorithm with both stochastic and adversarial bandits algorithms that achieves optimal asymptotic regret in their settings.
\begin{enumerate}
\allowdisplaybreaks
\item CUCB \cite{pmlr-v28-chen13a}. CUCB is the naive counterpart of our algorithm. The algorithm calculates an upper confidence interval for each arm and picks the best arm subset with the highest upper confidence interval.
\item TSCB \cite{wang2018thompson}. TSCB is the combinatorial version of the classical Thompson sampling algorithm. The algorithm maintains a prior beta distribution estimation for each arm and updates according to the received reward. At each time, the algorithm samples from the estimated distributions and pick actions according to the highest sample.
\item Combinatorial variant of EXP3 \cite{auer_nonstochastic_2002}. The algorithm maintains a weight for each arm and draws actions according to the normalized weight distribution. Upon receiving rewards, the algorithm update weight according to the classical EXP3 update rule.
\end{enumerate}
\subsection{Synthetic Experiments}
We conduct experiments presented in this section with synthetic data and compare our proposed algorithm with its naive counterpart. The approximation oracle is designed to succeed with probability $1$. Each bandit arm is modeled to follow a Bernoulli distribution with randomly populated $\mu \in [0,1]$ and all arms adapt LSI strategy.
The arms in the optimal arm subset have a strategic budget of $0$ since an additional strategic budget for optimal arms would only boost the performance of our algorithm. All other arms are equipped with a randomly allocated budget $B$, $ 0 \leq B \leq B_{max}$ by definition.
As is typical in the bandit literature, for example in \citet{auer2002finite}, we evaluate both the naive and strategic CUCB algorithms on their tuned versions, where the UCB exploration parameter is scaled by a constant factor $\gamma \in [0,1]$. To ensure reproducible results, each experiment is repeated for $10$ random seeds and the averaged result is presented.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 171.74 & 187.51 & 259.04 & 256.66 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{143.85} & \textbf{172.57} & \textbf{208.53} & \textbf{233.57} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 434.82 & 492.02 & 520.97 & 549.57 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{301.57} & \textbf{365.94} & \textbf{450.10} & \textbf{505.453} \\ \bottomrule
\end{tabularx}
\caption{Cumulative regret achieved by CUCB and SCUCB with synthetic data with various action size.}\label{table:1}
\end{table}
The first set of experiment is conducted with $K = \{10, 20\}$ arms through a time horizon of $T=5000$ time steps with maximum possible budget of $B_{max} = \{70, 90, 110, 130\}$. The algorithms are asked to select $\text{Action size} = 2$ arms as an arm subset at each time step. The cumulative regret incurred by CUCB and SCUCB algorithm is presented in the table \ref{table:1} where the best performance is highlighted in bold text. Clearly, SCUCB demonstrated its effectiveness as it achieves significantly smaller regrets in various possible maximum strategic budgets.
The next two tables reveals the advantage of SCUCB algorithm with various size of action set. The experiments are conducted with time horizon of $T=5000$ time steps with $K = \{10, 20\}$ arms and maximum possible budget of $B_{max} = \{50\}$. The algorithms are asked to select $\text{Action size} = \{2, 4, 6, 8\}$ arms as an arm subset at each time step. Once again, our SCUCB outperforms its naive counterpart by achieving much smaller cumulative regret,
the best numerical results across algorithms are highlighted in bold text.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 140.21 & 150.88 & 158.22 & 123.91 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{105.58} & \textbf{113.63} & \textbf{103.42} & \textbf{88.72} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 302.10 & 316.09 & 340.52 & 328.80 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{232.23} & \textbf{247.32} & \textbf{268.27} & \textbf{306.77} \\ \bottomrule
\end{tabularx}
\caption{Cumulative regret achieved by UCB and SCUCB with synthetic data with various $B_{max}$.}
\end{table}
To compare the two algorithms in detail, we plot the cumulative regret incurred by algorithms against the time steps. It becomes apparent that the SCUCB algorithm features a cumulative regret versus time steps line that is much smaller than the one caused by CUCB even under a higher maximum possible strategic budget. We compare the two algorithm under settings where the maximize possible budget is $B_{max} = 70, 90$ and $B_{max} = 110, 130$.
\begin{figure}[H]
\centering
\subfigure[]{
\label{Fig1}
\includegraphics[width=3.8cm]{figs/B7090_K20_N10000.png}}
\subfigure[]{
\label{Fig2}
\includegraphics[width=3.8cm]{figs/B110130_K20_N10000.png}}
\caption{\ref{Fig1} Comparison of CUCB and SCUCB with $B_{max} = 70, 90$. \ref{Fig2} Comparison of CUCB and SCUCB with $B_{max} = 110, 130$.}
\end{figure}
\subsection{Online Worker Selection in Crowdsourcing Systems}
We simulated an online crowdsourcing system that has a workflow resembled by figure \ref{crowdsys} and model it as a combinatorial bandits problem. We performed an extensive empirical analysis of our SCUCB algorithm against CUCB, the combinatorial version of Thompson sampling, and EXP3. The experiments were conducted with the Amazon sentiment dataset, where Amazon item reviews are labeled 'is/is not book' or 'is/is not negative \cite{Krivosheev:2018:CCM:3290265.3274366}. We split the dataset into two, where 'is book' contains data that only has labels 'is/is not book' and 'is negative' contains data that only has labels 'is/is not negative'. Both datasets consists of 7803 reviews, 284 workers, and 1011 tasks.
The correctness of workers and tasks can be visualized by the figures as below. Notice that most of the workers attain an accuracy of $60$ percent and above. This highlights that most of the workers we are interacting with has the ability to label most of the tasks correctly.
\begin{figure}[H]
\centering
\subfigure[]{
\label{worker}
\includegraphics[width=3.8cm]{figs/worker_visual.png}}
\subfigure[]{
\label{task}
\includegraphics[scale = 0.28]{figs/task_visual.png}}
\caption{\ref{worker} Visualization of correctness of worker's labels. \ref{task}Visualization of correctness of task's labels.}
\end{figure}
For each task, the data contains responses from 5 workers and the crowdsourcing platform is asked to select 2, 3, or 4 workers at each time step. To measure the performance of the algorithms, we made the assumption that the reward is monotonically correlated to the accuracy of the label compared to the ground truth label. Thus, we define the reward to be 1, if the worker's label is the same as ground truth label and 0 otherwise. To model the strategic behavior of the workers, we model each worker with a randomly allocated strategic level $s \in [0,1]$ such that the worker provides its honest response with probability $s$.
For the first set of experiments, we choose the maximum possible strategic budget to be $B_{max} = 500$. For actions size of $2, 3, 4$. We obtain the following cumulative rewards and cumulative regret results where the best performance is highlighted in bold text. To ensure reproducible results, each experiment is repeated 5 times and the averaged result is presented.
To further investigate the effect of the maximum possible strategic budget on the performance of algorithms, we fix the size of action size to be 2 and plot the cumulative reward/cumulative regret incurred by each algorithm again $B_{max} \in [100, 700]$. Regardless of the $B_{max}$ value, our proposed SCUCB algorithm consistently outperforms other methods. For better visualization, we omit the line for combinatorial Thompson sampling due to its relatively weaker performance.
\begin{figure}[H]
\centering
\subfigure[]{
\label{culre}
\includegraphics[width=3.8cm]{figs/culre.png}}
\subfigure[]{
\label{culreg}
\includegraphics[width=3.8cm]{figs/reg.png}}
\caption{\ref{culre} Cumulative rewards attained by algorithms with various level of $B_{max}$. \ref{culreg} Cumulative regrets attained by algorithms with various level of $B_{max}$. }
\end{figure}
\begin{table}[H]
\label{first}
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31672.6 & 46181.2 & 68244.6 \\
\multicolumn{1}{l|}{TSCB} & 19853.0 & 34226.2 & 60621.2 \\
\multicolumn{1}{l|}{EXP3} & 31569.6 & 46570.6 & 68482.2 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{32082.6} & \textbf{46623.6} & \textbf{68524.6} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 27787.4 & 42198.8 & 48125.4 \\
\multicolumn{1}{l|}{TSCB} & 39607.0 & 54153.8 & 55748.8 \\
\multicolumn{1}{l|}{EXP3} & 27890.4 & 41809.4 & 47887.8 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{27377.4} & \textbf{41756.4} & \textbf{47845.4} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31108.8 & 44757.6 & 59424.8 \\
\multicolumn{1}{l|}{TSCB} & 17240.6 & 32010.4 & 51109.0 \\
\multicolumn{1}{l|}{EXP3} & 30879.0 & 44767.2 & 58207.8 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{31759.0} & \textbf{44843.4} & \textbf{59441.4} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 28411.2 & 43566.2 & 56465.2 \\
\multicolumn{1}{l|}{TSCB} & 42279.4 & 56279.6 & 64781.0 \\
\multicolumn{1}{l|}{EXP3} & 28641.0 & 43523.4 & 57682.2 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{27761.0} & \textbf{43446.6} & \textbf{56448.6} \\ \bottomrule
\end{tabularx}
\caption{Cumulative rewards and cumulative regret attained by SCUCB and baseline algorithms with various action sizes on isbook dataset and isnegative dataset.}
\end{table}
The results of our crowdsourcing experiments reveal the robustness of our SCUCB algorithm under strategic manipulations. It also indicates the performance of adversarial and stochastic bandits algorithms under strategic manipulations. The combinatorial variant of Thompson sampling is vulnerable under manipulations, which agrees with our expectation given that the algorithm is a Bayesian algorithm in nature and is heavily relying on estimating the underlying distribution of arms. Its estimation can be easily perturbed under strategic manipulation, and thus leads to undesired performance results. It is also expected that the SCUCB algorithm far outperforms its naive counterpart, CUCB algorithm. The combinatorial version of EXP3 algorithm is the most competitive algorithm with our SCUCB. As the EXP3 algorithm was originally designed for a pure adversarial setting, the algorithm has some robustness under strategic manipulations. Shown in the Figure \ref{culre} and Figure \ref{culreg}, the combinatorial EXP3 algorithm consistently shows robustness across various levels of maximum possible strategic budget. However, as our SCUCB algorithm considers the maximum possible strategic budget and thus more adaptive, SCUCB far outperforms EXP3CB as $B_{max}$ increases.
\subsection{Online Recommendation System}
The online recommendation is a classic example of combinatorial bandits in applications. We evaluated algorithms in the latest MovieLens dataset which contains 9742 movies and 100837 ratings of the movies with each rating being $1 - 5$ \cite{10.1145/2827872}. We compare the algorithms based on the total values of ratings received, i.e. recommending a movie and receiving a rating of 5 is more desired than recommending a movie and receiving a rating of 1.
As the dataset may consist of unbalanced data for each movie, we adapted collaborative filtering and $k$-means clustering into our evaluation system. We used the collaborative filtered for training and testing and instead of recommending movies, we clustered the movies by $k$-means clustering and asked the bandits algorithm to choose one cluster at each time. The movie with the highest rating in the chosen cluster is recommended. The workflow of the experiment setup is summarized in the above diagram.
\begin{figure}[H]
\centering
\includegraphics[width=7cm]{figs/arch.png}
\caption{Experiment set up of online recommendation.}
\end{figure}
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 10 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2297.60 & 2494.36 & 2631.02 \\
\multicolumn{1}{l|}{TS} & 2160.14 & 2314.76 & 2468.26 \\
\multicolumn{1}{l|}{EXP3} & 2181.11 & 2449.28 & 2430.96 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2305.75} & \textbf{2314.76} & \textbf{2636.07} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 20 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2469.93 & 2336.92 & 2159.25 \\
\multicolumn{1}{l|}{TS} & 2260.20 & 1850.55 & 1905.42 \\
\multicolumn{1}{l|}{EXP3} & 2380.74 & 2317.82 & 2025.39 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2474.69} & \textbf{2341.19} & \textbf{2177.70} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 30 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\\midrule
\multicolumn{1}{l|}{UCB} & 2443.57 & 2393.88 & 2472.31 \\
\multicolumn{1}{l|}{TS} & 2132.16 & 2248.56 & 1855.14 \\
\multicolumn{1}{l|}{EXP3} & 2368.42 & 2265.32 & 2339.49 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2436.43} & \textbf{2397.72} & \textbf{2476.52} \\ \bottomrule
\end{tabularx}
\caption{Cumulative rewards attained by algorithms with different number of clusters $10, 20, 30$ and different level of $\{B_{max} = 30, 50, 70\}$ across 500 time steps. }
\end{table}
We performed the evaluation with various number of clusters (10, 20 or 30) and $B_{max} = 30, 50, 70$ through a time horizon of $T = 500$. The results presented below are averages over 5 runs to ensure reproducibility. The best performance across algorithms is highlighted in bold text. From the tables below, we conclude the effectiveness of SCUCB algorithm.
\subsection{Online Influence Maximization}
We implemented the online influence maximization with an offline influence maximization algorithm TIM as our oracle \cite{10.1145/2588555.2593670}. TIM is one of the offline influence maximization algorithms that achieve asymptotic optimality. We perform the experiments on two datasets. One is a simulation dataset with 16 nodes and 44 edges, the other is a 100 nodes subset from Digg dataset, where each node represents a user from Digg website \cite{nr}. We simulate the connectivity of edges at each time step by randomly assigning each edge a connectivity probability at the start of the experiment.
\begin{figure}[hbt!]
\centering
\subfigure[]{
\label{test_graph}
\includegraphics[width=3.6cm]{figs/test_graph.png}}
\subfigure[]{
\label{16c2}
\includegraphics[width=3.6cm]{figs/test.png}}
\caption{\ref{test_graph} Visualization of synthetic graph for evaluation. \ref{16c2} Mean rewards attained by algorithms.}
\end{figure}
Figure \ref{test_graph} visualizes the synthetic graph we created for evaluation. We then tested the algorithms over $T = 800$ time steps with action size of 2 and maximum possible strategic budget of $B_{max} = 300$. Figure \ref{16c2} shows the effectiveness of algorithms by evaluating based on the averaged nodes influenced. To ensure reproducible results, we calculate the mean of influence spread over 100 trials. For better visualization, we kept a running mean of over 50 time steps to smooth the lines.
We then investigate the effect of action size and the results are summarized in the following tables \ref{table:5}.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYYY@{}}
\toprule
\multicolumn{5}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Synthetic, 16 nodes, 44 edges} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 4 & 6 & 8 \\ \midrule
\multicolumn{1}{c|}{CUCB} & 5.46 & 6.54 & 10.4 & 12.5 \\
\multicolumn{1}{c|}{TSCB} & 6.7 & 8.50 & 10.44 & 12.7 \\
\multicolumn{1}{c|}{EXP3CB} & 5.86 & 8.14 & 10.34 & 12.46 \\
\multicolumn{1}{c|}{SCUCB} & \textbf{7.44} & \textbf{9.32} & \textbf{11.16} & \textbf{13.04} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{digg, 100 nodes} \\ \midrule
\multicolumn{1}{l|}{Action size} & 10 & 15 & 20 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 15.64 & 20.64 & 25.32 \\
\multicolumn{1}{l|}{TSCB} & 16.30 & 21.62 & 26.08 \\
\multicolumn{1}{l|}{EXP3} & 12.16 & 17.76 & 21.72 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{16.56} & \textbf{22.16} & \textbf{26.14} \\ \bottomrule
\end{tabularx}
\caption{Averaged final rewards attained by algorithms with synthetic and Digg dataset with action size of 2, 800 time steps and $B_{max} = 200$.}\label{table:5}
\end{table}
\section{Conclusion}
We investigate the problem of combinatorial MAB under strategic manipulations of rewards. We propose a variant of the UCB algorithm, SCUCB, which attains a regret at most $O(m \log T + m B_{max})$, with a lower bounds on the strategic budget for a malicious arm to incur a $\omega(\log T)$ regret of the bandit algorithm.
Compared to previous studies on the bandit problems under strategic manipulations, we relax on the assumption that the algorithm has access to the cumulative strategic budget spent by each arm at each time step.
For the robustness of bandit algorithms, we present lower bounds on the strategic budget for a malicious arm to incur a $\omega(\log T)$ regret of the bandit algorithm.
We provide extensive empirical results on both synthetic and real datasets with a range of applications to verify the effectiveness of the proposed algorithm. Our algorithm consistently outperforms baseline algorithms that were designed for stochastic and adversarial settings.
\subsection{Baseline Algorithms}
We compare our proposed SCUCB algorithm with both stochastic and adversarial bandits algorithms that achieves optimal asymptotic regret in their settings.
\begin{enumerate}
\item CUCB \cite{pmlr-v28-chen13a}. CUCB is the naive counterpart of our algorithm. The algorithm calculates a upper confidence interval for each arm and pick the best arm subset with the highest upper confidence interval.
\item TSCB \cite{wang2018thompson}. TSCB is the combinatorial version of the classical Thompson sampling algorithm. The algorithm maintain a prior beta distribution estimation for each arm and updates according to the received reward. At each time, the algorithm samples from the estimated distributions and pick actions according to the highest sample.
\item Combinatorial variant of EXP3 \cite{auer_nonstochastic_2002}. The algorithm maintains a weight for each arm and draw actions according to the normalized weight distribution. Upon receiving rewards, the algorithm update weight according to the classical EXP3 update rule.
\end{enumerate}
\subsection{Synthetic Experiments}
We conduct experiments presented in this section with synthetic data and compare our proposed algorithm with its naive counter part. The approximation oracle is designed to succeed with probability $1$. Each bandit arm is modeled to follow a Bernoulli distribution with randomly populated $\mu \in [0,1]$ and all arms adapt LSI strategy.
The arms in the optimal arm subset have a strategic budget of $0$ since additional strategic budget for optimal arms would only boost the performance of our algorithm. All other arms are equipped with a randomly allocated budget $B$,$ 0 \leq B \leq B_{max}$ by definition.
As is typical in the bandit literature, for example in \citet{auer2002finite}, we evaluate both the naive and strategic CUCB algorithms on their tuned versions, where the UCB exploration parameter is scaled by a constant factor $\gamma \in [0,1]$. To ensure reproducible result, each experiment is repeated for $10$ random seeds and the averaged result is presented.
\begin{table}[H]
\centering
\label{table:1}
\begin{tabular}{@{}lcccc@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 171.74 & 187.51 & 259.04 & 256.66 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{143.85} & \textbf{172.57} & \textbf{208.53} & \textbf{233.57} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lcccc@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 434.82 & 492.02 & 520.97 & 549.57 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{301.57} & \textbf{365.94} & \textbf{450.10} & \textbf{505.453} \\ \bottomrule
\end{tabular}
\caption{Cumulative regret achieved by UCB and SCUCB with synthetic data with various action size.}
\end{table}
The first set of experiment is conducted with $K = \{10, 20\}$ arms through a time horizon of $T=5000$ time steps with maximum possible budget of $B_{max} = \{70, 90, 110, 130\}$. The algorithms are asked to select $\text{Action size} = 2$ arms as an arm subset at each time step. The cumulative regret incurred by CUCB and SCUCB algorithm are presented in the table \ref{table:1} where the performance of our proposed SCUCB algorithm is highlighted in bold text. Clearly, SCUCB demonstrated its effectiveness as it achieves significantly smaller regrets in various possible maximum strategic budget.
The next two tables reveals the advantage of SCUCB algorithm with various size of action set. The experiments are conducted with time horizon of $T=5000$ time steps with $K = \{10, 20\}$ arms and maximum possible budget of $B_{max} = \{50\}$. The algorithms are asked to select $\text{Action size} = \{2, 4, 6, 8\}$ arms as an arm subset at each time step. Once again, our SCUCB outperforms its naive counterpart by achieving much smaller cumulative regret, for which numerical results are highlighted in bold text.
\begin{table}[H]
\centering
\begin{tabular}{@{}lcccc@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 140.21 & 150.88 & 158.22 & 123.91 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{105.58} & \textbf{113.63} & \textbf{103.42} & \textbf{88.72} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lcccc@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 302.10 & 316.09 & 340.52 & 328.80 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{232.23} & \textbf{247.32} & \textbf{268.27} & \textbf{306.77} \\ \bottomrule
\end{tabular}
\caption{Cumulative regret achieved by UCB and SCUCB with synthetic data with various $B_{max}$.}
\end{table}
To compare the two algorithms in details, we plot the cumulative regret incurred by algorithms against the time steps. It becomes apparent that the SCUCB algorithm features a cumulative regret versus time steps line that is much smaller than the one caused by CUCB even under a higher maximum possible strategic budget. We compare the two algorithm under settings where the maximize possible budget is $B_{max} = 70, 90$ and $B_{max} = 110, 130$.
\begin{figure}[H]
\centering
\subfigure[]{
\label{Fig1}
\includegraphics[width=3.8cm]{figs/B7090_K20_N10000.png}}
\subfigure[]{
\label{Fig2}
\includegraphics[width=3.8cm]{figs/B110130_K20_N10000.png}}
\caption{\ref{Fig1} Comparison of CUCB and SCUCB with $B_{max} = 70, 90$. \ref{Fig2} Comparison of CUCB and SCUCB with $B_{max} = 110, 130$.}
\end{figure}
\subsection{Online Worker Selection in Crowdsourcing Systems}
We simulated a online crowdsourcing system that has a workflow resembled by \ref{crowdsys} and model it as a combinatorial bandits problem. We performed extensive empirical analysis of our SCUCB algorithm against CUCB, the combinatorial version of Thompson sampling and EXP3. The experiments were conducted with the Amazon sentiment dataset, where Amazon items reviews are labeled 'is/is not book' or 'is/is not negative \cite{Krivosheev:2018:CCM:3290265.3274366}. We split the dataset into two, where 'is book' contains data that only has labels 'is/is not book' and 'is negative' contains data that only has labels 'is/is not negative'. Both dataset consists of 7803 reviews, 284 workers, and 1011 tasks.
The correctness of workers and tasks can be visualized by the figures as below.
\begin{figure}[H]
\centering
\subfigure[]{
\label{worker}
\includegraphics[width=3.8cm]{figs/worker_visual.png}}
\subfigure[]{
\label{task}
\includegraphics[scale = 0.28]{figs/task_visual.png}}
\caption{\ref{worker} Visualization of correctness of worker's labels. \ref{task}Visualization of correctness of task's labels.}
\end{figure}
For each task, the data contains responses from 5 workers and the crowdsourcing platform is asked to select 2, 3, or 4 workers at each time step. To measure the performance of the algorithms, we made the assumption that the reward is monotonically correlated to the accuracy of label compared to ground truth label. Thus, we define the reward to be 1, if the worker's label is the same with ground truth label and 0 otherwise. To model the strategic behavior of the workers, we model each worker with a randomly allocated strategic level $s \in [0,1]$ such that the worker provides its honest response with probability $s$.
\begin{table}[H]
\label{first}
\centering
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31672.6 & 46181.2 & 68244.6 \\
\multicolumn{1}{l|}{TSCB} & 19853.0 & 34226.2 & 60621.2 \\
\multicolumn{1}{l|}{EXP3} & 31569.6 & 46570.6 & 68482.2 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{32082.6} & \textbf{46623.6} & \textbf{68524.6} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 27787.4 & 42198.8 & 48125.4 \\
\multicolumn{1}{l|}{TSCB} & 39607.0 & 54153.8 & 55748.8 \\
\multicolumn{1}{l|}{EXP3} & 27890.4 & 41809.4 & 47887.8 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{27377.4} & \textbf{41756.4} & \textbf{47845.4} \\ \bottomrule
\end{tabular}
\end{table}
\begin{table}[H]\centering
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31108.8 & 44757.6 & 59424.8 \\
\multicolumn{1}{l|}{TSCB} & 17240.6 & 32010.4 & 51109.0 \\
\multicolumn{1}{l|}{EXP3} & 30879.0 & 44767.2 & 58207.8 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{31759.0} & \textbf{44843.4} & \textbf{59441.4} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 28411.2 & 43566.2 & 56465.2 \\
\multicolumn{1}{l|}{TSCB} & 42279.4 & 56279.6 & 64781.0 \\
\multicolumn{1}{l|}{EXP3} & 28641.0 & 43523.4 & 57682.2 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{27761.0} & \textbf{43446.6} & \textbf{56448.6} \\ \bottomrule
\end{tabular}
\caption{Cumulative rewards and cumulative regret attained by SCUCB and baseline algorithms with various action size on isbook dataset and isnegative datase.}
\end{table}
For the first set of experiments, we choose the maximum possible strategic budget to be $B_{max} = 500$. For actions size of $2, 3, 4$. We obtain the following cumulative rewards and cumulative regret results where SCUCB's performance is highlight in bold text. To ensure reproducible results, each experiment is repeated for 5 times and the averaged result is presented.
To further investigate the effect of maximum possible strategic budget on the performance of algorithms, we fix the size of action size to be 2 and plot the cumulative reward/cumulative regret incurred by each algorithm again $B_{max} \in [100, 700]$. Regardless of the $B_{max}$ value, our proposed SCUCB algorithm consistently outperforms other methods. For better visualization, we omit the line for combinatorial Thompson sampling due to its relatively weaker performance.
\begin{figure}[H]
\centering
\subfigure[]{
\label{culre}
\includegraphics[width=3.8cm]{figs/culre.png}}
\subfigure[]{
\label{culreg}
\includegraphics[width=3.8cm]{figs/reg.png}}
\caption{\ref{culre} Cumulative rewards attained by algorithms with various level of $B_{max}$. \ref{culreg} Cumulative regrets attained by algorithms with various level of $B_{max}$. }
\end{figure}
The results of our crowdsourcing experiments reveals the robustness of our SCUCB algorithm under strategic manipulations. It also indicates the performance of adversarial and stochastic bandits algorithms under strategic manipulations. The combinatorial variant of Thompson sampling is vulnerable under manipulations, which agrees with our expectation given that the algorithm is a Bayesian algorithm in nature and is heavily relying on estimating the underlying distribution of arms. Its estimation can be easily perturbed under strategic manipulation, and thus leads to undesired performance results. It is also expected that the SCUCB algorithm far outperform its naive counterpart, CUCB algorithm. The combinatorial version of EXP3 algorithm is the most competitive algorithm with our SCUCB. As the EXP3 algorithm was originally designed for pure adversarial setting, the algorithm has some robustness under strategic manipulations. Shown in the Figure \ref{culrereg}, the combinatorial EXP3 algorithm consistently show robustness across various level of maximum possible strategic budget. However, as our SCUCB algorithm considers the maximum possible strategic budget and thus more adaptive, SCUCB far outperforms EXP3CB as $B_{max}$ increases.
\subsection{Online Recommendation System}
Online recommendation is a classic example of combinatorial bandits in applications. We evaluated algorithms in the latest MovieLens dataset which contains 9742 movies and 100837 rating of the movies with each rating being $1 - 5$ \cite{10.1145/2827872}. We compare the algorithms based on the total values of ratings received, i.e. recommending a movie and receiving a rating of 5 is more desired than recommending a movie and receiving a rating of 1.
\begin{figure}[H]
\centering
\includegraphics[width=7cm]{figs/arch.png}
\caption{Experiment set up of online recommendation.}
\end{figure}
As the dataset may consists of unbalanced data for each movie, we adapted collaborative filtering and $k$-means clustering into our evaluation system. We used the collaborative filtered for training and testing and instead of recommending movies, we clustered the movies by $k$-means clustering and asked bandits algorithm to choose one cluster at each time. The movie with the highest rating in the chosen cluster is recommended. The workflow of the experiment set up is summarized in above diagram.
We performed the evaluation with various number of clusters (10, 20 or 30) and $B_{max} = 30, 50, 70$ through a time horizon of $T = 500$. The results presented below are averages over 5 runs to ensure reproducibility. From the tables below, we concludes the effectiveness of SCUCB algorithm, which performance is highlighted in bold text.
\begin{table}[H]
\centering
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 10 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2297.60 & 2494.36 & 2631.02 \\
\multicolumn{1}{l|}{TS} & 2160.14 & 2314.76 & 2468.26 \\
\multicolumn{1}{l|}{EXP3} & 2181.11 & 2449.28 & 2430.96 \\
\multicolumn{1}{l|}{\textbf{SUCB}} & \textbf{2305.75} & \textbf{2314.76} & \textbf{2636.07} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 20 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2469.93 & 2336.92 & 2159.25 \\
\multicolumn{1}{l|}{TS} & 2260.20 & 1850.55 & 1905.42 \\
\multicolumn{1}{l|}{EXP3} & 2380.74 & 2317.82 & 2025.39 \\
\multicolumn{1}{l|}{\textbf{SUCB}} & \textbf{2474.69} & \textbf{2341.19} & \textbf{2177.70} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 30 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\\midrule
\multicolumn{1}{l|}{UCB} & 2443.57 & 2393.88 & 2472.31 \\
\multicolumn{1}{l|}{TS} & 2132.16 & 2248.56 & 1855.14 \\
\multicolumn{1}{l|}{EXP3} & 2368.42 & 2265.32 & 2339.49 \\
\multicolumn{1}{l|}{\textbf{SUCB}} & \textbf{2436.43} & \textbf{2397.72} & \textbf{2476.52} \\ \bottomrule
\end{tabular}
\caption{Cumulative rewards attained by algorithms with different number of clusters $10, 20, 30$ and different level of $\{B_{max} = 30, 50, 70\}$ across 500 time steps. }
\end{table}
\subsection{Online Influence Maximization}
We implemented the online influence maximization with an offline influence maximization algorithm TIM as our oracle \cite{10.1145/2588555.2593670}. TIM is one of the offline influence maximization algorithm that achieves asymptotic optimality. We perform the experiments on two datasets. One is a simulation dataset with 16 nodes and 44 edges, the other is a 100 nodes subset from Digg dataset, where each nodes represents a user from Digg website \cite{nr}. We simulate the connectivity of edges at each time step by randomly assigning each edge a connectivity probability at the start of the experiment.
\begin{figure}[hbt!]
\centering
\subfigure[]{
\label{test_graph}
\includegraphics[width=3.6cm]{figs/test_graph.png}}
\subfigure[]{
\label{16c2}
\includegraphics[width=3.6cm]{figs/test.png}}
\caption{\ref{test_graph} Visualization of synthetic graph for evaluation. \ref{16c2} Mean rewards attained by algorithms.}
\end{figure}
Figure \ref{test_graph} visualizes the synthetic graph we created for evaluation. We then tested the algorithms over $T = 800$ time steps with action size of 2 and maximum possible strategic budget of $B_{max} = 300$. Figure \ref{16c2} show the effectiveness of algorithms by evaluating based on the averaged nodes influenced. To ensure reproducible results, we calculate the mean of nodes influence over 100 trials. For better visualization, we kept a running mean over 50 time steps to smooth the lines.
We then investigate the effect of action size and the results are summarized in the following two tables.
\begin{table}[hbt!]
\centering
\begin{tabular}{@{}ccccc@{}}
\toprule
\multicolumn{5}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Synthetic, 16 nodes, 44 edges} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 4 & 6 & 8 \\ \midrule
\multicolumn{1}{c|}{CUCB} & 5.46 & 6.54 & 10.4 & 12.5 \\
\multicolumn{1}{c|}{TSCB} & 6.7 & 8.50 & 10.44 & 12.7 \\
\multicolumn{1}{c|}{EXP3CB} & 5.86 & 8.14 & 10.34 & 12.46 \\
\multicolumn{1}{c|}{\textbf{SCUCB}} & \textbf{7.44} & \textbf{9.32} & \textbf{11.16} & \textbf{13.04} \\ \bottomrule
\end{tabular}
\begin{tabular}{@{}lccc@{}}
\toprule
\multicolumn{4}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{digg, 100 nodes} \\ \midrule
\multicolumn{1}{l|}{Action size} & 10 & 15 & 20 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 15.64 & 20.64 & 25.32 \\
\multicolumn{1}{l|}{TSCB} & 16.30 & 21.62 & 26.08 \\
\multicolumn{1}{l|}{EXP3} & 12.16 & 17.76 & 21.72 \\
\multicolumn{1}{l|}{\textbf{SCUCB}} & \textbf{16.56} & \textbf{22.16} & \textbf{26.14} \\ \bottomrule
\end{tabular}
\caption{Averaged final rewards attained by algorithms with synthetic and Digg dataset with action size of 2, 800 time steps and $B_{max} = 200$.}
\end{table}
\iffalse
\paragraph{Experiment setting}
We conduct the experiments with 10 arms where each arm subset constitutes $2$ arms unless otherwise specified. The time horizon is set to be $T = 5\times 10^3$ and the approximation oracle is designed to succeed with probability $1$. Each bandit arm follows a Bernoulli distribution with randomly populated $\mu \in [0,1]$ and all arms adapt LSI except in the discussion of arm strategies.
The arms in the optimal arm subset have a strategic budget of $0$ and all other arms are equipped with the same strategic budget, which is $B_{max}$ by definition. The setting of heterogeneous budgets is discussed in the appendix.
As is typical in the bandit literature, for example in \citet{auer2002finite}, we evaluate both naive and strategic CUCB algorithms on their tuned versions, where the UCB exploration parameter is scaled by a constant factor $\gamma \in [0,1]$. Each experiment is repeated $10$ times and the averaged result is presented.
\begin{figure*}[htbp]
\centering
\subfigure[]{
\label{Fig1}
\includegraphics[scale=0.3]{figs/diff_B.png}}
\subfigure[]{
\label{Fig2}
\includegraphics[scale=0.24]{figs/linearB.png}}
\subfigure[]{
\label{Fig3}
\includegraphics[scale=0.3]{figs/LSI2.png}}
\subfigure[]{
\label{Fig4}
\includegraphics[scale=0.3]{figs/50arm.png}}
\caption{\ref{Fig1} Cumulative regret v.s. time horizon under different maximum budgets; \ref{Fig2} Cumulative regret v.s. strategic budget; \ref{Fig3} Cumulative regret under the LSI strategy and random strategy; \ref{Fig4} Cumulative regret v.s. time horizon under different number of arms in each arm subset ($B_{max} = 50$).}
\end{figure*}
\paragraph{Cumulative regret v.s. maximum budgets} We experiment with $B_{max} \in \{0, 10, 50, 100\}$ and $\gamma = 0.2$. Figure \ref{Fig1} highlights the failure of naive CUCB as the regret grows linearly with respect to $T$. In comparison, the regret grows in a logarithmic fashion with strategic CUCB, which agrees with our theoretical finding of the $O(\log T)$ regret bound.
\paragraph{Cumulative regret v.s strategic budget} We conduct our empirical study with and the strategic budget ranging from $1$ to $100$ under $\gamma = 0.2$. Figure \ref{Fig2} shows that the regret is linear with the largest strategic budget $B_{max}$ under both strategic CUCB and naive CUCB, which agrees with our linear term on the budget in Theorem \ref{appendixA}.
We observe that, though at a small value of $B_{max}$ CUCB display unstable performance, strategic CUCB can achieve much lower regret than that of naive CUCB in general. We, therefore, conclude the superiority of strategic CUCB in small-budget and large-budget regimes.
\paragraph{Cumulative regret under LSI strategy and random strategy} We simulate the random strategy for arms as arms randomly spend a budget that is generated from a standard uniform distribution in each round.
It is illustrated in Figure \ref{Fig3} that regardless of the choice of the bandit learning algorithm, the regret under LSI is higher than the regret under a random strategy. As is discussed previously, LSI strategy is one of the most powerful strategies that strategic arms can perform to hamper regret. The effect of strategies with collusion is demonstrated in the appendix.
\paragraph{Cumulative regret v.s. different arm subset size} We relax the number of arms in each arm subset from $2$ to $4$ and $6$ to observe the sensitivity with respect to this parameter. We choose naive CUCB with $4$ arms in each arm subset as a baseline and the maximum budget is set to be $50$. As shown in Figure \ref{Fig4}, our algorithm outperforms naive CUCB regardless of the number of arms in each arm subset, which indicates the robustness of strategic CUCB on the arm subset size.
\fi
\section{Regret analysis of strategic UCB}
\begin{namedtheorem}[Lemma \ref{lambda} (Re-statement)]
Let $\Lambda_{i,t} = \sqrt{\frac{3 \log{t} }{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}
, where $\rho_{i,t-1}$ is the total strategic budget spent by arm $i$ up to time $t-1$ and $K_{i,t-1}$ be the total number of pulls of arm $i$ up to time $t-1$. Define the event $E_t = \{ |\ \widetilde{\mu}_{i,t-1} - \mu_{i}| \leq \Lambda_{i,t}, \forall i \in [m]\}$, where $\mu_i$ is the true mean of arm $i$'s underlying distribution.
Then, $\mathbb{P}(\lnot E_t)\leq 2m \cdot t^{-2}$.
\end{namedtheorem}
\begin{proof}
Denote $\hat{\mu}_{i,t-1} = \frac{\sum_{j=1}^{t-1} x_{i,j}}{K_{i,t-1}}$ as the estimation of the expected reward without the manipulated variable $z_{i,t}$. By the definition of $\hat{\mu}_{i,t}$ and $\widetilde{\mu}_{i,t}$,
\begin{align}
& \mathbb{P}\left(|\widetilde{\mu}_{i,t-1} - \mu_i| > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}\right) \notag \\
=\ & \mathbb{P}\left(\left|\hat{\mu}_{i,t-1} + \frac{\rho_{i,t-1}}{K_{i,t-1}}- \mu_i\right| > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}} \ \right) \notag \\
=\ & \mathbb{P}\left(\hat{\mu}_{i,t-1} - \mu_i > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} \ \right) + \mathbb{P}\left(\hat{\mu}_{i,t-1} - \mu_i < -\sqrt{\frac{3\log{t}}{2K_{i,t-1}}} -2 \frac{\rho_{i,t-1}}{K_{i,t-1}} \right) \notag \\
\leq\ & \mathbb{P}\left(\hat{\mu}_{i,t-1} - \mu_i > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} \ \right) + \mathbb{P}\left(\hat{\mu}_{i,t-1} - \mu_i < -\sqrt{\frac{3\log{t}}{2K_{i,t-1}}}\right) \notag \\
=\ & \sum_{s=1}^{t-1} \mathbb{P}\left(|\hat{\mu}_{i,t-1} - \mu_i|> \sqrt{\frac{3\log{t}}{2s}}, K_{i,t-1} = s \ \right) \notag \\
\leq\ & \sum_{s=1}^{t-1} \mathbb{P}\left(|\hat{\mu}_{i,{t-1}} - \mu_i| > \sqrt{\frac{3\log{t}}{2s}}\ \right)\notag\\
\leq\ & 2t\cdot \exp{(-3\log{t})} = \frac{2}{t^2} \,, \nonumber
\end{align}
where the last inequality follows the Chernoff-Hoeffding bound. By the union bound,
\[
\quad \mathbb{P}(\lnot E_t) = \mathbb{P}\left( \{ |\widetilde{\mu}_{i,t-1} - \mu_{i}| > \Lambda_{i,t}, \forall i \in [m]\} \right)\leq 2m \cdot t^{-2} \,. \qedhere
\]
\end{proof}
\begin{namedtheorem}[Theorem \ref{appendixA} (Re-statement)]
The regret of the strategic CUCB algorithm with m strategic arms in time horizon $T$ using an $(\alpha,\beta)$-approximation oracle is at most
\begin{align}
m\cdot\Delta_{max}\left( \frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}+1 \right)\,, \notag \nonumber
\end{align}
where $f^{-1}(\cdot)$ is the inverse bounded smoothness function.
\end{namedtheorem}
\begin{proof}
We start by introducing a few notations. Let $F_t$ to be the event where the $(\alpha,\beta)$-approximation oracle fails to produce an $\alpha$-approximation answer with respect to the input $\bm{\Bar{\mu}} = (\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$ at time $t$. By definition of a $(\alpha,\beta)$-approximation oracle, we have $\mathbb{P}(F_t) \leq 1 - \beta$. Observe that at an arbitrary time $t$, a suboptimal arm subset may be selected due to two reasons, i) the $(\alpha, \beta)$-approximation oracle fails to provide an $\alpha$-approximate arm subset with respect to the input vector $\bm{\Bar{\mu}} = (\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$ and ii) the estimated mean vector $\bm{\bar{\mu}} = (\bar{\mu}_{1,t}, \bar{\mu}_{2,t},......,\bar{\mu}_{m,t})$ deviates from true values by a significant amount.
To account for $\sum^T_{t=1} \mathbb{I}(S_t \in S_B)$,where $\mathbb{I}(S_t \in S_B)$ is 1 when algorithm choose a suboptimal arm subset, $S_t \in S_B$, at time $t$, we introduce a counter $N_i$ for each arm $i\in[m]$ after the $m$-round initialization and let $N_{i,t}$ be the value of $N_i$ at time $t$. We initialize $N_{i,m} = 1$. By definition, $\sum_{i \in [m]} N_{i,m} = m$. For $t > m$, the counter $N_{i,t}$ is updated as follows.
\begin{itemize}
\item If $S_t \in S_B$, then $N_{{i'},t} = N_{{i'},t-1} + 1$ where $i' = \argmin_{i \in S_t} N_{i,t-1}$. In the case that $i'$ is not unique, we break ties arbitrarily.
\item If $S_t \not \in S_B$, no counters will be updated then.
\end{itemize}
As such, the total number of pulls of suboptimal arm subsets is less than or equal to $\sum_{i=1}^m N_{i,T}$.
Define $\psi_t = \frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{t}} {\left(f^{-1}(\Delta_{min})\right)^2} > c\,, $
where $c$ is the larger solution of the following equation,
\begin{align}
&\frac{\left(f^{-1}(\Delta_{min})\right)^2}{4} c^2
- \left(2B_{max} f^{-1}(\Delta_{min}) + \frac{3\log{t}}{2}\right)c + 4B_{max}^2 = 0\,. \label{square}
\end{align}
By solving Equation \eqref{square}, we have
\begin{align}
c &= \frac{2\left(2B_{max} f^{-1}(\Delta_{min}) + \frac{3\log{t}}{2}\right)} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{2\sqrt{\left(2B_{max} f^{-1}(\Delta_{min}) + \frac{3\log{t}}{2}\right)^2 -4(f^{-1}(\Delta_{min})) B_{max}}} {\left(f^{-1}(\Delta_{min})\right)^2}\,. \label{psi}
\end{align}
We then decompose the total number $\sum_{i=1}^m N_{i,T}$ of pulls of suboptimal arm subsets as
\begin{align}
\quad\sum_{i=1}^m N_{i,T} =\ & m +\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B\} \notag \\
=\ & m +\sum_{t=m+1}^T \sum_{i\in[m]}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} \leq \psi_t\} \notag \\
& + \sum_{t=m+1}^T \sum_{i\in[m]}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} > \psi_t\}\notag \\
\leq\ & m + m\psi_T +\sum_{t=m+1}^T \sum_{i\in[m]}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} > \psi_t\}\notag \\
=\ & m + m\psi_T+\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B , \forall i \in S_t, N_{i,t-1} > \psi_t\} \,.\label{8}
\end{align}
The first inequality follows as $\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B, \forall i \in S_t, N_{i,t-1} \leq \psi_t\}$ can be trivially upper bounded by $\psi_t$ and the second equality holds by our rule of updating the counters.
The third term of Equation \eqref{8} can be further decomposed according to whether the oracle fails,
\begin{align}
&\quad \sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B, N_{i,t-1} > \psi_t, \forall i \in S_t\}\notag\\
&\leq \sum_{t=m+1}^T (\mathbb{I}\{F_t\} + \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\})\label{ke} \\
&= (T-m)(1-\beta)+\sum_{t=m+1}^T \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t , \forall i \in S_t\}\,. \notag
\end{align}
Let $\Lambda_{i,t} = \sqrt{\frac{3 \log{t} }{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}$ where $\rho_{i,t-1}$ is the total strategic budget spent by arm $i$ up to time $t-1$. Define event $E_t = \{ |\ \widetilde{\mu}_{i,t-1} - \mu_{i}| \leq \Lambda_{i,t}, \forall i \in [m]\}$, where $\mu_i$ is the true mean of arm $i$'s underlying distribution without manipulation. We continue the proof under $\mathbb{P}(\{ E_t, \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t \}) = 0$ and prove it afterwards.
Since $\mathbb{P}\left(\{ E_t, \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t , \forall i \in S_t\} \right) = 0$, by inclusion-exclusion principle,
$ \mathbb{P}(\{\lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\}) \leq \mathbb{P}(\lnot E_t)\,.$ Denote $\hat{\mu}_{i,t-1} = \frac{\sum_{j=1}^{t-1} x_{i,j}}{K_{i,t-1}}$ as the estimation of the expected reward without the manipulated variable $z_{i,t}$. By the definition of $\hat{\mu}_{i,t}$ and $\widetilde{\mu}_{i,t-1}$,
\begin{align}
&\quad \mathbb{P}\left(|\widetilde{\mu}_{i,t-1} - \mu_i| > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}} + \frac{B_{max}}{K_{i,t-1}} \ \right) \notag \notag \\
&= \mathbb{P}\left(\left|\hat{\mu}_{i,t-1} + \frac{\rho_{i,t-1}}{K_{i,t-1}}- \mu_i\right| > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}+\frac{B_{max}}{K_{i,t-1}} \right) \notag \\
&\leq \mathbb{P}\left(\left|\hat{\mu}_{i,t-1} + \frac{\rho_{i,t-1}}{K_{i,t-1}}- \mu_i\right| > \sqrt{\frac{3\log{t}}{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}} \ \right) \notag \\
&\leq 2t^{-2}\,,
\end{align}
where the last inequality holds due to Lemma \ref{lambda}.
By the union bound,
\begin{align}
&\ \quad \mathbb{P}(\lnot E_t) = \mathbb{P}( \{ |\widetilde{\mu}_{i,t-1} - \mu_{i}| > \Lambda_{i,t}, \forall i \in [m] \} \ ) \leq 2m \cdot t^{-2}\,. \notag
\end{align}
Hence,
\begin{align}
\mathbb{P}(\{\lnot F_t, S_t \in S_B, \forall i \in S_t, K_{i,t-1} > \psi_t, \forall i \in S_t\}) \leq \mathbb{P}(\lnot E_t) \leq 2m \cdot t^{-2}\,. \notag
\end{align}
We now show that $\mathbb{P}(\{ E_t, \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t \}) = 0$. Let $\Lambda = \sqrt{\frac{3 \log{t}}{2\psi_t}}+ \frac{2B_{max}}{\psi_t} $, which is not a random variable, and $B_{max} = max_{i \in [m]} B_i$, where $B_i$ is the strategic budget for arm $i$. For variable $\Lambda_{i,t}$, let $\Lambda_{i,t}^{\ast} = \max \{\Lambda_{i,t}\}$. Since $K_{i,t-1} > \psi_t$ and $B_{max} \geq B_i \geq \rho_i$, we have $\Lambda > \Lambda_{i,t}^{\ast}$. According to line 7 of Algorithm 1, we have $\Bar{\mu}_{i,t}= \widetilde{\mu}_{i,t-1} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}} + \frac{B_{max}}{K_{i,t-1}}$ and $\bar{\mu}_{i,t} = \hat{\mu}_{i,t-1} + \frac{\rho_{i,t-1}}{K_{i,t-1}} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}}+ \frac{B_{max}}{K_{i,t-1}}$. Thus, $|\hat{\mu}_{i,t-1} - \mu_i| \leq \Lambda_{i,t}$ implies that $0 < \bar{\mu}_{i,t-1} - \mu_i \leq 2\Lambda_{i,t} \leq 2\Lambda_{i,t}^{\ast} \leq 2\Lambda$, for $ i \in S_t$.
Recall $S^{\ast} = \argmax_{S \in \mathcal{S}} r_{\bm{\mu}}(S)$ and ${OPT}_{\bm{\mu}} = \max_{S\in \mathcal{S}} r_{\bm{\mu}}(S)$. Suppose $\{E_t, \lnot F_t, S_t \in S_B, \forall i \in S_t, K_{i,t-1} > \psi_t \} $ happens at time $t$, the following holds
\begin{align}
\quad r_{\bm{\mu}}(S_t) + f(2\Lambda) \notag & \geq r_{\bm{\mu}}(S_t) + f(2\Lambda_{i,t}^{\ast}) \geq r_{\bm{\bar{\mu}}}(S_t) \\ \notag
& \geq \alpha \cdot {OPT}_{\bar{\bm{\mu}}}\geq \alpha \cdot r_{\bar{\bm{\mu}}}(S^{\ast}_{\bm{\mu}}) \geq \alpha \cdot r_{{\bm{\mu}}} (S^{\ast}_{\bm{\mu}}) \\ \notag
& = \alpha \cdot {OPT}_{\bm{\mu}}\,. \notag
\end{align}
The first inequality is due to the strict monotonicity of $f(\cdot)$ and $\Lambda > \Lambda_{i,t}^{\ast}$. The second inequality is due to the bounded smoothness property and $|\bar{\mu}_{i,t-1} - \mu_i | \leq 2\Lambda_{i,t}$. The third inequality is due to the fact that $\lnot F_t$ implies $S_t \geq \alpha \cdot opt_{\bm{\mu}}$. The forth inequality is by the definition of $opt_{\bm{\mu}}$. The last inequality is is due to the monotonicity of $r_{\bm{\mu}}(S)$ and $0 < \bar{\mu}_{i,t-1} - \mu_i$.
Let $\kappa = \sqrt{\frac{3 \log{t}}{2c}} + \frac{2B_{max}}{c}$ where $c$ takes the value defined in Equation \eqref{square}. Given $\psi_t > c$, we have $\Lambda = \sqrt{\frac{3 \log{t}}{2\psi_t}} + \frac{2B_{max}}{\psi_t} < \kappa$. By Equation \eqref{psi}, we have $f(2\Lambda) < f(2\kappa) = \Delta_{min}$ and $\Delta_{min} > \alpha \cdot opt_{\bm{\mu}} - r_{\bm{\mu}}(S_t)$, which contradicts the definition of $\Delta_{min}$ and the fact that $S_t \in S_B$. Therefore,
\begin{equation}
\mathbb{P}(E_t, \lnot F_t, S_t \in S_B, \forall i \in S_t, K_{i,t-1} > \psi_t )= 0\,. \notag
\end{equation}
Based on the above analysis, we can upper bound the total number of suboptimal pulls as
\begin{align}
\mathbb{E}\left[\sum_{i=1}^m N_{i,T} \right]
&\leq m(1+\psi_T) + (T-m)(1-\beta) + \sum_{t=m+1}^T \frac{2m}{t^2}\notag \\
&= m\left(1+\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2}\right) + (T-m)(1-\beta) + \sum_{t=m+1}^T \frac{2m}{t^2}\,.
\end{align}
Since the cumulative regret relate closely to the total number of suboptimal pulls $\mathbb{E}[\sum_{i=1}^m N_{i,T}]$, the upper bound of cumulative regret is thus
\begin{align}
\quad Regret_{\bm{\mu},\alpha,\beta}(T)
\leq\ & T \cdot \alpha \beta \text{OPT}_{\bm{\mu}} - \left( T \cdot \alpha \text{OPT}_{\bm{\mu}} - \mathbb{E} \left[ \sum_{i=1}^m N_{i,T} \right] \Delta_{max} \right)
\notag \\
=\ & (\beta-1) T \cdot \alpha \text{OPT}_{\bm{\mu}} + \Delta_{max}\left( m \left(1+\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2}\right)\notag \right) \notag \\
&+ \Delta_{max}\left((T-m) (1-\beta) +\sum_{t=m+1}^T \frac{2m}{t^2}\right)\notag\\
\leq\ & ((T-m) \cdot \Delta_{max}- T \cdot \alpha \text{OPT}_{\bm{\mu}})(1-\beta)
\notag \\
& + m \cdot \Delta_{max}\left(1+\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}\right)\,.\notag
\end{align}
Each time the algorithm pull a suboptimal arm subset $S_t \in S_B$ at time $t$, the algorithm incur an additional regret of at most $\Delta_{max}$, which is less than or equal to $\alpha \cdot opt_{\bm{\mu}}$. Thus,
\begin{align}
&\quad (T-m) \Delta_{max}-T\alpha \cdot \text{OPT}_{\bm{\mu}} \notag\\
&\leq (T-m) \alpha \cdot \text{OPT}_{\bm{\mu}}- T\alpha\cdot \text{OPT}_{\bm{\mu}}\notag\\
&=-m \alpha\cdot \text{OPT}_{\bm{\mu}} < 0\,.\notag
\end{align}
As a result, the regret of the strategic CUCB algorithm under strategic manipulations of reward is at most
\begin{align}
&Regret_{\bm{\mu},\alpha,\beta}(T)\leq m \cdot \Delta_{max} \left(\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3} + 1\right)\notag \,. \qedhere
\end{align}
\end{proof}
\section{Lower bound on the strategic budget}
\begin{namedtheorem}[Theorem \ref{lower}]
In stochastic multi-armed bandit problems, for a strategic suboptimal arm $i$, under time horizon $T$ and without access to other arms' information, the minimum strategic budget needed for it to be pulled $\omega(\log{T})$ is $\omega(\log{T})$. The regret incurred for any bandits learning algorithm is thus $\omega(\log{T})$.
\end{namedtheorem}
\begin{proof}
Let time $t \in [1,T]$ be the time step arm $i$ is last pulled under UCB algorithm and $\eta$ is a parameter chosen by the algorithm. The following inequality must stands at time $t$,
\begin{align*}
& \hat{\mu}_{i,t} + \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}} + \frac{\rho_i}{K_{i,t}}\geq \hat{\mu}_{i^{\ast},t} + \sqrt{\frac{2\log(K_{i^{\ast},t}^2/\eta^2)}{K_{i^{\ast},t}}}\,.
\end{align*}
By the Chernoff-Hoeffding bound and the union bound,
\begin{align*}
&\quad \mathbb{P} \left(\hat{\mu}_{i,t} - \mu_i \geq \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}}\right)\\
&\leq \sum^{t}_{s=1} \mathbb{P} \left( \hat{\mu}_{i,t} - \mu_i \geq \sqrt{\frac{2\log(s^2/\eta^2)}{s}}, K_{i,t} = s \right) \\
&\leq\sum^{t}_{s=1} \mathbb{P} \left( \hat{\mu}_{i,t} - \mu_i \geq \sqrt{\frac{2\log(s^2/\eta^2)}{s}}\right) \\
& \leq \sum^{t}_{s=1} \frac{\eta^2}{s^2} = \eta^2\sum^{t}_{s=1} \frac{1}{s^2} \leq \frac{\pi^2}{6} \eta^2 \,.
\end{align*}
Thus $\hat{\mu}_{i,t} - \mu_i \leq \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}}$ and similarly $\mu_{i^{\ast},t} - \hat{\mu}_{i^{\ast}} \leq \sqrt{\frac{2\log(K_{i^{\ast},t}^2/\eta^2)}{K_{i^{\ast},t}}} $, each with probability of at least $1 - \frac{\pi^2}{6}\eta^2$. Hence with probability $1 - 2 \frac{\pi^2}{6}\eta^2$, we have
\begin{align}
\mu_i + \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}} + \frac{\rho_i}{K_{i,t}} \geq \mu_{i^{\ast}}\,. \notag
\end{align}
When arm $i$ is pulled, arm $i$ wants to ensure the following holds
\begin{align}
\sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}} + \frac{\rho_i}{K_{i,t}} \geq \delta_i\,,\notag
\end{align}
where $\delta_i = \mu_{i^{\ast}} - \mu_i$. Then,
\begin{equation*}
B_i \geq \rho_i \geq \left(\delta_i - \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}}\right) \cdot K_{i,t}\,. \qedhere
\end{equation*}
\end{proof}
\section{Collusion strategy with the objective of regret maximization}
Stemming from the strong correlation between the cumulative regret and the number of suboptimal pulls, the optimal collusion strategy under the regret maximization objective is a variant of the combinatorial optimization problem defined earlier. Specifically, the maximization objective of the optimization problem takes a priority weight of each $Y_i$ based on the $\delta_i = \mu_i - \mu_{i^\ast}$, as
\begin{equation*}
\begin{aligned}
& \underset{Y_{i},t_i,\ i\in [m]}{\mathrm{maximize}}
& & \delta_1 \cdot Y_{1}+\dots+ \delta_m \cdot Y_{m} \\
& \mathrm{subject}\text{ }\mathrm{to}
& & \frac{B_i}{Y_{i}} + \sqrt{\frac{3\log t_i}{Y_{i}}} \ge \delta_i + \sqrt{3\log t_i},\ i \in [m]\,, \\
& & & t_{i_j} \ge \sum_{l=1}^j Y_{i_l}, \ j\in [m]\,,\\
& & & \textup{for some permutation}\ i_1,\dots\ i_m\ \textup{of}\ [m]\,.\\
\end{aligned}
\end{equation*}
\section{Additional simulation results}
\begin{figure*}[htbp]
\centering
\subfigure[]{
\label{Fig7}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/diff_B_card4.png}}
\subfigure[]{
\label{Fig8}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/diff_scale.png}}
\caption{\ref{Fig7} Cumulative regret v.s. time horizon with different level of strategic budget; \ref{Fig8} Effects of tuning UCB exploration parameter.}
\end{figure*}
\begin{figure*}[htbp]
\centering
\subfigure[]{
\label{Fig9}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/diff_arm_1010.png}}
\subfigure[]{
\label{Fig10}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/diff_arm_100.png}}
\caption{\ref{Fig9} Cumulative regret v.s. time horizon with different number of arms in each arm subset under $B_{max} = 10$; \ref{Fig10} Cumulative regret v.s. time horizon with different number of arms in each arm subset under $B_{max} = 100$.}
\end{figure*}
\begin{figure*}[htbp]
\centering
\subfigure[]{
\label{Fig5}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/pulls_final_final.png}}
\subfigure[]{
\label{Fig6}
\includegraphics[width=8cm,height=6cm,keepaspectratio]{figs/regret_final_final.png}}
\caption{\ref{Fig5} The number of suboptimal pulls v.s. collusion strategies; \ref{Fig6} Regret v.s. collusion strategies}
\end{figure*}
\bigbreak
\paragraph{Experiment setting} All of the experiment settings remain the same as the experiments section unless otherwise indicated.
\paragraph{Cumulative regret v.s. maximum budget} We experiment with $B_{max} = \{0, 10, 50, 100\}$ and $\gamma$ = $0.2$. Previously we have empirically evaluated the performance of strategic CUCB when each suboptimal arm process the same amount of strategic budget in Figure \ref{Fig1}. We now simulate the situation when suboptimal arms have strategic budgets of different levels by uniformly randomly assigning budgets from $[0,B_{max}]$ to each suboptimal arm. As shown in Figure \ref{Fig7}, the performance of strategic CUCB is still robust when there is a diversity of strategic budget owned by arms. Meanwhile, our strategic CUCB algorithm still outperforms naive CUCB under this scenario.
\paragraph{Effects of tuning UCB exploration parameter} As shown in Figure \ref{Fig8}, our strategic CUCB achieves the best performance with $\gamma = 0.2$. This determines the discount parameter used in the experiment section.
\paragraph{More experiments on arm subset size} Similar to the experiment section, we relax the number of arms in each arm subset from $2$ to $4$ and $6$. We choose naive CUCB with $4$ arms in each arm subset as a baseline. We experiment on $B_{max} \in \{10, 50, 100\}$, respectively. Figure \ref{Fig9} and Figure \ref{Fig10} showcase that our algorithm outperforms naive CUCB regardless of the number of arms in each arm subset under all budget values.
\paragraph{Performance of different collusion strategies}
We experiment with three collusion strategies that are variants of the lump sum investing (LSI) strategy, which is the dominant Nash equilibrium strategy for the subgame between strategic arms. The maximum budget is set to be $50$. In prioritized budget LSI (PB-LSI) strategy, each strategic arm $i$ spends $\delta_i = \mu_{i^{\ast}} - \mu_i$ in the initialization rounds and then spends the rest of their budget all at once by the order of the value of their strategic budget. The prioritized delta LSI (PD-LSI) and prioritized budget-delta LSI (PBD-LSI) operate in similar fashions, but strategic arms spend budget according to the increasing order of their suboptimality gap $\delta_i = \mu_{i^{\ast}} - \mu_i$ and the increasing order of $B_i - \delta_i$, respectively. Figure \ref{Fig5} and Figure \ref{Fig6} showcase scenarios where the LSI strategy is suboptimal under the common objective of maximizing suboptimal pulls and maximizing regret. The figures show that considering the budget or suboptimal gap might not be sufficient for the decision of the collusion strategy, under both the objectives. It is clear that PBD-LSI outperforms the other three strategies.
However, as the suboptimality gaps are constants, the effects of different collusion strategies on cumulative regret is limited and is fully observed in a short horizon.
\subsubsection*{References}}
\usepackage{mathtools}
\usepackage{tikz}
\DeclareSymbolFont{extraup}{U}{zavm}{m}{n}
\DeclareMathSymbol{\varheart}{\mathalpha}{extraup}{86}
\DeclareMathSymbol{\vardiamond}{\mathalpha}{extraup}{87}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\mathbb{P}}{\mathbb{P}}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{\mathbb{I}}{\mathbb{I}}
\newcommand{\mathbb{N}}{\mathbb{N}}
\newcommand{\mathbb{B}}{\mathbb{B}}
\newcommand{\mathbb{Q}}{\mathbb{Q}}
\newcommand{\mathbb{Z}}{\mathbb{Z}}
\newcommand{{O}}{{O}}
\newcommand{{o}}{{o}}
\newcommand{\mathbf{e}}{\mathbf{e}}
\newcommand{\mathcal{S}}{\mathcal{S}}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{\mathcal{D}}{\mathcal{D}}
\newcommand{\mathcal{M}}{\mathcal{M}}
\newcommand{\mathcal{N}}{\mathcal{N}}
\newcommand{\mathcal{I}}{\mathcal{I}}
\newcommand{\mathcal{Y}}{\mathcal{Y}}
\newcommand{\mathcal{T}}{\mathcal{T}}
\newcommand{\mathcal{R}}{\mathcal{R}}
\newcommand{\overline{R}}{\overline{R}}
\newcommand{\boldsymbol{a}}{\boldsymbol{a}}
\newcommand{\mathcal{S}}{\mathcal{S}}
\newcommand{\mathbb{A}}{\mathbb{A}}
\newcommand{\mathbbm{1}}{\mathbbm{1}}
\def\mathcal{A}{\mathcal{A}}
\def\mathcal{B}{\mathcal{B}}
\def\mathcal{D}{\mathcal{D}}
\def\bm{V}{\bm{V}}
\def\bm{X}{\bm{X}}
\def\bm{\overline{X}}{\bm{\overline{X}}}
\def\bm{Y}{\bm{Y}}
\def\bm{Z}{\bm{Z}}
\def\bm{W}{\bm{W}}
\def\bm{A}{\bm{A}}
\def\bm{\hat{A}}{\bm{\hat{A}}}
\def\bm{B}{\bm{B}}
\def\bm{\hat{B}}{\bm{\hat{B}}}
\def\bm{x}{\bm{x}}
\def\bm{y}{\bm{y}}
\def\bm{z}{\bm{z}}
\def\bProofm{u}{\bProofm{u}}
\def\underline{X}{\underline{X}}
\def\underline{Y}{\underline{Y}}
\def\underline{Z}{\underline{Z}}
\def\underline{x}{\underline{x}}
\def\underline{y}{\underline{y}}
\def\underline{z}{\underline{z}}
\def{\varepsilon}{{\varepsilon}}
\def\{0,1\}{\{0,1\}}
\def\mathrm{NP}{\mathrm{NP}}
\def\mathrm{co}{\mathrm{co}}
\def\varepsilon{\varepsilon}
\def\oplus{\oplus}
\def\mid{\mid}
\def\hat{\ell}{\hat{\ell}}
\providecommand\floor[1]{\lfloor#1\rfloor}
\providecommand\ceil[1]{\lceil#1\rceil}
\providecommand\blog[1]{\log_2\ceil{#1}}
\providecommand\abs[1]{\lvert#1\rvert}
\providecommand\bigabs[1]{\bigl\lvert#1\bigr\rvert}
\providecommand\ip[1]{\langle#1\rangle}
\providecommand\absip[1]{\abs{\ip{#1}}}
\DeclareMathOperator\Var{\mathrm{Var}}
\DeclareMathOperator\Cov{\mathrm{Cov}}
\DeclareMathOperator\poly{\mathrm{poly}}
\DeclareMathOperator{\pr}{\mathrm{Pr}}
\newcommand{\red}[1]{\textcolor{red}{#1}}
\newcommand{\blue}[1]{\textcolor{blue}{#1}}
\newcommand{\orange}[1]{\textcolor{orange}{#1}}
\newcommand{\violet}[1]{\textcolor{violet}{#1}}
\def\mathrm{P}{\mathrm{P}}
\providecommand\norm[1]{\|#1\|}
\newcommand{\bracket}[1]{\left(#1\right)}
\newcommand{\sbracket}[1]{\left[#1\right]}
\makeatletter
\newcommand*{\rom}[1]{\expandafter\@slowromancap\romannumeral #1@}
\usepackage{booktabs}
\section{Introduction}
\label{sec:intro}
Sequential learning methods feature prominently in a range of real applications such as online recommendation systems, crowdsourcing systems, and online influence maximization problems. Among those methods, the multi-armed bandits problem serves as a fundamental framework. Its simple yet powerful model characterizes the dilemma of exploration and exploitation which is critical to the understanding of online sequential learning problems and their applications \citep{10.2307/2332286, Robbins:1952,auer2002finite,lattimore_szepesvari_2020}.
The model describes an iterative game constituted by a bandit algorithm and many arms. The bandit algorithm is required to, through a horizon $T$, choose an arm to pull at each time step. As the objective is to maximize the cumulative reward over time, the algorithm balances between exploiting immediate rewards based on the information collected or pulling less explored arms to gain more information about arms \citep{anantharam_asymptotically_1987,auer_nonstochastic_2002,cesa-bianchi_prediction_2006}.
Out of the real applications, many motivate the extension of MAB towards combinatorial multi-armed bandits (CMAB), where multiple arms can be selected in each round \citep{pmlr-v28-chen13a,DBLP:conf/nips/CombesSPL15,pmlr-v48-lif16,pmlr-v97-zimmert19a,pmlr-v117-rejwan20a}. CMAB demonstrates its effectiveness on problems like online social influence maximization, viral marketing, and advertisement placement, within which many offline variants are NP-hard. However, existing MAB and CMAB algorithms are often developed either under benign assumptions on the arms \citep{pmlr-v28-chen13a,pmlr-v38-kveton15,pmlr-v75-wei18a} or purely adversarial arms \citep{auer2002nonstochastic}. In the former setting, the arms are commonly assumed to report their reward signals truthfully without any strategic behavior, under which the drawbacks are apparent. In the latter setting, arms can attack any deployed algorithm to regret of $O(T)$ with this capability of reward manipulations, which is catastrophic for a bandit algorithm.
This assumption is stringent and rarely realistic.
In this paper, we adapt the combinatorial UCB (CUCB) algorithm with a carefully designed UCB-based exploration term.
A major difficulty stems from not knowing the manipulation term made by the arms, while our algorithm overcomes this by depending only on the knowledge of the maximum possible strategic budget.
Previous results by \citep{pmlr-v119-feng20c} only implies robustness of UCB style algorithm under stochastic multi armed bandits setting under only full knowledge of step wise deployment of strategic budget.
New tail bounds over the proposed exploration term and a new trade-off parameter that balances exploration and exploitation are utilized to facilitate the analysis of our algorithm.
We further establish results on the robustness of our UCB variant under strategic arms with an algorithm-dependent budget lower bound.
Our proposed algorithms are also evaluated empirically through an extensive set of synthetic environments and real datasets.
The real applications include reliable workers selection in online crowdsourcing systems, where workers might misrepresent a result for a better chance to be selected in the future; online information maximization, where nodes modify the spread to include itself into the seed set; and online recommendation systems, which characterizes the ``click framing'' behavior. Through a wide range of tasks and parameters, the experiment results corroborate our theoretical findings and demonstrate the effectiveness of our algorithms.
\subsection{Motivating Examples}
The setting of strategic manipulations describes the strategic behavior found in a variety of real applications.
Consider a crowdsourcing platform that provides a data labeling service for payment. The platform interacts with a group of customers and a pool of workers. The customers request tasks to the platform for labeling and the platform is then responsible for selecting workers from the worker pool to complete the tasks. This process repeats, during which the platform learns the best deployment. We maintain a mild assumption that the payments from customers are non-decreasing with the quality of labels. To maximize its profit, it is desired for the platform to select workers that provides the most reliable labels. The workflow of the platform can be described in the diagram below.
While the platform and the customers desire quality labels, it may not entirely be in the worker's interest to exert the highest effort and thus report quality labels each time, which factors into a range of reasons. Workers may adapt strategic behaviors to maximize their own utility instead. Thus, it becomes crucial to identify reliable workers to prevent strategic behaviors that jeopardize profits.
This naturally translates to a multi-armed bandits problem under strategic manipulations where the workers are the strategic arm and the payments act as the rewards. Under ideal assumptions it had shown the effectiveness of bandits algorithms on such problems \citep{jain2014quality, tran2014efficient,rangi2018multi}.
\begin{figure}[H]
\includegraphics[width=7.5cm]{figs/crowdsys.png}
\caption{Online crowdsourcing system}\label{crowdsys}
\end{figure}
\section{Related Work}
The problem of multi-armed bandits (MAB) was first investigated back in 1952 while some techniques utilized were developed back in 1933 \citep{10.2307/2332286, Robbins:1952, berry,auer2002finite,lattimore_szepesvari_2020}. Since then it has been extensively explored and serves as the foundation of many modern areas, including reinforcement learning, recommendation systems, graph algorithms, etc. \citep{DBLP:conf/iconip/BouneffoufBG12,pmlr-v70-vaswani17a, pmlr-v85-durand18a,10.5555/3367243.3367445,li2020stochastic}. With the need to model the selection of multiple arms in one round, MAB is then extended to combinatorial multi-armed bandits (CMAB), which see many deployments in real applications like news/goods recommendation, medical trials, routing, and so forth \citep{Wang_Ouyang_Wang_Chen_Asamov_Chang_2017,8022964,pmlr-v84-sankararaman18a, li2020online}.
Real applications motivate the community to derive algorithms in more adaptive environments. The first of which considers adversarial bandits with the classic EXP3 algorithm. The arms are assumed to be non-stationary but non-adaptive (which means that algorithms will adapt to the adversarial) \cite{auer2002nonstochastic}. Despite that adversarial bandits do not fall into the scale of this paper's related work, it leads tremendous effort to the following topics in adaptive arms.
\paragraph{Adversarial corruptions}
The adversary is given the ability to replace the observations of the bandit algorithm with arbitrary values within some constraints.
\citet{pmlr-v99-gupta19a} discuss the constraint that rewards are modified for at most $T_0$ rounds out of the total $T$ rounds. The asymptotic regret bound $O(mT_0)$ achieved in the work is shown to match the negative result, where $m$ is the number of arms. \citet{lykouris_stochastic_2018} discussed a different constraint where the corruption $z_t$ at time $t$ is cumulative up to some constant $B$.
The regret achieved under this constrained corruption is bounded by $O(mB\log(\frac{mT}{\delta}))$ with probability at least $1 - \delta$. \citet{JMLR:v20:18-395} restricts corruptions as independent Bernoulli events with probability $b$, while if corruption happens the reward becomes arbitrary and adversarial. This problem is addressed with median statistics with gap-dependent matching regret upper and lower bounds of $O(\sum_{i \neq i^{\ast}} \frac{1}{\Delta_i}\log(\frac{K}{\Delta_i}))$, where $i^{\ast}$ denotes the optimal arm and $\Delta_i$ is the suboptimality of arm $i$.
Several lines of research also discuss corruption topics in frequentist inference and partial monitoring \citep{pmlr-v83-gajane18a}, best arm identification \citep{pmlr-v99-gupta19a}, and adversarial bandits \citep{NEURIPS2020_e655c771}.
\paragraph{Adversarial attacks}
Different from arms in adversarial corruptions who intend to maximize the regret of the bandit algorithm, arms in the adversarial attack setting have the goal to maximize their number of pulls. The setting is first considered by \citet{NEURIPS2018_85f007f8}, where it shows that the attacker may spend a budget of $O(\log T)$ to deceit UCB and $\epsilon$-greedy algorithms to pull the target arms for at least $T - o(T)$ times. Stochastic and contextual MAB algorithms also suffer from undesired loss from this adversarial attack, under a variety of attack strategies \citep{pmlr-v97-liu19e}. Linear MAB algorithms, however, are shown to be near-optimal up to some linear or quadratic terms on the attack budget $B$ \citep{garcelon_adversarial_2020,bogunovic_stochastic_2020}.
\paragraph{Strategic manipulations}
The setting of strategic manipulations further weakens the capability of the adaptive arms and prevents the bandit algorithms from being overcautious.
The objective of an arm is still utility maximization under strategic manipulations, while each arm works on its utility individually.
Instead of having a global coordinator for the adversarial attack strategy, the strategic arms seek the best response via the dominant Nash equilibrium.
The strategic manipulation setting is first studied by \citet{pmlr-v99-braverman19b} where each arm is assumed to keep a portion of the reward in each round and arms maximize the cumulative reward it keeps.
The performance of bandit algorithms will then be catastrophic under the dominant Nash equilibrium, receiving an expected cumulative reward of $0$ almost surely.
With the utility of maximizing the number of pulls for each strategic arm, common bandits algorithms are proved to be robust with guaranteed $O(\log T)$ regret bound, but only under constant strategic budgets \citep{pmlr-v119-feng20c}.
For $\omega(\log T)$ budgets, bandits under strategic manipulations remain an open problem and will be discussed in this paper.
\section{Problem Formulation}
We consider the problem of combinatorial multi-armed bandits (CMAB) under the setting of strategic arms. In strategic CMAB, each arm is given a budget and the arm can strategically increase its emitted reward signals cumulatively up to this budget \emph{for its own interest}. This problem is a Stackelberg game that involves two parties. The bandit learning algorithm is deployed first to maximize its cumulative reward under the best response of the followers. The $m$ strategic bandits arms then deplete their budget where each of them aims to maximize its expected number of pulls. Knowing the principal's strategy and other followers' budget information, the followers are assumed to place their strategies according to the dominant Nash equilibrium between them.
Formally, given a time horizon $T$, the the bandit learning algorithm is asked to pull a subset of arms, denoted by an arm subset $S_t \in \mathcal{S} $ at each time $t\in [T]$, where $\mathcal{S} = { \{0,1\}^{m}} $ is the set of all possible arm subsets. At the time $t$, based on the principal's choice of arm subset, stochastic rewards $x_{i,t}$ are generated by arm $i$ from the underlying $1$-sub-Gaussian distribution with mean $\mu_i$, independent of other arms in $S_t$. The principal does not observe these rewards. Instead, each arm can increase the emitted reward signal by an arbitrary amount $z_{i,t}\ge 0$, as long as the cumulative manipulation over the horizon does not exceed a given budget $B_i$. The principal then receives the semi-bandit feedback $\{x_{i,t}+z_{i,t}\}_{i\in S_t}$.
Let $\bm{\mu} = (\mu_1, \mu_2, ..., \mu_i) $ be the vector of expectations of all arms. The expected reward of playing any arm subset S in
any round $r_{\bm{\mu}}(S) = \mathbb{E}[R_t(S)]$,which is a function of arm subset $S$ and $\bm{\mu}$. The reward function $r_{\bm{\mu}}(S)$ is assumed to satisfy two axioms:
\begin{itemize}
\item \textbf{Monotonicity.} The expected reward of playing any arm subset $S \in \mathcal{S}$ is monotonically non-decreasing with respect to the expected reward vector. That is, if for all $i \in [m], \mu_i \leq \mu_{i}'$, then $r_{\bm{\mu}}(S) \leq r_{\bm{\mu}'}(S)$.
\item \textbf{Bounded smoothness.} There exists a strictly increasing function $f(\cdot)$, termed the bounded smoothness function, such that for any two expected reward vectors $\bm{\mu}$ and $\bm{\mu'}$ satisfying $\|\bm{\mu}-\bm{\mu'}\|_{\infty}\le \Lambda $, we have $|r_{\bm{\mu}} - r_{\bm{\mu'}}| \leq f({\Lambda})$.
\end{itemize}
These axioms cover a rich set of reward functions and the explicit forms of $R_t(S)$ and $r_{\bm{\mu}}(S)$ are not needed to be specified \citep{pmlr-v28-chen13a}.
Without loss of generality, assume that $S^\ast = \argmax_{S\subseteq \mathcal{S}} r_{\bm{\mu}}(S)$ is the unique optimal subset of arms. When placing reward manipulations, each strategic arm has access to its own history $h_{i,t} = \{I_{i,t'}, x_{i,t'}, z_{i,t'}\}_{t'\ge 1}$, where $I_{i,t'}$ is the indicator of whether arm $i$ is pulled at time $t'$ and $t' < t$. The strategy of arm $i$ is determined by a function that maps this history to a manipulation $z_{i,t}$, as $Z_{i,t} \colon h_{i,t-1} \to \mathbb{R}$. Without loss of generality, we assume that arms in the optimal arm subset $i\in S^\ast$ have strategic budgets of $0$, which restricts their $z_{i,t}$ to be $0$.
In the combinatorial setting, even with the exact reward vector $\bm\mu$ provided, it can be hard to exactly compute the optimized $r_{\bm{\mu}}(S)$. In view of this, many have studied probabilistic approximation algorithms in combinatorial problems, which indicates that an ($\alpha,\beta$)-approximation oracle defined below can be usually available.
\begin{definition}[Approximation oracle]
Let $0\le \alpha,\beta \leq 1$ and define ${OPT}_{\bm{\mu}} = \max_{S\in \mathcal{S}} r_{\bm{\mu}}(S)$.
An oracle is called an ($\alpha,\beta$)-approximation oracle if it takes an expected reward vector $\bm{\mu}$ as input and outputs an arm subset $S \in \mathcal{S}$ such that $\mathbb{P}(r_{\bm{\mu}}(S) \geq \alpha \cdot {OPT}_{\bm{\mu}}) \geq \beta$. That is, the oracle gives an arm subset $S$ that is at least as good as $\alpha$ times the reward of an optimal arm subset with probability at least $\beta$.
\end{definition}
Denote ${S_B} = \{ r_{\bm{\mu}}(S) < \alpha \cdot {OPT}_{\bm{\mu}} | S \in \mathcal{S}\}$ to be the set of suboptimal arm subsets under the approximation oracle.
Note that a suboptimal arm subset can be given by the oracle for two reasons. The ($\alpha,\beta$)-approximation oracle can fail, which happens with probability at most $1-\beta$. The estimation of $\bm{\mu}$ can deviate from the true value by a significant amount, resulting in accurate input to the oracle.
The objective of the principal is to maximize the expected cumulative reward before manipulation over the time horizon $T$. Equivalently, the principal minimizes the regret, the cumulative difference between the scaled optimal reward and expected actual reward, as defined below.
\begin{definition}[Regret]
With access to an ($\alpha,\beta$)-\textit{approximation oracle}, the regret of a combinatorial bandit algorithm for $T$ rounds is
\[
Regret_{\bm{\mu},\alpha,\beta}(T) = T \cdot \alpha \cdot \beta \cdot {OPT}_{\bm{\mu}} - \mathbb{E}\big[\sum^T_{t=1}r_{\bm{\mu}}(S_t)\big]\,,
\]
where the randomness in $\mathbb{E}[\sum^T_{t=1}r_{\bm{\mu}}(S_t)]$ involves the stochasticity of the bandit algorithm and the oracle.
\end{definition}
The objective of each strategic arm $i\in [m]\setminus S^{\ast}$, however, is to maximize the number $\sum_{t=1}^T I_{i,t}$ of times it is pulled over the time horizon. To achieve this, the arm needs to confuse the principal by deviating the emitted reward signals up to the possessed budget.
\section{Strategic Combinatorial UCB}
We now propose a variant of combinatorial upper confidence bound algorithm that is robust to strategic manipulations of rewards in Algorithm \ref{algo}. The only mild assumption we maintain is that the learning algorithm has the knowledge of the largest budget possessed among all bandits arms, i.e, $B_{max}$. This is a relaxation of the assumptions made by \citet{pmlr-v119-feng20c} on the strategic UCB algorithm, in which the learning algorithm has access to the cumulative use of budget at every time step. We start with a detailed description of the algorithm and then analyze the theoretical upper bound of regret, which enjoys $O(m \log{T} + m B_{max})$.
For each arm $i$, our algorithm maintains a counter $K_{i,t-1}$ as the total number of times arm $i$ has been pulled up to time $t-1$ and $\widetilde{\mu}_{i,t} = \frac{\sum_{j=1}^{t-1} (x_{i,j}+ z_{i,j})}{K_{i,t-1}}$ as the empirical mean estimation based on the observations. At each time step, the algorithm computes the UCB estimation $\Bar{\mu}_{i,t}= \widetilde{\mu}_{i,t} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}} + \frac{B_{max}}{K_{i,t-1}}$ for $i \in [m]$. With $\Bar{\mu}_{i,t}$, the $(\alpha, \beta)$-approximation oracle then outputs an approximately optimal arm subset $S_t \in \mathcal{S}$. The algorithm plays the return arm subset and update the counter $K_{i,t}$ and the estimation $\widetilde{\mu}_{i,t}$ accordingly.
\begin{algorithm}[htb]
\DontPrintSemicolon
{\bf Input:}
Horizon $T$, number $m$ of arms, maximum budget $B_{max}$\;
{\bf Output:} Arm subset $S_t$\;
Initialize $K_{i,0} = 0$ and $\widetilde{\mu}_{i,0} = 0$ for $i \in [m]$\;
\For{t = 1 $\to$ m}{
Play an arbitrary arm subset $S_t \in \mathcal{S}$ such that $t \in S_t$\;
$K_{i,t} = K_{i,t-1} + 1$ for $i \in S_t$\;
$\widetilde{\mu}_j = x_{i,t}+z_{i,t}$ for $i \in S_t$ \;}
\For{t = m+1 $\to$ T}{
For $i \in [m]$, compute $\Bar{\mu}_{i,t}= \widetilde{\mu}_{i,t-1} + \sqrt{\frac{ 3 \log{t} }{2K_{i,t-1}}} + \frac{B_{max}}{K_{i,t-1}} $\;
$S_t$ = {\textbf{oracle}}$(\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$\;
Play arm subset $S_t$ and update
$K_{i,t} = K_{i,t-1} + 1$ and $ \widetilde{\mu}_{i,t} = \frac{\sum_{j=1}^{t} (x_{i,j} + z_{i,j})}{K_{i,t}}$ for $i \in S_t$ \;
}
\caption{SCUCB}
\label{algo}
\end{algorithm}
We first introduce a few notations that are used in our results. Define, for any arm $i\in[m]$, the suboptimality gaps as
\begin{align*}
&\Delta^i_{min} = \alpha \cdot {OPT}_{\bm{\mu}} - \max\{r_{\bm{\mu}}(S) \mid S \in S_B, i \in S\}\,, \notag \\
&\Delta^i_{max} = \alpha \cdot {OPT}_{\bm{\mu}} - \min\{r_{\bm{\mu}}(S) \mid S \in S_B, i \in S\}\,.
\end{align*}
We then denote the maximum and minimum of the suboptimality gaps as $\Delta_{max} = \max_{i \in [m]} \Delta^i_{max}\notag $ and $\Delta_{min} = \min_{i \in [m]} \Delta^i_{min}$.
The following lemma re-establish the canonical tail bound inequality in UCB under the setting of strategic manipulations.
\begin{lemma}\label{lambda}
Let ${\Lambda}_{i,t} = \sqrt{\frac{3 \log{t} }{2K_{i,t-1}}} + \frac{\rho_{i,t-1}}{K_{i,t-1}}$, where $\rho_{i,t-1}$ is the total strategic budget spent by arm $i$ up to time $t-1$ and $K_{i,t-1}$ be the total number of pulls of arm $i$ up to time $t-1$. Define the event $E_t = \{ |\ \widetilde{\mu}_{i,t-1} - \mu_{i}| \leq \Lambda_{i,t}, \forall i \in [m]\}$, where $\mu_i$ is the true mean of arm $i$'s underlying distribution.
Then, $\mathbb{P}(\lnot E_t)\leq 2m \cdot t^{-2}$.
\end{lemma}
Armed with Lemma \ref{lambda}, we present one of our main theorems, Theorem \ref{appendixA}, which gives the regret bound of $O(\log T)$ of SCUCB. The outline of the proof follows that of CUCB by \citet{pmlr-v28-chen13a}. To complete the proof, we carefully choose $\psi_t$, which controls the trade-off between the exploration and exploitation periods.
\begin{theorem}
\label{appendixA}
The regret of the SCUCB algorithm with $m$ strategic arms in time horizon $T$ using an $(\alpha,\beta)$-approximation oracle is at most
\begin{align}
m\cdot\Delta_{max}\left( \frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}+1 \right)\,, \notag \nonumber
\end{align}
where $f^{-1}(\cdot)$ is the inverse bounded smoothness function.
\end{theorem}
\begin{proof}[Proof sketch]
We introduce a counter $N_i$ for each arm $i\in[m]$ after the $m$-round initialization and let $N_{i,t}$ be the value of $N_i$ at time $t$. We initialize $N_{i,m} = 1$. By definition, $\sum_{i \in [m]} N_{i,m} = m$. For $t > m$, the counter $N_{i,t}$ is updated as follows:
\begin{itemize}
\item If $S_t \in S_B$, then $N_{{i'},t} = N_{{i'},t-1} + 1$ where $i' =$ $\argmin_{i \in S_t}$ $N_{i,t-1}$. In the case that $i'$ is not unique, we break ties arbitrarily;
\item If $S_t \notin S_B$, then no counters will be updated.
\end{itemize}
As such, the total number of pulls of suboptimal arm subsets is less than or equal to $\sum_{i=1}^m N_{i,T}$.
Define $\psi_t = \frac{8B_{max}f^{-1}(\Delta_{min}) + 6\log{t}} {\left(f^{-1}(\Delta_{min})\right)^2} > c $, where $c$ is the larger solution of
\begin{align*}
\label{square}
&(f^{-1}(\Delta_{min}))^2 c^2 + 16B_{max}^2 - ((8B_{max} f^{-1}(\Delta_{min})- 6\log{t}) c = 0 \,.
\end{align*}
Then, we decompose the total number $\sum_{i=1}^m N_{i,T}$ of pulls of suboptimal arm subsets as
\begin{align}
\sum_{i=1}^m N_{i,T} = \ & m +\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B\} \notag \\
=\ &m +\sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} \leq \psi_t\} \notag \\
&+ \sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} > \psi_t\}\notag \\
\leq \ & m + m\psi_T+\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}\,. \notag
\end{align}
The inequality follows as $\sum_{t=m+1}^T \sum_{i=1}^m \mathbb{I}\{ S_t \in S_B, N_{i,t} > N_{i,t-1}, N_{i,t-1} \leq \psi_t\}$ can be trivially bounded by $m\psi_T$. Thus the key to bound the total number of pulls of suboptimal arm subset is to upper bound $\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}$. Let $F_t$ denotes the event that the oracle fails to provide an $\alpha$-approximate arm subset with respect to the input vector $\Bar{\bm{\mu}} = (\Bar{\mu}_{1,t},\Bar{\mu}_{2,t},.....,\Bar{\mu}_{m,t})$.
Then we can decompose $\sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}$ as
\begin{align}
&\quad \sum_{t=m+1}^T \mathbb{I}\{ S_t \in S_B,N_{i,t-1} > \psi_t, \forall i \in S_t\}\notag\\
&\leq \sum_{t=m+1}^T (\mathbb{I}\{F_t\} + \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\})\notag \\
&\leq (T-m)(1-\beta) + \mathbb{I}\{ \lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\}\notag \,. \notag
\end{align}
By leveraging the monotonicity and smoothness assumptions of the reward function, we show that $\mathbb{P}(\{\lnot F_t, E_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\}) = 0$. Therefore, by the inclusion-exclusion principle,
\begin{align}
\mathbb{P}(\{\lnot F_t, S_t \in S_B, K_{i,t-1} > \psi_t, \forall i \in S_t\})
&\leq \mathbb{P}(\lnot E_t) \leq 2m \cdot t^{-2} \,.\notag
\end{align}
Leveraging the upper bound of total number of pulls of suboptimal arm subsets, the regret is bounded by
\begin{align}
& \quad Regret_{\bm{\mu},\alpha,\beta}(T)
\notag \\
& \leq T\alpha \beta \cdot \text{OPT}_{\bm{\mu}} - \left( T\alpha\cdot \text{OPT}_{\bm{\mu}} - \mathbb{E} \left[ \sum_{i=1}^m N_{i,T} \right]\cdot \Delta_{max} \right)
\notag \\
&\leq m\cdot \Delta_{max}\left(\frac{8B_{max} f^{-1}(\Delta_{min}) + 6\log{T}} {\left(f^{-1}(\Delta_{min})\right)^2} + \frac{\pi^2}{3}+1\right)\,. \notag \qedhere
\end{align}
\end{proof}
It remains in question whether a bandit algorithm can achieve a regret upper bound sublinear in $B_{max}$.
Our conjecture is negative.
In fact, under strategic manipulations of rewards, the design of robust bandit algorithms, e.g. UCB and $\epsilon$-greedy, is analogous to the design of outlier-robust mean estimation algorithms. Existing works on robust mean estimation, such as \citet{steinhardt_et_al:LIPIcs:2018:8368}, argue that from an information theoretical point of view mean estimation error must be depending on the variance of the data. Casting this argument to bandit with the strategic manipulation setting, we believe that a tight regret bound is unlikely to be independent of the strategic budget $B_{max}$. Moreover, Proposition 4 in \citet{steinhardt_et_al:LIPIcs:2018:8368} implies that the dependency on $B_{max}$ is linear. This corresponds to our linear dependency of regret in Theorem \ref{appendixA}.
\section{Lower Bounds for strategic budget}
To fully understand the effects of strategic manipulations, we investigate the relationship between the strategic budget and the performance of UCB-based algorithms. Our results provide the dependency between an arm's strategic budget and the number of times it is pulled, which influence the regret of the algorithm incurred by the arm. Our analysis gains some insight from \citet{zuo_near_2020}, which limits the discussion to the 2-armed bandit setting and cannot be applied to the general MAB setting directly.
We first define some notations used in the theorem.
Without loss of generality, let arm $i^{\ast}$ be the optimal arm and arm $i, i \neq i^{\ast}$ be an arbitrary strategic suboptimal arm. Assume that arm $i$ has no access to the information regarding other arms. Let $K_{i,t}$ and $K_{i^{\ast},t}$ denote the number of times arm $i$ and arm $i^{\ast}$ have been pulled up to time $t$, respectively. Denote $\hat{\mu}_{i,t}$ as the empirical estimate of the underlying mean $\mu$ without manipulations, i.e., $\hat{\mu}_{i,t} = \frac{\sum^{t-1}_{j=1}x_{i,j}}{K_{i,t-1}}$. Define the suboptimality gap for each strategic arm $i$ to be $\delta_i = \mu_{i^{\ast}} - \mu_i$. We use a slightly revised UCB algorithm as the basic algorithm, where the UCB estimation term for arm $i$ is $\hat{\mu}_{i,t} + \sqrt{\frac{2\log(K_{i,t}^2/\eta^2)}{K_{i,t}}}$ and $\eta$ is a confidence parameter chosen by the algorithm.
\begin{theorem}\label{lower}
In stochastic multi-armed bandit problems, for a strategic suboptimal arm $i$ without access to other arms' information, to be pulled for $\omega(k)$ in $T$ steps under the UCB algorithm where $k\ge O(\log{T})$, the minimum strategic budget is $\omega(k)$.
\end{theorem}
This dependency of $k$ can be extended to CMAB and CUCB straightforwardly when arms within an arm subset collude. Counter-intuitively, the dependency between the number of pulls of a strategic arm and its strategic budget is linear and subsequently, this infers that for a strategic arm to manipulate the algorithm to suffer an undesired regret of $\omega(k)$, where $k\ge O(\log T)$, the strategic budget must be at least $\omega(k)$.
\section{Experiments}
In this section, we evaluate our SCUCB algorithm empirically on synthetic data and three real applications, namely online worker selection in crowdsourcing, online recommendation, and online influence maximization. We highlight the best performance among all algorithms with bold text.
\subsection{Baseline Algorithms}
We compare our proposed SCUCB algorithm with both stochastic and adversarial bandits algorithms that achieves optimal asymptotic regret in their settings.
\begin{enumerate}
\allowdisplaybreaks
\item CUCB \cite{pmlr-v28-chen13a}. CUCB is the naive counterpart of our algorithm. The algorithm calculates an upper confidence interval for each arm and picks the best arm subset with the highest upper confidence interval.
\item TSCB \cite{wang2018thompson}. TSCB is the combinatorial version of the classical Thompson sampling algorithm. The algorithm maintains a prior beta distribution estimation for each arm and updates according to the received reward. At each time, the algorithm samples from the estimated distributions and pick actions according to the highest sample.
\item Combinatorial variant of EXP3 \cite{auer_nonstochastic_2002}. The algorithm maintains a weight for each arm and draws actions according to the normalized weight distribution. Upon receiving rewards, the algorithm update weight according to the classical EXP3 update rule.
\end{enumerate}
\subsection{Synthetic Experiments}
We conduct experiments presented in this section with synthetic data and compare our proposed algorithm with its naive counterpart. The approximation oracle is designed to succeed with probability $1$. Each bandit arm is modeled to follow a Bernoulli distribution with randomly populated $\mu \in [0,1]$ and all arms adapt LSI strategy.
The arms in the optimal arm subset have a strategic budget of $0$ since an additional strategic budget for optimal arms would only boost the performance of our algorithm. All other arms are equipped with a randomly allocated budget $B$, $ 0 \leq B \leq B_{max}$ by definition.
As is typical in the bandit literature, for example in \citet{auer2002finite}, we evaluate both the naive and strategic CUCB algorithms on their tuned versions, where the UCB exploration parameter is scaled by a constant factor $\gamma \in [0,1]$. To ensure reproducible results, each experiment is repeated for $10$ random seeds and the averaged result is presented.
The first set of experiment is conducted with $K = \{10, 20\}$ arms through a time horizon of $T=5000$ time steps with maximum possible budget of $B_{max} = \{70, 90, 110, 130\}$. The algorithms are asked to select $\text{Action size} = 2$ arms as an arm subset at each time step. The cumulative regret incurred by CUCB and SCUCB algorithm is presented in the table \ref{table:1} where the best performance is highlighted in bold text. Clearly, SCUCB demonstrated its effectiveness as it achieves significantly smaller regrets in various possible maximum strategic budgets.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 171.74 & 187.51 & 259.04 & 256.66 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{143.85} & \textbf{172.57} & \textbf{208.53} & \textbf{233.57} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Action size = 2, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 70 & 90 & 110 & 130 \\
\multicolumn{1}{l|}{CUCB} & 434.82 & 492.02 & 520.97 & 549.57 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{301.57} & \textbf{365.94} & \textbf{450.10} & \textbf{505.453} \\ \bottomrule
\end{tabularx}
\caption{Cumulative regret achieved by CUCB and SCUCB with synthetic data with various action size.}\label{table:1}
\end{table}
The next two tables reveals the advantage of SCUCB algorithm with various size of action set. The experiments are conducted with time horizon of $T=5000$ time steps with $K = \{10, 20\}$ arms and maximum possible budget of $B_{max} = \{50\}$. The algorithms are asked to select $\text{Action size} = \{2, 4, 6, 8\}$ arms as an arm subset at each time step. Once again, our SCUCB outperforms its naive counterpart by achieving much smaller cumulative regret,
the best numerical results across algorithms are highlighted in bold text.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 10} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 140.21 & 150.88 & 158.22 & 123.91 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{105.58} & \textbf{113.63} & \textbf{103.42} & \textbf{88.72} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYY@{}}
\toprule
\multicolumn{5}{c}{Cumulative regret, Bmax = 50, K = 20} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{4}{c}{Synethetic} \\ \midrule
\multicolumn{1}{l|}{Action Size} & 2 & 4 & 6 & 8 \\
\multicolumn{1}{l|}{CUCB} & 302.10 & 316.09 & 340.52 & 328.80 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{232.23} & \textbf{247.32} & \textbf{268.27} & \textbf{306.77} \\ \bottomrule
\end{tabularx}
\caption{Cumulative regret achieved by UCB and SCUCB with synthetic data with various $B_{max}$.}
\end{table}
To compare the two algorithms in detail, we plot the cumulative regret incurred by algorithms against the time steps. It becomes apparent that the SCUCB algorithm features a cumulative regret versus time steps line that is much smaller than the one caused by CUCB even under a higher maximum possible strategic budget. We compare the two algorithm under settings where the maximize possible budget is $B_{max} = 70, 90$ and $B_{max} = 110, 130$.
\begin{figure}[H]
\centering
\subfigure[]{
\label{Fig1}
\includegraphics[width=3.8cm]{figs/B7090_K20_N10000.png}}
\subfigure[]{
\label{Fig2}
\includegraphics[width=3.8cm]{figs/B110130_K20_N10000.png}}
\caption{\ref{Fig1} Comparison of CUCB and SCUCB with $B_{max} = 70, 90$. \ref{Fig2} Comparison of CUCB and SCUCB with $B_{max} = 110, 130$.}
\end{figure}
\subsection{Online Worker Selection in Crowdsourcing Systems}
We simulated an online crowdsourcing system that has a workflow resembled by figure \ref{crowdsys} and model it as a combinatorial bandits problem. We performed an extensive empirical analysis of our SCUCB algorithm against CUCB, the combinatorial version of Thompson sampling, and EXP3. The experiments were conducted with the Amazon sentiment dataset, where Amazon item reviews are labeled 'is/is not book' or 'is/is not negative \cite{Krivosheev:2018:CCM:3290265.3274366}. We split the dataset into two, where 'is book' contains data that only has labels 'is/is not book' and 'is negative' contains data that only has labels 'is/is not negative'. Both datasets consists of 7803 reviews, 284 workers, and 1011 tasks.
The correctness of workers and tasks can be visualized by the figures as below. Notice that most of the workers attain an accuracy of $60$ percent and above. This highlights that most of the workers we are interacting with has the ability to label most of the tasks correctly.
\begin{figure}[H]
\centering
\subfigure[]{
\label{worker}
\includegraphics[width=3.8cm]{figs/worker_visual.png}}
\subfigure[]{
\label{task}
\includegraphics[scale = 0.28]{figs/task_visual.png}}
\caption{\ref{worker} Visualization of correctness of worker's labels. \ref{task}Visualization of correctness of task's labels.}
\end{figure}
For each task, the data contains responses from 5 workers and the crowdsourcing platform is asked to select 2, 3, or 4 workers at each time step. To measure the performance of the algorithms, we made the assumption that the reward is monotonically correlated to the accuracy of the label compared to the ground truth label. Thus, we define the reward to be 1, if the worker's label is the same as ground truth label and 0 otherwise. To model the strategic behavior of the workers, we model each worker with a randomly allocated strategic level $s \in [0,1]$ such that the worker provides its honest response with probability $s$.
For the first set of experiments, we choose the maximum possible strategic budget to be $B_{max} = 500$. For actions size of $2, 3, 4$. We obtain the following cumulative rewards and cumulative regret results where the best performance is highlighted in bold text. To ensure reproducible results, each experiment is repeated 5 times and the averaged result is presented.
To further investigate the effect of the maximum possible strategic budget on the performance of algorithms, we fix the size of action size to be 2 and plot the cumulative reward/cumulative regret incurred by each algorithm again $B_{max} \in [100, 700]$. Regardless of the $B_{max}$ value, our proposed SCUCB algorithm consistently outperforms other methods. For better visualization, we omit the line for combinatorial Thompson sampling due to its relatively weaker performance.
\begin{figure}[H]
\centering
\subfigure[]{
\label{culre}
\includegraphics[width=3.8cm]{figs/culre.png}}
\subfigure[]{
\label{culreg}
\includegraphics[width=3.8cm]{figs/reg.png}}
\caption{\ref{culre} Cumulative rewards attained by algorithms with various level of $B_{max}$. \ref{culreg} Cumulative regrets attained by algorithms with various level of $B_{max}$. }
\end{figure}
The results of our crowdsourcing experiments reveal the robustness of our SCUCB algorithm under strategic manipulations. It also indicates the performance of adversarial and stochastic bandits algorithms under strategic manipulations. The combinatorial variant of Thompson sampling is vulnerable under manipulations, which agrees with our expectation given that the algorithm is a Bayesian algorithm in nature and is heavily relying on estimating the underlying distribution of arms. Its estimation can be easily perturbed under strategic manipulation, and thus leads to undesired performance results. It is also expected that the SCUCB algorithm far outperforms its naive counterpart, CUCB algorithm. The combinatorial version of EXP3 algorithm is the most competitive algorithm with our SCUCB. As the EXP3 algorithm was originally designed for a pure adversarial setting, the algorithm has some robustness under strategic manipulations. Shown in the Figure \ref{culre} and Figure \ref{culreg}, the combinatorial EXP3 algorithm consistently shows robustness across various levels of maximum possible strategic budget. However, as our SCUCB algorithm considers the maximum possible strategic budget and thus more adaptive, SCUCB far outperforms EXP3CB as $B_{max}$ increases.
\begin{table}[H]
\label{first}
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31672.6 & 46181.2 & 68244.6 \\
\multicolumn{1}{l|}{TSCB} & 19853.0 & 34226.2 & 60621.2 \\
\multicolumn{1}{l|}{EXP3} & 31569.6 & 46570.6 & 68482.2 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{32082.6} & \textbf{46623.6} & \textbf{68524.6} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_book} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 27787.4 & 42198.8 & 48125.4 \\
\multicolumn{1}{l|}{TSCB} & 39607.0 & 54153.8 & 55748.8 \\
\multicolumn{1}{l|}{EXP3} & 27890.4 & 41809.4 & 47887.8 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{27377.4} & \textbf{41756.4} & \textbf{47845.4} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Reward), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 31108.8 & 44757.6 & 59424.8 \\
\multicolumn{1}{l|}{TSCB} & 17240.6 & 32010.4 & 51109.0 \\
\multicolumn{1}{l|}{EXP3} & 30879.0 & 44767.2 & 58207.8 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{31759.0} & \textbf{44843.4} & \textbf{59441.4} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Crowdsourcing (Cumulative Regret), Bmax = 500} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{is\_negative} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 3 & 4 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 28411.2 & 43566.2 & 56465.2 \\
\multicolumn{1}{l|}{TSCB} & 42279.4 & 56279.6 & 64781.0 \\
\multicolumn{1}{l|}{EXP3} & 28641.0 & 43523.4 & 57682.2 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{27761.0} & \textbf{43446.6} & \textbf{56448.6} \\ \bottomrule
\end{tabularx}
\caption{Cumulative rewards and cumulative regret attained by SCUCB and baseline algorithms with various action sizes on isbook dataset and isnegative dataset.}
\end{table}
\subsection{Online Recommendation System}
The online recommendation is a classic example of combinatorial bandits in applications. We evaluated algorithms in the latest MovieLens dataset which contains 9742 movies and 100837 ratings of the movies with each rating being $1 - 5$ \cite{10.1145/2827872}. We compare the algorithms based on the total values of ratings received, i.e. recommending a movie and receiving a rating of 5 is more desired than recommending a movie and receiving a rating of 1.
As the dataset may consist of unbalanced data for each movie, we adapted collaborative filtering and $k$-means clustering into our evaluation system. We used the collaborative filtered for training and testing and instead of recommending movies, we clustered the movies by $k$-means clustering and asked the bandits algorithm to choose one cluster at each time. The movie with the highest rating in the chosen cluster is recommended. The workflow of the experiment setup is summarized in the above diagram.
We performed the evaluation with various number of clusters (10, 20 or 30) and $B_{max} = 30, 50, 70$ through a time horizon of $T = 500$. The results presented below are averages over 5 runs to ensure reproducibility. The best performance across algorithms is highlighted in bold text. From the tables below, we conclude the effectiveness of SCUCB algorithm.
\begin{table}[tbh]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 10 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2297.60 & 2494.36 & 2631.02 \\
\multicolumn{1}{l|}{TS} & 2160.14 & 2314.76 & 2468.26 \\
\multicolumn{1}{l|}{EXP3} & 2181.11 & 2449.28 & 2430.96 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2305.75} & \textbf{2314.76} & \textbf{2636.07} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 20 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\ \midrule
\multicolumn{1}{l|}{UCB} & 2469.93 & 2336.92 & 2159.25 \\
\multicolumn{1}{l|}{TS} & 2260.20 & 1850.55 & 1905.42 \\
\multicolumn{1}{l|}{EXP3} & 2380.74 & 2317.82 & 2025.39 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2474.69} & \textbf{2341.19} & \textbf{2177.70} \\ \bottomrule
\end{tabularx}
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{Recsys (Cumulative Reward), T=500, 30 clusters} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Movielens} \\ \midrule
\multicolumn{1}{l|}{Bmax} & 30 & 50 & 70 \\\midrule
\multicolumn{1}{l|}{UCB} & 2443.57 & 2393.88 & 2472.31 \\
\multicolumn{1}{l|}{TS} & 2132.16 & 2248.56 & 1855.14 \\
\multicolumn{1}{l|}{EXP3} & 2368.42 & 2265.32 & 2339.49 \\
\multicolumn{1}{l|}{SUCB} & \textbf{2436.43} & \textbf{2397.72} & \textbf{2476.52} \\ \bottomrule
\end{tabularx}
\caption{Cumulative rewards attained by algorithms with different number of clusters $10, 20, 30$ and different level of $\{B_{max} = 30, 50, 70\}$ across 500 time steps. }
\end{table}
\begin{figure}[tbh]
\centering
\includegraphics[width=7cm]{figs/arch.png}
\caption{Experiment set up of online recommendation.}
\end{figure}
\subsection{Online Influence Maximization}
We implemented the online influence maximization with an offline influence maximization algorithm TIM as our oracle \cite{10.1145/2588555.2593670}. TIM is one of the offline influence maximization algorithms that achieve asymptotic optimality. We perform the experiments on two datasets. One is a simulation dataset with 16 nodes and 44 edges, the other is a 100 nodes subset from Digg dataset, where each node represents a user from Digg website \cite{nr}. We simulate the connectivity of edges at each time step by randomly assigning each edge a connectivity probability at the start of the experiment.
\begin{figure}[hbt!]
\centering
\includegraphics[width=0.3\textwidth]{figs/test_graph.png}
\caption{Visualization of synthetic graph for evaluation.}
\label{test_graph}
\end{figure}
\begin{figure}[hbt!]
\centering
\includegraphics[width=0.35\textwidth]{figs/test.png}
\caption{ Mean rewards attained by algorithms.}\label{16c2}
\end{figure}
Figure \ref{test_graph} visualizes the synthetic graph we created for evaluation. We then tested the algorithms over $T = 800$ time steps with action size of 2 and maximum possible strategic budget of $B_{max} = 300$. Figure \ref{16c2} shows the effectiveness of algorithms by evaluating based on the averaged nodes influenced. To ensure reproducible results, we calculate the mean of influence spread over 100 trials. For better visualization, we kept a running mean of over 50 time steps to smooth the lines.
We then investigate the effect of action size and the results are summarized in the following tables \ref{table:5}.
\begin{table}[H]
\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YYYY@{}}
\toprule
\multicolumn{5}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{Synthetic, 16 nodes, 44 edges} \\ \midrule
\multicolumn{1}{l|}{Action size} & 2 & 4 & 6 & 8 \\ \midrule
\multicolumn{1}{c|}{CUCB} & 5.46 & 6.54 & 10.4 & 12.5 \\
\multicolumn{1}{c|}{TSCB} & 6.7 & 8.50 & 10.44 & 12.7 \\
\multicolumn{1}{c|}{EXP3CB} & 5.86 & 8.14 & 10.34 & 12.46 \\
\multicolumn{1}{c|}{SCUCB} & \textbf{7.44} & \textbf{9.32} & \textbf{11.16} & \textbf{13.04} \\ \bottomrule
\end{tabularx}
\end{table}
\begin{table}[H]\centering
\begin{tabularx}{0.47\textwidth}{@{}lY@{\extracolsep{\fill}}YY@{}}
\toprule
\multicolumn{4}{c}{OIM (Averaged Final Reward), Bmax = 200} \\ \midrule
\multicolumn{1}{l|}{Dataset} & \multicolumn{3}{c}{digg, 100 nodes} \\ \midrule
\multicolumn{1}{l|}{Action size} & 10 & 15 & 20 \\ \midrule
\multicolumn{1}{l|}{CUCB} & 15.64 & 20.64 & 25.32 \\
\multicolumn{1}{l|}{TSCB} & 16.30 & 21.62 & 26.08 \\
\multicolumn{1}{l|}{EXP3} & 12.16 & 17.76 & 21.72 \\
\multicolumn{1}{l|}{SCUCB} & \textbf{16.56} & \textbf{22.16} & \textbf{26.14} \\ \bottomrule
\end{tabularx}
\caption{Averaged final rewards attained by algorithms with synthetic and Digg dataset with action size of 2, 800 time steps and $B_{max} = 200$.}\label{table:5}
\end{table}
\section{Conclusion}
We investigate the problem of combinatorial MAB under strategic manipulations of rewards. We propose a variant of the UCB algorithm, SCUCB, which attains a regret at most $O(m \log T + m B_{max})$.
Compared to previous studies on the bandit problems under strategic manipulations, we relax on the assumption that the algorithm has access to the cumulative strategic budget spent by each arm at each time step.
For the robustness of bandit algorithms, we present lower bounds on the strategic budget for a malicious arm to incur a $\omega(\log T)$ regret of the bandit algorithm.
We provide extensive empirical results on both synthetic and real datasets with a range of applications to verify the effectiveness of the proposed algorithm. Our algorithm consistently outperforms baseline algorithms that were designed for stochastic and adversarial settings.
|
\section*{References}}
\usepackage{amssymb}
\usepackage{amsthm}
\usepackage{amsmath,amssymb,amsopn,amsfonts,mathrsfs,amsbsy,amscd}
\usepackage{longtable}
\usepackage{caption}
\usepackage{multirow}
\usepackage{lineno}
\setlength{\parindent}{0pt}
\newcommand{\we}{\wedge}
\newcommand{\signa}{({\mathbf s},{\mathbf t})}
\newcommand{\g}{\mathfrak{g}}
\newcommand{\p}{{\mathfrak{p} }}
\newcommand{\q}{{\mathfrak{q} }}
\newcommand{\prs}{\langle\;,\;\rangle}
\newcommand{\im}{\mathrm{Im}}
\newcommand{\too}{\longrightarrow}
\newcommand{\om}{\omega}
\newcommand{\esp}{\quad\mbox{and}\quad}
\newcommand{\an}{{\pi_{\#}}}
\newcommand{\met}{ metric contravariant connection
associated to $(\pi,<,{\rangle})$}
\newcommand{\fr}{\mathfrak}
\def\br{[\;,\;]}
\newcommand{\e}{\dot{e}}
\newcommand{\Se}{{\cal S}}
\newcommand{\Ua}{{\cal U}}
\newcommand{\X}{{\cal X}}
\newcommand{\G}{{\mathfrak{g}}}
\newcommand{\U}{{\mathfrak{U}}}
\newcommand{\A}{{\mathcal{A}}}
\newcommand{\F}{{\cal F}}
\newcommand{\h}{{\mathfrak{h}}}
\newcommand{\ad}{{\mathrm{ad}}}
\newcommand{\tr}{{\mathrm{tr}}}
\newcommand{\ric}{{\mathrm{ric}}}
\newcommand{\Li}{{\cal L}}
\newcommand{\spa}{{\mathrm{span}}}
\newcommand{\B}{{\cal B}}
\newcommand{\M}{{\cal M}}
\newcommand{\C}{{\cal C}}
\newcommand{\D}{{\cal D}}
\newcommand{\E}{{\cal E}}
\newcommand{\vi}{\emptyset}
\newcommand{\di}{\displaystyle}
\newcommand{\Om}{\Omega}
\newcommand{\na}{\nabla}
\newcommand{\si}{\sigma}
\newcommand{\al}{\alpha}
\newcommand{\be}{\beta}
\newcommand{\ga}{\gamma}
\newcommand{\Ga}{\Gamma}
\newcommand{\vare}{\varepsilon}
\newcommand{\ch}{\check{e}}
\newcommand{\la}{\lambda}
\newcommand{\De}{\Delta}
\newcommand{\de}{\delta}
\newcommand{\R}{\mathbb{R}}
\newcommand{\entraine}{\Longrightarrow}
\newcommand{\inj}{\hookrightarrow}
\newcommand{\recip}{\Longleftarrow}
\newcommand{\ssi}{\Longleftrightarrow}
\newtheorem{Def}{Definition}[section]
\newtheorem{theo}{Theorem}[section]
\newtheorem{pr}{Proposition}[section]
\newtheorem{Le}{Lemma}[section]
\newtheorem{co}{Corollary}[section]
\newtheorem{exo}{Exercice}
\newtheorem{exem}{Example}
\newtheorem{remark}{Remark}
\newtheorem{idee}{Idee}
\newtheorem{Formule}{Formula}
\newtheorem{Resultat}{Result}
\newtheorem{problem}{Problem}
\newtheorem{conjecture}{Conjecture}
\newenvironment{eqgauche
\stepcounter{equation}
\begin{equation*}}
\leqno(\arabic{equation})
\end{equation*}
}
\begin{document}
\begin{frontmatter}
\title{The classification of left-invariant para-K\"ahler structures on simply connected four-dimensional Lie groups
}
\author[]{ M. W. Mansouri and A. Oufkou}
\address[]{Universit\'e Ibno Tofail\\ Facult\'e des Sciences. K\'enitra-Maroc\\e-mail: [email protected]\\ <EMAIL>}
\begin{abstract} We give a complete classification of left
invariant para-K\"ahler structures on four-dimensional simply connected
Lie groups up to an automorphism. As an application we discuss some curvatures properties of the canonical connection associated to these structures as flat, Ricci flat and existence of Ricci solitons.
\end{abstract}
\begin{keyword}
Symplectic Lie algebras, para-K\"ahler structures, Ricci soliton\\
\MSC 53D05 \sep \MSC 53C30 \sep \MSC 53C25.
\end{keyword}
\end{frontmatter}
\section{Introduction and main results}
An almost para-complex structure on a $2n$-dimensional manifold $M$ is a field $K$ of endomorphisms of the tangent
bundle $TM$ such that $K^2 = Id_{TM}$ and the two eigendistributions $T^{\pm}M := ker(Id\pm K)$ have the same rank. An almost
para-complex structure $K$ is said to be integrable if the distributions $T^{\pm}M$ are involutive. This is equivalent to the
vanishing of the Nijenhuis tensor $N_K$ defined by
\[N_K(X, Y ) = [X,Y] + [KX,KY] -K[KX,Y]-K[X,KY ],\]
for vector fields $X$, $Y$ on $M$. In such a case $K$ is called a para-complex structure.
A para-K\"ahler structure on a manifold $M$ is a pair $(\langle .,.\rangle,K)$ where $\langle .,.\rangle$ is a
pseudo-Riemannian metric and $K$ is a parallel skew-symmetric para-complex
structure. If $(\langle .,.\rangle,K)$ is a para-K\"ahler structure on $M$, then $\om=\langle .,.\rangle\circ K$ is a symplectic
structure and the $\pm 1-$eigendistributions $T^\pm M$ of $K$ are two integrable
$\om$-Lagrangian distributions. Due to this, a para-K\"ahler structure can be
identified with a bi-Lagrangian structure $(\om, T^\pm M)$ where $\om$ is a symplectic
structure and $T^\pm M$ are two integrable Lagrangian distributions. Moreover the Levi-Civita connection associate to
neutral metric $\langle .,.\rangle$ coincides with the canonical connection associate to bi-Lagrangian structure (the unique
symplectic connection with parallelizes both foliations \cite{He}). For a survey on paracomplex geometry see [6] and for
background on bi-Lagrangian structures and their associated
connections, the survey \cite{E-S-T} is a good starting point and contains
further references (See as well \cite{A-M-T} and \cite{Bo}).
Suppose now that $M$ is a Lie group $G$ and $\omega$, $\langle .,.\rangle$ and $K$ are left invariant. If we denote
by $\g$ the Lie algebras of $G$, then $(\langle .,.\rangle,K)$ is determined by is restrictions to the Lie algebra $\g$. In this situation, $(\g, \langle .,.\rangle_e, K_e)$ or $(\g, \omega_e, K_e)$ is called a para-K\"ahler Lie algebra (e is unit of
$G$), in the rest of this paper a para-K\"ahler Lie algebra will be noted $(\g,\omega,K)$. Recall that two para-K\"ahler Lie algebras $(\g_1,\omega_1,K_1)$ and $(\g_2,\omega_2,K_2)$ are said to be
equivalent if there exists an isomorphism of Lie algebras
$T :\g_1\longrightarrow \g_1$ such as $T^*\omega_2=\omega_1$ and $T_*K_1=K_2$.
Para-K\"ahler (bi-Lagrangian) structures on Lie algebras in general have been studied, for example, in \cite{Ba}, \cite{B-B} and \cite{B-M}. In \cite{Ha}, there is a study the
existences of bi-Lagrangian structures on symplectic nilpotent Lie algebras of dimension $2$,$4$ and $6$.
A first classification of para-K\"ahler structures on four-dimensional Lie algebras was obtained by Calvaruso in
\cite{Ca}. Another description based on the classification of symplectic Lie algebras is proposed by Smolentsev and Shagabudinova in \cite{S-S}.
Benayadi and Boucetta provide in \cite{B-B} a new characterization of para-K\"ahler Lie algebras using left symmetric bialgebras inroduced by Bai in \cite{Ba}. Based on this characterization we propose in this paper, the classification of para-K\"ahler structures on four-dimensional Lie algebras. Notice that our classification is more complete and precise than the other classifications existing in the literature.
\textit{Notations}: For $\{e_1, e_2,e_3,e_4\}$ a basis of $\G$, we denote by $\{e^1, e^2, e^3,e^4\}$ the
dual basis on $\G^\ast$ and $e^{ij}$ the two-form $e^i\wedge e^j$, $\e^{ij}$ is the symmetric two-form $e^i \odot e^j$ and $E_{ij}$ is the endomorphism which sends $e_j$ to $e_i$ and vanishes on $e_k$ for $k\not=j$.
The para-K\"ahler Lie algebras $(\g, \langle .,.\rangle, K)$ is necessarily symplectic Lie algebra $(\g,\om)$. It is well known that a symplectic four-dimensional Lie
algebra is necessarily solvable. The classification of symplectic
four-dimensional Lie algebras $(\G,\omega)$ is given by the following
Table (see \cite{O}).
{\renewcommand*{\arraystretch}{0.9 }
\begin{center}
\begin{longtable}{llc}
\hline
Case & No vanishing brackets & $\omega$ \\
\hline
$\mathfrak{r} \mathfrak{h}_3$ &$[e_1,e_2]=e_3$&$e^{14}+e^{23}$ \\
\hline
$\mathfrak{r}\mathfrak{r}_{3,0}$ & $[e_1,e_2]= e_2$&$e^{12}+e^{34}$ \\
\hline
$\mathfrak{r}\mathfrak{r}_{3,-1}$ & $[e_1,e_2]= e_2$, $[e_1, e_3] =-e_3$&$e^{14}+e^{23}$ \\
\hline
$\mathfrak{r}\mathfrak{r}^\prime_{3,0}$ &$[e_1,e_2]=-e_3$, $[e_1,e_3]=e_2$&$e^{14}+e^{23}$ \\
\hline
$\mathfrak{r}_2\tau_2$ &$[e_1,e_2] =e_2$, $[e_3,e_4]=e_4$&$e^{12}+\mu e^{13}+e^{34}$ \\
\hline
\multirow{2}{*}{$\mathfrak{r}^\prime_2$} &$[e_1,e_3]= e_3$, $[e_1,e_4]=e_4$, &\multirow{2}{*}{$e^{14}+e^{23}$} \\
&$[e_2,e_3] = e_4$, $[e_2,e_4] = -e_3$&\\
\hline
$\mathfrak{n}_4$& $[e_4,e_1]=e_2$, $[e_4,e_2]=e_3$ &$e^{12}+e^{34}$ \\
\hline
$\mathfrak{r}_{4,0}$ &$[e_4,e_1]=e_1$, $[e_4,e_3]=e_2$ &$e^{14}\mp e^{23}$ \\
\hline
$\mathfrak{r}_{4,-1}$ &$[e_4,e_1]=e_1$, $[e_4,e_2]=-e_2$, $[e_4,e_3]=e_2-e_3$&$e^{13}+e^{24}$ \\
\hline
$\mathfrak{r}_{4,-1,\beta}$ &$[e_4,e_1]=e_1$, $[e_4,e_2]=-e_2$, $[e_4,e_3]=\beta e_3$&$e^{12}+e^{34}$ \\
\hline
$\mathfrak{r}_{4,\alpha,-\alpha}$&$[e_4,e_1]=e_1$, $[e_4,e_2]=\alpha e_2$, $[e_4,e_3]=-\alpha e_3$& $e^{14}+e^{23}$\\
\hline
$\mathfrak{r}^\prime_{4,0,\delta}$ &$ [e_4, e_1]=e_1$, $[e_4, e_2]=-\delta e_3$, $[e_4, e_3]=\delta e_2$&$e^{14}\mp e^{23}$ \\
\hline
\multirow{2}{*}{$\mathfrak{d}_{4,1}$}&$[e_1,e_2]=e_3$, $[e_4,e_3]= e_3$, &$e^{12}-e^{34}$\\
&$[e_4, e_1]=e_1$&$e^{12}-e^{34}+e^{24}$\\
\hline
\multirow{2}{*}{$\mathfrak{d}_{4,2}$}&$[e_1,e_2]=e_3$, $[e_4,e_3]= e_3$,& $e^{12}-e^{34}$\\
& $[e_4, e_1]=2e_1$, $[e_4, e_2]=-e_2$&$e^{14}\mp e^{23}$\\
\hline
\multirow{2}{*}{$\mathfrak{d}_{4,\lambda}$}&$[e_1,e_2]=e_3$, $[e_4,e_3]= e_3$, &\multirow{2}{*}{$e^{12}-e^{34}$} \\
&$[e_4, e_1] =\lambda e_1$, $[e_4, e_2]= (1-\lambda)e_2$&\\
\hline
\multirow{2}{*}{$\mathfrak{d}^\prime_{4,\delta}$}&$[e_1,e_2] = e_3$, $[e_4, e_1] = \frac{\delta}{2}e_1-e_2$,&\multirow{2}{*}{$\mp(e^{12}-\delta e^{34})$}\\
& $[e_4,e_3]=\delta e_3$, $[e_4,e_2] = e_1+\frac{\delta}{2}e_2$&\\
\hline
\multirow{2}{*}{$\mathfrak{h}_4$}&$[e_1,e_2]=e_3$, $[e_4, e_3]=e_3$, &\multirow{2}{*}{$\mp(e^{12}-e^{34})$}\\
&$[e_4, e_1]=\frac{1}{2}e_1$, $[e_4,e_2] = e_1 +\frac{1}{2}e_2$&\\
\hline
\caption{ Symplectic four-dimensional Lie algebras\\ $(\mu\geq0$, $-1\leq\beta<1$, $-1<\alpha<0$, $\delta>0$ and $\lambda\geq\frac{1}{2}$, $\lambda\not=1,2 ) $.}
\label{tab1}
\end{longtable}
\end{center}
Our main result is the following.
\begin{theo}\label{1-1}Let $(\g,\om,K)$ be a four-dimensional para-K\"ahler Lie algebra. Then $(\g,\om,K)$ is isomorphic to one of the following Lie algebras with the given para-K\"ahler structures:
\textbf{Lie algebra $\mathfrak{r} \mathfrak{h}_3$}
\begin{enumerate}
\item[]For $\om=e^{14}+e^{23}$
\begin{itemize}
\item[]$K_1=-E_{11}+E_{21}+E_{22}-E_{33}-E_{43}+E_{44}$
\item[]$K_2=\mp(E_{11}-E_{22}+E_{33}-E_{44})$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{r} \mathfrak{r}_{3,0}$}
\begin{enumerate}
\item[]For $\om=e^{12}+e^{34}$
\begin{itemize}
\item[]$K_1=-E_{11}+E_{22}-E_{33}+E_{44}$
\item[]$K_2=E_{11}+xE_{12}-E_{22}+E_{33}-E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{r}\mathfrak{r}_{3,-1}$}
\begin{enumerate}
\item[]For $\om=e^{14}+e^{23}$
\begin{itemize}
\item[]$K_1=\mp(E_{11}+E_{22}\pm E_{23}-E_{33}-E_{44})$
\item[]$K_2=E_{11}+E_{22}+xE_{14}-E_{33}-E_{44}$
\item[]$K_{3}=E_{11}-E_{22}\pm E_{32}+E_{33}-E_{44}$
\item[]$K_{4}=E_{11}-E_{22}+E_{33}-E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{r}_{2}\mathfrak{r}_{2}$}
\begin{enumerate}
\item[]For $\omega=e^{12}+\mu e^{13}+e^{34},\; (\mu>0)$
\begin{itemize}
\item[]$K_1=-E_{11}+E_{22}+E_{33}-E_{44}$
\item[]$K_2=\mp(E_{11}-2E_{13}-E_{22}-E_{33}+2E_{42}+E_{44})$
\end{itemize}
\item[]For $\omega=e^{12}+e^{34}$
\begin{itemize}
\item[]$K_1=E_{11}-E_{22}-2E_{24}+2E_{31}-E_{33}+E_{44}$
\item[]$K_2=-E_{11}+xE_{12}+E_{13}+xE_{14}+E_{22}+xE_{32}+xE_{34}-E_{42}+\frac{1}{x}E_{43}$
\item[]$K_3=-E_{11}+E_{22}-E_{33}+E_{44}$
\item[]$K_4=-E_{11}+E_{22}+E_{33}+xE_{43}-E_{44}$
\item[]$K_5=E_{11}-2E_{13}-E_{22}-E_{33}+2E_{42}+E_{44}$
\item[]$K_6=E_{11}+xE_{12}-E_{22}+E_{33}+yE_{34}-E_{44}$
\item[]$K_7=E_{11}+xE_{12}+xE_{14}-E_{22}+xE_{32}+E_{33}+xE_{34}-E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{r}_{2}'$}
\begin{enumerate}
\item[]For $\omega=e^{14}+e^{23} $
\begin{itemize}
\item[]$K_{1}=E_{11}+xE_{14}-E_{22}-\frac{4}{x}E_{32}+E_{33}-E_{44}$
\item[]$K_{2}=-E_{11}-E_{22}+E_{33}+E_{44}$
\item[]$K_{3}=xE_{11}+2yE_{12}+(1-x)E_{13}-2yE_{14}-2yE_{21}+xE_{22}+2yE_{23}+(1-x)E_{24}+(1+x)E_{31}+2yE_{32}-xE_{33}-2yE_{34}-2yE_{41}+(1+x)E_{42}+2yE_{43}-xE_{44}$
\item[]$K_{4}=E_{11}+xE_{14}+E_{22}-E_{33}-E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{r}_{4,0}$}
\begin{enumerate}
\item[]For $\omega=e^{14}+ e^{23}$ or $\omega=e^{14}-e^{23}$
\begin{itemize}
\item[]$K=\mp(E_{11}-E_{22}+E_{33}-E_{44})$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{r}_{4,-1}$}
\begin{enumerate}
\item[]For $\omega=e^{13}+e^{24}$
\begin{itemize}
\item[]$K_{1}=E_{11}+xE_{13}-E_{22}-E_{33}+E_{44}$
\item[]$K_{2}=-E_{11}+E_{22}+E_{33}-E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{r}_{4,-1,\beta} \;(-1<\beta<1)$}
\begin{enumerate}
\item []For $\omega =e^{12}+e^{34}$
\begin{itemize}
\item []$K_{1}=E_{11}\mp E_{12}-E_{22}-E_{33}+E_{44}$
\item []$K_{2}=E_{11}-E_{22}\mp E_{33}\pm E_{44}$
\item []$K_{3}=E_{11}-E_{22}+xE_{34}+\frac{1}{x}E_{43}$
\item []$K_{4}=E_{11}-E_{22}+E_{33}-E_{44}$
\item []$K_{5}=-E_{11}\mp E_{12}+E_{22}+E_{33}-E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{r}_{4,-1,-1}$}
\begin{enumerate}
\item []For $\omega=e^{12}+e^{34} $
\begin{itemize}
\item []$K_{1}=-E_{11}+xE_{21}+E_{22}+E_{23}-E_{33}+E_{41}+E_{44}$
\item []$K_{2}=-E_{11}+E_{21}+E_{22}-E_{33}+E_{44}$
\item []$K_{3}=-E_{11}-E_{21}+E_{22}-E_{33}+E_{44}$
\item []$K_{4}=-E_{11}+E_{22}-E_{33}+xE_{43}+E_{44}$
\item []$K_{5}=E_{11}\mp E_{12}-E_{22}-E_{33}+E_{44}$
\item []$K_{6}=E_{11}-E_{22}-E_{33}+xE_{43}+E_{44}$
\item []$K_{7}=-E_{11}+E_{22}+E_{33}-E_{44}$
\item []$K_{8}=E_{11}\mp E_{21}-E_{22}+E_{33}-E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{r}_{4,\al,-\al}$}\;($-1<\al<0$)
\begin{enumerate}
\item [] For $\omega=e^{14}+e^{23} $
\begin{itemize}
\item [] $K_{1}=-E_{11}+E_{22}+E_{23}-E_{33}+E_{44}$
\item [] $K_{2}=-E_{11}+E_{22}-E_{23}-E_{33}+E_{44}$
\item [] $K_{3}=\mp(-E_{11}+E_{22}-E_{33}+xE_{41}+E_{44})$
\item [] $K_{4}=E_{11}+E_{22}-E_{33}-E_{44}$
\item [] $K_{5}=\mp(E_{11}+E_{22}\mp E_{32}-E_{33}-E_{44})$
\item [] $K_{6}=E_{11}-E_{22}-E_{23}+E_{33}-E_{44}$
\item [] $K_{7}=E_{11}-E_{22}+E_{23}+E_{33}-E_{44}$
\item [] $K_{8}=-E_{11}-E_{22}-E_{32}+E_{33}+E_{44}$
\item [] $K_{9}=-E_{11}-E_{22}+E_{33}+E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{d}_{4,1}$}
\begin{enumerate}
\item[]For $\om=e^{12}-e^{34}$
\begin{itemize}
\item[]$K_1=E_{11}\mp E_{12}-E_{22}-E_{33}+E_{44}$
\item[]$K_2=\mp E_{11}\pm E_{22}-E_{33}+xE_{43}+E_{44}$
\item[]$K_3=E_{11}-E_{22}+E_{33}-E_{44}$
\item[] $K_4=-E_{11}\mp E_{12}+E_{22}+E_{33}-E_{44}$
\item[]$K_5=E_{11}+E_{21}-E_{22}+xE_{23}+E_{33}-xE_{41}-E_{44}$
\item[]$K_6=-E_{11}\mp E_{21}+E_{22}-E_{33}+E_{44}$
\end{itemize}
\item[]For $\om=e^{12}-e^{34}+e^{24}$
\begin{itemize}
\item[]$K_{1}=\mp(E_{11}+xE_{12}-E_{22}-E_{33}+E_{44})$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{d}_{4,2}$}
\begin{enumerate}
\item[]For $\om=e^{12}-e^{34}$
\begin{itemize}
\item[]$K_1=E_{11}\mp E_{12}-E_{22}-E_{33}+E_{44}$
\item[]$K_2=E_{11}-E_{22}-E_{33}+E_{44}$
\item[]$K_3=E_{11}-E_{22}+E_{33}+xE_{43}-E_{44}$
\end{itemize}
\item[]For $\om=e^{14}-e^{23}$
\begin{itemize}
\item[]$K_1=-E_{11}-E_{22}+\frac{1}{x}E_{32}+E_{33}-2xE_{14}+E_{44}$
\item[]$K_2=-E_{11}-E_{22}-2E_{31}+xE_{32}+E_{33}+2E_{24}+E_{44}$
\item[]$K_3=E_{11}+E_{22}-2E_{31}+xE_{32}-E_{33}+2E_{24}-E_{44}$
\item[]$K_4=E_{11}-E_{22}+xE_{12}+xE_{32}+E_{33}+xE_{14}+xE_{34}-E_{44}$
\item[]$K_5=-E_{11}+E_{22}-E_{33}+xE_{41}+E_{44}$
\item[]$K_6=E_{11}-E_{22}-2xE_{23}+E_{33}+xE_{41}-E_{44}$
\item[] $K_7=-E_{11}+2xE_{21}+E_{22}-2xE_{23}-E_{33}-2xE_{41}-2xE_{43}+E_{44}$
\end{itemize}
\item[]For $\om=e^{14}+e^{23}$
\begin{itemize}
\item[]$K_1=-E_{11}-E_{22}+\frac{1}{x}E_{32}+E_{33}-2xE_{14}+E_{44}$
\item[]$K_2=-E_{11}+E_{22}-E_{33}+xE_{14}+E_{44}$
\item[]$K_3=E_{11}-E_{22}+xE_{23}+E_{33}-\frac{1}{2}xE_{14}-E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{d}_{4,\frac{1}{2}}$}
\begin{enumerate}
\item[]For $\omega =e^{12}-e^{34}$
\begin{itemize}
\item[]$K_1=E_{11}-E_{22}-E_{33}+xE_{43}+E_{44}$
\item[]$K_2=E_{11}-E_{22}+E_{33}-E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{d}_{4,\lambda}$}\; ($\lambda>\frac12$, $\lambda\not=1,2$)
\begin{enumerate}
\item[]For $\omega =e^{12}-e^{34}$
\begin{itemize}
\item[]$K_1=E_{11}\mp E_{12}-E_{22}-E_{33}+E_{44}$
\item[]$K_2=-E_{11}+E_{21}+E_{22}-E_{33}+E_{44}$
\item[]$K_3=-E_{11}-E_{21}+E_{22}-E_{33}+E_{44}$
\item[]$K_4=E_{11}-E_{22}-E_{33}+xE_{43}+E_{44}$
\item[]$K_5=-E_{11}+E_{22}-E_{33}+xE_{43}+E_{44}$
\item[]$K_6=\mp E_{11}+xE_{12}\pm E_{22}+E_{33}-E_{44}$
\end{itemize}
\end{enumerate}
\textbf{Lie algebra $\mathfrak{h}_{4}$}
\begin{enumerate}
\item[]For $\omega =\pm(e^{12}-e^{34})$
\begin{itemize}
\item[]$K_1=\mp(E_{11}-E_{22}-E_{33}+E_{44})$
\end{itemize}
\end{enumerate}
\end{theo}
\begin{co}
The symplectic Lie algebras $\mathfrak{r}\mathfrak{r}^\prime_{3,0}$ , $\mathfrak{n}_{4}$, $\mathfrak{r}_{4,0,\delta}^{\prime}$ and $\mathfrak{d}_{4,\delta}^{\prime}$ does not admit a para-K\"ahler structure.
\end{co}
The paper is organized as follows. Section 2 contains the basic results which are essential to the classification of four-dimensional para-K\"ahler Lie algebras (proof of the Theorem \ref{1-1}). Theorem \ref{2.1} and Theorem \ref{main2} are the key steps in this proof.
Section 3 is devoted to some curvature properties of four-dimensional para-K\"ahler metrics. Section 4 contains the tables of Theorems \ref{main2} and the isomorphisms tables used in the proof of Theorem \ref{1-1}.
The software Maple $18^\circledR$ has been used to check all needed calculations.
\section{Proof of The Theorem \ref{1-1}}
In this section we begin with a reminder of the new approach introduced by Benayadi and Boucetta in \cite{B-B}, which characterizes the para-K\"ahler Lie algebras.
Recall that, a para-K\"ahler Lie algebra $(\g,\langle.,. \rangle,K)$ is carries a Levi-Civita product, the product characterized by Koszul's formula:
\[2\langle u.v,w\rangle= \langle[u, v], w\rangle+ \langle [w,u], v\rangle+ \langle[w, v], u\rangle.\]
The subalgebras $\g^{1}=\ker(K-Id_\g)$ and $\g^{-1}=\ker(K+Id_\g)$ have the following properties, $\g^{1}$ and $\g^{-1}$ are isotropic with respect to $\langle.,. \rangle$, Lagrangian with respect to $\om$ and checking that $\g=\g^{1}\oplus \g^{-1}$, moreover the restriction of the Levi-Civita product on $\g^{1}$ and $\g^{-1}$ induces a left symmetric structures. i.e. for any $u$, $v$, $w\in\g^{1}$ (resp. $\g^{-1}$),
\[ass(u,v,w) = ass(v,u,w)\]
where $ass(u,v,w) = (u.v).w -u.(v.w)$. In particular, $\g^{1}$ and $\g^{-1}$ are left symmetric algebras.
For any $u\in\g^{-1}$, let $u^*$ denote the element of $(\g^1)^*$ given by $u^*(v)=\langle u,v \rangle$. The map $u\longmapsto u^*$ realizes an isomorphism between $\g^{-1}$ and $(\g^1)^*$. Thus, we can identify $(\g,\langle.,. \rangle,K)$ relative to the phase space $(\g^1\oplus(\g^1)^*,\langle.,.\rangle_0,K_0)$, where $\langle.,.\rangle_0$ and $K_0$ are given by:
\[\langle u+\al,v+\be\rangle_0 = \al(v)+\be(u)\esp K_0(u+\al) =u-\al.\]
Both $\g^{1}$ and $(\g^{1})^*$ carry a left symmetric algebra structure. For any $u\in\g^{1}$ and for any $\al\in(\g^{1})^*$, we denote $L_u : \g^{1}\to\g^{1}$ and $L_\al: (\g^{1})^*\to (\g^{1})^*$ as the left multiplication by $u$ and $\al$, respectively, i.e., for any $v\in\g^{1}$ and any $\be\in(\g^{1})^*$,
\[L_uv=u.v\esp L_\al\be=\al.\be.\]
The Levi-Civita product (and the Lie bracket) on $\g$ is determined entirely by their restrictions to $(\g^{1})^*$ and $\g^{1}$: For any $u\in\g^{1}$ and for any $\al\in\g^{1})^*$,
\[u.\al=L^t_X\al \esp \al.u=- L^t_\al X.\]
Conversely, let $U$ be a finite dimensional vector space and $U^*$ is its dual space. We suppose that both $U$ and $U^*$ have the structure of a left symmetric algebra. We extend the products on $U$ and $U^*$ to $U\oplus U^*$ for any $X$, $Y\in U$ and for any $\al$, $\be\in U^*$, by putting
\begin{equation}\label{1}
(X+\al).(Y+\be) = X.Y - L^t_\al Y- L^t_X\be + \al.\be.
\end{equation}
We say that two left symmetric products on $U$ and $U^*$ is Lie-extendible if the commutator of the product on $U\oplus U^*$ given by (1) is a Lie bracket. In this case we have the following theorem:
\begin{theo} \label{2.1} \cite{B-B} Let $(U, .)$ and $(U^*, .)$ be two Lie-extendible left symmetric products. Then, $(U\oplus U^*,\langle.,.\rangle_0,K_0)$, endowed with the Lie algebra bracket associated with the product given by $(1)$ is a para-K\"ahler Lie algebra.
Where $\om_{0},\; \langle.,.\rangle_0$ and $K_0$ are given by:
\[\om_{0}(u+\al,v+\beta)=\beta(u)-\al(v),\quad \langle u+\al,v+\be\rangle_0 = \al(v)+\be(u)\esp K_0(u+\al) =u-\al.\] Moreover, all para-K\"ahler Lie algebras are obtained in this manner.
\end{theo}
Let now $U$ be a 2-dimensional vector space and $U^*$ its dual space and let $\{e_1,e_2\}$ , $\{e_3,e_4\}$ be a basis of $U$ and $U^*$. We base on the previous theorem and the classification of real left-symmetric algebras in dimension 2 listed below (see Theorem 1.2. of \cite{K}),
\begin{longtable}{l|l}
$\mathfrak{b}_{1,\al}$ : $e_2.e_1=e_1,\,e_2.e_2=\al.e_2$ & $\mathfrak{b}_4$ : $e_1.e_2=e_1,\,e_2.e_2=e_1.e_2$ \\
$\mathfrak{b}_2$ : $e_2.e_1=e_1,\,e_2.e_2= e_1 +e_2$& $\mathfrak{b}_5^+$ : $e_1.e_1=e_2,\,e_2.e_1=-e_1,\,e_2.e_2=-e_2$\\
$\mathfrak{b}_{3,\al}$ $\al\not=0$: $e_1.e_2=e_1,\,e_2.e_1=(1-\frac{1}{\al})e_1,\,e_2.e_2=e_2$&$\mathfrak{b}_5^-$ : $e_1.e_1=-e_2,\,e_2.e_1=-e_1,\,e_2.e_2=-e_2$\\
&\\
$\mathfrak{c}_1$ : Trivial left-symmetric algebra & $\mathfrak{c}_2$ : $e_2.e_2=e_2$\\
$\mathfrak{c}_5^+$ : $e_2.e_2=e_2,\,e_2.e_1=e_1,\,e_1.e_2=e_1,\,e_1.e_1=e_2$ &$\mathfrak{c}_3$ : $e_2.e_2=e_1$ \\
$\mathfrak{c}_5^-$ : $e_2.e_2=e_2,\,e_2.e_1=e_1,\,e_1.e_2=e_1,\,e_1.e_1=e_2$&$\mathfrak{c}_4$ : $e_2.e_2=e_2,\,e_2.e_1=e_1,\,e_1
e_2=e_1$.\\
\end{longtable}
\begin{remark}
$\mathfrak{b}$ stands for algebras with non-commutative associated Lie algebra and $\mathfrak{c}$ stands for algebras with commutative associated Lie algebra .
\end{remark}
\begin{theo}\label{main2}Let $(\g,\langle.,.\rangle,K)$ be a four-dimensional para-K\"ahler Lie algebra.
Then there exists a basis $\{e_1,e_2,e_3,e_4\}$ of $\g$ such that
\[\om=e^{13}+e^{24},\quad\langle.,.\rangle=\e^{13}+\e^{24} \esp K=E_{11}+E_{22}-E_{33}-E_{44}\]
and the non vanishing Lie brackets as listed in the Table \ref{tab4} and \ref{tab5}.
\end{theo}
\begin{proof}
We will give the proof in the case $\B_2$ since all cases should be handled in a similar way.
In that case the left-symmetric product in $U$ is given by $e_2.e_1=e_1$, $e_2.e_2=e_1+e_2$ and
let
\begin{eqnarray*}
e_3.e_3 &=& a_{33} e_3+ b_{33} e_4 \\
e_3.e_4&=& a_{34} e_3+ b_{34} e_4\\
e_4.e_3 &=& a_{43} e_3+ b_{43} e_4 \\
e_4.e_4 &=&a_{44} e_3+ b_{44} e_4
\end{eqnarray*}
be an arbitrary product in $U^*$, let's look for products in $U^*$ which satisfy the Jacobi identity $\oint[[e_i,e_j],e_k]=0$ with $1\leq i<j<k\leq 4$, where $\oint$ denotes the cyclic sum.
The identity $\oint[[e_1,e_2],e_3]=0$ and $\oint[[e_1,e_2],e_4]=0$ is equivalent to
\[ \begin{cases}
b_{34}+a_{33}+a_{43}=0\\
a_{44}=0\\
b_{44}+a_{43}=0\\
\end{cases}\]
suppose that $a_{44}=0$, $b_{44}+a_{43}=0$ and $b_{34}=-a_{33}-a_{43}$, the identity $\oint[[e_1,e_3],e_4]=0$ and $\oint[[e_2,e_3],e_4]=0$ is equivalent to
\[\begin{cases}
a_{33}a_{34}-2a_{33}a_{43}-a_{34}d_{43}-a_{43}^2-a_{43}d_{43}=0\\
a_{34}(a_{34}+a_{43})=0\\
a_{34}=0\\
\end{cases} \esp \begin{cases}
(a_{34}-3a_{43})d_{33}+d_{43}(a_{33}-d_{43})=0\\
2a_{43}^2+(2a_{33}-a_{34}+d_{43})a_{43}-a_{34}(a_{33}-d_{43})=0\\
a_{34}=0\\
a_{34}+d_{43}+2a_{33}=0\\
\end{cases}\]
we get $a_{34}= 0$, $a_{43}= 0$, $a_{33}= 0$, and $d_{43} = 0$. Then the product in $U^*$ is given by $e_3.e_4=e_4$
(who is indeed a left-symmetric product) and the Lie bracket in $U\oplus U^*$ is given by
\[[e_{1},e_{2}]=-e_{1}, \, [e_{2},e_{3}]=xe_{1}-e_{3}-e_{4}, \, [e_{2},e_{4}]=-e_{4}.\]
\end{proof}
\begin{proof}\textit{of the Theorem \ref{1-1}}.
The Theorem \ref{main2} confirms that for each Lie algebra $\G$ of the tables \ref{tab4} and \ref{tab5} there exist a base $B_0=(e_1,e_2,e_3,e_4)$ such that the para-k\"ahler structure is given by
\[\om=e^{13}+e^{24} \esp K=E_{11}+E_{22}-E_{33}-E_{44}\]
ant the Lie brackets depend on some parameters.
In Tables \ref{tab6} and \ref{tab7} we build a family of isomorphisms (depending on the values of parameters) from $\G$ ($\mathfrak{B_{i,j}} \; or \; \mathcal{C}_{i,j}$) onto a four-dimensional Lie algebra, (say $A$) of the Table \ref{tab1}.
Each isomorphism is given by the passage matrix $P$ from $B_0$ to $B= ( f_1, f_2, f_3, f_4)$. The image by $P$ of the para-K\"ahler structure $(\omega,K)$ is given by the matrices of its component in the bases $B$ and $B^*$ by
\[^tP\circ\omega\circ P=\omega_i\quad and \quad P^{-1}\circ K\circ P=K_i.\]
In this way we collect all the possible para-k\"ahler structures $(\omega_i,K_i)$ on $A$. Thereafter, we proceed
to the classification in $A$ (up to automorphism).
We will give the proof in the case $\mathfrak{r} \mathfrak{r}_{3,0}$ since all cases should be handled in a similar way. We will show that the Lie algebra $\mathfrak{r} \mathfrak{r}_{3,0}$ admits two non-equivalent para-K\"ahler structures. Note that in this case the non vanishing Lie bracket is
\[[f_1,f_2]=f_2\]
the symplectic form is $\omega_0=f^{12}+f^{34}$ and the automorphisms is
\[T=\left( \begin {array}{cccc}
1&0&0&0\\
a_{2,1}&a_{2,2}&0&0\\ a_{3,1}&0&a_{3,3}&a_{3,4}\\ a_{4,1}&0&a_{4,3}&a_{4,4}\end {array}
\right).\]
The groups of automorphisms of four dimensional Lie algebras were
given in \cite{O}.
From Table \ref{tab6} and Table \ref{tab7}, $\mathfrak{r} \mathfrak{r}_ {3.0} $ is obtained four times.
\begin{enumerate}
\item The transformation: $f_1=-e_4, \;f_2=e_2, \;f_3=e_3, \; f_4=e_1 $ gives an isomorphism from $\mathcal{C}_{1,6}$
to $\mathfrak{r} \mathfrak{r}_ {3,0}$ and the para-K\"ahler structure obtained on $\mathfrak{r} \mathfrak{r}_ {3,0}$ is
\[\omega_1=f^{12}-f^{34}\quad and\quad K_1=-E_{11}+E_{22}-E_{33}+E_{44}.\]
\item The transformation: $f_1=-e_2, \;f_2=-ye_2+e_4, \;f_3=e_1, \; f_4=e_3$
gives an isomorphism from $\mathcal{C}_{2,1}$ with $x=0$ to $\mathfrak{r} \mathfrak{r}_ {3,0}$ and the para-K\"ahler structure obtained on $\mathfrak{r} \mathfrak{r}_ {3,0}$ is
\[\omega_2=-f^{12}+f^{34}\quad and\quad K_2=E_{11}-2yE_{12}-E_{22}+E_{33}-E_{44}.\]
\item The transformation: $f_1=-e_2, \;f_2=e_4, \;f_3=e_1, \; f_4=e_3 $ gives an isomorphism from
$\mathcal{C}_{2,2}$ with $ x=0,\; y=0$ to $\mathfrak{r} \mathfrak{r}_ {3,0}$ and the para-K\"ahler structure obtained on $\mathfrak{r} \mathfrak{r}_ {3,0}$ is
\[\omega_3=-f^{12}+f^{34}\quad and\quad K_3=E_{11}-E_{22}+E_{33}-E_{44}.\]
\item The transformation:
$f_1=e_1-e_2, \;f_2=e_4, \;f_3=e_1 ,\; f_4=e_3 $ gives an isomorphism from
$\mathcal{C}_{2,3} $ with $ x=0,\; y=0$
to $\mathfrak{r} \mathfrak{r}_ {3,0}$ and the para-K\"ahler structure obtained on $\mathfrak{r}\mathfrak{r}_{3,0}$ is
\[\omega_4=-f^{12}+f^{24}+f^{34}\quad and\quad K_4=E_{11}-E_{22}+E_{33}-E_{44}.\]
\end{enumerate}
the algebra $\mathfrak{r}\mathfrak{r}_{3,0}$ support $\omega_0$ as a unique symplectic structure (up to automorphism), therefore there are four families of automorphisms $T_i$, $i\in\{1,...,4\}$ such that,$T^*_i\omega_i=\omega_0$ for $i\in\{1,...,4\}$, a direct calculation gives us
\[T_1=\left( \begin {array}{cccc} 1&0&0&0\\ a_{2,1}&1&0&0\\
0&0&\frac{a_{3,4}a_{4,3}-1}{a_{4,4}}&a_{3,4}\\ 0&0&a_{4,3}&a_{4,4}\end {array}
\right),\;T_2=\left( \begin {array}{cccc} 1&0&0&0\\ a_{2,1}&-1&0&0\\
0&0&\frac{a_{3,4}a_{4,3}+1}{a_{4,4}}&a_{3,4}\\ 0&0&a_{4,3}&a_{4,4}\end {array}
\right)
\]
\[T_3=\left( \begin {array}{cccc} 1&0&0&0\\ a_{2,1}&1&0&0\\
0&0&\frac{a_{3,4}a_{4,3}-1}{a_{4,4}}&a_{3,4}\\ 0&0&a_{4,3}&a_{4,4}\end {array}
\right),\;T_4=\left( \begin {array}{cccc} 1&0&0&0\\ a_{2,1}&-1&0&0\\
-1&0&\frac{a_{3,4}a_{4,3}+1}{a_{4,4}}&a_{3,4}\\ 0&0&a_{4,3}&a_{4,4}\end {array}
\right).
\]
Thus we obtain four para-K\"ahler structures on $\mathfrak{r}\mathfrak{r}_{3,0}$ given by $(\omega_0,K_{0i})$, $i\in{1,...,4}$ with $K_{0i}=T^{-1}_i\circ K_i\circ T_i$ a direct calculation gives us
\begin{itemize}
\item[]$K_{01}=-E_{11}+E_{22}-E_{33}+E_{44}$
\item[]$K_{02}=E_{11}+2yE_{12}-E_{22}+E_{33}-E_{44}$
\item[]$K_{03}=E_{11}-E_{22}+E_{33}-E_{44}$
\item[]$K_{04}=-E_{11}+E_{22}+E_{33}-E_{44}$
\end{itemize}
Noticing that $K_{03}$ is a sub-case of $K_{02}$ and that $(\omega_0,K_{04})$ is isomorphic to $(\omega_0,K_{01})$. Indeed we have $L^* \omega_0=\omega_0$ and $L^{-1}_i\circ K_{04}\circ L_i=K_{01}$
with
\[L=\left(\begin {array}{cccc}
1&0&0&0\\ 0&1&0&0
\\ 0&0&0&1\\ 0&0&-1&0\end {array}
\right).
\]
We complete the proof by showing that $(\omega_0,K_{01})$ is not isomorphic to $(\omega_0,K_{02})$. Indeed, the symplectomorphism group of $\omega_0$ is generated by
\[L_1=\left( \begin {array}{cccc} 1&0&0&0\\ a_{{2,1}}&1&0&0\\ 0&0&{\frac {a_{{3,4}}a_{{4,3}}+1}{a_{{4,4}}}}&a
_{{3,4}}\\ 0&0&a_{{4,3}}&a_{{4,4}}\end {array}
\right)
\quad and \quad L_2=\left( \begin {array}{cccc} 1&0&0&0\\ \noalign{\medskip}a_{{2,1}}&1&0
&0\\ \noalign{\medskip}0&0&a_{{3,3}}&a_{{3,4}}\\ \noalign{\medskip}0&0
&-{a_{{3,4}}}^{-1}&0\end {array} \right)
\]
a simple calculation gives us \[f^2((L_1^{-1}\circ K_{01}\circ L_1-K_{02})(f_2))=2\quad and \quad f^1((L_2^{-1}\circ K_{01}\circ L_2-K_{02})(f_1))=-2\] so $L_1^{-1}\circ K_{01}\circ L_1\not=K_{02}$ and $L_2^{-1}\circ K_{01}\circ L_2\not=K_{02}$.
\end{proof}
\section{Application: Curvature properties of four-dimensional para-K\"ahler Lie algebras}
Let now $(\G,\omega,K)$ denote a four-dimensional para-K\"ahler Lie algebra.
Let $\nabla : \G\times\G\longrightarrow\G$ be the Levi-Civita product associated to a left-invariant pseudo-Riemannian metric $h(X,Y)=\omega(KX,Y)$. The connection $\nabla$ is also called Hess connection. The curvature tensor is then described in terms of the map
\begin{equation}
\begin{array}{rcl}
R :\quad \G\times\G& \longrightarrow&gl(\G) \\
(X,Y) & \longmapsto& R(X,Y)=\nabla_{[X,Y]}-[\nabla_X, \nabla_Y]
\end{array}.
\end{equation}
The Ricci tensor is the symmetric tensor $ric$ given by
$ric(X,Y) = tr(Z\longmapsto R(X,Z)Y)$
and the Ricci operator $Ric : \G\longrightarrow\G$ is given by the relation $h(Ric(X),Y) = ric(X,Y).$ The scalar curvature is defined in the standard way by $s=tr(Ric)$.
Recall that: $(\G,h)$ is called flat if $R= 0$, Ricci flat if $Ric= 0$ and Ricci soliton if
\begin{equation}
\mathcal{L}_X h +ric=\lambda h,
\end{equation}
where $X=x_1e_1+x_2e_2+x_3e_3+x_4e_4$ is a vector field and $\lambda$ is a real constant, in that case if $X=0$ then $h$ is called Einstein metric and if $\lambda$ is positive, zero, or negative then $h$ is called a shrinking, steady, or expanding Ricci soliton, respectively. We give in the following theorem some geometrical situations for the left
invariant four-dimensional dimensional para-K\"ahler
Lie groups.
\begin{theo}
Let $(\G,\omega,K)$ be a class of para-K\"ahler Lie algebras obtained in Theorem \ref{1-1}. The associated para-K\"ahler metric and some of his properties are given in the following tables
\begin{longtable}{clcccc}
\hline
\multirow{2}{*}{Lie algebra} &\multirow{2}{*}{ Para-K\"ahler metric} & \multirow{2}{*}{R=0} &\multirow{2}{*}{Ric=0} &\multicolumn{2}{c}{Ricci soliton} \\
&&&&$\lambda$&$X$\\
\hline
\multirow{2}{*}{$\mathfrak{rh}_{3}$}&$\mp(\e^{14}-\e^{23})$&Yes&Yes&0&$(0,0,x_3,x_4)$\\
&$\e^{13}-\e^{14}+\e^{23}$&Yes&Yes&0&$(0,0,x_3,x_4)$\\
\hline
\multirow{2}{*}{$ \mathfrak{rr}_{3,0}$}&$\pm(\e^{12}+\e^{34})$&Yes&Yes&0&$(0,0,x_3,x_4)$\\
&$\e^{12}+x\e^{22}+\e^{34}$, $x\not=0$&No&No&&No\\
\hline
\multirow{2}{*}{$ \mathfrak{rr}_{3,-1}$}&$\pm(\e^{14}+\e^{23}\pm\e^{33})$&No&Yes&0&$(0,0,0,x_4)$\\
&$\e^{14}+\e^{23}+x\e^{44}$&Yes&Yes&0&$(x_1,0,0,x_4)$\\
&$\e^{14}\pm \e^{22}-\e^{23}$&No&Yes&$0$&$(0,0,0,x_{4})$\\
&$\e^{14}-\e^{23}$&Yes&Yes&0&$(x_{1},0,0,x_{4})$\\
\hline
$ \mathfrak{r}_{2}\mathfrak{r}_{2}$ &$\pm(-\e^{12}-\mu\e^{13}+2\e^{23}+2\mu\e^{33}+\e^{34})$&No&Yes&$0$&$(0,0,0,0)$\\
$\mu>0$&$-\e^{12}-\mu\e^{13}+\e^{34}$&No&Yes&$0$&$(0,0,0,0)$\\
\hline
&$\e^{12}+2\e^{14}-\e^{34}$&Yes&Yes&$-x_3$&$(x_3,0,x_3,0)$\\
&$-\e^{12}+x\e^{22}+\e^{23}+x\e^{24}-\frac1x\e^{33}+x\e^{44}$&No&No&$\frac{3}{2}x$&$(0,0,0,0)$\\
&$-\e^{12}-\e^{34}$&Yes&Yes&$-x_3$&$(x_3,0,x_3,0)$\\
$\mathfrak{r}_{2}\mathfrak{r}_{2}$&$-\e^{12}-x\e^{33}+\e^{34}$&Yes&Yes&$-x_1$&$(x_1,0,x_1,\frac{x}{2}x_1)$\\
$\mu=0$&$\e^{12}-\e^{23}-\e^{34}$&Yes&Yes&$-x_3$&$(x_3,0,x_3,0)$\\
&$\e^{12}+x\e^{22}+\e^{34}+y\e^{44}$, $xy\not=0$, $x\not=y$ &No&No&&No\\
&$\e^{12}+x\e^{22}+\e^{34}+x\e^{44}$, $x\not=0$ &No&No&$x$&$(0,0,0,0)$\\
&$\e^{12}+x\e^{22}+\e^{34}$, $x\not=0$&No&No&$-x$&$(0,0,x,0)$\\
&$\e^{12}+\e^{34}+y\e^{44}$, $y\not=0$ &No&No&$y$&$(-y,0,0,0)$\\
&$\e^{12}+\e^{34}$&Yes&Yes&$-x_3$&$(x_3,0,x_3,0)$\\
&$\e^{12}+x\e^{22}+x\e^{24}+\e^{34}+x\e^{44}$, $x\not=0$&No&No&$\frac{3}{2}x$&$(0,0,0,0)$\\
\hline
&$\e^{14}+\frac4x\e^{22}-\e^{23}+x\e^{44}$&No&No&$\frac32x$&$(0,0,0,0)$\\
&$-\e^{14}-\e^{23}$&Yes&Yes&$-x_1$&$(x_1,0,0,0)$\\
&$h_1$,\qquad $xy\not=0$&No&No&&No\\
$\mathfrak{r}_{2}'$&$h_2$,\qquad $y\not=0$&No&No&$-2y$&$(0,0,0,0)$\\
&$-(2+x)\e^{12}+(x+1)(\e^{14}+\e^{23})-x\e^{34}$, $x\not=0$&No&No&&No\\
&$-2\e^{12}+\e^{14}+\e^{23}$&Yes&Yes&$-x_3$&$(x_3,0,x_3,0)$\\
&$\e^{14}+\e^{23}+\e^{33}+\e^{41}+x\e^{44}$, $x\not=0$&No&No&$\frac32x$&$(-\frac32x,0,0,0)$\\
&$\e^{14}+\e^{23}+\e^{33}+\e^{41}$&Yes&Yes&$-x_1$&$(x_1,0,0,0)$\\
\hline
$\mathfrak{r}_{4,0}$ &$\pm(\e^{14}\pm\e^{23})$&No&Yes&$0$&$(0,x_2,0,0)$\\
\hline
\multirow{2}{*}{$\mathfrak{r}_{4,-1}$} &$\e^{13}-\e^{24}+x\e^{33}$, $x\not=0$&No&Yes&$0$&$(0,0,0,0)$\\
&$\pm(\e^{13}-\e^{24})$&Yes&Yes&$0$&$(0,0,0,0)$\\
\hline
\multirow{6}{*}{$\mathfrak{r}_{4,-1,\beta}$} &$\mp(\e^{12}\pm\e^{22}-\e^{34})$,\quad$\beta\not=0$&No&Yes&$0$&$(0,0,0,0)$\\
&$\e^{12}\mp\e^{34}$,\qquad\qquad $\beta\not=0$&Yes&Yes&$0$&$(0,0,0,0)$\\
&$\e^{12}-\frac1x\e^{33}+x\e^{44}$,\quad $\beta\not=0$&No&No&&No\\
&$\mp(\e^{12}\pm\e^{22}-\e^{34})$,\quad$\beta=0$&No&Yes&$0$&$(0,0,x_3,0)$\\
&$\e^{12}\mp\e^{34}$,\qquad\qquad $\beta=0$&Yes&Yes&$0$&$(0,0,x_3,x_4)$\\
&$\e^{12}-\frac1x\e^{33}+x\e^{44}$,\quad $\beta=0$&No&Yes&$0$&$(0,0,x_3,x_4)$\\
\hline
\multirow{6}{*}{$\mathfrak{r}_{4,-1,-1}$} &$-x\e^{11}-\e^{12}-\e^{13}-\e^{34}$,\quad$x\not=0$&No&Yes&$0$&$(0,0,0,0)$\\
&$-\e^{12}-\e^{13}-\e^{34}$&Yes&Yes&$0$&$(0,0,0,0)$\\
&$\mp(\mp\e^{11}-\e^{12}-\e^{34})$&No&Yes&$0$&$(0,0,0,0)$\\
&$-\e^{12}\mp\e^{34}$&Yes&Yes&$0$&$(0,0,0,0)$\\
&$\e^{12}\mp\e^{22}-\e^{34}$&No&Yes&$0$&$(0,0,0,0)$\\
&$\pm\e^{12}+ x\e^{33}-\e^{34}$,\quad$x\not=0$&No&No&$0$&$(0,0,0,0)$\\
&$\e^{12}-\e^{34}$&Yes&Yes&$0$&$(0,0,0,0)$\\
\hline
\multirow{4}{*}{$\mathfrak{r}_{4,\alpha,-\alpha}$ }&$\mp(\e^{14}-\e^{23}\mp\e^{33})$&No&Yes&$0$&$(0,0,0,0)$\\
&$\mp(x\e^{11}-\e^{14}+\e^{23})$,\qquad$x\not=0$&No&No&&No\\
&$\mp(\e^{14}\mp\e^{23})$&Yes&Yes&$0$&$(0,0,0,0)$\\
&$-\e^{14}\mp\e^{22}-\e^{23}$&No&Yes&$0$&$(0,0,0,0)$\\
\hline
&$\mp(\e^{12}\mp\e^{22}+\e^{34})$&Yes&Yes&$0$&$(x_1,0,0,0)$\\
&$\e^{12}+x\e^{33}+\e^{34}$,$x\not=0$&No&Yes&$\frac{3}{2}x$&$(0,0,0,0)$\\
&$\e^{12}+\e^{34}$&Yes&Yes&$-x_4$&$(x_1,0,0,x_4)$\\
&$\mp(\e^{12}-\e^{34})$&Yes&Yes&$-x_4$&$(0,0,0,x_4)$\\
$ \mathfrak{d}_{4,1}$&$-\e^{11}+\e^{12}-x\e^{13}-\e^{34}$, $x\not=0$&No&No&&No\\
&$\mp(\e^{11}-\e^{12}+\e^{34})$&No&Yes&$0$&$(0,0,0,0)$\\
&$-\e^{11}-\e^{12}+\e^{34}$&No&Yes&$0$&$(0,0,0,0)$\\
&$-\e^{12}+x\e^{33}+\e^{34}$, $x\not=0$&No&No&$\frac{3}{2}x$&$(0,0,0,0)$\\
&$\pm(\e^{12}+x\e^{22}-\e^{24}+\e^{34})$&Yes&Yes&$0$&$(x_1,0,0,0)$\\
\hline
&$\e^{12}\mp\e^{22}+\e^{34}$&No&Yes&$0$&$(\mp x_1,x_1,0,0)$\\
&$\e^{12}+\e^{34}$&Yes&Yes&$-x_4$&$(0,x_2,0,x_4)$\\
&$\e^{12}+x\e^{33}-\e^{34}$, $x\not=0$&No&No&$\frac32x$&$(0,0,0,0)$\\
&$\e^{12}-\e^{34}$&Yes&Yes&$-x_4$&$(0,0,0,x_4)$\\
&$2x\e^{11}-\e^{14}\mp\frac1x\e^{22}\mp\e^{23}$&No&No&&No\\
$\mathfrak{d}_{4,2}$ &$-2\e^{12}\mp\e^{14}+x\e^{22}\pm\e^{23}$&No&No&&No\\
&$\e^{14}+\e^{23}+x\e^{33}$, &No&No&$0$&$(0,0,0,0)$\\
&$\mp(\e^{14}+\e^{23})$ &Yes&Yes&$0$&$(0,0,0,0)$\\
&$\mp(\e^{14}-\e^{23})$ &Yes&Yes&$0$&$(0,0,0,0)$\\
&$-x\e^{11}-\e^{14}\mp\e^{23}$, $x\not=0$ &No&No&$0$&$(0,0,0,0)$\\
&$\mp x\e^{11}+\e^{14}\pm\e^{23}+2x\e^{33}$, $x\not=0$ &No&No&$0$&$(0,0,0,0)$\\
&$2x\e^{11}-2x\e^{13}-\e^{14}-\e^{23}+2x\e^{33}$, $x\not=0$ &No&No&$0$&$(0,0,0,0)$\\
\hline
\multirow{2}{*}{ $\mathfrak{d}_{4,\frac{1}{2}}$}&$\e^{12}+x\e^{33}-\e^{34}$, $x\not=0$&No&No&$\frac32x$&$(0,0,0,0)$\\
&$\e^{12}\mp\e^{34}$&Yes&Yes&$-x_4$&$(0,0,0,x_4)$\\
\hline
\multirow{5}{*}{$\mathfrak{d}_{4,\lambda}$}&$\e^{12}\mp\e^{22}+\e^{34}$&No&Yes&$0$&$(0,0,0,0)$\\
&$\mp\e^{11}-\e^{12}+\e^{34}$&No&Yes&$0$&$(0,0,0,0)$\\
&$\mp\e^{12}+x\e^{33}+\e^{34}$, $x\not=0$&No&No&$\frac32x$&$(0,0,0,0)$\\
&$\mp\e^{12}+x\e^{22}-\e^{34}$, $x\not=0$&No&Yes&$0$&$(0,0,0,0)$\\
&$\mp(\e^{12}\mp\e^{34})$&Yes&Yes&$-x_4$&$(0,0,0,x_4)$\\
\hline
$ \mathfrak{h}_{4}$ &$\pm(\e^{12}-\e^{34})$&No&Yes&$0$&$(0,0,0,0)$\\
\hline
\caption{Curvature properties of four-dimensional para-K\"ahler Lie algebras}
\end{longtable}
$h_1=y(\e^{11}-\e^{13}-\e^{22}+\e^{24}+\e^{33}-\e^{44})-(2+x)\e^{12}+(x+1)(\e^{14}+\e^{23})-x\e^{34}$\\
$h_2=y(\e^{11}-\e^{13}-\e^{22}+\e^{24}+\e^{33}-\e^{44})-2\e^{12}+\e^{14}+\e^{23}$
\end{theo}
\begin{proof}
We report below the details for the case of $\mathfrak{d}_{4,\frac{1}{2}}$ the other cases are treated in the same way. Let $\{e_1, e_2, e_3, e_4\}$ denotes the basis used in theorem \ref{1-1} for $\mathfrak{d}_{4,\frac{1}{2}}$. The non isomorphic para-K\"ahler structures in $\mathfrak{d}_{4,\frac{1}{2}}$ are
$(\omega,K_1)$ and $(\omega ,K_2)$ with
$\omega =e^{12}-e^{34}$,
$K_1=E_{11}-E_{22}-E_{33}+xE_{43}+E_{44}$ and $K_2=E_{11}-E_{22}+E_{33}-E_{44}$.
The corresponding compatible metric to $(\omega,K_i)$ is uniquely determined by $h_i(X, Y) = (K_iX, Y)$. Hence, para-K\"ahler metrics in $\mathfrak{d}_{4,\frac{1}{2}}$ are of the form
\[h_1=\left(\begin {array}{cccc} 0&1&0&0\\1&0&0&0
\\ 0&0&x&1\\ 0&0&1&0\end {array}
\right)\quad x\in\R \qquad and \qquad h_2=\left( \begin {array}{cccc} 0&1&0&0\\ 1&0&0&0
\\ 0&0&0&-1\\ 0&0&-1&0\end {array}
\right).
\]
For $h_1$, $x\not=0$,
using the Koszul formula, the Levi-Civita connection is described
\[\nabla_{e_1}=\left(\begin {array}{cccc}
0&0&-\frac12x&-1\\
0&0&0&0\\
0&1&0&0\\
0&-\frac12x&0&0\end {array}
\right), \qquad \qquad\nabla_{e_2}=\left( \begin {array}{cccc}
0&0&0&0\\
0&0&\frac12x&0\\
0&0&0&0\\
-\frac12x&0&0&0\end {array}
\right).
\]
\[\nabla_{e_3}=\left(\begin {array}{cccc}
-\frac12x&0&0&0\\
0&\frac12x&0&0\\
0&0&x&0\\
0&0&-x^2&-x\end {array}
\right), \qquad \qquad\nabla_{e_4}=\left( \begin {array}{cccc}
-\frac12&0&0&0\\
0&\frac12&0&0\\
0&0&1&0\\
0&0&-x&-1\end {array}
\right).
\]
Then we calculate the curvature matrices $R(e_i,e_j)$ (for $1\leq i<j\leq 4$) and we find
\[R(e_1,e_2)=\left( \begin {array}{cccc} -x&0&0&0\\ 0&x&0&0\\ 0&0&\frac{x}{2}&0\\
0&0&-\frac12x^{2}&-\frac12x\end {array} \right),\qquad R(e_1,e_3)=\left(\begin {array}{cccc}
0&0&-\frac14\,{x}^{2}&-\frac{x}{2}
\\0&0&0&0\\ 0&\frac{x}{2}&0&0
\\ 0&-\frac14 {x}^{2}&0&0\end {array} \right)
\]
\[R(e_2,e_3)=\left( \begin {array}{cccc} 0&0&0&0\\ \noalign{\medskip}0&0&-\frac14{x}
^{2}&0\\ \noalign{\medskip}0&0&0&0\\\frac14{x}^{2}&0&0
&0\end {array} \right),\qquad R(e_2,e_4)=\left(\begin {array}{cccc} 0&0&0&0\\ 0&0&-\frac x2&0
\\0&0&0&0\\ \frac x2&0&0&0\end {array}
\right)
\]
\[R(e_3,e_4)=\left(\begin {array}{cccc} \frac x2&0&0&0\\ 0&-\frac x2&0&0
\\ 0&0&-x&0\\0&0&x^2&x
\end {array} \right)\qquad and\quad R(e_1,e_4)=0.
\]
The Ricci tensor $ric$ and the Ricci operator $Ric$ are given
by
\[ric=\left( \begin {array}{cccc} 0&\frac32x&0&0\\ \frac32x&0
&0&0\\ 0&0&\frac32{x}^{2}&\frac32x\\ \noalign{\medskip}0
&0&\frac32x&0\end {array} \right)
\qquad and \qquad Ric= \left( \begin {array}{cccc} \frac32x&0&0&0\\ \noalign{\medskip}0&\frac32x
&0&0\\0&0&\frac32x&0\\ 0&0&0&\frac32x
\end {array} \right)
\]
The Lie derivative $\mathcal{L}_Xh_1$ of the metric $h_1$ with respect to an arbitrary vector field $X=x_1e_1+x_2e_2+x_3e_3+x_4e_4\in\G$ is given by
\[\mathcal{L}_Xh_1= \left( \begin {array}{cccc} 0&-x_4&x_2x&\frac32x_2
\\-x_4&0&-x_1x&-\frac12x_1
\\ x_2x&x_1x&-2x_4x&xx_3-x_4
\\ \frac 32x_2&-\frac12x_1&xx_3-x_4&2x_3\end {array} \right).\]
Then, solving equation $\mathcal{L}_X h +ric=\lambda h$, for $x\not=0$ we obtain
\[\lambda=\frac32x\qquad and\qquad X=0.\]
Notice that, in this case, the para-K\"ahler metric is a Einstein metric not Ricci flat.
For $h_1$ with $x=0$ and $h_2$
\[\nabla_{e_1}e_2=e_3,\;\nabla_{e_1}e_4=-e_1,\;\nabla_{e_4}e_1=-\frac12e_1,\;\nabla_{e_4}e_2=\frac12e_2,\;\nabla_{e_4}e_3=e_3,\;\nabla_{e_4}e_4=-e_4\]
This para-K\"ahler structure is flat ($R(e_i,e_j)=0$ for $1\leq i<j\leq 4$). The Lie derivative $\mathcal{L}_Xh_1$ of the metric $h_1$, is given by
\[\mathcal{L}_Xh_1=\left(\begin {array}{cccc} 0&-x_{{4}}&0&\frac32x_2
\\-x_4&0&0&-\frac12x_1\\0&0
&0&-x_4\\ \frac32x_2&-\frac12x_1&-x_4&2
x_3\end {array} \right).
\]
Then, solving equation $\mathcal{L}_X h=\lambda h$, for $x=0$ (or for $h_2$) we obtain
\[\lambda=-x_4\qquad and\qquad X=x_4e_4.\]
\end{proof}
\section{Tables}
{\renewcommand*{\arraystretch}{1.2}
\begin{center} \begin{longtable}{|l|l|}
\hline
Lie algebra &\multicolumn{1}{c|}{No zero brackets}\\
\hline
$\B_{1,\al}^1$
\begin{tiny}
$\al\not\in\{-2,-1,1\}$
\end{tiny} &$[e_1,e_2]=-e_1,\, [e_2,e_3]=xe_1-e_3,\, [e_2,e_4]=-\al e_4$ \\
\hline
$\B_{1,\al}^2$
\begin{tiny}
$\al\not\in\{-2,-1,0,1\}$
\end{tiny}
& $[e_1,e_2]=-e_1,\, [e_1,e_4]=-\frac{x}{\al} e_1,\, [e_2,e_3]=-e_3,\, [e_2,e_4]=xe_2-\al e_4$,\, $ [e_3,e_4]=\frac{x}{\al} e_3$\\
\hline
{$\B_{1,-2}^1$}& $ [e_1,e_2]=-e_1,\, [e_1,e_4]=xe_1,\, [e_2,e_3]=ye_1-e_3,\;[e_2,e_4]=2xe_2+2e_4,\;[e_3,e_4]=-xe_3$\\
\hline
$\B_{1,-2}^2$&$ [e_1,e_2]=-e_1,\, [e_1,e_3]=xe_1,\, [e_2,e_3]=ye_1-xe_2-e_3,\;[e_2,e_4]=xe_1+2e_4,\;[e_3,e_4]=-2xe_4$\\
\hline
$\B_{1,-1}^1$& $[e_1,e_2]=-e_1,\, [e_1,e_3]=-xe_1,\, [e_2,e_3]=ye_1+xe_2-e_3, \, [e_2,e_4]=e_4, \, [e_3,e_4]=xe_4$\\
\hline
$\B_{1,-1}^2$& $[e_1,e_2]=-e_1,\, [e_1,e_4]=xe_1,\, [e_2,e_3]=-e_3, \, [e_2,e_4]=xe_2+e_4, \, [e_3,e_4]=-xe_3$\\
\hline
$\B_{1,0}$&$[e_1,e_2]=-e_1, \, [e_1,e_4]=xe_{1}, \, [e_2,e_3]=-e_3, \, [e_3,e_4]=-xe_3$\\
\hline
\multirow{2}{*}{$\B_{1,1}^1$\; $x\not=0$}&$[e_1,e_2]=-e_1, \, [e_{1},e_{3}]=-\frac{y}{2}e_1, \, [e_1,e_4]=-xe_1, \,
[e_2,e_3]=\frac{y^2}{2x}e_1+\frac{y}{2}e_{2}-e_{3}$\\& $[e_{2},e_{4}]=ye_{1}+xe_{2}-e_{4}, \, [e_{3},e_{4}]=xe_{3}-\frac{y}{2}e_{4}$\\
\hline
$\B_{1,1}^2$&$[e_1,e_2]=-e_1,\, [e_2,e_3]=xe_1-e_3, \, [e_2,e_4]=-e_4$\\
\hline
$\B_{2}$&$[e_{1},e_{2}]=-e_{1}, \, [e_{2},e_{3}]=xe_{1}-e_{3}-e_{4}, \, [e_{2},e_{4}]=-e_{4}$ \\
\hline
$\B_{3,\alpha}^1 \; \alpha\neq0$& $[e_1,e_2]=\frac{1}{\alpha} e_1,\, [e_1,e_3]=[e_2,e_4]=-e_4,\, [e_2,e_3]=xe_1+\frac{1-\alpha}{\alpha}e_3$\\
\hline
$\B_{3,\alpha}^2\; \alpha\neq0$& $[e_1,e_2]=\frac{1}{\alpha} e_1,\, [e_1,e_3]=[e_2,e_4]=x\alpha e_2-e_4,\, [e_1,e_4]=xe_1, \, [e_2,e_3]=\frac{1-\alpha}{\alpha}e_3, \, [e_3,e_4]=x(\alpha-1)e_3$\\
\hline
$\B_{3,\frac{1}{2}}^1$& $[e_1,e_2]=2e_1,\, [e_1,e_3]=x e_1-e_4, \, [e_2,e_3]=ye_1-\frac{x}{2}e_2+e_3, \,[e_2,e_4]=-e_4,\; [e_3,e_4]=-\frac{x}{2} e_4$\\
\hline
\multirow{2}{*}{$\B_{3,\frac{1}{2}}^2$ $y\neq0$}& $[e_1,e_2]=2e_1,\, [e_1,e_3]=-2x e_1+\frac{y}{2} e_2-e_4, \,[e_1,e_4]=ye_1,\; [e_2,e_3]=-\frac{3x^{2}}{y} e_1+\frac{x}{2}e_2+e_3$\\
&$[e_2,e_4]=xe_1+\frac{y}{2}e_2-e_4,\; [e_3,e_4]=-\frac{y}{2} e_3-\frac{x}{2} e_4$\\
\hline
$\B_{3,\frac{1}{2}}^3$& $[e_1,e_2]=2e_1,\, [e_1,e_3]=[e_2,e_4]=x e_1-e_4, \, [e_2,e_3]=ye_1-xe_2+e_3, \, [e_3,e_4]=-2x e_4$\\
\hline
$\B_{3,1}$& $[e_1,e_2]=e_1,\, [e_1,e_3]=x e_1+ye_2-e_4, \,[e_1,e_4]=ye_1,\; [e_2,e_3]=ze_1, \, [e_2,e_4]=y e_2-e_4$\\
\hline
$\B_{4}$&$[e_{1},e_{2}]=e_{1}, \, [e_{1},e_{3}]=xe_{1}-e_{4}, \, [e_{2},e_{3}]=ye_{1}-e_{4}, \, [e_{2},e_{4}]=-e_{4}$\\
\hline
$\B_{5,1}^{+}$&$[e_{1},e_{2}]=e_{1}, \, [e_{1},e_{4}]=x e_{1}-e_3, \, [e_{2},e_{3}]=e_{3}, \, [e_{2},e_{4}]=-2xe_{2}+2e_4,\;[e_3,e_4]=-xe_3$\\
\hline
$\B_{5,2}^{+}$&$[e_{1},e_{2}]=e_{1}, \, [e_{1},e_{4}]=-\frac{x}{4} e_{1}-e_3, \, [e_{2},e_{3}]=xe_1+e_{3}, \, [e_{2},e_{4}]=-\frac{x}{2} e_{2}+2e_4,\;[e_3,e_4]=\frac{x}{4} e_3$\\
\hline
$\B_{5,3}^{+}$&$[e_{1},e_{2}]=e_{1}, \, [e_{1},e_{3}]=-x e_{1}, \, [e_{1},e_{4}]=-2xe_1-xe_2-e_{3}, \, [e_{2},e_{3}]=2xe_1+xe_2+e_{3}$\\&$[e_2,e_4]=3x e_1+2xe_2+2e_4,\;[e_3,e_4]=2xe_3-2xe_4$\\
\hline
$\B_{5,4}^{+}$&$[e_{1},e_{2}]=e_{1}, \, [e_{1},e_{3}]=-x e_{1}, \, [e_{1},e_{4}]=2xe_1-xe_2-e_{3}$\\&$ [e_{2},e_{3}]=-2xe_1+xe_2+e_{3},\;[e_2,e_4]=3x e_1-2xe_2+2e_4,\;[e_3,e_4]=-2xe_3-2xe_4$\\
\hline
$\B_{5,1}^{-}$&$[e_{1},e_{2}]=e_{1}, \, [e_{1},e_{4}]=x e_{1}+e_3, \, [e_{2},e_{3}]=e_{3},\;[e_2,e_4]=-2x e_2+2e_4,\;[e_3,e_4]=-xe_3$\\
\hline
$\B_{5,2}^{-}$&$[e_{1},e_{2}]=e_{1}, \, [e_{1},e_{4}]=\frac{x}{4} e_{1}+e_3, \, [e_{2},e_{3}]=xe_1+e_{3},\;[e_2,e_4]=\frac{x}{2} e_2+2e_4,\;[e_3,e_4]=-\frac{x}{4} e_3$\\
\hline
\caption{Four dimensional Para-K\"ahler Lie algebras coming from $\mathfrak{b}$}
\label{tab4}
\end{longtable}
\end{center}}
\newpage
{\renewcommand*{\arraystretch}{1.4}
\begin{center} \begin{longtable}{|l|l|}
\hline
Lie algebra &\multicolumn{1}{c|}{No zero brackets}\\
\hline
$\mathcal{C}_{1,1}$&$[e_{1},e_{4}]=e_{1}, \, [e_2,e_4]=\alpha e_2 , \; [e_3,e_4]=-e_3$\\
\hline
$\mathcal{C}_{1,2}$&$[e_{1},e_{4}]=e_{1}+e_2, \, [e_2,e_4]= e_2 , \; [e_3,e_4]=-e_3$\\
\hline
$\mathcal{C}_{1,3}$&$[e_{1},e_{3}]=[e_{2},e_{4}]=e_{2}, \, [e_1,e_4]=(1-\frac{1}{\alpha})e_1, \; [e_{3},e_{4}]=\frac{1}{\alpha}e_3$\\
\hline
$\mathcal{C}_{1,4}$&$[e_{1},e_{3}]=[e_{1},e_{4}]=[e_{2},e_{4}]=e_{2}, \, [e_3,e_4]=e_3$\\
\hline
$\mathcal{C}_{1,5}^{+}$&$[e_{1},e_{4}]=-e_{1}, \; [e_{2},e_{3}]=e_1 , \, [e_2,e_4]=-2 e_2,\;[e_3,e_4]=e_3$\\
\hline
$\mathcal{C}_{1,5}^{-}$&$[e_{1},e_{4}]=[e_{2},e_{3}]=-e_{1}, \, [e_2,e_4]=-2 e_2,\;[e_3,e_4]=e_3$\\
\hline
$\mathcal{C}_{1,6}$&$[e_{2},e_{4}]=e_{2}$\\
\hline
$\mathcal{C}_{1,7}$&$[e_{1},e_{4}]=e_{2}$\\
\hline
$\mathcal{C}_{1,8}$&$[e_{1},e_{3}]=[e_{2},e_{4}]=e_{2}, \, [e_1,e_4]= e_1$\\
\hline
$\mathcal{C}_{1,9}$&$[e_{1},e_{3}]=[e_{2},e_{4}]=e_{2}, \, [e_1,e_4]=[e_2,e_3]= e_1$\\
\hline
$\mathcal{C}_{1,10}$&$[e_{1},e_{3}]=[e_{2},e_{4}]=e_{2}, \, [e_1,e_4]= e_1 , \; [e_2,e_3]=-e_1$\\
\hline
$\mathcal{C}_{2,1}$&$[e_{1},e_{3}]=xe_{1}, \, [e_2,e_4]=ye_2-e_4$\\
\hline
$\mathcal{C}_{2,2}$&$[e_{1},e_{3}]=xe_{1}, \,[e_2,e_3]=ye_1,\; [e_2,e_4]=-e_4$\\
\hline
$\mathcal{C}_{2,3}$&$[e_{1},e_{3}]=xe_{1}, \,[e_2,e_3]=ye_1,\; [e_2,e_4]=xe_1-e_4,\;[e_3,e_4]=-xe_4$\\
\hline
$\mathcal{C}_{3,1}$&$[e_{1},e_{3}]=xe_{1}, \,[e_2,e_3]=ye_1+ze_2-e_4,\; [e_3,e_4]=ze_4$\\
\hline
$\mathcal{C}_{3,2}$&$[e_{1},e_{3}]=[e_2,e_4]=xe_{1}, \,[e_2,e_3]=ye_1+ze_2-e_4,\; [e_3,e_4]=(z-x)e_4$\\
\hline
$\mathcal{C}_{4,1}$&$[e_{1},e_{3}]=[e_2,e_4]=xe_{1}-e_4, \,[e_2,e_3]=ye_1+xe_2-e_3$\\
\hline
$\mathcal{C}_{4,2}$&$[e_{1},e_{3}]=[e_2,e_4]=xe_{2}-e_4, \,[e_2,e_3]=-e_3,\; [e_3,e_4]=xe_3$\\
\hline
$\mathcal{C}_{5,1}^{+}$&$[e_{1},e_{3}]=[e_{2},e_{4}]=xe_{1}+ye_2-e_4, \,[e_1,e_4]=[e_{2},e_{3}]=ye_1+xe_2-e_3$\\
\hline
$\mathcal{C}_{5,2}^{+}$&$[e_{1},e_{3}]=[e_2,e_4]=xe_{2}-e_4, \,[e_1,e_4]=[e_2,e_3]=-e_3,\;[e_3,e_4]=xe_3$\\
\hline
$\mathcal{C}_{5,1}^{-}$&$[e_{1},e_{3}]=[e_2,e_4]=xe_{1}+ye_2-e_4, \,[e_1,e_4]=ye_1-xe_2+e_3,\;[e_2,e_3]=-ye_1+xe_2-e_3$\\
\hline
$\mathcal{C}_{5,2}^{-}$&$[e_{1},e_{3}]=[e_2,e_4]=xe_{2}-e_4, \,[e_1,e_4]=e_3,\;[e_2,e_3]=-e_3,\;[e_3,e_4]=xe_3$\\
\hline
\caption{Four dimensional Para-K\"ahler Lie algebras coming from $\mathfrak{c}$}
\label{tab5}
\end{longtable}
\end{center}}
{\renewcommand*{\arraystretch}{1.4}
\begin{center}
\begin{longtable}{|ll|p{9.5cm}|c|}
\hline
Source && \multicolumn{1}{c|}{Isomorphism}&Target \\
\hline
$\B_{1,\al}^1$& $\vert\al\vert<1$&$f_1=e_1,\,f_2=-\frac{x}{2}e_1+e_3,\,f_3=e_4,\,f_4=e_2$& $\mathfrak{r}_{4,-1,-\al}$\\
\hline
$\B_{1,\al}^1$&$ \vert\al\vert>1,\;\al\neq-2$&$f_1=e_4,\,f_2=e_1,\,f_3=-\frac{x}{2}e_1+e_3,\,f_4=-\frac{1}{\vert\al\vert} e_2$& $\mathfrak{r}_{4,-\frac{1}{\vert\al\vert},\frac{1}{\vert\al\vert}}$\\
\hline
$\B_{1,\al}^2$&$ \vert\al\vert<1,\; \al\neq0$&$f_1=e_1,\,f_2=e_3,\,f_3=-\frac{x}{\al}e_2+e_4,\,f_4=\frac{-x+\al}{\al}e_2+e_4$& $\mathfrak{r}_{4,-1,-\al}$\\
\hline
$\B_{1,\al}^2$&$ \vert\al\vert>1,\; \al\neq-2$&$f_1=-\frac{x}{ \vert\al\vert} e_2+e_4,\,f_2=e_1,\,f_3=e_3,\,f_4=-\frac{x+1}{ \vert\al\vert}e_2+e_4$& $\mathfrak{r}_{4,-\frac{1}{\vert\al\vert},\frac{1}{\vert\al\vert}}$\\
\hline
$\B_{1,-2}^1$&$ x,y\neq0$& $f_1=-ye_1+e_2+\frac{1}{x}e_4 ,\; f_2=-\frac{1}{2}ye_1+e_3 ,\; f_3=ye_1 ,\; f_4=e_2+e_3$&$\mathfrak{d}_{4,2}$\\
\hline
$\B_{1,-2}^1$&$ x\neq0,\; y=0$& $f_1=e_2+\frac{1}{x}e_4 ,\; f_2=e_3 ,\; f_3=e_1 ,\; f_4=e_2+\frac{1}{2x}e_4 $&$\mathfrak{r}_{4,-\frac{1}{2},\frac{1}{2}}$\\
\hline
$\B_{1,-2}^1$&$ x=0$& $f_1=e_4 ,\; f_2=-\frac{y}{2}e_1+e_3 ,\; f_3=e_1 ,\; f_4=\frac{1}{2}e_2 $&$\mathfrak{r}_{4,-\frac{1}{2},\frac{1}{2}}$\\
\hline
$\B_{1,-2}^2$&$ x\neq0$&$f_1=e_4, \; f_2=\frac{1}{2x}e_1+e_2+\frac{1}{x} e_3, \; f_3=-xe_1, \; f_4=-\frac{1}{x}e_3+(\frac{y+1}{x^2})e_4$&$\mathfrak{d}_{4,2}$\\
\hline
$\B_{1,-2}^2$&$ x=0$&$f_1=e_4, \; f_2=-\frac{y}{2}e_1+e_3, \; f_3=e_1, \; f_4=\frac{1}{2}e_2$&$\mathfrak{r}_{4,-\frac{1}{2},\frac{1}{2}} $\\
\hline
$\B_{1,-1}^1$&&$f_1=-\frac{1}{2}ye_1-xe_2+e_{3} ,\; f_2=e_1 ,\; f_3=e_4 ,\; f_4=-e_2$&$\mathfrak{r}_{4,-1,-1}$\\
\hline
$\B_{1,0}^1$&&$f_1=e_2,\,f_2=e_1,\,f_3=-\frac{x}{2}e_1+e_3,\,f_4=e_4$&$\mathfrak{r}\mathfrak{r}_{3,-1}$\\
\hline
$\B_{1,0}^2$&&$f_1=e_2,\,f_2=e_1,\,f_3=e_3,\,f_4=xe_2+e_4$&$\mathfrak{r}\mathfrak{r}_{3,-1}$\\
\hline
$\B_{1,0}^3$&&$f_1=-e_2,\,f_2=-\frac{x}{2} e_1+e_{3},\,f_3=e_1,\,f_4=e_4$&$\mathfrak{r}\mathfrak{r}_{3,-1}$\\
\hline
$\B_{1,1}^1$& $x\neq0$& $f_1=e_{1}, \;f_2=xe_{3}-\frac{y}{2}e_{4},\; f_{3}=\frac{y}{2}e_{1}+xe_{2}-e_{4} , \; f_4=e_2 $& $\mathfrak{r}_{4,-1,-1}$\\
\hline
$\B_{1,1}^2$&& $f_1=e_1, \;f_2=-\frac{x}{2} e_1+e_3, \;f_3=e_{4}, \; f_4=e_2 $& $\mathfrak{r}_{4,-1,-1}$\\
\hline
$\B_{2}$& &$f_1=e_1, \;f_2=-e_4, \;f_3=-\frac{x}{2}e_1+e_{3}, \; f_4=e_2 $& $\mathfrak{r}_{4,-1}$\\
\hline
$\B_{3,\alpha}^1$&$ \frac{1}{\alpha}>\frac{1}{2}$& $f_1=e_1, \;f_2=-xe_1+\frac{\alpha-2}{\alpha} e_3, \;f_3=-\frac{\alpha-2}{\alpha}e_4, \; f_4=-e_2 $& $\mathfrak{d}_{4,\frac{1}{\alpha}}$\\
\hline
$\B_{3,\alpha}^1$&$ \frac{1}{\alpha}<\frac{1}{2}$& $f_1=-xe_1+\frac{\alpha-2}{\alpha}e_3, \;f_2=e_1, \;f_3=\frac{\alpha-2}{\alpha}e_4, \; f_4=-e_2 $& $\mathfrak{d}_{4,\frac{\alpha-1}{\alpha}}$\\
\hline
$\B_{3,2}^1$&$ x\neq0$&$f_1=-xe_1, \;f_2=e_3, \;f_3=xe_4, \; f_4=-e_2 $& $\mathfrak{h}_{4}$\\
\hline
$\B_{3,2}^1$&$ x=0$&$f_1=e_3, \;f_2=e_1, \;f_3=e_4, \; f_4=-e_2 $& $\mathfrak{d}_{4,\frac{1}{2}}$\\
\hline
$\B_{3,\alpha}^2$&$\frac{1}{\alpha}>\frac{1}{2}$& $f_1=e_1+x \alpha e_2-e_4, \;f_2=e_3, \;f_3=x \alpha e_2-e_4, \; f_4=-e_2+\frac{\alpha-1}{\alpha}e_3 $& $\mathfrak{d}_{4,\frac{1}{\alpha}}$\\
\hline
$\B_{3,\alpha}^2$&$\frac{1}{\alpha}<\frac{1}{2}$& $f_1=-x \alpha e_2+e_3+e_4, \;f_2=e_1, \;f_3=-x \alpha e_2+e_4, \; f_4=\frac{1}{\alpha}e_1-e_2 $& $\mathfrak{d}_{4,\frac{\alpha -1}{\alpha}}$\\
\hline
$\B_{3,2}^2$&& $f_1=e_3, \;f_2=e_1, \;f_3=-2x e_2+e_4, \; f_4=-e_2 $& $\mathfrak{d}_{4,\frac{1}{2}}$\\
\hline
$\B_{3,\frac{1}{2}}^1$& &$f_1=e_1, \;f_2=\frac{1}{3}y e_1-\frac{1}{2}xe_2+e_3, \;f_3=-e_4, \; f_4=-e_2 $& $\mathfrak{d}_{4,2}$\\
\hline
$\B_{3,\frac{1}{2}}^2$&$ y\neq0$& $f_1=e_1, \;f_2=-e_3-\frac{x}{y}e_4, \;f_3=xe_1-\frac{1}{2}ye_2+e_4, \; f_4=-\frac{2}{y} e_4 $& $\mathfrak{d}_{4,2}$\\
\hline
$\B_{3,1}$&$ x=0,\; y=0$& $f_1=e_1, \;f_2=ze_1+e_3, \;f_3=-e_4, \; f_4=-e_2 $& $\mathfrak{d}_{4,1}$\\
\hline
$\B_{3,1}$&$ x\neq0,\; y=0$& $f_1=\frac{z}{x}e_1-e_2+\frac{1}{x}e_3, \;f_2=e_4, \;f_3=-\frac{z}{x} e_1-\frac{1}{x}e_3, \; f_4=e_1-\frac{1}{x}e_4 $& $\mathfrak{r}_{2}\mathfrak{r}_{2}$\\
\hline
$\B_{3,1}$&$ x=0,\; y\neq0,\; z\neq0$& $f_1=-\frac{\sqrt{yz}}{2y}e_1-\frac{1}{2} e_2-\frac{1}{2\sqrt{yz}}e_3, \;f_2=-\sqrt{yz}e_1-ye_2+ e_4, \;f_3=\frac{\sqrt{yz}}{2y}e_1-\frac{1}{2} e_2+\frac{1}{2\sqrt{yz}}e_3, \; f_4=\sqrt{yz}e_1-ye_2+ e_4 $& $\mathfrak{r}_{2}\mathfrak{r}_{2}$\\
\hline
$\B_{3,1}$&$x=0,\; y\neq0,\; z=0$& $f_1=e_1, \,f_2=e_3, \,f_3=ye_2-e_4, \, f_4=-e_2 $& $\mathfrak{d}_{4,1}$\\
\hline
$\B_{3,1}$&$xy\neq0,\, x^{2}+4yz=0$& $f_1=-ye_2+e_4, \,f_2=-\frac{x^{2}}{4y}e_1-\frac{x}{2}e_2+e_3, \,f_3=\frac{x^{2}}{4}e_1+\frac{xy}{2} e_2-\frac{x}{2}e_4, \, f_4=-e_2 $& $\mathfrak{d}_{4,1}$\\
\hline
$\B_{3,1}$&$ xy\neq0,\, x^{2}+4yz>0$& $f_1=\frac{z}{\sqrt{x^{2}+4yz}}e_1-\frac{x+\sqrt{x^{2}+4yz}}{2\sqrt{x^{2}+4yz}}e_2+\frac{1}{\sqrt{x^{2}+4yz}}e_3, \,f_2=(-x+\sqrt{x^{2}+4yz})e_1-2ye_2+2e_4, \,f_3=-\frac{z}{\sqrt{x^{2}+4yz}}e_1+\frac{x-\sqrt{x^{2}+4yz}}{2\sqrt{x^{2}+4yz}}e_2-\frac{1}{\sqrt{x^{2}+4yz}}e_3, \, f_4=-\frac{x+\sqrt{x^{2}+4yz}}{2}e_1-ye_2+e_4 $& $\mathfrak{r}_{2}\mathfrak{r}_{2}$\\
\hline
$\B_{3,1}$&$ xy\neq0,\, x^{2}+4yz<0$
& $f_1=\frac{\sqrt{-x^{2}-4yz}}{2}e_1-e_2, \,f_2=-\frac{x\sqrt{-x^{2}-4yz}-4z}{2\sqrt{-x^{2}-4yz}} e_1-\frac{y\sqrt{-x^{2}-4yz}+x}{\sqrt{-x^{2}-4yz}}e_2+\frac{2}{\sqrt{-x^{2}-4yz}} e_3+e_4, \,f_3=\frac{\sqrt{-x^{2}-4yz}}{2}e_1, \, f_4=-\frac{x}{2}e_1-ye_2+e_4 $& $\mathfrak{r}_{2}'$\\
\hline
$\B_{4}$&$ x=0$& $f_1=-e_1, \,f_2=ye_1+e_3-e_4, \,f_3=e_4, \, f_4=- e_2 $& $\mathfrak{d}_{4,1}$\\
\hline
$\B_{4}$&$ x\neq0$& $f_1=\frac{x-y}{x} e_1-\frac{1}{x}e_3, \,f_2=-xe_1+e_4, \,f_3=\frac{y}{x} e_1-e_2+\frac{1}{x}e_3, \, f_4=e_4 $& $\mathfrak{r}_{2}\mathfrak{r}_{2}$\\
\hline
$\B_{5,1}^+$&& $f_1=-xe_2-e_3+e_4, \,f_2=e_1, \,f_3=e_3, \, f_4= e_1+e_2 $& $\mathfrak{d}_{4,2}$\\
\hline
$\B_{5,2}^+$&$ x\neq0$& $f_1=\frac{1}{2} e_2-\frac{2}{x} e_4, \,f_2=e_3, \,f_3=\frac{x}{2}e_1 +e_3, \, f_4=\frac{4}{x} e_4 $& $\mathfrak{d}_{4,2}$\\
\hline
$\B_{5,2}^+ $&$ x=0$& $f_1=e_3+e_4, \,f_2=e_1, \,f_3=e_3, \, f_4=- e_1+e_2 $& $\mathfrak{d}_{4,2}$\\
\hline
$\B_{5,3}^+ $&$ x\neq0$& $f_1=-e_3-e_4, \,f_2=e_1, \,f_3=xe_1+xe_2+e_3, \, f_4=- \frac{1}{x} e_3 $& $\mathfrak{d}_{4,2}$\\
\hline
$\B_{5,3}^+ $&$ x=0$& $f_1=e_4, \,f_2=e_1, \,f_3=e_3, \, f_4= e_2 $& $\mathfrak{d}_{4,2}$\\
\hline
$\B_{5,4}^+ $&$ x\neq0$&$f_1=e_3+e_4, \,f_2=e_1, \,f_3=-xe_1+xe_2+e_3, \, f_4=- e_1+e_2 $ & $\mathfrak{d}_{4,2}$\\
\hline
$\B_{5,4}^+$&$ x=0$& $f_1=-e_3+e_4, \,f_2=e_1, \,f_3=e_3, \, f_4=e_1+e_2 $& $\mathfrak{d}_{4,2}$\\
\hline
$\B_{5,1}^- $&& $f_1=-xe_2+e_4, \,f_2=e_1, \,f_3=-e_3, \, f_4=e_2 $& $\mathfrak{d}_{4,2}$\\
\hline
$\B_{5,2}^- $&$ x\neq0$& $f_1=e_2+\frac{4}{x} e_4, \,f_2=e_3, \,f_3=xe_1+2e_3, \, f_4=-\frac{4}{x} e_4 $& $\mathfrak{d}_{4,2}$\\
\hline
$\B_{5,2}^- $&$ x=0$& $f_1=e_3+e_4, \,f_2=e_1, \,f_3=-e_3, \, f_4= e_1+e_2 $& $\mathfrak{d}_{4,2}$\\
\hline
\caption{Isomorphisms from the Lie algebras obtained in Table \ref{tab4} onto the Lie algebras in Table \ref{tab1}}
\label{tab6}
\end{longtable}
\end{center}}
{\renewcommand*{\arraystretch}{1.4}
\begin{center}
\begin{longtable}{|ll|p{8cm}|c|}
\hline
Source && \multicolumn{1}{c|}{Isomorphism}&Target \\
\hline
$\mathcal{C}_{1,1} $&$ -1\le \alpha <1$& $f_1=e_1, \;f_2=e_3, \;f_3=e_2, \; f_4=-e_4 $& $\mathfrak{r}_{4,-1,\alpha}$\\
\hline
$\mathcal{C}_{1,1} $&$ \alpha <-1$& $f_1=e_2, \;f_2=e_1, \;f_3=e_3, \; f_4=-\frac{1}{\alpha}e_4 $& $\mathfrak{r}_{4,\frac{1}{\alpha},-\frac{1}{\alpha}}$\\
\hline
$\mathcal{C}_{1,1} $&$ \alpha >1$& $f_1=e_2, \;f_2=e_3, \;f_3=e_1, \; f_4=-\frac{1}{\alpha}e_4 $& $\mathfrak{r}_{4,-\frac{1}{\alpha},\frac{1}{\alpha}}$\\
\hline
$\mathcal{C}_{1,1} $&$ \alpha=1$& $f_1=e_3, \;f_2=e_1, \;f_3=e_2, \; f_4=e_4 $& $\mathfrak{r}_{4,-1,-1}$\\
\hline
$\mathcal{C}_{1,2}$&& $f_1=e_3, \;f_2=-e_2, \;f_3=e_1, \; f_4=e_4 $& $\mathfrak{r}_{4,-1}$\\
\hline
$\mathcal{C}_{1,3}$&$ 0<\alpha \leq 2$& $f_1=e_2-e_3, \;f_2=e_1, \;f_3=e_2, \; f_4=\frac{\alpha-1}{\alpha}e_1 -e_4 $& $\mathfrak{d}_{4,\frac{1}{\alpha}}$\\
\hline
$\mathcal{C}_{1,3}$&$ \alpha <0\; or \; \alpha>2$& $f_1=e_1+\alpha e_2, \;f_2=e_3, \;f_3=e_2, \; f_4=e_3 -e_4 $& $\mathfrak{d}_{4,\frac{\alpha-1}{\alpha}}$\\
\hline
$\mathcal{C}_{1,4}$&& $f_1=e_3, \;f_2=e_1-e_2, \;f_3=-e_2, \; f_4=-e_4 $& $\mathfrak{d}_{4,1}$\\
\hline
$\mathcal{C}_{1,5}^{+}$&& $f_1=e_2, \;f_2=e_3, \;f_3=e_1, \; f_4=e_4 $& $\mathfrak{d}_{4,2}$\\
\hline
$\mathcal{C}_{1,5}^{-}$&& $f_1=e_2, \;f_2=-e_3, \;f_3=e_1, \; f_4=e_4 $& $\mathfrak{d}_{4,2}$\\
\hline
$\mathcal{C}_{1,6}$&& $f_1=-e_4, \;f_2=e_2, \;f_3=e_3, \; f_4=e_1 $& $\mathfrak{r} \mathfrak{r}_{3,0}$\\
\hline
$\mathcal{C}_{1,7}$&& $f_1=e_1, \;f_2=e_4, \;f_3=e_2, \; f_4=e_3 $& $\mathfrak{r} \mathfrak{h}_{3}$\\
\hline
$\mathcal{C}_{1,8}$&& $f_1=e_1, \;f_2=e_3, \;f_3=e_2, \; f_4=-e_4 $& $\mathfrak{d}_{4,1}$\\
\hline
$\mathcal{C}_{1,9}$&& $f_1=\frac{1}{2} e_3-\frac{1}{2}e_4, \;f_2=e_1-e_2, \;f_3=-\frac{1}{2} e_3-\frac{1}{2}e_4, \; f_4=e_1+e_2 $& $\mathfrak{r}_{2}\mathfrak{r}_{2}$\\
\hline
$\mathcal{C}_{1,10}$&& $f_1=-e_4, \;f_2=e_3, \;f_3=e_1, \; f_4=-e_2 $& $\mathfrak{r}_{2}'$\\
\hline
$\mathcal{C}_{2,1}$&$ x=0$& $f_1=-e_2, \;f_2=-ye_2+e_4, \;f_3=e_1, \; f_4=e_3 $& $\mathfrak{r}\mathfrak{r}_{3,0}$\\
\hline
$\mathcal{C}_{2,1} $&$ x\neq0$& $f_1=-\frac{1}{x} e_3, \;f_2=e_1, \;f_3=-e_2, \; f_4=-xye_2+xe_4 $& $\mathfrak{r}_{2}\mathfrak{r}_{2}$\\
\hline
$\mathcal{C}_{2,2} $&$ x\neq0$& $f_1=-\frac{1}{x} e_3, \;f_2=e_1, \;f_3=\frac{y}{x}e_1 -e_2, \; f_4=e_4 $& $\mathfrak{r}_{2}\mathfrak{r}_{2}$\\
\hline
$\mathcal{C}_{2,2} $&$ x=0,\; y\neq0$& $f_1=e_4, \;f_2=-ye_1, \;f_3=e_3, \; f_4=-e_2 $& $\mathfrak{r}_{4,0}$\\
\hline
$\mathcal{C}_{2,2}$&$ x=0,\; y=0$& $f_1=-e_2, \;f_2=e_4, \;f_3=e_1, \; f_4=e_3 $& $\mathfrak{r}\mathfrak{r}_{3,0}$\\
\hline
$\mathcal{C}_{2,3} $&$ x\neq0$& $f_1=\frac{y}{x}e_1 -e_2, \;f_2=-xe_1+ e_4, \;f_3=e_2-\frac{1}{x}e_3 ,\; f_4=e_1 $& $\mathfrak{r}_{2}\mathfrak{r}_{2}$\\
\hline
$\mathcal{C}_{2,3} $&$ x=0,\; y\neq0$& $f_1=e_4, \;f_2=-ye_1, \;f_3=e_3 ,\; f_4=-e_2 $& $\mathfrak{r}_{4,0}$\\
\hline
$\mathcal{C}_{2,3} $&$ x=0,\; y=0$& $f_1=e_1-e_2, \;f_2=e_4, \;f_3=e_1 ,\; f_4=e_3 $& $\mathfrak{r}\mathfrak{r}_{3,0}$\\
\hline
$\mathcal{C}_{3,1} $&$z=0 ,\; x\neq0$& $f_1=e_1, \;f_2=-\frac{1}{x}e_4, \;f_3=-\frac{y}{x}e_1+e_2 ,\; f_4=-\frac{1}{x}
e_3 $& $\mathfrak{r}_{4,0}$\\
\hline
$\mathcal{C}_{3,1} $&$z=0 ,\; x=0$& $f_1=-e_2, \;f_2=e_2-e_3, \;f_3=ye_1-e_4 ,\; f_4=
e_1 $& $ \mathfrak{r} \mathfrak{h}_{3}$\\
\hline
$\mathcal{C}_{3,1} $&$z\neq0 ,\; x=0$& $f_1=\frac{1}{z} e_3, \;f_2=e_4, \;f_3=-2ye_1-2ze_2+e_4 ,\; f_4=
e_1 $& $ \mathfrak{r} \mathfrak{r}_{3,-1}$\\
\hline
$\mathcal{C}_{3,1} $&$z\neq0,\; -1\leq \frac{-x}{z}<1$& $f_1=e_4, \;f_2=\frac{2yz}{x-z}e_1-2ze_2+e_4, \;f_3=e_1 ,\; f_4=\frac{1}{z}
e_3 $& $ \mathfrak{r}_{4,-1,-\frac{x}{z}}$\\
\hline
$\mathcal{C}_{3,1} $&$z\neq0,\; x\neq0,\; \frac{-x}{z}>1$& $f_1=e_1, \;f_2=\frac{2yz}{x-z}e_1-2ze_2+e_4, \;f_3=e_4, \; f_4=-\frac{1}{x}
e_3 $& $ \mathfrak{r}_{4,\frac{z}{x},-\frac{z}{x}}$\\
\hline
$\mathcal{C}_{3,1} $&$z\neq0,\; x\neq0,\; \frac{-x}{z}<-1$& $f_1=e_1, \;f_2=e_4, \;f_3= \frac{2yz}{x-z}e_1-2ze_2+e_4, \; f_4=-\frac{1}{x}
e_3 $& $ \mathfrak{r}_{4,-\frac{z}{x},\frac{z}{x}}$\\
\hline
$\mathcal{C}_{3,1} $&$z\neq0,\; \frac{-x}{z}=1$& $f_1=-\frac{y}{2x}e_1+e_{2}+\frac{1}{2x}e_{4}, \;f_2=e_4, \;f_3= e_1, \; f_4=\frac{1}{x}
e_3 $& $ \mathfrak{r}_{4,-1,-1}$\\
\hline
$\mathcal{C}_{3,2}$&$z=0 ,\; x=0$& $f_1=-e_2, \;f_2=e_2-e_3, \;f_3=ye_1-e_4 ,\; f_4=
e_1 $& $ \mathfrak{r} \mathfrak{h}_{3}$\\
\hline
$\mathcal{C}_{3,2} $&$z=0 ,\; x\neq0$& $f_1=e_4, \;f_2=-\frac{y}{x}e_1+ e_2+\frac{1}{x} e_4, \;f_3=-xe_1 \; f_4=
-\frac{1}{x} e_3 $& $\mathfrak{d}_{4,1}$\\
\hline
$\mathcal{C}_{3,2} $&$z\neq0 ,\; x=0$& $f_1=\frac{1}{z} e_3, \;f_2=e_4, \;f_3=-2ye_1-2ze_2+e_4 \; f_4=
e_1 $& $ \mathfrak{r} \mathfrak{r}_{3,-1}$\\
\hline
$\mathcal{C}_{3,2} $&$z\neq0 ,\; x\neq0,\; \frac{z}{x}=\frac{1}{2}$& $f_1=-\frac{1}{2z} e_4, \,f_2=-\frac{y}{z} e_1+e_2, \,f_3=e_1 \, f_4=-\frac{1}{2z}
e_3 $& $\mathfrak{h}_{4}$\\
\hline
$\mathcal{C}_{3,2} $&$\neq0,\; x\neq0,\, \frac{z}{x}<\frac{1}{2}$& $f_1=\frac{y(x-2z)}{z}e_1+e_4, \,f_2=(x-2z)e_2+e_4, \;f_3= -x(x-2z)e_1,\, f_4=-\frac{y(x-2z)}{x^{2}}
e_2-\frac{1}{x}e_3 $& $ \mathfrak{d}_{4,\frac{x-z}{x}}$\\
\hline
$\mathcal{C}_{3,2} $&$z\neq0,\, x\neq0,\,\frac{z}{x}>\frac{1}{2}$& $f_1=(x-2z)e_2+e_4, \,f_2=\frac{y(x-2z)}{z}e_1+e_4, \,f_3= x(x-2z)e_1,\, f_4=-\frac{y(x-2z)}{x^{2}}
e_2-\frac{1}{x}e_3 $& $ \mathfrak{d}_{4,\frac{z}{x}}$\\
\hline
$\mathcal{C}_{4,1}$&& $f_1=-ye_1-xe_2+e_3, \;f_2=e_1, \;f_3=-xe_1+e_4 \; f_4=
-xe_1-e_2+e_4 $& $ \mathfrak{d}_{4,1}$\\
\hline
$\mathcal{C}_{4,2}$&& $f_1=e_3, \;f_2=e_1, \;f_3=-xe_2+e_4 \; f_4=
-e_2 $& $ \mathfrak{d}_{4,1}$\\
\hline
$\mathcal{C}_{5,1}^{+}$&& $f_1=\frac{1}{2} e_1-\frac{1}{2}e_2, \;f_2=(-x+y)e_1+(x-y)e_2-e_3+e_4, \;f_3=-(\frac{1}{2}+x+y)e_1-(\frac{1}{2}+x+y)e_2+e_3+e_4, \; f_4=
-(x+y)e_1-(x+y)e_2+e_3+e_4$& $\mathfrak{r}_{2} \mathfrak{r}_{2}$\\
\hline
$\mathcal{C}_{5,2}^{+}$&& $f_1=\frac{1}{2} e_1-\frac{1}{2}e_2, \,f_2=-xe_2-e_3+e_4, \;f_3=-\frac{1}{2}e_1-\frac{1}{2}e_2,\, f_4=
-xe_2+e_3+e_4 $& $ \mathfrak{r}_{2}\mathfrak{r}_{2}$\\
\hline
$\mathcal{C}_{5,1}^{-}$&& $f_1=ye_1-(x+1)e_2+e_3, \;f_2=-(x+1)e_1-ye_2+e_4, \;f_3=ye_1-xe_2+e_3 \; f_4=
-xe_1-ye_2+e_4 $& $ \mathfrak{r}_{2}'$\\
\hline
$\mathcal{C}_{5,2}^{-}$&& $f_1=-e_2-e_3, \;f_2=e_1-xe_2+e_4, \;f_3=-e_3 \; f_4=
-xe_2+e_4 $& $ \mathfrak{r}_{2}'$\\
\hline
\caption{Isomorphisms from the Lie algebras obtained in Table \ref{tab5} onto the Lie algebras in Table \ref{tab1}}
\label{tab7}
\end{longtable}
\end{center}}
\section*{Acknowledgments:} The authors would like to thank sincerely Professor Mohamed Boucetta for his many suggestions which were of great help to improve our work.
|
\section{Introduction}
\label{Sec:Introduction}
The empirical hadron spectrum is rich \cite{Zyla:2020zbs}, even without including the array of recently discovered exotic states \cite{Richard:2016eis, Esposito:2016noz, Liu:2019zoy, Barabanov:2020jvn, Eichmann:2020oqt}. Yet, contemporary theory still predicts more states than have been observed. This is especially true when one goes beyond systems comprised of light $\{u,d\}$-quarks and includes hadrons seeded by all flavours of valence quarks and/or antiquarks whose lifetime is long enough to produce measurable bound states, \emph{i.e}.\ also includes systems containing $\{s, c, b\}$ quarks and/or related antiquarks. The empirically ``missing'' states are being sought in a worldwide effort \cite{Galanti:2018bnp, KETZER2020103755, Carman:2020qmb, Brodsky:2020vco, NStar2020}.
Regarding calculations of hadron spectra, numerical simulations of lattice-regularised quantum chromodynamics (lQCD) provide a direct connection with the Standard Model Lagrangian. Many collaborations are tackling the problem. There are successes and challenges \cite{Bazavov:2009bb, Fodor:2012gf, Briceno:2017max, Hansen:2019nir}; and spectrum calculations are described, \emph{e.g}.\ in Refs.\,\cite{Durr:2008zz, Edwards:2011jj, Brown:2014ena, Lubicz:2017asp, Bali:2017pdv, Mathur:2018epb, Bahtiyar:2020uuj}. Notably, few results are available on negative parity states and hadron radial excitations.
Explaining the mass and structure of parity partners in hadron spectra is crucial to understanding strong interactions because they would be degenerate if chiral symmetry were not dynamically broken, as discussed for the $a_1$- and $\rho$-mesons in Ref.\,\cite{Weinberg:1967kj}. This dynamical chiral symmetry breaking (DCSB) is a corollary of emergent hadronic mass (EHM), the mechanism responsible for the $m_p \sim 1\,$GeV scale characteristic of visible matter \cite{Roberts:2020udq, Roberts:2020hiw, Roberts:2021xnz, Roberts:2021nhw}. Current-quark masses enter into quantum chromodynamics (QCD) via the Higgs mechanism of explicit chiral symmetry breaking; so studying the spectra of states built from $\{u,d,s, c, b\}$ quarks and/or related antiquarks opens the way to exploring constructive interference between Nature's two distinct mass generating mechanisms.
Quark models have also been widely employed in the calculation of hadron spectra; see, \emph{e.g}.\
Refs.\,\cite{Aznauryan:2012ba, Crede:2013sze, Segovia:2013wma, Giannini:2015zia, Fernandez:2021zjq}.
Owing to complications introduced by the need to preserve chiral symmetry and its pattern of breaking in QCD, \emph{i.e}.\ to faithfully express DCSB, quark models are most reliable for baryons and mesons containing heavy quarks. In such systems, Higgs-induced current-quark masses play the dominant role and differences between current- and constituent-quarks are least noticeable; see, \emph{e.g}.\ Ref.\,\cite[Fig.\,2.5]{Roberts:2021nhw}.
During the past decade, continuum Schwinger function methods, particularly Dyson-Schwinger equations, have increasingly been used in spectrum calculations following improvements in both \cite{Horn:2016rip, Eichmann:2016yit, Burkert:2017djo, Fischer:2018sdj, Barabanov:2020jvn, Qin:2020rad}: (\emph{i}) understanding the capacities and limitations of the approach; and (\emph{ii}) the quality and range of the description of hadron properties.
Recently, predictions for meson and baryon spectra in some of the low-lying flavour-SU$(N_f=5)$ multiplets were delivered \cite{Qin:2018dqp, Qin:2019hgk}. They were made using the rainbow-ladder (RL) truncation of the continuum bound-state equations. This is the leading-order in a systematic scheme \cite{Binosi:2016rxz, El-Bennich:2016qmb, Binosi:2016wcx, Fu:2017azw, Williams:2018adr, Xu:2018cor, Oliveira:2018fkj, Qin:2020jig}, which enables both: (\emph{i}) a unified, symmetry-preserving description of mesons and baryons, accounting properly for DCSB; and (\emph{ii}) a traceable connection to quantum chromodynamics (QCD). A next challenge here is to advance beyond the leading-order truncation. That effort will likely benefit from the use of high-performance computing.
Complementing such studies, Refs.\,\cite{Yin:2019bxe, Gutierrez-Guerrero:2019uwa} followed a different, largely algebraic path to calculating the spectra of ground-state pseudoscalar and vector $(f\bar g)$ mesons and $J^P=1/2^+, 3/2^+$ $(fgh)$ baryons, where $f,g,h \in \{u,d,s,c,b\}$. Exploiting the fact that a had-ron's mass is a volume-integrated (long-wavelength) \linebreak quantity, so not very sensitive to details of the system's wave function, Refs.\,\cite{Yin:2019bxe, Gutierrez-Guerrero:2019uwa} used a symmetry-preserving treatment of a vector$\times$vector contact interaction (CI) \cite{Roberts:2011wy} to deliver insights into features of these systems that can be obscured in approaches that rely heavily on computer resources. For instance, the analysis found diquark correlations to be important in all baryons studied; and owing to the dynamical character of the diquarks, the lightest allowed diquark correlation is typically that which defines the dominant component of a baryon's Faddeev amplitude.
Herein, we extend Refs.\,\cite{Yin:2019bxe, Gutierrez-Guerrero:2019uwa} and explore the capacity of the contact interaction to also explain the spectra of scalar and pseudovector mesons and $J^P=1/2^-$, $3/2^-$ baryons.
The meson spectrum is canvassed in Sec.\,\ref{Sec:Meson}, along with a discussion of leptonic decay constants, including those of scalar and pseudovector states. The masses and decay constants of $\pi, K, \eta_c, \eta_b$ mesons are used to determine the current-masses of the $\{u=d,s,c,b\}$ quarks and basic CI parameters. Predictions are then made for another sixty-six quantities.
Connections with diquark masses and correlation amplitudes are drawn in Sec.\,\ref{SecDiquarks}. They are important because soft diquarks seemingly play an important role in hadron structure \cite{Barabanov:2020jvn}; hence, serve usefully in developing a quark+diquark approximation to the baryon three-body problem \cite{Cahill:1988dx, Burden:1988dt, Reinhardt:1989rw, Efimov:1990uz}.
Results for the spectra of ground-state $J=1/2^\pm$, $3/2^\pm$ baryons are discussed in Sec.\,\ref{Sec:Baryons}. Following Refs.\,\cite{Roberts:2011cf, Eichmann:2016hgl, Lu:2017cln}, eight parameters are introduced to complete the CI definition in the baryon sector, with four serving to correct for omission of resonant (meson cloud) contributions in the quark+diquark Faddeev equation. Masses for eighty baryons subsequently emerge as predictions along with their CI spin-flavour Faddeev amplitudes.
A summary and perspective is provided in Sec.\,\ref{Sec:Epilogue}.
Appendices are included to sketch the CI, \ref{Sec:Contact}; and detail the structure of flavour-SU$(5)$ baryon Faddeev amplitudes, \ref{AppFaddeev}.
\section{Meson Spectrum}
\label{Sec:Meson}
Ref.\,\cite{Yin:2019bxe} calculated the masses and leptonic decay constants of ground-state pseudoscalar and vector $(f\bar g)$ me-sons, $f,g \in \{u,d,s,c,b\}$. Herein, we expand the coverage to include kindred scalar and pseudovector mesons, \emph{viz}.\ their parity partners. All CI bound-states satisfy a Bethe-Salpeter equation of the form in Eq.\,\eqref{LBSEI}. The pseudoscalar solution amplitude is given in Eq.\,\eqref{PSBSA} and the other systems are described by the following amplitudes:
\begin{subequations}
\label{fDecayConstant}
\begin{align}
\Gamma_\mu^{1^{-}}(Q) &= \gamma_\mu^\perp E_{1^{-}}(Q)\,,\\
\Gamma^{0^{+}}(Q) &= \textrm{I}_{D}E_{0^{+}}(Q)\,,\\
\Gamma_\mu^{1^{+}}(Q) &= \gamma_5 \gamma_\mu^\perp E_{1^{+}}(Q)\,,
\end{align}
\end{subequations}
where $Q_\mu \gamma_\mu^\perp = 0$ and $\textrm{I}_{D}$ is the $4\times 4$ identity matrix in spinor space. When a momentum-dependent interaction is used, the amplitudes have a richer structure \cite{LlewellynSmith:1969az, Krassnigg:2009zh}; but the terms in Eqs.\,\eqref{fDecayConstant} are dominant \cite{Maris:2000ig, Chang:2011ei, Qin:2011xq}.
In terms of the amplitudes in Eqs.\,\eqref{fDecayConstant}, canonically normalised via Eq.\,\eqref{normcan} and its analogues, the leptonic decay constants are calculated using $(t_+ = t+Q)$
\begin{subequations}
\begin{align}
f_{1^{-}}m_{1^{-}} &= \frac{N_{c}}{3}\int\frac{\textrm{d}^{4}t}{(2\pi)^{4}}\textrm{tr}\big[\gamma_{\mu}S_{f}(t_+)
\Gamma_{\mu}^{1^{-}}(Q)S_{g}(t)\big]\,, \\
f_{0^{+}}P_{\mu}&= N_{c}\int\frac{\textrm{d}^{4}t}{(2\pi)^{4}}\textrm{tr}\big[i\gamma_{\mu}S_{f}(t_+)\Gamma^{0^{+}}(Q)
S_{g}(t)\big]\,,\\
f_{1^{+}}m_{1^{+}}&=\frac{N_{c}}{3}\int\frac{\textrm{d}^{4}t}{(2\pi)^{4}} \textrm{tr}\big[\gamma_5\gamma_{\mu}S_{f}(t_+)
\Gamma_{\mu}^{1^{+}}(Q)S_{g}(t)\big]\,,
\end{align}
\end{subequations}
with $N_c=3$, and $f_{0^-}$ given in Eq.\,\eqref{ffg}. In $f=g$ systems \cite{Bhagwat:2006py}: $f_{0^{+}}\equiv 0$.
\begin{table}[t]
\caption{\label{Tab:MesonSpectrum}
Mesons: computed masses, Bethe-Salpeter amplitudes, and decay constants.
Empirical masses from Ref.\,\cite{Zyla:2020zbs};
entries marked by ``$\ast$'',
from Ref.\,\cite{Mathur:2018epb}.
Empirically unknown decay constants quoted from lQCD \cite{Becirevic:1998ua, Chiu:2007bc, Davies:2010ip, McNeile:2012qf, Donald:2012ga, Colquhoun:2015oha, Bali:2017pdv}.
(Dimensioned quantities in GeV.
Underlined entries from Table~\ref{Tab:DressedQuarks}.
``---'' indicates no empirical/lQCD results available for comparison.)
}
\begin{tabular*
{\hsize}
{c@{\extracolsep{0ptplus1fil}}
l@{\extracolsep{0ptplus1fil}}||
l@{\extracolsep{0ptplus1fil}}
l@{\extracolsep{0ptplus1fil}}|
l@{\extracolsep{0ptplus1fil}}
l@{\extracolsep{0ptplus1fil}}|
l@{\extracolsep{0ptplus1fil}}
l@{\extracolsep{0ptplus1fil}}
}
\hline\hline
$J^{P}$ & Meson & $m^{\rm CI}$ & $m^{\rm e/l}$ & $E$ & $F$ & $f^{\rm CI}$ & $f^{\rm e/l}$ \\
\hline
$0^{-}$ & $\pi(u\bar{d})$ & \underline{0.14} & 0.14 & 3.59 & 0.47$\ $ & \underline{0.10} & 0.092 \\
& $K(u\bar{s})$ & \underline{0.50} & 0.50 & 3.82 & 0.59$\ $ & \underline{0.11} & 0.11 \\
& $D(u\bar{c})$ & 1.93 & 1.87 & 3.09 & 0.36$\ $ & 0.16 & 0.15 \\
& $D_{s}(s\bar{c})$ & 2.01 & 1.97 & 3.23 & 0.48$\ $ & 0.17 & 0.18 \\
& $\eta_{c}(c\bar{c})$ & \underline{2.98} & 2.98 & 3.25 & 0.72$\ $ & \underline{0.24} & 0.24 \\
& $B(u\bar{b})$ & 5.42 & 5.28 & 1.67 & 0.095$\ $ & 0.17 & 0.14 \\
& $B_{s}(s\bar{b})$ & 5.50 & 5.37 & 1.79 & 0.14$\ $ & 0.18 & 0.16 \\
& $B_{c}(c\bar{b})$ & 6.28 & 6.28 & 3.38 & 0.61$\ $ & 0.27 & 0.35 \\
& $\eta_{b}(b\bar{b})$ & \underline{9.40} & 9.40 & 3.18 & 0.81$\ $ & \underline{0.41} & 0.41 \\
\hline
$1^{-}$ & $\rho(u\bar{d})$ & 0.93 & 0.78 & 1.53 & & 0.13 & 0.15 \\
& $K^{\ast}(u\bar{s})$ & 1.03 & 0.89 & 1.63 & & 0.12 & 0.16 \\
& $\phi(s\bar{s})$ & 1.13 & 1.02 & 1.74 & & 0.12 & 0.17 \\
& $D^{\ast}(u\bar{c})$ & 2.14 & 2.01 & 1.20 & & 0.15 & 0.17 \\
& $D_{s}^{\ast}(s\bar{c})$ & 2.23 & 2.11 & 1.22 & & 0.16 & 0.19 \\
& $J/\psi(c\bar{c})$ & 3.19 & 3.10 & 1.19 & & 0.20 & 0.29 \\
& $B^{\ast}(u\bar{b})$ & 5.47 & 5.33 & 0.70 & & 0.16 & 0.12 \\
& $B_{s}^{\ast}(s\bar{b})$ & 5.56 & 5.42 & 0.71 & & 0.16 & 0.15 \\
& $B_{c}^{\ast}(c\bar{b})$ & 6.39 & 6.33$^\ast$ & 1.37 & & 0.23 & 0.30 \\
& $\Upsilon(b\bar{b})$ & 9.49 & 9.46 & 1.48 & & 0.38 & 0.46 \\
\hline
$0^{+}$ & $\sigma(u\bar{d})$ & 1.22 & --- & 0.66 & & 0 & --- \\
& $\kappa(u\bar{s})$ & 1.34 & --- & 0.65 & & 0.023 & --- \\
& $D_{0}^{*}(u\bar{c})$ & 2.39 & 2.30 & 0.37 & & 0.11 & --- \\
& $D_{s0}^{*}(s\bar{c})$ & 2.51 & 2.32 & 0.35 & & 0.083 & $0.081(6)$ \\
& $\chi_{c0}(c\bar{c})$ & 3.46 & 3.42 & 0.23 & & 0 & --- \\
& $B_{0}^{\ast}(u\bar{b})$ & 5.64 & --- & 0.22 & & 0.19 & --- \\
& $B_{s0}^{\ast}(s\bar{b})$& 5.74 & --- & 0.21 & & 0.17 & --- \\
& $B_{c0}^{\ast}(c\bar{b})$& 6.63 & 6.71$^\ast$ & 0.18 & & 0.068 & --- \\
& $\chi_{b0}(b\bar{b})$ & 9.78 & 9.86 & 0.10 & & 0 & --- \\
\hline
$1^{+}$ & $a_{1}(u\bar{d})$ & 1.37 & 1.23 & 0.32 & & 0.29 & --- \\
& $K_{1}(u\bar{s})$ & 1.48 & 1.25 & 0.32 & & 0.27 & --- \\
& $f_{1}(s\bar{s})$ & 1.59 & 1.43 & 0.32 & & 0.25 & --- \\
& $D_{1}(u\bar{c})$ & 2.48 & 2.42 & 0.19 & & 0.32 & --- \\
& $D_{s1}(s\bar{c})$ & 2.59 & 2.46 & 0.18 & & 0.31 & $0.14(1)$ \\
& $\chi_{c1}(c\bar{c})$ & 3.51 & 3.51 & 0.12 & & 0.28 & --- \\
& $B_{1}(u\bar{b})$ & 5.70 & 5.73 & 0.11 & & 0.39 & --- \\
& $B_{s1}(s\bar{b})$ & 5.79 & 5.83 & 0.11 & & 0.37 & --- \\
& $B_{c1}(c\bar{b})$& 6.66 & 6.74$^\ast$ & 0.096 & & 0.24 & --- \\
& $\chi_{b1}(b\bar{b})$ & 9.79 & 9.89 & 0.055 & & 0.22 & --- \\
\hline\hline
\end{tabular*}
\end{table}
The explicit CI form for the vector-meson Bethe-Salpeter equation may be obtained simply by generalising Eqs.\,(18)\,--\,(20) in Ref.\,\cite{Chen:2012qr}; and those for the scalar and pseudovector channels from Ref.\,\cite[Appendices\,B.3, B.4]{Chen:2012qr}.
The latter, however, deserve additional attention.
It has long been known that RL truncation has some defects in the $0^+$, $1^+$ channels because these systems possess significant orbital angular momentum and RL fails to generate sufficient spin-orbit repulsion \cite{Chang:2009zb, Fischer:2009jm, Chang:2011ei, Qin:2020jig}. Nonperturbatively generated corrections to RL truncation, such as those originating in EHM-induced dressed-quark anomalous chromomagnetic moments \cite{Chang:2010hb}, can remedy this defect and generate the empirical splitting between parity partners \cite{Chang:2011ei} and radial excitations \cite{Qin:2020jig}. (Such corrections largely cancel in $0^-$ and $1^-$ channels.) An ameliorating expedient was introduced in \linebreak Ref.\,\cite{Roberts:2011cf}, and refined in Refs.\,\cite{Eichmann:2016hgl, Lu:2017cln}. Namely, one includes a multiplicative factor in the Bethe-Salpeter kernel for each of these channels, $g_{\rm SO}^{0^+}$, $g_{\rm SO}^{1^+}$, with $g_{\rm SO}^{0^+}$ chosen to produce a mass difference of approximately $0.3\,$GeV between the quark-core of the $0_{u\bar d}^+$ and that of the $\rho$-meson (as obtained with beyond-RL kernels \cite{Chang:2011ei}) and $g_{\rm SO}^{1^+}$ fixed to obtain the empirical size of the $a_1$-$\rho$ mass-splitting:
\begin{equation}
\label{SOfactors}
g_{\rm SO}^{0^+}=0.32\,, \quad g_{\rm SO}^{1^+} = 0.25\,.
\end{equation}
(\emph{N.B}.\ In a Poincar\'e-covariant treatment, no bound state is purely $S$- or $P$-wave; and $g_{\rm SO}=1$ indicates no additional interaction beyond that generated by the RL kernel.)
At this point, all current-quark masses are determined; the coupling and cutoff in a given flavour channel are fixed using Eqs.\,\eqref{alphaLambda}, \eqref{alphaIRMass}, with the argument being the empirical pseudoscalar meson mass in this channel; and the spin-orbit correction factors are defined in Eq.\,\eqref{SOfactors}.
Consequently, apart from the four underlined entries, every CI result in Table~\ref{Tab:MesonSpectrum} is a prediction. Recalling Ref.\,\cite{Yin:2019bxe}, the $F_{0^-}$ (pseudovector) component of each pseudoscalar meson is nonzero. On average, it is 15(6)\% of the $E_{0^-}$ (pseudoscalar) piece; so, the pseudovector component is quantitatively important in all cases.
To assist with understanding the comparisons in Table~\ref{Tab:MesonSpectrum}, we also represent them in Fig.\,\ref{FigMeson}.
Considering Fig.\,\ref{FigMeson}A, one sees that the CI delivers good estimates for the masses of both positive and negative parity ground-state flavour-SU$(5)$ mesons. Aspects of its symmetry-preserving formulation skew results toward overestimates in most cases \cite{Roberts:2010rn}, but the mean relative-difference between theory and experiment/lQCD is only 5(6)\%. (This value is obtained by comparing columns~1 and 2 in Table~\ref{Tab:MesonSpectrum}, omitting the underlined entries in column~1.)
\begin{figure*}[t]
\vspace*{2ex}
\leftline{\hspace*{0.5em}{\large{\textsf{A}}}}
\vspace*{-3ex}
\includegraphics[width=0.95\textwidth]{F1A.pdf}
\vspace*{1ex}
\leftline{\hspace*{0.5em}{\large{\textsf{B}}}}
\vspace*{-3ex}
\includegraphics[width=0.95\textwidth]{F1B.pdf}
\caption{\label{FigMeson}
\emph{Upper panel}\,--\,{\sf A}.
Comparison between CI predictions for meson masses and available experiment \cite{Zyla:2020zbs} and $m_{B^\ast_c}$, $m_{B_{c0}^{\ast}}$, $m_{B_{c1}}$ from Ref.\,\cite{Mathur:2018epb}.
\emph{Lower panel}\,--\,{\sf B}.
CI predictions for meson leptonic decay constants compared with experiment \cite{Zyla:2020zbs}, where known, and lQCD otherwise \cite{Becirevic:1998ua, Chiu:2007bc, Davies:2010ip, McNeile:2012qf, Donald:2012ga, Colquhoun:2015oha}.
In both panels, contact-interaction predictions are depicted as (blue) circles and comparison values by (green) bars.
(Pictorial representation of results in Table~\ref{Tab:MesonSpectrum}.)
}
\end{figure*}
Furthermore, as found with pseudoscalar and vector mesons in Ref.\,\cite{Yin:2019bxe}, the computed masses of scalar and pseudovector mesons neatly follow a pattern prescribed by equal spacing rules (ESRs) \cite{Okubo:1961jc, GellMann:1962xb, Qin:2018dqp, Chen:2019fzn, Qin:2019hgk}, \emph{e.g}.\
\begin{subequations}
\label{ESRexamples}
\begin{align}
m_{K_1} & \approx (m_{f_1}+m_{a_1})/2 \,,\\
m_{B_{s0}^{\ast}} - m_{B_{0}^\ast} & \approx m_{D_{s0}^\ast} - m_{D_{0}^\ast} \nonumber \\
& \approx m_{B_{s1}} - m_{B_{1}}
\approx m_{D_{s1}} - m_{D_{1}} \,,\\
m_{B_{c1}}-m_{B_{s1}} & \approx m_{B_{c0}^{\ast}}-m_{B_{s0}^{\ast}}\,,\\
(m_{\chi_{b1}} - m_{\chi_{c1}})/2 & \approx (m_{\chi_{b0}} - m_{\chi_{c0}})/2 \nonumber \\
& \approx m_{B_{s1}} - m_{D_{s1}}
\!\!\approx m_{B_{s0}^{\ast}} - m_{D_{s0}^\ast} . \label{MbMc}
\end{align}
\end{subequations}
These relations expose the scales that may be identified with the splittings between the spectrum-generating constituent-like quark masses, \emph{e.g}.\ Eqs.\,\eqref{MbMc} reveal a $b-c$ spectrum mass difference ${\mathsf d}_M^{bc} = 3.18(4)\,$GeV. This is a fair match with $M_b-M_c = 3.23\,$GeV from Table~\ref{Tab:DressedQuarks}. As evident in this table, the Higgs mechanism of mass generation is the origin of such splittings.
Interesting patterns are also found by comparing parity partners, \emph{e.g}.\
{\allowdisplaybreaks
\begin{subequations}
\begin{align}
0.44(1)\,{\rm GeV:} & \ m_{a_1} - m_{\rho} \approx m_{K_1} - m_{K^\ast} \nonumber \label{a1}\\
& \quad \approx m_{f_1}-m_\phi\,, \\
0.34(2)\,{\rm GeV:} & \ m_{\chi_{c1}} - m_{J/\psi} \approx m_{D_{s1}} - m_{D_s^\ast} \nonumber \\
& \quad \approx m_{D_1} - m_{D^\ast} \,, \label{Xc1}\\
0.48(2)\,{\rm GeV:} & \ m_{\chi_{c0}} - m_{\eta_c} \approx m_{D_{s0}^\ast} - m_{D_s} \nonumber \\
& \quad \approx m_{D_{0}^\ast} - m_{D} \,, \label{Xc0}\\
0.25(4)\,{\rm GeV:} & \ m_{\chi_{b1}} - m_{\Upsilon} \approx m_{B_{c1}}-m_{B_c^\ast} \nonumber \\ & \quad \approx m_{B_{s1}}-m_{B_s^\ast} \approx m_{B_1}-m_{B^\ast} \,, \label{Xb1}\\
0.30(8)\,{\rm GeV:} & \ m_{\chi_{b0}} - m_{\eta_b} \approx m_{B_{c0}^{\ast}}-m_{B_c} \nonumber \\ & \quad \approx m_{B_{s0}^{\ast}}-m_{B_s} \approx m_{B_{0}^{\ast}}-m_{B}\,. \label{Xb0}
\end{align}
\end{subequations}}
\hspace*{-0.4\parindent}These results reveal the scale of DCSB in meson spectra. Overall, parity partners are split by $0.35(10)\,$GeV.
However, comparing the impact of Higgs couplings into QCD, the expression of DCSB becomes weaker with increasing current-quark mass: Eq.\,\eqref{a1} \emph{cf}.\ \eqref{Xc1} \emph{cf}.\ \eqref{Xb1}; and Eq.\,\eqref{Xc0} \emph{cf}.\ \eqref{Xb0}.
Moreover, DCSB generates larger scalar-pseudoscalar splittings than pseudovector-vector: Eq.\,\eqref{Xb0} cf.\ \eqref{Xb1}; and Eq.\,\eqref{Xc0} \emph{cf}.\ \eqref{Xc1}.
CI results for meson leptonic decay constants are also listed in Table~\ref{Tab:MesonSpectrum} and drawn in Fig.\,\ref{FigMeson}B. These quantities describe quark+antiquark annihilation at a single spacetime point, so they are sensitive to ultraviolet physics. In QCD, this is expressed through the appearance of a logarithmic ultraviolet divergence in integrals like those in Eqs.\,\eqref{fDecayConstant}, which is compensated by the dressed-quark wave-function renormalisation constant \cite{Maris:1997hd, Ivanov:1998ms}. It is therefore unsurprising that the cutoff-regularised CI supplies a poorer description of meson decay constants than it does of their masses. It would be worse if Eqs.\,\eqref{alphaLambda}, \eqref{alphaIRMass} were not implemented. Nevertheless, in comparison with known empirical or lQCD values, the picture is fair: trends are generally reproduced and the mean-absolute-relative-difference between the entries in columns~5 and 6 of Table~\ref{Tab:MesonSpectrum} is 15(11)\%.
(Given that very little is known about the decay constants of scalar and pseudovector mesons, they are excluded from this comparison.)
Peculiarities of the CI's formulation mean that the description is better for pseudoscalars than it is for vector mesons \cite{Roberts:2010rn}.
Further improvement can be achieved by additional tuning of Eq.\,\eqref{alphaLambda}, including its extension into the light-quark sector and consideration of the particular properties of heavy+light mesons.
Notwithstanding these observations, as with pseudoscalar and vector mesons, analogues of Eqs.\,\eqref{ESRexamples} are applicable, \emph{e.g}.\
{\allowdisplaybreaks
\begin{subequations}
\label{ESRfexamples}
\begin{align}
f_{K_1} & \approx [ f_{a_1} + f_{f_1} ]/2 \,,\\
f_{B_{c1}} & \approx [ f_{\chi_{c1}} + f_{\chi_{b1}} ]/2 \,,\\
f_{B_{0}^\ast}-f_{B_{s0}^{\ast}} & \approx f_{D_{0}^\ast}- f_{D_{s0}^\ast} \nonumber \\
& \approx f_{B_{1}}- f_{B_{s1}}
\approx f_{D_{1}}-f_{D_{s1}} \,,\\
f_{B_{s1}}-f_{B_{c1}} & \approx f_{B_{s0}^{\ast}}-f_{B_{c0}^{\ast}}\,,\\
f_{B_{s1}} - f_{D_{s1}} & \approx f_{B_{s0}^{\ast}} - f_{D_{s0}^\ast} \,.
\end{align}
\end{subequations}}
Such relations are not surprising when one recalls that such decay constants are order parameters for chiral symmetry breaking \cite{Bender:1996bm}.
Regarding Fig.\,\ref{FigMeson}B, one finds that using the CI the decay constants of vector mesons and those of pseudovector mesons have the opposite tendency with increasing current-quark mass, \emph{e.g}.\
\begin{align}
& f_{B^\ast} < f_{B_s^\ast}< f_{B_c^\ast} < f_\Upsilon \nonumber \\
\mbox{\emph{cf}.}\ & f_{B_1} > f_{B_{s1}}> f_{B_{c1}} > f_{\chi_{b1}}.
\end{align}
It is currently unclear whether this is a CI artefact. However, it does follow the trend set by the meson Bethe-Salpeter amplitudes in Table~\ref{Tab:MesonSpectrum} and match qualitatively with the behaviour of scalar-meson decay constants. The question could be answered by using a mo-mentum-dependent quark+antiquark scattering kernel \cite{Qin:2011xq}.
\section{Diquark Mass Spectrum}
\label{SecDiquarks}
In approaching the problem of calculating the spectrum of flavour-SU$(5)$ ground-state baryons with $J^P = 1/2^\pm, 3/2^\pm$, we use a quark+diquark approximation to the Faddeev equation introduced elsewhere \cite{Cahill:1988dx, Burden:1988dt, Reinhardt:1989rw, Efimov:1990uz}. As apparent in Fig.\,\ref{figFaddeev}, here the diquarks are fully dynamical, appearing in a Faddeev kernel that requires their continual breakup and reformation. In consequence, they are vastly different from the static, pointlike degrees-of-freedom considered in early models of baryons \cite{Anselmino:1992vg}.
\begin{figure}[t]
\centerline{%
\includegraphics[clip, width=0.45\textwidth]{F2.pdf}}
\caption{\label{figFaddeev}
Covariant Faddeev equation: linear integral equation satisfied by the matrix-valued Faddeev amplitude, $\Psi$, for a baryon of total momentum $P= p_d+p_q$, which expresses the relative-momentum correlation between the quarks and nonpointlike-diquarks within the baryon. The shaded rectangle is the Faddeev equation kernel:
\emph{single line}, dressed-quark propagator (Appendix~\ref{Sec:Contact}); $\Gamma$, diquark correlation amplitude (Sec.\,\ref{SecDiquarks}); and \emph{double line}, diquark propagator (Eqs.\,\eqref{diquarkprops}.) }
\end{figure}
To proceed, therefore, one must calculate the masses and amplitudes for all diquark correlations that can exist in $J^P = 1/2^\pm, 3/2^\pm$ baryons. Using the CI, this means the following colour-$\bar 3$ correlations: flavour-$\overline{10}$ scalar; flavour-${15}$ pseudovector; flavour-$\overline{10}$ pseudoscalar; and flavour-$\overline{10}$ vector.\footnote{A $\bar 3_c$ flavour-$15$ vector-diquark is generally possible, but is not supported by a RL-like treatment of the CI \cite{Lu:2017cln}. Using momentum-dependent interactions, flavour-symmetric vector correlations remain strongly suppressed \cite{Chen:2017pse}.}
This is a straightforward exercise because the RL Bethe-Salpeter equation for a $J^P$ diquark is generated from that for a $J^{-P}$ meson by simply multiplying the meson kernel by a factor of $1/2$ \cite{Cahill:1987qr}. For instance, working from Eq.\,\eqref{LBSEI}, the mass and amplitude for a flavour-$\overline{10}$ scalar diquark is obtained from the following equation:
\begin{equation}
\Gamma^C_{[fg]}(Q) = - \frac{8 \pi}{3} \frac{\alpha_{\rm IR}}{m_G^2}
\int \!\! \frac{d^4t}{(2\pi)^4} \! \gamma_\mu S_f(t_+) \Gamma^C_{[fg]}(Q) S_g(t) \gamma_\mu \,,
\label{LBSEqq}
\end{equation}
where the correlation amplitude is $\Gamma_{[fg]}(Q)$ and
\begin{subequations}
\begin{align}
\Gamma^C_{[fg]}(Q) & := \Gamma_{[fg]}(Q)C^\dagger\\
& = \gamma_5 \left[ i E_{[fg]}(Q) + \frac{1}{2 M_{f g}} \gamma\cdot Q F_{[fg]}(Q)\right]\,,
\end{align}
\end{subequations}
with $C=\gamma_2\gamma_4$ being the charge-conjugation matrix. The canonical normalisation conditions are likewise \linebreak amended, with the multiplicative factor being $2/3$ in this case.
In a study of flavour-SU$(3)$ baryon spectra \cite{Lu:2017cln}, an additional step was found to be necessary in order to arrive at realistic results. Namely, considering that valence-quarks within a diquark are less tightly correlated than the valence-quark and -antiquark in a bound-state meson, then spin-orbit repulsion in diquarks should be weaker than it is in mesons. This was effected by writing
\begin{equation}
\label{eqsSO}
g_{\rm SO}^{1^-_{qq},0^-_{qq}} = g_{\rm SO}^{1^+_{q\bar q},0^+_{q\bar q}} \times {\mathpzc s}_{\rm SO} \,,\;
{\mathpzc s}_{\rm SO}=1.8\,,
\end{equation}
so that the modification factor in the RL-like diquark Bethe-Salpeter equations is nearer unity; hence, generates less repulsion. The value of ${\mathpzc s}_{\rm SO}$ was chosen in concert with
$g_{DB}^{\pi_\Psi \pi_d}$, Eq.\,\eqref{gDBval}, in order to ensure that the dressed-quark core mass of the nucleon's parity partner is not lower than that of its first radial excitation.
\begin{table}[t]
\caption{\label{diquarkspectrum}
Diquark correlations: computed masses (GeV) and Bethe-Salpeter amplitudes of scalar, axial-vector, pseudoscalar and vector diquarks. \emph{N.B}.\ We work in the isospin-symmetry limit; so, e.g.\ $[ds]_{0^{+}}=[us]_{0^{+}}$.}
\begin{center}
\begin{tabular*
{\hsize}
{l@{\extracolsep{0ptplus1fil}}
l@{\extracolsep{0ptplus1fil}}|
l@{\extracolsep{0ptplus1fil}}
l@{\extracolsep{0ptplus1fil}}|
l@{\extracolsep{0ptplus1fil}}
l@{\extracolsep{0ptplus1fil}}
}
\hline\hline
$J^{P}$ & Diquark$\ $ & $m_{qq}^{CI}$ & $m_{q\bar{q}}^{CI}$ $\ $ & $E_{qq}$ & $F_{qq}$ \\
\hline
$0^{+}$ & $[ud]_{0^{+}}$ & 0.78 & 0.14$\ $ & 2.74 & 0.31 \\
& $[us]_{0^{+}}$ & 0.93 & 0.50$\ $ & 2.88 & 0.39 \\
& $[uc]_{0^{+}}$ & 2.15 & 1.93$\ $ & 1.97 & 0.22 \\
& $[sc]_{0^{+}}$ & 2.26 & 2.01$\ $ & 1.99 & 0.29 \\
& $[ub]_{0^{+}}$ & 5.51 & 5.42$\ $ & 1.05 & 0.059 \\
& $[sb]_{0^{+}}$ & 5.60 & 5.50$\ $ & 1.05 & 0.083 \\
& $[cb]_{0^{+}}$ & 6.48 & 6.28$\ $ & 1.42 & 0.25 \\
\hline
$1^{+}$ & $\{uu\}_{1^{+}}$ & 1.06 & 0.93$\ $ & 1.31 & \\
& $\{us\}_{1^{+}}$ & 1.16 & 1.03$\ $ & 1.36 & \\
& $\{ss\}_{1^{+}}$ & 1.26 & 1.13$\ $ & 1.43 & \\
& $\{uc\}_{1^{+}}$ & 2.24 & 2.14$\ $ & 0.89 & \\
& $\{sc\}_{1^{+}}$ & 2.34 & 2.23$\ $ & 0.87 & \\
& $\{cc\}_{1^{+}}$ & 3.30 & 3.19$\ $ & 0.69 & \\
& $\{ub\}_{1^{+}}$ & 5.53 & 5.47$\ $ & 0.51 & \\
& $\{sb\}_{1^{+}}$ & 5.62 & 5.56$\ $ & 0.50 & \\
& $\{cb\}_{1^{+}}$ & 6.50 & 6.39$\ $ & 0.62 & \\
& $\{bb\}_{1^{+}}$ & 9.68 & 9.49$\ $ & 0.48 & \\
\hline
$0^{-}$ & $[ud]_{0^{-}}$ & 1.15 & 1.22$\ $ & 1.07 & \\
& $[us]_{0^{-}}$ & 1.28 & 1.34$\ $ & 1.05 & \\
& $[uc]_{0^{-}}$ & 2.35 & 2.39$\ $ & 0.62 & \\
& $[sc]_{0^{-}}$ & 2.48 & 2.51$\ $ & 0.56 & \\
& $[ub]_{0^{-}}$ & 5.61 & 5.64$\ $ & 0.36 & \\
& $[sb]_{0^{-}}$ & 5.72 & 5.74$\ $ & 0.33 & \\
& $[cb]_{0^{-}}$ & 6.62 & 6.63$\ $ & 0.28 & \\
\hline
$1^{-}$ & $[ud]_{1^{-}}$ & 1.33 & 1.37$\ $ & 0.51 & \\
& $[us]_{1^{-}}$ & 1.44 & 1.48$\ $ & 0.51 & \\
& $[uc]_{1^{-}}$ & 2.45 & 2.48$\ $ & 0.30 & \\
& $[sc]_{1^{-}}$ & 2.56 & 2.59$\ $ & 0.28 & \\
& $[ub]_{1^{-}}$ & 5.67 & 5.70$\ $ & 0.18 & \\
& $[sb]_{1^{-}}$ & 5.77 & 5.79$\ $ & 0.17 & \\
& $[cb]_{1^{-}}$ & 6.65 & 6.66$\ $ & 0.15 & \\
\hline\hline
\end{tabular*}
\end{center}
\end{table}
\begin{figure*}[t]
\includegraphics[width=0.95\textwidth]{F3.pdf}
\caption{\label{Fig:Diquark}
Computed masses of diquark correlations and their symmetry-related meson counterparts: diquarks -- (blue) stars and mesons -- (green) bars. (Pictorial representation of Table~\ref{diquarkspectrum}.)
}
\end{figure*}
Following this approach, using parameters determined as described here, in Sec.\,\ref{Sec:Meson} and Appendix~\ref{Sec:Contact}, one obtains the diquark masses and amplitudes listed in Table~\ref{diquarkspectrum}. Consistent with RL studies using realistic interactions \cite{Maris:2002yu, Maris:2004bp}, in the Bethe-Salpeter equation for a given $J^P$ diquark correlation we use the values of $\alpha_{\rm IR}$, $\Lambda_{\rm uv}$ associated with its $J^{-P}$ meson partner. A $\pm 10$\% change in ${\mathpzc s}_{\rm SO}$ alters the $0^-$, $1^-$ values by $\mp 2$\%.
Evidently, the antisymmetric Dirac-scalar combination of any two quark flavours (scalar diquark) is always lighter than the symmetric $J^P=1^+$ combination (pseudovector diquark, denoted $\{fg\}_{1^+}$), followed by the pseudoscalar and then the pseudovector diquarks. Just like mesons, the pattern of masses can be understood in terms of ESRs.
It is worth reiterating that diquark correlations are coloured and it is only in connection with partnering coloured objects (quark, another diquark, etc.) that a colour singlet system is obtained. Hence, diquarks are confined. That is not true in RL truncation \cite{Maris:2002yu}; but corrections to this leading-order analysis have been studied using an infrared-dominant interaction \cite{Munczek:1983dx}. In fully self-consistent symmetry-preser\-ving treatments, such corrections eliminate diquark bound-state poles from the quark+quark scattering matrix, whilst preserving the strong correlations \cite{Bhagwat:2004hn}. In these cases, one still has a physical interpretation of the RL masses; to wit, the quantity $\ell_{(fg)^{\!J^P}}:=1/m_{(fg)^{\!J^P}}$ is a spacetime distance that the diquark correlation can propagate before fragmentation.
Fig.\,\ref{Fig:Diquark} compares calculated diquark masses with those of their partner mesons. The level ordering of diquark correlations is precisely the same as that for mesons. Scalar and pseudovector diquarks are heavier than their partner mesons. Owing to the introduction of ${\mathpzc s}_{\rm SO}$ in Eq.\,\eqref{eqsSO}, this ordering is reversed for pseudoscalar and vector diquark correlations. Notwithstanding this, omitting the $[ud]_{0^+}$ and $[us]_{0^+}$ correlations, the mass of a diquark's partner meson is a reasonable guide to the diquark's mass: the mean difference in absolute value is $0.08(7)\,$GeV.
Owing to DCSB, the light-quark scalar diquark channels are atypical because their partner mesons are the almost-Nambu-Goldstone modes, $\pi$ and $K$. In two-colour chiral-limit QCD, scalar diquarks are also Nambu-Gold-stone modes \cite{Bloch:1999vk}. A hint of this remains in the three-colour theory, expressed in the relatively low masses of $[ud]_{0^+}$ and $[us]_{0^+}$, even though they are split widely from the $\pi$ and $K$.
It is the canonically normalised diquark Bethe-Sal-peter amplitudes that appear in baryon Faddeev equations. Listed in Table~\ref{diquarkspectrum}, they exhibit a clear ordering: in any given flavour sector, scalar diquarks have the largest amplitudes $\leftrightarrow$ strongest couplings. This means that the lighter flavour-$\overline{10}$ scalar diquark correlations are typically favoured in all $J=1/2$ baryon amplitudes because the Faddeev equations involve the diquark-amp-litude-squared. As will become apparent, under certain circumstances, \emph{e.g}.\ in baryons whose valence-quarks have widely different masses, this preference can be overcome.
Notably, the flavour-exchange symmetries of $J^P=3/2^\pm$ baryons mean that such CI states may only contain flavour-$15$ pseudovector diquark correlations, \emph{viz}.\ positive-parity diquarks. Using momentum-dependent interactions, flavour-$15$ vector diquarks are possible. \linebreak However, these correlations are weak; hence, even then, are unlikely to play a significant role. Nevertheless, this expectation should be checked.
\section{Baryon Spectrum}
\label{Sec:Baryons}
Following Ref.\,\cite{Yin:2019bxe}, we use the Faddeev equation drawn in Fig.\,\ref{figFaddeev} to calculate the spectrum of ground-state fla-vour-SU$(5)$ $J^P=1/2^\pm, 3/2^\pm$ baryons. In the isospin-symmetry limit, $88$ distinct states are supported. Details are presented in \ref{AppFaddeev}.
\begin{table*}[t]
\caption{\label{ResultsHalf}
Computed mass and Faddeev amplitude for each ground-state flavour-SU$(5)$ $J^P=1/2^+$ baryon.
The last column highlights the baryon's dominant spin-flavour correlation. The arrays of possibilities are given in Eqs.\,\eqref{halfnucleon}, \eqref{half2q1Q}, \eqref{HalfOneTwo}, \eqref{OneHalfQQQ}.
Empirical mass values, $M^{e}$, are taken from Ref.\,\cite{Zyla:2020zbs}; and lQCD results, $M^{l}$, from Refs.\,\cite{Durr:2008zz, Brown:2014ena, Mathur:2018epb}.
The results under heading $M^3$ are the three-body Faddeev equation predictions in Ref.\,\cite{Qin:2019hgk}. This analysis did not include SU$(3)$-flavour-symmetry breaking effects.
(Masses in GeV.)
}
\begin{center}
\begin{tabular*
{\hsize}
{l@{\extracolsep{0ptplus1fil}}|
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}|
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}|
c@{\extracolsep{0ptplus1fil}}}
\hline\hline
& $M^{\rm CI}$ & $M^{e}$ & $M^{l}$ & $M^{3}$ & $s^{r_{1}}$ & $s^{r_{2}}$ & $a_{1}^{r_{2}}$ & $a_{2}^{r_{2}}$ & $a_{1}^{r_{3}}$ & $a_{2}^{r_{3}}$ & $p^{r_{4}}$ & $p^{r_{5}}$ & $v_{1}^{r_{5}}$ & $v_{2}^{r_{5}}$ & $v_{1}^{r_{6}}$ & $v_{2}^{r_{6}}$ & $v_{1}^{r_{7}}$ & $v_{2}^{r_{7}}$ & dom.\ corr.\ \\
\hline
$N$ & 0.98 & 0.94 & 0.94 & 0.95\; & 0.88 & & -0.38 & -0.06 & 0.27 & 0.05 & -0.02 & & -0.02 & 0 & & & & & $u[ud]_{0^{+}}$ \\
$\Lambda$ & 1.10 & 1.12 & 1.11 & 1.11\; & 0.66 & 0.59 & & & -0.45 & -0.08 & -0.02 & -0.03 & & & -0.01 & 0 & -0.01 & 0\; & $s[ud]_{0^{+}}$ \\
$\Sigma$ & 1.20 & 1.19 & 1.17 & 1.11\; & 0.84 & & -0.45 & 0.13 & 0.26 & 0.01 & -0.01 & & -0.02 & 0 & & & & & $u[us]_{0^{+}}$ \\
$\Xi$ & 1.27 & 1.32 & 1.32 & 1.28\; & 0.89 & & -0.31 & 0.04 & 0.33 & 0.05 & -0.03 & & -0.02 & 0.01 & & & & & $s[us]_{0^{+}}$ \\
\hline
$\Lambda_{c}$ & 2.40 & 2.29 & 2.25 & 2.18\; & 0.18 & 0.86 & & & -0.35 & -0.33 & -0.02 & -0.05 & & & -0.02 & 0.03 & -0.02 & 0.01\; & $d[uc]_{0^{+}}-u[dc]_{0^{+}}$ \\
$\Sigma_{c}$ & 2.45 & 2.45 & 2.47 & 2.18\; & 0.46 & & -0.20 & 0.86 & 0.09 & 0.06 & 0.02 & & 0 & 0 & & & & & $c\{uu\}_{1^{+}}$ \\
$\Xi_{c}$ & 2.55 & 2.47 & 2.43 & 2.35\; & 0.18 & 0.84 & & & -0.36 & -0.35 & -0.02 & -0.06 & & & -0.02 & 0.03 & -0.02 & 0.01\; & $s[uc]_{0^{+}}-u[sc]_{0^{+}}$ \\
$\Xi_{c}^{\prime}$ & 2.59 & 2.58 & 2.57 & 2.35\; & 0.48 & & -0.21 & 0.85 & 0.09 & 0.06 & 0.02 & & 0 & 0 & & & & & $c\{us\}_{1^{+}}$ \\
$\Omega_{c}$ & 2.73 & 2.70 & 2.68 & 2.52\; & 0.50 & & -0.21 & 0.83 & 0.09 & 0.06 & 0.01 & & 0 & 0 & & & & & $c\{ss\}_{1^{+}}$ \\
\hline
$\Lambda_{b}$ & 5.62 & 5.62 & 5.63 & 5.39\; & 0.10 & 0.93 & & & -0.31 & -0.13 & 0 & 0.03 & & & -0.01 & 0.04 & -0.01 & -0.01\; & $d[ub]_{0^{+}}-u[db]_{0^{+}}$ \\
$\Sigma_{b}$ & 5.75 & 5.81 & 5.86 & 5.39\; & 0.33 & & -0.12 & 0.93 & 0.04 & 0.10 & 0.01 & & 0 & 0 & & & & & $b\{uu\}_{1^{+}}$ \\
$\Xi_{b}$ & 5.75 & 5.79 & 5.77 & 5.56\; & 0.11 & 0.94 & & & -0.31 & -0.13 & 0 & -0.03 & & & -0.01 & 0.03 & -0.01 & 0\; & $s[ub]_{0^{+}}-u[sb]_{0^{+}}$ \\
$\Xi_{b}^{\prime}$ & 5.88 & 5.94 & 5.93 & 5.56\; & 0.37 & & -0.11 & 0.92 & 0.05 & 0.12 & 0.01 & & 0 & 0 & & & & & $b\{us\}_{1^{+}}$ \\
$\Omega_{b}$ & 6.00 & 6.05 & 6.06 & 5.73\; & 0.42 & & -0.11 & 0.89 & 0.05 & 0.14 & 0.02 & & 0 & -0.01 & & & & & $b\{ss\}_{1^{+}}$ \\
\hline
$\Xi_{cc}$ & 3.64 & 3.62 & 3.61 & 3.42\; & 0.88 & & -0.34 & 0.30 & 0.10 & 0.06 & -0.05 & & -0.03 & 0.03 & & & & & $c[uc]_{0^{+}}$ \\
$\Xi_{cb}$ & 6.97 & & 6.94 & 6.63\; & 0.07 & -0.77 & & & 0.62 & -0.01 & 0.12 & 0.01 & & & 0.02 & -0.01 & 0.01 & 0\; & $u[cb]_{0^{+}}$ \\
$\Xi_{cb}^{\prime}$& 6.89 & & $6.96$ & 6.63\; & 0.89 & & -0.33 & -0.14 & 0.13 & 0.25 & -0.06 & & -0.03 & 0.02 & & & & & $b[uc]_{0^{+}}+c[ub]_{0^{+}}$ \\
$\Xi_{bb}$ & 10.22 & & 10.14 & 9.84\; & 0.85 & & -0.36 & 0.37 & 0.04 & 0.06 & -0.06 & & -0.05 & 0.04 & & & & & $b[ub]_{0^{+}}$ \\
\hline
$\Omega_{cc}$ & 3.79 & & 3.74 & 3.59\; & 0.86 & & -0.35 & 0.33 & 0.15 & 0.07 & -0.05 & & -0.03 & 0.03 & & & & & $c[sc]_{0^{+}}$ \\
$\Omega_{cb}$ & 7.08 & & 7.00 & 6.80\; & 0.05 & -0.84 & & & 0.50 & -0.19 & 0.08 & 0.01 & & & 0.01 & 0 & 0.01 & 0\; & $s[cb]_{0^{+}}$ \\
$\Omega_{cb}^{\prime}$ & 7.01 & & 7.03 & 6.80\; & 0.83 & & -0.32 & -0.21 & 0.18 & 0.36 & -0.06 & & -0.03 & 0.02 & & & & & $b[sc]_{0^{+}}+c[sb]_{0^{+}}$ \\
$\Omega_{bb}$ & 10.33 & & $\rule{-0.5em}{0ex}$10.27 & 10.01\; & 0.83 & & -0.38 & 0.39 & 0.07 & 0.09 & -0.07 & & -0.05 & 0.04 & & & & & $b[sb]_{0^{+}}$ \\
\hline
$\Omega_{ccb}\ $ & 8.04 & & 8.01 & 7.87\; & 0.40 & & -0.17 & 0.89 & 0.07 & 0.08 & 0.01 & & 0 & 0 & & & & & $b\{cc\}_{1^{+}}$ \\
$\Omega_{cbb}\ $ & 11.22 & & 11.20 & 11.08\; & 0.80 & & -0.39 & 0.40 & 0.16 & 0.14 & -0.05 & & -0.03 & 0.03 & & & & & $b[cb]_{0^{+}}$ \\
\hline\hline
\end{tabular*}
\end{center}
\end{table*}
\subsection{Completing the Faddeev kernels}
As highlighted above, the kernel in Fig.\,\ref{figFaddeev} introduces binding through diquark breakup and reformation via exchange of a dressed-quark. Again following Ref.\,\cite{Yin:2019bxe}, we exploit an oft used simplification, \emph{viz}.\ in the Faddeev equation for a baryon of type $B$, the quark exchanged between the diquarks, Eq.\,\eqref{AppendixFE}, is represented as
\begin{equation}
S_g^{\rm T}(k) \to \frac{g_B^2}{M_g}\,,
\label{staticexchange}
\end{equation}
where $g=l,s,c,b$ is the quark's flavour and $g_B$ is a coupling constant. This is a variation on the ``static approximation'' introduced in Ref.\,\cite{Buck:1992wz}. It makes the Faddeev amplitudes momentum-independent, just like the diquark Bethe-Salpeter amplitudes. Calculations reveal that it has little impact on the calculated masses \cite{Xu:2015kta}.
The couplings $g_{N}$, $g_{\Delta}$ are treated as parameters, with values chosen to obtain desired masses for the nucleon and $\Delta$-baryon, $m_N$, $m_\Delta$.
\begin{figure*}[t]
\vspace*{2ex}
\leftline{\hspace*{0.5em}{\large{\textsf{A}}}}
\vspace*{-3ex}
\includegraphics[width=0.99\textwidth]{F4A.pdf}
\vspace*{1ex}
\leftline{\hspace*{0.5em}{\large{\textsf{B}}}}
\vspace*{-3ex}
\includegraphics[width=0.99\textwidth]{F4B.pdf}
\caption{\label{figResultsHalf}
\emph{Upper panel}\,--\,{\sf A}.
CI computed masses (in GeV) of ground-state flavour-SU$(5)$ $J^P=1/2^+$ baryons in Table~\ref{ResultsHalf} compared with either experiment (first 15) \cite{Zyla:2020zbs} or lQCD (last 9) \cite{Brown:2014ena, Mathur:2018epb}.
\emph{Lower panel}\,--\,{\sf B}.
Similar comparison for ground-state flavour-SU$(5)$ $J^P=3/2^+$ baryons in Table~\ref{ResultsThreeHalf}: experiment (first 9) \cite{Zyla:2020zbs}; or lQCD (last 11) \cite{Brown:2014ena, Mathur:2018epb}.
Both panels: CI results -- (blue) circles; reference values -- (green) bars.
}
\end{figure*}
The Faddeev equation, Fig.\,\ref{figFaddeev}, generates what may be called a baryon's dressed-quark core \cite{Hecht:2002ej}: it omits what are typically described as meson-cloud contributions to the baryon masses, which work to reduce $m_N$, $m_\Delta$. The sizes of such corrections has been estimated: for the nucleon, the reduction is roughly $0.2\,$GeV and for the $\Delta$ it is $0.16\,$GeV. The choices \cite{Roberts:2011cf}
\begin{equation}
\label{gNgD}
g_N = 1.18\,,\; g_\Delta=1.56\,,
\end{equation}
produce $m_N=1.14\,$GeV and $m_\Delta=1.39\,$GeV, \emph{i.e}.\ inflated masses that leave room for correction by meson cloud effects. (\emph{N.B}.\ The values in Eq.\,\eqref{gNgD} are deliberately different from those used in Ref.\,\cite{Yin:2019bxe}.)
Using the framework outlined above, Ref.\,\cite{Lu:2017cln} studied flavour-SU$(3)$ baryons, finding that the ground-state positive-parity octet baryons are primarily constituted from like-parity diquarks, with negligible contributions from negative-parity correlations. Somewhat surprisingly, it found that their parity partners are also dominated by positive-parity diquark correlations; hence, too light. As with mesons and diquarks, the missing element was identified as too little spin-orbit repulsion generated by RL-like kernels. This was remedied by inserting an additional parameter into the Faddeev equation for $J^P=(1/2)^P$ baryons, \emph{i.e}.\ $g_{DB}^{\pi_\Psi \pi_d}$, a linear multiplicative factor, expressing EHM-induced enhancement of spin-orbit repulsion, attached to each diquark amplitude in the baryon's Faddeev equation kernel, Eq.\,\eqref{AppendixFE}:
\begin{equation}
\label{gDBval}
g_{DB}^{\pi_\Psi \pi_d} =
\left\{
\begin{array}{cl}
1.0 & \pi_\Psi = \pi_d \\
0.1 & \pi_\Psi = - \pi_d
\end{array}\right.\,,
\end{equation}
\emph{i.e}.\ nothing is done when the parity of the diquark correlation, $\pi_d $, matches that of the host baryon, $\pi_\Psi$, but suppression is introduced when the parity is different. The magnitude of the effect was chosen so that, in concert with Eqs.\,\eqref{eqsSO}, the dressed-quark core mass of the nucleon's parity
partner is not lower than that of its first radial excitation \cite{Lu:2017cln}.
Every element in the Faddeev kernel is now specified. It remains only to select the channel, derive the explicit form for the associated algebraic equation, then solve the resulting matrix equation to obtain the baryon mass and amplitude.
Specific examples of the Faddeev equations for the various light-quark systems, along with their derivations, are presented in Ref.\,\cite{Chen:2012qr}; and equations for $\Xi_c^+$, $\Xi_c^{\prime +}$, $\Xi_c^{\ast +}$ are given in Ref.\,\cite{Yin:2019bxe}. The equations solved herein are not any more complex, although they do involve larger kernel matrices because we also allow all baryons to contain pseudoscalar and vector diquarks.
As in Ref.\,\cite{Yin:2019bxe}, the ultraviolet cutoff in each baryon channel is identified with that of the lightest diquark in the system. This is always the smallest value; hence, the dominant regularising influence.
\begin{table*}[t]
\caption{\label{ResultsThreeHalf}
Computed mass and Faddeev amplitude for each ground-state flavour-SU$(5)$ $J^P=3/2^+$ baryon.
The last column highlights the baryon's dominant spin-flavour correlation. The possibilities are given in Eqs.\,\eqref{ThreeHalfDelta}, \eqref{ThreeHalfqQQ}, \eqref{ThreeHalfOneTwo}, \eqref{ThreeHalfQQQ}.
Empirical mass values, $M^{e}$, are taken from Ref.\,\cite{Zyla:2020zbs}; and lQCD results, $M^{l}$, from Refs.\,\cite{Durr:2008zz, Brown:2014ena, Mathur:2018epb}.
The results in column $M^3$ are the three-body Faddeev equation predictions in Ref.\,\cite{Qin:2019hgk}.
(Masses in GeV.)
}
\begin{center}
\begin{tabular*
{\hsize}
{l@{\extracolsep{0ptplus1fil}}|
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}|
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}|
c@{\extracolsep{0ptplus1fil}}}
\hline\hline
& $M^{\rm CI}$ & $M^{e}$ & $M^{l}$ & $M^{3}$\; & $a^{r_{1}}$ & $a^{r_{2}}$ & dom.\ corr.\ \\
\hline
$\Delta$ & 1.27 & 1.23 & 1.23 & 1.21\; & 1 & & $u\{uu\}_{1^{+}}$ \\
$\Sigma^{*}$ & 1.39 & 1.38 & 1.40 & 1.36\; & 0.61 & 0.79\; & $u\{us\}_{1^{+}}$ \\
$\Xi^{*}$ & 1.51 & 1.53 & 1.56 & 1.52\; & 0.85 & 0.52\; & $s\{us\}_{1^{+}}$ \\
$\Omega$ & 1.63 & 1.67 & 1.67 & 1.67\; & 1 & & $s\{ss\}_{1^{+}}$ \\
\hline
$\Sigma_{c}^{*}$ & 2.54 & 2.52 & 2.55 & 2.39\; & 0.63 & 0.78\; & $u\{uc\}_{1^{+}}$ \\
$\Xi_{c}^{*}$ & 2.67 & 2.65 & 2.65 & 2.55\; & 0.61 & 0.79\; & $s\{uc\}_{1^{+}}+u\{sc\}_{1^{+}}$ \\
$\Omega_{c}^{*}$ & 2.80 & 2.77 & 2.76 & 2.70\; & 0.59 & 0.81\; & $s\{sc\}_{1^{+}}$ \\
\hline
$\Sigma_{b}^{*}$ & 5.79 & 5.83 & 5.88 & 5.60\; & 0.66 & 0.75\; & $u\{ub\}_{1^{+}}$ \\
$\Xi_{b}^{*}$ & 5.92 & 5.95 & 5.96 & 5.75\; & 0.61 & 0.79\; & $s\{ub\}_{1^{+}}+u\{sb\}_{1^{+}}$ \\
$\Omega_{b}^{*}$ & 6.04 & & 6.09 & 5.90\; & 0.55 & 0.84\; & $s\{sb\}_{1^{+}}$ \\
\hline
$\Xi_{cc}^{*}$ & 3.73 & & 3.69 & 3.58\; & 0.98 & 0.20\; & $c\{uc\}_{1^{+}}$ \\
$\Xi_{cb}^{*}$ & 7.00 & & 6.99 & 6.78\; & 0.97 & 0.24\; & $b\{uc\}_{1^{+}}+c\{ub\}_{1^{+}}$ \\
$\Xi_{bb}^{*}$ & 10.26 & & 10.18 & 9.98\; & 0.99 & 0.08\; & $b\{ub\}_{1^{+}}$ \\
\hline
$\Omega_{cc}^{*}$ & 3.88 & & 3.82 & 3.73\; & 0.96 & 0.28\; & $c\{sc\}_{1^{+}}$ \\
$\Omega_{cb}^{*}$ & 7.12 & & 7.06 & 6.93\; & 0.94 & 0.35\; & $b\{sc\}_{1^{+}}+c\{sb\}_{1^{+}}$ \\
$\Omega_{bb}^{*}$ & 10.37 & & 10.31 & 10.14\; & 0.99 & 0.12\; & $b\{sb\}_{1^{+}}$ \\
\hline
$\Omega_{ccc}$ & 4.90 & & 4.80 & 4.76\; & 1 & & $c\{cc\}_{1^{+}}$ \\
$\Omega_{ccb}^{*}$ & 8.08 & & 8.04 & 7.96\; & 0.62 & 0.79\; & $c\{cb\}_{1^{+}}$ \\
$\Omega_{cbb}^{*}$ & 11.26 & & 11.23 & 11.17\; & 0.96 & 0.28\; & $b\{cb\}_{1^{+}}$ \\
$\Omega_{bbb}$ & 14.45 & & 14.37 & 14.37\; & 1 & & $b\{bb\}_{1^{+}}$ \\
\hline\hline
\end{tabular*}
\end{center}
\end{table*}
\subsubsection{$J^P=1/2^+$}
\label{subsubHalf}
Our results for the masses and amplitudes of flavour-SU$(5)$ $J^P = 1/2^+$ baryons are listed in Table~\ref{ResultsHalf}.
As remarked in association with Eq.\,\eqref{gNgD}, the Faddeev kernel employed omits resonant contributions that serve to reduce baryon masses. From each of the values in column~1 we have therefore subtracted
\begin{equation}
{\mathpzc z}_{N^+} = 0.16\,{\rm GeV}
\label{Nrescale}
\end{equation}
in order to express an empirically informed estimate of such corrections. This value is the average difference between the unmodified CI result and experiment in the first four rows. It matches the expectations described in connection with Eq.\,\eqref{gNgD}. As indicated after Eq.\,\eqref{gNgD}, Ref.\,\cite{Yin:2019bxe} achieved a similar effect by inflating the values of $g_{N,\Delta}$, but we find that can produce unexpected (perhaps unphysical) modifications of the Faddeev amplitudes.
The computed masses in Table~\ref{ResultsHalf} are also compared with empirical or lQCD values in Fig.\,\ref{figResultsHalf}A: the mean absolute-relative-difference is $1.3(1.3)$\%.
Compared with the results in column~4, drawn from Ref.\,\cite{Qin:2019hgk}, the analogous difference is $5.2(2.8)$\%. In that study, the Faddeev equations were solved in a fully-consistent RL truncation, eschewing a quark+diquark approximation. Moreover, the results are \emph{ab initio} predictions, whereas we introduced $g_N$, ${\mathpzc z}_{N^+}$ -- Eqs.\,\eqref{staticexchange}, \eqref{Nrescale} -- to adjust the overall scale of the $J^P=1/2^+$ spectrum.
Notwithstanding these things, the level of agreement confirms the validity of the ESRs used to complete the spectrum calculations in Ref.\,\cite{Qin:2019hgk}; and the fact that our results exceed the accuracy of Ref.\,\cite{Qin:2019hgk} indicates that we have implemented a phenomenologically efficacious CI formulation.
The $1/2^+$ Faddeev amplitudes are represented in Table~\ref{ResultsHalf} by the strength of the coefficient that multiplies the flavour and Dirac structure specified by Eqs.\,\eqref{SAPD}, \eqref{halfnucleon}, \eqref{half2q1Q}, \eqref{HalfOneTwo}, \eqref{OneHalfQQQ}. For instance, the $\Lambda$ entry in column~6 is ``$-0.45$'', indicating that this baryon contains a pseudovector diquark component $[d\{us\}_{1^{+}}-u\{ds\}_{1^{+}}]/\sqrt{2}$ with $i\gamma_5\gamma_\mu$ coefficient $a_1^+=-0.45$ in \linebreak Eq.\,\eqref{Aa1}.
The amplitudes in Table~\ref{ResultsHalf} confirm the conclusions in Ref.\,\cite{Yin:2019bxe}. Namely, concerning ground-state flavour-$SU(5)$ $J^P=1/2^+$ baryons:
($\mathpzc a$) the lightest participating diquark correlation usually defines the most important component of a baryon's Faddeev amplitude and this is true even if a pseudovector diquark is the lightest channel available; and
($\mathpzc b$) light-diquark dominance may be overcome in flavour channels for which the bound-state's spin-flavour structure and the quark-exchange character of the kernel in Fig.\,\ref{figFaddeev} lead dynamically to a preference for mixed-flavour correlations.
\begin{table*}[t]
\caption{\label{ResultsHalfMinus}
Computed mass (in GeV) and Faddeev amplitude for each ground-state flavour-$SU(5)$ $J^P=1/2^-$ baryon: the last column identifies the baryon's dominant spin-flavour correlation. The arrays of possibilities are given in Eqs.\,\eqref{halfnucleon}, \eqref{half2q1Q}, \eqref{HalfOneTwo}, \eqref{OneHalfQQQ}.
Empirical mass values, $M^{e}$, are taken from Ref.\,\cite{Zyla:2020zbs}; and lQCD results, $M^{l}$, from Ref.\,\cite{Bahtiyar:2020uuj}.
The results under heading $M^3$ are the three-body Faddeev equation predictions in Ref.\,\cite{Qin:2019hgk}, which did not include SU$(3)$-flavour-symmetry breaking effects.
}
\begin{center}
\begin{tabular*
{\hsize}
{l@{\extracolsep{0ptplus1fil}}|
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}|
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}|
c@{\extracolsep{0ptplus1fil}}}
\hline\hline
& $M^{\rm CI}$ & $M^{e}$ & $M^{l}$ & $M^{3}$ & $s^{r_{1}}$ & $s^{r_{2}}$ & $a_{1}^{r_{2}}$ & $a_{2}^{r_{2}}$ & $a_{1}^{r_{3}}$ & $a_{2}^{r_{3}}$ & $p^{r_{4}}$ & $p^{r_{5}}$ & $v_{1}^{r_{5}}$ & $v_{2}^{r_{5}}$ & $v_{1}^{r_{6}}$ & $v_{2}^{r_{6}}$ & $v_{1}^{r_{7}}$ & $v_{2}^{r_{7}}$ & dom.\ corr.\ \\
\hline
$N$ & 1.55 & 1.54 & & 1.54\; & 0.35 & & 0.04 & 0 & -0.03 & 0 & -0.92 & & -0.06 & 0.18 & & & & & $u[ud]_{0^{-}}$ \\
$\Lambda$ & 1.65 & 1.67 & & 1.58\; & 0.27 & 0.24 & & & 0.03 & 0 & -0.66 & -0.62 & & & -0.06 & 0.14 & -0.05 & 0.13\; & $s[ud]_{0^{-}}$ \\
$\Sigma$ & 1.70 & 1.68 & & 1.58\; & 0.46 & & 0.17 & -0.13 & -0.08 & 0.06 & -0.85 & & 0.06 & 0.13 & & & & & $u[us]_{0^{-}}$ \\
$\Xi$ & 1.78 & & & 1.62\; & 0.39 & & 0.04 & -0.01 & -0.04 & 0.02 & -0.90 & & -0.05 & 0.19 & & & & & $s[us]_{0^{-}}$ \\
\hline
$\Lambda_{c}$ & 2.67 & 2.59 & 2.67 & 2.67\; & 0.07 & 0.45 & & & 0.06 & -0.03 & -0.21 & -0.83 & & & -0.09 & 0.14 & 0.03 & 0.20\; & $d[uc]_{0^{-}}-u[dc]_{0^{-}}$ \\
$\Sigma_{c}$ & 2.84 & & 2.81 & 2.67\; & 0.13 & & 0.14 & -0.19 & 0.08 & -0.24 & -0.17 & & -0.27 & 0.88 & & & & & $u[uc]_{1^{-}}$ \\
$\Xi_{c}$ & 2.79 & 2.79 & 2.77 & 2.71\; & 0.07 & 0.51 & & & 0.09 & -0.05 & -0.20 & -0.78 & & & -0.06 & 0.07 & 0.09 & 0.25\; & $s[uc]_{0^{-}}-u[sc]_{0^{-}}$ \\
$\Xi_{c}^{\prime}$ & 2.94 & & 2.93 & 2.71\; & 0.15 & & 0.13 & -0.19 & 0.08 & -0.24 & -0.16 & & -0.27 & 0.87 & & & & & $s[uc]_{1^{-}}+u[sc]_{1^{-}}$ \\
$\Omega_{c}$ & 3.03 & & 3.04 & 2.74\; & 0.18 & & 0.13 & -0.19 & 0.08 & -0.25 & -0.13 & & -0.27 & 0.87 & & & & & $s[sc]_{1^{-}}$ \\
\hline
$\Lambda_{b}$ & 6.03 & 5.91 & & 5.92\; & 0 & -0.22 & & & 0.12 & -0.03 & 0 & 0.15 & & & 0.14 & -0.72 & 0.44 & 0.43\; & $b[ud]_{1^{-}}$ \\
$\Sigma_{b}$ & 6.32 & & & 5.92\; & 0.01 & & 0 & 0 & -0.05 & 0.25 & -0.01 & & 0.24 & -0.94 & & & & & $u[ub]_{1^{-}}$ \\
$\Xi_{b}$ & 6.15 & & & 5.96\; & 0 & -0.21 & & & 0.13 & -0.04 & 0 & 0.14 & & & 0.13 & -0.67 & 0.47 & 0.48\; & $b[us]_{1^{-}}$ \\
$\Xi_{b}^{\prime}$ & 6.40 & & & 5.96\; & 0.01 & & 0 & 0 & -0.05 & 0.26 & -0.03 & & 0.22 & -0.94 & & & & & $s[ub]_{1^{-}}+u[sb]_{1^{-}}$ \\
$\Omega_{b}$ & 6.49 & & & 5.99\; & 0.02 & & 0 & 0 & -0.05 & 0.27 & -0.06 & & 0.21 & -0.94 & & & & & $s[sb]_{1^{-}}$ \\
\hline
$\Xi_{cc}$ & 3.81 & & 3.93 & 3.79\; & 0.54 & & 0.09 & -0.07 & -0.02 & 0.02 & -0.83 & & 0.03 & 0.10 & & & & & $c[uc]_{0^{-}}$ \\
$\Xi_{cb}$ & 7.04 & & & 7.04\; & 0.01 & -0.19 & & & -0.29 & 0.21 & 0.51 & -0.29 & & & -0.24 & -0.64 & -0.19 & -0.01\; & $b[uc]_{1^{-}}-c[ub]_{1^{-}}$ \\
$\Xi_{cb}^{\prime}$& 6.98 & & & 7.04\; & 0.48 & & 0.11 & -0.03 & -0.05 & 0.02 & -0.75 & & 0.16 & 0.40 & & & & & $b[uc]_{0^{-}}+c[ub]_{0^{-}}$ \\
$\Xi_{bb}$ & 10.22 & & & 10.29\; & 0.62 & & 0.15 & -0.14 & -0.01 & 0.01 & -0.74 & & 0.15 & -0.04 & & & & & $b[ub]_{0^{-}}$ \\
\hline
$\Omega_{cc}$ & 3.93 & & 4.04 & 3.83\; & 0.66 & & 0.16 & -0.15 & -0.05 & 0.04 & -0.70 & & 0.17 & -0.02 & & & & & $c[sc]_{0^{-}}$ \\
$\Omega_{cb}$ & 7.15 & & & 7.08\; & 0.01 & -0.28 & & & -0.23 & 0.17 & 0.68 & -0.31 & & & -0.11 & -0.48 & -0.20 & -0.02\; & $b[sc]_{0^{-}}-c[sb]_{0^{-}}$ \\
$\Omega_{cb}^{\prime}$ & 7.16 & & & 7.08\; & 0.53 & & 0.14 & -0.05 & -0.08 & 0.04 & -0.69 & & 0.21 & 0.40 & & & & & $b[sc]_{0^{-}}+c[sb]_{0^{-}}$ \\
$\Omega_{bb}$ & 10.42 & & & 10.33\; & 0 & & 0 & 0 & -0.01 & 0.01 & 0.01 & & 0.72 & -0.69 & & & & & $b[sb]_{1^{-}}$ \\
\hline
$\Omega_{ccb}$ & 8.10 & & & 8.16\; & 0.24 & & 0.12 & -0.15 & 0.08 & -0.34 & -0.02 & & -0.21 & 0.86 & & & & & $c[cb]_{1^{-}}$ \\
$\Omega_{cbb}$ & 11.23 & & & 11.41\; & 0.01 & & -0.09 & 0.33 & -0.07 & 0.14 & 0.54 & & 0.56 & -0.52 & & & & & $b[cb]_{1^{-}}$ \\
\hline\hline
\end{tabular*}
\end{center}
\end{table*}
In corroborating these conclusions, we reaffirm the arguments against treatments of the baryon problem which assume they can be described as effectively two-body in nature, \emph{e.g}.\ as being built from a constituent-quark and static constituent-diquark. The dynamical character of diquark correlations is paramount because their breakup and reformation are crucial in defining baryon structure. This is confirmed for light-quark baryons by the fact that lQCD calculations produce a spectrum whose richness cannot be explained by a two-body model \cite{Edwards:2011jj}. The implications extend to baryons involving one or more heavy quarks, challenging both (\emph{i}) the treatment of singly-heavy baryons as two-body light-diquark+heavy-quark bound-states and (\emph{ii}) analyses of doubly-heavy baryons which assume such systems can be considered as two-body light-quark+heavy-diquark bound-states. These observations may also be relevant to few-body studies of tetra- and penta-quark problems; in particular, those involving $\{u,d,s\}$ quarks.
\subsubsection{$J^P=3/2^+$}
\label{subsubThreeHalf}
Our results for the masses and amplitudes of flavour-SU$(5)$ $J^P = 3/2^+$ baryons are listed in Table~\ref{ResultsThreeHalf}.
Akin to Eq.\,\eqref{Nrescale}, from each of the values in column~1 we have subtracted
\begin{equation}
{\mathpzc z}_{\Delta^+} = 0.12\,{\rm GeV}
\label{Drescale}
\end{equation}
so as to express an empirically informed estimate of meson-cloud corrections: this value is the average difference between the unmodified CI result and experiment in the first four rows. Again, it matches expectations described in connection with Eq.\,\eqref{gNgD}.
The computed $J^P=3/2^+$ masses are compared with empirical/lQCD values in Fig.\,\ref{figResultsHalf}B: the mean-abso-lute-relative-difference is $1.0(0.8)$\%.
Furthermore, the results compare well with the three-body calculation described in Ref.\,\cite{Qin:2019hgk}, for which the analogous difference is $2.6(1.6)$\%.
Notwithstanding the fact that we used $g_\Delta$, ${\mathpzc z}_{\Delta^+}$ -- Eqs.\,\eqref{staticexchange}, \eqref{Drescale} -- to adjust the overall scale of the $J^P=3/2^+$ spectrum, this improvement over the results in Ref.\,\cite{Qin:2019hgk} again highlights the utility of our CI formulation.
The CI Faddeev amplitudes of $J=3/2^+$ baryons are also listed in Table~\ref{ResultsThreeHalf}. Once more, the fully dynamical nature of the diquarks and the character of the Faddeev kernel work together to ensure a continual reshuffling of the dressed-quarks within the diquark correlations. Consequently, in all cases involving more than one quark flavour, the diquark combination with maximal flavour shuffling is favoured because it is fed by twice as many exchange processes as the less-mixed correlation.
This differs from the result for the $\Sigma_b^\ast$-baryon in Ref.\,\cite{Yin:2019bxe}. Therein, however, $g_\Delta$ was 26\% larger than here, thereby providing additional enhancement for the single $u$-quark exchange process over the doubly active, but $b-u$ mass-splitting-suppressed, $b$-quark exchange contribution which enables the shuffling.
\begin{figure*}[t]
\vspace*{2ex}
\leftline{\hspace*{0.5em}{\large{\textsf{A}}}}
\vspace*{-3ex}
\includegraphics[width=0.99\textwidth]{F5A.pdf}
\vspace*{1ex}
\leftline{\hspace*{0.5em}{\large{\textsf{B}}}}
\vspace*{-3ex}
\includegraphics[width=0.99\textwidth]{F5B.pdf}
\caption{\label{figResultsHalfNegative}
\emph{Upper panel}\,--\,{\sf A}.
CI computed masses (in GeV) of ground-state flavour-SU$(5)$ $J^P=1/2^-$ baryons in Table~\ref{ResultsHalfMinus} compared with experiment \cite{Zyla:2020zbs} (green bars), lQCD \cite{Bahtiyar:2020uuj} (gold triangles), or three-body Faddeev equation results \cite{Qin:2019hgk} (orange asterisks).
\emph{Lower panel}\,--\,{\sf B}.
Similar comparison for ground-state flavour-SU$(5)$ $J^P=3/2^-$ baryons, compiled using Table~\ref{ResultsThreeHalfMinus}.
Both panels: the states are labelled as parity partners of the lightest same-$J$ state, e.g. $p^-$ denotes the $N(1535)\,(1/2)^-$ baryon and $\Delta^-$ indicates $\Delta(1700)\,(3/2)^-$.
}
\end{figure*}
\subsubsection{$J^P=1/2^-$}
\label{subsubHalf}
CI results for the masses and amplitudes of flavour-SU$(5)$ $J^P = 1/2^-$ baryons are listed in Table~\ref{ResultsHalfMinus}.
From each of the values in column~1, we subtracted
\begin{equation}
{\mathpzc z}_{N^-} = 0.27\,{\rm GeV}
\label{NrescaleM}
\end{equation}
in order to express an empirically informed estimate of meson cloud corrections: this value is the average difference between the unmodified CI result and experiment in the first three rows.
The computed masses in Table~\ref{ResultsHalfMinus} are also depicted in Fig.\,\ref{figResultsHalfNegative}A. Compared with empirical values, or lQCD results when empirical values are unavailable, the mean absolute-relative-difference is $1.4(1.1)$\%; and compared with the results in column~4, drawn from Ref.\,\cite{Qin:2019hgk}, the analogous difference is $3.7(3.5)$\%.
The CI Faddeev amplitudes of $J^P=1/2^-$ baryons are also listed in Table~\ref{ResultsHalfMinus}. They are readily understood because they follow the pattern of the $J=1/2^+$ amplitude in almost all cases: typically, the lightest like-parity $J=1$ diquark correlation dominates. The exception is $\Xi_{cb}^-$/$\Omega_{cb}^-$, wherein the lightest like-parity diquark correlations dominate, which is a $0^-$ combination in this case.
\begin{table*}[t]
\caption{\label{ResultsThreeHalfMinus}
Computed mass (in GeV) and Faddeev amplitude for each ground-state flavour-$SU(5)$ $J^P=3/2^-$ baryon: the last column identifies the baryon's dominant spin-flavour correlation. The possibilities are given in Eqs.\,\eqref{ThreeHalfDelta}, \eqref{ThreeHalfqQQ}, \eqref{ThreeHalfOneTwo}, \eqref{ThreeHalfQQQ}.
Empirical mass values, $M^{e}$, are taken from Ref.\,\cite{Zyla:2020zbs}; and lQCD results, $M^{l}$, from Ref.\,\cite{Bahtiyar:2020uuj}.
The results under heading $M^3$ are the three-body Faddeev equation predictions in Ref.\,\cite{Qin:2019hgk}.
}
\begin{center}
\begin{tabular*
{\hsize}
{l@{\extracolsep{0ptplus1fil}}|
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}|
r@{\extracolsep{0ptplus1fil}}
r@{\extracolsep{0ptplus1fil}}|
c@{\extracolsep{0ptplus1fil}}}
\hline\hline
& $M^{\rm CI}$ & $M^{e}$ & $M^{l} $ & $M^{3}$\; & $a^{r_{1}}$ & $a^{r_{2}}$\; & dom.\ corr.\ \\
\hline
$\Delta$ & 1.59 & 1.67 & & 1.73\; & 1 & & $u\{uu\}_{1^{+}}$ \\
$\Sigma^{*}$ & 1.72 & 1.66 & & 1.79\; & 0.72 & 0.69\; & $s\{uu\}_{1^{+}}$ \\
$\Xi^{*}$ & 1.84 & 1.82 & & 1.84\; & 0.91 & 0.42\; & $s\{us\}_{1^{+}}$ \\
$\Omega$ & 1.95 & & & 1.90\; & 1 & & $s\{ss\}_{1^{+}}$ \\
\hline
$\Sigma_{c}^{*}$ & 2.80 & & 2.80 & 2.83\; & 0.85 & 0.53\; & $c\{uu\}_{1^{+}}$ \\
$\Xi_{c}^{*}$ & 2.91 & 2.82 & 2.80 & 2.89\; & 0.74 & 0.67\; &
$c\{us\}_{1^{+}}$ \\
$\Omega_{c}^{*}$ & 3.01 & & 3.07 & 2.94\; & 0.62 & 0.79\; & $s\{sc\}_{1^{+}}$ \\
\hline
$\Sigma_{b}^{*}$ & 6.03 & & & 6.07\; & 0.84 & 0.54\; & $b\{uu\}_{1^{+}}$ \\
$\Xi_{b}^{*}$ & 6.13 & & & 6.13\; & 0.69 & 0.72\; & $s\{ub\}_{1^{+}}+u\{sb\}_{1^{+}}$ \\
$\Omega_{b}^{*}$ & 6.23 & & & 6.19\; & 0.52 & 0.86\; & $s\{sb\}_{1^{+}}$ \\
\hline
$\Xi_{cc}^{*}$ & 3.93 & & 4.01 & 3.93\; & 0.98 & 0.18\; & $c\{uc\}_{1^{+}}$ \\
$\Xi_{cb}^{*}$ & 7.13 & & & 7.18\; & 0.96 & 0.28\; & $b\{uc\}_{1^{+}}+c\{ub\}_{1^{+}}$ \\
$\Xi_{bb}^{*}$ & 10.35 & & & 10.42\; & 0.99 & 0.08\; & $b\{ub\}_{1^{+}}$ \\
\hline
$\Omega_{cc}^{*}$ & 4.04 & & 4.12 & 3.99\; & 0.96 & 0.26\; & $c\{sc\}_{1^{+}}$ \\
$\Omega_{cb}^{*}$ & 7.23 & & & 7.23\; & 0.92 & 0.39\; & $b\{sc\}_{1^{+}}+c\{sb\}_{1^{+}}$ \\
$\Omega_{bb}^{*}$ & 10.44 & & & 10.48\; & 0.99 & 0.11\; & $b\{sb\}_{1^{+}}$ \\
\hline
$\Omega_{ccc}$ & 5.01 & & 5.08 & 5.03\; & 1 & & $c\{cc\}_{1^{+}}$ \\
$\Omega_{ccb}^{*}$ & 8.17 & & & 8.28\; & 0.63 & 0.77\; & $c\{cb\}_{1^{+}}$ \\
$\Omega_{cbb}^{*}$ & 11.32 & & & 11.52\; & 0.96 & 0.28\; & $b\{cb\}_{1^{+}}$ \\
$\Omega_{bbb}$ & 14.49 & & & 14.77\; & 1 & & $b\{bb\}_{1^{+}}$ \\
\hline\hline
\end{tabular*}
\end{center}
\end{table*}
\subsubsection{$J^P=3/2^-$}
\label{subsubThreeHalf}
These states are somewhat unusual. Given that they must be completely symmetric under permutations of flavour labels but the only negative-parity diquarks supported by the CI are flavour-$\overline{10}$ (antisymmetric) states, then the $J=3/2^-$ baryons can only contain positive parity pseudovector diquarks, Eq.\,\eqref{ThreeHalfQQQ}. Owing to this peculiarity, we judged that opposite-parity diquark correlations should not receive additional suppression in these systems;
so we solved the associated Faddeev equations with
\begin{equation}
g_{DB}^{\pi_{3/2^-} \pi_{1^+}}=g_{DB}^{\pi_{3/2^+} \pi_{1^+}}=1\,.
\end{equation}
Working with the results thus obtained, the average difference between the CI result and experiment in the first three rows of Table~\ref{ResultsThreeHalfMinus} is
\begin{equation}
z_{\Delta^-} = 0.13\,{\rm GeV}.
\end{equation}
Consequently, in column~1 of Table~\ref{ResultsThreeHalfMinus} we list the result obtained by subtracting $z_{\Delta^-}$ from the directly calculated CI result. Comparing column~1 with experiment, where known, and available lQCD results otherwise, the mean absolute-relative-difference is $2.2(1.4)$\%, and with the results in column~4, drawn from Ref.\,\cite{Qin:2019hgk}, the analogous difference is $1.4(1.9)$\%.
The computed masses in Table~\ref{ResultsThreeHalfMinus} are also depicted in Fig.\,\ref{figResultsHalfNegative}B alongside, as available, in this order: experiment, lQCD, or three-body Faddeev equation results.
The CI Faddeev amplitudes of $J=3/2^-$ baryons are also listed in Table~\ref{ResultsThreeHalfMinus}. They are similar to those of the $J=3/2^+$ states in Table~\ref{ResultsThreeHalf}, with the dominant diquark correlation being the same in all cases except $\Sigma^\ast$, $\Sigma^\ast_c$, $\Sigma^\ast_b$. Considering Eq.\,\eqref{SigmaQ}, one sees that in these cases there is a competition between the lightest diquark correlation and the heavy-light diquark, which is doubly-fed by the exchange kernel. The former wins in the negative-parity systems, whereas the latter is dominant in the positive-parity systems. This ordering reverses in the $3/2^-$ states as $g_{DB}^{\pi_{3/2^-} \pi_{1^+}}$ is reduced from unity. The $\Sigma^\ast$, $\Sigma^\ast_c$, $\Sigma^\ast_b$ Faddeev amplitudes should therefore be considered as somewhat uncertain.
\section{Summary and Outlook}
\label{Sec:Epilogue}
A confining, symmetry-preserving treatment of a vec-tor$\times$vector contact interaction (CI) was used to compute spectra of ground-state $J^P = 0^\pm, 1^\pm$ $(f\bar g)$ mesons and $J^P=1/2^\pm, 3/2^\pm$ $(fgh)$ baryons, where $f,g,h \in \{u,d,s,c,b\}$. The calculated meson masses agree well with experiment [Sec.\,\ref{Sec:Meson}]: the mean-relative-difference for 33 states is 5(6)\%. Expressing effects tied to the emergence of hadronic mass (EHM) was crucial to achieving this level of agreement. Regarding meson leptonic decay constants, empirical (or lattice QCD -- lQCD) values are reproduced with an accuracy of 15(11)\%; and predictions were made for the currently unknown decay constants of 19 $0^+$, $1^+$ states [Table~\ref{Tab:MesonSpectrum}].
A quark+diquark approximation to the baryon Faddeev equation was used herein. Its formulation required the calculation of masses and correlation strengths for all 38 distinct participating diquarks [Sec.\,\ref{SecDiquarks}]. As in all studies to date, the level ordering of the $J^{P}$ diquarks matches that of their $J^{-P}$ meson partners. Scalar and pseudovector diquarks are heavier than their partner mesons; but in our CI formulation, this ordering is reversed for pseudoscalar and vector diquark correlations. Nevertheless, omitting the diquarks partnered with\linebreak would-be Nambu-Goldstone mode mesons, the mass of a diquark's partner meson is a reasonable guide to the diquark's mass: the mean difference in absolute value is $0.08(7)\,$GeV.
A static approximation to the quark exchange kernel was used to solve the Faddeev equation. It produces momentum-independent Faddeev amplitudes, thereby ensuring a level of consistency with CI two-body bound-state amplitudes, and introduces four parameters. Correcting also for the omission of meson cloud contributions to the quark-core Faddeev kernels, four additional parameters were necessary to complete the CI definition in the baryon sector. In total, the CI predicts 88 distinct baryons, \emph{i.e}.\ every possible three-quark $1/2^\pm, 3/2^\pm$ ground-state baryon is realised. Of this number, 34 states are already known empirically and lQCD-computed masses are available for another 30. For this collection of 64 states, the mean absolute-relative-diffe-rence between CI prediction and experiment/lQCD mass is $\overline{\rm ard} = 1.4(1.2)$\% [Figs.\,\ref{figResultsHalf}, \ref{figResultsHalfNegative}]. Implementation of EHM-induced effects associated with spin-orbit repulsion in $1/2^-$ baryons was important to achieving this outcome. The same 88 ground-states are also predicted by a three-body Faddeev equation \cite{Qin:2019hgk}; and in comparison with those results, the analogous difference is 3.4(3.0)\%.
A primary merit of the framework employed herein is its simplicity, enabling all analyses and calculations to be completed algebraically. In total, there are twelve parameters: four used to define the interaction and its scale dependence via $\pi$, $K$, $\eta_c$, $\eta_b$ properties; and eight introduced to complete the baryon Faddeev equations.
From this foundation, the CI delivers predictions for 164 distinct quantities. Thus far, 114 of these observables have either been measured or computed using lQCD; and a comparison on this subset yields $\overline{\rm ard} = 4.5(7.1)$\%.
This level of quantitative success suggests that some credibility be given to the qualitative conclusions that follow from our CI analysis.
(\emph{I}) Nonpointlike, dynamical diquark correlations play an important role in all baryons; and, typically, the lightest allowed diquark is the most important component of a baryon's Faddeev amplitude.
(\emph{II}) Positive-parity diquarks dominate in positive-parity baryons, with $J=1^+$ diquarks being prominent in all of them.
(\emph{III}) Negative-parity diquarks can be neglected when studying positive-parity baryons; but owing to EHM, they are significant, even dominant, in $J=1/2^-$ baryons.
(\emph{IV}) On the other hand, $J=3/2^-$ baryons are built (almost) exclusively from $J=1^+$ diquark correlations.
Naturally, these conclusions should be checked using more sophisticated Faddeev equations with momentum-dependent exchange interactions, \emph{e.g}.\ extending Refs.\,\cite{Chen:2017pse, Chen:2019fzn}.
An extension of this analysis to radial excitations of the states considered herein is possible. However, that requires additional intervention because one must, by-hand, force zeros into what would otherwise be momen-tum-independent Faddeev amplitudes \cite{Roberts:2011cf}.
A potentially more worthwhile direction would be to adapt the framework to the challenge of understanding tetra- and penta-quark states \cite{Barabanov:2020jvn, Eichmann:2020oqt}.
A key longer-term goal is extension of the rainbow-ladder truncation three-body Faddeev equation study in Ref.\,\cite{Qin:2019hgk} so that it can: (a) directly handle non-degene-rate valence-quark systems (reducing/eliminating the need to use equal spacing rules); and (b) include beyond-rainbow-ladder contributions and thus both search for signals indicating the appearance of diquark correlations and deliver improved \emph{ab initio} predictions. In such an effort, the use of high-performance computing resources will be necessary.
\begin{acknowledgements}
We are grateful for constructive comments from C.~Chen, Y.~Lu and S-.S.~Xu.
Work supported by:
National Natural Science Foundation of China (under grant nos.\,11747140, 11805097);
Jiangsu Provincial Natural Science Foundation of China (under grant nos.\,BK20180738, BK20180323);
Nanjing University of Posts and Telecommunications Natural Science Foundation (under grant nos.\,NY218132, NY219063);
Jiangsu Province Innovation Program (under grant no.\,CZ0070619001);
Jiangsu Province \emph{Hundred Talents Plan for Professionals};
Chinese Ministry of Science and Technology \emph{International Expert Involvement Programme};
Ministerio Espa\~nol de Ciencia e Innovaci\'on, grant no. PID2019-107844GB-C22;
and Junta de Andaluc\'ia, contract nos.\ P18-FRJ-1132 and Operativo FEDER Andaluc\'ia 2014-2020 UHU-1264517.
\end{acknowledgements}
|
\section{Introduction}
\label{sec:introduction}
Recent machine learning breakthroughs help solve many tasks including facial recognition \cite{aajfacialadversarial}, surveillance \cite{aplkc20}, natural language processing tasks \cite{daelemans2002evaluation}, materials discovery \cite{edwardkimmaterialsdiscovery} and bio-authentication systems \cite{deng2019arcface}. One of these tasks where machine learning models can be utilized is to detect whether a driver is drowsy inside the car and help prevent accidents with an alert system. National Highway Traffic Safety Administration (NHTSA) data shows 37,461 people were killed in 34,436 motor vehicle crashes, an average of 102 per day in 2016. Of all these accidents, alcohol-impaired driving fatal crashes totaled 9477 in 2016 \cite{nhtsa2016}. A potential alert system that detects driver disengagements while driving could prevent these accidents where the cause was alcohol or drowsiness. An estimated 1 in 25 adult drivers (aged 18 or older) report having fallen asleep while driving in the previous 30 days \cite{drowsydriving1} \cite{drowsydriving2}.
Additionally, drowsy driving was responsible for 72,000 crashes, 44,000 injuries, and 800 deaths according to the National Highway Traffic Safety Administration (NHTSA) in 2013 \cite{drowsydriving3}. NHTSA also reports that up to 6,000 fatal crashes each year may be caused by drowsy drivers \cite{drowsydriving4} \cite{drowsydriving5} \cite{drowsydriving6}. Moreover, driving after going more than 20 hours without sleep is the equivalent of driving with a blood-alcohol concentration of 0.08\%, which is the U.S. legal limit. A person is three times more likely to be in a car crash if fatigued \cite{drowsydriving1}.
Many car manufacturers incorporate assisted driving or drowsiness alert systems to their cars. Such cars may have an eye tracking camera built-in that records driver's face while driver is on the wheel. Other form of such alerting systems currently implemented to production level cars by car manufacturers may include pressure sensors on the steering wheel. Pressure sensors can be used to measure any long duration where the driver's hands are not on the steering wheel. In the event that this scenario happens, a car's alerting system would alert the driver via sound, or apply emergency brakes in case the driver is sleepy, not conscious or passed out due to a health incident. Currently such implementations exist for high-end luxury cars \cite{mercedes}. Even though such steering interventions exist for high-end cars, the average car doesn't have such sophisticated drowsiness detection systems or intervention mechanisms. A general solution scalable to not only high-end luxury cars but also to general public cars could be a mobile application which could detect when driver closes their eyes for a certain amount of time and alert the driver via sound.
A previous work \cite{Alparslan2020} by the authors resulted in an initial prototype published as a mobile app for Android users. See \href{https://play.google.com/store/apps/details?id=driveup.facedetector}{here} for a working prototype of the said alerting system deployed and published as a mobile application.
Such mobile application would utilize a state-of-the-art neural network trained on images of people with closed and open eyes. Due to the nature of the task, it is crucial that such a neural network is agnostic to poor light conditions, out of zoom image captures, blurs, shadows, low resolution images, etc. Therefore, it is essential to have a robust model.
The title of this study, ``Robust SleepNets'', come from the idea of being able to detect fatigued or drowsy drivers (or any driver that might disengaged with driving for long duration) at the steering wheel. In this study, we assume eye closedness is a direct indicator of drowsiness so we focus on implementing neural network architectures that would detect eye closedness. We attack the models with adversarial attacks to study the robustness of the networks under adversarial conditions. The robustness acts as a proxy to real life conditions which might exhibit poor lightning, out-of-focus camera zoom, overexposed/underexposed shutter, height or width shift in the frame etc. We also investigate augmenting train data with the parameters described in \autoref{tab:data_augmentation_training_parameters} to study the impact of data augmentation on accuracy and adversarial defense.
In this study, we assume the following two problems are functionally equivalent.
\begin{enumerate}
\item Detecting drowsy driving
\item Detecting whether driver closes their eyes. (longer than some threshold that could be determined empirically. such threshold is not the focus of this study)
\end{enumerate}
In other words, we focus on detecting eye closedness to detect drowsy driving. Then, we \textbf{treat the real-word driving conditions such as poor light, out-of-focus camera etc as black-box adversarial attacks}. With these assumptions and motivations, we investigate the usage of adversarial training as a means to creating robust models that could detect eye closedness against said adversarial attacks. Additionally, we study the effect of training data augmentation.
This research is organized so that \autoref{sec:introduction} introduces the concept of drowsy driving and \autoref{sec:relatedwork} explores what has been done in this field. \autoref{sec:methodology} explains the dataset, the models, adversarial attacks, adversarial training and limitations that we use/have in this study. We report the results in \autoref{sec:results} and conclude the study in \autoref{sec:conclusion} and \autoref{sec:futurework} with summarizing what we have done in this study and discussing where the research might go in the future.
\section{Related Work}
\label{sec:relatedwork}
There has been previous results where eye detection was used as a gateway to detect drowsiness \cite{chen2017}. There has been little study to investigate the robustness of these models with adversarial attacks.
Adversarial attacks are inputs that look like the original images but with perturbations added to result in misclassifications in the classifier \cite{AdvAttacks} \cite{kim2020modeling} \cite{edwardkimregularization} \cite{kim2018deep}. Adversarial attacks can be created in the image domain \cite{aajfacialadversarial} as well as audio domain \cite{aab20audio}.
Adversarial training is one way of defending against these attacks since using adversarial attacks \cite{madry2017towards}. we can generate adversarial samples and then use these samples in our training to develop high accuracy models. Adversarial training as a defense depends on model and task at hand significantly.
Alparslan et al. \cite{Alparslan2020} investigated the eye and face models on driver detection and explored data augmentation to simulate real-world black-box adversarial image settings. Even though they did not apply adversarial attacks, or adversarial training, they claimed that adding a robust and systematic data augmentation to the training datasets would represent black-box attacks in a real-world scenario where driver face might be in too much light, or shadow, or it might be blurred if the camera angle is not adjusted.
In this study, we include data augmentation that Alparslan et al. included in their study. In addition to the data augmentation, we also investigate adversarial robustness by attacking the models with Projected Gradient Descent, Fast Gradient Sign and DeepFool attacks. Once we report the accuracy on the adversarially generated dataset, we feed the adversarial inputs back into the training dataset to defend against the attacks to study the possibility of creating robust models that would detect eye closedness and drowsiness in the presence of attacks.
\section{Methodology}
\label{sec:methodology}
We develop two models: first on eye images and second on face images. We use Eye-blink dataset \cite{Pan2007EyeblinkbasedAI} for the eye model and we use Closed-Eye in the Wild dataset \cite{SONG20142825} for the face model.
For the eye model, we use 3,108 images belonging to 2 classes (open and closed) to train and we use 776 images to cross-validate during training. We use 962 images that the model has never seen before to just test the data. For the eye model, all classification results in \autoref{table:all_results} are reported from these 962 images of test dataset.
For face model, we use 1,559 images belonging to 2 classes (open and closed) to train and we use 389 images to cross-validate during training. We use 485 images that the model has never seen before to just test the data. For the face model, all classification results in \autoref{table:all_results} are reported from these 485 images of test dataset.
Additionally, in this study, we investigate the possible impact of data augmentation against adversarial attacks. For both models, we report results for the model trained on augmented data as well as non-augmented data. The data augmentations that we apply are summarized below:
\begin{enumerate}
\label{list:data_augmentation}
\item \textbf{Rotation:} Image can be rotated randomly depending on the driver's position and the camera angle.
\item \textbf{Width Shift:} Image width might depend on the camera angle. The model needs to mitigate this randomness .
\item \textbf{Height Shift:} Image height might depend on the camera angle. The model needs to mitigate this randomness.
\item \textbf{Shear angle Shift:} Drivers plane intersects with the plane in which camera is mounted on a car. This creates additional randomness and the model needs to mitigate this randomness.
\item \textbf{Zoom:} The camera can be close or far to the driver and the model needs to mitigate this randomness.
\item \textbf{Horizontal Flip:} This doesn't correlate to real life setting, but the idea is that the driver's window will be always at its left side, which means the lightning conditions from the left side of the camera will be always poor compared to the right side. The model needs to detect fatigue regardless of the lightning hence the flip.
\item \textbf{Image Fill:} Image can be scaled down or up. The model needs to detect fatigue regardless.
\item \textbf{Scaling:} Image can be scaled down or up. The model needs to detect fatigue regardless.
\end{enumerate}
Additionally, we apply Projected Gradient Descent \cite{madry2017towards}, Fast gradient Sign Method \cite{goodfellow2014explaining} and DeepFool \cite{moosavi2015deepfool} to attack the models and report their accuracies.
\subsection{Adversarial Attacks}
\subsubsection{PGD}
\label{attack:pgd}
Projected Gradient Descent~\cite{madry2017towards} is a strategy for finding an adversarial example $x'$ for an input $x$ that satisfies a given norm-bound $\|x' -x\|_p \leq \epsilon$.
Let $B$ denote the $\ell_p$-ball of radius $\epsilon$ centered at $x$.
The attack starts at a random point $x_0 \in B$, and repeatedly sets
\begin{align*}
x_{i+1} &= \text{Proj}_{B}(x_i + \alpha \cdot g) \\
\quad \text{for } g &= \argmax_{\|v\|_p \leq 1} v^\top \nabla_{x_i} L(x_i, y) \;.
\end{align*}
Here, $L(x, y)$ is a suitable loss-function (e.g., cross-entropy), $\alpha$ is a step-size, $\texttt{Proj}_B$ projects an input onto the norm-ball $B$, and $g$ is the \emph{steepest ascent} direction for a given $\ell_p$-norm. E.g., for the $\ell_\infty$-norm, $\texttt{Proj}(z)$ is a clipping operator and $g = \texttt{sign}(\nabla_{x_i} L(x_i, y))$ .
\subsubsection{Fast Gradient Sign}
\label{attack:fgs}
The fast gradient sign \cite{goodfellow2014explaining} method optimizes for the $L_{\infty}$ distance metric and its advantage is fast running tim, which comes at the expense of generating images that are very similar to the original image.
Given an image $x$ the fast gradient sign method sets
\begin{equation*}
x' = x - \epsilon \cdot \text{sign}(\nabla \text{loss}_{F,t}(x)),
\end{equation*}
where $\epsilon$ is chosen to be sufficiently small so as to be undetectable,
and $t$ is the target label.
Intuitively, for each pixel, the fast gradient sign method uses the gradient
of the loss function to determine in which direction the pixel's intensity should be changed (whether it should be increased or decreased) to minimize the loss function; then, it shifts all pixels
simultaneously.
\subsubsection{DeepFool}
Deepfool \cite{moosavi2015deepfool} optimizes over $L_2$ distance metric with the assumption that neural networks are linear. Since neural networks are not linear, once a hyperplane is found (if found any) that seperates two classes, search terminates. DeepFool takes about 10x more than PGD and FGSM takes on average when we apply for our eye and face models.
We refer the authors to the work of Moosavi et al. \cite{moosavi2015deepfool} for more in-depth explanation.
\subsection{Adversarial Training}
Adversarial training \cite{fgm} \cite{kannan2018adversarial} \cite{madry2017towards} is the process where the adversarially generated samples are included in the training data in the hopes that the model will recognize the attacks next time sees it. In the current literature, adversarial training is one of the stronger defenses against adversarial attacks especially if it is combined with other defenses \cite{madry2017towards} \cite{tramer2020adaptive}. We adversarially train our augmented eye model, non-augmented eye model, augmented face model and finally non-augmented face model. During our adversarial training, we attack entirety of the training dataset via Projected Gradient Descent, Fast Gradient Sign Method and DeepFool attacks. Since we use one attack to adversarially train one model and there are three attacks, we report three adversarially trained models for the Eye-blink dataset and three adversarially trained models for the Closed Eyes in the Wild dataset. We repeat the entire process once for the augmented data case and once for non-augmented data case, which doubles our combinations.
\subsection{Data Augmentation}
Data augmentation is the process of generating new training data from the existing samples. Some data augmentation methods include adding rotation, adding text, adding zoom, height shift, width shift.
In this study, we also investigate the effect of data augmentation in the presence of adversarial attacks. The data augmentation parameters are the same for both eye and face model and can be seen in \autoref{tab:data_augmentation_training_parameters}. These parameters are the same as the parameters described in \cite{Alparslan2020}
\begin{table}[!htbp]
\caption{Noisified training data parameters are shown below. Adding noise to the training data to augment it simulates black-box settings in real-world scenarios where a driver face might blurred, occluded, underexposed/overexposed etc. In these scenarios, the data augmentation helps simulate an adversary under black box settings.}
\label{tab:data_augmentation_training_parameters}
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Change Type} & \textbf{Change Amount} \# \\ [0.5ex]
\hline
rotation range & 40$^o$\\
\hline
width shift range & 0.2 \\
\hline
height shift range & 0.2 \\
\hline
shear range & 0.2\\
\hline
zoom range & 0.2\\
\hline
horizontal flip & True\\
\hline
fill mode & 'nearest' \\
\hline
rescale & 1./255\\
\hline
\end{tabular}
\end{table}
\subsection{Models}
We use the same model architecture for the Eye-blink dataset and the Closed Eyes in the Wild dataset in order to eliminate the differences in accuracy that might arise due to architecture configurations. Since both dataset represents a binary classification problem, we argue that using the same architecture for both datasets doesn't cause problem to detect eye closedness.
For eye model, we use 3,108 images belonging to 2 classes as training data and 776 images belonging to 2 classes as validation data to train and cross validate our models (See \autoref{fig:accuracy_and_loss_augmented_eye}, \autoref{fig:accuracy_and_loss_nonaugmented_eye} to see the training accuracy and loss plots). Then, we report the testing accuracy on the testing dataset 962 images belonging to 2 classes (All numbers are from testing accuracy in \autoref{table:all_results}). The test dataset consist of images that are never seen before by the model.
For face model, we use 1,559 images belonging to 2 classes as training data and 389 images belonging to 2 classes as validation data to train and cross validate our models (See \autoref{fig:accuracy_and_loss_augmented_face},
\autoref{fig:accuracy_and_loss_nonaugmented_face} to see the training accuracy and loss plots). Then, we report the testing accuracy on the testing dataset 485 images belonging to 2 classes (All numbers are from testing accuracy in \autoref{table:all_results}). The test dataset consist of images that are never seen before by the model.
Model architectures can be seen in \autoref{tab:model_architecture}.
\begin{table}[!htbp]
\caption{Model Architecture for both models. We used the same architecture on the Eye-Blink and CWE dataset. The eye model and the face model both use binary cross entropy with Adam's optimizer on iterative gradient descent.}
\vspace{2em}
\label{tab:model_architecture}
\centering
\begin{tabular}{|c|c| c|}
\hline
\textbf{Layer type} & \textbf{Output Shape} & \textbf{Param} \# \\ [0.5ex]
\hline
Conv2D & (98, 98,6) & 60 \\
\hline
Average Pooling & (49,49,6) & 0 \\
\hline
Conv2D & (47,47,16) & 880 \\
\hline
Average Pooling & (23,23,16) & 0 \\
\hline
Flatten & 8464 & 0 \\
\hline
Dense & 120 & 1015800 \\
\hline
Dense & 84 & 10164 \\
\hline
Dense & 1 & 85 \\
\hline
\end{tabular}
\end{table}
\subsection{Dataset}
In this study, we use Eye-blink Dataset \cite{f14} data set to train our eye model. Eye Blink dataset has 2,100 closed and open eye images that are black and white. They are 24x24 pixels and only show eye patches. We use Closed Eyes In The Wild (CEW) \cite{f14} dataset to train our face model. CEW dataset has 1,192 subjects with both eyes closed and 1,231 subjects with eyes open. Some challenges of this set include amateur photography, occlusions, problematic lighting, pose, and motion blur.
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{figures/cew_dataset.png}
\caption{Samples from Closed Eyes in the Wild dataset. This dataset includes 1,231 opened-eye images of people and 1192 closed-eye images of people. Some difficulties with this dataset include blur, fade, shade and over/underexposing. Due to these innate difficulties, this dataset represents well the environment an actual driver might be in while driving in a real-world scenario.}~\label{fig:cew_dataset}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.9\columnwidth]{figures/eye_blink_dataset.png}
\caption{Samples from Eye-blink dataset. This dataset consists 2100 closed and open 24x24 pixel grayscale eye patch images. Some difficulties with this dataset include very low resolution and shade.}~\label{fig:figure2}
\end{figure}
\subsection{Limitations}
In this study, we examine many different combinations to see the full effect of attacks and defenses. We have two models and for each model, we repeat the experiments for the case when the training data is augmented and for the case where it is not augmented. We use three attacks on each model and also apply adversarial training for each of the attacks. Because we have about 16 different configurations when all combined in this study, we are generating grayscale adversarial images even though the input can be colored (applying the attack for each channel). This helps us speed up the training, attacking and defending duration. The readers for example will see from the \autoref{table:all_results} that PGD doesn't defend successfully when attacked. Researchers who might reproduce our combinations in the future might get different results if this nuance is not taken into consideration.
\section{Experiment Results and Evaluation}
\label{sec:results}
We report all the results in \autoref{table:all_results}. In our analysis, we need to consider two aspects of the attacks.
\begin{itemize}
\item How successful is the adversarial attack?
\item How successful is the defense?
\end{itemize}
In order to answer the first part, (i.e how successful the attacks are), we attack the entire test data and report how accurate the model is against them. The accuracies for this case are represented in parenthesis in the \autoref{table:all_results}. For example, for the case of Eye model \textbf{without} the data augmentation, PGD attack reports 81.70\% accuracy. This means the model is able to accurately assign a class label in the images adversarially altered via PGD attack, 81.70\% of the time. Since this accuracy value is fairly high, we conclude that PGD attack was not able to fool the classifier and failed as an \emph{attack}. The same model is then adversarially trained with the images that were altered via PGD attack, and reports 81.70\% accuracy, which suggests that the adversarial training doesn't have any success as a defense at all. Readers might be curious why the same accuracy is reported (81.70\% vs 81.70\%). This is quite possible if adversarial training is an identity function. (i.e adversarial training is not a successful defense since model reports the same accuracy for the attacks before the adversarial training and after the adversarial training).
A high accuracy value in the accuracy column of the \autoref{table:all_results} does not signify any conclusion regarding the success of the adversarial training. Mainly, the increase from the accuracy value inside the parenthesis tells how successful the adversarial training was in \autoref{table:all_results}. To answer the second part (i,e how successful the defenses), we calculate and compare that difference.
\subsubsection{PGD}
All PGD rows reports very high accuracy values ($> 81\%$) for all models after the adversarial training, but this does \emph{not} mean the adversarial training was successful since this can be attributed to the fact that PGD attack cannot lower the accuracy of the models to begin with. (All accuracies inside the parenthesis for the PGD rows are ($>79\%$). We conclude PGD cannot fool the classifier successfully. So, adversarial training via PGD attack doesn't offer significant improvements. On average PGD attack decreases the classifier accuracy 9.44\% for the eye model and 0.1\% for the face model.
On average PGD defense increases the classifier accuracy -1.2\% for the eye model and 0.2\% for the face model (Please note -1.2\% increase means 1.2\% decrease). This can be empirically checked via the samples generated by PGD in \autoref{sec:appendix} since adversarially generated samples that used PGD method look very similar to the original images. PGD attack/defense failure can be attributed to the fact that eye closedness detection is a very localized task on an image and PGD can't successfully alter such information.
\subsubsection{FGSM and DeepFool}
In the case of FGSM and DeepFool, as it can be seen in \autoref{table:all_results}, these two attacks can lower the accuracy of the models significantly compared to the baseline. (around 43\% )
On average FGSM attack decreases the classifier accuracy 42.24\% for the eye model and 32.22\% for the face model. On average FGSM defense increases the classifier accuracy 1.1\% for the eye model and 0\% for the face model.
On average DeepFool attack decreases the classifier accuracy 31.43\% for the eye model and 28.77\% for the face model. On average DeepFool defense increases the classifier accuracy -9.81\% for the eye model and -23.5\% for the face model. (Please note a -9.81\% increase means 9.81\% decrease so adversarially training a model via DeepFool attack actually worked the opposite way of its intent.)
For these attacks, however, we conclude the defense is not successful since adversarially trained models cannot produce accuracy values greater than the accuracy values on the adversarial attacks before the training. We conclude for FGSM and DeepFool that the attacks are extremely strong and defenses are not. We invite the reader to examine the adversarially generated samples in \autoref{sec:appendix}.
\subsubsection{Data augmentation}
For the face model, accuracy drops 17\% when training data is augmented compared to the case when training data is \emph{not} augmented.
However, we also see that data augmentation helps the classifier become more robust. On average for all the attacks, when training data is augmented, classifier reports 1.3\% higher accuracy when attacked.
For the eye model, accuracy drops 2\% when training data is augmented compared to the case when training data is \emph{not} augmented.
However, we also see that data augmentation helps the classifier become more robust. On average for all the attacks, when training data is augmented, classifier reports 6.14\% higher accuracy when attacked.
\subsubsection{Eye vs Face Model}
Accuracy values indicate that eye closedness detection is easier to achieve when done on only eye images. On average, eye models classify 12\% more accurately than face model for the base cases. This can be attributed to the fact that face image has redundant information such as hair, mouth, ears etc (i.e all part of the face other than eyes) that the classifier needs to learn whereas eye model does not have such overhead. Because eye detection is such localized task, this could also explain why data augmentation helps the eye model 5x more than the face model when attacked. Defending the area of eyes via augmentation is easier on the eye images than face images.
\begin{figure}[htbp!]%
\centering
\subfloat[Eye model accuracy for training and testing dataset versus epoch count ]{\includegraphics[scale=0.5]{figures/eye_nonaugmented_accuracy.png}}
\qquad
\subfloat[Eye model binary cross entropy loss for training and testing datasets versus epoch count]{\includegraphics[scale=0.5]{figures/eye_nonaugmented_loss.png} }%
\caption{Training and testing of eye model with Eye-blink dataset when training data is \textbf{non-augmented}.}
\label{fig:accuracy_and_loss_nonaugmented_eye}
\end{figure}
\begin{figure}[htbp!]%
\centering
\subfloat[Eye model accuracy for training and testing dataset versus epoch count ]{\includegraphics[scale=0.5]{figures/eye_augmented_accuracy.png}}
\qquad
\subfloat[Eye model binary cross entropy loss for training and testing datasets versus epoch count when training data is \textbf{augmented}]{\includegraphics[scale=0.5]{figures/eye_augmented_loss.png} }%
\caption{Training and testing of eye model with Eye-blink dataset when training data is \textbf{augmented}.}
\label{fig:accuracy_and_loss_augmented_eye}
\end{figure}
\begin{figure}[!htbp]%
\centering
\subfloat[Face model accuracy for training and testing dataset versus epoch count ]{\includegraphics[scale=0.5]{figures/face_nonaugmented_accuracy.png}}
\qquad
\subfloat[Face model binary cross entropy loss for training and testing datasets versus epoch count]{\includegraphics[scale=0.5]{figures/face_nonaugmented_loss.png} }%
\caption{Training and testing of face model with CEW dataset when training data is \textbf{non-augmented}.}
\label{fig:accuracy_and_loss_nonaugmented_face}
\end{figure}
\begin{figure}[!htbp]%
\centering
\subfloat[Face model accuracy for training and testing dataset versus epoch count ]{\includegraphics[scale=0.5]{figures/face_augmented_accuracy.png}}
\qquad
\subfloat[Face model binary cross entropy loss for training and testing datasets versus epoch count]{\includegraphics[scale=0.5]{figures/face_augmented_loss.png} }%
\caption{Training and testing of face model with CEW dataset when training data is \textbf{augmented}.}
\label{fig:accuracy_and_loss_augmented_face}
\end{figure}
\begin{table}[htbp!]
\centering
\caption{Accuracy results for the adversarially trained eye and face model. The third column (Config) reports the configuration of the attack type used during adversarial training. The accuracy column represents accuracy on newly generated adversarial samples after adversarial training. The increase compared to the accuracy value inside the parenthesis tells how successful the adversarial training was. Inside the parenthesis, it reports accuracy on original adversarial samples, which are generated on the model before its adversarial training. A high accuracy value inside the parenthesis means the attack couldn't lower the accuracy successfully. Overall, the eye model has performed better than the face model and data augmentation reduces accuracies more for the face model than eye model. }
\begin{adjustbox}{max width=\columnwidth}
\begin{tabular}{cc|c|c|c|c|c|}
\hline
\multicolumn{1}{|c|}{\textbf{Model}} &
\textbf{D.A} &\textbf{ Config } & \textbf{Accuracy} & \textbf{Precision} & \textbf{Recall} & \textbf{F-1 Score} \\ \hline
\multicolumn{1}{|c|}{\multirow{8}{*}{\textbf{Eye}}} &
{\multirow{4}{*}{\textbf{ W/O }}} & \textbf{Base} & 95 & 96 & 96 & 95.5\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{PGD} & 81.70 (81.70) & 84 & 82& 81\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{FGSM } & 54.47 (52.18) & 73 & 54& 42\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{DeepFool } & 55.09 (55.69) & 64 & 55& 45\\
\cline{2-7}
\multicolumn{1}{|c|}{} & {\multirow{4}{*}{\textbf{ W/ }}} & \textbf{Base} & 93 & 92 & 94& 93\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{PGD} & 85 (87.42) & 85 & 85& 85\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{FGSM } & 51.14 (51.14) & 26 & 51& 35\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{DeepFool } & 50.42 (69.44) & 45 & 50& 36\\
\hline
\multicolumn{1}{|c|}{\multirow{8}{*}{\textbf{Face}}} &
{\multirow{4}{*}{\textbf{ W/O }}} & \textbf{Base} & 90 & 90 & 90 & 89\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{PGD} & 81.03 (79.18) & 81 & 81& 81\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{FGSM } & 49.28 (49.28) & 24 & 49& 33\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{DeepFool } & 52.58(62.47) & 53 & 53& 48\\
\cline{2-7}
\multicolumn{1}{|c|}{} & {\multirow{4}{*}{\textbf{ W/ }}} & \textbf{Base} & 73 & 74 & 73& 73\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{PGD} & 82.68 (84.12)& 83 & 83& 83\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{FGSM } & 50.72 (50.72) & 26 & 51 & 34\\
\multicolumn{1}{|c}{} & \multicolumn{1}{|c|}{} & \textbf{DeepFool } & 49.90 (60.0) & 54 & 50 & 37\\
\hline
\end{tabular}
\end{adjustbox}
\label{table:all_results}
\end{table}
\section{Conclusion}
\label{sec:conclusion}
In this paper, we assumed eye closedness was a gateway to detecting driver fatigue. We trained two deep convolutional neural network models to detect eye closedness: one based on Eye-blink dataset and other based on Closed Eyes in the Wild dataset (CEW). Later, we crafted adversarial attacks via Fast Gradient Sign Method, Projected Gradient Descent Method and DeepFool attacks. Highest models for eye detection were baseline models without data augmentation with 95\% accuracy for the eye model and 90\% for face model. We conclude that PGD attack is not able to decrease classifier accuracy as much as FGSM and DeepFool attacks decrease (9.44\%, 42.24\%, 31.44\% decrease for eye model and 0.1\%, 32.22\%, -28.77\% decrease for the face model, respectively). Additionally, adversarially training the model with PGD attack or FGSM attack donot increase classifier accuracy. PGD reports 1.2\% (eye) and 0.2\% (face) accuracy decrease, FGSM reports 1.1\% (eye) and 0\% (face) accuracy increase and DeepFool reports 9.81\% (eye) and 23.5\% (face) accuracy \emph{decrease} which prove that DeepFool does not succeed at defending when used as adversarial training. Finally, eye model reports 6.14\% higher accuracy when it is trained on augmented data when attacked compared to the case where it is trained on non-augmented data. Face model reports 1.3\% higher accuracy when it is trained on augmented data when attacked compared to the case where it is trained on non-augmented data. We hope that our robustness study help emphasize the need for robust machine learning models in mission-critical systems in the presence of adversarial attacks. We also hope that this study gives more insight on robust eye closedness detection methods as well as the effect of data augmentation and adversarial attacks as defense tools.
\section{Future Work}
\label{sec:futurework}
In the future, combining the two models in an ensemble learning setting would yield different results for a given image and might be worthwhile to examine.
Additionally, another future work might include improving the inference duration to enable eye closedness detection in real-time or on a video.
\printbibliography
\newpage
\section{Appendix}
\vspace{3.5cm}
\label{sec:appendix}
\begin{figure}[htbp!]
\centering
\subfloat[Eye-blink dataset non-augmented samples ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/eye_non_augmented_base.png}}
\qquad
\subfloat[Eye-blink dataset non-augmented samples after applying PGD]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/eye_non_augmented_pgd.png} }%
\qquad
\subfloat[Eye-blink dataset non-augmented samples after applying FGSM ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/eye_non_augmented_fgsm.png}}
\qquad
\subfloat[Eye-blink dataset non-augmented samples after applying DeepFool ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/eye_non_augmented_deepfool.png}}
\caption{Baseline images, FGSM, PGD and DeepFool adverserial attacks on non-augmented Eye-blink dataset.}
\label{fig:eye_blink_without_augmentation}
\end{figure}
\begin{figure}[htbp!]
\centering
\subfloat[Eye-blink dataset augmented samples ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/eye_augmented_base.png}}
\qquad
\subfloat[Eye-blink dataset augmented samples after applying PGD]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/eye_augmented_pgd.png} }%
\qquad
\subfloat[Eye-blink dataset augmented samples after applying FGSM ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/eye_augmented_fgsm.png}}
\qquad
\subfloat[Eye-blink dataset augmented samples after applying DeepFool ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/eye_augmented_deepfool.png}}
\caption{Baseline images, FGSM, PGD and DeepFool adverserial attacks on augmented Eye-blink dataset.}
\label{fig:eye_blink_with_augmentation}
\end{figure}
\begin{figure}[htbp!]
\centering
\subfloat[CEW dataset non-augmented samples ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/face_non_augmented_base.png}}
\qquad
\subfloat[CEW dataset non-augmented samples after applying PGD]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/face_non_augmented_pgd.png} }%
\qquad
\subfloat[CEW dataset non-augmented samples after applying FGSM ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/face_non_augmented_fgsm.png}}
\qquad
\subfloat[CEW dataset non-augmented samples after applying DeepFool ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/face_non_augmented_deepfool.png}}
\caption{Baseline images, FGSM, PGD and DeepFool adverserial attacks on non-augmented Closed Eyes in the Wild (CEW) dataset.}
\label{fig:cew_without_augmentation}
\end{figure}
\begin{figure}[htbp!]
\centering
\subfloat[CEW dataset augmented samples ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/face_augmented_base.png}}
\qquad
\subfloat[CEW dataset augmented samples after applying PGD]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/face_augmented_pgd.png} }%
\qquad
\subfloat[CEW dataset augmented samples after applying FGSM ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/face_augmented_fgsm.png}}
\qquad
\subfloat[CEW dataset augmented samples after applying DeepFool ]{\includegraphics[width=0.7\columnwidth, scale=0.4]{figures/face_augmented_deepfool.png}}
\caption{Baseline images, FGSM, PGD and DeepFool adverserial attacks on augmented Closed Eyes in the Wild dataset.}
\label{fig:cew_with_augmentation}
\end{figure}
\end{document}
|
\section{Introduction}
The light nuclei have been studied within the view of the cluster feature for more than five
decades~\cite{Brink1966,Fujiwara1980,Tohsaki2001}, and various nuclear theories have been developed
for the study of nuclear clustering~\cite{Oertzen2006,Horiuchi1991,Ito2012}.
By assuming the cluster structure, various cluster states of light nuclei have been investigated
explicitly~\cite{Lyu2016,Suhara2010,Zhou2012}. However, as the number of the constituent clusters
and nucleons increases or nuclear system becomes dilute, the number of required basis wave functions
increases very quickly. Therefore, a method which can efficiently sift out the basis is highly
desired. For this purpose, many efforts have been made, such as the stochastic
sampling~\cite{Suzuki1998,Itagaki2003,Mitroy2013} and the imaginary-time development
method~\cite{Fukuoka2013}.
Recently, a newly time-dependent many-body theory has been developed in
Refs.~\cite{Imai2019,Zhou2020,Shin2021} for the calculations of Be and C isotopes. This real-time
evolution method (REM) generates the basis wave function using the equation of motion (EOM)
which has been applied in the study of heavy-ion collisions~\cite{Zhang2011,Ono1996,Ono2019}, but
now is found to be very effective in searching the basis wave functions for the microscopic
calculations because of its ergodic nature.
We intend to apply the REM on the $0^+$ ground state of $^6$He nucleus ($\alpha$+n+n). It is
a Borromean nucleus consisting of loosely bound and spatially extended three-body systems, typically
composed of a compact core plus two weakly bound neutrons
(n+n+core)~\cite{Ogan1999,Danilin1998,Zhukov1993}. These properties
can lead to the huge computational difficulties despite of its simple physical structure. Meanwhile,
the low lying states of $^6$Li ($\alpha$+n+p) also can be a good comparison, where the more compact
states ($1^+$ and $3^+$ states with $T=0$) and the dilute state ($0^+$ state with $T=1$) present
simultaneously. In this study, aiming to explore the applicability of REM, we will calculate the
$0^+$ ground state of $^6$He as well as the low lying states of $^6$Li to reproduce the halo and
un-halo properties of these states.
This paper is organized as follows: Section~\ref{sec:wavefunction} explains the framework of the
wave function and the real-time evolution method (REM). The numerical results including the energy
and the root-mean-square (r.m.s.) radius are presented and discussed in Sec.~\ref{sec:results}. The
conclusion is summarized in Sec.~\ref{sec:conclusion}.
\section{Framework}
\label{sec:wavefunction}
\subsection{Hamiltonian}
We begin with the Hamiltonian given below
\begin{equation}
\hat{H}=\sum_{i=1}^A \hat{t}_i - \hat{T}_{c.m.} + \sum_{i<j}^A \hat{v}_N + \sum_{i<j}^A \hat{v}_{C} + \sum_{i<j}^A \hat{v}_{LS}
\end{equation}
where $\hat{t}_i$ and $\hat{T}_{c.m.}$ denote the kinetic energies of each nucleon and the center of mass, respectively. $\hat{v}_N$ denotes the effective nucleon-nucleon interaction and $\hat{v}_C$ denotes the Coulomb interaction. $\hat{v}_{LS}$ denotes the spin-orbit interaction.
For the nucleon-nucleon interaction, we take the Volkov No.2 interaction~\cite{volkov1965} as
\begin{equation}
\begin{split}
V(\mathbf{r})=&(W - M\hat{P}^\sigma \hat{P}^\tau + B\hat{P}^\sigma - H\hat{P}^\tau)\\
&\times [V_1\text{exp}(-r^2/c_1^2)+V_2\text{exp}(-r^2/c_2^2)] ~.
\end{split}
\end{equation}
The corresponding exchange parameters are, $W=0.4$, $M=0.6$ and $B=H=0.125$. The parameters in the
Gaussian terms are, $V_1 = -60.65$ MeV, $V_2 = 61.14$ MeV, $c_1 = 1.80$ fm and $c_2 = 1.01$ fm.
We take the G3RS potential~\cite{Tamagaki1968,Yamaguchi1979} as the spin-orbit interaction,
\begin{equation}
V_{ls} = V_0(e^{d_1 r^2} - e^{d_2 r^2})\hat{P}_{31}\hat{L}\cdot \hat{S}~.
\end{equation}
The strength parameter $V_0$ is set to be 2000 MeV. The Gaussian parameters $d_1$ and $d_2$ are set
to be $5.0$ fm$^{-2}$ and $2.778$ fm$^{-2}$, respectively.
\subsection{Generator coordinate method}
In the current work, the single-particle wave function $\phi(\mathbf{r},Z)$ are expressed
in a Gaussian form multiplied by the spin-isospin part $\chi_{\tau,\sigma}$ as
\begin{equation}
\phi(\mathbf{r},Z) =
(\frac{2\nu}{\pi})^{3/4}\text{exp}[-\nu(\mathbf{r}-\frac{\bm{z}}{\sqrt{\nu}})^2+\frac{1}{2}z^2]\chi_{\tau,\sigma}.
\end{equation}
Here the coordinate $Z$ represents the generator coordinates, which includes the three-dimensional
coordinate $\bm{z}$ for the spatial part of the wave function as well as the spinor $a$ and $b$ for
the spin part $\chi_\sigma = a\ket{\uparrow}+b\ket{\downarrow}$. In this work, the spinor $a$ and
$b$ are also regarded as time-dependent variables which will be generated similarly to the spatial
coordinates as introduced later. The harmonic oscillator parameter $b=\sqrt{1/(2\nu)}=1.46$ fm,
which is same with that used in Refs.~\cite{Itagaki2003,Furumoto2018}.
We describe the $^6$He and $^6$Li as the $\alpha$-cluster plus two valence nucleon systems in
the wave function. Thus the corresponding wave function can be written as
\begin{equation}
\Phi(Z_1,Z_2,\bm{z}_\alpha)=\mathcal{A}\{\phi(\mathbf{r}_1,Z_1)\phi(\mathbf{r}_2,Z_2)\Phi_\alpha(\mathbf{r}_{3-6},\bm{z}_\alpha)\}~.
\end{equation}
Here $\Phi_\alpha$ is the wave function of the $\alpha$-cluster with $(0s)^4$ configuration. $\phi$
are the single-particle wave functions as introduced above, which are used to describe the valence
nucleons in $^6$He and $^6$Li. Thus, the coordinates $\mathbf{r}_{1}$ and $\mathbf{r}_{2}$ represent
the real spatial position of valence nucleons while $\mathbf{r}_{3-6}$ are for the nucleons in the
$\alpha$-cluster.
Within the framework of generator coordinate method (GCM), the final wave function is the
superposition of the basis wave functions with different sets of generator coordinates
($Z_1,Z_2,\bm{z}_\alpha$):
\begin{equation}
\Psi = \sum_i f_i \hat{P}^{J^\pi}_{MK}\Phi_i(Z_{1,i},Z_{2,i},\bm{z}_{\alpha,i})
\end{equation}
where $\hat{P}^{J^\pi}_{MK}$ is the parity and the angular momentum projector. The generator
coordinates $\mathbf{Z}$ can be obtained by solving the equation of motion in REM as explained in
the next subsection. The corresponding coefficients $f_i$ will be determined by the diagonalization
of the Hamiltonian.
\subsection{Real-time evolution method}
In the quantum system, the wave function should satisfy the Schrodinger equation at all times. Thus, the time-dependent variational principle holds for the intrinsic wave function mathematically:
\begin{equation}
\label{eq:timVarPrin}
\delta\int dt \frac{\bra{\Phi(\mathbf{Z}_1,\mathbf{Z}_2,\mathbf{Z}_\alpha)}i\hbar~d/dt - \hat{H} \ket{\Phi(\mathbf{Z}_1,\mathbf{Z}_2,\mathbf{Z}_\alpha)}}{\langle\Phi(\mathbf{Z}_1,\mathbf{Z}_2,\mathbf{Z}_\alpha)|\Phi(\mathbf{Z}_1,\mathbf{Z}_2,\mathbf{Z}_\alpha)\rangle} = 0
\end{equation}
Regarding the coordinate $Z$ as the function of the time $t$, we obtain the
equation of the motion (EOM) as
\begin{equation}
\label{eq:EOM1}
i\hbar\sum_{j=1,2,\alpha}\sum_{\sigma=x,y,z,a} C_{i\rho
j\sigma}\frac{dZ_{j\sigma}}{dt}=\frac{\partial \mathcal{H}_\text{int}}{\partial Z^*_{i\rho}}
\end{equation}
\begin{equation}
\label{eq:EOM2}
\mathcal{H}_\text{int}\equiv\frac{\bra{\Phi(\mathbf{Z}_1,\mathbf{Z}_2,\mathbf{Z}_\alpha)}\hat{H}\ket{\Phi(\mathbf{Z}_1,\mathbf{Z}_2,\mathbf{Z}_\alpha)}}{\langle\Phi(\mathbf{Z}_1,\mathbf{Z}_2,\mathbf{Z}_\alpha)|\Phi(\mathbf{Z}_1,\mathbf{Z}_2,\mathbf{Z}_\alpha)\rangle}
\end{equation}
\begin{equation}
\label{eq:EOM3}
C_{i\rho j\sigma}\equiv\frac{\partial^2 \text{ln}\langle\Phi(\mathbf{Z}_1,\mathbf{Z}_2,\mathbf{Z}_\alpha)|\Phi(\mathbf{Z}_1,\mathbf{Z}_2,\mathbf{Z}_\alpha)\rangle}{\partial Z^*_{i\rho}\partial Z_{j\sigma}}
\end{equation}
By following the EOM, from an initial wave function at $t=0$, the sets of the generator coordinates
($Z_1,Z_2,\bm{z}_\alpha$) for GCM can be yielded as a function of time $t$. The ensemble of the
basis wave functions $\Phi_i(Z_{1,i},Z_{2,i},\bm{z}_{\alpha,i})$ denoted by these sets of
the generated coordinates will hold the information of the quantum system. Thus, effective basis can
be generated.
In practical calculations, we choose the proper initial excitation energy (the definition can be
found in Ref.~\cite{Imai2019,Zhou2020}) for obtaining various cluster configurations in the
evolution. To avoid the clusters or valence nucleons move to unphysical regions, the rebound
condition is imposed in our REM calculations. By following the work in Ref.~\cite{Ono1996}, we add a
potential barrier to the Hamiltonian during the REM procedure with the form:
\begin{equation}
\begin{split}
V_\text{reb}=\frac{k}{2}\sum_i &f(|\mathbf{R}_i-\mathbf{R}_\text{c.m.}|)\\
f(x) = (x&-d)^2\theta(x-d) \\
\mathbf{R}_i=\frac{\text{Re}(\bm{z}_i)}{\sqrt{\nu}},~&\mathbf{R}_\text{c.m.}=\frac{4}{6}~\mathbf{R}_\alpha+\frac{1}{6}\sum_{j=1}^2
\mathbf{R}_j~.
\end{split}
\end{equation}
Here $\mathbf{R}_i$ and $\mathbf{R}_\text{c.m.}$ represent the spatial position of the $i$th valence
nucleon and the center of mass, respectively, so that $|\mathbf{R}_i-\mathbf{R}_{c.m.}|$ is the
distance between them. Because of the step function $\theta(x-d)$, the evolving valence nucleon will
face potential barrier when it is $d$ fm far from the center of mass, and be
smoothly pushed back in later evolution. We set the strength of the potential barrier $k=6$
MeV/fm$^2$, which determines how rapidly the height of the barrier increases. This value is not
physically important as long as it is not too large or too small. The rebound radius parameter $d$
is set to be $8$ fm in our calculations, which is large enough for the current work.
We perform the above REM process for the intrinsic wave function of $^6$He and obtain an ensemble of
basis. This ensemble of basis are used for both the calculations of $^6$He and $^6$Li.
\section{Results}\label{sec:results}
We firstly show the energy spectra for the low-lying states of $^6$He and $^6$Li nuclei in
Fig.~\ref{fig:enespec}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.9\hsize]{enespec.eps}
\caption{\label{fig:enespec} The energy spectra of $^6$He and $^6$Li. Ref. denotes the results
from the reference works~\cite{Itagaki2003,Furumoto2018}. REM denotes the results from the
current work. Exp. denotes the corresponding experimental data~\cite{Tilley2004}. The energy
is measured relative to the $\alpha$ + n + n threshold, which is set as 0 level with the dotted
line. For the calculated results of both in this work and the reference works, the energy of
$^4$He is $-27.57$ MeV, while it is $-28.30$ MeV in the experimental data.}
\end{center}
\end{figure}
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=0.6\hsize]{enecon.eps}
\caption{\label{fig:enecon} The energy convergence of $^6$He and $^6$Li from the REM calculations
concerning the successive addition of bases. The dash lines are the corresponding results from the
reference works~\cite{Itagaki2003,Furumoto2018}.}
\end{center}
\end{figure*}
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=0.7\hsize]{rmscon.eps}
\caption{\label{fig:rmscon}The r.m.s. radii of $^6$He and $^6$Li from the REM calculations
concerning the successive addition of bases. The dotted lines denote the results in the reference
works~\cite{Itagaki2003,Furumoto2018}.}
\end{center}
\end{figure*}
\begin{table*}[htbp]
\begin{center}
\caption{The numerical results of the $0^+$ ground state of $^6$He, as well as the $1^+$, $3^+$
and $0^+$ states of $^6$Li from the calculations of REM.} \label{table:num}
\vspace{2mm}
\begin{tabular*}{14cm}{ @{\extracolsep{\fill}} l c c c c}
\hline
&Energy (MeV) &Point matter (fm) &Point proton (fm) &Point neutron (fm)\\
\hline
$^6$He ($0^+$) &-28.37 &$2.71$ &$2.03$ &$2.99$\\
$^6$Li ($1^+$) &-30.92 &$2.65$ &$2.66$ &$2.65$\\
$^6$Li ($3^+$) &-29.87 &$2.42$ &$2.43$ &$2.41$\\
$^6$Li ($0^+$) &-27.58 &$2.79$ &$2.81$ &$2.77$\\
\hline
\end{tabular*}
\end{center}
\end{table*}
The experimental data and the corresponding results in the referenced
works~\cite{Itagaki2003,Furumoto2018} are also included for comparison. It should be noted that we
are using the same Hamiltonian and the same form of the basis wave functions. In
Fig.~\ref{fig:enespec}, it clearly shows that our REM method provides the almost consistent
results for the $0^+$ states of $^6$He and $^6$Li nuclei as the references. Besides, for the $1^+$
ground state and the $3^+$ excited state of $^6$Li, the wave functions from our REM procedure
provide better results than the reference work, which means that we have found more sufficient wave
function through the evolution with the EOM. These results support the validity of the
REM. Furthermore, it should be noted that we are using one ensemble of the basis for both of the
$^6$He and $^6$Li calculations, it is interesting that one EOM can reproduce both the $\text{T}=0$
states and $\text{T}=1$ states, and it indicates that the REM may have the potential for the
investigation of the isospin mixing states in the future study.
Next we shall check the accuracy of our calculations. We show the energy convergences with the
increasing number of basis in Fig.~\ref{fig:enecon}. It shows that the huge number of the basis have
been included and the binding energies of all these states are well converged. These results prove
that the number of basis in our calculations is sufficient to converge the energy
results. Furthermore, we can see that the converged results of $1^+$ and $3^+$ states in our
calculation are much lower than the results from the reference works. It denotes that the REM
procedure have found more effective basis, which should be included to the total wave function.
It is also an essential topic to investigate the halo property of the $^6$He nucleus as well as the
$^6$Li nucleus. The $0^+$ ground state of $^6$He is the well known two-neutron halo. Likewise, the
$0^+$ excited states of $^6$Li also has the controversial halo
property~\cite{Li2002}. To investigate the halo property in these two nuclei, we calculate the
root-mean-square (r.m.s.) radii of $^6$He and $^6$Li with the wave function from REM. The
corresponding results are shown in Fig.~\ref{fig:rmscon}.
In the left panel of this figure, the calculated r.m.s. radii of point matter, point proton and
point neutron of the $0^+$ state of $^6$He are $2.71$ fm, $2.03$ fm, and $2.99$ fm,
respectively. These results are showing the explicit halo property of the ground state of
$^6$He. From the right panel of Fig.~\ref{fig:rmscon}), one can also find that the r.m.s. radius of
point matter of the $0^+$ state of $^6$Li ($2.79$ fm) is larger than the radii of its $1^+$ ($2.65$
fm) and $3^+$ ($2.42$ fm). It implies that the $0^+$ state of $^6$Li can be treated as a halo state,
which is consistent with the experimental conclusion~\cite{Li2002}. These results show that the halo
property of these states can be naturally included in the ensemble of the basis from the
REM. Comparing with the reference works, we notice that our results on the r.m.s. radii are larger
than the results in the reference works, which are denoted by the dotted lines in
Fig.~\ref{fig:rmscon}. It indicates that our ensemble of basis from REM includes the basis, where
valence nucleons spread far from the core, so that we provide more dilute structure for the halo
states of $^6$He and $^6$Li nuclei than theirs.
In the end, the detailed numerical results are summarized in Table~\ref{table:num}. The current
results should be the most accuracy calculation on $^6$He and $^6$Li nuclei within the GCM
framework.
\section{Conclusion}
\label{sec:conclusion}
We perform the calculations for $^6$He and $^6$Li nuclei with a recently developed model named REM,
which can generate the ergodic ensemble of the basis wave functions. During this work, we generate
the basis wave functions from the procedure of REM and superpose them to construct the total wave
functions. The converged results for the energy and the r.m.s. radius of the $0^+$ state of $^6$He
as well as the $1^+$, $0^+$ and $3^+$ states of $^6$Li nuclei have been obtained in this work. The
halo properties of $^6$He and $^6$Li are well described in the current work, which indicates that
the REM can search the basis more efficiently. The current works on $^6$He and $^6$Li nuclei could
be the most accurate calculations within the GCM framework to date. The benchmark calculations
performed in this work can be instructive for further calculation with REM.
\begin{acknowledgement}
One of the author (M.K.) acknowledges that this work was supported by the JSPS KAKENHI Grant No. 19K03859
and by the COREnet program at RCNP Osaka University.
\end{acknowledgement}
\bibliographystyle{spphys}
|
\section{Introduction}
\label{sec:intro}
Topological insulators \cite{Asboth2016} (TIs) are one of the most intensively studied topics in condensed matter in recent years.
Paradigmatic examples of one-dimensional (1D) TIs, such as the Su-Shrieffer-Heeger (SSH) model \cite{Su1979}, exhibit midgap zero-energy edge states under open boundary conditions (OBC), which can be related to a nontrivial and quantized topological index characterizing the bulk bands below that energy gap, in what is commonly referred to as the bulk-boundary correspondence \cite{Hasan2010}.
Inspired by Dirac's derivation of his eponymous equation from taking the square-root of the Klein-Gordon equation, Arkinstall \textit{et al.} \cite{Arkinstall2017} proposed a scheme of relating the topological properties of a given 1D model with finite energy edge states to those of its squared model, a conventional TI with zero-energy edge states.
Accordingly, these models came to be known as square-root TIs ($\sqrt{\text{TIs}}$).
Later on it was realized \cite{Kremer2020} that the concept of $\sqrt{\text{TIs}}$ could be extended to bipartite models, since their squared versions appear in a block diagonal form, with one of them corresponding to the TI from which the topological features are inherited in the starting $\sqrt{\text{TI}}$.
Square-root topology was quickly extended to other 1D models \cite{Pelegri2019} including topological superconductors (TSs) and non-Hermitian systems \cite{Ezawa2020,Ke2020,Lin2021}, to higher-order TIs \cite{Song2020,Mizoguchi2020,Yan2020} [$d$-dimensional lattices hosting topological edge states in $(d-j)$-dimensions, with $j\geq 2$ \cite{Benalcazar2017,Benalcazar2017b,Pelegri2019c}], to topological semimetals \cite{Mizoguchi2021} and Chern insulators \cite{Ezawa2020}.
The general recipe for the construction of $\sqrt{\text{TIs}}$ and square-root TSs ($\sqrt{\text{TSs}}$) from their topological squared counterparts was developed by Ezawa \cite{Ezawa2020}.
It relies on the realization that, upon treating the tight-binding chain as a connected graph, one can construct the square-root versions of a given TI/TS by subdividing its tight-binding graph and renormalizing the resulting hopping parameters.
The main idea of this method is that subdivision of the tight-binding graph guarantees that it will become bipartite, even if it was not so before.
In turn, the bipartite property guarantees that the squared Hamiltonian can be written in a block diagonal form.
Focusing on 1D models, we show here how a further elaboration of the method in [\onlinecite{Ezawa2020}] allows one to contruct TIs/TSs of root degree $2^n$ ($\sqrt[2^n]{\text{TIs}}$/$\sqrt[2^n]{\text{TSs}}$), with $n$ any positive integer.
Furthermore, since the distance between the original TI/TS and the constructed $\sqrt[2^n]{\text{TIs}}$/$\sqrt[2^n]{\text{TSs}}$, measured by the number of successive squaring operations that have to be applied to the Hamiltonian of the latter in order to get to the former, grows with $n$, we codified the relation between the two by introducing the ``arborescence of $\sqrt[2^n]{\text{TIs}}$/$\sqrt[2^n]{\text{TSs}}$'', a term taken from graph theory, that enables one to keep track of the original topological features which are inherited by the edge states present in the starting model.
In a recent work \cite{Dias2021}, we already identified a specific subclass of 1D linear bipartite models, labeled sine-cosine models, which, after each squaring operation, retrieve a smaller self-similar version of themselves as one of the diagonal blocks, in what we described as a Matryoshka sequence.
The results analyzed there can be viewed as a notable subset appearing within the general framework of 1D $\sqrt[2^n]{\text{TIs}}$ that we draw here.
The rest of the paper is organized as follows.
In Sec.~\ref{sec:diamondchain}, we review the properties of a 1D $\sqrt{\text{TI}}$, namely the diamond chain with $\pi$-flux per plaquette, which we will take as our toy model from which $2^n$-root topology is derived.
In Sec.~\ref{sec:quarticroot}, we show how to construct the quartic-root TI from the $\sqrt{\text{TI}}$ introduced before, highlighting the relation between the edge states of the $\sqrt[4]{\text{TI}}$ and the topological state of the original TI.
In Sec.~\ref{sec:2nroottis}, we show how the method followed in the previous section to find the $\sqrt[4]{\text{TI}}$ can be replicated an arbitrary number of times to find the $\sqrt[2^n]{\text{TIs}}$ with a higher $n>2$ value.
The arborescence of $\sqrt[2^n]{\text{TIs}}$ is introduced here, as an intuitive way of relating any $\sqrt[2^n]{\text{TI}}$ with the original TI.
In Sec.~\ref{sec:2nrootss}, we derive the $\sqrt[2^n]{\text{TSs}}$ from the original TS, taken to correspond to the Kitaev chain \cite{Kitaev2001} mapped into its single-particle tight-binding analog model.
Finally, in Sec.~\ref{sec:conclusions} we present our conclusions.
\section{Diamond chain with $\pi$-flux per plaquette}
\label{sec:diamondchain}
Let us begin with a recapitulation of the properties of the diamond chain model with a $\pi$-flux per plaquette, depicted in Fig.~\ref{fig:diamond}(a), a known example of a 1D $\sqrt{\text{TI}}$ \cite{Kremer2020}. Under periodic boundary conditions (PBC), the square-root bulk Hamiltonian of this model, written in the $\{\ket{A(k)},\ket{B(k)},\ket{C(k)}\}$ basis, is given by
\begin{eqnarray}
H_{\sqrt{\text{TI}}}&=&
\sqrt{t}\begin{pmatrix}
0&h^\dagger_{\sqrt{\text{TI}}}
\\
h_{\sqrt{\text{TI}}}&0
\end{pmatrix},
\label{eq:hamilt_diamond}
\\
h^\dagger_{\sqrt{\text{TI}}}&=&
\begin{pmatrix}
1-e^{-ik}&\ \ \ 1+e^{-ik}
\end{pmatrix},
\end{eqnarray}
where the lattice constant is set to $a\equiv 1$ everywhere in this paper and the $k$-dependency is hidden in the terms and will remain so hereafter for convenience, except when deemed necessary.
The block antidiagonal form of (\ref{eq:hamilt_diamond}) indicates that the model is bipartite and, therefore, enjoys chiral-symmetry defined as $CH_{\sqrt{\text{TI}}}C^{-1}=-H_{\sqrt{\text{TI}}}$, with $C=\text{diag}(-1,1,1)$.
Diagonalization of the Hamiltonian in (\ref{eq:hamilt_diamond}) yields an all-bands flat energy spectrum \cite{Pelegri2020}, shown in Fig.~\ref{fig:diamond}(c), as a consequence of an Aharonov-Bohm caging effect induced by the $\pi$-flux in each plaquette \cite{Liberto2019,Pelegri2019b,Gligoric2020,Chang2021}.
Even though it would be more correct to speak of a \textit{square-root} energy spectrum, we will treat all $2^n$-root energy spectra in this paper as simply energy spectra.
In what follows, we set $t=1$.
Under open boundary conditions (OBC), there is a chiral pair of topological edge states with finite energies $E_{\text{edge}}^\pm=\pm\sqrt{2}$ [see in-gap states in the energy spectrum of an open chain in Fig.~\ref{fig:diamond}(d)], located around the end that terminates with a spinal A site.
A nonequivalent termination at both sites of the BC sublattice occurs at the oppposite end of the open chain, provided an integer number $N$ of unit cells is considered.
This feature is related to the fact that, even though the model is inversion ($\mathcal{I}$) symmetric, its $\mathcal{I}$-axis is shifted in relation to the center of the unit cell \cite{Marques2018,Pelegri2019,Madail2019} [see Fig.~\ref{fig:diamond}(a)].
As a consequence, the $\mathcal{I}$-symmetry operator becomes $k$-dependent \cite{Marques2019},
\begin{equation}
\mathcal{I}:\ \ \ \ P(k)H_{\sqrt{\text{TI}}}(k)P^{-1}(k)=H_{\sqrt{\text{TI}}}(-k),
\end{equation}
with $P(k)=\text{diag}(e^{ik},1,1)$.
The Zak phase of each band, a $Z_2$ topological index for 1D models \cite{Asboth2016}, can be expressed in this case as
\begin{equation}
\gamma_n=\arg(P_0^nP_\pi^n)-\int_{0}^{\pi}dk|u_{n,A}(k)|^2\ \ \mod 2\pi,
\label{eq:zakphase}
\end{equation}
where $P_0^n,P_\pi^n=\pm 1$ are the parity eigenvalues of band $n$ at the $\mathcal{I}$-invariant momenta $k=0,\pi$, respectively, and $u_{n,A}(k)$ is the A-component of the corresponding eigenstate $\ket{u_n(k)}$.
The second term in (\ref{eq:zakphase}) appears as a consequence of the mismatch between the $\mathcal{I}$-axis and the center of the unit cell \cite{Marques2019}.
Since finite energy eigenstates in bipartite sublattices have half their weight in each sublattice, one can immediately see that $|u_{\pm,A}(k)|^2=1/2$ for all $k$, where $n=+(-)$ is the index of the top (bottom) band.
The first term in (\ref{eq:zakphase}) is quantized to either 0 or $\pi$ such that, modulo $2\pi$, one necessarily has $\gamma_\pm=\frac{\pi}{2}$.
It should be noted that this fractional quantization is an artifact of $\sqrt{\text{TIs}}$ where the correction term in (\ref{eq:zakphase}) goes over an entire sublattice (as will be shown to occur also for the $\sqrt{\text{TS}}$ analyzed later on).
However, this does not hold in general for all 1D $\sqrt{\text{TIs}}$, that is, the Zak phases of their bands can have arbitrary values between 0 and $2\pi$, which will be the case for the $\sqrt{\text{TI}}$ in Appendix~\ref{app:2nrootssh}, as shown elsewhere \cite{Marques2019}.
The states of the zero-energy flat band $n=0$, on the other hand, have no weight on the A sublattice \cite{Pelegri2020}, so the usual $\pi$-quantization holds for $\gamma_0$.
\begin{figure*}[ht]
\begin{centering}
\includegraphics[width=0.9 \textwidth,height=9cm]{fig1}
\par\end{centering}
\caption{(a) Diamond chain with $\pi$-flux per plaquette, corresponding to the $\sqrt{\text{TI}}$; (b) squared model of (a), yielding the Creutz ladder on top (the original TI) with a perturbed right end (extra orange hopping) and an independent residual chain of decoupled A sites at the bottom.
Shaded regions represent the unit cell [only for the Creutz ladder in (b)], the blue cross its center and the vertical dashed green line the Inversion-axis.
(c) Energy spectrum of the model at the left, with $t\equiv 1$, as a function of the momentum under PBC and (d) the state index for the same chain with OBC and $N=10$ unit cells.
(e),(f) Same as in (c),(d), respectively, but for the model in (b).
The Zak phases for each band are indicated in the bulk spectra [cumulative for the degenerate bands in (e), with $\pi$ for the top band of the Creutz ladder and 0 for the band of decoupled A sites].}
\label{fig:diamond}
\end{figure*}
Upon squaring the square-root Hamiltonian in (\ref{eq:hamilt_diamond}) one arrives at
\begin{eqnarray}
H_{\sqrt{\text{TI}}}^2&=&
\begin{pmatrix}
H_{\text{res}}&0
\\
0&H_{\text{par}}
\end{pmatrix},
\\
H_{\text{res}}&=&h^\dagger_{\sqrt{\text{TI}}} h_{\sqrt{\text{TI}}}=2c_1,
\\
H_{\text{par}}&=&h_{\sqrt{\text{TI}}} h^\dagger_{\sqrt{\text{TI}}}=c_1\sigma_0+H_{\text{TI}},
\label{eq_hamilt_par}
\\
H_{\text{TI}}&=&\mathbf{d}\cdot\boldsymbol{\sigma},
\end{eqnarray}
where $\sigma_0$ is the $2\times 2$ identity matrix, $\boldsymbol{\sigma}=(\sigma_x,\sigma_y,\sigma_z)$ is the vector of Pauli matrices, $\mathbf{d}=c_1(0,\sin{k},-\cos{k})$ and $c_1=2t$ is a constant energy shift. We label $H_{\text{res}}$ ($H_{\text{par}}$) as the residual (parent) Hamiltonian and $H_{\text{TI}}$ the Hamiltonian of the original TI.
The real-space representation of $H_{\sqrt{\text{TI}}}^2$ with a general energy downshift of $c_1$, such that $H_{\text{par}}\to H_{\text{TI}}$ and $H_{\text{res}}\to c_1$, is depicted in Fig.~\ref{fig:diamond}(b), where it can be seen that $H_{\text{TI}}$ models a Creutz ladder with $\pi$-flux in each triangle \cite{Creutz1999,Creutz2001,Zurita2020,Flannigan2020,Kuno2020,Kuno2020b} in the BC sublattice, known to host a nondecaying zero-energy topological edge state at each end for all $|t|\geq 0$ under OBC, and $H_{\text{res}}$ represents a chain of decoupled A sites.
Notice that the $\mathcal{I}$-axis of the Creutz ladder can be placed at the center of the unit cell, such that the Zak phases of both its bands recover the usual $\pi$-quantization [see Fig.~\ref{fig:diamond}(e)].
Through a suitable basis rotation, $H_{\text{TI}}$ can be transformed into the Hamiltonian of a fully dimerized SSH model in the topological phase \cite{Kremer2020}.
The edge spectrum of the open squared chain, downshifted by $c_1$, for $N=10$ complete unit cells is shown in Fig.~\ref{fig:diamond}(f).
At first sight, one might be tempted to identify both midgap states with the topological states of the open Creutz ladder, our original TI.
There is, however, a subtle point that will prove to be of crucial importance throughout the rest of this paper: under OBC and for $N$ integer, the end sites of the starting square-root chain [see $\ket{A,1}$, $\ket{B,N}$, and $\ket{C,N}$ in Fig.~\ref{fig:diamond}(a)], have a lower coordination number than the rest of their equivalent bulk sites.
This will create an on-site energy offset upon squaring the Hamiltonian, akin to the offset in the self-energies for states with weight at the edge sites in second-order perturbation theory \cite{Marques2017},
\begin{eqnarray}
\bra{A,n}H_{\sqrt{\text{TI}}}^2\ket{A,n}&=&2c_1,\ \ n=,2,3,\dots,N,
\\
\bra{A,1}H_{\sqrt{\text{TI}}}^2\ket{A,1}&=&c_1,
\label{eq:impstate}
\\
\bra{\nu,n}H_{\sqrt{\text{TI}}}^2\ket{\nu,n}&=&c_1,\ \ n=,1,2,\dots,N-1,
\\
\bra{\nu,N}H_{\sqrt{\text{TI}}}^2\ket{\nu,N}&=&\frac{c_1}{2},
\end{eqnarray}
with $\nu=B,C$, and an extra hopping at the right edge of the Creutz ladder depicted in orange in Fig.~\ref{fig:diamond}(b),
\begin{equation}
\bra{B,N}H_{\sqrt{\text{TI}}}^2\ket{C,N}=t=\frac{c_1}{2}.
\end{equation}
As a result, the left and right topological states of the Creutz ladder, written as $\ket{T_L}=\frac{1}{\sqrt{2}}(\ket{B,1}+\ket{C,1})$ and $\ket{T_R}=\frac{1}{\sqrt{2}}(\ket{B,N}-\ket{C,N})$, respectively, will have different energies, on account of the correction terms at the right edge,
\begin{eqnarray}
H_{\sqrt{\text{TI}}}^2\ket{T_L}&=&c_1\ket{T_L},
\label{eq:topostateleft}
\\
H_{\sqrt{\text{TI}}}^2\ket{T_R}&=&0.
\label{eq:topostateright}
\end{eqnarray}
There is an impurity state at the left edge of the A chain, labeled $\ket{I_{\text{res}}}\equiv\ket{A,1}$, which is degenerate with $\ket{T_L}$ by comparing (\ref{eq:impstate}) and (\ref{eq:topostateleft}).
These are the two midgap states of Fig.~\ref{fig:diamond}(f), while $\ket{T_R}$ becomes degenerate with the bulk states of the lower band of the Creutz ladder, becoming a nontopological edge state due to the effect of the extra hopping term at the right end.
Returning now to the open square-root model, the finite energy states shown in Fig.~\ref{fig:diamond}(d) can be written, in the $\{\ket{A,1},\ket{B,1},\ket{C,1}\}$ basis, as
\begin{equation}
\ket{\text{Edge},\pm}=\frac{1}{2}
\begin{pmatrix}
\pm \sqrt{2}
\\
1
\\
1
\end{pmatrix}
=
\frac{1}{\sqrt{2}}
\begin{pmatrix}
\pm \ket{I_{\text{res}}}
\\
\ket{T_L}
\end{pmatrix},
\label{eq:ti2topoweight}
\end{equation}
that is, in relation to the squared model, each edge state of the starting square-root model can be said to be ``half'' topological and ``half'' impurity.
As we will see below, for $\sqrt[2^n]{\text{TIs}}$ the ratio of topological-to-impurity component for each edge state of the $2^n$-root model, relative to the original TI, will change in favor of the impurity component as $n$ increases.
\section{Quartic-root topological insulator}
\label{sec:quarticroot}
A $\sqrt[4]{\text{TI}}$ of our original TI [the Creutz ladder along the B and C sites of Fig.~\ref{fig:diamond}(b)] would have to be a given model that, upon squaring, would contain the $\sqrt{\text{TI}}$ model [the diamond chain of Fig.~\ref{fig:diamond}(a)] as one of its diagonal blocks, apart from an overall constant energy shift.
In other words, one has to ensure that the $\sqrt[4]{\text{TI}}$ chain: (i) keeps the same flux pattern as the $\sqrt{\text{TI}}$ model; (ii) is bipartite, and therefore has chiral symmetry and can be written in a block antidiagonal form; (iii) has a sublattice composed of the sites that will become the $\sqrt{\text{TI}}$ chain upon squaring; (iv) the sites in this sublattice have the same onsite energy upon squaring (constant energy shift).
The first three conditions are met by following the method outlined by Ezawa\cite{Ezawa2020} which, in essence, treats the problem as one of graph theory: one constructs the split graph \cite{Ma2020} through subdivision of the original graph (chain) by adding nodes (sites) at the middle of each link (hopping term), as illustrated by the added gray sites in Fig.~\ref{fig:quarticti}, which guarantees that the higher-degree root TI is bipartite, with one of its sublattices corresponding to the sites of the starting model (the blue nodes in the $\sqrt[4]{\text{TI}}$ of Fig.~\ref{fig:quarticti} form a sublattice given by the sites of the $\sqrt{\text{TI}}$).
In the process, the hopping terms are doubled in number and transformed as $\sqrt{t}e^{i\phi}\to\sqrt[4]{t}e^{i\phi/2}$, \textit{i.e.}, one takes the square-root of the magnitude and divides the phase by two, in order to keep the same $\pi$-flux per plaquette.
\begin{figure}[ht]
\begin{centering}
\includegraphics[width=0.45 \textwidth]{fig2}
\par\end{centering}
\caption{Construction of the $\sqrt[4]{\text{TI}}$ from the $\sqrt{\text{TI}}$.
The shaded regions indicate the respective unit cell.
The gray sites appear from subdivision of the $\sqrt{\text{TI}}$ (a site is included in the middle of each link). The resulting hopping terms have, in relation to the corresponding ones at the $\sqrt{\text{TI}}$, the square root of the (constant) magnitude and halved phase factors.
Green extra sites are included in the $\sqrt[4]{\text{TI}}$ to keep the same coordination number for all sites of the blue sublattice, which implies that sites $\ket{4,N+1}$ and $\ket{5,N+1}$ have to be included at the right end under OBC.}
\label{fig:quarticti}
\end{figure}
If we view the squaring of the Hamiltonian as a two-step quantum walk, where the on-site potential is given by the weighted sum of the paths that a particle can make from a given site to its connected neighbors in the first step and then hop back in the second step, it becomes clear that the last condition listed above is not met considering the blue and grey sites alone in Fig.~\ref{fig:quarticti} since the coordination number of the spinal sites $\ket{1,l}$ is four, where $\ket{j,l}$, with $j=1,2,\dots,11$ and $l=1,2,\dots,N$, is the state of a particle occupying site $j$ of unit cell $l$, while it is two for the same sublattice sites $\ket{2,l}$ and $\ket{3,l}$, resulting in
\begin{eqnarray}
\bra{1,l}H^2_{\sqrt[4]{\text{TI}}} \ket{1,l}&=&4\sqrt{t},
\label{eq:spinalsites}
\\
\bra{\beta,l}H^2_{\sqrt[4]{\text{TI}}}\ket{\beta,l}&=&2\sqrt{t},
\label{eq:topbottomsites}
\end{eqnarray}
with $\beta=2,3$.
As such, the on-site potentials at the relevant blue sublattice do not form a constant energy shift upon squaring.
\begin{figure}[ht]
\begin{centering}
\includegraphics[width=0.45 \textwidth]{fig3}
\par\end{centering}
\caption{Energy spectrum, with $t\equiv1$, as a function of the momentum for: (a) the $\sqrt[4]{\text{TI}}$ model given in (\ref{eq:hamilt_ti4}); (b),(c) the diagonal block of the $\sqrt[4]{\text{TI}}^2$ model corresponding to the $\sqrt{\text{TI}}$ model given in (\ref{eq:hamilt_diamond}) and to the $\sqrt{\text{res},2}^\prime$ model given in (\ref{eq:hamilt_res2}), respectively. The five-fold degenerate bands are indicated in the plots.}
\label{fig:espectrumti4k}
\end{figure}
There are two fundamental methods that allow one to circumvent this limitation and impose a constant energy shift upon squaring the Hamiltonian: (i) by renormalizing some of the hopping terms, which is more economical at the level of the number of sites per unit cell, at the cost of requiring successive renormalizations of the hopping terms after each squaring operation;
(ii) the magnitude of the hopping term is kept constant (that is, $\sqrt[2^n]{t}$ for all hoppings in $\sqrt[2^n]{\text{TI}}$) and extra sites are added and connected to the sites of the relevant sublattice with lower coordination number, in order to compensate for the difference on their on-site energies upon squaring the Hamiltonian \footnote{An intermediate way that combines the distinguishing aspects of both fundamental methods (renormalized hopping terms and extra added sites) is also possible.}.
Since method (ii) is valid for a general model, while method (i) can only be applied to specific models such as the diamond chain, the sawtooth chain, etc., we will use method (ii) in the following sections of the main text and refer the reader to Appendix~\ref{app:alternativemethod} for further details on method (i), which is expected to become the easiest of the two to handle as the root degree of the $\sqrt[2^n]{\text{TI}}$ increases.
The effect on the energy spectrum of the extra green sites in the unit cell will simply be to originate the same number of extra zero-energy flat bands which, according to Lieb's theorem \cite{Lieb1989}, is given by the imbalance in the number of sites in each sublattice (eight for the gray and green sublattice and three for the blue sublattice, yielding a total of five zero-energy bands).
Application of method (ii) leads to the inclusion of the extra green sites depicted in Fig.~\ref{fig:quarticti}, such that (\ref{eq:topbottomsites}) becomes now
\begin{equation}
\bra{\beta,l}H^2_{\sqrt[4]{\text{TI}}}\ket{\beta,l}=4\sqrt{t},
\end{equation}
yielding the same result as (\ref{eq:spinalsites}), as required.
\begin{figure*}[ht]
\begin{centering}
\includegraphics[width=0.9 \textwidth,height=9cm]{fig4}
\par\end{centering}
\caption{Energy spectrum, in units of $t\equiv 1$, as a function of state index $i$, obtained from diagonalization of: (a) $H_{\sqrt[4]{\text{TI}}}$, the Hamiltonian of the open $\sqrt[4]{\text{TI}}$ chain at the bottom of Fig.~\ref{fig:quarticti} with $N=10$ complete unit cells plus extra sites 4 and 5 at unit cell $N+1$; (b) $H_{\sqrt[4]{\text{TI}}}^{2^\prime}=H_{\sqrt[4]{\text{TI}}}^{2}-c_2$, with $c_2=4\sqrt{t}$; (c) $H_{\sqrt[4]{\text{TI}}}^{4^\prime}=H_{\sqrt[4]{\text{TI}}}^{2^\prime}H_{\sqrt[4]{\text{TI}}}^{2^\prime}-c_1$, with $c_1=2t$. Bulk (edge) states are colored in blue (red). (d) Profile of the edge states in the $\sqrt[4]{\text{TI}}$ chain, where the radius of the circle represents the amplitude of the wavefunction at the respective site and the color represents its phase, coded by the color bar at the right.}
\label{fig:espectrumti4r}
\end{figure*}
Under PBC, the bulk Hamiltonian of the $\sqrt[4]{\text{TI}}$, in the ordered $\{\ket{j(k)}\}$ basis, where $j=1,2,\dots,11$ refers to the $j^\text{th}$ component within the unit cell, has the form
\begin{eqnarray}
H_{\sqrt[4]{\text{TI}}}&=&
\sqrt[4]{t}\begin{pmatrix}
0&h^\dagger_{\sqrt[4]{\text{TI}}}
\\
h_{\sqrt[4]{\text{TI}}}&0
\end{pmatrix},
\label{eq:hamilt_ti4}
\\
h^\dagger_{\sqrt[4]{\text{TI}}}&=&
\begin{pmatrix}
i&1&0&1&1&0&0&0
\\
-ie^{ik}&0&1&1&0&0&1&0
\\
0&e^{ik}&0&0&1&1&0&1
\end{pmatrix},
\end{eqnarray}
whose squared version therefore becomes
\begin{eqnarray}
H_{\sqrt[4]{\text{TI}}}^2&=&
\begin{pmatrix}
H_{\sqrt{\text{par},2}}&0
\\
0&H_{\sqrt{\text{res},2}}
\end{pmatrix},
\label{eq:squaredti4}
\\
H_{\sqrt{\text{par},2}}&=&h_{\sqrt{\text{TI}}}^\dagger h_{\sqrt{\text{TI}}}=c_2I_3+H_{\sqrt{\text{TI}}},
\label{eq_hamilt_par}
\\
H_{\sqrt{\text{res},2}}&=&h_{\sqrt{\text{TI}}} h_{\sqrt{\text{TI}}}^\dagger=c_2I_8 +H_{\sqrt{\text{res},2}^\prime},
\label{eq:hamilt_res2}
\end{eqnarray}
where $c_2=4\sqrt{t}$ is a constant energy shift, explicitly included also in $H_{\sqrt{\text{res},2}}$ to keep the energy spectrum of $H_{\sqrt{\text{TI}}}$ and $H_{\sqrt{\text{res},2}^\prime}$ leveled, $I_m$ is the $m\times m$ identity matrix and $H_{\sqrt{\text{TI}}}$ is given in (\ref{eq:hamilt_diamond}) with $\{\ket{1(k)},\ket{2(k)},\ket{3(k)}\}\to\{\ket{A(k)},\ket{B(k)},\ket{C(k)}\}$.
Diagonalization of $H_{\sqrt[4]{\text{TI}}}$, $H_{\sqrt{\text{TI}}}$ and $H_{\sqrt{\text{res},2}^\prime}$ yields the energy spectra of Fig.~\ref{fig:espectrumti4k}.
It can be seen that the $\sqrt{\text{TI}}$ and $\sqrt{\text{res},2}^\prime$ models share the same spectrum, apart from the five-fold degenerate lower band in the latter, corresponding to squaring and downshifting by $c_2$ the degenerate zero-energy bands of the $\sqrt[4]{\text{TI}}$ model, originated by the imbalance in the number of sublattice sites, as explained above.
As expected, the spectrum of Fig.~\ref{fig:espectrumti4k}(b) exactly matches the one on Fig.~\ref{fig:diamond}(c), since both of them correspond to the same $\sqrt{\text{TI}}$ model.
Finally, there is an ambiguity in the definition of the unit cell of $\sqrt[4]{\text{TI}}$, since sites $\ket{4,1}$ and $\ket{5,1}$ can be added either to the left, as in Fig.~\ref{fig:quarticti}, or to the right.
The ambiguity is somewhat resolved, however, when OBC are considered, in which case one has to add these two sites at the right of the last unit cell also (see sites $4$ and $5$ at the incomplete $N+1$ unit cell in Fig.~\ref{fig:quarticti}), in order to keep the coordination number at the blue sites of both edges the same as for the bulk blue sites.
Since the extra sites perturb the right edge physics, preventing in general the emergence of right edge states, one could argue that a direct bulk-edge correspondence is lost, since the lower number of edge states under OBC will not have a correspondence with the cummulative Zak phases of the bulk bands below the energy gap where each of them lies.
On the other hand, even without the extra sites at the right edge under OBC the usual bulk-edge correspondence is still broken, since the $\mathcal{I}$-axis does not cross the center of the unit cell of the $\sqrt[4]{\text{TI}}$, meaning that the Zak phases of the bulk bands are not $\pi$-quantized in general.
The energy spectrum of the open $\sqrt[4]{\text{TI}}$ chain at the bottom of Fig.~\ref{fig:quarticti}, for $N=10$ complete unit cells plus sites 4 and 5 of the $N+1$ unit cell, is shown in Fig.~\ref{fig:espectrumti4r}(a), where two chiral pairs of nondegenerate edge states with energies
\begin{eqnarray}
E_{\text{edge}}^{\sqrt[4]{\text{TI}}}&=&\pm\sqrt{E_{\text{edge}}^{\sqrt{\text{TI}}}+c_2} \nonumber
\\
&=&\pm\sqrt[4]{2t}\sqrt{2\sqrt{2}\pm1},
\end{eqnarray}
where $E_{\text{edge}}^{\sqrt{\text{TI}}}=\pm\sqrt{2t}$ [see edge states in Fig.~\ref{fig:diamond}(d)].
The spatial profile of these edge states is shown in Fig.~\ref{fig:espectrumti4r}(d).
After squaring this spectrum and taking out the constant shift $c_2$, we arrive at the spectrum of Fig.~\ref{fig:espectrumti4r}(b), where the edge states now appear in two doubly degenerates pairs and the $E=0$ states in Fig.~\ref{fig:espectrumti4r}(a) become the $E=-4\sqrt{t}=-c_2$ states in Fig.~\ref{fig:espectrumti4r}(b).
Lastly, this squared and shifted spectrum is in turn squared and downshifted by $c_1=2t$ in Fig.~\ref{fig:espectrumti4r}(c).
All four edge states are now degenerate at zero energy, with only one of them corresponding to the topological state $\ket{T_L}$ of the original TI model, defined above (\ref{eq:topostateleft}), while the other three are impurity states, one stemming from the $H_{\text{res}^\prime}=H_{\text{res}}-c_1$ [the $\ket{I_{\text{res}}}$ state defined below (\ref{eq:topostateright})] and the other two, labeled $\ket{I_{\text{res},2}^1}$ and $\ket{I_{\text{res},2}^2}$, from $H_{\text{res},2^\prime}=H_{\sqrt{\text{res},2}^\prime}^2-c_1$, with $H_{\sqrt{\text{res},2}^\prime}^2$ defined in (\ref{eq:hamilt_res2}), involving only the sites in the green and gray sublattice depicted at the bottom of Fig.~\ref{fig:quarticti}.
We define the ancestor Hamiltonian as the direct sum of all the terms contained in Fig.~\ref{fig:espectrumti4r}(c),
\begin{equation}
H_\text{anc}:=H_{\text{TI}}\oplus H_{\text{res}^\prime}\oplus H_{\text{res},2^\prime},
\end{equation}
which is the Hamiltonian describing the ancestor chain.
The four edge states of the $\sqrt[4]{\text{TI}}$ chain in Fig.~\ref{fig:espectrumti4r}(a), labeled $\ket{T_4,j}$, with $j=1,2,3,4$ in an increasing energy order, can be written as linear combinations of the degenerate edge states in Fig.~\ref{fig:espectrumti4r}(c) with equal weight in all of them, resulting in
\begin{equation}
|\braket{T_L}{T_4,j}|^2=\frac{1}{4},
\label{eq:ti4topoweight}
\end{equation}
that is, the edge states of the $\sqrt[4]{\text{TI}}$ can be said to be, in relation to the ancestor chain, ``one quarter'' topological [notice the same $\ket{T_L}$ component appearing in all edge states of Fig.~\ref{fig:espectrumti4r}(d)] and ``three quarters'' impurity, which
should be compared with the case of the edge states of the $\sqrt{\text{TI}}$ chain in (\ref{eq:ti2topoweight}).
\section{$2^n$-root topological insulators}
\label{sec:2nroottis}
\begin{figure*}[ht]
\begin{centering}
\includegraphics[width=0.8 \textwidth,height=11cm]{fig5}
\par\end{centering}
\caption{Outwards directed rooted tree (arborescence) of $\sqrt[2^n]{\text{TIs}}$.
The root node on top corresponds to the Hamiltonian of the starting $\sqrt[2^n]{\text{TI}}$ model.
From each level to the next the Hamiltonian is squared and shifted down in energy by $c_{2^{n-j}}$, with $j=1,2,\dots,n$ the level index.
Each link selects one of the diagonal blocks of the resulting Hamiltonian, with the lower root degree TIs placed along the hypotenuse (each generating two children nodes from its block diagonal squared form) and the successive residual Hamiltonians placed at the right of those (each generating a single child node).
The process is stopped at the level where $H_\text{TI}$ is reached along the hypotenuse. The ancestor Hamiltonian $H_\text{anc}$ is defined as the direct sum over all leaf nodes.
$\#_{\text{edge}}$, indicated on the root and leaf nodes, is the number of edge states appearing from diagonalization of the Hamiltonian of the respective node.
In $H_\text{anc}$ only the red encircled edge state is topological, while all others are impurity states.}
\label{fig:tree}
\end{figure*}
The procedure followed in the last section to find the $\sqrt[4]{\text{TI}}$ model from the $\sqrt{\text{TI}}$ can be readily generalized.
More concretely, one can find the $\sqrt[2^n]{\text{TI}}$, for any $n\in \mathbb{N}$, from the $\sqrt[2^{n-1}]{\text{TI}}$ in four steps:
\\
1. One starts by subdividing the graph made by the $\sqrt[2^{n-1}]{\text{TI}}$ chain, that is, a new node (site) is introduced at the middle of each link (hopping term).
The resulting chain is therefore bipartite.
\\
2. The magnitude of the new hopping parameters is given by the square-root of the corresponding one in the $\sqrt[2^{n-1}]{\text{TI}}$ chain, while their phases are divided in half to keep the same flux pattern in the chain,
\begin{equation}
e^{i\phi}\sqrt[2^{n-1}]{t}\to e^{i\phi/2}\sqrt[2^{n}]{t}.
\end{equation}
Note that each link in the $\sqrt[2^{n-1}]{\text{TI}}$ chain generates two after subdivision.
\\
3. One has to ensure that the same squared on-site potential appears at the sites in the sublattice that decouples from the other one upon squaring of the Hamiltonian to generate the $\sqrt[2^{n-1}]{\text{TI}}$ chain.
First, one has to identify the sites of the relevant sublattice with the highest squared on-site potential (defined as $c_{2^{n-1}}$) and the others, with a lower squared on-site potential. To level the squared on-site potentials of the sites of these two subsets, one has to compensate for the difference of those in the latter subset by (i) renormalizing the hopping terms (the approach followed in Appendix~\ref{app:alternativemethod}), and/or (ii) introducing connections to extra sites, the approach followed in the previous sections.
The phases of the extra hopping terms constitute an additional degree of freedom that one can use to guarantee that the relevant sublattice remains in itself bipartite upon squaring the model, as will be illustrated in the next section for a $\sqrt[4]{\text{TS}}$.
\\
4. This last step only applies under OBC.
The extra sites of the $\sqrt[2^{n}]{\text{TI}}$ that result from subdivision of the $\sqrt[2^{n-1}]{\text{TI}}$ create an ambiguity in the definition of the unit cell, since the outermost new sites can be placed either to the right or to the left in the unit cell.
We chose to place these sites at the leftmost region within the unit cell (see sites $\ket{4,1}$ and $\ket{5,1}$ of the $\sqrt[4]{\text{TI}}$ chain at the bottom of Fig.~\ref{fig:quarticti}).
However, under OBC these outermost sites have to be included also at the right of the last complete unit cell, in order to avoid a lower squared on-site potential at the rightmost sites of the relevant sublattice (see sites $\ket{4,N+1}$ and $\ket{5,N+1}$ at the bottom of Fig.~\ref{fig:quarticti}, keeping at four the coordination number of the two rightmost blue sites).
\\
By applying $n$ successive squaring operations to the Hamiltonian of the $\sqrt[2^n]{\text{TI}}$ chain, and taking out the constant energy shift given by the on-site potential energy at the relevant sublattice after each operation (in order to keep this sublattice bipartite), a cascade of block diagonal Hamiltonians appear, containing both the residual blocks and the lower degree roots of $\sqrt[2^n]{\text{TI}}$.
A way to visualize this is by constructing the outwards directed rooted tree (arborescence \cite{Deo2017}) of $\sqrt[2^n]{\text{TIs}}$, depicted in Fig.~\ref{fig:tree}.
The root node at the top represents the Hamiltonian of the starting $\sqrt[2^n]{\text{TI}}$ chain, hosting $\#_{\text{edge}}=2^n$ non-degenerate edge states.
Each link connects a Hamiltonian to one of its block diagonal terms upon squaring and shifting down by the constant energy term of the respective level, with $n+1$ levels in total.
Nodes along the hypotenuse correspond to a given root degree of the original TI and have two child nodes, since their squared Hamiltonians yield, apart from a constant energy shift, a residual block and the block with the lower root degree TI [see, e.g., the squared $\sqrt[4]{\text{TI}}$ model in (\ref{eq:squaredti4})].
Nodes living outside the hypotenuse, however, correspond to the successive residual Hamiltonians that one has to continue squaring and shifting down by the constant of the respective level, such that each originates only one child node.
The process stops when we reach the TI model along the hypotenuse.
We define the ancestor Hamiltonian as the direct sum of all leaf nodes,
\begin{equation}
H_\text{anc}:=H_{\text{TI}}\oplus H_{\text{res}^\prime}\oplus H_{\text{res},2^\prime} \oplus \dots \oplus H_{\text{res},{2^{n-1}}^\prime}.
\label{eq:ancestorhamilt}
\end{equation}
The number of zero-energy edge states $\#_{\text{edge}}$ originating from each of the terms in (\ref{eq:ancestorhamilt}) is indicated at the bottom of Fig.~\ref{fig:tree}.
Only the encircled one stemming from $H_\text{TI}$ is topological in nature [the $\ket{T_L}$ state defined in (\ref{eq:topostateleft})], while all others are impurity states of the residual Hamiltonians that appear due to the detuned on-site potentials at their respective left edge sites.
As expected, there is a total of $2^n$ degenerate zero-energy states in $H_\text{anc}$, the same as in the root node $H_{\sqrt[2^n]{\text{TI}}}$, with each of the $\ket{T_{2^n},j}$ states, with $j=1,2,\dots,2^n$ in an increasing energy order, appearing in a different energy gap of the latter.
If we define the projector onto the impurity subspace of the ancestor chain,
\begin{equation}
\hat{I}=\sum\limits_{m=0}^n \sum\limits_{j=1}^{2^m} \ket{I_{\text{res,}2^m},j}\bra{I_{\text{res,}2^m},j},
\end{equation}
with $\ket{I_{\text{res,}2^0},1}\equiv\ket{I_{\text{res}}}$ defined below (\ref{eq:topostateright}), then we can determine the topological and impurity weights of the edge states of our starting $\sqrt[2^n]{\text{TI}}$ chain, relative to the edge states of the ancestor chain, as
\begin{eqnarray}
|\braket{T_L}{T_{2^n},j}|^2&=&\frac{1}{2^n},
\\
\bra{T_{2^n},j}\hat{I}\ket{T_{2^n},j}&=&\frac{2^n-1}{2^n},
\end{eqnarray}
for all $j=1,2,\dots,2^n$.
We can conclude that, as $n$ increases, the weight of the starting edge states on the topological state of the ancestor chain gets diluted, in favor of an increasing weight over the sum of all impurity states.
Focusing on the subspace of $2^n$ edge states of the starting $\sqrt[2^n]{\text{TI}}$ chain, a method for the layered decomposition of each of these states in terms of the edge states of the successive residual chains plus the topological state of the original TI chain is presented in Appendix~\ref{app:holographic}.
Interestingly, this decomposition is shown there to be reminiscent of the holographic duality that defines the anti-de-Sitter (AdS)/conformal field theory (CFT) correspondence.
For the case of the model we have been considering, whose original TI is given by the Creutz ladder depicted in Fig.~\ref{fig:diamond}(b), the values of the constant energy shifts, corresponding to the on-site potential at the sites forming the diagonal block in the squared Hamiltonian yielding the lower root degree TI, can be readily found to be
\begin{eqnarray}
c_1&=&2t,
\\
c_{2^m}&=&4\sqrt[2^m]{t},\ \ m=1,2,\dots,n-1,
\end{eqnarray}
which, setting $t\equiv 1$, simplifies to $c_1=2$ and $c_{2^m}=4$.
The energy of the $2^n$ edge states coming from the diagonalization of the root node $H_{\sqrt[2^n]{\text{TI}}}$ can be obtained recursively through
\begin{eqnarray}
E_{\sqrt[2^n]{\text{TI}}}&=&\pm\sqrt{E_{\sqrt[2^{n-1}]{\text{TI}}}+c_{2^{n-1}}},
\label{eq:eneredgeti2n}
\\
E_\text{TI}&=&0.
\label{eq:eneredgeti}
\end{eqnarray}
As a final example, let us follow the four steps outlined above to construct the $\sqrt[8]{\text{TI}}$ model.
Its unit cell is depicted in Fig.~\ref{fig:ti8ucell}.
\begin{figure}[ht]
\begin{centering}
\includegraphics[width=0.4 \textwidth]{fig6}
\par\end{centering}
\caption{43-sites unit cell of the $\sqrt[8]{\text{TI}}$ model. Blue sites form a sublattice yielding the $\sqrt[4]{\text{TI}}$ model at the bottom of Fig.~\ref{fig:quarticti} upon squaring, gray sites come from the subdivision of $\sqrt[4]{\text{TI}}$ and green sites are extra sites introduced to keep the coordination number constant for sites in the blue sublattice. The 10 open sites at the right belong to the next unit cell and have to be included at unit cell $N+1$ under OBC.}
\label{fig:ti8ucell}
\end{figure}
The blue sites form a sublattice corresponding to the unit cell of the $\sqrt[4]{\text{TI}}$ shown at the bottom of Fig.~\ref{fig:quarticti}, while the gray sites are a consequence of step 1, the subdivision of $\sqrt[4]{\text{TI}}$.
The magnitude of the new hopping parameters becomes $\sqrt[8]{t}$, while the phases of the dashed hopping terms are now $\frac{\pi}{4}$, that is, half the phase of the corresponding hoppings in the $\sqrt[4]{\text{TI}}$, in agreement with step 2, which keeps the same $\pi$-flux per plaquette pattern.
The green extra sites are introduced in order to keep the same coordination number at the sites in the blue sublattice such that, upon squaring $H_{\sqrt[8]{\text{TI}}}$, the on-site energy is the same for all blue sites, as required by step 3.
As a general rule of thumb, one identifies the sites within a unit cell that are positioned at the left of the spinal site as the set of sites that has to be included also at the right end under OBC, such that step 4 is fulfilled when, under OBC, the open sites in Fig.~\ref{fig:ti8ucell} are included in the incomplete $N+1$ unit cell for it to yield exactly, as one of its diagonal block and apart the $c_4$ energy shift, the $\sqrt[4]{\text{TI}}$ chain at the bottom of Fig.~\ref{fig:quarticti}.
\begin{figure*}[ht]
\begin{centering}
\includegraphics[width=0.95 \textwidth,height=5cm]{fig7}
\par\end{centering}
\caption{Energy spectrum, in units of $t\equiv 1$, as a function of state index $i$, obtained from diagonalization of: (a) $H_{\sqrt[8]{\text{TI}}}$, the Hamiltonian of the open $\sqrt[8]{\text{TI}}$ chain with the unit cell of Fig.\ref{fig:ti8ucell}, with $N=5$ complete unit cells plus the 10 open extra sites at unit cell $N+1$; (b) $H_{\sqrt[8]{\text{TI}}}^{2^\prime}=H_{\sqrt[8]{\text{TI}}}^{2}-c_4$, with $c_4=4\sqrt[4]{t}$; (c) $H_{\sqrt[8]{\text{TI}}}^{4^\prime}=H_{\sqrt[8]{\text{TI}}}^{2^\prime}H_{\sqrt[8]{\text{TI}}}^{2^\prime}-c_2$, with $c_2=4\sqrt{t}$; (d) $H_{\sqrt[8]{\text{TI}}}^{8^\prime}=H_{\sqrt[8]{\text{TI}}}^{4^\prime}H_{\sqrt[8]{\text{TI}}}^{4^\prime}-c_1$, with $c_1=2t$. Bulk (edge) states are colored in blue (red). Only a partial spectrum is shown in (b)-(d) for clarity, with some high energy bulk bands, in absolute value, outside the range of the y-axis.}
\label{fig:espectrumti8r}
\end{figure*}
The energy spectrum for the total Hamiltonian at each level of the arborescence of the $\sqrt[8]{\text{TI}}$, \textit{i.e.}, with $H_{\sqrt[8]{\text{TI}}}$ as the root node in Fig.~\ref{fig:tree}, for $N=5$ complete unit cells plus 10 extra sites at unit cell $N+1$, is shown in Fig.~\ref{fig:espectrumti8r}.
The progression shows how the eight nondegenerate edge states coming from diagonalization of $H_{\sqrt[8]{\text{TI}}}$ [in-gap red states in Fig.~\ref{fig:espectrumti8r}(a)] end up as the eight-fold degenerate zero-energy state coming from diagonalization of $H_{\text{anc}}$ [midgap states in Fig.~\ref{fig:espectrumti8r}(d)].
By applying steps 1 through 4, one can construct the $\sqrt[16]{\text{TI}}$ from the $\sqrt[8]{\text{TI}}$ in Fig.~\ref{fig:ti8ucell}.
We have been assuming in the main text that step 3 is satisfied by adding extra sites, while keeping the magnitude of the hopping terms the same everywhere, such that the number of sites per unit cell for the $\sqrt[2^n]{\text{TI}}$ follows the recurrence relation
\begin{eqnarray}
\#_{\text{uc}}^{2^n}&=&\#_{\text{uc}}^{2^{n-1}}+2^{2n-1}, \ \ n\geq 2,
\\
\#_{\text{uc}}^{2}&=&3,
\end{eqnarray}
meaning that one has $\#_{\text{uc}}^{16}=\#_{\text{uc}}^{2^4}=171$, $\#_{\text{uc}}^{32}=683$, etc.
In other words, as $n$ increases it becomes impractical to follow this method, with the one followed in Appendix~\ref{app:alternativemethod}, which relies on renormalizing the hopping parameters in place of adding extra sites, progressively gaining traction as the most tractable of the two.
It should be noted that the $\sqrt[2^n]{\text{TIs}}$ studied so far provide a practical way of generating 1D models with an all-bands-flat spectrum \cite{Danieli2020,Danieli2020b,Ichinose2021}, alternatively to a recent proposal of generating these spectra by taking advantage of a non-Abelian AB caging effect, related to the periodicity with which a $\pi$-flux appears in the plaquettes of the diamond chain \cite{Li2020,Mukherjee2020}.
Another type of $\sqrt[2^n]{\text{TI}}$, whose energy spectrum is in general dispersive, is analyzed in Appendix~\ref{app:2nrootssh}. This other type of $\sqrt[2^n]{\text{TI}}$ is labeled there as $\sqrt[2^n]{\text{SSH}}$, since the original TI is in this case taken to be the SSH model.
\section{$2^n$-root topological superconductors}
\label{sec:2nrootss}
In this section, we show how the same method for constructing $\sqrt[2^n]{\text{TIs}}$ can be applied for the construction of $\sqrt[2^n]{\text{TSs}}$, using the Kitaev model \cite{Kitaev2001} as the original TS, whose $\sqrt{\text{TS}}$ is already known \cite{Ezawa2020}.
We will go one step further and determine the $\sqrt[4]{\text{TS}}$, from which one can straightforwardly generalize for $n>2$.
\begin{figure*}[ht]
\begin{centering}
\includegraphics[width=0.95 \textwidth,height=8.3cm]{fig8}
\par\end{centering}
\caption{Single-particle tight-binding analog representation of the: (a) Kitaev TS model; (b) $\sqrt{\text{TS}}$ model; (c) $\sqrt[4]{\text{TS}}$ model.
Shaded regions represent the unit cell, and in (b) the blue cross marks its center and the vertical dashed green line the Inversion-axis.
Gray sites come from the subdivision of the chain to the left, while the green sites in $\sqrt[4]{\text{TS}}$ were included to keep the on-site potential constant in the blue sublattice sites upon squaring $H_{\sqrt[4]{\text{TS}}}$.
(d)-(f) Energy spectrum as a function of the momentum for the respective models on top under PBC and with $\Delta=t\equiv 1$.
The Zak phase of each band is indicated in (e) (cumulative for the degenerate zero-energy bands).}
\label{fig:ts4model}
\end{figure*}
The Hamiltonian of a periodic Kitaev chain, modeling a 1D quantum wire placed on top of a $p$-wave superconductor, reads as
\begin{equation}
H_\text{TS}=\sum\limits_k(c^\dagger_k\ \ c_{-k})
\begin{pmatrix}
2t\cos k-\mu&2\Delta\sin k
\\
2\Delta\sin k & -2t\cos k+\mu
\end{pmatrix}
\begin{pmatrix}
c_k
\\
c^\dagger_{-k}
\end{pmatrix},
\label{eq:hamiltkitaev}
\end{equation}
where $c_k$ is the particle annihilation operator acting on the state with momentum $k$, $\Delta$ is the superconducting pairing term induced in the chain by proximity effect with the superconductor, and $\mu$ the chemical potential.
Diagonalization of this Hamiltonian yields two symmetric flat bands for $t=\Delta$, as shown in Fig.~\ref{fig:ts4model}(d).
The presence of an annihilation operator in the Nambu pseudospinor $\mathbf{\Psi}_k=(c^\dagger_k\ \ c_{-k})$ prevents a real-space representation of the model by direct application of the inverse Fourier transform to (\ref{eq:hamiltkitaev}).
However, one can construct the single-particle tight-binding analog of the Kitaev chain through the mapping $\mathbf{\Psi}_k=(c^\dagger_k\ \ c_{-k})\to \mathbf{\Psi}_k=(c^\dagger_{1,k}\ \ c^\dagger_{2,k})$, that is, one treats each component of $\mathbf{\Psi}_k$ as if corresponding, e.g., to different atomic species or internal states.
Under this mapping, the real-space representation of the single-particle tight-binding analog model can be readily found to have the form of the Creutz ladder in Fig.~\ref{fig:ts4model}(a), with its ``electron channel'' at the top leg and its ``hole channel'' at the bottom leg.
Notice that the superconducting pairing is translated into an interleg crossed hopping term carrying a $\frac{\pi}{2}$ phase factor.
Thus, this system becomes amenable to the same four steps treatment outlined in the previous section for the construction of the $\sqrt[2^n]{\text{TI}}$.
We will further set $\mu=0$ below, otherwise the successive $\sqrt[2^n]{\text{TS}}$ models would not be bipartite and would therefore lack chiral symmetry, such that their squared models could not be written in a block diagonal form and our method would not be applicable.
For convenience we will also set $t\equiv 1$ as the energy unit and keep $\Delta$ as our degree of freedom.
Subdivision of the graph of the Kitaev model in Fig.~\ref{fig:ts4model}(a), along with the corresponding changes to the magnitudes and phases of the hopping parameters (steps 1 and 2), is enough to generate the $\sqrt{\text{TS}}$ of Fig.~\ref{fig:ts4model}(b), whose bulk Hamiltonian, written in the ordered $\{\ket{j(k)}\}$ basis, where $j=1,2,\dots,6$ refers to the $j^{th}$ component within the unit cell, is given by
\begin{eqnarray}
\mathcal{H}_{\sqrt{\text{TS}}}&=&
\begin{pmatrix}
0&h^\dagger_{\sqrt{\text{TS}}}
\\
h_{\sqrt{\text{TS}}}&0
\end{pmatrix},
\label{eq:hamilt_ts2}
\\
h_{\sqrt{\text{TS}}}&=&
\begin{pmatrix}
\sqrt{t}(1+e^{-ik})&\ \ \ 0
\\
e^{-i\frac{\pi}{4}}\sqrt{\Delta}&\ \ \ e^{-i(k-\frac{\pi}{4})}\sqrt{\Delta}
\\
e^{-i(k-\frac{\pi}{4})}\sqrt{\Delta}&\ \ \ e^{-i\frac{\pi}{4}}\sqrt{\Delta}
\\
0&\ \ \ -i\sqrt{t}(1-e^{-ik}).
\end{pmatrix},
\end{eqnarray}
whose diagonalization, for $t=\Delta$, yields the energy spectrum in Fig.~\ref{fig:ts4model}(e), where the two-fold degenerate zero-energy band comes as a consequence of Lieb's theorem\cite{Lieb1989}, equaling in number the imbalance within the unit cell between sites in the gray sublattice (four) and in the blue sublattice (two).
Given the noncentered $\mathcal{I}$-axis within the unit cell, the Zak phase of band $n$ in (\ref{eq:zakphase}) can be rewritten as
\begin{equation}
\gamma_n=\arg(P_0^nP_\pi^n)-\sum\limits_{j=3}^6\int_{0}^{\pi}dk|u_{n,j}(k)|^2\ \ \mod 2\pi,
\label{eq:zakphasets2}
\end{equation}
since the correction term on the right-hand side goes over the components of the eigenstates in the gray sublattice \cite{Marques2019}.
Similarly to the case of the diamond chain analyzed in Sec.~\ref{sec:diamondchain}, the Zak phases of all finite energy bands of the $\sqrt{\text{TS}}$ yield $\frac{\pi}{2}$, that is, they are not $\pi$-quantized, as shown in Fig.~\ref{fig:ts4model}(e).
Let us suppose we want to construct the $\sqrt[4]{\text{TS}}$ from the $\sqrt{\text{TS}}$ in Fig.~\ref{fig:ts4model}(b), whose sites become now the relevant blue sublattice of $\sqrt[4]{\text{TS}}$.
We start by applying steps 1 and 2, subdivision of $\sqrt{\text{TS}}$ and subsequent renormalization of the magnitudes and phases of the hopping parameters (remember that the superconducting pairings were converted into hopping terms), leading to the model depicted in Fig.~\ref{fig:ts4model}(c) \textit{without} the green sites.
Upon squaring the Hamiltonian, the on-site potential in sites 1 and 2 in the blue sublattice can be easily seen to yield $c_2=2\sqrt{\Delta}+2\sqrt{t}$, while for sites 3 and 6 (4 and 5) it yields only $2\sqrt{t}$ ($2\sqrt{\Delta}$) because they are missing, in relation to sites 1 and 2, two $\sqrt[4]{\Delta}$ ($\sqrt[4]{t}$) hopping connections.
Since we want the same $c_2$ on-site potential at all sites in the blue sublattice in order to satisfy step 3, we introduce the extra green sites in Fig.~\ref{fig:ts4model}(c) with judiciously chosen hopping terms to sites 3, 4, 5 and 6 such that they exactly compensate for the difference in their respective on-site potential, upon squaring, in relation to $c_2$.
Notice the $\frac{\pi}{2}$ included in the $\sqrt[4]{\Delta}$ hoppings connected to the green sites, introducing a $\pi$-flux around these small loops.
This is required in order to prevent finite couplings between sites 3 and 4 (5 and 6) upon squaring the Hamiltonian (for instance, there are two two-step hopping processes between sites 3 and 4, with $\sqrt[4]{t\Delta}$ magnitude but symmetric $\pm\frac{\pi}{2}$ phase factors, thus canceling out), which would break the bipartite condition and, therefore, not yield the $\sqrt{\text{TS}}$ in Fig.~\ref{fig:ts4model}(b) as one of its diagonal blocks.
The bulk Hamiltonian of the $\sqrt[4]{\text{TS}}$, written in the ordered $\{\ket{j(k)}\}$ basis, where $j=1,2,\dots,18$ refers to the $j^{th}$ component within the unit cell, is given by
\begin{widetext}
\begin{eqnarray}
\mathcal{H}_{\sqrt[4]{\text{TS}}}&=&
\begin{pmatrix}
0&h^\dagger_{\sqrt[4]{\text{TS}}}
\\
h_{\sqrt[4]{\text{TI}}}&0
\end{pmatrix},
\label{eq:hamilt_ts4}
\\
h_{\sqrt[4]{\text{TS}}}&=&
\begin{pmatrix}
e^{-ik}\sqrt[4]{t}&0&\sqrt[4]{t}&0&0&0
\\
e^{-i(k-\frac{\pi}{8})}\sqrt[4]{\Delta}&0&0&0&e^{-i\frac{\pi}{8}}\sqrt[4]{\Delta}&0
\\
0& e^{-i(k-\frac{\pi}{8})}\sqrt[4]{\Delta}&0& e^{-i\frac{\pi}{8}}\sqrt[4]{\Delta}&0&0
\\
0&e^{-i(k-\frac{\pi}{4})}\sqrt[4]{t}&0&0&0&e^{-i\frac{\pi}{4}}\sqrt[4]{t}
\\
0&0&-i\sqrt[4]{\Delta}&\sqrt[4]{t}&0&0
\\
0&0&0&0&\sqrt[4]{t}&i\sqrt[4]{\Delta}
\\
0&0&i\sqrt[4]{\Delta}&\sqrt[4]{t}&0&0
\\
0&0&0&0&\sqrt[4]{t}&-i\sqrt[4]{\Delta}
\\
\sqrt[4]{t}&0&\sqrt[4]{t}&0&0&0
\\
e^{-i\frac{\pi}{8}}\sqrt[4]{\Delta}&0&0&e^{i\frac{\pi}{8}}\sqrt[4]{\Delta}&0&0
\\
0&e^{-i\frac{\pi}{8}}\sqrt[4]{\Delta}&0&0&e^{i\frac{\pi}{8}}\sqrt[4]{\Delta}&0
\\
0&e^{-i\frac{\pi}{4}}\sqrt[4]{t}&0&0&0&e^{i\frac{\pi}{4}}\sqrt[4]{t}
\end{pmatrix},
\end{eqnarray}
\end{widetext}
whose diagonalization, for $t=\Delta$, yields the energy spectrum in Fig.~\ref{fig:ts4model}(f), where the six-fold degenerate zero-energy band comes, similarly to the $\sqrt{\text{TS}}$, as a consequence of Lieb's theorem (gray and green sites belong to the same sublattice).
\begin{figure*}[ht]
\begin{centering}
\includegraphics[width=0.95 \textwidth,height=4.5cm]{fig9}
\par\end{centering}
\caption{Energy spectrum, in units of $t\equiv 1$, as a function of $\Delta$ obtained from diagonalization of: (a) $H_{\sqrt[4]{\text{TS}}}$, the Hamiltonian of the open $\sqrt[4]{\text{TS}}$ chain with the unit cell of Fig.~\ref{fig:ts4model}(c), with $N=50$ complete unit cells plus extra sites 3 to 18 at unit cell $N+1$; (b) $H_{\sqrt[4]{\text{TS}}}^{2^\prime}=H_{\sqrt[4]{\text{TS}}}^{2}-c_2$, with $c_2=2\sqrt{t}+2\sqrt{\Delta}$; (c) $H_{\sqrt[4]{\text{TS}}}^{4^\prime}=H_{\sqrt[4]{\text{TS}}}^{2^\prime}H_{\sqrt[4]{\text{TS}}}^{2^\prime}-c_1$, with $c_1=2t+2\Delta$. Bulk (edge) states are colored in blue (red). Each red edge band is two-fold degenerate in (a), four-fold degenerate in (b) and eight-fold degenerate in (c).}
\label{fig:espectrumts4r}
\end{figure*}
As with the $\sqrt[2^n]{\text{TI}}$ studied above, there remains an ambiguity in the definition of the unit cell of the $\sqrt[2^n]{\text{TS}}$.
For instance, in the $\sqrt[4]{\text{TS}}$ of Fig.~\ref{fig:ts4model}(c), sites 3 through 18 have been placed at the left of sites 1 and 2 [those ultimately forming the original TS in Fig.~\ref{fig:ts4model}(a)] within the unit cell, but they could just as well been placed at the right.
If choosing one of the options dissipates the ambiguity, one must remember that step 4 requires that, under OBC, sites other than 1 and 2 have to appear at both ends.
Since we chose to grow the unit cell to the left, extra sites 3 through 18 have to be included at the incomplete unit cell $N+1$ at the right edge under OBC.
The energy spectrum of the open $\sqrt[4]{\text{TS}}$ with $N=50$ complete unit cell plus the 16 extra sites at unit cell $N+1$ is shown in Fig.~\ref{fig:espectrumts4r}(a).
Four doubly degenerate (red) bands of edge states are present at different energy gaps for finite $\Delta$.
Diagonalization of the squared Hamiltonian of the open $\sqrt[4]{\text{TS}}$ chain, followed by a $c_2=2\sqrt{t}+2\sqrt{\Delta}$ energy shift, yields the energy spectrum of Fig.~\ref{fig:espectrumts4r}(b), where the two red bands of edge states become four-fold degenerate.
Due to the $c_2$ shift, the six-fold degenerate band at $E=0$ of Fig.~\ref{fig:espectrumts4r}(a) becomes the lower $E=-c_2$ band in Fig.~\ref{fig:espectrumts4r}(b), which looses its flatness since $c_2$ depends on $\Delta$.
Upon squaring once more the Hamiltonian corresponding to the spectrum of Fig.~\ref{fig:espectrumts4r}(b) and subtracting a constant energy shift given by $c_1=2t+2\Delta$, one obtains the ancestor Hamiltonian $H_{\text{anc}}$ of the $\sqrt[4]{\text{TS}}$.
Its diagonalization yields the energy spectrum of Fig.~\ref{fig:espectrumts4r}(c).
The six-fold degenerate lowest energy band has $E=-c_1$ and the four-fold degenerate highest energy one has $E=c_2^2-c_1$.
The bands of edge states all collapse onto the eight-fold degenerate Majorana zero-energy red band.
For a finite $\Delta$, the $\sqrt[4]{\text{TS}}$ chain is in the topological phase and harbors $\#_{\text{edge}}=8$ edge states, double the number of the $\sqrt[4]{\text{TI}}$ chain [compare with Fig.~\ref{fig:espectrumti4r}(a)].
In general, one has $\#_{\text{edge}}=2^{n+1}$ edge states in the $\sqrt[2^n]{\text{TS}}$ chain.
The arborescence of $\sqrt[2^n]{\text{TSs}}$ is the same as in Fig.~\ref{fig:tree}, only doubling the number of edge states present in the root node and, consequently, also in all the leaf nodes.
Whereas the original TI we considered previously only contributed with one topological edge state to all $\sqrt[2^n]{\text{TI}}$, localized at the left edge, since the impurity hopping and on-site potential terms appearing at the right end send the right edge state to the bulk [see the Creutz ladder in Fig.~\ref{fig:diamond}(b) and energy spectrum of Fig.~\ref{fig:diamond}(f)], our original TS, the Kitaev chain depicted in Fig.~\ref{fig:ts4model}(a), contributes with one topological edge state at each end to all $\sqrt[2^n]{\text{TS}}$.
Note that these two oppositely located edge states would correspond, in the Majorana basis of $\gamma$ operators, to the two free $\gamma$ operators living at each end of the chain that combine to form a Majorana fermion.
In fact, it is because we want to keep this correspondence that, in contrast with the right edge perturbed Creutz ladder in Fig.~\ref{fig:diamond}(b), the original ladder in Fig.~\ref{fig:ts4model}(a) is unperturbed, requiring that gray sites from the incomplete unit cell $N+1$ to be already present at the right edge at the square-root level shown in Fig.~\ref{fig:ts4model}(b).
This is why we find double the number of edge states here, in relation to the same level of the TI studied in the previous sections.
It is clear that a $\sqrt[8]{\text{TS}}$ can be constructed from the $\sqrt[4]{\text{TS}}$ in Fig.~\ref{fig:ts4model}(c) by application of steps 1 to 4 and, by further repetition of the process, any $\sqrt[2^n]{\text{TS}}$ can be constructed.
The energy of the edge states can be determined recursively, as in (\ref{eq:eneredgeti2n}), by changing only the degeneracy of the initial value in (\ref{eq:eneredgeti}),
\begin{eqnarray}
E_{\sqrt[2^n]{\text{TS}}}&=&\pm\sqrt{E_{\sqrt[2^{n-1}]{\text{TS}}}+c_{2^{n-1}}},
\label{eq:eneredgets2n}
\\
E_\text{TS}&=&0\ \ (\times 2),
\label{eq:eneredgets}
\end{eqnarray}
where the constant energy shifts are now given by
\begin{equation}
c_{2^m}=2\sqrt[2^m]{t}+2\sqrt[2^m]{\Delta},
\end{equation}
with $m=0,1,2,\dots,n-1$.
Analogously to the case of a $\sqrt[2^n]{\text{TI}}$, all $\#_{\text{edge}}=2^{n+1}$ edge states present in a $\sqrt[2^n]{\text{TS}}$ have a diluting weight in the two topological edge states of the TS as $n$ increases and, conversely, an increasing weight in the impurity edge states that stem from the residual Hamiltonian terms in $H_{\text{anc}}$.
Remarkably, for $\Delta=t$ the model of Fig.~\ref{fig:ts4model}(a) becomes the Creutz ladder in Fig.~\ref{fig:diamond}(b) \text{without} the right edge hopping and on-site potential perturbations.
This shows that modifications to the same original TI/TS can generate different types of $\sqrt[2^n]{\text{TIs}}$/$\sqrt[2^n]{\text{TIs}}$, given that before the diamond chain with $\pi$-flux per plaquette was identified as the square-root version of the right edge perturbed Creutz ladder combined with a chain of decoupled sites, whereas here we constructed the $\sqrt{\text{TS}}$ and the $\sqrt[4]{\text{TS}}$ from the unperturbed Creutz ladder only.
Since it would be hard to guess \textit{a priori} that the very specific combination of Fig.~\ref{fig:diamond}(b) would be the one to generate a $\sqrt{\text{TI}}$, sometimes it is best in practice to start with a presumptive candidate for square-root topology, whose admission criteria are (i) being bipartite, (ii) having finite energy edge states and (iii) having nonquantized topological indices, and work from there in both directions, finding its original TI/TS along one and the higher root degree versions along the other.
\section{Conclusions}
\label{sec:conclusions}
We outlined here the procedure to go beyond square-root topology \cite{Ezawa2020}, towards topology with a root degree of any positive integer power of two ($2^n$-root topology, with $n\in \mathbb{N}$).
This was demonstrated for one-dimensional topological insulators and supercondutors, where the original models from which higher root degree versions were constructed were taken to be the Creutz ladder, for the former, which generated a family of extended diamond chains with a $\pi$-flux per plaquette, and the Kitaev chain mapped onto a single-particle tight-binding model, for the latter.
The connection between a $2^n$-root topological insulator/superconductor and its original topological insulator/superconductor was established through a graph shaped as an outwards directed rooted tree, where the Hamiltonian of the former is identified as the root node and the Hamiltonian of the latter as one of the leaf nodes.
Concerning the experimental realization of the $2^n$-root topological insulators studied here, several platforms appear as suitable candidates for their implementation, most notably photonic \cite{Baboux2016,Mukherjee2018,Zhang2019,XIa2020,Kremer2020,Jorg2020}, solid-state \cite{Huda2020} and optical lattices \cite{Tale2015}.
In what concerns the $2^n$-root topological superconductor model introduced here, on the other hand, its direct implementation in a superconducting system seems out-of-reach with current technology, as it would require some very fine local tuning of the phases of the superconducting and hopping terms.
Nevertheless, and as we explained in Sec.~\ref{sec:2nrootss}, one can simulate this model by translating it into its single-particle tight-binding analog model, which can be realized in the venues listed above or, as recently proposed \cite{Ezawa2019,Ezawa2020b}, in topoelectrical circuits \cite{Song2020,Li2020b,Olekhno2020,Wu2020,Yang2021}.
Our work invites future studies searching for deeper connections between a Hamiltonian and any of its positive integer powers.
We expect $2^n$-root topology to be generalizable to simply $n$-root topology.
In particular, odd-root topology would come as a natural extension of our results, since a block antidiagonal Hamiltonian will remain so when raised to any odd power (one should only be careful in taking out the constant energy shifts after each squaring operation).
In turn, raising this odd-power Hamiltonian to an even power will render it block diagonal.
For instance, one can treat six-root topology as a product of cubic-root and square-root topology.
It may be possible to construct, on this basis, a network of topological insulators such that, as with mathematical theorems, the set of those we elevate to an ``axiomatic'' status, like that implicitly granted to the Su-Schrieffer-Heeger model, would become a matter of convention.
The results concerning the extension of $2^n$-root topology to encompass two-dimensional Chern, weak and higher-order topological insulators, as well as topological semimetals, are being finalized and will be the subject of a forthcoming paper.
\section*{Acknowledgments}
\label{sec:acknowledments}
This work was developed within the scope of the Portuguese Institute for Nanostructures, Nanomodelling and Nanofabrication (i3N) projects UIDB/50025/2020 and UIDP/50025/2020 and funded by FCT - Portuguese Foundation for Science and Technology through the project PTDC/FIS-MAC/29291/2017. A.M.M. acknowledges financial support from the FCT through the work contract CDL-CTTRI-147-ARH/2018.
LM acknowledges financial support from the FCT through Grant No. SFRH/BD/150640/2020.
A.M.M. would like to thank Rui Am\'erico Costa for useful clarifications on graph theory.
|
\section{Introduction}
\label{sec:intro}
The production of lepton pairs through the Drell–-Yan~(DY) mechanism~\cite{Drell:1970wh} is the most {\it classic} hard-scattering process at hadron colliders.
It provides large production rates and offers clean experimental signatures, given the presence of at least one lepton with large transverse momentum in the final state.
Studies of the resonant production of $W$ and $Z$ bosons at the Tevatron and at the LHC
lead, for instance, to precise determinations of the $W$ mass~\cite{Group:2012gb,Aaboud:2017svj} and of the weak mixing angle \cite{Aaltonen:2018dxj,ATLAS:2018gqq}.
It is therefore essential to count on highly accurate theoretical predictions for the DY
cross sections and the associated kinematical distributions. This in turn implies the inclusion
of radiative corrections to sufficiently high perturbative orders in the strong and EW couplings $\as$ and $\alpha$.
The DY process was one of the first hadronic hard-scattering reactions for which radiative corrections were computed.
The pioneering calculations of the next-to-leading order~(NLO)~\cite{Altarelli:1979ub}
and next-to-next-to-leading order~(NNLO)~\cite{Hamberg:1990np,Harlander:2002wh} QCD corrections to the total cross section were followed by
(fully) differential NNLO computations including the leptonic decay of the vector boson~\cite{Anastasiou:2003yy,Anastasiou:2003ds,Melnikov:2006kv,Catani:2009sm,Catani:2010en}.
The complete EW corrections for $W$ production have been computed in Refs.~\cite{Dittmaier:2001ay,Baur:2004ig,Zykunov:2006yb,Arbuzov:2005dd,CarloniCalame:2006zq}
and for $Z$ production in Refs.~\cite{Baur:2001ze,Zykunov:2005tc,CarloniCalame:2007cd,Arbuzov:2007db,Dittmaier:2009cr}.
Very recently, the next-to-next-to-next-to-leading order~(N$^3$LO) QCD radiative corrections to the inclusive production of
a virtual photon~\cite{Duhr:2020seh} and of a $W$ boson~\cite{Duhr:2020sdp} became available.
Since the high-precision determination of EW parameters requires control over the kinematical distributions at very high accuracy,
the next natural step is the evaluation of mixed QCD--EW corrections. In the community, significant efforts have recently been devoted to this topic.
The mixed QCD--QED corrections to the inclusive production of an on-shell $Z$ boson were obtained
in Ref.~\cite{deFlorian:2018wcj} through an abelianisation procedure from the NNLO QCD results~\cite{Hamberg:1990np,Harlander:2002wh}.
This calculation was extended to the fully differential level for off-shell $Z$ boson production and decay into a pair of neutrinos (i.e.\ without final-state radiation) in Ref.~\cite{Cieri:2020ikq}.
A similar calculation was carried out in Ref.~\cite{Delto:2019ewv} in an on-shell approximation for the $Z$ boson, but including the factorised NLO QCD corrections to $Z$ production and the NLO QED corrections to the leptonic $Z$ decay.
Complete ${\cal O}(\as\alpha)$ computations for the production of on-shell $Z$ and $W$ bosons have been presented
in Refs.~\cite{Bonciani:2020tvf} and~\cite{Behring:2020cqi}, respectively.
Beyond the on-shell approximation, the most important results have been obtained in the {\it pole} approximation (see Ref.~\cite{Denner:2019vbn} for a general discussion).
Such approximation is based on a systematic expansion of the cross section around the $W$ or $Z$ resonance, in order to split the radiative corrections
into well-defined, gauge-invariant contributions. Contrary to the {\it narrow-width} approximation, the pole approximation (PA) describes off-shell effects of the
massive vector boson around the resonance.
Such method has been applied in Refs.~\cite{Dittmaier:2014qza,Dittmaier:2015rxo} to evaluate the so-called {\it factorisable} contributions of {\it initial--final} and {\it final--final} type. In these calculations the {\it initial--initial} contributions were neglected, and the so-called {\it non-factorisable} corrections were shown to be very small.
Given the relevance of mixed QCD--EW corrections for precision studies of DY production, and for an accurate measurement of the $W$ mass \cite{CarloniCalame:2016ouw},
it is important to go beyond these approximations.
New-physics effects, in particular, could show up in the tails of kinematical distributions, where the PA is not expected to work.
A first step in this direction has been carried out in Ref.~\cite{Dittmaier:2020vra}, where complete
results for the ${\cal O}(n_F\as \alpha)$ contributions to the DY cross section were presented.
One of the missing ingredients in the complete ${\cal O}(\as\alpha)$ calculation is the corresponding two-loop virtual amplitude.
The evaluation of the $2\to 2$ two-loop Feynman diagrams with internal masses is indeed at the frontier of current computational techniques.
Progress on the evaluation of the corresponding two-loop master integrals has been reported in Refs.~\cite{Bonciani:2016ypc,Heller:2019gkq,Hasan:2020vwn}. Very recently,
the computation of the complete two-loop amplitude for the neutral current dilepton production process was discussed in Ref.~\cite{Heller:2020owb}.
On the other hand, the required tree-level and one-loop amplitudes can nowadays be obtained with automated tools.
In this paper we focus on the charged-current DY production process
\begin{equation}
\label{eq:proc}
pp\to \ell^+\nu_\ell+X
\end{equation}
and present a new calculation of the mixed QCD--EW corrections.
At variance with previous work~\cite{Dittmaier:2014qza,Dittmaier:2015rxo}, we use the PA only to evaluate the finite part of the genuine two-loop contribution.
To this purpose, we carry out the required on-shell projection
by using the $W$ boson two-loop form factor recently presented in Ref.~\cite{Behring:2020cqi}, and we further improve our approximation through a reweighting technique.
All the remaining real and virtual ${\cal O}(\as\alpha)$ contributions are evaluated without any approximation. The corresponding tree-level and one-loop scattering amplitudes are computed with {\sc Openloops}~\cite{Cascioli:2011va, Buccioni:2017yxi, Buccioni:2019sur} and {\sc Recola}~\cite{Actis:2016mpe,Denner:2017wsf}, finding complete agreement.
The required phase space generation and integration is carried out within the {\sc Matrix} framework~\cite{Grazzini:2017mhc}.
The core of {\sc Matrix} is the Monte Carlo program {\sc Munich}\footnote{{\sc Munich}, which is the abbreviation of “MUlti-chaNnel Integrator at Swiss (CH) precision”, is an automated parton-level NLO\
generator by S. Kallweit.}, which contains a fully automated implementation of the dipole subtraction method for massless and massive partons
at NLO QCD~\cite{Catani:1996jh,Catani:1996vz,Catani:2002hc} and NLO EW~\cite{Kallweit:2017khh,Dittmaier:1999mb,Dittmaier:2008md,Gehrmann:2010ry,Schonherr:2017qcj}.
The cancellation of the remaining infrared singularities is achieved by using a formulation of the $q_T$ subtraction formalism~\cite{Catani:2007vq} derived from the NNLO QCD computation of heavy-quark production~\cite{Catani:2019iny,Catani:2019hip,Catani:2020kkl} through a suitable abelianisation procedure~\cite{deFlorian:2018wcj,Buonocore:2019puv}.
The paper is organised as follows. In Sec.~\ref{sec:HV-PA} we set up the formalism and detail the use of the PA and the reweighting procedure exploited in the computation. In Sec.~\ref{sec:num} we validate our method and present our results for the mixed QCD--EW corrections to the charged-current DY process. Our findings are summarised in Sec.~\ref{sec:summa}.
\section{Hard virtual function in the pole approximation}
\label{sec:HV-PA}
The differential cross section for the process in Eq.~(\ref{eq:proc}) can be written as
\begin{equation}
\label{eq:exp}
d{\sigma}=\sum_{m,n=0}^\infty d{\sigma}^{(m,n)}\, ,
\end{equation}
where $d{\sigma}^{(0,0)}\equiv d{\sigma}_{\rm LO}$ is the Born level contribution and $d{\sigma}^{(m,n)}$ the ${\cal O}(\as^m\alpha^n)$ correction.
The mixed QCD--EW corrections correspond to the term $m=n=1$ in this expansion.
According to the $q_T$ subtraction formalism~\cite{Catani:2007vq} $d{\sigma}^{(m,n)}$ can be evaluated as
\begin{equation}
\label{eq:master}
d{\sigma}^{(m,n)}={\cal H}^{(m,n)}\otimes d{\sigma}_{\rm LO}+\left[d\sigma_{\rm R}^{(m,n)}-d\sigma_{\rm CT}^{(m,n)}\right]\, .
\end{equation}
The first term in Eq.~(\ref{eq:master}) is obtained through a convolution (denoted by the symbol $\otimes$),
with respect to the longitudinal-momentum fractions $z_1$ and $z_2$ of the colliding partons,
of the perturbatively computable function ${\cal H}^{(m,n)}$ with the LO cross section $d{\sigma}_{\rm LO}$.
The second term is the {\it real} contribution $d\sigma_{\rm R}^{(m,n)}$, where the charged lepton and the corresponding neutrino
are accompanied by additional QCD and/or QED radiation that produces a recoil with finite transverse momentum $q_T$.
For $m+n=2$ such contribution can be evaluated by using the dipole subtraction formalism~\cite{Catani:1996jh,Catani:1996vz,Catani:2002hc,Kallweit:2017khh,Dittmaier:1999mb,Dittmaier:2008md,Gehrmann:2010ry,Schonherr:2017qcj}.
In the limit $q_T\to 0$ the real contribution
$d\sigma_{\rm R}^{(m,n)}$ is divergent, since the recoiling radiation becomes soft and/or collinear to the initial-state partons.
The role of the third term, the counterterm $d\sigma_{\rm CT}^{(m,n)}$, is to cancel the singular behaviour in the limit $q_T\to 0$, thereby rendering the cross section in Eq.~(\ref{eq:master}) finite.
Besides the numerous applications at NNLO QCD for the production of colourless final-state systems (see Ref.~\cite{Grazzini:2017mhc} and references therein), and for heavy-quark production~\cite{Catani:2019iny,Catani:2019hip,Catani:2020kkl}, which correspond to the case $m=2$, $n=0$, the method has been applied in Ref.~\cite{Buonocore:2019puv} to study NLO EW corrections to the Drell--Yan process, which represents the case $m=0$, $n=1$.
The structure of the coefficients ${\cal H}^{(0,1)}$ and $d\sigma_{\rm CT}^{(0,1)}$ can indeed be straightforwardly worked out from the corresponding coefficients
entering the NLO QCD computation of heavy-quark production through a suitable abelianisation procedure~\cite{Buonocore:2019puv}.
The structure of the coefficients ${\cal H}^{(1,1)}$ and $d\sigma_{\rm CT}^{(1,1)}$ can also be derived from those controlling
the NNLO QCD computation of heavy-quark production.
The initial-state soft/collinear and purely collinear contributions were already presented in Ref.~\cite{Cieri:2020ikq}.
The fact that the final state is colour neutral implies that final-state radiation is of pure QED origin.
Therefore, the purely soft contributions (one-loop diagrams with a soft photon or gluon and tree-level diagrams with a soft photon and a soft gluon)
have a much simpler structure compared to the corresponding contributions entering the
NNLO QCD computation of Refs.~\cite{Catani:2019iny,Catani:2019hip,Catani:2020kkl} and can be directly constructed from
those appearing at ${\cal O}(\alpha)$.\footnote{The only additional technical complication is the fact that the heavy quark is replaced by a much lighter lepton whose mass regularises final state QED collinear singularities.}
The only missing perturbative ingredient is therefore the ${\cal O}(\as\alpha)$ two-loop amplitude entering the coefficient ${\cal H}^{(1,1)}$ in Eq.~(\ref{eq:master}).
The coefficient ${\cal H}^{(m,n)}$ can be decomposed as
\begin{equation}
{\cal H}^{(m,n)}=H^{(m,n)}\delta(1-z_1)\delta(1-z_2)+\delta{\cal H}^{(m,n)}\,,
\end{equation}
where the hard contribution $H^{(m,n)}$ contains the $(m+n)$-loop virtual corrections.
More precisely, we define
\begin{equation}
\label{eq:H01def}
H^{(0,1)}=\frac{2{\rm Re}\left({\cal M}_{\rm fin}^{(0,1)}{\cal M}^{(0,0)*}\right)}{|{\cal M}^{(0,0)}|^2}
\end{equation}
and
\begin{equation}
\label{eq:H11def}
H^{(1,1)}=\frac{2{\rm Re}\left({\cal M}_{\rm fin}^{(1,1)}{\cal M}^{(0,0)*}\right)}{|{\cal M}^{(0,0)}|^2} \, .
\end{equation}
In Eqs.~(\ref{eq:H01def},\ref{eq:H11def}) ${\cal M}^{(0,0)}$ is the Born amplitude, while ${\cal M}_{\rm fin}^{(0,1)}$ and ${\cal M}_{\rm fin}^{(1,1)}$ (defined through an expansion analogous to Eq.~(\ref{eq:exp})) are the finite parts of the renormalised virtual amplitudes entering the NLO EW and the mixed QCD--EW calculations, respectively. Their explicit expressions in terms of the renormalised virtual amplitudes ${\cal M}^{(1,0)}$, ${\cal M}^{(0,1)}$, ${\cal M}^{(1,1)}$ after subtraction of the infrared poles in $d=4-2\epsilon$ dimensions read
\begin{align}
{\cal M}^{(1,0)}_{\rm fin}=&\;{\cal M}^{(1,0)}+ \f{1}{2}\left(\frac{\as}{\pi}\right)C_F\left[\f{1}{\epsilon^2} + \left(\f{3}{2} + i\pi\right)\f{1}{\epsilon} - \f{\pi^2}{12} \right]\amp^{(0)}\,,
\label{eq:M10}\\
{\cal M}^{(0,1)}_{\rm fin}=&\;{\cal M}^{(0,1)}+ \f{1}{2}\left(\frac{\alpha}{\pi}\right)\left\{ \left[ \f{1}{\epsilon^2} + \left(\f{3}{2} + i\pi\right)\f{1}{\epsilon} - \f{\pi^2}{12}\right] \frac{e^2_u+e^2_d}{2} -\f{2\Gamma_t}{\epsilon}\right\}\amp^{(0)}\,,\label{eq:M01}\\
\label{eq:M11}
\ampF^{(1,1)}=&\;{\cal M}^{(1,1)}-\left(\frac{\as}{\pi}\right)\left(\frac{\alpha}{\pi}\right)\Bigg\{\f{1}{8\epsilon^4}(e_u^2+e^2_d)C_F+ \f{1}{2\epsilon^3}C_F\left[\left(\frac{3}{2} + i\pi \right)\frac{e_u^2+e^2_d}{2}-\Gamma_t\right]\Bigg\}\amp^{(0)}\nn\\
&+ \f{1}{2\epsilon^2}\bigg\{\left(\frac{\alpha}{\pi}\right)\frac{e_u^2+e^2_d}{2}\ampF^{(1,0)} + C_F\left(\frac{\as}{\pi}\right) \ampF^{(0,1)} \nn\\
&\hspace{1.1cm}+ C_F\left(\frac{\as}{\pi}\right)\left(\frac{\alpha}{\pi}\right)\left[ \left(\f{7}{12}\pi^2 -\f{9}{8} -\f{3}{2}i\pi \right)\frac{e_u^2+e^2_d}{2} + \left(\f{3}{2}+i\pi\right)\Gamma_t\right] \amp^{(0)} \bigg\} \nn\\
&+\f{1}{2\epsilon} \bigg\{\left(\frac{\alpha}{\pi}\right)\left[ \left( \f{3}{2}+i\pi \right) \frac{e_u^2+e^2_d}{2}-2\Gamma_t \right]\ampF^{(1,0)} + \left(\frac{\as}{\pi}\right)C_F\left[ \f{3}{2}+i\pi \right]\ampF^{(0,1)}\nn\\
& \hspace{1cm}+\f{1}{8}C_F\left(\frac{\as}{\pi}\right)\left(\frac{\alpha}{\pi}\right)\left[ \left( \f{3}{2} -\pi^2 +24\zeta(3)+\f{2}{3}i\pi^3\right)\frac{e_u^2+e^2_d}{2} -\f{2}{3}\pi^2\Gamma_t \right]\amp^{(0)} \bigg\}\,.
\end{align}
In Eqs.~(\ref{eq:M01}) and (\ref{eq:M11}) the function $\Gamma_t$ is the NLO EW soft anomalous dimension, which is
an operator acting on the charge indices of the initial-state quarks and the final-state lepton. Its explicit expression can be derived from the corresponding NLO QCD expression~\cite{Catani:2014qha} through
an abelianisation procedure~\cite{Buonocore:2019puv}. For the relevant Born level process
\begin{equation}
u(p_1)\bar{d}(p_2)\to \ell^+(p_3) \nu_{\ell}(p_4)
\end{equation}
$\Gamma_t$ takes the form
\begin{equation}
\Gamma_t= -\f{1}{4}\bigg\{ e^2_\ell (1-i\pi) + \sum_{i=1,2}e_ie_3\ln\f{(2p_i \cdot p_3)^2}{Q^2m_\ell^2}\bigg\}\,,
\end{equation}
where $Q^2=(p_1+p_2)^2$, $m_\ell$ is the mass of the charged lepton and the electric charges are given by $e_1=e_u=2/3$, $e_2=e_{\bar d}=1/3$, $e_3=-e_{\ell^+}=-1$.
Since our calculation of the coefficient $H^{(1,1)}$ will be carried out by using the PA, in Sec.~\ref{sec:numPA} we are going to validate this approximation by studying its equivalent $H^{(0,1)}$ at ${\cal O}(\alpha)$.
We therefore define two different approximations of the $H^{(0,1)}$ coefficient,
\begin{align}
\label{eq:H01PA}
H^{(0,1)}_{\rm PA}&=\frac{2{\rm Re}\left({\cal M}_{\rm fin}^{(0,1)}{\cal M}^{(0,0)*}\right)_{\rm PA}}{|{\cal M}^{(0,0)}|^2}\,,\\
\label{eq:H01PArwt}
H^{(0,1)}_{\rm PA,rwg}&=\frac{2{\rm Re}\left({\cal M}_{\rm fin}^{(0,1)}{\cal M}^{(0,0)*}\right)_{\rm PA}}{|{\cal M}^{(0,0)}_{\rm PA}|^2}\, .
\end{align}
In the pure PA of Eq.~(\ref{eq:H01PA})
the $ H^{(0,1)}_{\rm PA}$ coefficient is simply computed by evaluating the interference of the tree-level and the one-loop amplitude in the PA. Since the $H^{(0,1)}$ coefficient is eventually multiplied with
the Born cross section $d\sigma_{\rm LO}$ (see the first term on the right-hand side of Eq.~(\ref{eq:master})), this procedure corresponds to the standard PA.
An alternative procedure is to evaluate also the denominator in Eq.~(\ref{eq:H01def}) in the PA. This approach effectively reweights the virtual--tree interference in PA with the full squared Born amplitude, and is thus labelled ${\rm PA,rwg}$ in Eq.~(\ref{eq:H01PArwt}).
We now turn to ${\cal O}(\as\alpha)$. Since, as we will show, the PA with the reweighting procedure works rather
well for $H^{(0,1)}$, we define two approximations for the $H^{(1,1)}$ coefficient
involving different reweighting factors,
\begin{align}
\label{eq:H11PA1}
H^{(1,1)}_{{\rm PA,rwg_B}}&=H^{(1,1)}_{{\rm PA}}\times \frac{|{\cal M}^{(0,0)}|^2}{|{\cal M}^{(0,0)}_{\rm PA}|^2}=
\frac{2{\rm Re}\left({\cal M}_{\rm fin}^{(1,1)}{\cal M}^{(0,0)*}\right)_{\rm PA}}{|{\cal M}^{(0,0)}_{\rm PA}|^2}\,,\\
\label{eq:H11PA2}
H^{(1,1)}_{{\rm PA,rwg_V}}&=H^{(1,1)}_{{\rm PA}} \times \frac{H^{(0,1)}}{H^{(0,1)}_{\rm PA}}=\frac{2{\rm Re}\left({\cal M}_{\rm fin}^{(1,1)}{\cal M}^{(0,0)*}\right)_{\rm PA}}{|{\cal M}^{(0,0)}|^2}\times\frac{2{\rm Re}\left({\cal M}_{\rm fin}^{(0,1)}{\cal M}^{(0,0)*}\right)_{\phantom{\rm PA}}}{2{\rm Re}\left({\cal M}_{\rm fin}^{(0,1)}{\cal M}^{(0,0)*}\right)_{\rm PA}}\,.
\end{align}
In the first approximation, Eq.~(\ref{eq:H11PA1}), the interference of the two-loop amplitude with the Born amplitude is computed in the PA, and the same is done for the denominator.
This is analogous to what is performed in Eq.~(\ref{eq:H01PArwt}) at NLO EW. In the second approximation, the expression of the $H^{(1,1)}_{\rm PA}$ coefficient is multiplied by the ratio $H^{(0,1)}/H^{(0,1)}_{\rm PA}$, which corresponds to a reweighting with the full EW-virtual--Born interference.
This reweighting procedure can be motivated by assuming a factorised approach: if we could write $H^{(1,1)}\sim H^{(1,0)}\times H^{(0,1)}$, such additional factor would improve the $H^{(0,1)}$ coefficient with the exact
one-loop EW virtual amplitude and, therefore, with the correct Sudakov EW logarithmic contributions \cite{Denner:2000jv,Accomando:2004de} at large values of transverse momenta. In the following we will use the prescription in Eq.~(\ref{eq:H11PA2}) to define our central prediction for the ${\cal O}(\as\alpha)$ correction. The difference between the results obtained with the two approximations
can be used as an estimate of the uncertainty of our procedure.
\section{Numerical results}
\label{sec:num}
Having described our calculation and defined our approximations, we can now move on to present our results.
We consider the process $pp\to \mu^+\nu_\mu+X$ at centre-of-mass energy \mbox{$\sqrt{s}=14$\,TeV}.
As for the EW couplings, we follow the setup of Ref.~\cite{Dittmaier:2015rxo}.
In particular, we use the $G_\mu$ scheme with $G_F=1.1663787\times 10^{-5}$\,GeV$^{-2}$ and set the \textit{on-shell} values of masses and widths to $m_{W,{\rm OS}}=80.385$\,GeV, $m_{Z, {\rm OS}}=91.1876$\,GeV, $\Gamma_{W, {\rm OS}}=2.085$\,GeV, $\Gamma_{Z, {\rm OS}}=2.4952$\,GeV. Following the default mass scheme in {\sc Recola}, those values are translated to the corresponding \textit{pole} values $m_{V}=m_{V,{\rm OS}}/\sqrt{1+\Gamma^2_{V,{\rm OS}}/m^2_{V{,\rm OS}}}$ and $\Gamma_{V}=\Gamma_{V,{\rm OS}}/\sqrt{1+\Gamma^2_{V,{\rm OS}}/m_{V,{\rm OS}}^2}$, $V=W,Z$, from which $\alpha=\sqrt{2}\,G_F m_{W}^2(1-m_{W}^2/m_{Z}^2)/\pi$ is derived,
and we use the complex-mass scheme \cite{Denner:2005fg} throughout.
The muon mass is fixed to $m_\mu=105.658369$\,MeV, and the pole masses of the top quark and the Higgs boson to $m_t=173.07$\,GeV and $m_H=125.9$\,GeV, respectively.
The CKM matrix is taken to be diagonal. We use the \texttt{NNPDF31$\_$nnlo$\_$as$\_$0118$\_$luxqed} set of parton distributions~\cite{Bertone:2017bme}, which is based on the LUXqed methodology~\cite{Manohar:2016nzj}
for the determination of the photon flux.
The QCD coupling $\as$ is evaluated at the corresponding (3-loop) order.
The renormalisation and factorisation scales are fixed to $\mu_R=\mu_F=m_W$.
At Born level this process proceeds through quark--antiquark annihilation. NLO QCD (EW) corrections additionally involve quark--gluon (quark--photon) channels, and the mixed ${\cal O}(\as\alpha)$ corrections additionally involve the gluon--photon channel and (anti)quark--(anti)quark interference channels. All the partonic channels are included in our calculation, although, as will be shown below,
the dominant contribution is given by the quark--antiquark and quark--gluon channels.
We use the following selection cuts,
\begin{equation}
\label{eq:cuts}
p_{T,\mu}>25\,{\rm GeV}\,,\qquad |y_\mu|<2.5\,,\qquad p_{T,\nu}>25\,{\rm GeV}\,,
\end{equation}
and work at the level of {\it bare muons}, i.e., no lepton recombination with close-by photons is carried out.
\subsection{Validation of the pole approximation}
\label{sec:numPA}
Before presenting our results, we study the quality of the PA for the $H^{(0,1)}$ coefficient.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.46\textwidth]{{plots/HVEW_pT_lep_OS}.pdf}\hfill
\includegraphics[width=0.46\textwidth]{{plots/HVEW_pT_lep_OFFS}.pdf}\\
\end{center}
\vspace{-2ex}
\caption{\label{fig:H1}
Contributions $d\sigma^{(0,1)_{\rm H}}_{\mathrm{PA}}$ and $d\sigma^{(0,1)_{\rm H}}_{\mathrm{PA,rwg}}$ of the hard coefficient $H^{(0,1)}$ in the approximations introduced in Eqs.~(\ref{eq:H01PA}) and (\ref{eq:H01PArwt}), respectively, compared with the exact result $d\sigma^{(0,1)_{\rm H}}$: the region of the Jacobian peak (left) and the tail of the distribution in the muon $p_T$ (right).
The upper panels show the absolute predictions, and the lower panels display the relative difference to $d\sigma^{(0,1)_{\rm H}}$.
}
\end{figure}
In Fig.~\ref{fig:H1} we show the contribution of the coefficient $H^{(0,1)}$ to the NLO EW correction as a function of the muon transverse momentum ($p_{T}$).
In particular, we compare the two approximations of Eqs.~(\ref{eq:H01PA}) and (\ref{eq:H01PArwt}) to the exact result.
The left panel depicts the region around the Jacobian peak, and the right panel the high-$p_T$ region.
In the region below the Jacobian peak the PA works relatively well and reproduces the exact result up to a few percent.
On the contrary, as $p_T$ increases, $H^{(1,0)}_{\rm PA}$ significantly undershoots the exact result.
However, when the PA is improved through the reweighting procedure, the $H^{(0,1)}$ coefficient is reproduced sufficiently well. The relative differences are at the percent level in the low-$p_T$ region and remain under control up to high transverse momenta. To which extent the approximation is good, depends, however, on the actual impact of the $H^{(0,1)}$ contribution to the complete NLO EW correction.
This is, of course, not an issue at ${\cal O}(\alpha)$ since $H^{(0,1)}$ is available.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.46\textwidth]{{plots/HVMIX_pT_lep_OS}.pdf}\hfill
\includegraphics[width=0.46\textwidth]{{plots/HVMIX_pT_lep_OFFS}.pdf}\\
\end{center}
\vspace{-2ex}
\caption{\label{fig:H11} Impact of the $\mathcal{O}(\as \alpha)$ hard-virtual contribution on the mixed QCD--EW corrections, computed adopting the reweighting procedures introduced in Eqs.~(\ref{eq:H11PA1}) and (\ref{eq:H11PA2}), respectively: the region of the Jacobian peak (left) and the tail of the distribution in the muon $p_T$ (right).
The upper panels show the absolute predictions, with the contributions $d\sigma^{(1,1)_{\rm H}}_{\mathrm{PA,rwg_B}}$ and $d\sigma^{(1,1)_{\rm H}}_{\mathrm{PA,rwg_V}}$ multiplied by a factor of 10 (100) in the peak (tail) region.
The lower panels show the effect of $d\sigma^{(1,1)_{\rm H}}_{\mathrm{PA,rwg_{B/V}}}$ normalized to $d\sigma^{(1,1)}$.}
\end{figure}
We now consider the $H^{(1,1)}$ coefficient. In Fig.~\ref{fig:H11} we show our result for the ${\cal O}(\as\alpha)$ correction, obtained by using the approximation in Eq.~(\ref{eq:H11PA2}), and the contribution
of the $H^{(1,1)}$ coefficient in the two approximations.
The lower panel displays the relative impact of the $H^{(1,1)}$ contribution.
We first discuss the low-$p_T$ region. Here the relative impact of the $H^{(1,1)}$ contribution can be quite large, being of ${\cal O}(-30\%)$ for very low $p_T$ values
and reaching ${\cal O}(+30\%)$ in the regions $p_T\sim 38$--$39$\,GeV and $p_T\sim 40$--$41$\,GeV. However, these are the $p_T$ regions where the ${\cal O}(\as\alpha)$ correction is either very small or changes sign.
On the other hand, given our observations at ${\cal O}(\alpha)$, the reweighted pole approximation is expected to work well in these regions, and the two approximations defined above are indeed in good agreement.
The situation is very different in the high-$p_T$ region. Here the impact of the $H^{(1,1)}$ coefficient is smaller than $1\%$ of the entire ${\cal O}(\as\alpha)$ correction.
We also observe that the contribution of $H^{(1,1)}$ becomes even smaller as $p_T$ increases, being
at the permille level at very large $p_T$ values.
This is not unexpected since at large $p_T$ the resonant Born-like topologies are suppressed and the cross section is dominated by real contributions, where an on-shell $W$ boson is accompanied by
hard QCD or QED emissions. More precisely, we find that in this region the dominant contribution is given by the quark--gluon partonic channel, which is computed exactly.
Therefore, even if our approximations in Eqs.(\ref{eq:H11PA1}) and (\ref{eq:H11PA2}) should fail, say, by a factor of two in this region,
the impact on the computed correction would be negligible.
Since all the other contributions in Eq.~(\ref{eq:master}) are treated exactly,
we conclude that our calculation of the complete ${\cal O}(\as\alpha)$ correction can be considered reliable both at small and large values of $p_T$.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.46\textwidth]{{plots/MIX_pT_lep_OS}.pdf}\hfill
\includegraphics[width=0.46\textwidth]{{plots/MIX_pT_lep_OFFS}.pdf}\\
\end{center}
\vspace{-2ex}
\caption{\label{fig:corr}
Complete ${\cal O}(\as\alpha)$ correction to the differential cross section $d\sigma^{(1,1)}$ in the muon $p_T$, and its factorized approximation $d\sigma^{(1,1)}_{\rm fact}$, defined in Eq.~(\ref{eq:fact}).
The top panels show the absolute predictions, while the central (bottom) panels display the ${\cal O}(\as\alpha)$ correction normalized to the LO (NLO QCD) result.
}
\end{figure}
\subsection[Results on the ${\cal O}(\as\alpha)$ correction]{Results on the $\boldsymbol{{\cal O}(\as\alpha)}$ correction}
\label{sec:numres}
We now turn to our final result for the complete ${\cal O}(\as\alpha)$ correction in Fig.~\ref{fig:corr}.
In the main panels we show the absolute correction $d\sigma^{(1,1)}/dp_T$ as a function of the muon $p_T$.
The central (bottom) panels display the correction normalised to the LO (NLO QCD) result.
As discussed above, the central value for $d\sigma^{(1,1)}$ is obtained by computing the hard coefficient $H^{(1,1)}$ as in Eq.~(\ref{eq:H11PA2}),
and the difference with the result obtained by using the prescription in Eq.~(\ref{eq:H11PA1})
is taken as an uncertainty. However, such uncertainty is so small that it is not resolved on the scale of these plots.
Our results can be compared with those from an approach in which QCD and EW corrections are assumed to completely factorise.
Such approximation can be defined as follows:
For each bin, the QCD correction, $d\sigma^{(1,0)}/dp_T$, and the EW correction restricted to the $q{\bar q}$ channel,
$d\sigma_{q{\bar q}}^{(0,1)}/dp_T$, are computed, and the factorised ${\cal O}(\as\alpha)$ correction is calculated as
\begin{equation}
\label{eq:fact}
\frac{d\sigma^{(1,1)}_{\rm fact}}{dp_T}=\left(\frac{d\sigma^{(1,0)}}{dp_T}\right)\times\left(\frac{d\sigma_{q{\bar q}}^{(0,1)}}{dp_T}\right)\times\left(\frac{d\sigma_{\rm LO}}{dp_T}\right)^{-1}\, .
\end{equation}
This definition and, in particular, the exclusion of the photon-induced channels in the EW correction $d\sigma^{(0,1)}$ deserve some comments \cite{Kallweit:2019zez}. A factorised approach of mixed QCD--EW corrections is justified if the dominant sources of QCD and EW corrections factorise with respect to the hard $W$ production subprocess.
In phase space regions populated by the emission of a hard additional photon (jet), a factorised approach including the photon-induced contribution would effectively multiply {\it giant K-factors} \cite{Rubin:2010xp} of QCD and EW origin, and, therefore, is not expected to work.
\begin{figure}[t!]
\centering
\includegraphics[width=0.46\textwidth]{plots/pT-lep-OS-extrapolation-error-zoom.pdf}\hfill
\includegraphics[width=0.46\textwidth]{plots/pT-lep-OFFS-extrapolation-error.pdf}
\caption{Relative uncertainty of the ${\cal O}(\as\alpha)$ coefficient as a function of the muon $p_T$. Combined statistical and systematic uncertainties are shown, and the latter is associated to the $q_T\to0$ extrapolation procedure described in the main text, performed on a binwise level. In the grey-shaded regions, $d\sigma^{(1,1)}/dp_T$ is approximately zero and its relative error thus meaningless
}\vspace*{1ex}
\label{fig:syst}
\end{figure}
Our result for the ${\cal O}(\as\alpha)$ correction in the region of the Jacobian peak is reproduced relatively well by the factorised approximation.
Beyond the Jacobian peak, the factorised result tends to overshoot the complete result.
This is in line with what was observed in Ref.~\cite{Dittmaier:2015rxo}. As $p_T$ increases, the (negative) impact of the mixed QCD--EW corrections increases, and at $p_T=500$\,GeV it reaches about $-140\%$ with respect to the LO prediction and $-20\%$ with respect to the NLO QCD result.
The factorised approximation describes the qualitative behaviour of the complete correction relatively well, also in the tail of the distribution, but it tends to overshoot the full result as $p_T$ increases.
We add some details on the procedure used to obtain these results.
In order to numerically evaluate the contribution in the square bracket of Eq.~(\ref{eq:master}), a technical cut-off $r_{\rm cut}$ is introduced on the dimensionless variable $q_T / M$, where $M$ is the invariant mass of the lepton--neutrino system. The final result in each bin, which corresponds to the limit $r_{\rm cut} \to 0$, is extracted by computing $d\sigma^{(1,1)}/dp_T$ at fixed values of $r_{\rm cut}$ in the range $[0.01\%, r_{\rm max}]$. Quadratic least $\chi^2$ fits are performed for different values of $r_{\rm max}\in[0.5\%, 1\%]$. The extrapolated value is then extracted from the fit with lowest $\chi^2/$degrees-of-freedom, and the uncertainty is estimated by comparing the results obtained through the different fits. This procedure is the same as implemented in {\sc Matrix}~\cite{Grazzini:2017mhc}. The ensuing uncertainties of the computed correction (not shown in Fig.~\ref{fig:corr}), obtained combining statistical and systematic errors, are shown in Fig.~\ref{fig:syst}.
They range from the percent level at low $p_T$ values to ${\cal O}(3\%)$ at $p_T=500$\,GeV, with the exception of regions where $d\sigma^{(1,1)}/dp_T$
is approximately zero and thus the relative errors are artificially large. We have checked, however, that in these regions the error
is well below one permille of the respective cross section and thus phenomenologically irrelevant.
We finally present our predictions for the fiducial cross section corresponding to the selection cuts in Eq.~(\ref{eq:cuts}). In Table~\ref{tab:fid} we report
the contributions $\sigma^{(i,j)}$ to the cross section (see Eq.~(\ref{eq:exp})) in the various partonic channels. The numerical uncertainties are stated in brackets, and for the NNLO corrections
$\sigma^{(2,0)}$ and the mixed QCD--EW contributions $\sigma^{(1,1)}$ they include the systematic uncertainties from the $r_{\rm cut}\to 0$ extrapolation.
The contribution from the channels $u{\bar d},\,c{\bar s}$ is denoted by $q{\bar q}$.
\begin{table}
\renewcommand{\arraystretch}{1.4}
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
$\sigma$ [pb] & $\sigma_{\rm LO}$ & $\sigma^{(1,0)}$ & $\sigma^{(0,1)}$ & $\sigma^{(2,0)}$ & $\sigma^{(1,1)}$ \\
\hline
\hline
$q{\bar q}$ & $5029.2$ & $\phantom{+0}970.5(3)\phantom{00}$ & $-143.61(15)$ & $\phantom{+}251(4)\phantom{.0}$ & $\hspace*{-0.6ex}\phantom{0}-7.0(1.2)\phantom{00}\hspace*{-0.6ex}$\\
\hline
$qg$ & --- & $-1079.86(12)$ & --- & $-377(3)\phantom{.0}$ & $\hspace*{-0.6ex}\phantom{+}39.0(4)\phantom{.00}\hspace*{-0.6ex}$ \\
\hline
$q(g)\gamma$ & --- & --- & $\phantom{+00}2.823(1)$ & --- & $\hspace*{-0.6ex}\phantom{+0}0.055(5)\phantom{.0}\hspace*{-0.6ex}$\\
\hline
$q({\bar q})q^\prime$ & --- & --- & --- & $\phantom{+0}44.2(7)$ & $\hspace*{-0.6ex}\phantom{+0}1.2382(3)\phantom{.}\hspace*{-0.6ex}$\\
\hline
$g g$ & --- & --- & --- & $\phantom{+}100.8(8)$ & --- \\
\hline
\hline
tot & $5029.2$ & $\phantom{0}$$-109.4(4)\phantom{00}$ & $-140.8(2)\phantom{00}$ & $\phantom{00}19(5)\phantom{.0}$ & $\hspace*{-0.6ex}\phantom{+} 33.3(1.3)\phantom{00}\hspace*{-0.6ex}$ \\
\hline
\end{tabular}
\caption{\label{tab:fid}
The different perturbative contributions to the fiducial cross section (see Eq.~(\ref{eq:exp})). The breakdown into the various partonic channels is also shown.}
\renewcommand{\arraystretch}{1.0}
\end{table}
The contributions from the channels
$qg,\,{\bar q}g$ and $q\gamma,\,{\bar q}\gamma$, which enter at NLO QCD and EW, are labelled by $qg$ and $q\gamma$, respectively.
The contribution from all the remaining quark--quark channels
$qq',\, {\bar q}{\bar q}',\, q{\bar q}'$ (excluding $u{\bar d},\,c{\bar s}$) to the NNLO QCD and mixed corrections
is labelled by $q({\bar q})q^\prime$.
Finally, the contributions from the gluon--gluon and gluon--photon channels, which are relevant only at ${\cal O}(\as^2)$ and ${\cal O}(\as\alpha)$, are denoted by $gg$ and $g\gamma$, respectively.
We see that NLO QCD corrections are subject to large cancellations between the $q{\bar q}$ and the $qg$ channels.
As a consequence, NLO QCD and NLO EW corrections have a similar quantitative impact: the NLO QCD corrections amount to $-2.2\%$ with respect
to the LO result, while NLO EW corrections contribute $-2.8\%$. Because of similar cancellations, the NNLO QCD corrections are rather small and amount to $+0.4\%$.
The newly computed mixed QCD--EW corrections amount to $+0.6\%$ with respect to LO, and are dominated by the $qg$ channel, whereas the
photon-induced $q(g)\gamma$ channels give a negligible contribution.
As a final remark we note that the pattern of the higher-order QCD corrections to the perturbative series is strongly dependent on the choice of the renormalisation and factorisation scales. For example, the scale choice $\mu_R=\mu_F={m_W}/{2}$ leads to a more common perturbative pattern: $\sigma^{(1,0)}/\sigma_{\rm LO} = +10\%$, $\sigma^{(0,1)}/\sigma_{\rm LO} = -2.9\%$, $\sigma^{(2,0)}/\sigma_{\rm LO} =+4.2\%$, $\sigma^{(1,1)}/\sigma_{\rm LO} = +0.76\%$.
\section{Summary}
\label{sec:summa}
In this paper we have presented a new computation of the
mixed QCD--EW corrections to charged-lepton production at the LHC. The cancellation of soft and collinear singularities has been achieved by using a
formulation of the $q_T$ subtraction formalism derived from the NNLO QCD calculation for heavy-quark production through a suitable abelianisation procedure.
All the real and virtual contributions
due to initial- and final-state radiation
have been consistently included without any approximation, except for the finite part
of the two-loop virtual correction, which has been computed in the pole approximation
and suitably improved through a reweighting procedure. We have shown that our results are reliable in both on-shell and off-shell regions, thereby providing the
first prediction of the mixed QCD--EW corrections in the entire region of the lepton
transverse momentum.
Our calculation is fully differential in the momenta of the charged lepton, the corresponding neutrino, and the associated QED and QCD radiation.
Therefore, it can be used to compute arbitrary infrared-safe observables, and, in particular, we can also deal with {\it dressed leptons}, i.e.\ leptons recombined with close-by photons.
The method we have used is in principle applicable to other EW processes, as, for example, vector-boson pair production. More details on our calculation will be presented elsewhere.
\vskip 0.5cm
\noindent {\bf Acknowledgements}
\noindent
We would like to express our gratitude to Jean-Nicolas Lang and Jonas Lindert for their
continuous support on {\sc Recola} and {\sc OpenLoops}.
We wish to thank Stefano Catani and Alessandro Vicini for helpful discussions and comments on the manuscript.
This work is supported in part by the Swiss National Science Foundation (SNF) under contracts IZSAZ2$\_$173357 and 200020$\_$188464. The work of SK is supported by the ERC Starting Grant 714788 REINVENT. FT acknowledges support from INFN.
|
\section{Introduction}
\label{sec:introduction}
There has been immense research on developing technological solutions in order to address the energy requirements which are rising exponentially and thereby making the challenge of energy conservation harder day by day. The increasing energy demands not only affects a country's economy but also comes with significant negative implications on the environment. Hence, the only effective way to conserve energy now is to encourage its efficient usage. A fine-grained monitoring of energy consumption and communicating the same to the consumers can help in the noteworthy reduction of energy wastage ~\cite{shorrock2003domestic}, ~\cite{darby2006effectiveness}. The method for the disaggregation of electrical measurements was proposed by Hart \cite{192069} at the beginning of the 1980s and owing to the shift towards smart meters, ease of data availability in the last few years and amidst rising interests in climate change led to the boost of extensive research in this field. Energy disaggregation primarily aims to provide a piece of detailed information about energy sensing and a breakdown of the consumption patterns, which can further be analysed for the automation of energy management and the development of similar systems. Strategies such as rescheduling the appliances with high-power consumption rates to off-peak times will thus be feasible \cite{zoha2012non}.
It would also help appliance design companies acquire a more reliable understanding of the power-usage patterns of their devices and hence will enable them to develop sustainable products.
There are two primary approaches to energy disaggregation, in particular, Intrusive Load Monitoring (ILM) and Non-Intrusive Load Monitoring (NILM). Intrusive load monitoring consists of using a low-end metering device for the measurement of the electricity consumption of one or more appliances, usually employing at least one censoring device per appliance, whereas in the case of NILM it only requires a single meter per house or a building that is being monitored. Non-intrusive load monitoring (NILM) or energy disaggregation is aimed at separating the household energy measured at the aggregate level into constituent appliances. The presence of nilmtk-contrib ~\cite{10.1145/3360322.3360844}, an open-source, implementation of the energy disaggregation problem, has unfolded the means for comparisons of the different algorithms executing energy disaggregation. It has also enabled researchers to observe and evaluate NILM approaches for abstraction and varied evocation as they can be applied to multiple data sets accessible online. The versatility of NILM API makes experimentation in this field easy by lowering the entry barriers and making the implementation generic irrespective of the datsets and algorithms which helped to progress research in this area.
\begin{table*}[]
\footnotesize
\caption{Articles included in systematic literature review}
\begin{tabular}{|p{4cm}|p{4cm}|p{0.6cm}|p{0.6cm}|p{0.6cm}|p{0.6cm}|p{0.6cm}|p{0.6cm}|p{0.6cm}|}
\hline
Subject &
Purpose &
Dia-logue on framework &
Dis-cussion on NILM-API &
Related dataset references &
Algo-rithm implementation &
NILM-API implementation &
Empir-ical analysis &
Future research prosp-ects \\ \hline
Non-Intrusive Load Monitoring System Framework and Load Disaggregation Algorithms: A Survey\cite{8861646} &
presents a general NILM framework and reviews publicly available data sets. &
Yes &
Yes &
Yes &
No &
No &
No &
Yes \\ \hline
A Survey on Non-Intrusive Load Monitoring Methodies and Techniques for Energy Disaggregation Problem\cite{faustine2017survey} &
presents an overview of NILM, related methods and reviews state-of-the-art NILM algorithms.&
Yes &
Yes &
Yes &
No &
No &
No &
No \\ \hline
Building power consumption datasets: Survey, taxonomy and future directions\cite{HIMEUR2020110404} &
analyses the nature of building energy consumption datasets. &
Yes &
No &
Yes &
No &
No &
No &
Yes \\ \hline
On performance evaluation and machine learning approaches in non-intrusive load monitoring\cite{klemenjak2018performance} &
determines the accuracy and generalisation abilities of NILM algorithms using the data sets REDD, UK-DALE, and Dataport. &
Yes &
Yes &
Yes &
No &
No &
No &
Yes \\ \hline
Prospects of Appliance Level Load Monitoring in Off-the-Shelf Energy Monitors: A Technical Review\cite{haq2018prospects} &
encourages the incorporation of appliance-level e-monitoring and load disaggregation; sets requirements to implement load disaggregation in next-generation e-monitors. &
Yes &
No &
Yes &
No &
No &
No &
Yes \\ \hline
An Overview of Non-Intrusive Load Monitoring: Approaches, Business Applications, and Challenges\cite{zhuang2018overview} &
surveys NILM framework and load disaggregation algorithms; reviews load signature models, exhibits existing datasets and performance metrics. &
Yes &
Yes &
Yes &
Yes &
No &
Yes &
Yes \\ \hline
Literature Review of Power Disaggregation\cite{jiang2011literature} &
reviews the current state of the algorithms and
NILM systems. &
Yes &
No &
No &
No &
No &
No &
Yes \\ \hline
Load Disaggregation Technologies: Real World and Laboratory
Performance\cite{mayhorn2016load} &
reviews recent field studies and laboratory tests of NILM technologies. &
Yes &
No &
No &
Yes &
No &
Yes &
Yes \\ \hline
\textbf{Proposed~Comprehensive~review on energy disaggregation}&
\textbf{reviews~of~the state-of-the-art~techniques, implements NILMTK, presents analysis and draws conclusions}&
\textbf{Yes} &
\textbf{Yes} &
\textbf{Yes} &
\textbf{Yes} &
\textbf{Yes} &
\textbf{Yes} &
\textbf{Yes} \\ \hline
\end{tabular}
\label{Table: LR}
\end{table*}
\section{Scope and contribution of this paper}
There has been a good number of research and publications in energy disaggregation, non-intrusive load monitoring, home energy management and appliance classification. Recent discoveries in machine learning and computer vision has driven a lot of attraction to machine learning-based strategies for NILM. As it develops, energy consumption datasets serve as the basis for train and test data for NILM algorithms. These data sets are generated by using a piece of measurement equipment such as a smart plug and combing the energy readings given by smart-meters and appliance-level meters. The increasing popularity of research in this field has led to a large number of datasets being released in the past few years. NILMTK was designed specifically as an open-source implementation to unfold the means for comparisons and proliferation of the different algorithms executing energy disaggregation. ~\cite{10.1145/3360322.3360844}. Performance evaluation, contrast and comparison of different NILM algorithms is still a research challenge, for multiple reasons ~\cite{klemenjak2018performance} \cite{pereira2018performance} \cite{herrero2017non}. Table-\ref{Table: LR} is a tabular depiction of the literature review which integrates the information about the purpose of research and their major findings in this field, that has been compiled from similar survey papers. It can be observed that the prospects of NILM application are both, wide and versatile. There have been multiple attempts to devise the generalisation abilities of existing NILM algorithms in order to incorporate the methodology to appliance-level monitoring and disaggregation. It is also challenging to collect and store data at the required sample rates in order to make data sets available for these algorithms. This paper provides a concise and clear overview of the NILM framework and the NILMTK-API by implementing the API on three publicly available data sets and comparing the observed results.
This survey article exhibits the following concepts:
\begin{enumerate}
\item This paper aims to present an updated overview of the techniques used in the NILM framework including the recent developments in its strategies for the energy disaggregation problem.
\item The paper presents an experimental overview of the application of NILM-API, which was released with nilmtk-contrib, on three publicly available data sets, draws conclusions and highlights on future research directions.
\item A detailed overview of the energy disaggregation problem is presented. Here, we have shown the advantages of the NILM API which enables comparison of algorithms in this domain even without expertise in the field.
\end{enumerate}
\section{Organisation of this paper}
The rest of the paper is organised as follows. We start with a brief background on the load monitoring approaches which is followed by a formal introduction to the energy disaggregation problem and a discussion about the available implementations and their general framework. We then introduce the State-of-the-art Disaggregation Techniques, the NILMTK-contrib repository, and summarise the different methods that are being employed for NILM. Next, we describe the benchmark data sets, including the features taken into consideration for the experiment API. Later we, describe the advantages of this implementation by the means of an empirical comparison, before we sum up our analysis.
\section{Background on Load Monitoring approaches}
Load monitoring and identification is a mechanism for the evaluation and regulation of the usage of electrical energy and operating conditions of individual appliances, wherein the main power metre located in the house forms the basis of the analysis of the composite load measurement. They can supply the customer and the utility with information such as the type of load, the detail of the electricity consumption and the operating conditions of the appliances. The practice of load monitoring is quintessential for the management of energy as it gives us statistical insights on appliance energy consumption and their patterns which can be applied for scheduling heavy loads and optimal utilisation of energy resources.
\subsection{Intrusive load monitoring}
Intrusive load monitoring consists of using a low-end metering device for the measurement of the electricity consumption of one or more appliances. It is said to be 'intrusive' in order to indicate that the meter which is being used to employ energy monitoring is located inside the dwelling, typically close to the appliance that is monitored. In essence, an intrusive load monitoring system standardises a mechanism for measuring the energy usage of an appliance by employing a metering system that connects each appliance to power meters within the household. It requires, therefore, entering the house and is thus remarked as an intrusive system. The results obtained with intrusive load monitoring are accurate, however, it requires an elaborate installation incurring high costs which usually demands wiring and units for data storage to the households concerned ~\cite{aladesanmi2015overview}. The techniques for load monitoring can also be divided into direct or indirect monitoring techniques ~\cite{6486814}. Direct monitoring techniques are often referred to as physically intrusive procedures, measure the electrical properties of the power of each appliance requirement. The process involves a tiny low-end device that is connected to the power cord of each device, to generate physically intrusive signatures for measuring the energy consumed by the appliances. The state of operation of an appliance is signalled to the data collecting device, whenever the appliance is switched on. The electromagnetic field produced by the current flow through the wire is used to calculate the power drawn by the appliance. Although it provides an accurate measurement of energy consumption, it is not always economically feasible ~\cite{6486814}.
\subsection{Non-intrusive load monitoring}Non-intrusive load monitoring consists of measuring electricity consumption using a smart meter, which is usually placed at the meter panel.
It is also known as one-sensor metering since the whole process relies on a single point of measurement. This method qualifies to be non-intrusive meaning that no extra equipment is installed inside the house. Since this is a one-sensor metering procedure, the signatures of appliances get superposed in order to comprehend the contribution of single appliances, they're needed to be separated. This operation of separating the energy drawn by each appliance is called the disaggregation of the total electricity consumption. Non-intrusive load monitoring is a simple and reliable way to determine energy usage and therefore the status of operation of specific appliances, carrying the analysis of the aggregate load estimated by the building's main meter. The process of NILM also involves the analysis of individual appliance energy consumption, the deduction of the devices being utilised in a building as well as the changes happening within the voltage and current entering into a building ~\cite{aladesanmi2015overview}. It is considered non-intrusive since it does not involve an intrusion into the building or consumer premises when measuring the energy consumption of various appliances. Smart metres equipped with NILM technology are used by companies to survey the particular uses of electrical power in different homes. NILM is taken into account for the cost-effective alternative to intrusive monitoring techniques. The ideation behind analysing the flow of power in order to record the operating conditions of appliances in a household was coined by George W. Hart while he attempted to collect and examine residential load data as a part of a residential photo-voltaic system ~\cite{192069}. The acknowledgement of step-change in either active or reactive power out of the total load is the working principle for load monitoring. It is executed by modifying the operating state of the various customer’s appliances. A NILM is installed temporarily to analyse the characteristics of the appliances which may be used to recommend ways for reducing energy demand and costs.
\section{The Energy Disaggregation Problem}
Speeding urbanisation has led to some conventional changes in the housing structures which engrossed the need for the development of high-rise buildings all around the world. The maintenance of these buildings requires end-use appliance energy conservation in a synchronous manner. The transformation in housing patterns also came together with a mechanism for an appliance level energy break down of the building’s aggregate power consumption. The building's total energy reading can now be used for the estimation of appliance-specific power consumption which is facilitated by the use of smart meters. Nearly, two decades ago, a method for the disaggregation of electrical measurements was proposed by Hart, which required only the examination of the overall load data in order to identify signatures of the power consumption of each appliance among them. The proposed methodology did not require the installation of any equipment inside the property of the customer and is thus considered to be non-intrusive. The aggregated data for the building's energy usage can be obtained from outside the building or the residence using the main electrical panel. The separation of the whole-house building data into its principal energy-consumption constituents is the complete goal of this process. NILM algorithms have the ability to determine device-specific energy disaggregation using only the aggregated data collected from a single measurement point and thus Non-Intrusive Load Monitoring remains an attractive method for energy disaggregation.
\subsection{General Framework of NILM}
As mentioned in the nilm toolkit \cite{10.1145/3360322.3360844} \cite{10.1145/2602044.2602051}, the model on which NILM works is as follows: For an observed time series of aggregate measurements $Y = (Y_1, Y_2,..., Y_T )$, where $Y_t \in R^+$ represents the active power which is measured in Watt-hours or Watts using an electricity meter at time $t$. This signal represents the sum of the total energy consumed by all the constituent appliances present in a building. For the same scenario, the building facility is assumed to have $I$ appliances, wherein this equation $X_i = (x_{i1}, x_{i2}, ... , x_{iT} )$ where $x_{it} \in R^+$ is the energy-signal description for each appliance.
According to the disaggregation algorithm, the main readings of the buildings can be expressed as the sum total of the appliance signals along with $E_t$ where $E_t$ is an error term. Thus, when given the aggregate measurement of $Y$, the goal of the energy disaggregation problem is to retrieve the unknown signals $X_i$.
\begin{figure}[]
\includegraphics[width=0.75\columnwidth,keepaspectratio]{energy_did.pdf}
\centering
\caption{Energy dis aggregation}
\label{Fig: EN}
\end{figure}
Figure-\ref{Fig: EN} demonstrates one such scenario where a high-rise building H contains multiple apartments which are all equipped with a smart meter (SMX, where X is the apartment number) and the building itself has a smart-meter aggregator (SM0) that is connected to the electricity service. This building can be considered as a meter-group. If the electricity consumption data for this building is collected and sampled at a given rate, then applying NILM to this meter group will give us the appliance-wise energy consumption patterns for this group.
\section{State-of-the-art Dis aggregation Techniques}
The method for the disaggregation of electrical measurements was proposed by Hart \cite{192069} at the beginning of the 1980s. Owing to the efforts towards energy conservation and emission reduction, there has been extensive research in this field which also got boosted due to the rise in deployment of smart-meters across the globe. There are more than 10 publicly available data sets from across different geographies to support significant research activities. The growing interest in this field has led researchers to try and implement various algorithms for solving the energy disaggregation problem, involving neural networks, big data, soft-computing and statistical methods as shown in Figure-\ref{Fig: TAX}.
\begin{figure}[]
\centering
\includegraphics[width=0.6\columnwidth,keepaspectratio]{nilm_taxo_crop.pdf}
\centering
\caption{State-of-the-art Energy dis aggregation techniques}
\label{Fig: TAX}
\end{figure}
Probabilistic procedures have been gaining popularity for energy consumption modelling using a hidden Markov model (HMM) ~\cite{kolter2011redd}, ~\cite{parson2012models}, ~\cite{zhong2015latent}, ~\cite{zhong2014signal}. The problem has been approached by implementing both supervised and unsupervised methodologies including some, where signal processing methods were adopted to enable the use of appliance's features for the purpose of disaggregation ~\cite{7539273}, ~\cite{jin2011time}. Soft-computing techniques have also been utilized for the purpose of solving NILM by employing fuzzy clustering ~\cite{puente2020non}, ~\cite{lin2011applications} and ~\cite{shenavar2007novel}. Additionally, factorization procedures that aim to leverage the low-rank composition of energy consumption, also form a set of algorithms that are used to perform energy disaggregation ~\cite{batra2018transferring},\cite{batra2017matrix}, and \cite{kolter2010energy}.
Late in the year-2019, the open-source API for non-intrusive load monitoring was released as a toolkit (NILMTK). This was aimed at facilitating the means for comparisons of the different algorithms executing energy disaggregation and was intended to serve as their source library, the data set parsers needed to implement NILM, and as a reference benchmark for algorithm implementations.
The nilmtk-contrib which is the current code repository that contains the API documentation and usage examples ~\cite{batra2019towards}, lessens the amount of entry-level expertise required for algorithm developers and paves a way for algorithmic comparisons. It adds several functional enhancements to the API and eases its installation process as the disaggregation API gets rewritten.
\subsection{Neural Networks}
Neural networks have been successfully applied to a number of load scenarios and have achieved better scores in terms of accuracy as well as producing generalization for unseen houses ~\cite{kelly2015neural}, ~\cite{he2016empirical}. The most successfully employed neural network algorithms are Recurrent neural networks and Long Short Term Memory. There also exists an RNN based approach for NILM on small power office equipment ~\cite{9255127}.
\subsubsection{DAE}
Autoencoders are made of simplistic neural structures and share similarities with principal component analysis. If an autoencoder is composed of a linear activation within each of its layers, it is highly likely that the latent variables present at the bottleneck of the encoder network (the smallest layer in the network) might directly resemble to the principal components from PCA. They basically are a set of unsupervised machine learning concepts that result in reducing the inherent dimensionality of the projected data while preserving its essential features and enable the data from a higher dimension to be synthesized in a lower dimension using non-linear transformation. It comprises of two functions, specifically, encoder and decoder. The process of data compression and down-sampling the input-data into a fewer number of bits and mapping it back to the latent space is performed by the encoder. The decoder function then puts the input encoding to use, reconstructing the input and mapping it to the latent space. The latent space is present at the bottleneck. A special deep neural network design implemented to extract a particular input component from the received noisy data is termed a denoising autoencoder. As the name suggests, it subtracts the noise from a given input in order to extract meaningful data. It forces the network to undergo overfitting on the random noise by adding some amount of white noise to the input. The error incurred is then compared to produce the required output. Removal of grain from input images and the reverb of voice signals are some popular applications of DAE. For its usage in NILM, the mains signal has been treated as a noisy appliance-power signal and the mains signal is considered to be the summation of the total power consumed by the target appliance and some additional noise. Since DAE denoises one appliance at a time, it needs multiple trained models for a group of appliances to be disaggregated. Moreover, in this scenario, the DAE gets as input a window of the mains readings (length of time-window remains fixed) and produces the induced appliance consumption values for the same time-window as output. The architecture of the network remains as was proposed in nilmtk-contrib \cite{10.1145/3360322.3360844}.
\subsubsection{RNN}
Recurrent Neural Networks are a set of algorithms commonly used for time series prediction, natural language processing, and other sequence processing queries. It tackles the input elements one by one and retains all the previously encountered states. The algorithm works by saving the output of a state and feeding it back to the input layer which carries it for later predictions. Therefore, every element of the model takes into consideration not only the current inputs but also what it remembers from the preceding elements. Retaining the memory of previous states gives it the ability to learn a long-term dependency by processing a series of events and manipulate the entire context at once in order to make a prediction.
An RNN is similar to a chain-like recurrent module of a neural network but also carries a layer of memory cells. Holding onto information throughout the network leaves the RNNs with a few disadvantages, as they bring in the problem of vanishing gradients and exploding gradients. Since the gradients are the basic carriers of information in the RNN, they can sometimes either accumulate excessive information which may even contain error, or can become extremely small making the parameter updates insignificant. A very high parameter update to the neural model might weigh it down unnecessarily and remarkably small ones might make it ineffective. The resolution of this challenge is accomplished by using memory cells which can both, save a state and keep a carry to minimize information loss during sequence processing. These cells are LSTMs (Long Short Term Memory). It also has a chain-like structure but with multiple communicating neural network layers which decide the amount of data that is required to be retained, the significance of the data to be remembered and the part of the memory cell that impacts the output at the given time-step. Moreover, RNNs work well for sequential data as in the case of NILM, because they allow neuron connections in the same layer of the neural network. The employed RNN model is fed a sequence of mains readings as input and produces a single value output gives is the power consumed by the appliance at test. The network also utilizes some units of long short-term memory (LSTM) and stores values in the built-in memory cells to overcome the problem of vanishing gradients. The network architecture remains as was proposed in nilmtk-contrib \cite{10.1145/3360322.3360844}.
\subsubsection{Sequence-to-Sequence}
The sequence to sequence learning model is a deep learning concept that is used to convert from one sequence to another. It contains an encoder RNN to understand the input sequence and a decoder RNN to decode the thought vector thereby constructing an output sequence. The process of condensing the input into a vector is achieved by an encoder network which passes its output to the decoder network where the vector is unfolded to form a new sequence. There's a set of RNNs that operate commonly between the encoder and the decoder in a sequence-to-sequence model. At each step of recurrence in the encoder network, a new word is fed to the input which gets utilized in the subsequent step by the next state. Once the decoder receives the final state from the encoder, applies a discrete probability distribution (to the input at each step) to predict the output taking into consideration, a loss function. The model here tries to map the sequence of mains reading to the target appliance sequence by learning a regressive map between them. The seq2seq model uses a regression expression which is defined by $x_t:t+W−1 = f (Y_t:t+W−1, Q)+ E_t$ where $f$ is a neural network. The architecture of the network and other hyperparameters remain as they were proposed in nilmtk-contrib \cite{10.1145/3360322.3360844}.
\subsubsection{Sequence-to-Point}
Sequence to point learning (seq2point) operates by modifying the received network input to work as a mains window, while the output for the target appliance shows up functioning at the midpoint of its window.
The model believes that the midpoint of the target appliance is meant to be correlated with the received mains signal information, both before and after the point of time when it occurs. This training technique can also be considered to be a non-linear regression. The architecture of the network and other hyperparameters remain as they were proposed in nilmtk-contrib \cite{10.1145/3360322.3360844}.
\subsubsection{OnlineGRU}
Gated Recurrent Unit (GRU) is a new generation of Neural Networks that replace the LSTM units with lightweight Gated Recurrent Units (GRU) thus lessening the computational requirements while delivering the equivalent performance. They also attempt to reduce redundancy by optimising the recurrent layer sizes, as well as minimising the risk of a vanishing gradient. GRUs are equipped with an internal mechanism called gates which are capable of regulating the flow of information and provide an effective solution to short-term memory. These gates work by passing information on to the long chain of sequences as they learn which part of the data is important. GRUs are similar to LSTMs except that they do not have a cell state and instead, use the hidden state for information transfer. They usually contain two gates, one for reset and update each. The reset gate determines how much of the past knowledge to forget whereas the update gate decides what part of the information to be thrown away and what new information to be added to the unit. If the reset gate approaches zero, the hidden state is forced to disregard its state and gets reset with the current input. This permits the hidden state to discard any information that's found to be insignificant within the future. This result permits a more compact representation. Whereas the upgrade gate controls how much information from the past hidden state will be exchanged to the current hidden state. The actuation of the GRU at a specific time may be a straight addition between the past actuation and the candidate activation, where an updated door chooses how much the unit overhauls its activation or content. The Online GRU model for NILM, receives as input, the latest available mains readings of the test appliance $Y_t:t+W −1$ and outputs the calculated power consumption $x_j(t+W −1)$ of the appliance at test $j$, for the last point in time. The architecture of the network and other hyperparameters remain as they were proposed in nilmtk-contrib \cite{10.1145/3360322.3360844}.
\subsection{Big Data}
A good deal of big-data approaches has popped for energy disaggregation with the application of data analytics in smart meters ~\cite{zhang2018big} such as Neighbourhood NILM which works on the intuition that ‘similar’ homes have ‘similar’ energy consumption on a per-appliance basis ~\cite{batra2015neighbourhood} and a three-level learning framework that would provide different levels of knowledge abstractions for smart cities. It is scalable and also suits the nature of data generated by smart cities which can become hierarchical ~\cite{mohammadi2018enabling}.
\subsection{Soft Computing}
Non-intrusive load monitoring for residential electrical consumption has also seen success in feature extraction and pattern recognition using fuzzy logic. A set of soft-computing routines are being applied to learn and identify the power usage patterns of different devices from aggregated consumption records ~\cite{puente2020non}. A fuzzy classifier has been proposed which can recognise the distinctive appliance-energy states, such as energizing and de-energizing with the use of Fuzzy C-Means (FCM) clustering and optimization ~\cite{lin2011applications}.
\subsection{Bayesian Procedures}
Bayesian algorithms are being used for event detection and classifying the load data into the individual contribution of appliances. Findings for event detection method based on cepstrum smoothing ~\cite{de2017bayesian} and LSTM models ~\cite{jones2020stop} along
with a set of modules to address the issues of multi-dimensionality when the number of appliances increase ~\cite{kaselimi2019bayesian} have been devised.
\section{Data sets and Algorithms}
The following section elaborates on the data sets that have been used for the purpose of this research experiment and the algorithms that were implemented using the NILM-TK API. The API was implemented using five neural network algorithms on three publicly available data sets.
\begin{figure*}[!htbp]
\includegraphics[width=\textwidth,height=\textheight,keepaspectratio]{iawe_crop.pdf}\hfill
\caption{Predicted energy consumption: IAWE}
\label{Fig: IAWE}
\end{figure*}
\subsection{Data sets Considered}
This analysis uses three data sets, namely REDD, UK-DALE and IAWE. The data sets used for this experiment have been taken from open-source locations.
\subsubsection{REDD}The Reference Energy Disaggregation Data Set (REDD), is a freely available data set containing detailed power usage information from several homes, which has been published with the aim of promoting further research on energy disaggregation. REDD was the first public energy dataset that was released by MIT in 2011. REDD contains high and low-frequency energy readings from 6 households in the USA recorded for short period (between a few weeks and a few months). This data set is widely used for the evaluation of NILM algorithms.
\subsubsection{UK-DALE} UK-DALE is an open-access data set from the UK, that records Domestic Appliance-Level Electricity at a sample rate of $16 kHz$ for the whole-house and at $1/6 Hz$ for individual appliances. The data set contains 16 kHz current and voltage aggregate meter readings and 6 second sub-metered power data from individual appliances across 3 UK homes, as well as 1 second aggregate and 6 second sub-metered power data for 2 additional homes. An update to this data set was released in August 2015 which has expanded the data available for house 1 to 2.5 years. The updated data set has been utilized for this experiment.
\begin{table*}[!htbp]
\centering
\caption{MAE against Ground Truth for Iawe}
\begin{tabular}{cccccc}
\hline
\textbf{MAE- Iawe} & \textbf{WindowGRU} & \textbf{RNN} & \textbf{DAE} & \textbf{Seq2Point} & \textbf{Seq2Seq} \\ \hline
washing machine & 66.726196 & 65.665489 & 65.275597 & 65.736549 & 65.298805 \\ \hline
fridge & 53.850956 & 68.835884 & 51.917274 & 36.775997 & 54.116047 \\ \hline
computer & 24.112394 & 24.148933 & 24.46452 & 24.64801 & 24.25812 \\ \hline
air conditioner & 169.634262 & 139.516449 & 139.706543 & 120.483185 & 141.557159 \\ \hline
television & 1.347255 & 1.353503 & 1.347071 & 1.475708 & 1.449378 \\ \hline
\label{Table: IAWE}
\end{tabular}
\end{table*}
\begin{table*}[!htbp]
\centering
\caption{MAE against Ground Truth for REDD}
\begin{tabular}{cccccc}
\hline
\textbf{MAE- Redd} & \textbf{WindowGRU} & \textbf{RNN} & \textbf{DAE} & \textbf{Seq2Point} & \textbf{Seq2Seq} \\ \hline
washing machine & 49.6175 & 60.661507 & 49.541073 & 47.514114 & 52.302326 \\ \hline
fridge & 59.622471 & 68.384491 & 63.540127 & 55.803658 & 63.520336 \\ \hline
light & 30.128523 & 32.435764 & 31.222496 & 31.245703 & 31.557386 \\ \hline
sockets & 0.870423 & 0.888707 & 0.939906 & 0.827897 & 0.883221 \\ \hline
microwave & 28.598331 & 28.599768 & 34.998123 & 21.999201 & 24.803219 \\ \hline
\label{Table: REDD}
\end{tabular}
\end{table*}
\begin{table*}[!htbp]
\centering
\caption{MAE against Ground Truth for UKdale}
\begin{tabular}{cccccc}
\hline
\textbf{MAE-UKdale} & \textbf{WindowGRU} & \textbf{RNN} & \textbf{DAE} & \textbf{Seq2Point} & \textbf{Seq2Seq} \\ \hline
washing machine & 50.239201 & 55.136917 & 53.854832 & 49.03191 & 43.448345 \\ \hline
fridge & 44.169342 & 44.513477 & 45.257683 & 44.960304 & 44.8424 \\ \hline
light & 33.446453 & 34.21154 & 35.531647 & 33.719086 & 35.57534 \\ \hline
dish washer & 49.290272 & 47.874489 & 39.505829 & 49.656605 & 41.864983 \\ \hline
microwave & 21.169703 & 16.606073 & 25.805212 & 20.247181 & 19.278296 \\ \hline
\label{Table: UKdale}
\end{tabular}
\end{table*}
\subsubsection{IAWE} The IAWE data set was released by the Indraprastha Institute of Information Technology, India. This data contains the aggregated and sub-metered electricity and gas readings from 33 household sensors, captured at a resolution of one second. The data set covers 73 days of a single house in Delhi, India.
\subsection{Algorithms Considered}
For the purpose of this experiment, five neural network based algorithms have been utilised, namely: DAE, RNN, Sequence-to-Sequence, Sequence-to-Point and OnlineGRU, all of which have been described in detail in section 4.1 of this article.
\begin{figure*}[!htbp]
\includegraphics[width=\textwidth,height=\textheight,keepaspectratio]{redd_crop.pdf}\hfill
\caption{Predicted energy consumption: REDD}
\label{Fig: REDD}
\end{figure*}
\subsection{Experimental Scenario}
The tests were run on a machine with GeForce GTX 1660 Ti/PCIe/SSE2 GPUs with Intel® Core™ i7-9750H CPU @ 2.60GHz × 12 and 16 GB RAM. The batch size was 32, for all neural algorithms and they were trained for 5 epochs each. The sample period was kept as 60 seconds.
\subsection{Results and Discussion} For each data set, the network was trained for a period of 30 days based on the data and tested on the subsequent 20 days using the nilmtk-contrib API to predict the energy consumed per appliance and report the mean absolute error(MAE) in each case. The outputs from the tests on the three datasets under consideration do not collectively reveal the best performer out of the five algorithms used. Moreover, there is no clear winner in each case except UK-dale where Sequence-to-Sequence leads. This could be due to the differences in the data used.
Mean absolute error calculates the total magnitude of error in a set of predictions, not considering their direction.
It assumes that all individual differences are supposed to have equal weight and therefore, for a test sample, it gives out the absolute differences between the actual and predicted observation.
It can be used to represent the average prediction error in terms of the variable at the test. The metric ranges from 0 to $\infty$ and is unresponsive to the direction of errors. This score has a negative orientation, the lower the value, the better. It effectively describes the magnitude of residuals and does not indicate the underperformance or overperformance of the model. All elements are given equal weightage and hence contribute directly to the total value of the error and larger the value, linear is the contribution.
\begin{figure*}[!htbp]
\includegraphics[width=\textwidth,height=\textheight,keepaspectratio]{ukdale_crop.pdf}\hfill
\caption{Predicted energy consumption: UKdale}
\label{Fig: UKdale}
\end{figure*}
It is noticeable from Table-\ref{Table: IAWE} that there is no clear winner in the case of iAWE and the mean absolute errors are very high which could be because the data set is based on a single housing facility. The mean absolute errors remain similar in cases of washing machine, computer and television with the reported errors being the minimum for television. The case of air conditioner reports the maximum observed error from all three scenarios which is greater than 100 for all five algorithms and varies largely for every algorithm. The performance of Sequence-to-Sequence and DAE remains similar across all five appliances with DAE performing the best for the television and Sequence-to-Sequence performing the best for the washing machine. It can also be observed from Figure-\ref{Fig: IAWE}, the difference in the load cycles of different appliances.
It is evident from Table-\ref{Table: REDD} that none of the algorithms in consideration perform the best overall in the case of REDD although Sequence-to-Point leads in four of the five most used appliances. This could be due to the dissimilar usage trends of the appliances that the graphs clearly demonstrate in Figure-\ref{Fig: REDD}, hence, signifying the energy consumption patterns of the appliances. The mean absolute errors are the lowest in this case making it a preferred data set to base our conclusions on.
The mean absolute errors in for the sockets remain less than 1 which is the minimum error reported by an algorithm in all three scenarios. The errors show maximum variation in the case of the washing machine followed by the fridge. There is little variation in the reported values of mean absolute error for the sockets which is again followed by the light. The performance of WindowGRU and RNN is comparable across light, sockets and microwave but differ largely in cases of the washing machine and fridge.
Table-\ref{Table: UKdale} shows that Sequence-to-Sequence performs the best for most cases of UK-Dale and the mean absolute errors, although high, are similar for nearly all appliances. The predictions from the graph indicate how some appliances have clearly defined usage patterns (as seen in Figure-\ref{Fig: UKdale}) whereas others are used on a daily basis. The mean absolute error stays the lowest in the case of microwave with 16.60 being the minimum as reported by RNN. It can also be noticed that the mean absolute error for the fridge remains similar across all five algorithms followed by little variation in the case of light and shows the maximum variation for the washing machine.
Different appliances use different quantities of energy and, thus, compared to a high energy consumption load, the errors measured relative to low energy use might be less significant. Furthermore, some end devices operate less frequently than others, so if a statistically substantial number of run times is not recorded, metric results may not be indicative of efficiency. In view of these problems, it is recommended that every metric assessment reflects some leveling of the use of energy or other basis for appropriate comparisons of results across different end uses which coule be a fixed energy usage per end use, a fixed number of real events per end use, or a fixed period that is capable of capturing a variety of conditions.
\section{Conclusion and future work}In this paper, we demonstrated how nilmtk-contrib provides an interface to energy disaggregation problems and also the advantages of the NILM API which enables comparison of algorithms in this domain even without expertise in the field, thus enabling efficient experimental evaluations. Consequently, acing development within the field and supporting the progress of research in NILM.
Our experimental results carried out on three publicly available data sets namely IAWE, REDD and UKdale do not indicate the presence of any patterns in the output.
This could be a result of the large differences in the data sets. Although, it suggests that the API can be used for data sets from different geographies.
With the need for energy and resources rising each day, careful and efficient utilisation is the only way, to conserve it and energy dis-aggregation can be an essential element in the conservation of energy since it elaborates the energy usage tendencies.
The trends observed in the energy-usage patterns from a household can be used for the purpose of security as an anomaly in them might represent a sign of appliance failure or illegal use of supplied electricity. The appliance usage patterns can also be used to calculate and control the amount of carbon emissions.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
We consider the problem of finding $w \in {\mathbb{R}}^n$ that minimizes the cost
\begin{align} \label{eq:dist_opt_problem}
F(w) := \frac1N \sum_{i=1}^N f_i(w),
\end{align}
in a distributed way.\footnote{The scaling factor $1/N$ is simply for convenience in this paper because we are interested in the minimizer $w^*$ but not in the optimal cost.}
Here, $F(\cdot)$ is assumed to be strongly convex, but each $f_i \colon {\mathbb{R}}^n \rightarrow {\mathbb{R}}$ need not be convex.
Each $f_i(\cdot)$ is assumed to be continuously differentiable and its gradient $\nabla f_i(\cdot)$ is globally Lipschitz.
We consider a network of $N$ agents and each agent $i$ ($i=1,\dots,N$) knows the local cost function $f_i$ only but not others.
The main objective in this paper is the distributed minimization problem of \eqref{eq:dist_opt_problem}; that is, to design the agents in the network and their communication policy such that each agent finds the global minimizer of $F(w)$.
Distributed optimization problem has received much attention due to various applications such as resource allocation problem, distributed state estimation, or distributed machine learning.
It has been studied mostly in the discrete-time domain.
For example, early solutions to the problem can be found in \citep{Nedic2009} which proposes a consensus-based subgradient method.
Extensions are then made by numerous works, e.g., for fixed step size \citep{Yuan2016}, for asymptotic convergence under fixed step size \citep{Shi2015}, and for gradient tracking method \citep{Qu2018} to name a few.
A common approach in these works is to combine a consensus algorithm with the classical gradient descent method to obtain a distributed algorithm.
Finally, a common framework to analyze different variations of the distributed algorithms is also studied \citep{Alghunaim2020,Jakovetic2019}, which unifies various algorithms proposed in the literature.
Also see \citep{Yang2019} for a survey of distributed algorithms.
Recently, further improvements are made to the performance of the distributed algorithms.
From the optimization theory, it is well known that accelerated methods such as Nesterov gradient descent \citep{Nesterov2004} improve the convergence rate.
Consequently, accelerated optimization methods are combined with consensus algorithms to obtain distributed algorithms with improved performance.
For instance, \citet{Qu2020} proposed a distributed Nesterov gradient descent and \cite{Xin2019} proposed a distributed heavy-ball method.
In parallel to the studies in the discrete-time domain, continuous-time optimization algorithms have also attracted attention due to the insights it provides based on the rich knowledge from the classical stability analysis.
For instance, a continuous-time Nesterov gradient descent is analyzed by \citet{Su2016}, and Lyapunov analysis for momentum methods is done in the works such as \citep{Wilson2016,Shi2019}.
Accordingly, continuous-time {\em distributed} algorithms have also been developed.
Early works are done by \cite{Wang2010} which proposed a proportional-integral type algorithm.
This is extended in various manners, e.g., for discrete communications using event-triggered controls \citep{Kia2015}, for strongly connected and weight-balanced graphs \citep{Gharesifard2014}, for communication delays using passivity \citep{Hatanaka2018}, and for a constrained problem \citep{Yang2017}.
However, these works employ the gradient descent algorithm and accelerated methods are not adopted.
Most importantly, all of the distributed algorithms mentioned so far (both discrete and continuous cases) assume convexity of each local function $f_i$.
In this paper, we introduce a continuous-time distributed optimizer
and present a distributed heavy-ball method.
The advantages are listed as follows.
\begin{itemize}
\item Individual cost function $f_i(w)$ need not be convex as long as their sum, i.e., the global cost function $F(w)$, is strongly convex.
\item The convergence rate of the distributed algorithm is arbitrarily close to the convergence rate of the centralized one.
\item Two algorithms are presented using the proportional-integral(PI)-type coupling. The first one communicates $2n$-dimensional information with neighboring agents, and has the benefit of `initialization-free,' so that agents can join or leave the network during the operation. The second one communicates just $n$-dimensional information but a specific initialization is needed.
\item We present a general theorem that can also be used for designing other distributed algorithms, and demonstrate its use for designing a distributed gradient descent algorithm.
\end{itemize}
\begin{comment}
the proportional-integral (PI) based distributed optimization algorithms and analyze their convergence.
We extend the classical approach by considering the higher-order dynamical systems as well as output coupling.
Motivated by the practical synchronization of heterogeneous agents \citep{Lee2020}, a concept of \textit{blended dynamics} of the overall system is proposed which provides a novel perspective of the algorithms.
In particular, convergence is analyzed by using the stability of the blended dynamics instead of the stability of individual agents.
For example, we only require strong convexity of $F(w)$ and do not assume convexity of individual cost function $f_i(w)$.
Convergence rates of the proposed distributed algorithms are also analyzed.
The proposed approach shows that the convergence rates of the distributed algorithms can be decomposed into the convergence rate of the blended dynamics and the gains associated with proportional and integral feedback.
Hence, by designing gains appropriately, performance of the blended dynamics is recovered.
Using the proposed blended dynamics approach, we present a systematic method to construct distributed algorithms from a centralized algorithm.
In particular, a distributed heavy-ball algorithm is proposed based on the centralized heavy-ball algorithm.
We also obtain that the proposed distributed heavy-ball algorithm achieves the accelerated convergence rate with suitably chosen gains.
\end{comment}
In Section \ref{sec:main1}, we state the general theorem regarding the behavior of multi-agent system in which each agent exchanges their output information only by the PI-type couplings.
This result is utilized in Section \ref{sec:opt} for distributed optimization algorithms.
Numerical simulations and conclusions appear in Sections \ref{sec:simulation} and \ref{sec:conclusion}, respectively.
We use the following notation in this paper.
For given matrices $X_1,\ldots,X_N$, we denote $[X_1^\top ~ \cdots ~ X_N^\top]^\top$ by $[X_1;\cdots;X_N]$ and we let $1_N \coloneqq [1;\cdots;1] \in {\mathbb{R}}^N$.
For a vector $x$ and a matrix $X$, $|\cdot|$ represents their Euclidean 2-norm and induced 2-norm, respectively.
The Kronecker product is denoted by $\otimes$.
The minimum and the maximum singular values of a matrix $A$ are denoted by $\sigma_m(A)$ and $\sigma_M(A)$, respectively.
A (undirected) graph is defined as a tuple $({\mathcal{N}},{\mathcal{E}})$ where ${\mathcal{N}}\coloneqq\{1,\ldots,N\}$ is the node set and ${\mathcal{E}} \subseteq {\mathcal{N}} \times {\mathcal{N}}$ is the edge set.
The set of neighbors of agent $i$ is defined as ${\mathcal{N}}_i \coloneqq \{j \in {\mathcal{N}} \mid (j,i) \in {\mathcal{E}} \}$.
The graph is connected if two nodes can be joined by a path \citep{Godsil2001}.
The Laplacian matrix ${\mathfrak{L}}=[l_{ij}] \in {\mathbb{R}}^{N \times N}$ is defined as $l_{ij}\coloneqq-1$ if $(j,i)\in {\mathcal{E}}$ and $l_{ij} \coloneqq 0$ otherwise for $i \neq j$, and $l_{ii}\coloneqq-\sum_{j \neq i} l_{ij}$.
\section{Distributed Computation Algorithm with PI Coupling}\label{sec:main1}
In this section, we present a result on the behavior of multi-agent system under the proportional-integral (PI) type of couplings, which will lead to a few applications of distributed optimization in the next section.
Let us consider the multi-agent system consisting of $N$ agents with the node dynamics of agent $i \in {\mathcal{N}} := \{1,\ldots,N\}$ being given by
\begin{align} \label{eq:agent_dyn}
\begin{split}
\dot x_i & = h_i(x_i) + u_i \\
y_i & = E x_i,
\end{split}
\end{align}
where $x_i \in {\mathbb{R}}^n$ is the state of agent $i$, $h_i: {\mathbb{R}}^n \rightarrow {\mathbb{R}}^n$ is called the vector field of agent $i$ which is assumed to be globally Lipschitz, and $y_i \in {\mathbb{R}}^q$ is called the output of agent $i$ which is communicated with other agents.
(Moreover, in the next section, $y_i$ will be the variable that converges to the optimal value $w^* := \argmin_w F(w)$ for the minimization problem \eqref{eq:dist_opt_problem}.)
Typically, we have $q \le n$, and we assume that $E \in {\mathbb{R}}^{q \times n}$ has full row rank.
The input term $u_i \in {\mathbb{R}}^n$ is the coupling term whose value is determined with the received information from the neighboring agents.
We assume that {\em the communication graph is undirected and connected}, and consider the following PI-type coupling:
\begin{subequations}\label{eq:control}
\begin{align}
\dot \xi_i &= - \kappa K \sum_{j \in {\mathcal{N}}_i} (y_j - y_i) \label{eq:c} \\
\text{(A)}: \; u_i &= k_\mathrm P E^\top \sum_{j \in {\mathcal{N}}_i} (y_j - y_i)
+ k_\mathrm I E^\top \sum_{j \in {\mathcal{N}}_i} (\xi_j - \xi_i) \label{eq:ca}
\end{align}
where $\xi_i \in {\mathbb{R}}^q$ is the state, ${\mathcal{N}}_i$ is the index set of agents that send the information to the agent $i$, the coupling gains $\kappa$, $k_\mathrm P$ and $k_\mathrm I$ are positive numbers to be designed, and $K$ is a positive definite matrix to be designed as well.
On the other hand, we also consider another coupling input $u_i$ given by \eqref{eq:c} and
\begin{align}\label{eq:cb}
\begin{split}
\text{(B)}: \quad u_i &= k_\mathrm P E^\top \sum_{j \in {\mathcal{N}}_i} (y_j - y_i) - k_\mathrm I E^\top \xi_i \\
&{\color{black}\text{with} \;\; \sum_{i=1}^N \xi_i(0) = 0.}
\end{split}
\end{align}
\end{subequations}
An immediate observation is that the input (A) requires the communication of both $y_j$ and $\xi_j$ among the agents while only $y_j$ is communicated for the input (B).
However, when the input (B) is used, the initial conditions should satisfy $\sum_{i=1}^N \xi_i(0) = 0$, while this is not needed when the input (A) is used.
In this sense, the algorithm (A) (i.e., \eqref{eq:c} and \eqref{eq:ca}) is called to be {\em initialization-free}, which is desired if some agents leave and/or new agents join the network during the operation.
Define $x \in {\mathbb{R}}^{Nn}$ and $\xi \in {\mathbb{R}}^{Nq}$ as the column stack of $x_i$ and $\xi_i$, respectively.
Then the dynamics \eqref{eq:agent_dyn}, \eqref{eq:c}, and \eqref{eq:ca} or \eqref{eq:cb}, can be written compactly as
\begin{subequations} \label{eq:pi_dyn_stack}
\begin{align}
\dot x &= h(x) - k_\mathrm P ({\mathfrak{L}}_\mathrm P \otimes E^\top E) x - k_\mathrm I ({\mathfrak{D}}_\mathrm I \otimes E^\top) \xi \label{eq:pi_dyn_stack1} \\
\dot \xi &= \kappa ({\mathfrak{L}}_\mathrm P \otimes KE) x \label{eq:pi_dyn_stack2}
\end{align}
\end{subequations}
where $h(x) \coloneqq [h_1(x_1);\cdots;h_N(x_N)] \in {\mathbb{R}}^{Nn}$, ${\mathfrak{L}}_\mathrm P$ is the Laplacian matrix for the communication graph, and
\begin{align*}
{\mathfrak{D}}_\mathrm I = \begin{cases} {\mathfrak{L}}_\mathrm P, & \text{if input (A) is used,} \\
I_N, & \text{if input (B) is used.} \end{cases}
\end{align*}
In order to analyze the behavior of \eqref{eq:pi_dyn_stack}, two steps of state transformations, which are inspired from the work \citep{Lee2020}, are introduced.
First, from the given matrix $E \in {\mathbb{R}}^{q \times n}$, find two matrices $Z \in {\mathbb{R}}^{n \times (n-q)}$ and $W \in {\mathbb{R}}^{n \times q}$ such that the columns of $Z$ and $W$ are an orthonormal basis of $\ker(E)$ and $\ker(E)^\perp$, respectively.
Then, $T_1 := [Z \ W] \in {\mathbb{R}}^{n \times n}$ is an orthogonal matrix, $EZ = 0$, and $EW \in {\mathbb{R}}^{q \times q}$ is an invertible matrix.
As a result, with $Z$ and $W$, the individual state $x_i$ for $i \in {\mathcal{N}}$ can be transformed into $z_i$ and $w_i$ as
\begin{equation}
z_i = Z^\top x_i \;\; \in {\mathbb{R}}^{n-q}, \qquad w_i = W^\top x_i \;\; \in {\mathbb{R}}^{q}, \label{eq:ct1}
\end{equation}
so that we have
\begin{equation}\label{eq:x_i}
x_i = Z z_i + W w_i, \qquad \forall i \in {\mathcal{N}}.
\end{equation}
Next, define $R \in {\mathbb{R}}^{N \times (N-1)}$ such that the columns of $R$ is an orthonormal basis of $\ker(1_N^\top)$.
Then, we have the following property for a matrix $T_2 \in {\mathbb{R}}^{N \times N}$;
\begin{align*}
T_2 := \begin{bmatrix}
\frac{1}{N}1_N^\top \\ R^\top
\end{bmatrix}, \quad \text{then} \quad
T_2^{-1} =
\begin{bmatrix}
1_N & R
\end{bmatrix}.
\end{align*}
Let $w \in {\mathbb{R}}^{Nq}$ be the column stack of $w_i$.
Then, by the matrix $(T_2 \otimes I_q)$, the states $w$ and $\xi$ can be converted into $[\bar w;\widetilde{w}]$ and $[\bar \xi;\widetilde{\xi}]$ as
\begin{equation}
\begin{aligned}
\bar w &= \frac1N \sum_{i=1}^N w_i, & \bar \xi &= \frac1N \sum_{i=1}^N \xi_i, \\
\widetilde{w} &= (R^\top \otimes I_q) w, & \widetilde{\xi} &= (R^\top \otimes I_q) \xi
\end{aligned} \label{eq:ct2}
\end{equation}
where $\bar w, \bar \xi \in {\mathbb{R}}^{q}$ and $\widetilde w, \widetilde \xi \in {\mathbb{R}}^{q(N-1)}$.
Thus, we have
\begin{subequations}
\begin{align}
w &= (1_N \otimes I_q) \bar w + (R \otimes I_q) \widetilde w \label{eq:w} \\
\xi &= (1_N \otimes I_q) \bar \xi + (R \otimes I_q) \widetilde \xi \label{eq:xi} \\
x &= (I_N \otimes Z) z + (1_N \otimes W) \bar w + (R \otimes W) \widetilde w \label{eq:x1} \\
\text{or,} \;\; x_i &= Z z_i + W \bar w + (R_i \otimes W) \widetilde w, \quad \forall i \in {\mathcal{N}} \label{eq:x2}
\end{align}
\end{subequations}
where $z \in {\mathbb{R}}^{N(n-q)}$ is the column stack of $z_i$ and $R_i$ is the $i$-th row of $R$.
Now, through two consecutive linear coordinate changes by \eqref{eq:ct1} and \eqref{eq:ct2}, the system \eqref{eq:pi_dyn_stack} is converted into
\begin{subequations} \label{eq:pi_dyn_trans}
\begin{align}
\dot z_i &= Z^\top h_i(x_i), \quad i \in {\mathcal{N}}, \label{eq:pi_dyn_trans_z} \\
\dot{\bar w} &= \frac1N \sum_{i=1}^N W^\top h_i(x_i) \label{eq:pi_dyn_trans_w_bar}\\
\begin{split} \label{eq:pi_dyn_trans_w_tilde}
\dot{\widetilde{w}} &= (R^\top \otimes W^\top) h(x) - k_\mathrm P (\Lambda_{\mathrm P} \otimes W^\top E^\top E W) \widetilde{w} \\
& \quad - k_\mathrm I (\Lambda_\mathrm I \otimes W^\top E^\top) \widetilde{\xi}
\end{split} \\
\dot{\bar{\xi}} & = 0 \label{eq:pi_dyn_trans_xi_bar}\\
\dot{\widetilde{\xi}} & = \kappa (\Lambda_\mathrm P \otimes KEW) \widetilde{w} \label{eq:pi_dyn_trans_xi_tilde}
\end{align}
\end{subequations}
where $\Lambda_\mathrm P = R^\top {\mathfrak{L}}_\mathrm P R \in {\mathbb{R}}^{(N-1) \times (N-1)}$ is positive definite because the graph is connected, and $\Lambda_\mathrm I$ is either $\Lambda_\mathrm P$ or $I_{N-1}$ depending on the cases (A) and (B), respectively.
No coupling term appears in \eqref{eq:pi_dyn_trans_z} because $EZ = 0$.
To see how \eqref{eq:pi_dyn_trans_w_bar} is derived, note that $\bar w = (1/N)(1_N^\top \otimes W^\top) x$.
Then, it is seen that no coupling term appears in \eqref{eq:pi_dyn_trans_w_bar} in the case of (A) because $1_N^\top {\mathfrak{L}}_\mathrm P = 0$.
For the case of (B), we have $\dot {\bar w} = (1/N)\sum_{i=1}^N W^\top h_i(x_i) - k_\mathrm I E^\top \bar \xi$, but by \eqref{eq:pi_dyn_trans_xi_bar} and by $\bar \xi(0) = (1/N) \sum_{i=1}^N\xi_i(0) = 0$ from \eqref{eq:cb}, we have that $\bar \xi(t) \equiv 0$.
Equation \eqref{eq:pi_dyn_trans_w_tilde} is obtained with $\widetilde w = (R^\top \otimes W^\top) x$ by \eqref{eq:pi_dyn_stack1}, \eqref{eq:xi}, and \eqref{eq:x1}, and the fact that $\bar \xi(t) \equiv 0$ is used for the case of (B).
Equations \eqref{eq:pi_dyn_trans_xi_bar} and \eqref{eq:pi_dyn_trans_xi_tilde} are obtained from \eqref{eq:pi_dyn_stack2}, \eqref{eq:ct2}, and \eqref{eq:x1} since $\color{black} \bar \xi = (1/N) (1_N^\top \otimes I_q) \xi$.
It is worthwhile to emphasize again that the average of $\xi_i$ (i.e., $\bar{\xi}$) is constant in both cases of (A) and (B), and hence, $\bar{\xi}(t)$ is completely determined by the initial conditions of $\xi_i(0)$.
\begin{ass} \label{asm:exp_stab}
For a given multi-agent system \eqref{eq:agent_dyn}, define its {\em blended dynamics} as
\begin{align} \label{eq:blend_dyn}
\begin{split}
\dot z_i &= Z^\top h_i(Z z_i + W \bar w), \quad i \in {\mathcal{N}}, \\
\dot {\bar w} &= \frac1N \sum_{i=1}^N W^\top h_i(Z z_i + W \bar w) .
\end{split}
\end{align}
Assume that the blended dynamics \eqref{eq:blend_dyn} has a unique equilibrium $(z_1^*,\dots,z_N^*,\bar{w}^*)$ that is {\em globally exponentially stable with a rate} $\mu > 0$; that is,
\begin{equation}\label{eq:desiredrate}
|e_\delta(t)| \le c e^{-\mu t} |e_\delta(0)|
\end{equation}
where $e_\delta := [z - z^*; \bar w - \bar w^*]$ (with $z = [z_1; \cdots; z_N]$ and $z^* = [z_1^*; \cdots; z_N^*]$) and $c$ is a constant.
\end{ass}
Note that the blended dynamics \eqref{eq:blend_dyn} is nothing but the subsystem \eqref{eq:pi_dyn_trans_z} and \eqref{eq:pi_dyn_trans_w_bar} when $\widetilde w \equiv 0$\footnote{In fact, in \citep{Lee2020}, the blended dynamics is defined as the quasi-steady-state subsystem when \eqref{eq:pi_dyn_trans} is viewed as a singularly perturbed system. The difference is whether \eqref{eq:pi_dyn_trans_xi_bar} belongs to the blended dynamics, and here, we do not include it considering that the behavior of \eqref{eq:pi_dyn_trans_xi_bar} is trivial.}
(use \eqref{eq:x2}), and has the dimension of $N(n-q)+q$.
Since the blended dynamics \eqref{eq:blend_dyn} has the equilibrium $(z^*,\bar w^*)$ by Assumption \ref{asm:exp_stab}, the whole system \eqref{eq:pi_dyn_trans} has an equilibrium $(z^*, \bar w^*, \widetilde w^*, \bar \xi^*, \widetilde \xi^*)$ where $\widetilde w^* = 0$, $\bar \xi^* = (1/N) \sum_{i=1}^N \xi_i(0)$, and $\widetilde \xi^* = (1/k_\mathrm I) (\Lambda_\mathrm I \otimes W^\top E^\top)^{-1}(R^\top \otimes W^\top) h(x^*)$ in which $x^* = (I_N \otimes Z) z^* + (1_N \otimes W) \bar w^*$.
This equilibrium is unique with respect to the initial condition $\xi(0)$.
\begin{rmk}
The blended dynamics \eqref{eq:blend_dyn} can be seen as a {\em residual dynamics} of the multi-agent system \eqref{eq:agent_dyn} and \eqref{eq:control} (or, \eqref{eq:pi_dyn_trans}), which is left over when the output $y_i$ achieves consensus.
Indeed, if $y_i(t) \equiv y_j(t)$, then $w_i(t) \equiv w_j(t)$ because $y_i = E x_i = EW w_i$ by \eqref{eq:x_i} and $EW$ is invertible.
This yields that $\widetilde w(t) = (R^\top \otimes I_q)w(t) = 0$ because $R^\top 1_N = 0$.
Then, \eqref{eq:pi_dyn_trans_z} and \eqref{eq:pi_dyn_trans_w_bar} with $\widetilde w \equiv 0$ becomes the blended dynamics.
Note that the blended dynamics \eqref{eq:blend_dyn} contains the average of the vector fields $h_i$ (for its $\bar w$-dynamics), which differs from the dynamics of any individual agent as well as the overall dynamics.
Note that Assumption \ref{asm:exp_stab} asks stability of the blended dynamics (which consists of the averaged vector field), but not of individual agents, which will be the main ingredient how convexity of individual cost functions is not necessary.
\end{rmk}
For convenience, let us translate the equilibrium of \eqref{eq:pi_dyn_trans} into the origin through $z_\delta := z - z^*$, $\bar w_\delta := \bar{w} - \bar{w}^*$, and $\widetilde\xi_\delta = \widetilde{\xi} - \widetilde{\xi}^*$.
Then the state $x$ can be written as
\begin{align}\label{eq:xx}
\begin{split}
x &= (I_N \otimes Z) (z_\delta + z^*) \\
&\qquad + (1_N \otimes W) (\bar w_\delta + \bar w^*) + (R \otimes W) \widetilde{w}
\end{split}
\end{align}
so that
\begin{align} \label{eq:eq_pt_x}
x^* = (I_N \otimes Z) z^* + (1_N \otimes W) \bar{w}^*.
\end{align}
Then it can be verified that \eqref{eq:pi_dyn_trans} becomes
\begin{subequations} \label{eq:pi_dyn_final}
\begin{align}
\dot z_\delta &= (I_N \otimes Z^\top) h(x) \label{eq:origin1} \\
\dot {\bar w}_\delta &= \frac1N (1_N^\top \otimes I_n) h(x) \label{eq:origin2} \\
\begin{split}\label{eq:origin3}
\dot {\widetilde w} &= (R^\top \otimes W^\top) (h(x) - h(x^*)) \\
&\qquad - k_\mathrm P (\Lambda_\mathrm P \otimes W^\top E^\top E W) \widetilde{w} \\
&\qquad - k_\mathrm I (\Lambda_\mathrm I \otimes W^\top E^\top) \widetilde \xi_\delta
\end{split} \\
\dot {\bar \xi} &= 0 \label{eq:origin4} \\
\dot {\widetilde\xi}_\delta &= \kappa (\Lambda_\mathrm P \otimes KEW) \widetilde{w} \label{eq:origin5}
\end{align}
\end{subequations}
where $\Lambda_\mathrm I$ is $\Lambda_\mathrm P$ for (A) and $I_{N-1}$ for (B).
Now, we present the main result.
\begin{thm} \label{thm:main_convergence}
Consider the multi-agent system \eqref{eq:agent_dyn} and the PI-type coupling \eqref{eq:c} and \eqref{eq:ca}, or \eqref{eq:c} and \eqref{eq:cb}, with
$$K = E W W^\top E^\top.$$
Then, under Assumption \ref{asm:exp_stab}, the following results hold.
\begin{itemize}
\item [(a)] For any positive $\kappa$ and $k_\mathrm I$ such that $2\kappa > k_\mathrm I$, there exists\footnote{See \eqref{eq:k_rp1} in the proof for an explicit expression of $k^*_{\mathrm P}$.} $k^*_{{\mathrm P}}(\kappa,k_\mathrm I)$ such that, for each $k_\mathrm P > k^*_{\mathrm P}$, the origin of \eqref{eq:origin1}, \eqref{eq:origin2}, \eqref{eq:origin3}, and \eqref{eq:origin5} is globally exponentially stable.
Therefore, $x_i(t)$ of \eqref{eq:agent_dyn} converges to $x_i^* = Z z_i^* + W \bar w^*$ exponentially fast.
\item [(b)] Set
$$\kappa = k_\mathrm I = \phi^* k_\mathrm P$$
where $\phi^* > 0$ is sufficiently small.\footnote{In fact, $\phi^*$ can be chosen (see \eqref{eq:phistar}) such that
\begin{align*}
0< \phi^* &< \min \Big\{ \sqrt{2} \sqrt{\frac{\sigma_m(\Lambda_{\mathrm P})}{\sigma_M(\Lambda_\mathrm I)}},
\frac{\sigma_M(\Lambda_\mathrm P)}{\sigma_M(\Lambda_\mathrm I)} \sigma_M(E), \\
&\qquad \frac{2}{\sigma_M(E)}, \frac{\sigma_m(\Lambda_\mathrm P)\sigma_m^2(E)}{\sqrt{\sigma_M(\Lambda_\mathrm P) \sigma_M(\Lambda_\mathrm I)} \sigma_M^2(E)} \Big\} .
\end{align*}
}
For any positive $\upsilon < \mu$, there exists\footnote{See \eqref{eq:k_rp2} in the proof for an explicit expression of $k^{**}_{\mathrm P}$.} $k^{**}_{\mathrm P}(\upsilon,\phi^*) > 0$ such that, for each $k_\mathrm P > k^{**}_{\mathrm P}$, we have that
\begin{align*}
\left| [e_\delta(t); \widetilde{w}(t); \widetilde{\xi}_\delta(t)] \right| \leq \bar{c} e^{-(\mu - \upsilon)t} \left| [ e_\delta(0); \widetilde{w}(0); \widetilde{\xi}_\delta(0) ]\right|
\end{align*}
where $e_\delta = [z_\delta;\bar{w}_\delta]$ and $\bar c$ is a constant.
Therefore, $x_i(t)$ converges to $x_i^*$ exponentially fast at least with the rate $\mu - \upsilon$.
\end{itemize}
\end{thm}
Theorem \ref{thm:main_convergence}.(a) states that the proposed dynamics \eqref{eq:agent_dyn} with \eqref{eq:control} is exponentially stable for any (small) $\kappa$ and $k_\mathrm I$ if $2\kappa > k_\mathrm I$ and $k_\mathrm P$ is sufficiently large.
If $k_\mathrm P$ and $k_\mathrm I = \kappa$ are chosen as proposed in Theorem \ref{thm:main_convergence}.(b), then the convergence rate to the equilibrium can be made arbitrarily close to that of the blended dynamics.
For this result, existence of exponentially stable equilibrium for the blended dynamics \eqref{eq:blend_dyn} is enough, and each agent may even be unstable as long as the blended dynamics is stable.
\begin{rmk} \label{rem:gains}
In order to recover the convergence rate $\mu$ by choosing $\upsilon$ arbitrarily close to $\mu$, one has to choose $\kappa = k_\mathrm I = \phi^* k_\mathrm P$ with suitable choice of $\phi^*$ and $k_\mathrm P$, which may be tedious.
This can be simplified by choosing
$$\kappa = k_\mathrm I = \beta, \quad k_\mathrm P = \beta^{\frac32}, \quad \beta \gg 1.$$
Indeed, with the choice, $\phi^* = 1/\sqrt{\beta}$ so that large $\beta$ makes \eqref{eq:phistar} hold.
Moreover, it is seen from \eqref{eq:k_rp2} that $k^{**}_{\mathrm P} = \vartheta_2/(\phi^*)^2 = \vartheta_2 \beta$ (with some constant $\vartheta_2$) when $\phi^*$ is sufficiently small, i.e., $\beta$ is sufficiently large.
Therefore, it holds that $k_\mathrm P > k^{**}_\mathrm P$ if $\beta \gg 1$.
\end{rmk}
\begin{pf}
The first step is to obtain a Lyapunov function for the blended dynamics \eqref{eq:blend_dyn} that can characterize the convergence rate $\mu$.
For this purpose, we employ the converse Lyapunov theorem by \citet{Corless1998}, among others, which states under Assumption~\ref{asm:exp_stab} that, for any $\upsilon$ such that $0 < \upsilon < \mu$, there exists a Lyapunov function $\bar V(e_\delta)$, with $e_\delta = [z_\delta; \bar w_\delta]$, such that
\begin{gather*}
c_1 |e_\delta|^2 \leq \bar V(e_\delta) \leq c_2 |e_\delta|^2, \quad
\left|\frac{\partial \bar V}{\partial e_\delta}(e_\delta)\right| \leq c_3 |e_\delta|, \\
\dot {\bar V}(e_\delta)|_{\text{along \eqref{eq:blend_dyn}}} \le - (2\mu - \upsilon) \bar V(e_\delta)
\end{gather*}
where $c_1,c_2,c_3 > 0$.
Since $c_1|e_\delta(t)|^2 \le \bar V(e_\delta(t)) \le e^{-(2\mu-\upsilon)t} c_2 |e_\delta(0)|^2$, they just guarantee the convergence rate of $\mu - \upsilon/2$.
Because our goal is to recover the convergence rate $\mu$ arbitrarily closely as stated in Theorem \ref{thm:main_convergence}, the above statement is enough for our purpose.
With the function $\bar V$, let us consider
\begin{align*}
V(e_\delta,\widetilde w,\widetilde \xi_\delta) & = \bar V(e_\delta) +
\frac12 \begin{bmatrix} \widetilde w \\ \widetilde \xi_\delta \end{bmatrix}^\top
\begin{bmatrix} X & \phi Y \\ \phi Y^\top & 2U \end{bmatrix}
\begin{bmatrix} \widetilde w \\ \widetilde \xi_\delta \end{bmatrix}
\end{align*}
where
\begin{align*}
X &= \Lambda_\mathrm P \otimes W^\top E^\top E W, & U &= \Lambda_\mathrm I \otimes I_q, \\
Y &= \Lambda_\mathrm I \otimes W^\top E^\top, & \phi &= \phi(\kappa,k_\mathrm I,k_\mathrm P) = \frac{2\kappa - k_\mathrm I}{k_\mathrm P}.
\end{align*}
In order for the function $V$ to be a candidate Lyapunov function of \eqref{eq:pi_dyn_final}, it should be a positive definite function, which asks
\begin{align}\label{eq:theta1}
\phi < \sqrt{\frac{2\sigma_m(X)}{\sigma_M(YU^{-1}Y^\top)}} =: \theta_1.
\end{align}
On the other hand, we will need an upper bound of $V$ that is independent of $\phi$.
For this, noting that
$$\widetilde w^\top \phi Y \widetilde \xi_\delta \le \frac{\phi|Y|}{2}(|\widetilde w|^2 + |\widetilde \xi_\delta|^2),$$
it can be seen that, if $\phi |Y|/2 \le \max(|X|/2,|U|)$, or
\begin{align}\label{eq:theta2}
\phi \leq \frac{2}{|Y|} \max\left( \frac12 |X|, |U| \right) =: \theta_2,
\end{align}
then
\begin{align*}
\frac12 & \widetilde w^\top X \widetilde w + \widetilde w^\top \phi Y \widetilde \xi_\delta + \widetilde \xi_\delta^\top U \widetilde \xi_\delta \\
&\leq \max(|X|,2|U|) (|\widetilde{w}|^2 + |\widetilde \xi_\delta|^2) =: \eta (|\widetilde{w}|^2 + |\widetilde \xi_\delta|^2)
\end{align*}
in which, $\eta$ is independent of $\phi$.
With these in mind, let us take the time derivative of $V$, term by term.
For the first term, we have that
\begin{align*}
&\dot {\bar V}|_{\text{along \eqref{eq:origin1} and \eqref{eq:origin2}}} = \dot {\bar V}|_{\text{along \eqref{eq:blend_dyn}}} \\
&\quad + \frac{\partial \bar V}{\partial e_\delta} \cdot \begin{bmatrix} I_N \otimes Z^\top \\ \frac1N (1_N^\top \otimes W^\top) \end{bmatrix} \left( h(x) - h(x - (R \otimes W) \widetilde w) \right) \\
&\le - (2\mu-\upsilon) \bar V(e_\delta) + c_3 L |e_\delta| |\widetilde w| \\
&\le -2 (\mu-\upsilon) \bar V(e_\delta) - \upsilon c_1 |e_\delta|^2 + \frac{\upsilon c_1}{3} |e_\delta|^2 + \frac{3c_3^2 L^2}{4\upsilon c_1} |\widetilde w|^2
\end{align*}
where $L$ is a Lipschitz constant of $h$.
Here we note that
\begin{align*}
|h(x)-h(x^*)| &\le L |(I_N \otimes Z) z_\delta + (1_N \otimes W) \bar w_\delta \\
&\qquad + (R \otimes W) \widetilde w| \le 2L\sqrt{N} |e_\delta| + L |\widetilde w|.
\end{align*}
With this, we now have
\begin{align*}
\frac{d (\frac12 \widetilde w^\top X \widetilde w)}{d t} &= \widetilde w^\top X(R^\top\otimes W^\top) (h(x)-h(x^*)) \\
&\qquad - k_\mathrm P \widetilde w^\top X^2 \widetilde w - k_\mathrm I \widetilde w^\top XY \widetilde \xi_\delta \\
&\le 2 |X| L \sqrt{N} |e_\delta| |\widetilde w| + |X| L |\widetilde w|^2 \\
&\qquad - k_\mathrm P \sigma_m^2(X) |\widetilde w|^2 - k_\mathrm I \widetilde \xi_\delta^\top Y^\top X \widetilde w, \\
&\le \frac{\upsilon c_1}{3}|e_\delta|^2 + \frac{3 |X|^2 L^2 N}{\upsilon c_1} |\widetilde w|^2 + |X| L |\widetilde w|^2 \\
&\qquad - k_\mathrm P \sigma_m^2(X) |\widetilde w|^2 - k_\mathrm I \widetilde \xi_\delta^\top Y^\top X \widetilde w, \\
\frac{d (\widetilde \xi_\delta^\top U \widetilde \xi_\delta)}{d t} &= 2 \kappa \widetilde \xi_\delta^\top U (\Lambda_\mathrm P \otimes KEW) \widetilde w
= 2 \kappa \widetilde \xi_\delta^\top Y^\top X \widetilde w ,
\end{align*}
and
\begin{align*}
&\phi \frac{d (\widetilde w^\top Y \widetilde \xi_\delta)}{d t} = \phi \kappa \widetilde w^\top Y (\Lambda_\mathrm P \otimes KEW) \widetilde w \\
&\quad + \phi \widetilde \xi_\delta^\top Y^\top (R^\top \otimes W^\top) (h(x) - h(x^*)) \\
&\quad - \phi k_\mathrm P \widetilde \xi_\delta^\top Y^\top (\Lambda_\mathrm P \otimes W^\top E^\top E W) \widetilde w \\
&\quad - \phi k_\mathrm I \widetilde \xi_\delta^\top Y^\top (\Lambda_\mathrm I \otimes W^\top E^\top) \widetilde \xi_\delta \\
&\le \phi \kappa \widetilde w^\top Y (I_N \otimes EW) X \tilde w + 2 \phi |Y| L \sqrt{N} |e_\delta| |\widetilde \xi_\delta| \\
&\quad + \phi |Y| L |\widetilde w| |\widetilde \xi_\delta|
- \phi k_\mathrm P \widetilde \xi_\delta^\top Y^\top X \widetilde w - \phi k_\mathrm I \widetilde \xi_\delta^\top Y^\top Y \widetilde \xi_\delta \\
&\le \phi \kappa |Y||E||X| |\widetilde w|^2 + \frac{\upsilon c_1}{3}|e_\delta|^2 + \phi^2 \frac{3 |Y|^2 L^2 N}{\upsilon c_1} |\widetilde \xi_\delta|^2 \\
&\quad + \frac{\phi k_\mathrm I \sigma_m^2(Y)}{4} |\widetilde \xi_\delta|^2 + \frac{\phi|Y|^2L^2}{k_\mathrm I \sigma_m^2(Y)} |\widetilde w|^2 \\
&\quad - \phi k_\mathrm P \widetilde \xi_\delta^\top Y^\top X \widetilde w - \phi k_\mathrm I \sigma_m^2(Y) |\widetilde \xi_\delta|^2 .
\end{align*}
Then, it follows that
\begin{align}\label{eq:vdot}
\begin{split}
&\dot V \le -2(\mu - \upsilon) \bar V
-3\phi \left( \frac{k_\mathrm I \sigma_m^2(Y)}{4} - \frac{\phi |Y|^2 L^2 N}{v c_1} \right) |\widetilde \xi_\delta|^2 \\
&-\left( k_\mathrm P \sigma_m^2(X) - \phi \kappa |Y| |E| |X| - \frac{\phi}{k_\mathrm I} \frac{|Y|^2 L^2}{\sigma_m^2(Y)} - \theta_0 \right) |\widetilde w|^2
\end{split}
\end{align}
where
$$\theta_0 := \frac{3c_3^2 L^2}{4\upsilon c_1} + \frac{3 |X|^2 L^2 N}{\upsilon c_1} + |X| L.$$
Therefore, it is seen that, under the assumption that $\kappa$ and $k_\mathrm I$ are chosen such that $2\kappa-k_\mathrm I>0$ (so that $\phi = (2\kappa - k_\mathrm I)/k_\mathrm P > 0$), we can make $\dot V$ negative definite by letting $k_\mathrm P$ sufficiently large.
Indeed, the first big parenthesis in \eqref{eq:vdot} becomes positive when
\begin{equation}
k_\mathrm P > \frac{2\kappa-k_\mathrm I}{k_\mathrm I} \frac{4|Y|^2L^2N}{\sigma_m^2(Y)vc_1} =: \frac{2\kappa-k_\mathrm I}{k_\mathrm I} \theta_3.
\end{equation}
In addition, the second big parenthesis in \eqref{eq:vdot} becomes positive when
\begin{align}
\sigma_m^2(X) k_\mathrm P^2 - \theta_0 k_\mathrm P - (2\kappa-k_\mathrm I) \left( \kappa \theta_4 + \frac{\theta_5}{k_\mathrm I} \right) > 0
\end{align}
where $\theta_4 := |Y||E||X|$ and $\theta_5 := |Y|^2L^2/\sigma_m^2(Y)$, which is the case when
\begin{align*}
k_\mathrm P &> \frac{\theta_0 + \sqrt{\theta_0^2 + 4(2\kappa-k_\mathrm I)(\kappa\theta_4 + \theta_5/k_\mathrm I)\sigma_m^2(X)}}{2\sigma_m^2(X)} \\
&=: \varphi_1(\kappa,k_\mathrm I).
\end{align*}
Overall, the function $\dot V$ of \eqref{eq:vdot} becomes upper-bounded by a negative quadratic function if
\begin{align}
k_\mathrm P &> \max\left( \frac{2\kappa-k_\mathrm I}{\theta_1}, \frac{2\kappa-k_\mathrm I}{\theta_2}, \frac{2\kappa-k_\mathrm I}{k_\mathrm I}\theta_3, \varphi_1(\kappa,k_\mathrm I) \right) \notag \\
&=: k_{\mathrm P}^*(\kappa,k_\mathrm I,\Lambda_\mathrm P,\Lambda_\mathrm I,E,L,N,\upsilon,c_1,c_3) \label{eq:k_rp1}
\end{align}
where
\begin{align*}
\theta_1 &= \sqrt{2} \sqrt{\frac{\sigma_m(\Lambda_{\mathrm P})}{\sigma_M(\Lambda_\mathrm I)}} \frac{\sigma_m(E)}{\sigma_M(E)} \\
\theta_2 &= \max \left( \frac{\sigma_M(\Lambda_\mathrm P)}{\sigma_M(\Lambda_\mathrm I)} \sigma_M(E), \frac{2}{\sigma_M(E)} \right) \\
\theta_3 &= \frac{4 \sigma_M^2(\Lambda_\mathrm I) \sigma_M^2(E) L^2 N}{\sigma_m(\Lambda_\mathrm I)^2 \sigma_m(E)^2} \\
\theta_4 &= \sigma_M(\Lambda_\mathrm P) \sigma_M(\Lambda_\mathrm I) \sigma_M^4(E) \\
\theta_5 &= \frac{\sigma_M^2(\Lambda_\mathrm I) \sigma_M^2(E)}{\sigma_m^2(\Lambda_\mathrm I) \sigma_m^2(E)}L^2 \\
\theta_0 &= \frac{3c_3^2L^2}{4\upsilon c_1} + \frac{3\sigma_M^2(\Lambda_\mathrm P) \sigma_M^4(E) L^2 N}{\upsilon c_1} + \sigma_M(\Lambda_\mathrm P)\sigma_M^2(E)L,
\end{align*}
which is derived by the fact that $|X|=\sigma_M(\Lambda_\mathrm P) \sigma_M^2(E)$, $|Y| = \sigma_M(\Lambda_\mathrm I) \sigma_M(E)$, $\sigma_m(X) = \sigma_m(\Lambda_\mathrm P) \sigma_m^2(E)$, and $\sigma_m(Y) = \sigma_m(\Lambda_\mathrm I) \sigma_m(E)$.
This completes the proof of Theorem \ref{thm:main_convergence}.(a).
To prove Theorem \ref{thm:main_convergence}.(b), we will show that
\begin{align}\label{eq:tmp}
\begin{split}
\dot V &\le -2(\mu-\upsilon) \bar V - 2(\mu-\upsilon)\eta (|\widetilde w|^2 + |\widetilde \xi_\delta|^2 ) \\
&\le -2(\mu-\upsilon) V
\end{split}
\end{align}
by the choice of $\kappa = k_\mathrm I = \phi^* k_\mathrm P$ where $k_\mathrm P$ will be determined shortly and $\phi^*$ is a positive constant such that
\begin{equation}\label{eq:phistar}
\phi^* < \min \left( \theta_1, \theta_2, \frac{\sigma_m(X)}{\sqrt{\theta_4}} \right) .
\end{equation}
It follows from this choice that $\phi = (2\kappa - k_\mathrm I)/k_\mathrm P = \phi^*$, and so, the conditions \eqref{eq:theta1} and \eqref{eq:theta2} are satisfied.
Now, we want both the coefficients of $|\widetilde \xi_\delta|^2$ and $|\widetilde w|^2$ in \eqref{eq:vdot} to be less than $-2(\mu-\upsilon)\eta$ for \eqref{eq:tmp}.
For the first coefficient, we ask
\begin{equation}\label{eq:2a}
k_\mathrm P \frac{3 (\phi^*)^2 \sigma_m^2(Y)}{4} - \frac{3 (\phi^*)^2 |Y|^2L^2N}{\upsilon c_1} > 2(\mu-\upsilon)\eta.
\end{equation}
For the second, we ask
\begin{align}\label{eq:2b}
\begin{split}
&k_\mathrm P (\sigma_m^2(X) - (\phi^*)^2 |Y||E||X|) - \frac{1}{k_\mathrm P} \frac{|Y|^2L^2}{\sigma_m^2(Y)} - \theta_0 \\
&= k_\mathrm P (\sigma_m^2(X) - (\phi^*)^2 \theta_4) - \frac{\theta_5}{k_\mathrm P} - \theta_0 > 2(\mu-\upsilon)\eta.
\end{split}
\end{align}
Under \eqref{eq:phistar}, both inequalities \eqref{eq:2a} and \eqref{eq:2b} hold if $k_\mathrm P > k_{\mathrm P}^{**}$ where
\begin{align} \label{eq:k_rp2}
&k_{\mathrm P}^{**}(\phi^*,\Lambda_\mathrm P,\Lambda_\mathrm I,E,L,N,\upsilon,c_1,c_3,\mu) := \max \nonumber \\
&\Bigg\{ \vartheta_1, \frac{\vartheta_2}{(\phi^*)^2}, \frac{ (\theta_0 + 2(\mu\!-\!\upsilon)\eta) \!+\! \sqrt{(\theta_0 + 2(\mu\!-\!\upsilon)\eta)^2 + 4\vartheta_3 \theta_5}}{ 2\vartheta_3 } \Bigg\}
\end{align}
in which
\begin{align*}
\vartheta_1 &= 2 \cdot \frac{4 \sigma_M^2(\Lambda_\mathrm I) \sigma_M^2(E) L^2 N}{\sigma_m^2(\Lambda_\mathrm I) \sigma_m^2(E) \upsilon c_1}, \quad \vartheta_2 = 2 \cdot \frac{8 (\mu - \upsilon) \eta}{3 \sigma_m^2(\Lambda_\mathrm I) \sigma_m^2(E)}, \\
\vartheta_3 &= \sigma_m^2(\Lambda_\mathrm P) \sigma_m^4(E) - (\phi^*)^2 \theta_4, \\
\eta &= \max ( \sigma_M(\Lambda_\mathrm P)\sigma_M^2(E), 2 \sigma_M(\Lambda_\mathrm I) ) .
\end{align*}
This completes the proof.
\end{pf}
\begin{comment}
Noting that it is a second-order inequality for $k_\mathrm P$, it is satisfied when $k_\mathrm P > \vartheta_1$ where $\vartheta_1$ is the positive root of the second-order polynomial of $k_\mathrm P$.
Finally, we also want the coefficient of $|\widetilde \xi_\delta|^2$ in \eqref{eq:vdot} to be less than $-2(\mu-\upsilon)$; or,
which holds with $k_\mathrm P > \vartheta_2$ where $\vartheta_2$ is easily obtained from the above inequality.
Then, the proof completes with $k_\mathrm P > \max(\vartheta_1,\vartheta_2)$.
\begin{align*}
&\dot V \le -2(\mu - \upsilon) \bar V - k_\mathrm P\widetilde w^\top X^2 \widetilde w + \phi\kappa \widetilde w^\top Y (I_N \otimes EW)X \widetilde w \\
&+ \left( \frac{3c_3^2 L^2}{4\upsilon c_1} + \frac{3 |X|^2 L^2 N}{\upsilon c_1} + |X| L + \frac{\phi}{k_\mathrm I} \frac{|Y|^2L^2}{\lambda_{\min}(Y^\top Y)} \right) |\widetilde w|^2 \\
&- \phi \left( k_\mathrm I \frac{3\sigma_m^2(Y)}{4} - \phi \frac{3 |Y|^2 L^2 N}{\upsilon c_1} \right) |\widetilde \xi_\delta|^2 \\
& \le -2(\mu - \upsilon) \bar V - \frac{k_\mathrm P}{2} \widetilde{w} X^2 \widetilde{w} - \frac{\phi k_\mathrm I}{2} \lambda_{\min}(Y^\top Y) |\widetilde{\xi}_\delta|^2 \\
&-\bigg( \underbrace{ \frac{k_\mathrm P}{2}\widetilde w^\top X^2 \widetilde w - \phi \kappa \widetilde w^\top Y (I_N \otimes EW)X \widetilde w }_{\text{Term 1}} \bigg) \\
& \hspace{-5em} -\bigg( \underbrace{ \frac{k_\mathrm P \lambda_{\min}(X^2)}{2} - \frac{3c_3^2 L^2}{4\upsilon c_1} - |X| L - \frac{3 |X|^2 L^2 N}{\upsilon c_1} - \frac{\phi}{k_\mathrm I} \frac{|Y|^2L^2}{\lambda_{\min}(Y^\top Y)} }_{\text{Term 2}}\bigg) |\widetilde w|^2 \\
&- \phi \bigg( \underbrace{ \frac{k_\mathrm I}{4} \lambda_{\min}(Y^\top Y) - \phi \frac{3 |Y|^2 L^2 N}{\upsilon c_1}}_{\text{Term 3}} \bigg) |\widetilde \xi_\delta|^2.
\end{align*}
We claim that the terms 1,2, and 3 can be made positive with sufficiently large gain $k_\mathrm P$.
First, it can be checked that the term 1 is positive if
\begin{align*}
\frac{\sqrt{(2\kappa - k_\mathrm I)\kappa}}{k_\mathrm P} & \le \sqrt{\frac{\lambda_{\min}(X^2)}{2|Y||E| |X|} }\\
& = \frac{\lambda_{\min}(\Lambda_\mathrm P) }{\sqrt{2 \lambda_{\max}(\Lambda_\mathrm I) \lambda_{\max}(\Lambda_\mathrm P)}}
\frac{\sigma_{\min}(E)^2}{\sigma_{\max}(E)^2}
\\
& =: \theta_3.
\end{align*}
Secondly, the positivity of the term 2 is equivalent to the positivity of
\begin{align*}
& k_\mathrm P^2 - \left( \frac{3c_3^2 L^2}{2\upsilon c_1\lambda_{\min}(X^2)} + \frac{ 2|X| L}{\lambda_{\min}(X^2)} + \frac{6 |X|^2 L^2 N}{\upsilon c_1\lambda_{\min}(X^2)}\right) k_\mathrm P \\
& - \frac{2(2\kappa - k_\mathrm I)|Y|^2L^2}{k_\mathrm I\lambda_{\min}(Y^\top Y)\lambda_{\min}(X^2)}.
\end{align*}
Using the solution of the quadratic equation and the fact that $\sqrt{a^2 + b^2} \leq |a| + |b|$ for all $a,b \in {\mathbb{R}}$, we obtain that the term 2 is positive if
\begin{align*}
k_\mathrm P & \ge \frac{1}{\lambda_{\min}(X^2)} \left( \frac{3c_3^2 L^2}{2\upsilon c_1} + 2|X| L + \frac{6 |X|^2 L^2 N}{\upsilon c_1}\right) \\
& \quad + \sqrt{\frac{2(2\kappa - k_\mathrm I)}{k_\mathrm I}} \frac{ |Y| L}{ \sigma_{\min}(Y) \lambda_{\min}(X)} \\
& =: \kappa_1
\end{align*}
Finally, the term 3 is positive if
\begin{align*}
k_\mathrm P & \ge \frac{12 |Y|^2 L^2 N}{\upsilon c_1 \lambda_{\min}(Y^\top Y)} = \frac{12 L^2 N }{\upsilon c_1 } \frac{\lambda_{\max}(\Lambda_\mathrm I)^2}{\lambda_{\min}(\Lambda_\mathrm I)^2} \frac{\sigma_{\max}(E)^2}{\sigma_{\min}(E)^2} \\
& =: \kappa_2
\end{align*}
Now, for some given gains $k_\mathrm I$ and $\kappa$, let
\begin{align*}
k^*_{\mathrm P,1} := \max \left( \kappa_1, \kappa_2, \frac{2\kappa - k_\mathrm I}{\theta_1}, \frac{2\kappa - k_\mathrm I}{\theta_2}, \frac{\sqrt{(2\kappa - k_\mathrm I)k_\mathrm I}}{\theta_3} \right),
\end{align*}
Then, for any $k_\mathrm P \geq k^*_{\mathrm P,1}$, the terms 1,2, and 3 are positive and we obtain
\begin{align*}
\dot{V} & \le-2(\mu - \upsilon) \bar V - \frac{k_\mathrm P}{2} \lambda_{\min}(X^2) |\widetilde{w}|^2 \\
& \qquad - \frac{(2\kappa - k_\mathrm I)k_\mathrm I}{2 k_\mathrm P} \lambda_{\min}(Y^\top Y) |\widetilde{\xi}_\delta|^2 \\
& \le-2(\mu - \upsilon) \bar V - \min \left( \frac{k_\mathrm P}{2} \lambda_{\min}(X^2), \right.\\
& \qquad\qquad \left. \frac{(2\kappa - k_\mathrm I)k_\mathrm I}{2 k_\mathrm P} \lambda_{\min}(Y^\top Y) \right) \left( |\widetilde{w}|^2 + |\widetilde{\xi}_\delta|^2\right) \\
& \le-2(\mu - \upsilon) \bar V \\
& \quad - \frac{1}{\eta} \min \left( \frac{k_\mathrm P}{2} \lambda_{\min}(X^2), \frac{(2\kappa - k_\mathrm I)k_\mathrm I}{2 k_\mathrm P} \lambda_{\min}(Y^\top Y) \right) V_2(\tilde{w}, \tilde{\xi}_\delta) \\
& \le - \min \left( 2(\mu - \upsilon), \frac{k_\mathrm P}{2 \eta} \lambda_{\min}(X^2), \frac{(2\kappa - k_\mathrm I)k_\mathrm I}{2 k_\mathrm P \eta} \lambda_{\min}(Y^\top Y) \right) {V},
\end{align*}
where $\eta$ is defined in \eqref{eq:defn_eta}.
That is, the Lyapunov function ${V}$ decays with a rate
\begin{align}
& \min \left( 2(\mu - \upsilon), \frac{k_\mathrm P}{2 \eta} \lambda_{\min}(X^2), \frac{(2\kappa - k_\mathrm I)k_\mathrm I}{2 k_\mathrm P \eta} \lambda_{\min}(Y^\top Y) \right) \label{eq:conv_rate_simple} \\
&= \min \left( 2(\mu - \upsilon), \frac{k_\mathrm P}{2 \eta} \lambda_{\min}(\Lambda_\mathrm P)^2 \sigma_{\min}(E)^4, \right. \nonumber\\
& \hspace{10em} \left. \frac{(2\kappa - k_\mathrm I)k_\mathrm I}{2 k_\mathrm P \eta} \lambda_{\min}(\Lambda_\mathrm I)^2 \sigma_{\min}(E)^2 \right) \nonumber
\end{align}
Consequently, we obtain that $x^*$ is exponentially stable with the rate \eqref{eq:conv_rate}.
In order to recover the convergence rate of $\mu - \upsilon$, let $k_\mathrm I = \kappa$ and choose $k_\mathrm I$ such that
\begin{align*}
k_\mathrm I = \theta^* k_\mathrm P
\end{align*}
where $\theta^* := \min(\theta_1,\theta_2,\theta_3)$ and let
\begin{align*}
& k^*_{\mathrm P,2} := \max\left( \kappa^*, \frac{2\eta(\mu - \upsilon)}{\lambda_{\min}(X^2)}, \frac{2 \eta (\theta^*)^2 (\mu - \upsilon)}{\lambda_{\min}(Y^\top Y)} \right) \\
& =\max\left( \kappa^*, \frac{2 \eta(\mu - \upsilon) }{\lambda_{\min}(\Lambda_\mathrm P)^2 \sigma_{\min}(E)^4}, \frac{2 \eta (\theta^*)^2 (\mu - \upsilon) }{\lambda_{\min}(\Lambda_\mathrm I)^2 \sigma_{\min}(E)^2}\right).
\end{align*}
Then for any $k_\mathrm P \ge k^*_{\mathrm P,2}$, it follows from the construction of $k_{\mathrm P,2}^*$ that
\begin{align}
2(\mu - \upsilon) & \le \frac{k_\mathrm P}{2 \eta} \lambda_{\min}(X^2) \label{eq:exp_rate_cond_1}\\
2(\mu - \upsilon) & \le \frac{k_\mathrm P \lambda_{\min}(Y^\top Y)}{2 \eta (\theta^*)^2} = \frac{k_\mathrm I^2 \lambda_{\min}(Y^\top Y)}{2 k_\mathrm P \eta} \label{eq:exp_rate_cond_2}.
\end{align}
Additionally, it can be checked that $k_\mathrm P \geq k^*_{\mathrm P,1} = \max(\allowbreak\kappa_1, \kappa_2, (\theta^*/\theta_1) k_\mathrm P, (\theta^*/\theta_2) k_\mathrm P, (\theta^*/\theta_3)k_\mathrm P)$ holds.
Hence, it follows from \eqref{eq:conv_rate_simple}, \eqref{eq:exp_rate_cond_1} and \eqref{eq:exp_rate_cond_2} that the convergence rate becomes
\begin{align*}
& \min\left( 2(\mu - \upsilon), \frac{k_\mathrm P}{2 \eta} \lambda_{\min}(X^2), \frac{(2\kappa - k_\mathrm I)k_\mathrm I}{2 k_\mathrm P \eta} \lambda_{\min}(Y^\top Y) \right) \\
& = 2(\mu - \upsilon).
\end{align*}
In conclusion, we obtain
$$
\dot{{V}} \leq -2(\mu - \upsilon) {V}.
$$
Recalling that $V$ is a Lyapunov function in terms of $e_\delta, \widetilde{w}$, and $\widetilde{\xi}_\delta$, it follows that there exists a constant $\bar{c}$ such that \eqref{eq:sln_conv} holds.
After tedious calculations, it can also be verified that the time-derivative of $V_2^{\rm d} := (1/2) \tilde{w}^\top \Lambda_\mathrm P \tilde{w} + (k_\mathrm I/\kappa) \delta \tilde{\xi}^\top \otimesr{R}^\top {\mathfrak{D}}_\mathrm I \otimesr{R} \delta \tilde{\xi}$ along \eqref{eq:pi_dyn_final} becomes
\begin{align} \label{eq:lyap_diag}
\dot{V}_2^{\rm d}& \leq \beta_{13} |\delta z||\tilde{w}| + \beta_{23} |\delta \bar{w}| |\tilde{w}| - (\alpha_{\tilde{w}} + \frac{k_\mathrm P}{2}\lambda_{1}(\Lambda_\mathrm P)^2) |\tilde{w}|^2 \nonumber \\
& + k_\mathrm I \tilde{w}^\top \Lambda_\mathrm P (R^\top \otimes W^\top E^\top) {\mathfrak{D}}_\mathrm I \otimesr{R} \delta \tilde{\xi}
\end{align}
where the constants are defined as $\beta_{13} =\beta_{23} \coloneqq \sqrt{N}|\Lambda_\mathrm P| \bar{L}$ and $\alpha_{\tilde{w}} \coloneqq (1/2)k_\mathrm P \lambda_{1}(\Lambda_\mathrm P)^2 - |\Lambda_\mathrm P| \bar{L}$.
On the other hand, the time-derivative of $V_2^{\rm off} := \epsilon \tilde{w}^\top Y \delta \tilde{\xi}$ along \eqref{eq:pi_dyn_final} becomes
\begin{align}
\dot{V}_2^{\rm off} & \leq \alpha_{14}(\epsilon) |\delta z| |\delta \tilde{\xi}| + \alpha_{24}(\epsilon) |\delta \bar{w}| |\delta \tilde{\xi}| + \alpha_{34}(\epsilon)|\tilde{w}| |\delta \tilde{\xi}| \nonumber \\
&\hspace{-1em} + \alpha_{33}(\epsilon) |\tilde{w}|^2 - \epsilon k_\mathrm P \tilde{w}^\top \Lambda_\mathrm P Y \delta \tilde{\xi} - \epsilon k_\mathrm I \alpha_{\delta \tilde{\xi}}|\delta \tilde{\xi}|^2, \label{eq:lyap_off}
\end{align}
where the constants are defined as
$
\alpha_{14}(\epsilon) = \alpha_{34}(\epsilon) \coloneqq \epsilon \cdot {\bar{L} |Y|}
$,
$
\alpha_{24}(\epsilon) \coloneqq \epsilon \cdot \sqrt{N}\bar{L} |Y|
$,
$
\alpha_{33}(\epsilon) \coloneqq \epsilon \cdot 2 \kappa |Y| |\Lambda_\mathrm P||EW|
$, and
$
\alpha_{\delta \tilde{\xi}} \coloneqq \lambda_{1}((R_\otimes^\top {\mathfrak{D}}_\mathrm I R_\otimes) (I_N \otimes EWW^\top E^\top)(R_\otimes {\mathfrak{D}}_\mathrm I R_\otimes)).
$
Finally, combining \eqref{eq:lyap_1}, \eqref{eq:lyap_diag}, and \eqref{eq:lyap_off}, the time-derivative of $V$ along \eqref{eq:pi_dyn_final} can be written as
\begin{align}
\dot{V} & \leq - 2(\mu - \upsilon) \bar{V}(\delta_e) - \frac{k_\mathrm P}{2}\lambda_{1}(\Lambda_\mathrm P)^2 |\tilde{w}|^2 - \epsilon k_\mathrm I \alpha_{\delta \tilde{\xi}} |\delta \tilde{\xi}|^2 \nonumber \\
& + (k_\mathrm I - \epsilon k_\mathrm P) \tilde{w}^\top \Lambda_\mathrm P (R\otimes EW)^\top {\mathfrak{D}}_\mathrm I (R\otimes I_q) \delta \tilde{\xi}- \bar{e}^\top \Xi \bar{e} \label{eq:pf_lyap_with_V}\\
& \leq - \bar{e}^\top \Xi \bar{e}, \label{eq:pf_lyap}
\end{align}
if we let $\epsilon = k_\mathrm I / k_\mathrm P$, $\bar{e} := [|\delta z|; |\delta \bar{w}|; |\tilde{w}|; |\delta \tilde{\xi}|] \in {\mathbb{R}}^4$ and define the symmetric matrix $\Xi \in {\mathbb{R}}^{4 \times 4}$ as
\begin{align*}
\Xi \coloneqq -\frac{1}{2}
\begin{bmatrix}
-2 \upsilon c_1 & 0 & \alpha_{13} + \beta_{13} & \alpha_{14}(\epsilon) \\
0 & -2 \upsilon c_1 & \alpha_{23} + \beta_{23} & \alpha_{24}(\epsilon)\\
\alpha_{13} + \beta_{13} & \alpha_{23} + \beta_{23} & 2(\alpha_{33}(\epsilon) - \alpha_{\tilde{w}}) & \alpha_{34}(\epsilon)\\
\alpha_{14}(\epsilon) & \alpha_{24}(\epsilon) & \alpha_{34}(\epsilon) & - \alpha_{\delta \tilde{\xi}}(\epsilon)
\end{bmatrix}.
\end{align*}
Now, apply Young's inequality to \eqref{eq:pf_lyap}.
For example, we have
\begin{align*}
(a_{13} + \beta_{13}) |\delta z| |\tilde{w}| & \leq \frac{1}{2 \rho} |\delta z|^2 + \frac{\rho}{2} (\alpha_{13} + \beta_{13})^2 |\tilde{w}|^2 \\
& = \frac{\upsilon c_1}{2} |\delta z|^2 + \frac{1}{2 \upsilon c_1} (\alpha_{13} + \beta_{13})^2 |\tilde{w}|^2,
\end{align*}
where we used $\rho = 1/(\upsilon c_1)$.
Using the same value of $\rho$ for the rest of the terms, it can be verified that
$\Xi$ becomes positive definite if $k_\mathrm P > 0$ is sufficiently large such that
\begin{align} \label{eq:cond_k_p}
\hspace{-1em}|\Lambda_\mathrm P| \bar{L} + \frac{1}{\upsilon c_1} (\alpha_{13} + \beta_{13})^2 + \frac{\upsilon c_1}{2} < \frac{1}{8} k_\mathrm P \lambda_{1}(\Lambda_\mathrm P)^2
\end{align}
and $0 < \epsilon < \epsilon_1$ is sufficiently small such that
\begin{subequations} \label{eq:cond_eps}
\begin{align}
\kappa \epsilon |Y| |\Lambda_\mathrm P| |EW| & < \frac{1}{8} k_\mathrm P \lambda_{1}(\Lambda_\mathrm P)^2 \\
\frac{\epsilon }{2 \upsilon c_1} (N+2) \bar{L}^2 |Y|^2 & < \frac{1}{4} k_\mathrm I \alpha_{\delta \tilde{\xi}}.
\end{align}
\end{subequations}
In particular, let $k_\mathrm I$ and $\kappa$ be fixed positive values and let $k_\mathrm P$ be sufficiently large to satisfy \eqref{eq:cond_k_p} while $\epsilon = k_\mathrm I / k_\mathrm P$ is sufficiently small such that \eqref{eq:cond_eps} holds.
Then $\Xi > 0$ and the exponential convergence of $x^*$ follows from \eqref{eq:pf_lyap}.
For the convergence rate, suppose $k_\mathrm I = \kappa$.
Then, it can be checked that there exists $\bar{k}_\mathrm P^*$ and $\sigma^*$ such that for all $k_\mathrm P > \max(\sigma^* k_\mathrm I, \bar{k}_\mathrm P^*)$ the inequalities \eqref{eq:cond_k_p} and \eqref{eq:cond_eps} hold.
Hence, $-\bar{e}^\top \Xi \bar{e} \leq 0$ and \eqref{eq:pf_lyap_with_V} becomes
\begin{align*}
\dot{V}
& \leq - \min \left( 2(\mu - \upsilon), \frac{k_\mathrm P}{2\eta} \lambda_{1}(\Lambda_\mathrm P)^2, \frac{k_\mathrm I}{k_\mathrm P^2} \frac{1}{2\eta} \alpha_{\delta \tilde{\xi}}\right) V,
\end{align*}
where $\eta > 0$ is such that $\eta (|\tilde{w}|^2 + |\delta \tilde{\xi}|^2) \geq V_2(\tilde{w}, \delta \tilde{\xi})$.
In particular, $\eta$ can be chosen to be independent of $k_\mathrm P$ or $k_\mathrm I$ if $k_\mathrm P > \max(\sigma^* k_\mathrm I, \bar{k}_\mathrm P^*)$.
Since the Lyapunov function $V$ decays exponentially, this completes the proof.
\end{comment}
\begin{rmk}
By investigating the proof of Theorem \ref{thm:main_convergence} in detail, relationship between the convergence rate and the gains $k_\mathrm P$ and $k_\mathrm I$ can be inspected.
In particular, from the coefficients of the three terms in \eqref{eq:vdot}, it may be stated that, when $\kappa = k_\mathrm I$, increasing $k_\mathrm P$ with $k_\mathrm I$ kept fixed may degrade the convergence rate.
One can find that those three coefficients are proportional to $\mu - \upsilon$, $k_\mathrm I^2/k_\mathrm P$, and $k_\mathrm P$, respectively, and so, if $k_\mathrm P$ is too large compared to $\kappa = k_\mathrm I$, then the second coefficient gets smaller.
\end{rmk}
\begin{comment}
It can be seen from \eqref{eq:conv_rate} that the convergence rate is decomposed into three main components: 1) convergence rate related to the blended dynamics (e.g., $\mu - \upsilon$), 2) structure of the graph related to the proportional feedback (e.g., $\lambda_1(\Lambda_\mathrm P)$ and $k_\mathrm P$), and 3) structure of the graph related to the integral feedback (e.g., $\lambda_{q+1}({\mathfrak{D}}_\mathrm I)$).
Dependence of the convergence rate on the structure of the communication network is typical for distributed algorithms.
More interestingly, notice that the convergence rate also depends on the ratio $k_\mathrm I^2 / k_\mathrm P$.
This implies that only increasing the proportional gain $k_\mathrm P$ may hurt the performance and $k_\mathrm I$ must be increased as well to improve overall performance.
Finally, by increasing the coupling gain $k_\mathrm P$ and letting $k_\mathrm I = (1/\sigma^*) k_\mathrm P$, the overall convergence rate is ultimately limited by $\mu - \upsilon$.
This means that the convergence rate of \eqref{eq:pi_dyn_stack} can be made arbitrarily close to that of the blended dynamics.
\end{comment}
\section{Distributed Optimization Algorithms}\label{sec:opt}
In this section, we use Theorem \ref{thm:main_convergence} to obtain distributed algorithms for solving the minimization problem:
\begin{equation}\label{eq:prob}
\min_{w \in {\mathbb{R}}^n} F(w) = \frac1N \sum_{i=1}^N f_i(w)
\end{equation}
under the assumption:
\begin{ass}\label{asm:min}
The cost function $F(w)$ is continuously differentiable and strongly convex with parameter $\alpha$,\footnote{A function $F:{\mathbb{R}}^n \to {\mathbb{R}}$ is strongly convex with parameter $\alpha$ if $(\nabla F(w)-\nabla F(w_0))^\top (w-w_0) \ge \alpha |w-w_0|^2$ for all $w$ and $w_0$.} and $\nabla f_i$, $i \in {\mathcal{N}}$, are globally Lipschitz.
\end{ass}
It should be noted that the convexity of $f_i$ is not assumed while strong convexity of $F$ is required.
The proposed distributed algorithms are based on the PI-type coupling \eqref{eq:control}.
We present the case (A) in \eqref{eq:control} only because the case (B) yields the same convergence result as the case (A).
\subsection{Distributed Gradient Descent Method}
We first illustrate the utility of Theorem \ref{thm:main_convergence} by analyzing the classical distributed PI algorithm, which is given by
\begin{align}
\dot w_i &= - \nabla f_i(w_i) + k_\mathrm P \sum_{j \in {\mathcal{N}}_i} (w_j - w_i) + k_\mathrm I \sum_{j \in {\mathcal{N}}_i} (\xi_j - \xi_i) \notag \\
\dot \xi_i &= - k_\mathrm I \sum_{j \in {\mathcal{N}}_i} (w_j - w_i) . \label{eq:pi_simplified_ex}
\end{align}
This corresponds to the case (A) in \eqref{eq:control} where $\kappa = k_\mathrm I$ is used.
In particular, it is seen that the output $y_i = w_i$, or $E = I_n$.
This yields that $W = I_n$ and $Z$ is null, and so, $K = I_n$ and the blended dynamics \eqref{eq:blend_dyn} consists of $\bar{w}$ only and becomes
\begin{align} \label{eq:cgd_blend}
\dot{\bar{w}} & = -\frac{1}{N} \sum_{i=1}^N \nabla f_i(\bar{w}) = - \nabla F(\bar{w}).
\end{align}
Notice that \eqref{eq:cgd_blend} is exactly the centralized gradient descent method for minimizing \eqref{eq:prob}.
Under Assumption~\ref{asm:min}, the blended dynamics \eqref{eq:cgd_blend} has the exponentially stable equilibrium at the minimizer $w^*$ of $F$ with a rate $\alpha$.
Indeed, it follows with $V(\bar w) = (1/2)|\bar w - w^*|^2$ that $\dot V = -(\bar w - w^*)^\top (\nabla F(\bar w) - \nabla F(w^*)) \le -2\alpha V$ (since $\nabla F(w^*)=0$).
Thus, Assumption~\ref{asm:exp_stab} holds, and Theorem~\ref{thm:main_convergence} with $h_i = -\nabla f_i$ and Remark \ref{rem:gains} yield the following.
\begin{thm}\label{thm:1}
Under Assumption \ref{asm:min}, the distributed algorithm \eqref{eq:pi_simplified_ex} solves the problem \eqref{eq:prob} in the sense that (a) $w_i(t) \rightarrow w^*$, $\forall i \in {\mathcal{N}}$, when $k_\mathrm P$ is sufficiently large and $k_\mathrm I > 0$, and (b) its convergence rate can be made arbitrarily close to the convergence rate of the centralized gradient descent method if $k_\mathrm P=k_\mathrm I^{3/2}$ and $k_\mathrm I$ is sufficiently large.
\end{thm}
\subsection{Distributed Heavy-ball Method with State Coupling}
The argument used to obtain Theorem \ref{thm:1} may inspire a design paradigm of multi-agent system.
That is, the node dynamics of individual agents (e.g., \eqref{eq:pi_simplified_ex}) are designed such that their blended dynamics (e.g., \eqref{eq:cgd_blend}) is the system that performs the desired task.
In this sense, let us suppose that we want to solve \eqref{eq:prob} by the heavy-ball method \citep{Qian1999}:
\begin{align} \label{eq:cent_hb}
\begin{split}
\dot{\bar{w}} &= \bar{z} \\
\dot{\bar{z}} &= -2 \sqrt{\alpha} \bar{z} - \frac{1}{N} \sum_{i=1}^N \nabla f_i(\bar{w})
\end{split}
\end{align}
and raise the question what is a suitable node dynamics that solves \eqref{eq:prob} in a distributed way.
An immediate answer to the question is:
\begin{align}
\begin{bmatrix} \dot{w}_i \\ \dot{z}_i \end{bmatrix}
& = \begin{bmatrix}
z_i \\ -2 \sqrt{\alpha} z_i - \nabla f_i(w_i)
\end{bmatrix} \notag \\
&\quad + k_\mathrm P \sum_{j \in {\mathcal{N}}_i} \left(
\begin{bmatrix}
w_j \\ z_j
\end{bmatrix} - \begin{bmatrix}
w_i \\ z_i
\end{bmatrix}
\right) + k_\mathrm I \sum_{j \in {\mathcal{N}}_i} (\xi_j - \xi_i) \notag \\
\dot{\xi}_i & = - k_\mathrm I \sum_{j \in {\mathcal{N}}_i} \left(
\begin{bmatrix}
w_j \\ z_j
\end{bmatrix} - \begin{bmatrix}
w_i \\ z_i
\end{bmatrix}
\right) \label{eq:dist_hb_full}
\end{align}
where $w_i \in {\mathbb{R}}^n$, $z_i \in {\mathbb{R}}^n$, and $\xi_i \in {\mathbb{R}}^{2n}$.
Since the vectors $[w_i;z_i]$ are communicated, we have $E = I_{2n}$.
Therefore, $W = I_{2n}$ and $Z$ is null, and so, the blended dynamics of \eqref{eq:dist_hb_full} is nothing but \eqref{eq:cent_hb}.
Convergence property of \eqref{eq:cent_hb} is well studied by \citet{Siegel2019}.
That is, the equilibrium point of \eqref{eq:cent_hb} is given by $[w^*; 0]$, where $w^*$ is the minimizer of \eqref{eq:prob}, and the solution of \eqref{eq:cent_hb} converges to the equilibrium exponentially fast with the rate $\sqrt{\alpha}/2$.\footnote{In Lemma \ref{lem:heavy_ball_reduced_stab} of the next subsection, we extend the proof of \citep{Siegel2019}, which can also be used for justifying this claim.}
Hence, Assumption \ref{asm:exp_stab} holds, and Theorem \ref{thm:main_convergence} yields the following.
\begin{thm}\label{thm:2}
Under Assumption \ref{asm:min}, the states $w_i$ and $z_i$ of \eqref{eq:dist_hb_full} converge to the equilibrium of \eqref{eq:cent_hb}, i.e., $w_i(t) \rightarrow w^*$ and $z_i(t) \rightarrow 0$ when $k_\mathrm P$ is sufficiently large and $k_\mathrm I > 0$.
Moreover, the convergence rate of the distributed algorithm \eqref{eq:dist_hb_full} can be made arbitrarily close to $\sqrt{\alpha}/2$ with $k_\mathrm P = k_\mathrm I^{3/2}$ and sufficiently large $k_\mathrm I$.
\end{thm}
The heavy-ball method is known to outperform the gradient descent method in view of convergence rate for a class of problems (e.g., when $0 < \alpha \ll 1$).
This property is inherited to the distributed heavy-ball method because the loss of convergence rate can be arbitrarily small.
\subsection{Distributed Heavy-ball Method with Output Coupling}
If we let $x_i = [w_i;z_i]$ in view of \eqref{eq:agent_dyn}, it is seen that the algorithm \eqref{eq:dist_hb_full} communicates the full state $x_i$ of size $2n$, as well as $\xi_i$ of size $2n$, because $E = I_{2n}$.
However, we can reduce the amount of communication if we do not communicate $z_i$, or if we let $E = [I_n, 0_{n \times n}]$.
In fact, we propose the following node dynamics in this subsection:
\begin{align}
\begin{bmatrix}
\dot{w}_i \\ \dot{z}_i
\end{bmatrix}
& = \begin{bmatrix}
z_i \\ -2\sqrt{\alpha} z_i - \nabla f_i(w_i)
\end{bmatrix} \notag \\
&\quad + k_\mathrm P
\begin{bmatrix}
I_n \\ 0_n
\end{bmatrix}
\sum_{j \in {\mathcal{N}}_i} \left(
w_j - w_i
\right) + k_\mathrm I \begin{bmatrix}
I_n \\ 0_n
\end{bmatrix}\sum_{j \in {\mathcal{N}}_i} (\xi_j - \xi_i) \notag \\
\dot{\xi}_i & = - k_\mathrm I \sum_{j \in {\mathcal{N}}_i} \left(
w_j - w_i
\right) \label{eq:dist_hb_re}
\end{align}
where $\xi_i \in {\mathbb{R}}^n$.
The algorithm \eqref{eq:dist_hb_re} communicates $w_i$ and $\xi_i$ only, whose sizes are both $n$.
(If the case (B) of \eqref{eq:control} is used, then we can communicate $w_i$ only. However, in this case, initialization $\sum_{i=1}^N \xi_i(0)=0$ is necessary.)
It follows from $E = [I_n, 0_{n \times n}]$ that $W = [I_n; 0_{n \times n}]$ and $Z = [0_{n \times n}; I_n]$.
This leads to the blended dynamics of \eqref{eq:dist_hb_re} as
\begin{align} \label{eq:dist_hb_reduced}
\begin{split}
\dot{\bar{w}} &= \frac{1}{N} \sum_{i=1}^N z_i \\
\dot{z}_i &= -2 \sqrt{\alpha} z_i - \nabla f_i(\bar{w}), \quad i \in {\mathcal{N}} .
\end{split}
\end{align}
The following lemma asserts that Assumption \ref{asm:exp_stab} holds.
\begin{lem} \label{lem:heavy_ball_reduced_stab}
Under Assumption \ref{asm:min}, the blended dynamics \eqref{eq:dist_hb_reduced} has the unique equilibrium point $(w^*,z_1^*,\dots,z_N^*)$ where
\begin{equation}\label{eq:stars}
w^* = \argmin_{w} F(w) \quad \text{and} \quad z_i^* = -\frac1{2\sqrt{\alpha}} \nabla f_i(w^*)
\end{equation}
and the equilibrium is exponentially stable with convergence rate $\sqrt{\alpha}/2$.
\end{lem}
\begin{pf}
Let $\bar{z}:= (1/N)\sum_{i=1}^N z_i$ so that
\begin{align*}
\dot {\bar w} &= \bar z \\
\dot {\bar z} &= -2\sqrt{\alpha}\bar z - \nabla F(\bar w).
\end{align*}
Following the derivation of \citep{Siegel2019}, let the function $V(\bar{w},\bar{z})$ be defined as
\begin{align*}
V := F(\bar{w}) - F(w^*) + \frac{1}{2} \left| \sqrt{\alpha}(\bar{w} - w^*) + \bar{z} \right|^2.
\end{align*}
Then, we have
\begin{align*}
\dot V &= \nabla F(\bar w) \bar z + (\sqrt{\alpha}(\bar w - w^*) + \bar z)(-\sqrt{\alpha}\bar z - \nabla F(\bar w)) \\
&= -\sqrt{\alpha} \nabla F(\bar w) (\bar w - w^*) - \alpha \bar z (\bar w - w^*) - \sqrt{\alpha} \bar z^2 \\
&\le -\sqrt{\alpha} \left(F(\bar w)-F(w^*)+\frac{\alpha}{2} |\bar w - w^*|^2 \right) \\
&\quad - \alpha \bar z (\bar w - w^*) - \sqrt{\alpha} \bar z^2
\end{align*}
in which, strong convexity of $F$ is used.\footnote{Strong convexity of a function $F:{\mathbb{R}}^n \to {\mathbb{R}}$ with parameter $\alpha$ is equivalently characterized as: $F(x) \ge F(x_0) + \nabla F(x_0) (x-x_0) + (\alpha/2)|x-x_0|^2$, $\forall x, x_0$. Here, we put $x=w^*$ and $x_0=\bar w$.}
Now, pick an arbitrary positive number $\upsilon < \sqrt{\alpha}$.
Since $F(w) \ge F(w^*)$ for all $w$, we have
\begin{align*}
\dot V &\le -\left(\sqrt{\alpha}-\upsilon\right) \left(F(\bar w)-F(w^*)+\frac{\alpha}{2} |\bar w - w^*|^2 \right) \\
&\quad - \frac{\upsilon\alpha}{2}|\bar w - w^*|^2 - \alpha \bar z (\bar w - w^*) - \sqrt{\alpha} \bar z^2 \\
&\le -\left(\sqrt{\alpha}-\upsilon\right) \left( F(\bar w)-F(w^*) + \frac12 |\sqrt{\alpha} (\bar w - w^*) + \bar z|^2 \right) \\
&- \frac{\upsilon\alpha}{2}|\bar w - w^*|^2 - \upsilon\sqrt{\alpha}(\bar w - w^*) \bar z + \left(\frac{\sqrt{\alpha} - \upsilon}{2} - \sqrt{\alpha}\right) \bar z^2
\end{align*}
and, by Young's inequality for $\upsilon\sqrt{\alpha}(\bar w - w^*) \bar z$,
\begin{align*}
\dot V &\le -(\sqrt{\alpha}-\upsilon) V - \frac{\upsilon\alpha}{2}|\bar w - w^*|^2 \\
&\quad + \frac{\upsilon\alpha}{4}|\bar w - w^*|^2 + \left(\upsilon + \frac{\sqrt{\alpha} - \upsilon}{2} - \sqrt{\alpha}\right) \bar z^2 \\
&\le -(\sqrt{\alpha}-\upsilon) V - \frac{\upsilon\alpha}{4}|\bar w - w^*|^2 .
\end{align*}
Now, let $\widetilde z := R^\top ([z_1;\cdots;z_N]-[z_1^*;\cdots;z_N^*])$.
Then,
\begin{align*}
\dot {\widetilde z} &= -2\sqrt{\alpha} R^\top z - R^\top [\nabla f_1(\bar w); \cdots; \nabla f_N(\bar w)] \\
&= -2 \sqrt{\alpha} \widetilde z + R^\top \begin{bmatrix} \nabla f_1(w^*)-\nabla f_1(\bar w) \\ \vdots \\ \nabla f_N(w^*)-\nabla f_N(\bar w) \end{bmatrix}
\end{align*}
in which, \eqref{eq:stars} is used.
Now, let $L$ be the maximum of all Lipschitz coefficients of $\nabla f_i$'s.
With
\begin{align} \label{eq:lyap_W}
W = V + \frac{\gamma}{2} \widetilde z^\top \widetilde z
\end{align}
we have (recalling $|R| = 1$)
\begin{align*}
\dot W &\le -(\sqrt{\alpha} - \upsilon) V - \frac{\upsilon\alpha}{4}|\bar w - w^*|^2 - 2\sqrt{\alpha}\gamma |\widetilde z|^2 \\
&\qquad + \gamma L |\bar w - w^*| |\widetilde z| \\
&\le -(\sqrt{\alpha} - \upsilon) V - 2\sqrt{\alpha}\gamma |\widetilde z|^2 + \frac{\gamma^2 L^2}{\upsilon\alpha} |\widetilde z|^2 .
\end{align*}
With a sufficiently small $\gamma$ such that
$$\frac{3\sqrt{\alpha}+\upsilon}{2} - \frac{\gamma L^2}{\upsilon\alpha} \ge 0$$
we have
$$\dot W \le -(\sqrt{\alpha} - \upsilon) (V + (\gamma/2) |\widetilde z|^2) = -(\sqrt{\alpha} - \upsilon)W.$$
It can be checked that $W$ is a positive definite function in terms of $\bar{w} - w^*$ and $\widetilde{z}$.
Hence, we obtain that the unique equilibrium point is exponentially stable with a rate $(\sqrt{\alpha} - \upsilon) / 2$.
Since the choice of $\upsilon$ is arbitrary, we conclude that the convergence rate is $\sqrt{\alpha}/2$, which completes the proof.
\end{pf}
\begin{thm}\label{thm:3}
Under Assumption \ref{asm:min}, the states $w_i$ and $z_i$ of \eqref{eq:dist_hb_re} converge to the equilibrium in \eqref{eq:stars}, i.e., $w_i(t) \rightarrow w^*$ and $z_i(t) \rightarrow -\nabla f_i(w^*)/(2\sqrt{\alpha})$ when $k_\mathrm P$ is sufficiently large and $k_\mathrm I > 0$.
Moreover, the convergence rate of the distributed algorithm \eqref{eq:dist_hb_re} can be made arbitrarily close to $\sqrt{\alpha}/2$ with $k_\mathrm P = k_\mathrm I^{3/2}$ and sufficiently large $k_\mathrm I$.
\end{thm}
\begin{comment}
\begin{pf}
For the proof, notice that the function $W$ defined at \eqref{eq:lyap_W} can be taken as the Lyapunov function $\bar{V}$ obtained at the start of the proof of Theorem \ref{thm:main_convergence}, i.e., $W$ is a Lyapunov function for the blended dynamics of \eqref{eq:dist_hb_re} and guarantees the convergence rate that is arbitrarily close to $\sqrt{\alpha}/2$.
With this, rest of the proof can be proceeded as in the proof of Theorem \ref{thm:main_convergence}.
\end{pf}
\end{comment}
\begin{table*}[t]
\centering
\caption{}
\label{tab:1}
\begin{tabular}{lccccc}
\toprule
\multirow{2}{*}{Study} & Require & Initialization & Required & \multirow{2}{*}{Approach} & Convergence rate \\
& convexity of $f_i$ & -free & communication & & (if strongly convex) \\ \midrule
\citet{Wang2010} & Yes & Yes & $2n$ & Not available & Asymptotic\\ \midrule
\citet{Kia2015} & Yes & No & $n$ & Lyapunov & Exponential \\ \midrule
\citet{Yang2017} & Yes & No & $n$ & LaSalle & Asymptotic\\ \midrule
\citet{Hatanaka2018} & Yes & Yes & $2n$ & LaSalle & Asymptotic\\ \midrule
\citet{Xin2019} & Yes & No & $2n$ & Lyapunov & Exponential \\ \midrule
\citet{Qu2020} & Yes & No & $3n$ & Lyapunov & Exponential \\ \midrule
Proposed algorithm (A), or \eqref{eq:dist_hb_re} & No & Yes & $2n$ & Lyapunov & Exponential \\ \midrule
Proposed algorithm (B) & No & No & $n$ & Lyapunov & Exponential \\
\bottomrule
\end{tabular}
\end{table*}
Compared with the existing algorithms, the classical PI algorithms (e.g., \citet{Kia2015,Yang2017}) communicate $n$-dimensional information to its neighbors while requiring a specific initial condition (like the case (B) of \eqref{eq:control}).
Algorithms are proposed that do not require specific initializations (e.g., \citet{Hatanaka2018,Wang2010}), but these communicate $2n$-dimensional information.
Additionally, most works only prove asymptotic convergence while exponential convergence is a preferred property.
An exception is \citep{Kia2015} which proves exponential convergence but accelerated methods such as heavy-ball method are not studied.
For discrete-time algorithms, the distributed Nesterov method studied by \citet{Qu2020} communicates $3n$-dimensional information and requires initialization.
Distributed heavy-ball method proposed by \cite{Xin2019} communicates $2n$-dimensional information but still requires a specific initial condition.
Additionally, convergence rates of discrete-time algorithms did not match the rate of the corresponding centralized algorithms.
The proposed algorithm \eqref{eq:dist_hb_re} implements the distributed heavy-ball method while only communicating $2n$-dimensional information and is initialization-free.
In addition, if the proposed algorithm is implemented by the case (B) of \eqref{eq:control}, then they communicate only $n$-dimensional information while it is no longer initialization-free.
In both cases of (A) and (B), we recover the convergence rate of the centralized algorithm arbitrarily closely.
These discussions are summarized in Table \ref{tab:1}.
\section{Numerical Experiments} \label{sec:simulation}
For a numerical simulation, let us consider a distributed quadratic problem with $N=12$ agents.
The cost function of each agent is given by $f_i(w) = w^\top A_i w + b_i^\top w$ where $A_i \in {\mathbb{R}}^{6 \times 6}$ is a symmetric matrix and $b_i \in {\mathbb{R}}^{6}$.
It is supposed that $\sum_{i=1}^N A_i > 0$, while each $A_i$ may be indefinite.
The condition number $C$ of $(1/N)\sum_{i=1}^N f_i(x)$ is defined as the ratio of the maximum to the minimum eigenvalue of $(1/N)\sum_{i=1}^N A_i$.
Matrices $A_i$ and $b_i$ are generated randomly such that the maximum eigenvalue of $(1/N)\sum_{i=1}^N A_i$ is approximately $1.0$, while having a large condition number (i.e., small $\alpha$) in order to see the effect of heavy-ball methods.
The communication graph is generated randomly using the Erd\H{o}s-R\'{e}nyi model with each edge having a probability of $0.2$.
Simulation results\footnote{One can obtain and run the simulation code at {\tt https://doi.org/10.24433/CO.9155541.v1}.} are shown in Fig. \ref{fig:result_1}, where average error to the optimal value is plotted in vertical axis (i.e., $(1/N) \sum_{i=1}^N |w_i(t) - w^*|$ for distributed algorithms and $|w(t) - w^*|$ for centralized algorithms) and horizontal axis is the time $t$.
Two heavy-ball methods, by the state coupling in \eqref{eq:dist_hb_full} (denoted by `HB-State') and by the output coupling in \eqref{eq:dist_hb_re} (denoted by `HB-Output'), are implemented which correspond to the case (A).
Thus, \eqref{eq:dist_hb_full} communicates $4n$-dimensional information whereas \eqref{eq:dist_hb_re} communicates $2n$-dimensional information.
Results are compared with the PI algorithm (denoted by `PI') of \citep{Hatanaka2018}, the centralized gradient descent (`CGD'), and the centralized heavy-ball method \eqref{eq:cent_hb} (denoted by `CHB').
First graph of Fig. \ref{fig:result_1} shows the result when $k_\mathrm P = 0.3$ and $k_\mathrm I = 0.15$.
It can be seen that the heavy-ball methods outperform the gradient descent methods which have slow performance.
We can also see that the proposed distributed heavy-ball algorithms converge to the optimal value but do not recover the convergence rate of the centralized heavy-ball method.
On the other hand, the second graph of Fig. \ref{fig:result_1} shows the result when $k_\mathrm P = 1.0$ and $k_\mathrm I = 0.5$.
Since sufficiently high gains are used, we see that the performance of the proposed algorithms recover the convergence rate of the centralized heavy-ball method.
\begin{figure}
\centering
\includegraphics[scale=0.42]{sim_result_case_low.png}
\vspace{1em}
\includegraphics[scale=0.42]{sim_result_case_high.png}
\caption{Simulation results with low gains (top) and high gains (bottom).}
\label{fig:result_1}
\end{figure}
\section{Conclusion} \label{sec:conclusion}
We have studied distributed continuous-time algorithms, based on PI-type couplings, to solve distributed optimization problems when the global cost function $(1/N)\sum_{i=1}^N f_i(w)$ is $\alpha$-strongly convex while individual $f_i$'s are not necessarily convex.
A concept of blended dynamics is proposed to analyze the system and it is shown that the property of the blended dynamics is important in characterizing the behavior of the overall system.
It is also shown that the distributed algorithm recovers the convergence rate of the blended dynamics with suitably chosen coupling gains.
Using these results, distributed algorithms are constructed from the centralized algorithms, and in particular, distributed heavy-ball methods are proposed which achieve the convergence rate of $\sqrt{\alpha}/2$ arbitrarily closely.
Numerical simulations are done to verify the performance of the proposed algorithms.
The proposed Theorem \ref{thm:main_convergence} may also be used for designing other distributed algorithms.
|
\section{Introduction}
The Poisson-Dirichlet distribution $\mathrm{PD}(0,\theta),\theta>0,$ is proposed by Kingman \cite{King1} in the simplex $\overline{\nabla}_{\infty}=\{x\in[0,1]^{\infty}\mid x_{1}\geq x_{2}\geq\cdots\geq0,\sum_{i=1}^{\infty}x_i\leq 1\}$. Pitman generalized the one-parameter Poisson Dirichlet distribution $\mathrm{PD}(0,\theta)$ to the two-parameter Poisson-Dirichlet distribution $\mathrm{PD}(\alpha,\theta),\alpha\in(0,1),\theta+\alpha>0$, in \cite{EP}. They are also the representing measure of the Ewens-Pitman partition structure. Partition structure coined by Kingman in \cite{King} is an exchangeable partition distribution $M$ of the set $\mathbb{N}=\{1,2,\cdots,n,\cdots\}$. Due to its exchangeability, the restriction of $M$ to $\{1,2,\cdots,n\}$ is $M_n(\eta)=\dim(\eta)\varphi(\eta)$, where $\eta=(\eta_1,\cdots,\eta_l), |\eta|=:\sum_{i=1}^l\eta_i=n,$ is an integer partition of $n$. Moreover, $\varphi(\eta)$ is the probability of a single partition with cluster sizes $\eta$. Due to exchangeability, two partitions have the same probability as long as their cluster sizes are the same. So $\dim(\eta)$ is the multiplicities of such partitions with cluster size $\eta$. Then the family of distributions $\{M_{n},n\geq1\}$ will satisfy a natural consistent condition and they are uniquely determined by a representing measure in the Kingman simplex $\overline{\nabla}_{\infty}$ due to its exchangeability and the de Finetti theorem.
Let $\Gamma_n$ be the totality of integer partitions of $n$, then $\Gamma=\cup_{n\geq0}^{\infty}\Gamma_n$ exhausts all integer partitions, and $\Gamma_{0}=\emptyset$ is treated as an empty partition. Usually $\Gamma$ can be the vertex set of a graded branching diagram $(\Gamma,\chi)$ where $\chi(\eta,\omega)$ assigns positive weight to an edge joining $\eta$ and $\omega$, and $0$ otherwise (see Figure \ref{bd}). Each integer partition is represented by a Young diagram (see Figure \ref{yd}), we say $\eta\subset\zeta$ if the Young diagram of $\eta$ is contained in the Young diagram of $\zeta$. There are various kinds of edge weights. In the algebra $\cal{A}$ of symmetric functions with variables $(x_1,\cdots,x_n,\cdots)$, there are various kinds of linear bases $\{f_{\eta},\eta\in\Gamma\}$(refer to \cite{Mac} or Appendix), such as the monomial functions, the Shur functions and the Jack functions. These functions usually satisfy the Pieri formula
$$
\left(\sum_{i=1}^{\infty}x_i\right)f_{\eta}=\sum_{\eta\subset\zeta, |\eta|+1=|\zeta|}\chi(\eta,\zeta)f_{\zeta},
$$
where $\eta$ can be obtained from $\zeta$ by removing a box. The edge weights are chosen to be the coefficients in the above Pieri formula. The weights from the monomial functions, the Shur functions and the Jack functions define the Kingman graph, the Young graph and the Jack graph respectively. The specialization of the symmetric functions can be regarded as an algebra homomorphism $\Phi: \cal{A}\to \mathbb{R}$. In particular, a specialization mapping base functions to positive values will determine a positive harmonic function $\varphi(\eta)=\Phi(f_{\eta})$ on $\Gamma$, satisfying
$$
\varphi(\eta)=\sum_{\eta\subset\zeta, |\eta|+1=|\zeta|}\chi(\eta,\zeta)\varphi(\zeta).
$$
Under pointwise convergence topology, the space $\cal{C}$ of all positive harmonic functions becomes compact and convex. Therefore, one can expect that $\varphi(\eta)=\int_{E} K(\eta,x)\mu(dx)$ where $E$ is the Martin boundary of $\cal{C}$, $K(\eta,x)$ is the extremity and $\mu(dx)$ is a probability measure. For the Kingman graph, this representation is the Kingman's one-to-one correspondence of the partition structures \cite{King} and $K(\eta,x)$ is the continuous extension of the monomial functions. For the Jack graph, the representation is also established in \cite{KOO}, and $K(\eta,x)$ is the extended Jack functions. The $Z$-partition structure is a partition structure defined on the Jack graph, and its representing measure is called $Z$-measure, denoted as $\mathcal{Z}(z,z^{\prime},\vartheta), z,z^{\prime}\in\mathbb{C},\vartheta>0$.
Petrov established the two-parameter Poisson-Dirichlet diffusion with the stationary distribution $\mathcal{PD}(\alpha,\theta)$ through an up-down Markov chain on the Kingman graph \cite{Pet}.
Similarly, Olshanski \cite{Ol_2} has established a reversible diffusion on the Thoma simplex
\begin{align*}
\Omega=\{(\alpha,\beta)\in[0,1]^{\infty}\times[0,1]^{\infty}\mid& \alpha_1\geq\alpha_2\geq\cdots\geq0,\\
&\beta_1\geq\beta_2\geq\cdots\geq0,\sum_{i=1}^{\infty}\alpha_i+\sum_{j=1}^{\infty}\beta_j\leq1\}.
\end{align*}
Its stationary distribution is the measure $\mathcal{Z}(z,z^{\prime},\vartheta)$. In this paper, we will use the $Z$-measure diffusion to refer to the diffusion obtained by Olshanski in \cite{Ol_2}.
Interestingly, both types of diffusions have similar explicit transition densities, and the spectrums of their generators are also the same $\{0,\frac{n(n-1+\theta)}{2}\mid n\geq2\}$ where $\theta=\frac{zz^{\prime}}{\vartheta}>0$ for the $Z$-measure diffusion.
Recently, spectral expansion is applied to the $Z$-measure diffusion again in \cite{SY_K} to derive its explicit transition density. Similar method has previously been used to establish the transition density of the two-parameter Poisson-Dirichlet diffusion in \cite{FSWX}. In fact, this method was first adopted by Ethier in \cite{ET}. Surprisingly, by rearranging the density in \cite{FSWX}, Zhou \cite{Zhou} yields the following expression
\begin{equation}
p(t,x,y)=\widetilde{d_1^{\theta}}(t)+\sum_{n=2}^{\infty}d_n^{\theta}(t) p_{n}(x,y),\label{ex}
\end{equation}
where $p_n(x,y)$ is a transition kernel and $\{\widetilde{d}_1^{\theta}(t),d_n^{\theta}(t),n\geq2\}, \theta>-1$ is the distribution of a pure death process related to the Kingman coalescent in \cite{Zhou}.
In this paper, we will rearrange the transition density of the $Z$-measure diffusion as the treatment in \cite{Zhou}. We obtain that the $Z$-measure diffusion has a similar expression
\begin{equation}
q(t,\sigma,\omega)=\widetilde{d_1^{\theta}}(t)+\sum_{n=2}^{\infty}d_n^{\theta}(t) \mathcal{K}_{n}(\sigma,\omega)\label{zf},
\end{equation}
where $\mathcal{K}_{n}(\sigma,\omega)$ is also a transition kernel and $\{\widetilde{d}_1^{\theta}(t),d_n^{\theta}(t),n\geq2\}, \theta>0$ is the same as the coefficients in (\ref{ex}).
The next natural question would be why their transition densities have the same coefficients. This question has been resolved for the two-parameter Poisson-Dirichlet diffusion in \cite{GSMZ} by a dual process method.
In order to understand why the $Z$-diffusion also has the similar coefficients in its density expression (\ref{zf}), we will also apply the dual process method to the $Z$-measure diffusion $Y_t$. But we can only find the dual process of the the $Z$-measure diffusion when $\vartheta=1$ because we relies on the equation (5) in \cite{Ol_2}. When $\vartheta\neq 1$, similar equation is not known. For $\vartheta=1$, the dual process $\mathcal{D}_{t}$ is also a partition-valued jump process characterized by the generator
$$
\mathcal{L}_1^{J}f(\eta)=-\frac{n(n-1+\theta)}{2}f(\eta)+\frac{n(n-1+\theta)}{2}\sum_{\zeta\subset\eta,|\eta|=|\zeta|+1} p_1^{\downarrow}(\eta,\zeta)f(\zeta),~ \mathcal{L}_1^{J}1=0,
$$
where $p_1^{\downarrow}(\eta,\zeta)$ is the transition probability of the down Markov chain in the Jack graph discussed in \cite{Ol_2}, $f$ is the continuous function on the Thoma simplex. The dual relation is defined through a bivariate function $F(\eta, \omega)=\frac{s_{\eta}^{o}(\omega)}{\mathbb{E}_{z,z^{\prime},1}s_{\eta}^{o}}$ which is the normalized kernel in the representation of the Young graph. The notation $\mathbb{E}_{z,z^{\prime},1}$ should always be interpreted as the expectation with respect to the $Z$-measure $\mathcal{Z}(z,z^{\prime},1)$. We show that duality relation reads as $\mathbb{E}_{\omega}F(\eta, Z_t)=\mathbb{E}_{\eta}F(\mathcal{D}_{t}, \omega)$. Because the distribution of $\mathcal{D}_{t}$ is easier to calculate, then the expression (\ref{zf}) can be obtained. Now the radial process $|\mathcal{D}_{t}|$ of the dual process $\mathcal{D}_{t}$ will be exactly the same as that in the two-parameter Poisson-Dirichlet diffusion. This will eventually determine the coefficients in the density expression (\ref{zf}). Though when $\vartheta\neq1$, we can not find the find dual process, the result when $\vartheta=1$ encourage us to conjecture that the dual process of the $Z$-measure diffusion should also be a partition-valued jump process whose jump rate is $\frac{n(n-1+\theta)}{2}$ and its embedded chain is the down Markov chain in \cite{Ol_2}.
The plan of the paper is as follows. In section 2, we will introduce the branching diagrams and the up and down Markov chains. In section 3, we will talk about the $Z$-measure diffusion and its transition density. In section 4, we will use the dual process method to derive the transition density of the $Z$-measure diffusion when $\vartheta=1$. In the last section, a few conjectures will be presented.
\section{Branching diagram and up-down Markov chain}
\subsection{Branching Diagram}
For $n,l\in\mathbb{Z}_{+}$, $\eta=(\eta_1,\cdots,\eta_l)\in\mathbb{N}^{l}$ is called an integer partition of $n$ if $\eta_1\geq\eta_2\geq\cdots\eta_l>0$ and $|\eta|:=\sum_{i=1}^{l}\eta_i=n$. Define $\alpha_i(\eta)=\#\{1\leq j\leq l\mid \eta_j=i\},1\leq i\leq n$, then $(\alpha_1(\eta),\cdots,\alpha_n(\eta))$ is a different representation of the partition $\eta$. Denote $\Gamma_n$ to be the set of all integer partitions of $n$, then $\Gamma=\cup_{n\geq0}\Gamma_n$ is the set of all integer partitions, where $\Gamma_0=\{\emptyset\}$ and $\emptyset$ is an empty partition. An integer partition $\eta=(\eta_1,\cdots,\eta_l)$ may also be represented by the Young diagram (Figure \ref{yd}) defined by attaching boxes at position $(i,j)$ where $1\leq i\leq l, 1\leq j\leq \eta_i$. Here the row number increases from top to bottom and the column number increases from left to right.
\begin{figure}[htbp]
\begin{center}
\ytableausetup{boxsize=1.5em}
\ydiagram{5,4,1}
\caption{Young diagram $\eta=(5,4,1)$}
\label{yd}
\end{center}
\end{figure}
In this paper, we will interchangeably use $\eta$ to represent either an integer partition or its Young diagram. We define the diagonal line of the Young diagram $\eta$ as the set of boxes $\{(i,i)\mid i\leq \eta_{i},\eta^{\prime}_{i}\}$ and $r$ to be the length of the diagonal line of $\eta$. The transposition of $\eta$ with respect to its diagonal line will give us a new partition $\eta^{\prime}$ called conjugate of $\eta$. We say $\eta\subset \nu$ if the Young diagram of $\eta$ is contained in the Young diagram of $\nu$. Then $\subset$ is a partial order in $\Gamma$. For each box $b=(i,j)$ in the partition $\eta$, we will define its arm length as $a(b)=\eta_i-j$ and its leg length as $l(b)=\eta^{\prime}_j-i$. A partition $\eta$ may also be represented by its Frobenius coordinates $\widetilde{\eta}= (a(1,1)+\frac{1}{2},\cdots,a(r,r)+\frac{1}{2}; l(1,1)+\frac{1}{2},\cdots,l(r,r)+\frac{1}{2})$.
A branching diagram is a graded graph $(\Gamma,\chi)$ (Figure \ref{bd}), where $\Gamma=\cup_{n\geq0}\Gamma_n$ is the vertex set. There is an edge joining $\eta\in\Gamma_n$ and $\nu\in\Gamma_{n+1}$ if and only if $\eta\subset \nu$, and their edge weight is $\chi(\eta,\nu)$.
\begin{figure}
\begin{center}
\ytableausetup
{boxsize=0.5em}
\begin{tikzpicture}[scale=0.5]
\node(0) at (-0.5,0) {$\emptyset$};
\node(1) at (1,0) {\ydiagram{1}};
\node(2a)at (2.5,0.7){\ydiagram{2}};
\node(2b) at (2.5,-0.7){\ydiagram{1,1}};
\node(3a) at (4.5, 1.5) {\ydiagram{3}};
\node(3b) at (4.5,0) {\ydiagram{2,1}};
\node(3c)at (4.5,-1.5){\ydiagram{1,1,1}};
\node(4a) at (7.5,3){\ydiagram{4}};
\node(4b) at (7.5,1.5) {\ydiagram{3,1}};
\node(4c) at (7.5,0) {\ydiagram{2,2}};
\node(4d)at (7.5,-1.4){\ydiagram{2,1,1}};
\node(4e) at (7.5,-3){\ydiagram{1,1,1,1}};
\path(0) edge (1);
\path(1) edge (2a);
\path(1) edge (2b);
\path(2a) edge (3a);
\path(2a) edge (3b);
\path(2b) edge (3b);
\path(2b) edge (3c);
\path(3a) edge (4a);
\path(3a) edge (4b);
\path(3b) edge (4b);
\path(3b) edge (4c);
\path(3b) edge (4d);
\path(3c) edge(4d);
\path(3c) edge(4e);
\end{tikzpicture}
\caption{Branching Diagram}
\label{bd}
\end{center}
\end{figure}
When the edge weights are determined by the coefficients in the Pieri formula of the monomial symmetric functions, we will have the Kingman graph. Its edge weight is defined as $\chi^{K}(\eta,\zeta)=\alpha_{\zeta_{i}}(\zeta)$ if $\zeta-\eta=(i,j)$, i.e. $\zeta$ can be obtained from $\eta$ by attaching a box at $(i,j)$. If we choose the edge weight to be the coefficients in the Pieri formula of the Jack functions, then we end up with the Jack graph, whose edge weight, denoted as $\chi^{J}_{\vartheta}(\eta,\zeta)$, will be
$$
\chi^{J}_{\vartheta}(\eta,\zeta)=\prod_{b}\frac{(a(b)+(l(b)+2)\vartheta)(a(b)+1+l(b)\vartheta)}{(a(b)+1+(l(b)+1)\vartheta)(a(b)+(l(b)+1)\vartheta)}
$$
where $b$ runs over all boxes in the $j$-th column of $\eta$ if $\zeta-\eta=(i,j)$. When $\vartheta=1$, then $\chi^{J}_{1}=1$ and the Jack graph reduces to the Young graph.
We define the weight of a path $t: \eta=\eta(0)\subset\cdots\subset\eta(k)=\zeta$ as $\prod_{i=0}^{k-1}\chi(\eta(i),\eta(i+1))$. Then we define the total weight between $\eta$ and $\zeta$ in the Kingman graph as $\dim^K(\eta,\zeta)=\sum_{\eta=\eta(0)\subset\cdots\subset\eta(k)=\zeta}\prod_{i=0}^{k-1}\chi(\eta(i),\eta(i+1))$. Similarly, the total weight between $\eta$ and $\zeta$ in the Jack graph will be $\dim^J_{\vartheta}(\eta,\zeta)=\sum_{\eta=\eta(0)\subset\cdots\subset\eta(k)=\zeta}\prod_{i=0}^{k-1}\chi(\eta(i),\eta(i+1))$.
In particular, when $\eta=\emptyset$, we regard $\dim^K(\nu)=\dim(\emptyset,\nu)$ as the total weight of $\nu$ in the Kingman graph and $\dim^J_{\vartheta}(\nu)=\dim^{J}_{\vartheta}(\emptyset,\nu)$ as the total weight of $\nu$ in the Jack graph. Naturally, we have
\begin{equation}\label{rw}
\dim^{K}(\zeta)=\sum_{\eta\subset\zeta,|\zeta|=|\eta|+1}\chi^J(\eta,\zeta)\dim^K(\eta)
\end{equation}
and
\begin{equation}\label{jw}
\dim^{J}_{\vartheta}(\zeta)=\sum_{\eta\subset\zeta,|\zeta|=|\eta|+1}\chi^J(\eta,\zeta)\dim^J_{\vartheta}(\eta).
\end{equation}
In the Kingman graph, the total weight of $\eta$ is $\dim^{K}(\eta)=\frac{n!}{\eta_1!\cdots\eta_l!}$ if $\eta=(\eta_1,\cdots,\eta_l)\in\Gamma_n$. In the Jack graph, the total of $\eta$ is $\dim^{J}_{\vartheta}(\eta)=\frac{n!}{H(\eta;\vartheta)H^{'}(\eta;\vartheta)}$, where $H(\eta;\vartheta)=\prod_{b\in\eta}(a(b)+1+l(b)\eta), H^{'}(\eta;\vartheta)=\prod_{b\in\eta}(a(b)+(1+l(b))\eta)$. When $\vartheta=1$, the total weight of $\eta$ is $\dim_{1}^{J}(\eta)=\frac{n!}{H^2(\eta;1)}$ in the Young graph.
\subsection{Down-Up Markov Chain} Due to the equation (\ref{rw}), one can easily construct a down Markov chain $\{D_{n}^{K},n\geq1\}$ in the Kingman graph with the following transition probability
$$
p^{\downarrow,K}(\zeta,\eta)=\frac{\chi^K(\eta,\zeta)\dim^K(\eta)}{\dim^K(\zeta)}=\frac{\alpha_{\zeta_i}(\zeta)\zeta_i}{n}, ~\zeta-\eta=(i,j),\zeta\in\Gamma_n.
$$
This down Markov chain is the embedded chain of the dual process in \cite{GSMZ}.
Similarly, one can also construct a down Markov chain $\{D_{n}^{J,\vartheta},n\geq1\}$ in the Jack graph with the following transition probability
$$
p^{\downarrow,J}_{\vartheta}(\zeta,\eta)=\frac{\chi^J_{\vartheta}(\eta,\zeta)\dim^J_{\vartheta}(\eta)}{\dim^J_{\vartheta}(\zeta)}, ~\zeta\in\Gamma_n, \eta\in\Gamma_{n-1}.
$$
When $\vartheta=1$, this down Markov chain is the embedded chain of the dual process of the $Z$-measure diffusion that we are going to discuss in Section \ref{s4}.
As you may see, the down Markov chain depends only on the edge weights of the graph. But we can also construct up Markov chain if we have exchangeable partition structures in the branching diagram. The Ewens-Pitman partition structure can be used to construct an up Markov chain $\{U_n^{K,\theta,\alpha},n\geq1\}$ in the Kingman graph \cite{Pet}. Then we can define an up-down chain $\{UD^{K,\theta,\alpha,n}_{m},m\geq1\}$ on $\Gamma_n$, updating itself as Gibbs sampler,
\begin{align}
&\mathbb{P}(UD^{K,\alpha,\theta,n}_{m+1}=\zeta\mid UD^{K,\alpha,\theta,n}_{m}=\eta)\nonumber\\
=&\sum_{|\nu|=n+1}\mathbb{P}(U_n^{K,\theta,\alpha}=\nu\mid U_n^{K,\theta,\alpha}=\eta)\mathbb{P}(D_n^{K}=\zeta\mid D_n^{K}=\nu)\label{gibbs}
\end{align}
Naturally, the partition distribution $M_n^{K,\theta,\alpha}$ will serve as the stationary distribution of this up-down chain. The usual space and time scaling yields the two-parameter Poisson-Dirichlet diffusion in \cite{Pet}. The Ewens-Pitman partition structure can be replicated by the Blackwell-MacQueen urn model, and it has found many applications in classification problems through Bayesian statistics \cite{BE}.
For the Jack graph there is also a special $Z$ partition structure \cite{BO1}
\begin{equation}\label{zp}
M^{J,z,z^{\prime},\vartheta}_n(\eta)=\dim^{J}_{\vartheta}(\eta)\frac{(z)_{\eta,\vartheta}(z^{\prime})_{\eta,\vartheta}}{\theta_{(n)}H^{\prime}(\eta;\vartheta)}, ~\theta=\frac{zz^{\prime}}{\vartheta},\vartheta>0,
\end{equation}
where $z,z^{\prime}$ are either (i) $z\in\mathbb{C}-(\mathbb{Z}_{\leq0}+\vartheta \mathbb{Z}_{\geq0})$ and $z^{\prime}=\overline{z}$ (principal case)or (ii) $\vartheta$ is rational number and $z,z^{\prime}$ are real numbers belonging to an interval between two consecutive lattice points in $\mathbb{Z}+\vartheta\mathbb{Z}$(complementary case).
Here $(z)_{\eta;\vartheta}=\prod_{(i,j)\in\eta}(z+(j-1)-(i-1)\vartheta)$. The representing measure $\mathcal{Z}(z,z^{\prime},\vartheta)$ of the partition structure (\ref{zp}) is the $Z$-measure. Similarly, one can construct an up Markov chain $\{U^{J,z,z^{\prime},\vartheta}_{n},n\geq1\}$ \cite{Ol_2}, By mimicking the update in (\ref{gibbs}), one can also construct an up-down Markov chain $\{UD^{J,z,z^{\prime},\vartheta,n}_{m},m\geq1\}$ on $\Gamma_n$. Then the usual space and time scaling yields the $Z$-diffusion in \cite{Ol_2}.
As far as the author's knowledge, no quick replication of $Z$ partition structure $M^{J,z,z^{\prime},\vartheta}_n(\eta)$ has been spotted now. So whether the $Z$ partition structure can be applied to classification problems is still open.
\section{The $Z$-measure diffusion and its transition density}
\subsection{The $Z$-measure diffusion}
The diffusion approximation of the up-down Markov chain $\{UD^{J,z,z^{\prime},\vartheta,n}_{m},m\geq1\}$ on $\Gamma_n$ can be carried out by the following space scaling
$$
\pi: \eta\to \frac{\widetilde{\eta}}{n}=\left(\frac{a(1,1)+\frac{1}{2}}{n},\cdots,\frac{a(r,r)+\frac{1}{2}}{n}; \frac{l(1,1)+\frac{1}{2}}{n},\cdots,\frac{l(r,r)+\frac{1}{2}}{n}\right),
$$
where $\widetilde{\eta}$ is the Frobenius coordinates of $\eta$.
As $n\to\infty$, Olshanski in \cite{Ol_2} has shown that $\pi(UD^{J,z,z^{\prime},\vartheta,n}_{[n^2t]})$ converges to the $Z$-measure diffusion $Y_t$ on the Thoma simplex $\Omega$ with the following pre-generator
\begin{align*}
A_{z,z^{\prime},\vartheta}=&\frac{1}{2}\sum_{i,j\geq2}ij(\varphi_{i+j-1}^{o}-\varphi_{i}^{o}\varphi_j^{o})\frac{\partial^2}{\partial \varphi_i^{o}\partial \varphi_j^{o}}+\frac{\vartheta}{2}\sum_{i,j\geq1}(i+j+1)\varphi_i^{o}\varphi_j^{o}\frac{\partial}{\partial \varphi_{i+j+1}^{o}}\\
&+\frac{1}{2}\sum_{i\geq2}\left[(1-\vartheta)i(i-1)\varphi_{i-1}^{o}+(z+z^{'})i\varphi_{i-1}^{o}-i(i-1)\varphi_i^{o}-i\frac{zz^{'}}{\vartheta}\varphi_i^{o}\right]\frac{\partial}{\partial \varphi_i^{o}}.
\end{align*}
The core of $ A_{z,z^{\prime},\vartheta}$ is spanned by $\{\varphi_j^{o},j\geq1\}$, where $\varphi_j^{o}$ is the image of $\varphi_j(x)=\sum_{i=1}^{\infty}x_i^j$ under the special algebra homomorphisms $\Phi_{\omega}: \mathcal{A}\to\mathbb{R}, \omega=(\alpha,\beta)\in\Omega,$
$$
\Phi_{\omega}(\varphi_{n})=\sum_{i=1}^{\infty}\alpha_i^{n}+(-\vartheta)^{n-1}\sum_{j=1}^{\infty}\beta_j^{n}, ~n\geq2, ~\Phi_{\omega}(\varphi_{1})=1.
$$
Since the Jack functions $J_{\zeta}(x;\vartheta)$ can be written as linear combinations of $\varphi_{\eta_1}\cdots\varphi_{\eta_l}$, where $\eta=(\eta_1,\cdots,\eta_l)\subset \zeta$, then $J_{\zeta}^{o}(\omega;\vartheta):=\Phi_{\omega}(J_{\zeta}(x;\vartheta))$.
In particular, when $\vartheta=1$, we will have the extended Shur function $s^{o}_{\zeta}(\omega):=\Phi_{\omega}(J_{\zeta}(x;1))$. The $Z$-measure diffusion reduces to the diffusion in \cite{BO2}.
\subsection{Transition density of the $Z$-measure diffusion $Y_t$}
By spectral expansion of $A_{z,z^{\prime},\vartheta}$ in the Hilbert space $L^2(\Omega, \mathcal{Z}(z,z^{\prime},\vartheta))$, Korotkikh obtained the following explicit transition density of $Y_t$ in \cite{SY_K}.
\begin{pro}
The transition density of $Y_t$ is
\begin{equation}\label{spec_expan}
q(t,\sigma,\omega)=1+\sum_{m=2}^{\infty}e^{-t\lambda_m}G_m(\sigma,\omega),
\end{equation}
where $\theta=\frac{zz^{\prime}}{\vartheta}$, $\lambda_m=\frac{m(m-1+\theta)}{2}$,
$
G_m(\sigma,\omega)=\sum_{n=0}^{m}(-1)^{m-n}\binom{m}{n}\frac{(\theta+2m-1)(\theta+n)_{m-1}}{m!}\mathcal{K}_n(\sigma,\omega),
$
and
\begin{equation}\label{kernel}
\mathcal{K}_n(\sigma,\omega)=\sum_{|\eta|=n}\frac{j_{\eta}(\sigma;\vartheta)j_{\eta}(\omega;\vartheta)}{\mathbb{E}_{z,z^{\prime},\vartheta}j_{\eta}},~ j_{\eta}(\omega;\vartheta)=\dim_{\vartheta}^{J}(\eta)J^o_{\eta}(\omega;\vartheta).
\end{equation}
In particular, when $n=0,1$, $\mathcal{K}_n(\sigma,\omega)=1$.
\end{pro}
In \cite{SY_K},
$
K_n^{o}(\sigma,\omega)=\sum_{|\eta|=n}\frac{H^{\prime}(\eta;\vartheta)}{H(\eta;\vartheta)}\frac{J^{o}(\sigma;\vartheta)J^{o}(\omega;\vartheta)}{(z)_{\eta,\vartheta}(z^{\prime})_{\eta,\vartheta}}.
$
By Proposition \ref{jack}, Borodin and Olshanski \cite{BO1} have shown that the representing measure of the $Z$ partition structure is the $Z$ measure $\mathcal{Z}(z,z^{\prime},\vartheta)$ .
\begin{pro}\label{jack}
For a partition structure $\{M_{n},n\geq1\}$ on the Jack graph, there is a unique probability measure $\mu$ on the Thoma simplex $\Omega$ such that
$$
M_n(\eta)=\int_{\Omega}j_{\eta}(\omega;\vartheta)\mu(d\omega),
$$
Moreover, $\mu_n(d\omega)=\sum_{\eta\in\Gamma_n}M_n(\eta)\delta_{\frac{\widetilde{\eta}}{n}}(d\omega)$ will converge weakly to $\mu$. Here
$$
\widetilde{\eta}=\left(a(1,1)+\frac{1}{2},\cdots,a(r,r)+\frac{1}{2}; l(1,1)+\frac{1}{2},\cdots,l(r,r)+\frac{1}{2}\right)
$$
is the Frobenius coordinate of $\eta$.
\end{pro}
Therefore, similar to the Ewens sampling formula, we have the following sampling formula
\begin{equation}\label{sap}
\mathbb{E}_{z,z^{\prime},\vartheta}j_{\eta}=M_{n}^{z,z^{\prime},\vartheta}(\eta)
\end{equation}
and $K_n^{o}(\sigma,\omega)=\frac{\mathcal{K}_n(\sigma,\omega)}{n!(\theta)_{n}}$. Then one can easily see that (\ref{spec_expan}) is equivalent to the density in \cite{SY_K}.
In this paper, we will rearrange the right hand side of the equation (\ref{spec_expan}) to yield a new representation
\begin{thm}\label{main}
The transition density of $Y_t$ is
\begin{equation}\label{con}
q(t,\sigma,\omega)=d_0^{\theta}(t)+d_1^{\theta}(t)+\sum_{n=2}^{\infty}d_n^{\theta}(t)\mathcal{K}_n(\sigma,\omega)
\end{equation}
where
\begin{align*}
d_{0}^{\theta}(t)=&1-\sum_{m=1}^{\infty}\frac{2m-1+\theta}{m!}(-1)^{m-1}\theta_{(m-1)}e^{-\lambda_{m}t}\\
d_{n}^{\theta}(t)=&\sum_{m=1}^{\infty}\frac{2m-1+\theta}{m!}(-1)^{m-n}\binom{m}{n}(n+\theta)_{(m-1)}e^{-\lambda_{m}t},n\geq1.
\end{align*}
\end{thm}
Due to the estimation of $G_m$ in \cite{SY_K}, the proof of Theorem \ref{main} is the same as the proof of Theorem 2.1 in \cite{Zhou}.
\begin{proof}
By Proposition 16 in \cite{SY_K}, we know there exists positive constants $c,d$ such that
$$
\sup_{\sigma,\omega\in\Omega}|G_m(\sigma,\omega)|\leq c m^{dm}.
$$
Then one can show that the series in (\ref{spec_expan}) is uniformly convergent( see \cite{Zhou} or \cite{SY_K}). Then we can switch the order of summation in (\ref{spec_expan}).
\begin{align*}
q(t,\sigma,\omega)=&1+\sum_{m=2}^{\infty}e^{-\lambda_{m}t}\Big[\sum_{n=2}^{m}(-1)^{m-n}\binom{m}{n}\frac{(\theta+2m-1)(\theta+n)_{m-1}}{m!}\mathcal{K}_n(\sigma,\omega)\\
&+\frac{2m+\theta-1}{m!}(-1)^{m-1}(\theta+1)_{(m-1)}m\mathcal{K}_1(\sigma,\omega)\\
&+\frac{2m+\theta-1}{m!}(-1)^{m}(\theta)_{(m-1)}\mathcal{K}_0(\sigma,\omega)\Big]
\end{align*}
Since $\mathcal{K}_1(\sigma,\omega)=\mathcal{K}_0(\sigma,\omega)=1$, we have
\begin{align*}
q(t,\sigma,\omega)=&1+\sum_{m=2}^{\infty}e^{-\lambda_{m}t}\sum_{n=2}^{m}(-1)^{m-n}\binom{m}{n}\frac{(\theta+2m-1)(\theta+n)_{m-1}}{m!}\mathcal{K}_n(\sigma,\omega)\\
&+\sum_{m=2}^{\infty}e^{-\lambda_{m}t}\Big[\frac{2m+\theta-1}{m!}(-1)^{m-1}(\theta+1)_{(m-1)}m\\
&+\frac{2m+\theta-1}{m!}(-1)^{m}(\theta)_{(m-1)}\Big]\\
=&1+\sum_{m=2}^{\infty}e^{-\lambda_{m}t}\sum_{n=2}^{m}(-1)^{m-n}\binom{m}{n}\frac{(\theta+2m-1)(\theta+n)_{(m-1)}}{m!}\mathcal{K}_n(\sigma,\omega)\\
&+\sum_{m=2}^{\infty}e^{-\lambda_{m}t}\Big[\frac{2m+\theta-1}{m!}(-1)^{m-1}(\theta+1)_{(m-1)}m\\
&+\frac{2m+\theta-1}{m!}(-1)^{m}(\theta)_{(m-1)}\Big]\\
=&1+\sum_{m=2}^{\infty}e^{-\lambda_{m}t}\sum_{n=2}^{m}(-1)^{m-n}\binom{m}{n}\frac{(\theta+2m-1)(\theta+n)_{(m-1)}}{m!}\mathcal{K}_n(\sigma,\omega)\\
&+\sum_{m=2}^{\infty}e^{-\lambda_{m}t}\frac{2m+\theta-1}{m!}(-1)^{m-1}\Big[(\theta+1)_{(m-1)}m-(\theta)_{(m-1)}\Big]
\end{align*}
Note that $(\theta+1)_{(m-1)}m-(\theta)_{(m-1)}=0$ for $m=1$. Therefore,
\begin{align*}
q(t,\sigma,\omega)
=&1+\sum_{m=2}^{\infty}e^{-\lambda_{m}t}\sum_{n=2}^{m}(-1)^{m-n}\binom{m}{n}\frac{(\theta+2m-1)(\theta+n)_{(m-1)}}{m!}\mathcal{K}_n(\sigma,\omega)\\
&+\sum_{m=1}^{\infty}e^{-\lambda_{m}t}\frac{2m+\theta-1}{m!}(-1)^{(m-1)}\Big[(\theta+1)_{(m-1)}m-(\theta)_{(m-1)}\Big]\\
=&1-\sum_{m=1}^{\infty}e^{-\lambda_{m}t}\frac{2m+\theta-1}{m!}(-1)^{(m-1)}(\theta)_{(m-1)}\\
+&\sum_{m=1}^{\infty}e^{-\lambda_{m}t}\frac{2m+\theta-1}{m!}(-1)^{(m-1)}(\theta+1)_{(m-1)}m\\
+&\sum_{m=2}^{\infty}e^{-\lambda_{m}t}\sum_{n=2}^{m}(-1)^{m-n}\binom{m}{n}\frac{(\theta+2m-1)(\theta+n)_{(m-1)}}{m!}\mathcal{K}_n(\sigma,\omega)\\
=&d_0^{\theta}(t)+d_1^{\theta}(t)+\sum_{n=2}^{\infty} \mathcal{K}_n(\sigma,\omega)\sum_{m=n}^{\infty}e^{-\lambda_{m}t}(-1)^{m-n}\binom{m}{n}\frac{(\theta+2m-1)(\theta+n)_{(m-1)}}{m!}\\
=&\widetilde{d_1^{\theta}}(t)+\sum_{n=2}^{\infty}d_{n}^{\theta}(t)\mathcal{K}_n(\sigma,\omega)
\end{align*}
\end{proof}
\begin{cor}
The diffusion $Y_t$ satisfies the following ergodic inequality
\begin{align*}
\sup_{\omega\in \Omega}\|\mathbb{P}_{\omega}(Y_t\in\cdot)-\mathcal{Z}(z,z^{\prime},\vartheta)(\cdot)\|_{\mathrm{Var}}\leq \frac{(\theta+1)(\theta+2)}{2}e^{-(\theta+1)t},~\theta=\frac{zz^{\prime}}{\vartheta}>0.
\end{align*}
\end{cor}
This inequality can be easily derived from an inequality of tail probabilities (see \cite{Tav})
$$
\sum_{n=2}^{\infty}d_n^{\theta}(t)\leq \frac{(\theta+1)(\theta+2)}{2}e^{-(\theta+1)t}.
$$
\section{Dual process of the $Z$-measure diffusion when $\vartheta=1$}\label{s4}
The transition density of the $Z$-measure diffusion looks so similar to the transition density of the two-parameter Poisson-Dirichlet diffusion. Given their huge differences, it is surprising that their transition densities are both mixture of distributions with exactly the same coefficients. To figure out why they have the same coefficients, we adopt the dual process method used in \cite{GSMZ}. The duality between $Y_t$ and its dual process $\mathcal{D}_{t}$ is defined through a bivariate function $F(\eta,\omega)$ where $\eta\in\Gamma$ and $\omega\in\Omega$. The bivariate function $F(\eta,\omega)$ is usually chosen to be the normalized kernel $F(\eta,\omega)=\frac{j_{\eta}(\omega;\vartheta)}{\mathbb{E}_{z,z^{\prime},\vartheta}j_{\eta}}$ in Proposition \ref{jack}. Then $Y_t$ and its dual $\mathcal{D}_t$ satisfy
\begin{equation}\label{dual}
\mathbb{E}_{\eta}^*F(\mathcal{D}_t,\omega)=\mathbb{E}_{\omega}F(\eta, Y_t)
\end{equation}
where $\mathbb{E}_{\eta}^*$ is the expectation with respect to the distribution of $\mathcal{D}_t$. In this paper, we use the dual equation (\ref{dual}) to derive the transition density (\ref{con}) directly when $\vartheta=1$. This derivation will clearly explain why the coefficients $d_n^{\theta}(t)$ show up in the transition density (\ref{con}). However, when $\vartheta\neq1$, we fail to verify the dual process of $Y_t$ because we don't know whether the $Z$-measure diffusion has similar equation (5) in \cite{Ol_2}.
\subsection{Dual process} In this section, we will consider test functions $g_{\eta}(\omega)=\frac{j_{\eta}(\omega;1)}{\mathbb{E}_{z,z^{\prime},\vartheta}j_{\eta}}$, so $\mathbb{E}_{z,z^{\prime},\vartheta}g(\eta)=1$. Because when $\vartheta=1$, $J_{\eta}(x;1)$ is just the Shur function $s_{\eta}(x)$. So $j_{\eta}(\omega;1)$ is $\dim_{1}^{J}(\eta)\Phi_{\omega}(s_{\eta})=\dim_{1}^{J}(\eta)s^{o}_{\eta}(\omega)$. Moreover, by the equation (\ref{sap}), we know
$$
\mathbb{E}_{z,z^{\prime},1}\dim_{1}^{J}(\eta)s^{o}_{\eta}=\frac{n!}{H^2(\eta;1)}\frac{(z)_{\eta,1}(z^{\prime})_{\eta,1}}{(\theta)_{(n)}}
$$
We define $F(\eta,\omega)=g_{\eta}(\omega), \eta\in\Gamma,\omega\in\Omega$. So
\begin{equation}\label{sam-formula}
F(\eta,\omega)=g_{\eta}(\omega)=\frac{H(\eta;1)\theta_{(n)}s_{\eta}^{o}(\omega)}{(z)_{\eta,1}(z^{\prime})_{\eta,1}}.
\end{equation}
Now consider a jump process $\mathcal{D}_t$ defined by
\begin{equation}
\mathcal{L}^{J}_{1}f(\eta)=-\frac{n(n-1+\theta)}{2}f(\eta)+\frac{n(n-1+\theta)}{2}\sum_{\zeta\subset \eta,|\eta|=|\zeta|+1}\frac{\dim_{1}^{J}(\zeta)}{\dim^{J}_{1}(\eta)}f(\zeta),
\end{equation}
where $\frac{\dim_{1}^{J}(\zeta)}{\dim^{J}_{1}(\eta)}=p^{\downarrow,J}_{1}(\eta, \zeta)$ is the transition probability of the down Markov chain $D^{J,1}_n$ in the Young graph. Moreover $ \mathcal{L}^{J}_{1}1=0$, and because $s_{(1)}^{o}=1$ then $\mathcal{D}_t$ will be absorbed at state $\eta=(1)$. The radial process $|\mathcal{D}_t|$ of $\mathcal{D}_t$ is only determined by the jump rates $\frac{n(n-1+\theta)}{2},n\geq1$. Therefore, $|\mathcal{D}_t|$ is exactly the radial process in the Kingman coalescent by collapsing state $0$ and $1$ as a new state $1$. The distribution of $|\mathcal{D}_t|$ is obtained in \cite{Tav} and \cite{Zhou}. So as long as the jump rates are the same for the dual process, the coefficients in the their transition density, if exists, will be the same.
\begin{thm}\label{d1}
When $\vartheta=1$, the $Z$-measure diffusion $Y_t$ and $\mathcal{D}_t$ satisfy the following duality
\begin{equation}\label{dual-equation}
\mathbb{E}_{\eta}^*F(\mathcal{D}_t,\omega)=\mathbb{E}_{\omega}F(\eta, Y_t)
\end{equation}
\end{thm}
\begin{proof}
By the Lemma 5.4 in \cite{BO2}, we know
\begin{equation}\label{recur}
A_{z,z^{\prime},1}s_{\eta}^{o}(\omega)=-\frac{n(n-1+\theta)}{2}s_{\eta}^{o}(\omega)+\frac{1}{2}\sum_{\zeta\subset \eta, |\eta|=|\zeta|+1}\frac{(z)_{\eta,1}(z^{'})_{\eta,1}}{(z)_{\zeta,1}(z^{'})_{\zeta,1}}s^{0}_{\zeta}(\omega)
\end{equation}
Due to equation (\ref{sam-formula}), we know
$$
s^{o}_{\eta}(\omega)=g_{\eta}(\omega)\frac{(z)_{\eta,1}(z^{'})_{\eta,1}}{H(\eta;1)\theta_{(n)}},~~s^{o}_{\zeta}(\omega)=g_{\zeta}(\omega)\frac{(z)_{\zeta,1}(z^{'})_{\zeta,1}}{H(\zeta;1)\theta_{(n-1)}}
$$
Replacing $s^{o}_{\eta}$ and $s^{o}_{\zeta}$ in equation (\ref{recur}) yields
$$
A_{z,z^{\prime},1}g_{\eta}(\omega)=-\frac{n(n-1+\theta)}{2}g_{\eta}(\omega)+\frac{1}{2}\sum_{\zeta\subset \eta, |\eta|=|\zeta|+1}\frac{H(\eta;1)\theta_{(n)}}{H(\zeta;1)\theta_{(n-1)}}g_{\zeta}(\omega)
$$
Because $\frac{H(\eta;1)}{H(\zeta;1)}=n\frac{\dim_{1}^{J}(\zeta)}{\dim_{1}^{J}(\eta)}$ and $\frac{\theta_{(n)}}{\theta_{(n-1)}}=n-1+\theta$, we have
$$
A_{z,z^{\prime},1}g_{\eta}(\omega)=-\frac{n(n-1+\theta)}{2}g_{\eta}(\omega)+\frac{n(n-1+\theta)}{2}\sum_{\zeta\subset \eta, |\eta|=|\zeta|+1}\frac{\dim_{1}^{J}(\zeta)}{\dim_{1}^{J}(\eta)}g_{\zeta}(\omega).
$$
Therefore,
$$
A_{z,z^{\prime},1}F(\eta,\omega)=\mathcal{L}^{J}_{1}F(\eta,\omega).
$$
By Theorem 4.4.1 in \cite{EK}, one can show that the $Z$-measure diffusion $Y_t$ and $\mathcal{D}_t$ satisfy the dual equation (\ref{dual-equation}).
\end{proof}
\begin{pro}\label{dual-prob1}
The dual process $\mathcal{D}_t$ has the following transition probability
$$
\mathbb{P}(\mathcal{D}_t=\eta\mid \mathcal{D}_0=\nu)=d_{mn}^{\theta}(t)\mathcal{H}(\eta,\nu),~\eta\in\Gamma_{n}, \nu\in \Gamma_m, n\leq m.
$$
Here
$$
\mathcal{H}(\eta,\nu)=\frac{\dim^{J}_{1}(\eta)\dim^{J}_{1}(\eta,\nu)}{\dim^{J}_{1}(\nu)}
$$
and
\begin{equation}\nonumber
d_{mn}^\theta(t) =
\sum_{k=n}^{m}e^{-\frac{1}{2}k(k+\theta-1)t}
(-1)^{k-n}
\frac{(2k+\theta-1)(n+\theta)_{(k-1)}}{n!(k-n)!}
\frac{ m_{[k]} }{ (\theta+m)_{(k)} }.
\end{equation}
\end{pro}
\begin{proof} By the definition of $|\mathcal{D}_{t}|$ and Proposition 2.1 in \cite{Zhou}, we know
$$
\mathbb{P}(|\mathcal{D}_{t}|=n\mid|\mathcal{D}_{0}|=m)=d_{mn}^{\theta}(t).
$$
Then
\begin{align*}
\mathbb{P}_{\nu}(\mathcal{D}_t=\eta)=&\mathbb{P}(\mathcal{D}_{t}=\eta\mid \mathcal{D}_0=\nu)=\mathbb{P}(\mathcal{D}_t=\eta| |\mathcal{D}_t|=n,\mathcal{D}_0=\nu)\mathbb{P}(|\mathcal{D}_{t}|=n\mid \mathcal{D}_0=\nu)\\
=&d_{mn}^{\theta}(t)\mathbb{P}(\mathcal{D}_t=\eta| |\mathcal{D}_t|=n,\mathcal{D}_0=\nu).
\end{align*}
For $\nu\in\Gamma_{m}, \eta\in\Gamma_{n}$, there are paths of length $k=m-n$ joining $\nu$ and $\eta$. These paths are the realizations of the embedded down Markov chain $D^{J,1}_n$. Thus,
\begin{align*}
\mathbb{P}(\mathcal{D}_t=\eta| |\mathcal{D}_t|=n,\mathcal{D}_0=\nu)=&\sum_{\eta=\eta(k)\subset\cdots\subset \eta(0)=\nu}\prod_{i=0}^{k-1}p^{\downarrow,J}_{1}(\eta(i),\eta(i+1))\\
=&\sum_{\eta=\eta(k)\subset\cdots\subset \eta(0)=\nu}\prod_{i=0}^{k-1}\frac{\dim_{1}^{J}(\eta(i+1))\chi^{J}_{1}(\eta(i),\eta(i+1))}{\dim_{1}^{J}(\eta(i))}\\
=&\frac{\dim_1^{J}(\eta(k))}{\dim_{1}^{J}(\eta(0))}\sum_{\eta=\eta(k)\subset\cdots\subset \eta(0)=\nu}\prod_{i=0}^{k-1}\chi^{J}_{1}(\eta(i),\eta(i+1))\\
=&\frac{\dim_1^{J}(\eta)}{\dim_{1}^{J}(\nu)}\dim_{1}^{J}(\eta,\nu).
\end{align*}
\end{proof}
\begin{pro}\label{dual-prob2}
For $j_{\eta}(\omega;1), \eta\in\Gamma_m$, we have
\begin{equation}\label{dp2}
\mathbb{E}_{z,z^{\prime},1}[j_{\eta}(Y_t;1)\mid Y_0=\omega]=(\mathbb{E}_{z,z^{\prime},1}j_{\eta})\left[d_{m1}^{\theta}(t)+\sum_{n=2}^{m}d_{mn}^{\theta}(t)\sum_{\zeta\subset\eta, |\zeta|=n}\mathcal{H}(\eta,\zeta)\frac{j_{\zeta}(\omega;1)}{\mathbb{E}_{z,z^{'},1}j_{\zeta}}\right]
\end{equation}
\end{pro}
\begin{proof}
By the duality equation (\ref{dual-equation}), we know
\begin{align*}
\mathbb{E}_{\omega}g_{\eta}(Y_t)=\mathbb{E}_{\eta}^*g_{\mathcal{D}_t}(\omega)=\sum_{\zeta\subset\eta}\mathbb{P}_{\eta}(\mathcal{D}_t=\zeta)g_{\zeta}(\omega)=\sum_{n=1}^{m}d_{mn}^{\theta}(t)\sum_{|\zeta|=n,\zeta\subset \eta}\mathcal{H}(\zeta,\eta)g_{\zeta}(\omega).
\end{align*}
Then we have the equation (\ref{dp2}) if we replace $g_{\eta}(\omega)=\frac{j_{\eta}(\omega;1)}{\mathbb{E}_{z,z^{\prime},1}j_{\eta}}$ and $g_{\zeta}(\omega)=\frac{j_{\zeta}(\omega;1)}{\mathbb{E}_{z,z^{\prime},1}j_{\zeta}}$.
\end{proof}
\subsection{Proof of Theorem \ref{main} through the dual process method}Next we will use Proposition \ref{dual-prob2} to deduce the transition density (\ref{con}). By Proposition \ref{jack}, we know
$$
\mu_{m}(d\omega)=\sum_{|\eta|=m}\mathbb{E}[j_{\eta}(Y_t;1)\mid Y_0=\omega]\delta_{\frac{\widetilde{\eta}}{m}}(d\omega)
$$
will converge weakly to the distribution of $Y_t$, i.e. the transition probability $q(t,\omega, \cdot)= \mathbb{P}_{\omega}(Y_t\in\cdot)$. By the equation (\ref{dp2}), we know
\begin{equation}
\mu_{n}(d\omega)=d_{m1}^{\theta}(t)\mu_{1,m}(d\omega)+\sum_{n=2}^{m}d_{mn}^{\theta}(t)\sum_{|\zeta|=n}\mu_{n,m}(d\omega)\frac{j_{\zeta}(\omega;1)}{\mathbb{E}_{z,z^{\prime},1}j_{\zeta}}
\end{equation}
where
\begin{align*}
\mu_{1,m}(d\omega)=&\sum_{|\eta|=m}[\mathbb{E}_{z,z^{\prime},1}j_{\eta}]\delta_{\frac{\widetilde{\eta}}{m}}(d\omega)\\
\mu_{n,m}(d\omega)=&\sum_{|\eta|=m,\zeta\subset\eta}\mathcal{H}(\eta,\zeta)[\mathbb{E}_{z,z^{\prime},1}j_{\eta}]\delta_{\frac{\widetilde{\eta}}{m}}(d\omega),~n\geq2.
\end{align*}
As $m\to\infty$, we need to show that
\begin{equation}\label{claim1}
\mu_{1,m}(d\omega)\to \mathcal{Z}(z,z^{\prime},\vartheta)(d\omega)
\end{equation}
and
\begin{equation}\label{claim2}
\mu_{n,m}(d\omega)\to j_{\zeta}(\omega;1)\mathcal{Z}(z,z^{\prime},\vartheta)(d\omega)
\end{equation}
The claim (\ref{claim1}) can be directly obtained from Proposition \ref{jack}. Now we are going to show the claim (\ref{claim2}). For any $f\in C(\Omega)$, we have
\begin{align*}
\int_{\Omega}f(\omega)\mu_{n,m}(d\omega)=&\sum_{|\eta|=m,\zeta\subset\eta}f(\frac{\widetilde{\eta}}{m})\mathcal{H}(\eta,\zeta)(\mathbb{E}_{z,z^{\prime},1}j_{\eta})\\
=& \sum_{|\eta|=m,\zeta\subset\eta}f(\frac{\widetilde{\eta}}{m})j_{\zeta}(\frac{\widetilde{\eta}}{m};1)(\mathbb{E}_{z,z^{\prime},1}j_{\eta})+O(\frac{1}{\sqrt{m}})\\
\to& \int_{\Omega}f(\omega)j_{\zeta}(\omega;1)\mathcal{Z}(z,z^{\prime},\vartheta)(d\omega)
\end{align*}
where the second equality is due to Theorem 6.1 and Theorem 7.1 in \cite{KOO}. Since
$$
\lim_{m\to\infty}d_{mn}^{\theta}(t)=d_{n}^{\theta}(t),
$$
then we have derived the representation in Theorem \ref{main}.
\section{Further discussion}
\subsection{Conjectures on Diffusions with Given Stationary Measures} The conclusions in this paper indicate that the dual process is only determined by the weights in the branching diagram, whether it be the Kingman graph or the Jack graph. The dual processes are determined by a generator
$$
\mathcal{L}f(\eta)=-\frac{n(n-1+\theta)}{2}f(\eta)+\frac{n(n-1+\theta)}{2}\sum_{\zeta\subset \eta,|\eta|=|\zeta|+1}p^{\downarrow}(\eta,\zeta)f(\zeta).
$$
The diffusions, however, depend on the partition structures on the Kingman graph or the Jack graph. Due to the Kingman's representation theorem and the representation theorem in Proposition \ref{jack}, the partition structures are uniquely determined by their representing measures, which will be the stationary distribution of the diffusions constructed through the up-down Markov chains.
More generally, for a given probability measure $\mu$ on the Kingman simplex $\overline{\nabla}_{\infty}$, we can consider generator
$
\mathcal{B}_{\mu}
$
defined on an algebra $\mathcal{A}^{o}_{K}$ spanned by $\{1,\varphi_k,k\geq2\}$ as follows
\begin{equation}\label{conj1}
\mathcal{B}_{\mu}^Kg_{\eta}(x)=-\frac{n(n-1+\theta)}{2}g_{\eta}(x)+\frac{n(n-1+\theta)}{2}\sum_{\zeta\subset \eta,|\eta|=|\zeta|+1}\frac{\dim^K(\zeta)\chi^{K}(\zeta,\eta)}{\dim^K(\eta)}g_{\zeta}(x)
\end{equation}
where $\{g_{\eta}(x)=\frac{m_{\eta}(x)}{\mathbb{E}_{\mu} m_{\eta}}\mid \eta\in\Gamma\}$ is a linear base and $\mathbb{E}_{\mu}$ is the expectation with respect to the probability distribution $\mu$. It will uniquely determine the operation of $\mathcal{B}_{\mu}^K$ on $\mathcal{A}^{o}_{K}$.
\begin{thm}[Conjecture 1]
For a given probability measure $\mu$ on the Kingman simplex $\overline{\nabla}_{\infty}$, the generator defined in (\ref{conj1}) will determine a reversible diffusion $W_t^K$ with the stationary distribution $\mu$. Its transition density is
$$
p^K(t,x,y)=d_0^{\theta}(t)+d_1^{\theta}(t)+\sum_{n=2}^{\infty}d_n^{\theta}(t)p_n(x,y),
$$
where
$$
p_n(x,y)=\sum_{|\eta|=n}\frac{\dim^K(\eta)m_{\eta}^o(x)\dim^K(\eta)m_{\eta}^o(y)}{\mathbb{E}_{\mu}[\dim^K(\eta)m_{\eta}^o]}.
$$
\end{thm}
Moreover, for a given probability measure $\mu$ in the Thoma simplex $\Omega$, we can also consider generator
$
\mathcal{B}_{\mu}^{J}
$
defined on an algebra $\mathcal{A}^{o}_{J}$ spanned by $\{1,\varphi_k^{o},k\geq2\}$ as follows
\begin{equation}\label{conj2}
\mathcal{B}_{\mu}^Jg_{\eta}(x)=-\frac{n(n-1+\theta)}{2}g_{\eta}(x)+\frac{n(n-1+\theta)}{2}\sum_{\zeta\subset \eta,|\eta|=|\zeta|+1}\frac{\dim^J_{\vartheta}(\zeta)\chi^{J}_{\vartheta}(\zeta,\eta)}{\dim^J_{\vartheta}(\eta)}g_{\zeta}(x)
\end{equation}
where $\{g_{\eta}(x)=\frac{j_{\eta}(x;\vartheta)}{\mathbb{E}_{\mu} j_{\eta}}\mid \eta\in\Gamma\}$ is a linear base. It will uniquely determine the operation of $\mathcal{B}_{\mu}^J$ on $\mathcal{A}^{o}_{J}$.
\begin{thm}[Conjecture 2]
For a given probability measure $\mu$ on the Thoma simplex $\Omega$, the generator defined in (\ref{conj2}) will determine a reversible diffusion $W_t^J$ with the stationary distribution $\mu$. Its transition density is
$$
p^J(t,x,y)=d_0^{\theta}(t)+d_1^{\theta}(t)+\sum_{n=2}^{\infty}d_n^{\theta}(t)\mathcal{K}_n(x,y),
$$
where
$$
\mathcal{K}_n(x,y)=\sum_{|\eta|=n}\frac{j_{\eta}(x;\vartheta)j_{\eta}(y;\vartheta)}{\mathbb{E}_{\mu}[j_{\eta}]}.
$$
\end{thm}
\section{Appendix}
|
\section{Introduction}
Mrk~231, also known as UGC 08058 and IRAS 12540+5708 at a redshift $z = 0.042$ \citep{1991rc3..book.....D}, is a relatively nearby ultra-luminous infrared galaxy (ULIRG) \citep{1988ApJ...325...74S}. Discovery of strong hydroxyl (OH) megamasers in this galaxy supports an extreme starburst in the nuclear region \citep{1985Natur.315...26B,2003Natur.421..821K}, consistent with its high infrared luminosity of $\approx 3 \times 10^{12}~ L_\odot$ \citep{1989AJ.....98..766S}.
Mrk~231 is characterised by prominent multi-phase outflows extending over multiple physical scales \citep{2014ApJ...788..123L}. Optical broad absorption lines (BALs) are detected in both low and high ionisation species and indicate galactic-scale outflows arising at the pc-scales \citep{2013ApJ...764...15V,2016ApJ...825...42V}. The study of \citet{2015A&A...583A..99F} identifies a molecular gas outflow with a velocity $\geq 400$ \kms originating from the nuclear region and extending to $\approx 1$ kpc, and an ionised X-ray ultra-fast outflow (UFO) with a velocity of $\approx$ 20000 \kms originating from the accretion disc; an analysis (based on their energetics and momentum transfer) suggests that a bulk of the UFO kinetic energy is transferred to the mechanical energy of the kpc-scale molecular outflow. The study of \cite{2011ApJ...729L..27R} reports on the detection of fast ($\approx$ 1000 \kms) outflowing neutral gas (as traced by the Na I D absorption) extending from the nuclear region up to 3 kpc; they suggest that this is driven by radiative pressure or mechanical energy from the central AGN. Neutral and ionised gas outflows have also been observed with velocities up to $10^3$ \kms\ \citep{2005ApJ...632..751R,2011ApJ...729L..27R,2017ApJ...850...40R}. H{\sc i}\ 21 cm absorption features are detected \citep{1998AJ....115..928C, 2016A&A...593A..30M} on scales ranging from a few to a few hundred pc. The H{\sc i}\ line emission traces an east-west elongated disc possibly coincident with the inner part of the molecular disc at an even larger scale.
The multi-phase multi-scale outflows in active galactic nuclei (AGN) may be enabled by a mixture of driving mechanisms, including the central AGN (radiation pressure), a nuclear starburst (stellar winds and radiation pressure), and the radio jet (kinetic energy). While the role of AGN-driven winds and outflows (non-relativistic) on feedback at large scales has been investigated in some detail \citep[e.g., ][]{2017ApJ...850...40R,2021A&A...645A..21G}, the role of the radio jet in driving the outflow in the nuclear region requires clarification \cite[e.g.][]{2013Sci...341.1082M}. Mrk~231 is one of the few radio-quiet quasars containing a prominent jet. The source shows a triple (core and two lobes) radio morphology structure with an extent of $\sim$40 pc (50 mas) in the north-south direction, revealed by the European VLBI Network (EVN) observation at 1.6 GHz \citep{1988AJ.....96..841N} and the Very Long Baseline Array (VLBA) observations at multiple frequencies from 1.4 to 8.4 GHz \citep{1999ApJ...516..127U}. The higher-resolution VLBA images at 15 GHz resolve the central component into two compact components separated by $\sim$1 pc aligning in the northeast-southwest direction \citep{1999ApJ...516..127U,1999ApJ...517L..81U}. The northeastern component showed a remarkable variability \citep{1999ApJ...517L..81U} and was thus identified as the core (the location of the active nucleus). The inner 1-pc position angle of the jet is different from that of the larger 40-pc triple structure by $\sim60\degr$. \citet{2009ApJ...706..851R,2017ApJ...836..155R,2020ApJ...891...59R} observed Mrk~231 using the VLBA at higher frequencies of 22 and 43 GHz and obtained a similar core-jet morphology at 1-pc scale. The derived core brightness temperatures, $T_{\rm B} > 10^{10}$ K, are 2--3 orders of magnitude lower than that of typical blazars \citep{2005AJ....130.2473K}, but are consistent with radio-quiet quasars \citep{1996ApJ...468L..91B,2005ApJ...621..123U}. The kinematic property of the VLBI jet is still under debate. \citet{1999ApJ...517L..81U} determined a jet proper motion speed of $0.14 \pm 0.052\, c$ in the period from 1996.8 to 1998.7, while \citet{2017ApJ...836..155R} measured a higher jet apparent speed $>3.15 \,c$ during a flaring state in 2015. This can be explained if Mrk~231 contains episodic fast-moving ($v \approx 1-3\, c$) knots and a continuous slower background jet body \citep{2020ApJ...891...59R}.
In this paper, we use archival VLBI data of Mrk~231 to study the jet kinematics. We discuss the complex nature of the pc-scale jet and possible connection between the wide-angle outflows and the radio jet. Throughout, we adopt the following cosmological parameters: H$_0$ = 71 \kms\ Mpc$^{-1}$, $\Omega_\Lambda = 0.73$ and $\Omega_m = 0.27$. At $z = 0.042$, 1 mas angular separation corresponds to 0.818 pc linear size in projection on the plane of the sky, and 1 mas yr$^{-1}$ jet proper motion is converted to $2.78\,c$.
\section{Observations and data reduction}
\begin{table*}
\centering
\caption{Logs of the 8.4 GHz VLBI observations.}
\begin{tabular}{cccccccc}
\hline \hline
Date & Project Code & Bandwidth & On-source time &Beam & $\sigma_{\rm rms}$ & $S_{\rm peak}$ \\
(yyyy-mm-dd) & & (MHz) &(min) & (maj, min, PA) & (mJy beam$^{-1}$) &(mJy beam$^{-1}$) \\
\hline
1994-08-12 &BB023 &16 &3.9 &$1.9\times0.9$ mas$^2$, $15.9\degr$ &1.02 &82 \\
1998-09-15 &BU013 &32 &37.7 &$1.8\times1.0$ mas$^2$, $14.8\degr$ &0.14 &117 \\
1999-03-08 &RDV13 &32 &25.2 &$1.0\times0.8$ mas$^2$, $-47.5\degr$ &0.26 &86 \\
1999-12-16 &BU013 &16 &11.2 &$3.0\times1.2$ mas$^2$, $32.2\degr$ &0.16 &135 \\
2000-01-08 &BU013 &16 &13.3 &$3.0\times1.1$ mas$^2$, $57.4\degr$ &0.21 &138 \\
2006-09-03 &BA080 &32 &3.2 &$3.5\times1.0$ mas$^2$, $49.5\degr$ &0.36 &156 \\
2014-06-09 &BG219D &384 &4.0 &$1.5\times1.2$ mas$^2$, $-77.0\degr$ &0.16 &120 \\
2017-01-21 &UF001B &384 &4.7 &$1.1\times1.1$ mas$^2$, $4.0\degr$ &0.20 &142 \\
2018-12-04 &UG002U &384 &6.4 &$1.4\times1.1$ mas$^2$, $15.7\degr$ &0.15 &288 \\
\hline
\end{tabular} \\
Note: Column 5 gives the restoring beam; Columns 6 and 7 present the {\it rms} noise and peak intensity in the image, respectively.
\label{table1}
\end{table*}
\begin{table*}
\centering
\caption{Model fitting results.}
\begin{tabular}{ccccccc}
\hline \hline
Time & Comp. & $S_{\rm int}$ & R & PA &$\theta_{\rm FWHM}$ &$T_{\rm B}$ \\
(yyyy-mm-dd) & & (mJy) &(mas) &(deg) &(mas) &(10$^{10}$ K) \\
(1) & (2) & (3) & (4) & (5) & (6) & (7)
\\ \hline
1994-08-12 &J1 &92$\pm$14 &1.086$\pm$0.023 &$-112.4\pm0.9$ &0.43$\pm$0.05 & 0.9$\pm$0.3 \\
&C &48$\pm$7 &0 &0 &0.21$\pm$0.03 & 2.0$\pm$0.4 \\
1998-09-15 &J1 &132$\pm$9 &1.154$\pm$0.014 & $-111.7\pm0.7$ &0.45$\pm$0.03 & 1.2$\pm$0.2 \\
&C &27$\pm$2 &0 &0 &0.10$\pm$0.02 & $>4.9$ \\
1999-03-08 &J1 &96$\pm$8 &1.117$\pm$0.015 & $-112.4\pm0.8$ &0.29$\pm$0.02 & 1.9$\pm$0.3 \\
&C &49$\pm$5 &0 &0 &0.10$\pm$0.02 & $>8.3$ \\
1999-12-16 &J1 &130$\pm$9 &1.032$\pm$0.015 & $-115.6\pm0.8$ &0.41$\pm$0.03 & 1.4$\pm$0.2 \\
&C &28$\pm$2 &0 &0 &0.10$\pm$0.02 & $>5.0$ \\
2000-01-08 &J1 &121$\pm$9 &1.024$\pm$0.014 & $-113.6\pm0.8$ &0.36$\pm$0.02 & 1.6$\pm$0.2 \\
&C &33$\pm$2 &0 &0 &0.10$\pm$0.02 & $>5.9$ \\
2006-09-03 &J1 &149$\pm$12 &1.154$\pm$0.014 &$-113.5\pm0.7$ &0.51$\pm$0.04 & 1.1$\pm$0.2 \\
&C &51$\pm$4 &0 &0 &0.10$\pm$0.02 & $>6.6$ \\
2014-06-09 &J2 &124$\pm$9 &0.978$\pm$0.014 &$-116.5\pm0.8$ &0.47$\pm$0.03 & 1.0$\pm$0.2 \\
&C &36$\pm$3 &0 &0 &0.10$\pm$0.02 & $>6.1$ \\
2017-01-21 &J2 &155$\pm$12 &1.006$\pm$0.015 &$-118.3\pm0.8$ &0.39$\pm$0.03 & 1.7$\pm$0.3 \\
&C &98$\pm$8 &0 &0 &0.10$\pm$0.02 & $>16.5$ \\
2018-12-04 &J2 &291$\pm$17 &0.982$\pm$0.013 &$-115.9\pm0.8$ &0.36$\pm$0.02 & 3.8$\pm$0.5 \\
&C &117$\pm$7 &0 &0 &0.10$\pm$0.02 & $>19.9$ \\
\hline
\end{tabular}\\
Columns (3) to (7) present in sequence the integrated flux density, radial separation, and position angle with respect to the core C, component size (full width at half maximum of the fitted Gaussian component) and brightness temperature.
\label{table2}
\end{table*}
We obtained the 8-GHz VLBA data at nine epochs from the public archive of the National Radio Astronomy Observatory (NRAO) of the US\footnote{\url{https://science.nrao.edu/observing/data-archive}} and Astrogeo Center\footnote{\url{http://astrogeo.org/} maintained by L. Petrov.}. The data covers a 25-yr long interval from 1994 to 2019. The observational setup is presented in Table~\ref{table1}. The NRAO VLBA data was calibrated and analysed in Astronomical Image Processing System \citep[AIPS,][]{Greisen2003} following the standard procedure described in the cookbook. The Astrogeo data has been calibrated using the VLBI processing software PIMA \citep{Petrov_2011}. The calibrated visibility data sets were further analysed in the Caltech DIFMAP software package \citep{1995BAAS...27..903S} to handle the residual phase errors and imaging. Data with significant scatter were averaged over a time span of 30 s. Bad data points appearing as outliers were mostly caused by the radio frequency interference and thus flagged. The edited data was then self-calibrated and mapped following an automated pipeline and cross-verified through manual interaction. Natural weighting was utilised to create the final maps to obtain higher sensitivity. The images were de-convolved with the restoring beam using the CLEAN algorithm \citep{1974A&AS...15..417H}. The CLEAN-ing was done up to three times the rms noise level of the residual map.
We used circular Gaussian models to fit the self-calibrated visibility data and estimated the integrated flux densities and angular size of the core and jet components. First, a single Gaussian component was fitted on the source. Next, we extracted the model component from the data and created the residual map, from which we continued to fit a second component with a circular Gaussian model. Then, we took these two circular Gaussian components together to fit the original visibility data to ensure that the model best matches the data. The model fitting stopped when there was no visible feature brighter than 5$\sigma$ appearing in the residual map. The archive data was mostly from astrometric or geodetic observations carried out in snapshot mode. The on-source time ranged from 3.2 to 37.7 minutes, and the acquired images were of varying qualities (see Table \ref{table1}). Therefore, we only focused on these two bright components, and did not attempt to fit other weaker emission structures. We also used two elliptical Gaussian models to fit the visibility data and found consistency with the circular Gaussian models. Based on previous VLBI observations \citep{1999ApJ...517L..81U,2009ApJ...706..851R,2017ApJ...836..155R,2020ApJ...891...59R}, the northeast component is identified as the core, and the southwest component is a jet knot. The model-fitting results, including the flux density, size, radial separation with respect to the core and brightness temperature are presented in Table \ref{table2}.
The uncertainties in the model fitting parameters were estimated using the relations given in \citet{2012A&A...537A..70S}. The calculations were based on the approximations introduced by \citet{1999ASPC..180..301F} and were modified for the strong sidelobes induced by incomplete VLBI ($u,v$) coverage. This modified error estimate \citep{2012A&A...537A..70S} was suitable for the current data obtained from snapshot observations and thus had large sidelobes. When the fitted Gaussian size $d$ of a component was smaller than the minimum resolvable size $d_{\rm min}$ of a given interferometer, $d_{\rm min}$ was used in place of $d$. $d_{\rm min}$ can be estimated using the equation given by \citet{2005astro.ph..3225L}:
\begin{equation}
d_{\rm min}=\cfrac{2^{2-\beta}}{\pi}\left[\pi a b \ln2\ln\left(\cfrac{\rm SNR}{\rm SNR-1}\right)\right]^{1/2}
\end{equation}
where $a$, $b$ are the major and minor axes sizes of the point spread function (the synthesised beam), SNR is the signal-to-noise ratio of the component's peak brightness to the image noise, $\beta$ is the power-law index describing the visibility data weighting, i.e., $\beta = 0$ is used for the uniform weighting and $\beta = 2$ for the natural weighting. The uncertainties in the component sizes were estimated as the square-root of quadratic sum of their corresponding statistical and fitting errors. The statistical error in component size was estimated using the relation given in \citet{2012A&A...537A..70S}. The fitting error in component size, on the other hand, was estimated by varying the flux density and size parameters in small steps around the initial fitted values so that the fitted parameters in each step of the modified model were not very different from the initial values and the corresponding reduced chi-squares were close to one. During the fitting all parameters but the model type were kept variable. The standard deviation of the fitted component sizes obtained after each fitting iteration of the most offset modified model was taken as the fitting error in component size. The position errors were then estimated as the half of the component size errors.
\section{Results}\label{sec:3}
\subsection{Radio morphology}\label{sec:3.1}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{J1256+56_X1.pdf}
\includegraphics[width=0.5\textwidth]{J1256+56_X2.pdf}
\caption{VLBI images of Mrk~231 observed at 8~GHz on the epoch 1999 March 8 and 2018 December 4. The images are created with natural weighting. The negative red coloured contour is at $-2\sigma$ level and the positive black-coloured contour levels are in the series of 3$\sigma \times$(1, 2, 4, 8, 16, 32, 64), where $\sigma$ is the \textit{rms} noise in each image. The colour scale shows the intensity in the logarithmic scale.}
\label{VLBI}
\end{figure}
Figure \ref{VLBI} shows the 8-GHz VLBI images of Mrk~231 observed on 1999 March 8 and 2018 December 4. The images of other epochs show a similar morphology but a lower quality and thus are not shown here. The image parameters are presented in Table \ref{table1}. The emission structure has a largest extent of about two mas (corresponding to a projected size of about 1.6 pc) along the northeast--southwest direction, consistent with the previous VLBI observational results \citep[e.g., ][]{1999ApJ...517L..81U,2017ApJ...836..155R}. The northeastern component was identified as the core in previous studies \citep{1999ApJ...517L..81U,2009ApJ...706..851R}. We use the same nomenclature in this paper. In the 8-GHz images, the core is weaker but more compact than the jet component (see Table \ref{table2}); however, the core is brighter at higher frequencies ($\geq 15$ GHz), especially during flaring periods \citep[e.g., ][]{2017ApJ...836..155R}).
The jet component's peak emission in the 8 GHz images is about 1.2 mas ($\sim$1 pc) away from the core. The jet position angle (PA) and the core--jet separation (R) are plotted as a function of time in Figure~\ref{fig2}. We find that the jet features are not easily described with a single component from the image. Instead, two groups of jet components may be inferred: the components between 1994 and 2007 have similar separation ($\sim1.15$ mas) and position angle ($\sim-113\degr$), and are labelled as J1; other components between 2014 and 2019 are labelled as J2 (see Table \ref{table2}). \citet{2017ApJ...836..155R,2020ApJ...891...59R} also detected a jet component K1 at a similar distance and position angle with J2 during the 2015 and 2017 flares. The authors assumed K1 is a stationary component and determined the relative motion of the core toward K1, $0.97\, c$. This motion is attributed to a newly discrete jet ejection that also dominates the total nuclear flux density at 43 GHz. Determining this small displacement ($\sim$0.075 mas) in the inner jet is only possible by virtue of the high-accuracy astrometry achieved by their phase-referencing observations and the high-resolution VLBI images at 43 GHz. In contrast, J1 and J2 in our study are detected from lower-frequency, lower resolution 8-GHz images. Moreover, the 8.4-GHz core suffers from synchrotron self-absorption that hinders the detection of minor changes in the emission structure of the core region. Therefore, our present study focuses on the kinematics of the underlying jet.
\citet{1999ApJ...516..127U} reveals a north-south elongated triple structure with a total extent of 40 pc in their 1.4-GHz VLBA image. The southern lobe is brighter and longer, thus is identified to be associated with the advancing jet. The peak of the southern lobe is $\sim$25 pc ($\sim30$ mas) away from the core at a position angle of $\sim -172\degr$. Figure~\ref{VLBI} displays the inner-parsec jet extending to the southwest direction at a position angle $\sim -115\degr$. The misalignment between the inner 1-pc and the outer 25-pc southern jet is about $60\degr$, and the mean position angle of the VLBI jets is $-144\degr$. The jet geometry is consistent with that of the molecular outflow. The molecular CO (2--1) outflow shows a wide-angle biconical geometry with a size of $\sim$1 kpc along the northeast (red-shifted component) to southwest (blue-shifted component) directions \citep{2015A&A...583A..99F}. In the nuclear region, the OH megamaser emission reveals a dusty molecular torus or thick disc located between 30 and 100 pc from the central engine \citep{2003Natur.421..821K}. The dusty disc's projected position angle is $-145\degr$, in a good alignment with the central axis of the cone cleaned by the jet. The radio jet seems to follow a twisted path within the conical cone and is prominent at the boundaries of the cone.
Both the core and jet components show a compact morphology. The brightness temperatures ($\mathrm{T}_{\rm B}$) of the VLBI components are calculated using the equation 2 of \citet{1982ApJ...252..102C}. The core brightness temperature is remarkably higher than the jet brightness temperature (Table~\ref{table2}), reinforcing the core identification. We should note that, except for the first epoch, the core in the other epochs is unresolved; the size ($\theta_{\rm FWHM}$) is given as an upper limit, and thus the $\mathrm{T}_{\rm B}$ represents a lower limit. The core brightness temperature of Mrk~231 exceeds $10^{10}$~K, substantially higher than typical values for radio-quiet AGN \citep{2005ApJ...621..123U,1996ApJ...468L..91B}. In the flaring state, the core $\mathrm{T}_{\rm B}$ considerably increases, as also found in previous observations. The highest $\mathrm{T}_{\rm B} > 2.0 \times 10^{11}$ K is detected on 2018 October 4 when the flux density is at the maximum among these datasets. An even higher value $\sim 10^{12}$~K was obtained in \citet{2009ApJ...706..851R} during the 2006 flare. These values approach the equipartition temperature limit for flat-spectrum radio-loud quasars \citep[e.g., ][]{2006ApJ...642L.115H}, suggesting a blazar-like flare in Mrk~231 \citep{2013ApJ...776L..21R}. The peculiar radio core properties (variability, brightness temperature) place Mrk~231 at a position between the radio-loud and radio-quiet AGN populations.
\subsection{Jet Proper Motion}
The bottom panel of Figure \ref{fig2} shows the radial distance of the jet component as a function of time. In addition to the archival Astrogeo VLBI data, we also added the published 8-GHz data points from \citet{2009ApJ...706..851R,2017ApJ...836..155R,2020ApJ...891...59R}. We did not include the higher frequency ($\nu \geq 15$ GHz) data since the frequency-dependent opacity is likely to induce systematic errors to the core positions. Compared to the previous proper motions made by \citet{2017ApJ...836..155R,2020ApJ...891...59R}, the 8-GHz data used here cover a more extended time baseline of over 14 yr, allowing to determine (or constrain) the kinematics of the underlying jet flow. As we discussed early in Section \ref{sec:3.1}, J1 and J2 have distinctively different position angles, i.e., PA$_{\rm J1} = -113\degr\pm2\degr$, and PA$_{\rm J2} =-117\degr\pm 2\degr$. If they are associated with the same component, the resultant jet proper motion speed would be $-0.005\pm0.002$ mas yr$^{-1}$ ($-0.014\pm0.006\,c$). Negative proper motion lacks a physical ground; therefore J1 and J2 must be different jet components.
The historical light curve of Mrk~231 shows that the source is in an active state after 2010 \citep{2020ApJ...891...59R}. Several prominent flares are found, peaking in 2011.7, 2013.1, 2015.2 and 2017.9, roughly in a two-year time interval. In contrast, it remained at a relatively lower level of activity from 1994 to 2010. The Mrk~231 jet consists of a slowly moving, continuous background flow and discrete, fast-moving knots which are associated with large episodic flares. The latter might be short-lived and fade rapidly.
The relative positions of K1 and J2 shows a large scattering (Figure \ref{fig2} bottom panel). The possible reason is that the newly-formed flaring jet component in epochs 2015 and 2017 changed the emission structure in the core region. Since the new jet component is still blended with the core, using the core peak position as the reference point would lead to a systematic error at an uncertain level. For this reason, we do not attempt to calculate the proper motion of J2.
We use the data points of J1 from 1994 to 2007 to estimate the proper motion. The linear regression fitting gives the proper motion speed, $\mu_{\rm J1} = 0.004 \pm 0.007$ mas yr$^{-1}$ (or $0.011 \pm 0.020\,c$). This small proper motion value is consistent with those of many other radio-quiet AGN jets in magnitude and indicates a subluminal motion. We mention that J1's apparent speed is substantially lower than the previous measurements for the discrete jet knot \citep{2017ApJ...836..155R,2020ApJ...891...59R}. Our proper motion measurement reinforces the idea that the jet components J1 and J2 detected in the low radio frequency images are associated with the background jet flow. They are distinctive from the superluminal jet knot ejected from the core in the flaring state.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{MRK231_PA.pdf}
\includegraphics[width=0.45\textwidth]{MRK231_pm.pdf}
\caption{The jet position angle change with time (upper panel) and the relative jet distance with respect to the core versus the observing epoch (bottom panel). Symbols are as follows: blue circles -- Astrogeo database \citep[e.g.,][]{2008AJ....136..580P}, light blue triangles -- \citet{1999ApJ...516..127U}, pink hexagon -- \citet{2017ApJ...836..155R}, and green triangle -- \citet{2020ApJ...891...59R}. }
\label{fig2}
\end{figure}
\section{Discussion}\label{sec:4}
The complex jet structure from the inner pc to outer 25-pc scales indicates morphological and kinematic evolution. The sub-relativistic speed of the background jet flow is suggestive of outflows driven from the inner region, powered by the accretion disc winds \cite[e.g.,][]{2021MNRAS.500.2620Y}. This is especially the case for highly luminous AGN where the bolometric luminosity can be super-Eddington, indicative of a high accretion rate and hence a prominent accretion disc \cite[e.g.,][]{2020MNRAS.494.1744Y,2021MNRAS.500.2620Y}. The specific scenarios are elaborated below.
\subsection{Change of the jet structure}\label{sec:4.1}
As shown in Section \ref{sec:3.1}, the position angle of the southern advancing jet has a difference of $60\degr$ from the inner parsec to the outer 25 parsecs. This large change may be enabled by a variety of mechanisms.
\begin{itemize}
\item
deflection of the jet flow caused by a collision between the jet and a massive cloud in the ISM. Such deflections may cause enhanced brightness and polarisation at the sharp bending, especially if the density contrast between the jet and ISM is large \cite[e.g.][]{2019ApJ...873...11J,2020NatCo..11..143A}. Besides the jet structure change, jet-cloud collisions may also result in deceleration and distortion of the jet structure, as has been seen in the radio-loud quasars 3C 43 \citep[][]{2003A&A...403..537C} and 3C 48 \citep[][]{1991Natur.352..313W,2010MNRAS.402...87A}. Such disruption effects should be even prominent in low-power radio jets (e.g., NGC 1068 \citep{1996MNRAS.278..854M} and NGC 4258 \citep{1991ApJ...381..110P}). However, the southern jet/lobe in Mrk~231 at 25 pc shows a well-defined shape after the presumed collision without any signature of distortion \citep{1999ApJ...516..127U}. This scenario may thus not be operational at these scales.
\item
random wandering of the jet head over a large range of position angles. It is possible that the jet impacts on a dense ISM producing an advancing shock which is observed as a radio lobe. The active jet can be deflected by this action and encounter the external medium at different sites until it finds an eventual lower-density location to break through and proceed \cite[e.g.][]{2016A&A...593A..30M}. This wandering behaviour of the jet head is usually used to explain the observed appearance of multiple hotspots in the lobes of the Fanaroff-Riley type II galaxies \citep[e.g., Cygnus A,][]{1985Natur.313...34W}. However, the position angle difference $\sim60\degr$ is too large for the opening angle of a lobe, making this scenario less likely.
\item
rotation of the jet body as influenced by the circum-nuclear disc. This may happen when the jet is not aligned with the rotation axis of the circum-nuclear disc. When the (low-power) jet advances into this disc, a portion of the gas in the disc is entrained into the jet and increases the angular momentum in the perpendicular direction of the jet. The jet body is subject to the differential rotation of the disc through the entrained external gas. This scenario seems possible since the tilted nuclear disc is indeed observed in Mrk~231; the OH megamaser emission \citep{2003Natur.421..821K} reveals a clockwise rotating torus or thick disc located between 30 and 100 pc from the nucleus. The disc axis points at a position angle of $-145\degr$, in a good alignment with the axis of the cone traversed by the curved jet. The observed inner 1-pc and outer 25-pc radio jets coincidentally align with the boundaries between the cone cleaned by the jet and the disc; the mixture and entrainment of external gas into the jet flow at the boundary results in enhanced emission and bending of the jet body following the rotation of the disc. Moreover, the Figure \ref{fig2} upper panel highlights the change of the inner pc-scale jet position angle from $-113^\circ\pm2\degr$ to $-117^\circ\pm2\degr$ in a time interval of $\sim$25 year, providing a hint of the rotation. Continuous monitoring of the VLBI jet structure over a longer time scale would be essential to confirm a smooth and continuous change of the jet direction, as is expected from the rotating jet model.
\item
projection effect of a helical jet. The helical jet flow produces an S-shape in projection on the plane of the sky. The S-shaped structure is seen in nearby Seyfert galaxies \citep[e.g., NGC 4258, ][]{2005ApJ...622..178G} and in small-sized GHz-peaked spectrum (GPS) galaxies \citep[e.g., B0500+019,][]{2001A&A...377..377S}. The driving mechanism of helical jets may include a regular precession of the jet axis \cite[e.g.,][]{2005A&A...431..831L,2012MNRAS.421.1861V} near the base enabled by a sub--pc-scale SMBH binary system \cite[e.g.,][]{1980Natur.287..307B,2016MNRAS.463.1812M}, a tilted or warped disc caused by a spinning black hole \citep[e.g., ][]{1975ApJ...195L..65B,1992MNRAS.258..811P,2005MNRAS.363...49K,2005ApJ...635L..17L}, magneto-hydrodynamic instabilities \cite[e.g.,][]{1987ApJ...318...78H}, or due to developing disc-jet instabilities \cite[e.g.,][]{2003ApJ...584..135L,2010MNRAS.402...87A,2013MNRAS.434.3487A}. For a simple ballistic motion of the precessing jet, the twisted trajectory is a projection of the jet knots ejected at various position angles \citep{1981ApJ...250..464L}. It is likely that the component J1 (1994 -- 2007), the component J2 (2014 -- 2019), and the 25-pc southern lobe, which are at different position angles, are indicative of such episodic behaviour. A SMBH binary model was proposed by \citet{2015ApJ...809..117Y} to account for the deficit of near-UV continuum flux in Mrk 231. If the helical jet is driven by the tidal interaction between the primary and secondary black holes, the inferred precession period is on the order of $10^3$ yr and the mass ratio is less than 0.03, \textit{i.e.}, an intermediate-mass black hole orbits the primary $10^8\, M_\odot$ one. However, the presence of binary black holes lacks convincing observational evidence \citep[e.g.,][]{2016ApJ...825...42V,2016ApJ...829....4L}. Owing to the data used (PA variation) being sparse and spanning a very short time span, a precession of the jet can not be confirmed from the currently available VLBI data. .
\end{itemize}
In summary, the first two models (\textit{i.e.}, deflection due to jet-cloud collision, and random wandering of jet head) lack strong observational support from the existing data. High sensitivity VLBI polarimetric observations of Mrk~231 are necessary to explore the location of the jet bending and the polarisation structure to confirm or exclude these scenarios. Other mechanisms (jet rotation and helical jet) seem possible but warrant further observations and detailed studies. In either scenario, the interactions between the bending jet and the ISM result in large-aperture-angle shocks that sweep the ISM within a cone and expel the gas to move outward. This dynamic process happens at the onset of the large-scale ionised and molecular outflows and may participate in powering the outflows (see more discussion below).
\subsection{Stagnation of the low-power jet in the host galaxy}\label{sec:4.2}
Mrk~231 lacks a large-scale radio jet structure beyond a few kpc \citep{1998AJ....115..928C,1999ApJ...516..127U,1999ApJ...519..185T,2016A&A...593A..30M}. Although there is some extended emission a few hundred mas south of the nucleus, those diffuse features might not be directly related to the jet \citep{2016A&A...593A..30M}. The compact triple structure has a projected extent of $\sim 50$ mas ($\sim 40$ pc) in the north-south direction. The southern lobe is brighter and longer (25 pc in projection), suggesting it is associated with the advancing jet. The sub-kpc source size and symmetry morphology is reminiscent of a compact symmetric object (CSO). The radio power of Mrk~231 is $\log P_{\rm 1.4GHz} = 24.15$ W Hz$^{-1}$ \citep{2016A&A...593A..30M}, placing it in the low-power CSO population \citep{2012ApJ...760...77A}.
The subluminal jet speed measured for the background jet in Mrk~231 is consistent with its low-power nature. The proper motion vector component of the jet head in the south direction is $0.011\, c$. If we take this value as the upper limit of the lobe expansion speed, we can estimate the age of the radio structure to be $\gtrsim$7400 yr. The radio structure's actual age could be much longer than this value due to the obstruction of the jet growth.
The jet structure strongly influences the size and power of the radio source. When the jet head wiggles over a large range of directions, it continuously hits new regions of the ISM. At each direction, the time-averaged jet power is small.
The jet power ($L_{\rm j}$) is composed of radiative power ($L_{\rm j,rad}$) and kinetic power ($L_{\rm j,kin}$). The study of \citep{2014IJMPS..2860188F} employs a statistically viable sample of jetted AGN for a resulting empirical relationship between the constituent jet luminosities $L_{\rm j,rad}$ and $L_{\rm j,kin}$ and the core radio luminosity $L_{\rm radio}$,
\begin{align}
\log L_{\rm j,rad} &= (12\pm 2) + (0.75\pm 0.04)~\log L_{\rm radio}\\ \nonumber
\log L_{\rm j,kin} &= (6\pm 2) + (0.90\pm 0.04)~\log L_{\rm radio}.
\end{align}
Owing to the small physical separation between the jet and core components, and the unresolved nature of the latter, the putative radio core may be brighter. We then take the combined flux density of $\approx$ 400 mJy during the 2018 epoch, an upper limit as representative of that from the radio core. The corresponding radio luminosity is about $1.5\times10^{41}$erg s$^{-1}$, which converts to $\log L_{\rm j,rad} = 42.89$ and $\log L_{\rm j,kin} = 43.06$ using the above equations and hence, a luminosity ratio $L_{\rm j,kin}/L_{\rm j} = L_{\rm j,kin}/(L_{\rm j,kin}+L_{\rm j,rad})$ of 0.6. The balance of momentum flux between the jet and external ISM on interaction and the density contrast can be cast in terms of the kinematic, geometric, and energetic parameters governing the jet and ISM \cite[e.g.,][]{1998ARA&A..36..539F}.
\begin{align}\label{mombal1}
&\frac{L_{\rm j}}{(\beta_j-\beta_h) c} = \kappa \rho_j \beta^2_h c^2 A_h \\
&\kappa = \left(\frac{\beta_j}{\beta_h}-1\right)^2, \label{mombal2}
\end{align}
where $\beta_j$ is the speed of the jet flow, $\beta_h$ is the speed of the advancing jet head at the working surface, and $A_h = r^2_h \sin \theta_0$ is the surface area in contact at the working surface of size $r_h$ and for a jet half opening angle $\theta_0$, and $\kappa$ is the density contrast between the external ISM $\rho_{\rm ISM}$ and the jet flow $\rho_j$ \citep{2020NatCo..11..143A}. The density of the baryonic jet is
\begin{equation}
\rho_j = \frac{L_{\rm j, kin}}{(\pi r^2/2) \beta^3_j c^3 \Gamma^2_j},
\label{rhojet}
\end{equation}
where $r$ is the jet extent and Lorentz factor $\Gamma_j = (1-\beta^2_j)^{-1/2}$. The equations (\ref{mombal1}), (\ref{mombal2}) and (\ref{rhojet}) are used along with $\sin \theta_0 \approx r_h/r$ to cast the density contrast as
\begin{equation}
\kappa = \left(1-\frac{2}{\pi} \left(\frac{L_{\rm j, kin}}{L_{\rm j}}\right)^{1/3} \frac{\sin \theta_0}{\Gamma^{2/3}_j}\right)^{-2},
\end{equation}
which after some algebraic manipulation can be numerically solved for the density contrast $\kappa$ as
\begin{equation}
C \beta^2_h x^5+2 C \beta^2_h x^4+(1-C+C \beta^2_h) x^3+3 x^2+3 x+1 = 0,
\end{equation}
where $C = \left(\frac{2}{\pi} \frac{L_{\rm j,kin}}{L_{\rm j}}\right) \sin^3 \theta_0$ and $x = \kappa^{1/2}$. The contrast ratio $\kappa$ is now a function of jet properties that can be directly inferred from the VLBI observations or constrained. The measured $\beta_h = 0.013 \ c$ is used in the above equation to solve for $\kappa = \kappa \left(L_{\rm j,kin}/L_{\rm j},\theta_0\right)$ and the contours are plotted in \autoref{kappathetaL1}. The luminosity ratio is taken in the range $0.01 - 1.0$ and the jet half opening angle in the range $0.1^\circ - 60^\circ$ for the numerical solution. We obtain $\kappa$ in the range $1.1 - 28.6$ for the above range, indicating a generally denser surrounding medium. This argument of dense external ISM is consistent with the OH megamaser observations \citep{2003Natur.421..821K}. $\kappa$ is larger for wider opening angles, consistent with the expectation for a weaker, relatively un-beamed jet/outflow.
\begin{figure}
\centering
\includegraphics[scale=0.35]{kappathetaL1.pdf}
\caption{Contours of the density contrast as a function of the luminosity ratio $L_{\rm j,kin}/L_{\rm j}$ and jet half opening angle $\theta_0$. The vertical dashed line marks $L_{\rm j,kin}/L_{\rm j} = 0.6$ for Mrk 231 (see discussion in \ref{sec:4.3}) indicating $\kappa \gtrsim 1.2$.}
\label{kappathetaL1}
\end{figure}
Comparing the jet radiative power and kinetic power shows that a substantial fraction of the jet power is in the form of radiation. The jet-ISM interactions lead to a bulk of jet mechanical energy dissipation to the shocks, which in turn power and expel the large-scale nuclear outflow winds/outflows. As a result, the advancing motion of the jet flow is likely to inhibit or slow down. Thus, in a rotating, clumpy, dense ISM environment such as in Mrk~231, the growth of the low-power radio source could be stopped at some distance, as predicted by the frustrated model for compact radio sources \citep{1984AJ.....89....5V,1997ApJ...485..112B,2000MNRAS.319..445S}.
\subsection{Jet contributes to powering the wide-angle outflows}\label{sec:4.3}
Wide-angle outflows on sub-kpc to kpc scales are ubiquitous in radio-quiet quasars \citep{2005ARA&A..43..769V}. Sustenance of the wide-angled outflows at large scales requires the AGN energy and momentum to be transferred into the ISM over a substantial volume through radiative or mechanical processes \citep{2011ApJ...729L..27R,2012ApJ...750...55S}. Radio jets can enable this feedback and facilitate the expulsion of neutral and ionised gas from the nuclear region \citep{2005A&A...444L...9M,2009MNRAS.400..589H,2009ApJ...690..953F,2020arXiv200102675M}.
Mrk~231 hosts powerful multi-phase outflows extending at multiple scales, including the nuclear ultra-fast X-ray winds, ionised gas outflows arising from the broad-line region, and the galactic-scale molecular and atomic gas outflows \citep{2015A&A...583A..99F,2016A&A...593A..30M}. The inferred jet flow speed in this work is slightly lower than the X-ray ultra-fast winds (up to 20000 km s$^{-1}$). The inverse correlation between the radio and far-UV luminosity in Mrk~231 indicates a direct connection between the jet and the BAL winds \citep{2017ApJ...836..155R}.
Unlike the radio-loud quasars which usually have collimated and highly relativistic jet, the jets in radio-quiet quasars are less powerful and uncollimated. The subrelativistic jets are often regarded as winds or outflows \citep{1992ApJ...396..487S}. Large-scale (up to $\sim$10 kpc) loosely collimated outflows are found to be ubiquitous in radio-quiet Type 1 quasars, mostly along the minor axes of the host galaxies, and are thought to be quasar-driven \citep[e.g., ][]{2017ApJ...850...40R}. Mrk~231 sits between the radio-loud and radio-quiet quasars. It is one of the few radio-quiet Type 1 quasars having prominent jets; another such example is PG 1700+518 which shows strong evidence of a jet-driven outflow on kpc scales \citep{2012MNRAS.419L..74Y,2018ApJ...852....8R}. The PG 1700+518 jet lies along with the outflows. Similarly, the Mrk~231 jet and the BAL ionised outflows exist on the same scale. The change of the pc-scale jet structure over a broad range of directions results in advancing shocks that can participate in ionising and accelerating the BAL winds, and in the expulsion of gas at larger scales. The time-averaged kinetic energy flux of the jet is $\sim 10^{41-42}$ erg s$^{-1}$ in the quiescent state, and increases to $\sim 10^{43}$ erg s$^{-1}$ during the flaring state \citep{2009ApJ...706..851R}, which accounts for (1--35)\% of the energy rate of the molecular outflow \citep{2015A&A...583A..99F}. These clues suggest that the contribution of the jet to powering the outflows is not negligible when Mrk~231 is in its flaring state.
\section{Summary}
We analysed archive VLBI data of Mrk 231 and presented the milli-arcsecond resolution images that reveal the inner parsec jet structure. Comparing the published data with ours, we found a large misalignment ($\sim 60\degr$) between the inner pc jet and the outer 25-pc jet. A hint of jet position angle change is seen in the inner pc scale and warrants new observations to confirm. The jet structure can be explained by a scenario in which the jet lies in a large opening angle cone and expands in various directions. The cone swept by the jet is likely to be physically associated with the dusty disc revealed from the previous OH megamaser observations. The jet knots revealed in the 8-GHz VLBI images represent the active working surface of the jet head hitting on the external ISM. The entrainment of the gas from the rotating circum-nuclear disc in the jet may result in the rotation of the jet body, or the jet itself follows a helical trajectory driven by jet precession or other mechanisms. The expansion of the jet within the cone creates large-aperture-angle shocks that accelerate and expel material from the nuclear region. The propagation of the jet energy and momentum outwards further powers the molecular and atomic gas outflows. Although the quasar radiation might be the dominant power source of the AGN winds/outflows, the jet power accounts for a fraction of the energy rate of the molecular outflow in Mrk 231. Thus, the Mrk 231 jet's contribution to the feedback to its nuclear environment, especially in the flaring state, may not be negligible. The jet direction change within the large-opening-angle cone causes the time-averaged kinetic energy and momentum in a specific direction to be significantly lower than a collimated jet flow. Moreover, a substantial fraction of the jet power is dissipated in radiation. Consequently, the Mrk~231 jet is obstructed by the dense ISM within a few tens pc scale and fails to grow into a large-sized radio source.
\section*{Acknowledgements}
We thank the anonymous referee for her/his constructive comments which greatly improved the manuscript. This work is supported by the National Key R\&D Programme of China (2018YFA0404603) and the Chinese Academy of Sciences (CAS, 114231KYSB20170003). S.J. is supported by the CAS-PIFI (grant No. 2020PM0057) postdoctoral fellowship. Y.C.W. thanks the hospitality of Shanghai Astronomical Observatory during her summer internship. The authors acknowledge the use of Astrogeo Center database maintained by L. Petrov. The National Radio Astronomy Observatory are facilities of the National Science Foundation operated under cooperative agreement by Associated Universities, Inc.
\section*{Data availability}
The datasets underlying this article were derived from the public domain in NRAO archive (project codes: BU013, BA080;
\url{https://science.nrao.edu/observing/data-archive}) and Astrogeo archive (project codes: BB023, RDV13, BG219D, UF001B, UG002U;
\url{http://astrogeo.org/}). The calibrated visibility data can be shared on reasonable request to the corresponding author.
\bibliographystyle{mnras}
|
\section{Introduction}
The superfluidity of the liquid helium below the $\lambda$ point is discovered in 1938.
This phenomenon is explained as the Bose-Einstein condensation of the part of atoms.
The liquid helium is the physical object with strong interaction
while the Bose-Einstein condensation is predicted for the ideal Bose gas.
The Bose-Einstein condensation of collections of weakly interacting atoms is achieved in 1995.
The dynamics of the weakly interacting spin-0 scalar Bose-Einstein condensates (BECs) is described by the Gross-Pitaevskii equation
which is the effective single-particle nonlinear Schrodinger equation \cite{Dalfovo RMP 99}.
In the following years the Bose-Einstein condensate of atoms being in several spin states \cite{Stamper-Kurn RMP 13}, \cite{Kawaguchi PR 12}.
Corresponding model in form of the nonlinear Pauli equation is suggested.
The Gross-Pitaevskii equation can be represented in the form of quantum hydrodynamic equations.
The set of hydrodynamic equations consists of the continuity and Euler equation for the potential velocity field for the spin-0 BECs.
The velocity field becomes nonpotential for the spin-1 BECs.
Moreover, additional hydrodynamic functions appear to describe the hydrodynamic properties of the spin-1 BECs.
They are the spin density and the nematic tensor density.
The nematic tensor is proportional to the anticommutator of the spin operators.
It is an independent function for the spin-1 BECs
(except the ferromagnetic phase, where it is reduced to the spin density).
In contrast, the product of the spin operators reduces to its first degree for the spin-1/2 particles.
Overall, the set of hydrodynamic equations is obtained in the regime of small collisions
\cite{Fujimoto PRA 13}, \cite{Szirmai PRA 12}, \cite{Kudo PRA 10}.
The Gross-Pitaevskii equation can be derived from the microscopic quantum mechanics represented in the second quantization form \cite{Alon arXiv 21},
which is one of realizations for the transition from the microscopic to the macroscopic models.
The many-particle quantum hydrodynamics presents another approach giving
the representation of the many-particle wave function via the set of hydrodynamic functions
\cite{Maksimov QHM 99}, \cite{MaksimovTMP 2001}, \cite{Andreev PRA08}, \cite{Andreev IJMP B 13}, \cite{Andreev 2001}, \cite{Andreev 2005}.
The density-functional theory for spin-0 bosons is considered including the GP equation \cite{Saarikoski RMP 10}.
Earlier steps in the theory of weakly interacting BECs includes the derivation of spectrum of collective excitations
called the Bogoliubov spectrum obtained in 1947 \cite{Bogoliubov 47}.
This theory is based on the second quantization method.
Effects of the small temperatures are included by Popov in 1965 \cite{Popov JETP 65}, \cite{Griffin PRB 96}, \cite{Zaremba PRA 98}.
Variety of waves phenomena is observed in BECs.
It includes the density waves \cite{Nakamura PRA 14}, \cite{Roy PRA 14}, \cite{Trukhanova PS 20}
solitons \cite{Andreev LP 19}, \cite{Andreev LP 21}, \cite{Andreev 2009},
vortices
\cite{Gautam PRA 14},
turbulence,
shock waves \cite{Crosta PRA 12}, \cite{Kamchatnov PRA 12}.
Collective excitations are considered in Fermi gas either \cite{Korolyuk PRA 14}.
These phenomena demonstrate more complex behavior in the spinor BECs in comparison with spin-0 BECs
\cite{Fujimoto PRA 13}.
Finite temperature effects in ultracold gases are considered as well
\cite{Roy PRA 14}, \cite{Gautam PRA 14}, \cite{Ticknor PRA 12}, \cite{Liu PRA 03}, \cite{Liu PRA 03 2}.
Presence of the bosons in the excited states is modeled via the second fluid which called the normal fluid.
The interaction
between the atoms in the condensate and the surrounding gas of noncondensed particles
can be also modeled within the kinetic model
\cite{Morandi PRA 13}.
Finite temperature effects are considered in ultracold fermions \cite{Korolyuk PRA 14} either.
Depletion of BECs and quantum fluctuations make
the dynamical properties bosons intermixing with the finite temperature effects more complex \cite{Andreev 2005},
\cite{Blakie PRA 13}, \cite{Pitaevskii PRL 98}, \cite{Braaten PRL 99}, \cite{Astrakharchik PRL 05}, \cite{Xu PRL 06}, \cite{Andreev 2101}.
mean-field potential of the magnetic dipole-dipole interaction is included at the study of the collective excitations in spin-1 BECs \cite{Deng PRA 20}.
Instantaneous quench of quadratic Zeeman shift from
$q_{1}>0$ to $q_{2}<0$ through the quantum phase transition at $q=0$ corresponding to the transition from polar to the antiferromagnetic phases is made experimentally
for the antiferromagnetic spinor BEC \cite{Vinit PRA 17}.
This paper is organized as follows.
In Sec. II the structure of the hydrodynamic equations including basic definitions and major steps of derivation are described.
In Sec. III transition to the two-fluid quantum hydrodynamic model for the finite temperature spin-1 bosons is made.
In Sec. IV the spectrum of collective excitations is considered for the ferromagnetic phase.
In Sec. V a brief summary of obtained results is presented.
\section{Hydrodynamic equations}
The dynamic of the quantum system is governed by the Schrodinger equation
\begin{equation}\label{BECfinTSpin1 Schrodinger equation micro} \imath\hbar\partial_{t}\Psi_{S}=\hat{H}_{SS'}\Psi_{S'}, \end{equation}
where the coordinate representation is chosen.
Therefore, the wave function $\Psi_{S}=\Psi_{S}(R,t)$ is the function of coordinates of $N$ identical particles under consideration
$R=\{\textbf{r}_{1}, ..., \textbf{r}_{N}\}$ representing the configurational space of the system.
The spin-1 bosons are considered.
Hence, the distribution of the probability amplitude is determined by the wave function $\Psi_{S}$
along with the amplitude of the spin projection probability.
The spinor structure of the wave function is reflected in the subindex $\Psi_{S}$,
where spin index $S$ is the short notation for the $N$ indexes of spin of each particle $S=\{s_{1}, ..., s_{N}\}$.
More accurately, the spin-1 particles are the vector particles,
where single particle wave function is the three component column,
while the spinor is the mathematical object invented for the spin-1/2 fermions.
The spin structure of the wave function is the direct product of $N$ three component columns
and the Hamiltonian $\hat{H}=\hat{H}_{SS'}$ acting as the $3N$-rank matrix.
The action of matrix $\hat{H}_{SS'}$ on the "spinor-vector" $\Psi_{S'}$ includes summation on subindex $S'$:
$\sum_{S'=s_{1}',...,s_{N}'}\hat{H}_{SS'}\Psi_{S'}$,
with $s_{i}=0$, $\pm1$,
but equation (\ref{BECfinTSpin1 Schrodinger equation micro}) does not include the summation
since the summation on the repeating index is assumed.
To specify the physical system under consideration requires the explicit form of the Hamiltonian of the system:
$$\hat{H}=\sum_{i=1}^{N}\biggl(\frac{\hat{\textbf{p}}^{2}_{i}}{2m}
+V_{ext}(\textbf{r}_{i},t)-p\hat{F}_{i,z}+q\hat{F}_{i,z}^{2}\biggr)$$
\begin{equation}\label{BECfinTSpin1 Hamiltonian micro}
+\frac{1}{2}\sum_{i,j\neq i}\biggl(U_{1}(\textbf{r}_{i}-\textbf{r}_{j})
+U_{2}(\textbf{r}_{i}-\textbf{r}_{j})\hat{\textbf{F}}_{i}\cdot\hat{\textbf{F}}_{j} \biggr),\end{equation}
where $m$ is the mass of atom,
$\hat{\textbf{p}}_{i}=-\imath\hbar\nabla_{i}$ is the momentum of i-th atom.
The first group of terms is the superposition of the singe-particle Hamiltonians,
which includes the linear Zeeman term proportional to $p$
and the quadratic Zeeman term proportional to $q$.
The last group of terms in the Hamiltonian (\ref{BECfinTSpin1 Hamiltonian micro})
contains two terms.
The scalar spinless part of the short-range boson-boson interaction $U_{1,ij}$
which equally contributes in interaction of atoms with same spin projection.
The spindependent part of the short-range boson-boson interaction $U_{2,ij}$ demonstrates
deviation from potential $U_{1,ij}$ for the interaction of atoms with different spin projections.
The explicit form of the short-range interaction is not specified.
No distinguish between bosons in the BEC state and bosons in other states is made at the microscopic level.
Separation of all bosons on BEC and normal fluid is made in terms of collective variables.
Hamiltonian (\ref{BECfinTSpin1 Hamiltonian micro}) contains the single-particle spin matrixes
$$\hat{\textbf{F}}_{i}=\{\hat{F}_{x}, \hat{F}_{y}, \hat{F}_{z}\}$$
\begin{equation}\label{BECfinTSpin1 spin matrixes} =\begin{array}{ccc}
\frac{1}{\sqrt{2}}\left(
\begin{array}{ccc}
0 & 1 & 0 \\
1 & 0 & 1 \\
0 & 1 & 0 \\
\end{array}
\right),
&
\frac{1}{\sqrt{2}}\left(
\begin{array}{ccc}
0 & -\imath & 0 \\
\imath & 0 & -\imath \\
0 & \imath & 0 \\
\end{array}
\right),
&
\left(
\begin{array}{ccc}
1 & 0 & 0 \\
0 & 0 & 0 \\
0 & 0 & -1 \\
\end{array}
\right).
\end{array}
\end{equation}
The microscopic distribution of spin-1 bosons in the physical space can be obtained by the projection of the distribution in the configurational space
\cite{Maksimov QHM 99}, \cite{Andreev PRA08}, \cite{Andreev 2001}, \cite{Andreev 1912}, \cite{Andreev LPL 18}, \cite{Andreev 2007}
\begin{equation}\label{BECfinTSpin1 concentration n definition}
n=\int \Psi_{S}^{\dagger}(R,t)\sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})\Psi_{S}(R,t)dR. \end{equation}
This definition contains the integration over the configurational space and the summation over the spin indexes.
Same operations are included in the definitions of other hydrodynamic functions presented below.
Each boson has probability to be in quantum state with different projections of spin.
Considering all bosons in the quasi-classic terms we can refer to number of bosons with different spin projections.
Therefore, the concentration of all bosons can be separated on the partial concentrations $n=n_{1}+n_{0}+n_{-1}$,
where subindexes $0$, $\pm1$ refer to the spin projections of bosons.
We consider system of bosons at the arbitrary temperature.
Moreover, we derive equations for the nonequilibrium systems,
where temperature changes in space and time.
Temperature is the characteristic of the collective behavior like the concentration
(\ref{BECfinTSpin1 concentration n definition}) and other hydrodynamic functions presented below.
Therefore, it is not presented in the many-particle wave function describing the microscopic dynamics of quantum system.
Let us calculate the temporal evolution of the concentration (\ref{BECfinTSpin1 concentration n definition}).
We consider the time derivative of the concentration.
The time derivative acts on the wave functions under integral
while the time derivatives of the wave function are replaced by the Hamiltonian
in accordance with the Schrodinger equation (\ref{BECfinTSpin1 Schrodinger equation micro}).
The straightforward calculations lead to the continuity equation
\begin{equation}\label{BECfinTSpin1 continuity equation via j}
\partial_{t}n+\nabla\cdot \textbf{j}=0. \end{equation}
The appearance of the continuity equation is rather obvious conclusion.
However, the derivation of the continuity equation provides the explicit definition for the current in terms of the many-particle wave function:
\begin{equation}\label{BECfinTSpin1 j def}
\textbf{j}=\frac{1}{2m}\int \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})
[\Psi_{S}^{\dagger}(R,t)\hat{\textbf{p}}_{i}\Psi_{S}(R,t)+h.c.]dR, \end{equation}
where $h.c.$ stands for the Hermitian conjugation.
Vector $\textbf{j}$ is the probability current in the physical space
which reduces to the current of particles in the classical limit.
Moreover, vector $\textbf{j}$ is proportional to the momentum density of the quantum system.
Next step is the derivation of the momentum balance equation
which is the preliminary form of the Euler equation
\begin{equation}\label{BECfinTSpin1 j evol eq}
\partial_{t}j^{\alpha}+\partial_{\beta}\Pi^{\alpha\beta}=
-\frac{1}{m}n\partial^{\alpha}V_{ext}+\frac{1}{m}F_{int}^{\alpha}, \end{equation}
where
the summation on the repeating vector indexes is assumed,
and
$$\Pi^{\alpha\beta}=\frac{1}{4m^{2}}\int dR\sum_{i}\delta(\textbf{r}-\textbf{r}_{i})
[\Psi_{S}^{\dagger}(R,t)\hat{p}_{i}^{\alpha}\hat{p}_{i}^{\beta}\Psi_{S}(R,t)$$
\begin{equation} \label{BECfinTSpin1 Pi def}
+(\hat{p}_{i}^{\alpha}\Psi(R,t))_{S}^{\dagger}\hat{p}_{i}^{\beta}\Psi_{S}(R,t)+h.c.] \end{equation}
is the momentum flux,
and $F_{int}^{\alpha}$ is the force field describing the interaction between bosons.
The general form of the interparticle interaction force field appears in the following form:
$$F_{int}^{\alpha}=-\int (\partial^{\alpha}U_{1}(\textbf{r}-\textbf{r}'))n_{2}(\textbf{r},\textbf{r}',t)d\textbf{r}'$$
\begin{equation}\label{BECfinTSpin1 F via n2 S2}
-\int (\partial^{\alpha}U_{2}(\textbf{r}-\textbf{r}'))S_{2}^{\beta\beta}(\textbf{r},\textbf{r}',t)d\textbf{r}', \end{equation}
where
\begin{equation}\label{BECfinTSpin1 n 2 def} n_{2}(\textbf{r},\textbf{r}',t)=
\int \Psi_{S}^{\dagger}\sum_{i,j\neq i}\delta(\textbf{r}-\textbf{r}_{i})\delta(\textbf{r}'-\textbf{r}_{j})\Psi_{S}dR, \end{equation}
and
$$S_{2}^{\alpha\beta}(\textbf{r},\textbf{r}',t)=
\int \Psi_{S}^{\dagger}(R,t)\sum_{i,j\neq i}\delta(\textbf{r}-\textbf{r}_{i})\times$$
\begin{equation}\label{BECfinTSpin1 S 2 def} \times\delta(\textbf{r}'-\textbf{r}_{j})
[\hat{F}_{i}^{\alpha}\hat{F}_{j}^{\beta}\Psi(R,t)]_{S}dR, \end{equation}
with
$[\hat{\textbf{F}}_{i}\cdot\hat{\textbf{F}}_{j}\Psi(R,t)]_{S}$ stands for
$[\hat{\textbf{F}}_{s_{i}s_{i}'}\cdot\hat{\textbf{F}}_{s_{j}s_{j}'}\Psi_{s_{1},...,s_{i}',...,s_{j}',...,s_{N}}(R,t)$.
Approximate calculation of $F_{int}^{\alpha}$ for the short-range interaction in the first order by the interaction radius
\cite{Andreev PRA08}, \cite{Andreev IJMP B 13}, \cite{Andreev 2001}, \cite{Andreev LP 19}
leads to the following result
$$F_{int}^{\alpha}=-\frac{1}{2}g_{1}\partial^{\alpha}(2n_{n}^{2}+4n_{n}n_{B}+n_{B}^{2})$$
\begin{equation}\label{BECfinTSpin1 F via n n S S}
-\frac{1}{2}g_{2}\partial^{\alpha}(2\textbf{S}_{n}^{2}+4\textbf{S}_{n}\textbf{S}_{B}+\textbf{S}_{B}^{2}), \end{equation}
where
\begin{equation}\label{BECfinTSpin1 g i def} g_{i}=\int U_{i}(r)d\textbf{r} \end{equation}
are the interaction constants for the two parts of the short-range interaction potential
presented in the Hamiltonian (\ref{BECfinTSpin1 Hamiltonian micro}),
the subindex $B$ refers to the BEC, the subindex $n$ refers to the normal fluid,
and
\begin{equation}\label{BECfinTSpin1 S def} \textbf{S}(\textbf{r},t)=
\int \Psi_{S}^{\dagger}(R,t)\sum_{i}\delta(\textbf{r}-\textbf{r}_{i})
(\hat{\textbf{F}}_{i}\Psi(R,t))_{S}dR \end{equation}
is the spin density.
Moreover, expression (\ref{BECfinTSpin1 F via n n S S}) is presented in terms of two fluid model.
The functions describing the BEC and the normal fluid enter this equation in nonsymmetric form,
so they are not combined in the functions describing all bosons simultaneously.
This result for the force field gives additional motivation to split hydrodynamic equations on two subsystems.
Equation (\ref{BECfinTSpin1 F via n n S S}) has the following zero temperature limit
\begin{equation}\label{BECfinTSpin1 F via n n S S}
F_{int,BEC}^{\alpha}=-g_{1}n_{B}\partial^{\alpha}n_{B}
-g_{2}\textbf{S}_{B}\partial^{\alpha}\textbf{S}_{B}. \end{equation}
Presence of the spin density (\ref{BECfinTSpin1 S def}) leads to the necessity to derive the equation for its evolution
$$\partial_{t}S^{\alpha}+\partial_{\beta}J^{\alpha\beta}=
-\frac{p}{\hbar}\varepsilon^{\alpha z \gamma}S^{\gamma}
+2\frac{q}{\hbar} \varepsilon^{\alpha z \gamma}N^{z\gamma}$$
\begin{equation}\label{BECfinTSpin1 spin evolution}
+\frac{\varepsilon^{\alpha\beta\gamma}}{\hbar}
\int U_{2}(\textbf{r}-\textbf{r}')S_{2}^{\gamma\beta}(\textbf{r},\textbf{r}',t)d\textbf{r}',\end{equation}
where
\begin{equation}\label{BECfinTSpin1 J alpha beta def} J^{\alpha\beta}=\frac{1}{2m}\int \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})
[\Psi_{S}^{\dagger}\hat{F}_{i}^{\alpha}\hat{p}_{i}^{\beta}\Psi_{S}+h.c.]dR \end{equation}
is the spin current giving the redistribution of particles with no influence of interaction,
and
\begin{equation}\label{BECfinTSpin1 nematic tensor def} N^{\alpha\beta}=\frac{1}{2}\int \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})
\Psi_{S}^{\dagger}(\hat{F}_{i}^{\alpha}\hat{F}_{i}^{\beta}+\hat{F}_{i}^{\beta}\hat{F}_{i}^{\alpha})\Psi_{S}dR \end{equation}
is the density of the macroscopic nematic tensor.
It is defined as the quantum average of the nematic tensor operator
$\hat{N}_{i}^{\alpha\beta}=(1/2)(\hat{F}_{i}^{\alpha}\hat{F}_{i}^{\beta}+\hat{F}_{i}^{\beta}\hat{F}_{i}^{\alpha})$ of the $i$-th particle.
The nematic tensor is the independent function for the spin-1 bosons,
while for the spin-1/2 fermions it reduces to the concentration of particles.
The general form of the spin evolution equation (\ref{BECfinTSpin1 spin evolution}) contains the interparticle interaction in the last term.
However, its calculation in the first order by the interaction radius gives the zero value.
Hence, equation (\ref{BECfinTSpin1 spin evolution}) reduces to
\begin{equation}\label{BECfinTSpin1 spin evolution reduced}
\partial_{t}S^{\alpha}+\partial_{\beta}J^{\alpha\beta}=
-\frac{p}{\hbar}\varepsilon^{\alpha z \gamma}S^{\gamma}
+2\frac{q}{\hbar} \varepsilon^{\alpha z \gamma}N^{z\gamma}.\end{equation}
Therefore, no interaction gives contribution in the spin evolution equations.
The nontrivial evolution of spin density is related to the quadratic Zeeman effects presented via
novel hydrodynamic function: the nematic tensor $N^{\alpha\beta}$.
If we want to obtain the influence of interaction on the spin density evolution
we need to derive the nematic tensor evolution equation.
Hence, we consider the time derivative of function (\ref{BECfinTSpin1 nematic tensor def})
and use the Schrodinger equation (\ref{BECfinTSpin1 Schrodinger equation micro})
with Hamiltonian (\ref{BECfinTSpin1 Hamiltonian micro}) for the time derivatives of the microscopic many-particle wave function.
As the result we find the following equation:
$$\partial_{t}N^{\alpha\beta}+\partial_{\gamma}J_{N}^{\alpha\beta\gamma}
=\frac{p}{\hbar}\varepsilon^{z\alpha\gamma}N^{\beta\gamma}+\frac{p}{\hbar}\varepsilon^{z\beta\gamma}N^{\alpha\gamma}$$
$$-\frac{q}{2\hbar}(\varepsilon^{z\alpha\gamma}\delta^{\beta z}
+\varepsilon^{z\beta\gamma}\delta^{\alpha z})S^{\gamma}$$
$$+\frac{g_{2}}{\hbar}\biggl[ \varepsilon^{\beta\gamma\delta}
\biggl( 2N_{n}^{\alpha\delta}S_{n}^{\gamma} +2N_{n}^{\alpha\delta}S_{B}^{\gamma} +2N_{B}^{\alpha\delta}S_{n}^{\gamma}
+N_{B}^{\alpha\delta}S_{B}^{\gamma}\biggr)$$
\begin{equation}\label{BECfinTSpin1 N eq evol}
+\varepsilon^{\alpha\gamma\delta}
\biggl( 2N_{n}^{\beta\delta}S_{n}^{\gamma} +2N_{n}^{\beta\delta}S_{B}^{\gamma} +2N_{B}^{\beta\delta}S_{n}^{\gamma}
+N_{B}^{\beta\delta}S_{B}^{\gamma}\biggr)\biggr],
\end{equation}
where
$$J_{N}^{\alpha\beta\gamma}=\frac{1}{4m}\int \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})
[\Psi_{S}^{\dagger}(R,t)\times$$
\begin{equation}\label{BECfinTSpin1 J a b g def}
\times(\hat{F}_{i}^{\alpha}\hat{F}_{i}^{\beta}+\hat{F}_{i}^{\beta}\hat{F}_{i}^{\alpha})\hat{p}_{i}^{\gamma}\Psi_{S}(R,t)+h.c.]dR \end{equation}
is the flux of the nematic tensor,
the subindex $N$ refers to the nematic tensor,
and
the last group of terms proportional to the second interaction constant
appears at the calculation of the following term in the first order by the interaction radius
$$\int dR\sum_{i,k\neq i}\delta(\textbf{r}-\textbf{r}_{i})U_{2,ik}\Psi_{S}^{\dagger}(R,t)\times$$
\begin{equation}\label{BECfinTSpin1 interaction for N}
\times(\varepsilon^{\beta\gamma\delta}\hat{N}_{i}^{\alpha\delta}\hat{F}_{k}^{\gamma}
+\varepsilon^{\alpha\gamma\delta}\hat{N}_{i}^{\beta\delta}\hat{F}_{k}^{\gamma})\Psi_{S}(R,t). \end{equation}
The general structure of hydrodynamic equations is obtained above.
However, we need to introduce the velocity field to give these equations more traditionally appearance.
Moreover, this allows to consider the structure of fluxes of all variables.
\subsection{Madelung transformation for the single-particle wave function}
Before we present the structure of the many-particle wave function
let us consider the single particle wave function.
For the spin-1 boson it is the three component column presenting the vector wave function
\begin{equation}\label{BECfinTSpin1 psi general vector} \psi(\textbf{r},t)=\left(
\begin{array}{c}
\psi_{+1} \\
\psi_{0} \\
\psi_{-1} \\
\end{array}
\right),
\end{equation}
where the amplitude $a(\textbf{r},t)$ and phase $S(\textbf{r},t)$ of the wave function can be introduced
together with the unit vector showing the spin state of the particle
\begin{equation}\label{BECfinTSpin1 psi via a S z} \psi=a e^{iS} \hat{z}(\Theta), \end{equation}
where we have unit vector $\hat{z}(\Theta)$ describing the possible spin states.
Making a step ahead we consider the example of the unit vector $\hat{z}(\Theta)$ applied for the description
of the ferromagnetic phase of spin-1 BECs
\begin{equation}\label{BECfinTSpin1 z via angles F} \hat{z}=e^{-\imath\chi}\left(
\begin{array}{c}
\cos^{2}(\theta/2) e^{-\imath\varphi} \\
\frac{1}{\sqrt{2}}\sin(\theta) \\
\sin^{2}(\theta/2) e^{\imath\varphi} \\
\end{array}
\right),
\end{equation}
with
$\Theta=\{\chi,\theta,\varphi\}$,
and
$\chi(\textbf{r},t)$, $\theta(\textbf{r},t)$, $\varphi(\textbf{r},t)$ are the scalar fields,
\emph{and}
$\hat{z}^{\dag}\hat{z}=1$.
Vector (\ref{BECfinTSpin1 z via angles F}) can be obtain by rotation of vector
\begin{equation}\label{BECfinTSpin1 z 100}\hat{z}=\left(
\begin{array}{c}
1 \\
0 \\
0 \\
\end{array}
\right),
\end{equation}
where the rotation is made by matrix
\begin{equation}\label{BECfinTSpin1 Rotation Matrix}
\left(
\begin{array}{ccc}
e^{-\imath\chi}\cos^{2}\frac{\theta}{2} e^{-\imath\varphi} &
-\frac{e^{-\imath\varphi}}{\sqrt{2}} \sin\theta & e^{\imath\chi}\cos^{2}\frac{\theta}{2} e^{-\imath\varphi} \\
e^{-\imath\chi}\frac{1}{\sqrt{2}}\sin\theta & \cos\theta & e^{\imath\chi}\frac{1}{\sqrt{2}}\sin\theta \\
e^{-\imath\chi}\sin^{2}\frac{\theta}{2} e^{\imath\varphi} &
\frac{e^{\imath\varphi}}{\sqrt{2}} \sin\theta & e^{\imath\chi}\sin^{2}\frac{\theta}{2} e^{\imath\varphi} \\
\end{array}
\right).
\end{equation}
The second phase of the spin-1 BECs is the polar phase.
It corresponds to the unit vector obtained by the rotation of vector
\begin{equation}\label{BECfinTSpin1 z 010} \hat{z}=\left(
\begin{array}{c}
0 \\
1 \\
0 \\
\end{array}
\right),
\end{equation}
which leads to
\begin{equation}\label{BECfinTSpin1 z P}\hat{z}=\left(
\begin{array}{c}
-\frac{1}{\sqrt{2}} \sin(\theta)e^{-\imath\varphi} \\
\cos(\theta) \\
\frac{1}{\sqrt{2}} \sin(\theta)e^{\imath\varphi} \\
\end{array}
\right),
\end{equation}
However, these are examples of unit vector $\hat{z}$ from the macroscopic physics.
Or it can be considered as the wave function for the single particle.
We need to make step back to the microscopic description.
\subsection{Introduction of the velocity field and the average spin field via the Madelung transformation}
For the many-particle wave function we have $3N$ dimensional configurational space.
Moreover, the spin part of the wave function is the direct product of $N$ columns of three component each (\ref{BECfinTSpin1 psi general vector}).
Here, we can express the amplitude $a(R,t)$ and phase $S(R,t)$
\begin{equation}\label{BECfinTSpin1 Psi a S zzz} \Psi_{S}(R,t)= a(R,t) e^{\imath S(R,t)}\bigotimes_{i=1}^{N}\hat{z}_{i} \end{equation}
where
$\bigotimes_{i=1}^{N}\hat{z}_{i}=\hat{z}_{1}\otimes \hat{z}_{2} \otimes... \otimes \hat{z}_{N}$,
and
$\hat{z}_{i}=\hat{z}_{i}(\Theta_{i})$,
with
$\Theta_{i}=\Theta_{i}(\textbf{r}_{i},t)$ is the set of three angular fields defined in three dimensional subspace of the configurational space.
Generalized Madelung decomposition (\ref{BECfinTSpin1 Psi a S zzz}) gives the following representation of the concentration:
\begin{equation}\label{BECfinTSpin1 n def via a} n=\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}(R,t). \end{equation}
It is independent of the explicit form of unit vectors $\hat{z}_{i}$.
Other hydrodynamic functions contain the contribution of unit vectors $\hat{z}_{i}$.
The generalized Madelung decomposition (\ref{BECfinTSpin1 Psi a S zzz}) provides the modified expression for the current
$$\textbf{j}=\frac{\hbar}{m}
\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}(R,t)\times$$
\begin{equation}\label{BECfinTSpin1 j def via a S}
\times(\partial_{i}^{\alpha}S(R,t)-\imath\hat{z}_{i}^{\dag}\partial_{i}^{\alpha}\hat{z}_{i}), \end{equation}
where the gradient of phase $S(R,t)$ existing for the spinless particles is shifted by the gradient of the unit vectors.
We have the following specification of equation (\ref{BECfinTSpin1 j def via a S}) for the ferromagnetic phase
$$\textbf{j}=\frac{\hbar}{m}
\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}(R,t)\times$$
\begin{equation}\label{BECfinTSpin1 j def via a S for F}
\times(\partial_{i}^{\alpha}S(R,t)-\partial_{i}^{\alpha}\chi_{i}-\cos\theta_{i}\partial_{i}^{\alpha}\varphi_{i}), \end{equation}
where the velocity of $i$-th particle can be introduced in terms of components of unit vector $\hat{z}_{i}$ (\ref{BECfinTSpin1 z via angles F}):
$v_{i}^{\alpha}=\partial_{i}^{\alpha}S-\partial_{i}^{\alpha}\chi_{i}-\cos\theta_{i}\partial_{i}^{\alpha}\varphi_{i}$.
For the arbitrary unit vector $\hat{z}_{i}$ we also introduce the velocity of $i$-th particle
$v_{i}^{\alpha}=\partial_{i}^{\alpha}S-\imath\hat{z}_{i}^{\dag}\partial_{i}^{\alpha}\hat{z}_{i}$.
Substitution of the wave function via the arbitrary unit vector $\hat{z}_{i}$ contains the contribution of spin of $i$-th particle
$\textbf{s}_{i}=\hat{z}_{i}^{\dag}\hat{\textbf{F}}_{i}\hat{z}_{i}$ under the integral on the configurational space
\begin{equation}\label{BECfinTSpin1 S def via a z} \textbf{S}=\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})
(\hat{z}_{i}^{\dag}\hat{\textbf{F}}_{i}\hat{z}_{i}) a^{2}(R,t). \end{equation}
Other unit vectors $\hat{z}_{j\neq i}$ exclude themselves via the normalization.
At the full polarization of the spin of each particle $\textbf{s}_{i}$ appear via the unit vector:
\begin{equation}\label{BECfinTSpin1 S def via a z F} \textbf{S}=\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})\textbf{n}_{i}a^{2}(R,t), \end{equation}
where
$\textbf{s}_{i}=\textbf{n}_{i}=\textbf{n}(\textbf{r}_{i},t)$$=\{\cos\varphi_{i}\sin\theta_{i}, \sin\varphi_{i}\sin\theta_{i} ,\cos\theta_{i}\}$
is the unit vector
which shows the contribution of the spin of $i$-th particle in the spin density vector field.
We have unit module of each spin,
but spin can have different directions relatively each other.
Hence, it can correspond to the partial spin polarization.
Moreover, the partial spin polarization of the system can be caused by the partial spin polarization of each particle.
The ferromagnetic phase exists at the full polarization of the spin of each particle $\textbf{s}_{i}$
which have same spin polarization.
In the polar phase $\textbf{s}_{i}=0$.
Hence, $\textbf{S}_{p}=0$,
where subindex $p$ stands for the indication of the polar phase.
The last in the line of major hydrodynamic functions necessary for the description of the spin-1 BECs
is the density of the nematic tensor $N^{\alpha\beta}$.
It can be represented by the single-particle nematic tensor
$n^{\alpha\beta}_{i}=$$(1/2)\hat{z}_{i}^{\dag}(\hat{F}_{i}^{\alpha}\hat{F}_{i}^{\beta}$$+\hat{F}_{i}^{\beta}\hat{F}_{i}^{\alpha})\hat{z}_{i}$
under the integral on the configurational space
\begin{equation}\label{BECfinTSpin1 N def via a z} N^{\alpha\beta}=\frac{1}{2}\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})
(\hat{z}_{i}^{\dag}(\hat{F}_{i}^{\alpha}\hat{F}_{i}^{\beta}+\hat{F}_{i}^{\beta}\hat{F}_{i}^{\alpha})\hat{z}_{i}) a^{2}(R,t). \end{equation}
Let us consider the structure of the single-particle nematic tensor $n^{\alpha\beta}_{i}$.
At the full polarization of the spin of each particle $\textbf{s}_{i}=\textbf{n}_{i}$
we have the following simplification of the single-particle nematic tensor
$n^{\alpha\beta}_{i}=(\delta^{\alpha\beta}+n_{i}^{\alpha}n_{i}^{\beta})/2$.
Or it can be rewritten via the spin of $i$-th particle
$n^{\alpha\beta}_{i}=(\delta^{\alpha\beta}+s_{i}^{\alpha}s_{i}^{\beta})/2$.
The transition to the nematic tensor density requires to consider the quantum average of this function on the many-particle wave function.
Hence, in general, the second term does not give the product of two spin densities $\textbf{S}$.
Corresponding nematic tensor is discussed in this section below.
\subsection{Transformation of the fluxes at the Madelung transformation}
Above we consider the time evolution of the concentration $n$, the current $\textbf{j}$,
the spin density $\textbf{S}$, and the nematic tensor $N^{\alpha\beta}$.
The interaction influence their evolution.
However, the fluxes of these functions give their redistribution even at the zero interaction.
To understand the structure of the quantum hydrodynamic equations
we need to consider the representation of the fluxes of basic hydrodynamic variables
at the generalized Madelung decomposition (\ref{BECfinTSpin1 Psi a S zzz}).
The current $\textbf{j}$ is the flux of concentration $n$.
Hence, one flux is considered.
We need to consider the momentum flux $\Pi^{\alpha\beta}$ (\ref{BECfinTSpin1 Pi def}),
the spin current $J^{\alpha\beta}$ (\ref{BECfinTSpin1 J alpha beta def}),
and the flux of nematic tensor $J_{N}^{\alpha\beta\gamma}$ (\ref{BECfinTSpin1 J a b g def}).
We start this part of analysis with the representation of the spin current $J^{\alpha\beta}$ (\ref{BECfinTSpin1 J alpha beta def}):
$$J^{\alpha\beta}=\frac{\hbar}{m}\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})s_{i}^{\alpha}(\partial_{i}^{\beta}S(R,t))a^{2}(R,t)$$
\begin{equation}\label{BECfinTSpin1 J alpha beta def via a z I}
+\frac{1}{2m}\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}(R,t)(z_{i}^{\dag}\hat{F}^{\alpha}_{i}\hat{p}_{i}^{\beta}z_{i}+h.c.).
\end{equation}
The first term in the spin current contains the spin of $i$-th particle,
but it contains a part of the velocity of $i$-th particle.
Let us represent $\partial_{i}^{\beta}S(R,t)$ via the velocity $v_{i}^{\alpha}$.
Hence, we have term which has clear physical meaning of the spin current
via the product of spin $s_{i}^{\alpha}$ and velocity of the same particle $v_{i}^{\beta}$:
$$J^{\alpha\beta}=\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})s_{i}^{\alpha}v_{i}^{\beta}a^{2}(R,t)$$
$$+\frac{\imath\hbar}{m}\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})s_{i}^{\alpha}(z_{i}^{\dag}\partial_{i}^{\beta}z_{i})a^{2}(R,t)$$
\begin{equation}\label{BECfinTSpin1 J alpha beta def via a z II}
+\frac{1}{2m}\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}
(R,t)(z_{i}^{\dag}\hat{F}^{\alpha}_{i}\hat{p}_{i}^{\beta}z_{i}+h.c.).
\end{equation}
Hence, the first term can be called the quasi-classic part of the spin-current $J^{\alpha\beta}_{qc}$,
while other terms are the quantum corrections.
The second example of the fluxes is the momentum flux $\Pi^{\alpha\beta}$ (\ref{BECfinTSpin1 Pi def}).
Let us present the result of application of Madelung decomposition (\ref{BECfinTSpin1 Psi a S zzz}):
$$\Pi^{\alpha\beta}=\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}(R,t)v_{i}^{\alpha}v_{i}^{\beta}$$
$$+\frac{\hbar^{2}}{2m^{2}}\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})
(\partial^{\alpha}_{i}a\cdot\partial^{\beta}_{i}a-a\partial^{\alpha}_{i}\partial^{\beta}_{i}a)$$
$$+\frac{\hbar^{2}}{4m^{2}}\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}
(\partial^{\alpha}_{i}z_{i}^{\dag}\cdot\partial^{\beta}_{i}z_{i}$$
\begin{equation}\label{BECfinTSpin1 Pi def via a v z}
+\partial^{\beta}_{i}z_{i}^{\dag}\cdot\partial^{\alpha}_{i}z_{i}
+4z_{i}^{\dag}\partial^{\alpha}_{i}z_{i}\cdot z_{i}^{\dag}\partial^{\beta}_{i}z_{i}). \end{equation}
The first term presents the quasi-classic part of the momentum flux $\Pi_{qc}^{\alpha\beta}$
which exists in the classic limit(if $\hbar\rightarrow0$).
Other terms are the quantum corrections.
The second term on the right-hand side of equation (\ref{BECfinTSpin1 Pi def via a v z})
is the spinless quantum correction.
The last group of terms is the spin related quantum correction.
Last example of the fluxes is the flux of nematic tensor:
$$J_{N}^{\alpha\beta\gamma}
=\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}(R,t)n_{i}^{\alpha\beta}v_{i}^{\gamma}$$
$$+\frac{\imath\hbar}{m}\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})
n_{i}^{\alpha\beta}(z_{i}^{\dag}\partial_{i}^{\gamma}z_{i})a^{2}(R,t)$$
\begin{equation}\label{BECfinTSpin1 J a b g def via a z}
+\frac{1}{2m}\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})
a^{2}(R,t)(z_{i}^{\dag}\hat{N}^{\alpha\beta}_{i}\hat{p}_{i}^{\gamma}z_{i}+h.c.).
\end{equation}
It also has the quasi-classical and quantum parts.
The "velocities" of different quantum particles give different contributions in the combined current $\textbf{j}$ (\ref{BECfinTSpin1 j def via a S}).
The current $\textbf{j}$ allows to introduce the average velocity (the velocity field) $\textbf{v}=\textbf{j}/n$,
which is the vector field showing local average velocity.
The deviation of velocity of each particle from the average velocity $\textbf{u}_{i}=\textbf{v}_{i}-\textbf{v}$ is introduced.
Chaotic variations from the average velocity can be associated with the thermal motion,while the temperature itself is the average square of velocity of the chaotic part of motion.
Hence, the temperature is proportional to the average of $\textbf{u}_{i}^{2}$.
Similar steps can be made for the analysis of structure of the spin density (\ref{BECfinTSpin1 S def via a z}).
The average spin can be introduced $\textbf{s}=\textbf{S}/n$.
Therefore, the deviation can be introduced $\textbf{t}_{i}=\textbf{s}_{i}-\textbf{s}$ as well.
Moreover, same separation can be made for the nematic tensor $N^{\alpha\beta}$:
$n_{i}^{\alpha\beta}=N^{\alpha\beta}/n+\nu_{i}^{\alpha\beta}$,
where $\nu_{i}^{\alpha\beta}$ is the deviation of the nematic tensor of $i$-th particle from the average value $N^{\alpha\beta}/n$.
The definitions of the spin density (\ref{BECfinTSpin1 S def}) and the nematic tensor density (\ref{BECfinTSpin1 nematic tensor def})
show that the average values of the deviations $\textbf{t}_{i}$ and $\nu_{i}^{\alpha\beta}$ are equal to zero.
Let us consider the structure of the nematic tensor density in the ferromagnetic phase.
Above we find the expression for the nematic tensor of $i$-th particle
via the spin of $i$-th particle in the ferromagnetic phase of the spin-1 bosons
$n^{\alpha\beta}_{i}=(\delta^{\alpha\beta}+s_{i}^{\alpha}s_{i}^{\beta})/2$.
We substitute the spin via the average spin $\textbf{s}=\textbf{S}/n$ and the deviation from the average
$\textbf{t}_{i}=\textbf{s}_{i}-\textbf{S}/n$.
Hence, the nematic tensor of $i$-th particle transforms to
$n^{\alpha\beta}_{i}=(1/2)(\delta^{\alpha\beta}+S^{\alpha}S^{\beta}/n^{2}$
$+(S^{\alpha}/n)t_{i}^{\beta}+(S^{\beta}/n)t_{i}^{\alpha} +t_{i}^{\alpha}t_{i}^{\beta})$.
The last term is constructed of the parameters defined in the local frame.
So, it is associated with the $\nu_{i}^{\alpha\beta}$,
which is the deviation of the nematic tensor of $i$-th particle from the average value $N^{\alpha\beta}/n$
We use it to find the nematic tensor density
$N^{\alpha\beta}=\int dR \sum_{i}\delta(\textbf{r}-\textbf{r}_{i})n^{\alpha\beta}_{i}a^{2}$.
Functions $S^{\alpha}$, $n$ can be easily extracted from under the integral.
We also include
$\int dR \sum_{i}\delta(\textbf{r}-\textbf{r}_{i})t^{\alpha}_{i}a^{2}=0$.
Hence, we find
$N^{\alpha\beta}=(1/2)n(\delta^{\alpha\beta}+S^{\alpha}S^{\beta}/n^{2})$
$+(1/2)\int dR \sum_{i}\delta(\textbf{r}-\textbf{r}_{i})\nu^{\alpha\beta}_{i}a^{2}$.
It is considered above that the module of the spin of each particle $s^{\alpha}_{i}$ is equal to one in the ferromagnetic phase.
Moreover, if we need to get the ferromagnetic state constructed of number of spins all of them should be parallel to each other.
Hence, there is no deviation of each spin from the average $t^{\alpha}_{i}=0$.
Therefore, the last term in the expression for the nematic tensor density $N^{\alpha\beta}$ is equal to zero.
Finally we obtain the well-known expression
$N^{\alpha\beta}=(1/2)(n\delta^{\alpha\beta}+S^{\alpha}S^{\beta}/n)$
(see for instance \cite{Yukawa PRA 12} in text after equation 31).
The first term in equation (\ref{BECfinTSpin1 Pi def via a v z}) is the quasi-classic part of the momentum flux.
We substitute the velocity via two terms $\textbf{v}_{i}=\textbf{u}_{i}+\textbf{v}$.
Next, we use that the average of the velocity in the comoving frame $\textbf{u}_{i}$ is equal to zero.
Therefore, the quasi-classic part of the momentum flux appears as the sum of two terms:
\begin{equation}\label{BECfinTSpin1 Pi via n v p}
\Pi_{qc}^{\alpha\beta}=nv^{\alpha}v^{\beta}+p^{\alpha\beta}, \end{equation}
where $p^{\alpha\beta}$ is the pressure with the following definition
\begin{equation}\label{BECfinTSpin1 pressure def}
p^{\alpha\beta}=\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}(R,t)u_{i}^{\alpha}u_{i}^{\beta}. \end{equation}
The nonequilibrium temperature scalar field is the average of the square of the velocity of the chaotic motion.
Hence, equation (\ref{BECfinTSpin1 pressure def}) shows that
the temperature is proportional to the trace of pressure $p^{\alpha\beta}$:
temperature is $T=p^{\beta\beta}/3n$,
where
$p^{\beta\beta}=p^{xx}+p^{yy}+p^{zz}$,
and
$p^{\beta\beta}=3p$
for the equal equal diagonal elements presented by $p$.
Same separation appears for the quasi-classic parts of the spin current and the nematic tensor flux
\begin{equation}\label{BECfinTSpin1 J alpha beta via S v} J_{qc}^{\alpha\beta}=S^{\alpha}v^{\beta}+J_{th}^{\alpha\beta}, \end{equation}
where we have the thermal part of the spin current tensor
\begin{equation}\label{BECfinTSpin1 J th alpha beta} J_{th}^{\alpha\beta}=
\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}(R,t)t_{i}^{\alpha}u_{i}^{\beta}, \end{equation}
\emph{and}
\begin{equation}\label{BECfinTSpin1 J alpha beta gamma via N v}
J_{N,qc}^{\alpha\beta\gamma}=N^{\alpha\beta}v^{\gamma}+J_{N,th}^{\alpha\beta\gamma}, \end{equation}
where we have the thermal part of the nematic tensor flux
\begin{equation}\label{BECfinTSpin1 J th alpha beta gamma} J_{N,th}^{\alpha\beta\gamma}=
\int dR \sum_{i=1}^{N}\delta(\textbf{r}-\textbf{r}_{i})a^{2}(R,t)\nu_{i}^{\alpha\beta}u_{i}^{\gamma}. \end{equation}
Tensors $J_{th}^{\alpha\beta}$ and $J_{N,th}^{\alpha\beta\gamma}$ similarly to the pressure $p^{\alpha\beta}$
are related to the thermal motion of the quantum particles.
However, the thermal spin current and the thermal nematic tensor flux includes the chaotic motion of spin orientation.
Below, we present the set of two non-linear Schrodinger equations for the BEC and the normal fluid.
It presents the minimal coupling model similar
to the set of non-linear Schrodinger equations existing in literature for the spin-0 bosons \cite{Dalfovo RMP 99}.
So, the contribution of the pressure of the normal fluid is neglected as well.
Equations (\ref{BECfinTSpin1 Pi via n v p}), (\ref{BECfinTSpin1 J alpha beta via S v}) and (\ref{BECfinTSpin1 J alpha beta gamma via N v})
presents the flux of momentum, the flux of spin, and the flux of the nematic tensor in the quasi-classic limit,
so the quantum terms are neglected there.
These expressions are obtained for the arbitrary temperatures.
However, we are interested in the small temperature limit.
First, consider the zero temperature limit.
Hence, no deviations of the velocity, spin, and the nematic tensor from the local average exist.
In our notations the zero temperature leads to
$u_{i}^{\alpha}=0$, $t_{i}^{\alpha}=0$, and $\nu_{i}^{\alpha\beta}=0$.
It also leads to $p_{B}^{\alpha\beta}=0$, $J_{th,B}^{\alpha\beta}=0$, and
$J_{th,B}^{\alpha\beta\gamma}=0$,
where the subindex $B$ refers to the BEC state.
In the nonzero temperature regime in the small temperature limit
functions
$p^{\alpha\beta}$, $J_{th}^{\alpha\beta}$, and
$J_{th}^{\alpha\beta\gamma}$ have nonzero value.
Moreover, they are associated with the normal fluid.
The value of parameters
$u_{i}^{\alpha}$, $t_{i}^{\alpha}$, and $\nu_{i}^{\alpha\beta}$
have the following behavior.
If we consider their value from the quasi-classic point of view we can state that for some particles being in the BEC state these parameters are equal to zero.
While for other particles being in the excited states these parameters are non zero,
and it leads to nonzero value of functions $p^{\alpha\beta}$, $J_{th}^{\alpha\beta}$, and
$J_{th}^{\alpha\beta\gamma}$.
However, more accurate quantum picture is a bit different.
Each boson at small temperature has probability to be in the BEC state and in some excited states.
Hence, parameters $u_{i}^{\alpha}$, $t_{i}^{\alpha}$, and $\nu_{i}^{\alpha\beta}$ can be nonzero for all particles,
but their values are reduced by the partial probability to find each particle in the excited states.
\section{Transition to the two-fluid hydrodynamics}
Basic hydrodynamic functions are additive on the particles on the system.
Therefore, they are additive on the subsystems.
So, we have the separation of the concentration $n=n_{B}+n_{n}$,
where subindex $B$ refers to the Bose-Einstein condensate,
and subindex $n$ refers to the normal fluid.
Same is true for the current $\textbf{j}=\textbf{j}_{B}+\textbf{j}_{n}$,
the momentum flux $\Pi^{\alpha\beta}=\Pi^{\alpha\beta}_{B}+\Pi^{\alpha\beta}_{n}$,
the spin density $\textbf{S}=\textbf{S}_{B}+\textbf{S}_{n}$,
the spin current $J^{\alpha\beta}=J^{\alpha\beta}_{B}+J^{\alpha\beta}_{n}$,
the nematic tensor $N^{\alpha\beta}=N^{\alpha\beta}_{B}+N^{\alpha\beta}_{n}$,
and the flux of the nematic tensor
$J_{N}^{\alpha\beta\gamma}=J^{\alpha\beta\gamma}_{N,B}+J^{\alpha\beta\gamma}_{N,n}$.
Logically, first we separate the basic hydrodynamic functions on two subsystems.
Second, we introduce the velocity field in each subsystem in accordance with the method described above.
It leads to the following set of equations
\begin{equation}\label{BECfinTSpin1 continuity equation via j a}
\partial_{t}n_{a}+\nabla\cdot \textbf{j}_{a}=0, \end{equation}
\begin{equation}\label{BECfinTSpin1 j evol eq a}
\partial_{t}j_{a}^{\alpha}+\partial_{\beta}\Pi_{a}^{\alpha\beta}=
-\frac{1}{m}n_{a}\partial^{\alpha}V_{ext}+\frac{1}{m}F_{int,a}^{\alpha}, \end{equation}
\begin{equation}\label{BECfinTSpin1 spin evolution reduced a}
\partial_{t}S_{a}^{\alpha}+\partial_{\beta}J_{a}^{\alpha\beta}=
-\frac{p}{\hbar}\varepsilon^{\alpha z \gamma}S_{a}^{\gamma}
+\frac{2q}{\hbar} \varepsilon^{\alpha z \gamma}N_{a}^{z\gamma},\end{equation}
and
$$\partial_{t}N_{a}^{\alpha\beta}+\partial_{\gamma}J_{N,a}^{\alpha\beta\gamma}
=\frac{p}{\hbar}\varepsilon^{z\alpha\gamma}N_{a}^{\beta\gamma}+\frac{p}{\hbar}\varepsilon^{z\beta\gamma}N_{a}^{\alpha\gamma}$$
\begin{equation}\label{BECfinTSpin1 N eq evol a}
-\frac{q}{2\hbar}(\varepsilon^{z\alpha\gamma}\delta^{\beta z}
+\varepsilon^{z\beta\gamma}\delta^{\alpha z})S_{a}^{\gamma}+F_{N,a}^{\alpha\beta},
\end{equation}
where $a$ stands for $B$ and $n$,
$F_{int,a}^{\alpha}$ is the force field acting on species $a$ being caused by both species,
$F_{N,a}^{\alpha\beta}$ is the second rank force field tensor giving the contribution of the short-range interaction in the nematic tensor evolution equation
which acts on species $a$.
Partial fluxes can be represented via the partial velocity fields in accordance with equations derived above:
$\textbf{j}=\textbf{j}_{B}+\textbf{j}_{n}=n_{B}\textbf{v}_{B}+n_{n}\textbf{v}_{n}$,
$\Pi^{\alpha\beta}_{a}=n_{a}v_{a}^{\alpha}v_{a}^{\beta}+p_{a}^{\alpha\beta}$,
$p_{B}^{\alpha\beta}=0$,
$J^{\alpha\beta}_{a}=S_{a}^{\alpha}v_{a}^{\beta}$,
$J^{\alpha\beta\gamma}_{N,a}=N_{a}^{\alpha\beta}v_{a}^{\gamma}$,
where the thermal part of the spin current,
and the nematic tensor flux are neglected.
The force fields existing in the partial Euler equations
appear from the expression (\ref{BECfinTSpin1 F via n n S S}),
where the function under the derivative appears as the source of the force,
while the second function in the term present the subsystem moving under action of this force
$$F_{int,B}^{\alpha}=
-g_{1}n_{B}\partial^{\alpha}n_{B}
-2g_{1}n_{B}\partial^{\alpha}n_{n}$$
\begin{equation}\label{BECfinTSpin1}
-g_{2}\textbf{S}_{B}\partial^{\alpha}\textbf{S}_{B}
-2g_{2}\textbf{S}_{B}\partial^{\alpha}\textbf{S}_{n},
\end{equation}
and
$$F_{int,n}^{\alpha}=-2g_{1}n_{n}\partial^{\alpha}n_{B}
-2g_{1}n_{n}\partial^{\alpha}n_{n}$$
\begin{equation}\label{BECfinTSpin1}
-2g_{2}\textbf{S}_{n}\partial^{\alpha}\textbf{S}_{B}
-2g_{2}\textbf{S}_{n}\partial^{\alpha}\textbf{S}_{n}. \end{equation}
The second rank force field tensors have the following form in accordance with the separation of equation (\ref{BECfinTSpin1 N eq evol}):
$$F_{N,B}^{\alpha\beta}=\frac{g_{2}}{\hbar}\biggl[ \varepsilon^{\beta\gamma\delta}
N_{B}^{\alpha\delta}\biggl( 2S_{n}^{\gamma}+S_{B}^{\gamma} \biggr)$$
\begin{equation}\label{BECfinTSpin1 F a b BEC} +\varepsilon^{\alpha\gamma\delta}
N_{B}^{\beta\delta}\biggl( 2S_{n}^{\gamma}+S_{B}^{\gamma}\biggr)\biggr], \end{equation}
and
$$F_{N,n}^{\alpha\beta}=2\frac{g_{2}}{\hbar}\biggl[ \varepsilon^{\beta\gamma\delta}
N_{n}^{\alpha\delta}\biggl(S_{n}^{\gamma}+S_{B}^{\gamma} \biggr)$$
\begin{equation}\label{BECfinTSpin1 F a b NF} +\varepsilon^{\alpha\gamma\delta}
N_{n}^{\beta\delta}\biggl(S_{n}^{\gamma} +S_{B}^{\gamma} \biggr)\biggr]. \end{equation}
In equations
(\ref{BECfinTSpin1 F a b BEC}) and (\ref{BECfinTSpin1 F a b NF})
the spin densities $S_{B}$ and $S_{n}$
present the source of the force field acting on the nematic tensor of BEC in
(\ref{BECfinTSpin1 F a b BEC})
and the normal fluid in (\ref{BECfinTSpin1 F a b NF}).
Further calculation of the quantum terms in
(\ref{BECfinTSpin1 J alpha beta def via a z II}),
(\ref{BECfinTSpin1 Pi def via a v z}),
(\ref{BECfinTSpin1 J a b g def via a z})
allows to obtain their approximate explicit forms in term of the hydrodynamic functions.
Corresponding results for the BEC can be found in Ref. \cite{Yukawa PRA 12}
(see eq. 24 for the spin current,
eq. 26 for the current of the nematic tensor,
the divergence of the momentum flux is presented in the Euler equation (28),
its quantum part is proportional to the square of the Planck constant).
Same structures can be approximately used for the normal fluid as the equations of state for the corresponding functions.
It is necessary to mention that the quantum terms are required
to get NLSEs we need to use the explicit form of the quantum terms.
\subsection{Nonlinear Schrodinger equations}
The Gross-Pitaevskii equation is the famous example of the non-linear Schrodinger equations
used for the modeling of spin-0 neutral BECs.
If we consider the microscopic dynamics of spin-0 BECs
we can derive the quantum hydrodynamic equations by the method presented above (see also \cite{Andreev PRA08}).
Further construction of the macroscopic wave function
using the hydrodynamic functions allows to derive the Gross-Pitaevskii equation from the quantum hydrodynamic equations.
This step complites the derivation of the Gross-Pitaevskii equation from the microscopic quantum motion by the quantum hydrodynamic method
\cite{Maksimov QHM 99}, \cite{MaksimovTMP 2001}, \cite{Andreev PRA08}.
If we consider particles with the nonzero spin
we do not derive the non-linear Schrodinger equation from the single fluid (if we consider the BEC only) model.
However, we construct the non-linear Schrodinger equation
in order to give same hydrodynamic equations like the quantum hydrodynamics obtained from the microscopic motion.
Same method is used to obtain the pair of the non-linear Schrodinger equations for the BEC and the normal fluid of the spin-1 ultracold bosons.
Therefore, the non-linear Schrodinger equations have the following form:
$$\imath\hbar\partial_{t}\hat{\Phi}_{B}=\Biggl(-\frac{\hbar^{2}\nabla^{2}}{2m}+V_{ext}
-p\hat{F}_{z}+q\hat{N}_{zz}$$
\begin{equation}\label{BECfinTSpin1 GP B}
+g(n_{B}+2n_{n})+g_{2}(\textbf{S}_{B}+2\textbf{S}_{n})\hat{\textbf{F}}\Biggr)\hat{\Phi}_{B}
,\end{equation}
and
$$\imath\hbar\partial_{t}\hat{\Phi}_{n}=\Biggl(-\frac{\hbar^{2}\nabla^{2}}{2m}+V_{ext}
-p\hat{F}_{z}+q\hat{N}_{zz}$$
\begin{equation}\label{BECfinTSpin1 GP like nf}
+2g(n_{B}+n_{n})+2g_{2}(\textbf{S}_{B}+\textbf{S}_{n})\hat{\textbf{F}}\Biggr)\hat{\Phi}_{n},\end{equation}
where
the three component vector macroscopic wave functions have the following structure
$\hat{\Phi}_{a}=\sqrt{n_{a}}e^{\imath m\phi_{a}/\hbar}\hat{z}_{a}$.
The concentrations $n_{a}$ is the square of module of the macroscopic wave function
$n_{a}=\hat{\Phi}_{a}^{\dag}\hat{\Phi}_{a}$.
The spin densities $\textbf{S}_{a}$ are also represented via the macroscopic wave function
$\textbf{S}_{a}=\hat{\Phi}_{a}^{\dag}\hat{\textbf{F}}\hat{\Phi}_{a}$.
Other hydrodynamic functions like $j^{\alpha}$, $\Pi^{\alpha\beta}$, $N^{\alpha\beta}$, $J^{\alpha\beta}$, and $J_{N}^{\alpha\beta\gamma}$
are also expressed via the macroscopic wave functions $\hat{\Phi}_{a}$.
Corresponding equations are shown in Appendix.
\section{Collective excitations for the ferromagnetic phase: quasi-classic regime}
We focused on the part of the spectrum of the bulk collective excitations caused by the interparticle interaction.
Hence, the spin textures related to the quantum kinematic motion is neglected.
The ferromagnetic equilibrium state (phase) is considered.
Hence, the nematic tensor density is constructed of the concentration of bosons $n_{a}$ and the spin density $\textbf{S}_{a}$.
\subsection{Zero temperature limit}
Before we present the spectrum of the collective excitations existing at the finite temperatures
we calculate the zero temperature spectrum appearing from the presented model of spin-1 BECs.
Let us consider the equilibrium state of uniform $n_{B0}=const$
and the macroscopically motionless $\textbf{v}_{B0}=0$ BECs.
In the ferromagnetic phase the spin density is also nonzero $\textbf{S}_{B0}\neq0$.
In order to study the small amplitude collective excitations
we present the approximate form of the hydrodynamic equations containing terms up to first order on the small amplitude:
\begin{equation}\label{BECfinTSpin1 cont lin} \partial_{t}\delta n_{B}+n_{0B}\nabla\delta \textbf{v}_{B}=0, \end{equation}
\begin{equation}\label{BECfinTSpin1 Euler lin} mn_{0B}\partial_{t}\delta \textbf{v}_{B}
=-g_{1}n_{0B}\nabla n_{B}
-g_{2}S_{0B}^{\beta}\nabla\delta S_{B}^{\gamma}, \end{equation}
and
$$\partial_{t}\delta S_{B}^{\alpha}+S_{0B}^{\alpha}(\nabla\delta \textbf{v}_{B})$$
\begin{equation}\label{BECfinTSpin1 Spin evol lin}
=-\frac{p}{\hbar}\varepsilon^{\alpha z\gamma}(S^{\gamma}_{0B}+\delta S_{B}^{\gamma})
+2\frac{q}{\hbar}\varepsilon^{\alpha z\gamma}(N_{0B}^{z\gamma}+\delta N_{B}^{z\gamma}), \end{equation}
where
\begin{equation}\label{BECfinTSpin1 N a b via n S} N_{B}^{\alpha\beta}=\frac{1}{2}\delta^{\alpha\beta}n_{B}+\frac{1}{2}S_{B}^{\alpha}S_{B}^{\beta}/n_{B}. \end{equation}
It gives the equilibrium expression
\begin{equation}\label{BECfinTSpin1 N eq a b via n S}
N_{0B}^{\alpha\beta}=\frac{1}{2}\delta^{\alpha\beta}n_{0B}+\frac{1}{2}S_{0B}^{\alpha}S_{0B}^{\beta}/n_{0B} ,\end{equation}
and the linear perturbation
$$\delta N_{B}^{\alpha\beta}=\frac{1}{2}\delta^{\alpha\beta}\delta n_{B}
+\frac{1}{2}S_{0B}^{\alpha}\delta S_{B}^{\beta}/n_{0B} $$
\begin{equation}\label{BECfinTSpin1 N pert a b via n S}
+\frac{1}{2}S_{0B}^{\beta}\delta S_{B}^{\alpha}/n_{0B}
-\frac{1}{2}(S_{0B}^{\alpha}S_{0B}^{\beta}/n_{0B}^{2})\delta n_{B}. \end{equation}
Equilibrium part of equation (\ref{BECfinTSpin1 Spin evol lin}) existing on the right-hand side and has the following form
\begin{equation}\label{BECfinTSpin1} \varepsilon^{\alpha z\gamma}(-p S^{\gamma}_{0B}+q\delta^{z\gamma}+qS_{0B}^{z} S_{0B}^{\gamma}/n_{0B}). \end{equation}
The second term is equal to zero since $\varepsilon^{\alpha z\gamma}\delta^{z\gamma}= \varepsilon^{\alpha zz}=0$.
The rest should be equal to zero in order to satisfy the hydrodynamic equations with the chosen equation of state.
It is satisfied for $\textbf{S}_{0B}=S_{0B} \textbf{e}_{z}$.
Moreover, all spin are parallel to each other in the ferromagnetic phase,
hence $S_{0B}=n_{0B}$.
We consider the propagation of the plane waves
which travel in the arbitrary direction $\textbf{k}=\{k_{x}, k_{y}, k_{z}\}$.
Hence, the perturbations have the following structure illustrated within the concentration
$\delta n=N\cdot e^{-\imath\omega t+\imath \textbf{k} \textbf{r}}$,
where $N$ is the constant amplitude.
Therefore, equations are transformed to the algebraic equations:
\begin{equation}\label{BECfinTSpin1} \omega\delta n_{B}=n_{0B} (\textbf{k}\delta \textbf{v}_{B}), \end{equation}
\begin{equation}\label{BECfinTSpin1} mn_{0B}\omega\delta \textbf{v}_{B}=n_{0B}\textbf{k}(g_{1}\delta n_{B}+g_{2}\delta S_{B}^{z}), \end{equation}
and
\begin{equation}\label{BECfinTSpin1} \omega\delta S_{B}^{z}=S_{0B}^{z}(\textbf{k}\delta \textbf{v}_{B}),\end{equation}
\emph{and} two other equations
\begin{equation}\label{BECfinTSpin1} -\imath\omega\delta S_{B}^{x}=(p-q)\delta S_{B}^{y}, \end{equation}
and
\begin{equation}\label{BECfinTSpin1} -\imath\omega\delta S_{B}^{y}=-(p-q)\delta S_{B}^{x}. \end{equation}
Set of hydrodynamic equations splits on two sets of equation.
One for $\delta n_{B}$, $\delta \textbf{v}_{B}$, and $\delta S_{B}^{z}$
which gives the acoustic wave spectrum:
\begin{equation}\label{BECfinTSpin1} \omega^{2}=(g_{1}+g_{2})n_{0}k^{2}/m . \end{equation}
Another set is for $\delta S_{B}^{x}$ and $\delta S_{B}^{y}$
which give the constant frequency in the quasi-classic limit:
\begin{equation}\label{BECfinTSpin1} \omega=\mid q-p\mid/\hbar. \end{equation}
Since the nematic tensor density is reduced to the concentration and the spin density
we do not consider the nematic tensor evolution equation.
However, the nematic tensor evolution equation provides additional wave solution for the partially spin polarized BECs,
where the deformation mode appears.
It is related to elements $N^{xx}$, $N^{yy}$, and $N^{xy}$.
Moreover, the nematic tensor evolution contributes in the dynamics of the spin wave via the evolution of elements $N^{xz}$ and $N^{yz}$ for the partially spin polarized BECs.
\subsection{Spectrum in two-fluid hydrodynamics of the BEC and normal fluid}
We consider the plane wave small amplitude perturbations of the equilibrium state of the ferromagnetic phase of the finite temperature spin-1 bosons.
The equilibrium state is characterized by two constant concentrations $n_{0B}$, $n_{0n}$, zero velocity fields $\textbf{v}_{0B}=0$, $\textbf{v}_{0n}=0$,
and the spin densities $\textbf{S}_{0B}=S_{0B} \textbf{e}_{z}$ and $\textbf{S}_{0n}=S_{0n} \textbf{e}_{z}$,
where $S_{0B}=n_{0B}$ and $S_{0n}=n_{0n}$.
Hence, we have the following set of equations for the perturbations:
\begin{equation}\label{BECfinTSpin1} \omega\delta n_{a}=n_{0a} (\textbf{k}\delta \textbf{v}_{a}), \end{equation}
$$mn_{0B}\omega\delta \textbf{v}_{B}=n_{0B}\textbf{k}(g_{1}\delta n_{B}+g_{2}\delta S_{B}^{z})$$
\begin{equation}\label{BECfinTSpin1}
+2n_{0B}\textbf{k}(g_{1}\delta n_{n}+g_{2}\delta S_{n}^{z}), \end{equation}
$$mn_{0n}\omega\delta \textbf{v}_{n}=2n_{0n}\textbf{k}(g_{1}\delta n_{B}+g_{2}\delta S_{B}^{z})$$
\begin{equation}\label{BECfinTSpin1}
+2n_{0n}\textbf{k}(g_{1}\delta n_{n}+g_{2}\delta S_{n}^{z}), \end{equation}
and
\begin{equation}\label{BECfinTSpin1} \omega\delta S_{a}^{z}=S_{0a}^{z}(\textbf{k}\delta \textbf{v}_{a}),\end{equation}
\emph{and} for equations for the spin projection evolution
\begin{equation}\label{BECfinTSpin1} -\imath\omega\delta S_{a}^{x}=(p-q)\delta S_{a}^{y}, \end{equation}
and
\begin{equation}\label{BECfinTSpin1} -\imath\omega\delta S_{a}^{y}=-(p-q)\delta S_{a}^{x}. \end{equation}
The spin waves for the BEC and for the normal fluid have same frequency in the ferromagnetic phase:
\begin{equation}\label{BECfinTSpin1} \omega=\mid q-p\mid/\hbar. \end{equation}
We have the following set of equations for the perturbations of the concentrations of the BEC and the normal fluid
\begin{equation}\label{BECfinTSpin1} m\omega^{2}\delta n_{B}=n_{0B}k^{2}(g_{1}+g_{2})[\delta n_{B} +2\delta n_{n}], \end{equation}
and
\begin{equation}\label{BECfinTSpin1} m\omega^{2}\delta n_{n}=2n_{0n}k^{2}(g_{1}+g_{2})[\delta n_{B} +\delta n_{n}]. \end{equation}
Presence of the spin density projection in the direction of the equilibrium spin polarization $\delta S_{a}^{z}=\delta n_{a}$
changes the speed of two sounds via the shift of the interaction constant to $g_{1}+g_{2}$
in the same way as it is happens for the zero temperature.
\section{Conclusion}
The spin-1 BECs have been actively studied over 20 years.
Corresponding non-linear Schrodinger equations have been used in the large part of these studies.
This paper has been focused on the derivation of the well-known model from the microscopic motion of the quantum particles,
where the many-particle Schrodinger equation in the coordinate representation has been used as the starting point of the derivation.
Moreover, the generalization of the existing minimal coupling model has been made to include the contribution of the small temperatures.
Therefore, the system of spin-1 bosons in presented as the system of two quantum liquids
(more like two collective degrees of freedom of single species)
the BEC presenting particles in the quantum states with the lowest energy and the normal quantum fluid presenting particles being in the quantum states with energies above the minimal energy.
Hydrodynamic equations are derived for the arbitrary spin state of liquid.
However, the small temperature ferromagnetic phase and the polar phase are discussed as the limit regimes of the general model.
The set of hydrodynamic equations obtained from the Schrodinger equation leads to the set of $2\times N\times3$ functions
(scalar functions or the elements of tensors),
where $N$ is the number of particles,
the multiplier "2" comes from the fact that
spinless part of evolution of each particle is described by single complex function identical to two real functions,
the multiplier "3" comes from three projections of spin for each particle.
So, it is almost "infinite" number of equations.
Therefore, it should be truncated.
The step to make the truncation is made in accordance with the existing with the minimal coupling model
giving the non-linear Schrodinger equation with three component vector-spinor macroscopic wave function.
Therefore,the evolution of the concentration, velocity field, spin density, and the nematic tensor density are derived.
These equations are derived for the arbitrary temperature and the arbitrary interaction strength.
Next, the term describing the interaction are calculated
in the small temperature limit for the weakly interacting particles with the small range of action of the interaction potential.
Finally, each equation is separated on two parts:
one for the particles in the BEC state,
another for the normal fluid.
Hence, the two-fluid model of the small-temperature spin-1 bosons has been obtained.
These equations have been used to study the bulk small amplitude collective excitations in the form of the plane waves in the infinite medium.
\section{Acknowledgements}
Work is supported by the Russian Foundation for Basic Research (grant no. 20-02-00476).
This paper has been supported by the RUDN University Strategic Academic Leadership Program.
\section{Appendix: Hydrodynamic functions in terms of macroscopic wave functions}
During the derivation of the hydrodynamic equations from the microscopic model
we present the definitions of the hydrodynamic functions via the many-particle microscopic wave functions.
However, the introduction of the non-linear Schrodinger equations for the effective macroscopic wave function requires the presentation of the relation
of the hydrodynamic functions with this effective wave function.
The concentration and the spin density are presented above,
while other functions are given here in the appendix:
\begin{equation}\label{BECfinTSpin1} j_{a}^{\alpha}=\frac{1}{2m}(\hat{\Phi}_{a}^{\dag}\hat{p}^{\alpha}\hat{\Phi}_{a}+h.c.), \end{equation}
\begin{equation}\label{BECfinTSpin1} \Pi_{a}^{\alpha\beta}=\frac{1}{4m^{2}}(\hat{\Phi}_{a}^{\dag}\hat{p}^{\alpha}\hat{p}^{\beta}\hat{\Phi}_{a}
+(\hat{p}^{\alpha}\hat{\Phi})_{a}^{\dag}\hat{p}^{\beta}\hat{\Phi}_{a}+h.c.), \end{equation}
\begin{equation}\label{BECfinTSpin1} N_{a}^{\alpha\beta}=\frac{1}{2}\hat{\Phi}_{a}^{\dag}(\hat{F}^{\alpha}\hat{F}^{\beta}+\hat{F}^{\beta}\hat{F}^{\alpha})\hat{\Phi}_{a}, \end{equation}
\begin{equation}\label{BECfinTSpin1} J_{a}^{\alpha\beta}=\frac{1}{2m}(\hat{\Phi}_{a}^{\dag}\hat{F}^{\alpha}\hat{p}^{\beta}\hat{\Phi}_{a}+h.c.), \end{equation}
and
\begin{equation}\label{BECfinTSpin1} J_{N,a}^{\alpha\beta\gamma}
=\frac{1}{4m}(\hat{\Phi}_{a}^{\dag}(\hat{F}^{\alpha}\hat{F}^{\beta}+\hat{F}^{\beta}\hat{F}^{\alpha})\hat{p}^{\gamma}\hat{\Phi}_{a}+h.c.). \end{equation}
|
\section{Introduction}
We study voting in two-tier voting systems. Suppose the population of a state or union of states is subdivided
into $M$ groups (member states for example). Each group sends a representative
to a council which makes decisions for the union. The representatives cast their vote (`aye' or `nay') according to
the majority (or to what they believe is the majority) in their respective group. Since the groups
may differ in size, it is natural to assign different voting
weights to the representatives, reflecting the size of the respective group. When a parliament such as a the House of Representatives in the U.S. is elected, usually the country is subdivided into a number of districts of roughly equal population, each of which votes on a representative for a single seat. This procedure is feasible within a country but may not be possible in other situations. Even in the U.S., no effort has been made to divide the states and reassemble them into districts of roughly equal size so that each of them could have the same number of senators without giving rise to questions whether that is the `right' way to determine the number of senators. It is even less likely that sovereign countries -- such as the members of the United Nations or the European Union -- would be willing to submit to being divided into districts of equal size. Thus, it is not possible in practice to circumvent the question of how to assign voting weights to groups of different sizes.
To determine these weights is the problem of `optimal' weights. How should the weights be assigned? One objective
studied in the literature is to minimise the democracy deficit, i.e.\!
the deviation of the council vote from a hypothetical
referendum across the entire population. The democracy deficit was first studied for binary voting (the same setting which is considered in the present article) by Felsenthal and Machover \cite{FM1999}. Later on, it was also analysed in other settings by several authors (see e.g.\! \cite{FelsenthalM,HOEC,SlomZy,KiLa,MN2012,Toth}). Other notions of optimal weights are based on welfare considerations or the criterion of equalising the influence of all voters belonging to the overall population. In the latter vein, we find the seminal article by Penrose \cite{Pen1946}, where the square root law was first established as the assignation rule for voting weights that equalises the probability of each voter's being decisive in a two-tier voting system under the assumption of stochastically independent voting. Other contributions to the study of optimal voting weights under welfare and influence frameworks can be found in \cite{MN2012,BB2007,KMTL2013,KMN2017}. Correlated voting across groups was also analysed by Kaniovski and Zaigraev in \cite{KanZai2009}.
Suppose the overall population is of size $N$, whereas the group size is
$N_{\lambda}$, where the subindex $\lambda$ stands for
the group $\lambda\in\left\{ 1,\ldots,M\right\} $ . Let the two voting alternatives
be encoded as $\pm1$, $+1$ for `aye' and $-1$
for `nay'. The vote of voter $i\in\left\{ 1,\ldots,N_{\lambda}\right\} $ in group $\lambda $ will be denoted by $X_{\lambda i}$.
\begin{defn}
For each group $\lambda$, we define the voting margin $S_{\lambda}:=\sum_{i=1}^{N_{\lambda}}X_{\lambda i}$.
The overall voting margin is $S:=\sum_{\lambda=1}^{M}S_{\lambda}$.
\end{defn}
Each group casts a vote in the council:
\begin{defn}
The council vote of group $\lambda$ is given by
\[
\chi_{\lambda}:=\begin{cases}\,
\phantom{-}1, & \text{if }S_{\lambda}>0,\\
\,-1, & \text{otherwise.}
\end{cases}
\]
\end{defn}
The (representative of) group $\lambda$ votes `aye' if there is a majority in group $\lambda $ on the issue in question. Each group $\lambda$ is assigned a weight $w_\lambda$. The weighted sum $\sum_{\lambda=1}^{M}w_{\lambda}\chi_{\lambda}$
is the council vote. The council vote is in favour of a proposal if
$\sum_{\lambda=1}^{M}w_{\lambda}\chi_{\lambda}>0$. Weights $w_{1},\ldots, w_{M} $ together with a relative quota $q\in(0,1)$ constitute
a weighted voting system for the council, in which a coalition $C\subset \{ 1,2,\ldots,M \}$ is winning if
\begin{align*}
\sum_{i\in C}\,w_{i}~>~q\;\sum_{i=1}^{M}\,w_{i}.
\end{align*}
We will exclusively consider the majority rule with $q=1/2$ in this article.
It is reasonable to choose the voting weights $w_{\lambda}$ in the council in such a way, that
the raw democracy deficit
\begin{align*}
\left|\,S-\sum_{\lambda=1}^{M}\,w_{\lambda}\,\chi_{\lambda}\,\right|
\end{align*}
is as small as possible. For a given set of weights, each configuration of all $N$ votes induces a certain raw democracy deficit. It is immediately clear that in general there is no choice of weights which makes
this variable small uniformly over all possible distributions of Yes-No-votes across the overall population. All we can hope for is to
make it small `on average', more precisely we try to minimise the expected quadratic deviation
of $\sum_{\lambda=1}^M w_{\lambda}\chi_{\lambda}$ from $S$.
To follow this approach, we have to clarify what we mean by
`expected' deviation, i.e.\! there has to be some notion of randomness underlying the voting procedure.
While the votes cast are assumed to be deterministic and rational, obeying the
voters' preferences which we do not model explicitly, the proposal
put before them is assumed to be unpredictable, i.e.\! random. Since each yes/no
question can be posed in two opposite ways, one to which a given voter
would respond `aye' and one to which they would respond `nay', it is reasonable to assume that each voter votes `aye' with
the same probability they vote `nay'.
This leads us to the following definition:
\begin{defn}
A \emph{voting measure} is a probability measure $\IP $ on the space
of voting configurations $\left\{ -1,1\right\} ^{N}= \prod_{\lambda=1}^{M}\, \{ -1,1 \}^{N_{\lambda}}$ with the symmetry
property
\begin{align}
\mathbb{P}\left(X_{11}=x_{11},\ldots,X_{MN_{M}}=x_{MN_{M}}\right)~
=~\mathbb{P}\left(X_{11}=-x_{11},\ldots,X_{MN_{M}}=-x_{MN_{M}}\right)\label{eq:symmetry}
\end{align}
for all voting configurations $\left(x_{11},\ldots,x_{MN_{M}}\right)\in\left\{ -1,1\right\} ^{N}$.
By $\mathbb{E}$ we will denote the expectation with respect to $\mathbb{P}$.
\end{defn}
The simplest and widely used voting measure is the $N$-fold product of the measures
\begin{align*}
P_{0}(1)~=~P_{0}(-1)~=\frac{1}{2}
\end{align*}
which models independence between all the individual votes $X_{\lambda i}$. In this
special case, known as the \emph{Impartial Culture} (see e.g. \cite{GK1968}, \cite{GehrleinL}, or \cite{KurzMN}), we have
\begin{align*}
\mathbb{P}\left(X_{11}=x_{11},\ldots,X_{MN_{M}}=x_{MN_{M}}\right)~
=~\prod_{\lambda=1}^{M}\prod_{i=1}^{N_{\lambda}}\,P_{0}(X_{\lambda i}=x_{\lambda i})~=~\frac{1}{2^{N}}.
\end{align*}
This article treats the class of voting measures called the \emph{collective bias model} (or common belief model, CBM) which extends
the Impartial Culture considerably by allowing correlations both between voters in the same group
as well as correlations across group borders. We introduce and discuss the CBM in Section \ref{sec:CBM}.
Once a voting measure is given, the quantities $X_{\lambda i}$, $S_{\lambda}$, $\chi_{\lambda}$, and the raw democracy deficit are random variables defined on the same probability space $\{-1,1\}^{N}$.
Now we can define the concept of democracy deficit which is a measure of how well the council votes follow the public opinion:
\begin{defn} \label{def:democracy-deficit}The \emph{democracy deficit} given a voting
measure $\mathbb{P}$ and a set of weights $w_{1},\ldots,w_{M}$ is
defined by
\begin{align*}
\Delta_{1}~=\Delta_{1}(w_{1},\ldots,w_{M})~:=~\mathbb{E}\left[\left(S
-\sum_{\lambda=1}^{M}w_{\lambda}\chi_{\lambda}\right)^{2}\right]\,.
\end{align*}
We call $(w_{1},\ldots, w_{M}) $ \emph{optimal weights} if they minimise the democracy deficit, i.e.
\begin{align*}
\Delta_{1}(w_{1},\ldots,w_{M})~=~\min_{(v_{1},\ldots,v_{M})\in\mathbb{R}^{M}}\;\Delta_{1}(v_{1},\ldots,v_{M}).
\end{align*}
\end{defn}
Note that the democracy deficit depends on the voting measure. It is also worth pointing out that the democracy deficit is a differentiable function of the council weights. This facilitates the analysis required to find the optimal weights.
Instead of minimising the democracy deficit, we could ask the question of how to minimise the probability that the binary council decision differs from the decision made by a referendum. This would be a less strict criterion in the sense that for a favourable public opinion of 51\%, a 51\% percent vote in the council and a 100\% vote would both be considered equally good. However, one could argue that a 100\% vote in the council would not be a good representation of public opinion. In fact, the 49\% minority might feel they are not represented in the council at all, giving rise to populist anti-elite sentiment among the minority. We argue that adjusting the voting outcomes in the council in such a way that they follow the popular opinion as closely as possible is a worthwhile goal.
If we multiply each weight by the same positive constant and keep the relative quota $q$ fixed, we obtain an equivalent voting system. If the weights $w_{\lambda}$ minimise the democracy deficit $\Delta_{1} $, then
the (equivalent)
weights $ \frac{ w_{\lambda}}{\sigma}$ minimise the `renormalised' democracy deficit
$\Delta_{\sigma} $ defined by
\begin{align*}
\Delta_{\sigma}~=~\Delta_{\sigma}(v_{1},\ldots,v_{M})~:=~\mathbb{E}\left[\left(\frac{S}{\sigma}
-\sum_{\lambda=1}^{M}v_{\lambda}\chi_{\lambda}\right)^{2}\right]\,.
\end{align*}
It is, therefore, irrelevant whether we minimise $\Delta_{1}$ or $\Delta_{\sigma} $ as long as $\sigma>0 $.
In this article, we will compute optimal weights as $N$ tends to infinity. As a rule, in this limit the minimising weights
for $\Delta_{1}$ will also tend to infinity, it is therefore useful to minimise $\Delta_{\sigma}$ with an $N$-dependent $\sigma $ to keep the weights bounded. A particularly convenient choice is to
normalize the weights $w_{\lambda}$ in such a way that $\sum_\lambda w_{\lambda}=1$.
The rest of this paper is organised as follows: as a first step, in Section \ref{warm-up}, we recall the CBMs with independent groups studied in the past and give an example of a CBM with correlated voting across group boundaries. Then, we formally define the CBM and give several more examples in Section \ref{sec:CBM}. In Section \ref{sec:demdef}, we discuss the problem of determining the optimal weights in order to minimise the democracy deficit. Section \ref{sec:asymp} contains the results concerning the large population behaviour of CBMs. In Sections \ref{sec:optw} and \ref{sec:optwtc}, we calculate the optimal weights in the large population limit. Then, Section \ref{sec:specmod} discusses the optimal weights for some specific models introduced earlier, such as additive and multiplicative models. Sections \ref{sec:optw} to \ref{sec:specmod} contain the main results of this article concerning the optimal weights for a large set of CBMs. The second part of our analysis of optimal weights concerns their non-negativity. Under independence of the groups, the optimal weights are always strictly positive. Thus, this is a new aspect owed entirely to the relaxation of the independence assumption and not previously analysed in the literature. Section \ref{sec:nonneg} deals with the problem of negative optimal weights and conditions that rule them out. Section \ref{sec:extensions} presents an extension of the CBM with non-identical group bias distributions. Finally, Section \ref{sec:conclusion} concludes the paper.
\section{A Warm-Up}\label{warm-up}
Before defining the CBM in its full generality, we first recall the
CBMs treated in the past, where the biases in different groups are
independent of each other, and hence the voters belonging to different
groups act independently. Then, we give an example of a CBM with correlated
groups. It is our hope that the informal description of these special
cases before defining and analysing the CBM in its full generality
will make the model and the article more accessible to a wider range
of readers.
In \cite{HOEC}, one of us introduced the CBM with
groups still being independent. To distinguish it from the generalisation we are going to introduce
below, we refer to the CBM with independent groups as the \emph{simple CBM} for the rest of this paper.
Let $T_{1},\ldots,T_{M}$ be a collection of independent and identically
distributed random variables with support in $\left[-1,1\right]$.
We will refer to a realisation of each of these variables as $t_{\lambda}$.
Conditionally on $t_{\lambda}$, the voters in group $\lambda$ vote
independently of each other and the probability of a `yes' vote
is $\frac{1+t_{\lambda}}{2}$. Thus, a positive bias makes a $+1$
vote more likely from each of the voters. Since within each group
all voters are subject to the same bias, there is, in fact, positive
correlation between votes. There is, however, no correlation between
votes in different groups. The group bias reflects some real-world
influence on the voters' decisions, such as cultural norms, institutions
such as organised religions, or more recently social media and influencers, specific to each group.
Suppose each of these entities has some stance on each issue that
can be put to vote and these opinions aggregate to some public bias.
This bias can be quantified: a value of $-1$ or close to it reflects
a strong rejection; a value around 0 means neutrality or indifference;
a positive value close to $+1$ reflects strong support of an issue.
The bias affects the voting outcomes in such a way that, provided
the population is large enough, a strong negative bias will result
in a large negative vote $S_{\lambda}$. Similarly, a strong positive
bias induces a large positive vote, and the absence of a substantial
bias causes a small absolute voting margin $\left|S_{\lambda}\right|$,
i.e.\! a voting outcome close to a tie, because the individual votes are nearly
independent.
In order to obtain the distribution of votes in the overall population,
we have to average out the votes given all possible values of the
group bias variables. Due to the independence of the group bias variables
$T_{1},\ldots,T_{M}$, we can factor the probabilities and obtain
for each voting configuration $\left(x_{11},\ldots,x_{MN_{M}}\right)\in\left\{ -1,1\right\} ^{N}$
\[
\IP\left(X_{11}=x_{11},\ldots,X_{MN_{M}}=x_{MN_{M}}\right)=\prod_{\lambda=1}^{M}\IP\left(X_{\lambda1}=x_{\lambda1},\ldots,X_{\lambda N_{\lambda}}=x_{\lambda N_{\lambda}}\right)
\]
because the groups are independent. In accordance with the verbal
description given in the last paragraph, the probabilities for each
group's voting configuration can be expressed as follows:
\[
\IP\left(X_{\lambda1}=x_{\lambda1},\ldots,X_{\lambda N_{\lambda}}=x_{\lambda N_{\lambda}}\right)=E\left(\prod_{i=1}^{N_{\lambda}}P_{T_{\lambda}}\left(X_{\lambda i}=x_{\lambda i}\right)\right),
\]
where $P_{T_{\lambda}}$ is the probability measure on $\left\{ -1,1\right\} $
with $P_{T_{\lambda}}\left(1\right)=\frac{1+T_{\lambda}}{2}$ and
the expectation $E$ is taken with respect to the distribution of the group
bias variable $T_{\lambda}$. To recapitulate, we obtain the probability
of a voting configuration by taking the (random) probabilities $P_{T_{\lambda}}\left(X_{\lambda i}=x_{\lambda i}\right)$
for each individual vote and multiplying them together for all voters
belonging to group $\lambda$. This multiplicative form of the probabilities
is due to the conditional independence of the individual votes in
the group for a given realisation of the group bias variable $T_{\lambda}=t_{\lambda}$.
Next, we take the expectation of this product of probabilities over
all possible realisations of $T_{\lambda}$. This gives us the probability
of the voting configuration $\left(x_{\lambda1},\ldots,x_{\lambda N_{\lambda}}\right)$
in group $\lambda$. Finally, multiplying the probabilities of each
group's voting configuration yields the probability of the overall
voting configuration $\left(x_{11},\ldots,x_{MN_{M}}\right)$ under
the simple CBM.
For the simple CBM, we can calculate the optimal weights in the council
(see \cite{HOEC} for this result). These turn out to be proportional
to the expected absolute value of the group voting margins, $w_{\lambda}\propto\IE\left(\left|S_{\lambda}\right|\right)$.
It is known that as each group's population diverges to infinity,
we have $\IE\left(\left|S_{\lambda}\right|\right)/N_{\lambda}\rightarrow E\left|T_{\lambda}\right|$.
The latter expectation is a characteristic of the underlying distribution
of the bias variables $T_{\lambda}$. This implies (under the mild
assumption $E\left|T_{\lambda}\right|>0$) that in the large population
limit the optimal weights in the council are proportional to each
group's population, i.e.\! $w_{\lambda}=CN_{\lambda}$, with the
same positive multiplicative constant $C$ for each group.
Thus, the simple CBM yields the recommendation of assigning each group
a number of votes in the council which is proportional to its population.
This stands in contrast to Penrose's square root law, which prescribes
weights proportional to the square root of the population instead
of the population itself. Evidently, proportionality favours larger
groups at the expense of smaller ones and vice-versa for the square
root law.
Next, we consider a model with bias variables $T_{1},\ldots,T_{M}$
which are correlated, thus inducing correlated voting across groups
boundaries. Let $Z$ be a uniformly distributed random variable on
the interval $\left[-1/2,1/2\right]$. We will call $Z$ the global
bias variable. Let $Y_{1},\ldots,Y_{M}$ be i.i.d.\! copies of $Z$.
We define the group bias variables $T_{1},\ldots,T_{M}$ by setting
$T_{\lambda}:=Y_{\lambda}+Z$ for each group $\lambda$. This is a
special case of what we will call an `additive model' in later sections.
In an additive model, the global bias is modified by a group-specific bias which may reinforce or counteract the global bias. By assuming all these variables are identically distributed, we assign equal influence to the global bias and group-specific attitudes. Of course, it is also possible to assume a stronger global or local bias, a topic we will explore in Section \ref{sec:additive}.
Even though $Z,Y_{1},\ldots,Y_{M}$ are independent, due to the addition
of the same variable $Z$ in the definition of each $T_{\lambda}$,
the group bias variables $T_{\lambda}$ are \emph{not} independent.
Given a realisation $t_{\lambda}=y_{\lambda}+z$, the individual votes
in group $\lambda$ each turn out positive with probability $\frac{1+t_{\lambda}}{2}$.
Contrary to the simple CBM, we cannot factor the probabilities of
the overall voting configuration into the probabilities of the group
voting configurations. Instead, the probabilities can be expressed
as
\begin{align*}
\IP\left(X_{11}=x_{11},\ldots,X_{MN_{M}}=x_{MN_{M}}\right) & =E\left(\prod_{\lambda=1}^{M}\prod_{i=1}^{N_{\lambda}}P_{T_{\lambda}}\left(X_{\lambda i}=x_{\lambda i}\right)\right),
\end{align*}
where the expectation $E$ is taken with respect to the distributions of
$Z,Y_{1},\ldots,Y_{M}$. As the terms $P_{T_{\lambda}}\left(X_{\lambda i}=x_{\lambda i}\right)$
each depend on two different random variables, $Y_{\lambda}$ and
$Z$, there is no way to factor the expectation above.
In order to minimise the democracy deficit, we have to solve the linear
equation system given by \eqref{eq:LES}. We omit the calculations and refer the
reader to Section \ref{sec:additive}, where we will analyse a more general additive
CBM with uniformly distributed bias variables. By Theorem \ref{opt_weights_uniform}, the
optimal weight for each group $\lambda$ is asymptotically given by
\[
w_{\lambda}=D+C\frac{N_{\lambda}}{N},
\]
where we have simplified and normalised the weights. The positive
constants $C$ and $D$ are common for all groups. Contrary to the
simple CBM with independent groups, in this correlated example, we
have a summand which is proportional to the size of the group, but
we also have a constant summand $D$ which is the same for each group
and hence independent of the group's size. This is qualitatively the
same formula as the one employed for the composition of the U.S.\!
Electoral College, where $D$ equals 2, the number of senators for each state,
and $C=435$ is the number of representatives, of which each state
receives a number roughly proportional to its population. The new
feature of the general CBM compared to the simple CBM concerning the
problem of optimal council weights is the presence of the constant
term $D$. This functional form for the optimal weights applies not
just to the special case discussed in this section but in general
to CBMs with correlated groups. See Theorem \ref{thm:asweights} for the general result.
As far as the authors of this article know, this is the first theoretical
justification of the formula that determines the number of electors
for each state in the U.S.\! Electoral College. We want to emphasise
that it is the correlation between votes in different groups that
induces the constant term $D$ in the formula for the optimal weights.
No matter what distribution we choose for the group bias variables
$T_{\lambda}$ in a simple CBM \textendash{} even allowing for $T_{\lambda}$
that follow different distributions \textendash{} the constant $D$
always equals 0. The formula $w_{\lambda}=D+C\frac{N_{\lambda}}{N}$
is more favourable to small groups than proportionality. However,
the square root law is even better for small groups in most cases,
the exception being when the difference in size between small and
large groups is minuscule.
\section{The Collective Bias Model}\label{sec:CBM}
We recall from the last section that in the simple CBM, the votes $X_{\lambda i}$ within a group $\lambda $ are correlated via a random
variable $T_{\lambda}$ with values in $[-1,1] $, the local `collective bias'. The random variables $T_{\lambda}$ model the influence
of a cultural tradition in the respective group or the leverage of a strong political party (or religious group, etc.) within the group $\lambda $. It is this central influence, which affects all voters within a given group equally, which induces positive correlation between votes within each group. Aside from this central influence, the voters make up their own minds. This is in contrast to models with interactions between voters such as those inspired by spin models from statistical mechanics, e.g.\! mean-field models (see \cite{KT_opt_weights_CW} for a discussion of a mean-field model and the determination of the optimal weights). In the simple CBM, there is no correlation between votes in different groups, only correlation within groups.
In what follows, we will define the simple CBM in the same measure-theoretic language we will also employ for the general CBM. Given the bias $T_{\lambda}=t_{\lambda}$, the per capita voting margin $S_\lambda/N_\lambda$ inside group $\lambda$ fluctuates around $t_{\lambda}$.
More precisely, suppose the bias variable $T_{\lambda} $ is distributed according to the probability
measure $\rho $ on $[-1,1] $. Then the simple CBM for the group $\lambda $ is given by
\begin{align}\label{eq:simpleCBM}
\mathbb{P}\Big(X_{\lambda 1}=x_{1}, X_{\lambda 2}=x_{2},\ldots,X_{\lambda N_{\lambda}}=x_{N_{\lambda}}\Big)~=~
\int P_{t}\big(x_1,\ldots, x_{N_{\lambda}}\big)\rho(\textup{d}t),
\end{align}
where
\begin{align*}
& P_{t}(x)~=~\left\{
\begin{array}{ll}
\eh(1+t), & \hbox{for $x=1$,} \\[2mm]
\eh(1-t), & \hbox{for $x=-1$,}
\end{array}
\right.\\
\text{and}\quad & P_{t}\big(x_1,\ldots, x_{N_{\lambda}}\big)~
=~P_{t}(x_{1})\,P_{t}(x_{2})\,\cdots\,P_{t}(x_{N_{\lambda}}).
\end{align*}
By $E_{t} $ we denote the expectation with respect to $P_{t} $. The definition of $P_{t}$ implies that $E_{t}(X)=t$.
We call $\rho $ the local bias measure of group $\lambda$.
We remark that, due to de Finetti's Theorem\footnote{De Finetti's Theorem states that an infinite sequence of exchangeable random variables can be represented as a mixture of i.i.d.\! random variables. The mixture is specified by a probability measure referred to as a de Finetti measure. De Finetti's Theorem has been considerably generalised. See e.g.\! \cite{DF1987}.}, the simple CBM is the most general voting measure that is `anonymous' in the sense that reordering the voters leaves the measure unchanged (see \cite{Klenke} or \cite{KiFinetti}).
The `Impartial Anonymous Culture', which underlies the Shapley-Shubik power index \cite{Straffin} (see also \cite{GehrleinL} or \cite{KurzMN}), is a particular case of \eqref{eq:simpleCBM} where $\rho $ is the
uniform distribution on $[-1,1] $. The Impartial Culture is another special case for which $\rho=\delta_{0}$, the Dirac measure\footnote{The Dirac measure (or point mass) at $x\in \IR^M$, $\delta_x$, is a probability measure which assigns any set $A\subset \IR^M$ the probability 1 if $x\in A$ and 0 otherwise.} at $t=0$.
In the simple CBM, the voting results in different groups are independent, so the corresponding voting measure on $\prod_{\lambda=1}^{M} \{ -1,1 \}^{N_{\lambda}} $ is given by the product of the probabilities \eqref{eq:simpleCBM}.
\begin{align*}
\mathbb{P}\left(\uX_{1}=\ux_{1},\ldots,\uX_{M}=\ux_{M}\right)~
=~\int P_{t_1}\big(\ux_{1}\big)\;\rho(\textup{d}t_{1})~\cdots
~\int P_{t_M}\big(\ux_{M}\big)\;\rho(\textup{d}t_{M}),
\end{align*}
where $\underline{X}_{\lambda}=(X_{\lambda 1},\ldots,X_{\lambda N_{\lambda}})$ and similarly for $\underline{x}_{\lambda}$.
In this paper, we study the \emph{generalised collective bias model} (CBMs with dependence across group boundaries were first analysed in \cite{Toth}). In this model, there
is an additional \emph{global} bias
variable $Z$ with values in $[-1,1] $ and with distribution $\mu $. The global bias $Z$ influences each of the groups in a similar way. This is implemented in the model by allowing the local bias measure $\rho $ to depend on the
value $Z=z$. More precisely, the (generalised) collective bias model is given by:
\begin{defn}\label{def:CBM}
Suppose $\mu $ is a probability measure on $[-1,1] $ and for every $z\in[-1,1]$ there is a
probability measure $\rho^{z}$ on $[-1,1] $. Then we define the probability measure $\IP_{\mu \rho}$
on $\{ -1,1 \}^{N}=\prod_{\lambda=1}^{M}\{ -1,1 \}^{N_{\lambda}}$ by
\begin{align}
\IP\big(\ux_{1},\ux_{2},\ldots,\ux_{M}\big)~
=~&\int \left( \int P_{t_{1}}(\ux_{1})\,\rho^{z}(\textup{d}t_{1})
\cdots\int P_{t_{M}}(\ux_{M})
\,\rho^{z}(\textup{d}t_{M}) \right)\;\mu(\textup{d}z),\label{eq:genCBM}
\end{align}
where $\ux_{\lambda}\in\{ -1,1 \}^{N_{\lambda}}$.
We call $\IP_{\mu\rho} $ the \emph{collective bias measure} with \emph{global bias measure} $\mu$ and \emph{local bias measure} $\rho = \rho^{z} $ or the CBM($\mu,\rho $) for short. If $\mu $ and $\rho $ are clear from the context, we simply write $\IP$ instead of $\IP_{\mu\rho}$.
\end{defn}
\begin{rem}\label{rem:kernel}
Technically speaking, $\rho^{z} $ is a stochastic kernel (see e.g. \cite{Klenke}), i.e.:
\begin{enumerate}
\item For every $z\in[-1,1]$, the quantity $\rho^{z}$ is a probability measure on $[-1,1] $.
\item For every Borel set $A\subset[-1,1]$, the function $z\mapsto \rho^{z}A$ is measurable.
\end{enumerate}
\end{rem}
We could allow the kernels $\rho^{z}$ to depend on
the group $\lambda $, and in Section \ref{sec:extensions} we will come back to this generalisation,
but for the moment we take the same local bias measure for all groups.
To ensure that $\IP_{\mu\rho}$ is a \emph{voting measure}, i.e.$\!\,$ to satisfy \eqref{eq:symmetry},
we assume the following sufficient condition in what follows:
\begin{ass}\label{ass:sym}
\begin{enumerate}
\item $\mu$ is symmetric, i.e. $\mu A=\mu(-A)$,
\item for all $z\in\left[-1,1\right]$, the distributions
$\rho^{z}$ satisfy $\rho^{z}A=\rho^{-z}\left(-A\right)$ for all measurable sets $A\subset\left[-1,1\right]$.
\end{enumerate}
\end{ass}
The general framework of a CBM is given by a set
of bias random variables that represent some cultural or political
influence that acts on all voters. There is a \emph{global bias variable}
$Z$ with distribution $\mu$ which induces correlation between voters
of different groups. Furthermore, there is a \emph{local bias variable} $T_{\lambda}$
for each group. Its conditional distribution given the realisation
$Z=z$ is $\rho^{z}$. The group bias variable $T_{\lambda}$
induces correlation between the voters belonging to that group. The result is correlated voting across group boundaries, as a rule with stronger
correlation within each group to account for shared culture and preferences.
Conditionally on the realisations of $Z=z$ according to $\mu$
and $T_{\lambda}=t_{\lambda}$ according to $\rho^{z}$, all
voters in group $\lambda$ cast their vote independently, with a probability
of voting `aye' equal to $\frac{1+t_{\lambda}}{2}$. Hence,
a value $t_{\lambda}=1$ implies that all voters belonging to
group $\lambda$ vote `aye' almost surely. Similarly, $t_{\lambda}=-1$
implies all vote `nay' almost surely. $t_{\lambda}=0$ means
there is no bias, and all voters in the group vote independently with probability $\tfrac{1}{2}$ for `aye' (and the same probability for `nay').
\begin{example}\label{ex:1}
{ }\quad We discuss various examples (or classes of examples) of CBMs.
\begin{enumerate}
\item If the measures $\rho^{z} $ are independent of $z$, then the (generalised)
CBM reduces to the simple CBM. The Impartial Anonymous Culture is a particular case of this class of examples.
\item If $\rho^{z}=\delta_{0} $, then all random variables
$X_{\lambda i}$ are independent reflecting Impartial Culture.
\item If $\rho^{z}=\delta_{z}$, then we have a simple CBM for the \emph{union}, i.e.\! for \emph{all} $X_{\lambda i}$.
\item \label{ex:add} In the class of \emph{additive models}, the `total bias' prevailing within each group $T_\lambda$ is the \emph{sum} of the global bias variable $Z$ and a local or group bias modifier variable $Y_{\lambda}$, i.e. $T_\lambda=Z+Y_\lambda$.
Assume the bias modifiers $Y_\lambda$ are independent and identically distributed according to a fixed symmetric probability measure $\rho$.
Then, for each realisation $Z=z$, the local measure $\rho^{z}$ is given by
\begin{align*}
\rho^{z}[a,b]~=~\rho[a-z,b-z].
\end{align*}
So for this model class we have
\begin{align}
\IP\big(\ux_{1},\ux_{2},\ldots,\ux_{M}\big)~=~\int \left( \int P_{z+y_{1}}(\ux_{1})\,\rho(\textup{d}y_{1})
\cdots\int P_{z+y_{M}}(\ux_{M})
\,\rho(\textup{d}y_{M}) \right)\;\mu(\textup{d}z). \label{eq:addCBM}
\end{align}
To ensure that $t_{\lambda}=z+y_\lambda\in[-1,1]$ we assume that $\supp\,\mu\,+\,\supp\,\rho\;\subset[-1,1]$, where $\supp\, \mu$ stands for the support of the measure $\mu$. This kind of additive CBM was first introduced and analysed in Section 4.2 of \cite{Toth}. Additive models are discussed in more detail in Section \ref{sec:additive}.
\item For a particular example of the additive model which we are going to discuss in some detail,
we choose $\mu $ and $\rho $ as the uniform distribution on $[-g,g] $ and on $[-\ell,\ell] $, respectively, with $0<g,\ell$ and $g+\ell\leq 1$.
In this case, the additive CBM-measure is given by
\begin{align*}
\frac{1}{2g}\int_{-g}^{+g}\;\left(\,\frac{1}{2\ell}\int_{z-\ell}^{z+\ell} P_{t_{1}}(\underline{x}_{1})\,\textup{d}t_{1}\cdots
\frac{1}{2\ell}\int_{z-\ell}^{z+\ell} P_{t_{M}}(\underline{x}_{M})\,\textup{d}t_{M}\,\right)\,\textup{d}z.
\end{align*}
This example may be considered a `hierarchical' version of Impartial Anonymous Culture.
\item\label{ex:mult} In the class of \emph{multiplicative models}, the total bias is the \emph{product} of the global bias variable $Z$ and the group bias modifier variable $Y_{\lambda}$, i.e. $T_\lambda=ZY_\lambda$. We assume the $Y_\lambda$ are independent and identically distributed according to a fixed probability measure $\rho$. Then the local measure is $\rho^0=\delta_0$ if $Z=0$, and for $Z=z\neq0$,
\begin{align*}
\rho^{z}[a,b]~=~\rho\left[\frac{a}{z}\wedge\frac{b}{z},\frac{a}{z}\vee\frac{b}{z}\right].
\end{align*}
Above, we used the notation $a\vee b:=\max\{a,b\}$ and $a\wedge b:=\min\{a,b\}$ for all real numbers $a$ and $b$.
This kind of multiplicative CBM was first introduced and analysed in Section 4.1 of \cite{Toth}. We discuss the multiplicative model in Section \ref{sec:mult}.
\item In the CBM($\mu,\rho$), the measure $\rho $ must have support in $[-1,1] $. Above, we assumed
without loss of generality the same for the measure $\mu $. In the following example, it is more convenient to have more freedom in the choice of $\mu $.
Suppose that $\rho^{z} $ is the beta distribution $B(z,z,-1,1) $, i.e. $\rho^z$ has the density
\begin{align*}
f_{z}(x):=\frac{\Gamma(2z)}{\Gamma(z)^{2}\; 2^{2z-1}}\;(1+x)^{z-1}\,(1-x)^{z-1}
\end{align*}
for $x\in[-1,1]$, where $\Gamma $ is the Gamma function. For $\mu$, we can take any probability distribution on $(0,\infty) $. Note that the symmetry condition \eqref{eq:symmetry} is satisfied.
For large $z$, the measures $\rho^{z} $ are more and more concentrated around $0$. For $z=1$, the measure
$\rho^{z} $ is the uniform distribution, and for small $z>0,$\;$\rho^{z} $ is more and more concentrated near the extreme positions $+1$ and $-1$. The measures $\rho^{z} $ are intimately connected with P\'olya urn models which are discussed, for example, in \cite{Berg} and \cite{KurzMN}.
In a sense, the parameter $z$ reflects the `polarisation' inside the society.
\item\label{ex:det} We end the presentation of examples with a rather pathological class, in fact one we are going to exclude below.
Suppose that for $\mu$-almost all $z$ either $\rho^{z}=\delta_{1}$ or $\rho^{z}=\delta_{-1}$. Then the popular vote is always unanimous. So, in a sense, there is little randomness in this example.
\end{enumerate}
\end{example}
\section{Democracy Deficit and Optimal Weights}\label{sec:demdef}
We want to choose the weights so that the democracy deficit is minimal.
By taking partial derivatives of $\Delta_{\sigma}$ with respect to each $w_\lambda$, we obtain a system of linear equations
that characterizes the optimal weights. Indeed, for $\lambda=1,\ldots, M$,
\begin{equation}
\sum_{\nu=1}^{M}\;\mathbb{E}\left(\chi_{\lambda}\chi_{\nu}\right)w_{\nu}~
=~\frac{1}{\sigma}\mathbb{E}\left(\chi_{\lambda}S\right)\,.\label{eq:LES}
\end{equation}
Defining the matrix $A$, the weight vector $w$ and the vector $b$ on the right hand side
of \eqref{eq:LES} by
\begin{align}
A~&:=~\left(A_{\lambda\nu}\right)_{\lambda,\nu=1,\ldots,M}~
:=~\mathbb{E}\left(\chi_{\lambda}\chi_{\nu}\right)\label{eq:A}\\
w~&:=~\left(w_{\lambda}\right){}_{\lambda=1,\ldots,M}\notag\\
b~&:=~\left(b_{\lambda}\right)_{\lambda=1,\ldots,M}~
:=~\frac{1}{\sigma}\,\mathbb{E}\left(\chi_{\lambda}S\right)\,.\label{eq:b}
\end{align}
we may write \eqref{eq:LES} in matrix form as
\begin{align}\label{eq:LESM}
A\;w~=~b\,.
\end{align}
Since the quantity $b$ depends on $\sigma $ (by a factor of $\frac{1}{\sigma} $),
the optimal weights do as well.
\medskip
A solution $w$ of \eqref{eq:LESM} is indeed a minimum if the matrix $A$, the Hessian of $\Delta $, is (strictly) positive definite.
In this case, the matrix $A$ is invertible and consequently there is a unique tuple of optimal weights, namely the unique solution of \eqref{eq:LESM}.
If the groups vote independently of each other, the matrix $A$ is diagonal. This happens for
CBM($\mu,\rho $)-measures for which $\rho $ is independent of $z$. These cases are treated in
\cite{HOEC}.
It turns out that in the general case the matrix $A$ is indeed invertible under rather mild conditions.
\begin{defn}\label{def:suffrand}
We say that a voting measure $\IP $ on $\prod_{\lambda=1}^{M}\,\{ -1,1 \}^{N_{\lambda}}$ is \emph{sufficiently random} if
\begin{align}\label{eq:suffrand}
\IP\, (\chi_{1}=c_{1},\ldots,\chi_{M}=c_{M})~>~0\qquad \text{for all } c_{1},\ldots,c_{M}\in\{ -1,1 \}.
\end{align}
\end{defn}
Note that \eqref{eq:suffrand} is not very restrictive. For example, if the support $\rm{supp}\,\IP$ of the measure $\IP$ is the whole space $\{ -1,1 \}^{N}$, then $\IP$ satisfies \eqref{eq:suffrand}.
Moreover, for CBMs, we have:
\begin{prop}\label{prop:suffrand}
Suppose that $\IP $ is a CBM($\mu,\rho$)-measure. Then $\IP$ is sufficiently random if and only if
\begin{align}\label{eq:noDelta}
\mu\,\Big\{z\;\Big\vert \; \rho^{z}\{ -1 \}=1 \quad\text{\rm or}\quad \rho^{z}\{ 1 \}=1\Big\}~<~1.
\end{align}
\end{prop}
\begin{rem}
If $\mu\{z\mid\rho^{z}\{ -1 \}=1 \;\text{\rm or}\; \rho^{z}\{ 1 \}=1\}=1 $, then the voting result in each
group is unanimous, so weights proportional to $N_{\lambda}$ are optimal weights (not necessarily
unique).
\end{rem}
\begin{prop}\label{prop:posdef} Let $\IP $ be a voting measure and let $A $ be defined by \eqref{eq:A}.
\begin{enumerate}
\item The matrix $A$ is positive semi-definite.
\item $A$ is positive definite if $\IP$ is sufficiently random.
\end{enumerate}
\end{prop}
\begin{proof}
For vectors $x,y\in\IR^M$, we will write $(x,y)$ for the Euclidean inner product. For any vector $x=(x_{1},\ldots,x_{M}) $, we have
\begin{align}\label{eq:psd}
\big(x,A\,x\big)~=~\IE\left(\left(\sum_{\lambda=1}^{M} x_{\lambda}\,\chi_{\lambda}\right)^{2} \right)~\geq~0\,.
\end{align}
So $A $ is positive semi-definite.
Suppose now that $(x,Ax)=0$. Then
\begin{align*}
\IE\left(\left(\sum_{\lambda=1}^{M} x_{\lambda}\,\chi_{\lambda}\right)^{2} \right)~=~0\,.
\end{align*}
This implies that
\begin{align}\label{eq:eigen}
\sum_{\lambda=1}^{M} x_{\lambda}\,\chi_{\lambda}~=~0 \qquad{\text{almost surely.}}
\end{align}
For a sufficiently random model, this is only possible if $x=0$.
\end{proof}
\begin{thm}\label{thm:weights}
If the voting measure $\IP $ is sufficiently random, the optimal weights minimising the democracy deficit $\Delta_{\sigma} $ are unique and given by
\begin{align}\label{eq:w}
w~=~A^{-1}\,b.
\end{align}
\end{thm}
\begin{defn}\label{def:wb}
If $w$ satisfies \eqref{eq:w}, we set
\begin{align*}
\overline{w}_{\nu}~:=~\frac{w_{\nu}}{\sum_{\lambda=1}^{M} w_{\lambda}},
\end{align*}
and call $\overline{w}_{\nu} $ the \emph{normalised optimal weights}.
\end{defn}
While the weights $w$ depend on $\sigma $ through $b=b_{\sigma}$, the normalised weights $\overline{w}$
are independent of $\sigma $. The $\overline{w}_{\nu} $ sum up to $1 $.
For the rest of this paper, we shall always assume that our models are sufficiently random.
Given Theorem \ref{thm:weights}, one is tempted to believe that the problem of optimal weights
is solved. Unfortunately, this is not the case, because it is practically impossible to compute the
ingredients like $\IE(\chi_{\lambda}\chi_{\nu})$ and $\IE(S\chi_{\lambda}) $ for finite (but fairly large) $N$.
A way out is to compute these quantities approximately for $N\to\infty$, and this is what we are doing throughout the rest of this paper.
\section{Asymptotics for the Collective Bias Model}\label{sec:asymp}
For given $\mu $ and $\rho^{z} $ and for $\uN=(N_{1},\ldots,N_{M})$, we denote by
$\IP_{\uN} $ the CBM($\mu,\rho$)-measure on $\prod_{\lambda=1}^{M}\{ -1,1 \}^{N_{\lambda}}$.
In the following, we try to compute optimal weights for large $N=\sum_\lambda N_{\lambda}$. More precisely, we consider
\eqref{eq:LESM} for $N\to\infty$. This limit is always taken in the sense that
\begin{align}\label{eq:Nlambda}
\lim_{N\to\infty} \frac{N_{\lambda}}{N}~=~\alpha_{\lambda}~>~0
\end{align}
for each $\lambda$, and we assume that each group's population diverges to infinity as $N$ goes to infinity. Observe that $\sum_\lambda \alpha_{\lambda}=1 $. The constants $\alpha_\lambda$ represent the population of each group as a fraction of the overall population (at least asymptotically). Whenever the $N_{\lambda} $ are clear from the context we write $\IP_{N}, \IE_{N}$ instead of $\IP_{\uN}, \IE_{\uN}$, etc. We also set
\begin{align}
\big(A_{N}\big)_{\lambda\nu}~&:=~\IE_{N}\big(\chi_{\lambda}\chi_{\nu}\big),\quad \big(b_{N}\big)_{\lambda}~:=~\IE_{N}\left(\frac{S}{N}\chi_{\lambda}\right), \notag\\\quad\text{and}\quad s_{N}~&:=~\IE_{N}\left(\left(\frac{S}{N}\right)^{2}\right). \label{eq:AN}
\end{align}
Then
\begin{align}
\Delta_{N}(w)~=~s_{N}\;-2\,(w,\, b_{N})\;+(w,A_{N}\,w). \label{eq:DeltaN}
\end{align}
In the above formulas, we set $\sigma:=N$.
From now on, we assume that $\IP$ is sufficiently random, i.e.\! that \eqref{eq:noDelta} holds.
Moreover, to avoid discussing different cases we also assume that $\rho $ is not trivial in the sense that
\begin{align}\label{eq:rhotriv}
\mu\{z\mid \rho^{z}=\delta_{0} \}~<~1\,.
\end{align}
If \eqref{eq:rhotriv} is violated, all voters act independently of each other. This is the `Impartial
Culture' and Penrose's square root law holds (see e.g.$\!\,$ \cite{FelsenthalM} or \cite{HOEC}).
The following result is the key observation which allows us to
evaluate important quantities asymptotically. This theorem explains the large population behaviour of a CBM.
\begin{thm}\label{thm:convthm}
Suppose that the functions $f_{\lambda}:[-1,1]\to\IR, \lambda=1,\ldots,M$, are continuous on $[-1,0)\cup(0,1] $, and assume that the limits $f_{\lambda}(0+)=\lim_{t\searrow 0}f_{\lambda}(\alpha_{\lambda}t)$ and
$f_{\lambda}(0-)=\lim_{t\nearrow 0}f_{\lambda}(\alpha_{\lambda}t)$ exist. Set
\begin{align*}
I_{z}\big(f_{\lambda}\big)~:=~\int_{[-1,0)\cup(0,1] } f_{\lambda}(\alpha_{\nu}\,t)\;\rho^{z}(\textup{d}t)\,+\,\frac{1}{2}\big(f_\lambda(0+)+f_\lambda(0-)\big)\; \rho^{z}\{ 0 \}.
\end{align*}
Then
\begin{align}\label{eq:convthm}
&\IE\,\left( f_{1}\left(\frac{1}{N}S_{1}\right)\,\cdot\,\ldots\,\cdot f_{M}\left(\frac{1}{N}S_{M}\right)\right)~\rightarrow~
\int\, I_{z}\big(f_{1}\big)\,\cdots\, I_{z}\big(f_{M}\big)
\; \mu(\textup{d}z).
\end{align}
\end{thm}
We could handle functions $f_{\lambda}$ with discontinuities (and left and right limits) in other points than $0$ as well, but we need the result only in the above form. The proof below, however, works for the more general case as well.
Theorem \ref{thm:convthm} says that the normalised voting margins $S_\lambda / N $ follow a distribution given by $\rho^z$ and $\mu$ in the large population limit. We can take transformations $f$ of these voting margins and their behaviour will be described by the distributions $\rho^z$ and $\mu$. Chief among these transformations will be the council vote $\chi_\lambda = \chi_\lambda\left(S_\lambda\right)$ cast by each group which presents a point of discontinuity at 0.
\begin{proof}
By the strong law of large numbers, we get
\begin{align*}
P_{t}\left(\lim_{n\to\infty} \frac{1}{N_{\lambda}}S_{\lambda}=t\right)~=~1.
\end{align*}
So, if $f$ is continuous on $[-1,1] $, it follows that
\begin{align*}
\int E_{t}\left(f\left(\frac{1}{N}S_{\lambda}\right) \right)\,\rho^{z}(\textup{d}t)~\to~\int f(\alpha_{\lambda} t)\,\rho^{z}(\textup{d}t)
\end{align*}
for all $z$. From this, \eqref{eq:convthm} follows for continuous $f_{\lambda}$.
To prove \eqref{eq:convthm} in the general case, we observe that
for $N\to\infty$
\begin{align*}
&P_{0}\left(\frac{1}{N_{\lambda}}S_{\lambda}>0\right)~\to~\frac{1}{2
\quad\text{and}\quad P_{0}\left(\frac{1}{N_{\lambda}}S_{\lambda}<0\right)~\to~\frac{1}{2}.
\end{align*}
\end{proof}
\begin{defn}\label{def:notation}
We introduce the following notation for further use:
\begin{align*}
& m_{1}(\rho)~=~\int t \,\rho^{z}(\textup{d}t), && m_{2}(\rho)~=~\int t^2 \,\rho^{z}(\textup{d}t),\\
& \om_{1}(\rho)~=~\int |t| \,\rho^{z}(\textup{d}t), && d(\rho)~=~\rho^{z} (0,1]-\rho^{z} [-1,0). \notag
\end{align*}
Note that the above quantities depend on $z $. These quantities are important characteristics of the measures $\mu$ and $\rho$. They measure the strength of the group bias for different values $z$ of the global bias. E.g.\! a positive $m_1(\rho)$ close to 1 implies that the group bias measure $\rho^z$ induces, on average, a strong bias in favour of the issue being considered. Whereas $\om_1(\rho)$ can be interpreted as a measure of intra-group cohesion, the product of $m_1$ and $d(\rho)$ is a measure of inter-group cohesion. These two measures will allow us to compare how strong the intra-group cohesion is versus the inter-group cohesion. These quantities will be used in the calculation of the optimal weights.
For any function $\varphi$ on $[-1,1] $, we introduce the shorthand notation
\begin{align*}
\langle \,\varphi\, \rangle~=~\int \varphi(z)\,\mu(\textup{d}z).
\end{align*}
\end{defn}
\begin{thm}\label{thm:limits}
Assume \eqref{eq:noDelta}, \eqref{eq:Nlambda} and \eqref{eq:rhotriv}. Then
\begin{align*}
(A_{N})_{\lambda \nu}~&\to~ a~:=~\langle d(\rho)^{2} \rangle,\;\lambda\neq\nu,\\
(b_{N})_{\lambda}~&\to ~b_{\lambda}~:=~ \langle\om_{1}(\rho)-m_{1}(\rho)d(\rho)\rangle\,\alpha_{\lambda}\;+\;\langle m_{1}(\rho)\,d(\rho) \rangle,\\
\textrm{s}_{N}~&\to~\textrm{s}~:=
~\sum_{\nu=1}^{M}\,\alpha_{\nu}^{2} \big(\langle m_{2}(\rho)\rangle-\langle {m_{1}(\rho)}^{2} \rangle\big) + \langle m_{1}(\rho)^{2} \rangle.
\end{align*}
\end{thm}
Theorem \ref{thm:limits} follows immediately from Theorem \ref{thm:convthm}.
Informally speaking, Theorem \ref{thm:limits} says that the minimisation problem \eqref{eq:DeltaN}
`converges' to the minimisation problem
\begin{align}\label{eq:DeltaInf}
\min\;\Delta_{\infty}(v_{1},\ldots,v_{M})~=~s\;-2\,(v,\, b)\;+(v,A\,v).
\end{align}
In the following, we try to explore the validity of this informal idea. The following theorem implies that for positive definite limiting coefficient matrices $A$ the optimal weights of the finite population problem converge to the optimal weights of the asymptotic problem.
\begin{thm}\label{cor:A}
The matrices $A_{N}$ converge (in operator norm) to the matrix
\begin{align}\label{eq:Ainf}
A_{\lambda\nu}~=~\left\{
\begin{array}{ll}
1, & \hbox{if $\lambda=\nu $,} \\
a, & \hbox{otherwise,}
\end{array}
\right.
\end{align}
with $a=\langle d(\rho)^{2} \rangle$.
Moreover, $A$ is positive semi-definite. $A$ is positive definite if $a<1$. In this case,
\begin{align}\label{eq:konvinv}
{A_{N}}^{-1}~\to~A^{-1}
\end{align}
and
\begin{align}\label{eq:Ainv}
{\left(A^{-1}\right)}_{\lambda\nu}~=~\frac{1}{D}\;\left\{
\begin{array}{ll}
1+(M-2)a, & \hbox{if $\lambda=\nu$,} \\
-a, & \hbox{otherwise,}
\end{array}
\right.
\end{align}
where $D=(1-a)\big((1+(M-1)a\big)$.
\end{thm}
\begin{proof}
We note that $0\leq a\leq 1$.
Since, for any $x\in\IR^{M}$,
\begin{align*}
\left(x,Ax\right)~=~(1-a)\,\sum_{\lambda=1}^{M}x_{\lambda}^{2}\;+\;a\,\left(\sum_{\lambda=1}^{M}x_{\lambda}\right)^{2},
\end{align*}
we see that $A$ is positive semi-definite in general and positive definite if $a<1$.
Let $I$ stand for the $M\times M$ identity matrix. To prove \eqref{eq:konvinv} we compute,
\begin{align}
\|{A_{N}}^{-1}-A^{-1}\|~&=~\Big\|\,A^{-1}\,\Big(\big(I+(A_{N}-A)A^{-1}\big)^{-1}-I\Big)\,\Big\|\notag\\
&\leq~\|A^{-1}\|\;\sum_{k=1}^{\infty}\,\|A_{N}-A\|^{k}\,\|A^{-1}\|^{k}\notag\\
&=~\frac{\|A^{-1}\|^2\,\|A_{N}-A\|}{1-\|A^{-1}\|\|A_{N}-A\|}\label{eq:exp}.
\end{align}
Since $\|A_{N}-A\|$ tends to 0, \eqref{eq:exp} goes to 0 as well.
The claim \eqref{eq:Ainv} follows by direct
calculation.
\end{proof}
\begin{defn}
We say that the collective bias model CBM$(\mu,\rho)$ is \emph{tightly correlated} if $a=\langle d(\rho)^{2} \rangle=1$.
\end{defn}
As we will see, tight correlation implies that all groups end up voting unanimously in the council. For now, we characterise tight correlation in terms of the probabilities assigned by $\rho$ for different values $z$ of the global bias. The key idea is that $\rho^z$ assigns probability 1 to either $(0,1]$ or $[0,1)$ for ($\mu$-almost) all $z$, and thus all group biases will be of the same sign, inducing the aforementioned unanimous council vote.
\begin{prop}\label{prop:tight}
The collective bias model $CBM(\mu,\rho)$ is tightly correlated if and only if
for $\mu$-almost all $z$ either $\rho^{z} (0,1]=1$ or $\rho^{z} [-1,0)=1$ holds.
\end{prop}
\begin{proof}
Since $0\leq d(\rho)^{2}\leq 1$ for all $z$, we have $\xi:=1-d(\rho)^{2}\geq 0$ and $\int \xi\,d\mu=0$ implies
$\xi=0$ $\mu$-almost surely. It follows that $|d(\rho)|=1$ for $\mu$-almost all $z $, so $\rho^{z} (0,1]=1$
or $\rho^{z} [-1,0)=1$.
\end{proof}
\section{Optimal Weights}\label{sec:optw}
In this section, we investigate the asymptotics of the optimal weights of CBMs
for large $N$. As above, we assume \eqref{eq:noDelta}, \eqref{eq:Nlambda}, and \eqref{eq:rhotriv} for the rest of this paper.
The tightly correlated case needs a different treatment, so we first assume that the model CBM$(\mu,\rho)$ is \emph{not} tightly correlated, i.e.$\!\,$ that $a=\langle d(\rho)^{2} \rangle<1$, in this section. Section \ref{sec:optwtc} discusses the tightly correlated case.
By Theorem \ref{thm:weights}, for fixed $N $, there are unique optimal weights $w_{N} $.
\begin{thm}\label{thm:asweights}
If the model $CBM(\mu,\rho) $ is not tightly correlated, then the optimal weights $w^{(N)}$, i.e.\!
the minima of $\Delta_{N} $, converge
for $N\to\infty$ to the minima of $\Delta_{\infty}$ (defined in \eqref{eq:DeltaInf}), and these weights $w_{\lambda}$ are given by
\begin{align}
w_{\lambda}~=~C_{1}\,\alpha_{\lambda}\;+\;C_{2}\label{eq:optw},
\end{align}
with coefficients depending on $\mu,\rho$, and $M$ but not on the $\alpha_{\lambda}$.
More precisely,
\begin{align}
C_{1}~&=~\frac{1}{1-a}\,\Big(\langle \om_{1}(\rho) \rangle-\langle m_{1}(\rho)\, d(\rho) \label{eq:C1} \rangle\Big)\\
\text{and}\quad C_{2}~&=~\frac{1}{1-a}\,\frac{\langle m_{1}(\rho)\,d(\rho) \rangle-a\langle \om_{1}(\rho)\rangle}{1+(M-1)\,a}.\label{eq:C2}
\end{align}
Moreover,
\begin{align}\label{eq:sumw}
\sum_\lambda w_{\lambda}~=~
\frac{\langle \om_{1}(\rho) \rangle\,+\,(M-1)\langle m_{1}(\rho)d(\rho) \rangle}{1+(M-1)\,a}.
\end{align}
\end{thm}
Theorem \ref{thm:asweights} follows from Theorems \ref{thm:limits} and \ref{cor:A} by a straightforward computation.
\begin{cor}Under the assumptions of Theorem \ref{thm:asweights},
the normalised weights $\overline{w}^{(N)} $ converge to
\begin{align}
\overline{w}_{\lambda}~=~\overline{C}_{1}\,\alpha_{\lambda}\;+\;\overline{C}_{2}\label{eq:normw}
\end{align}
\end{cor}
\begin{rem}
\begin{enumerate}
\item By Theorem \ref{thm:asweights}, the optimal weights are always the sum of a term
proportional to the size of the population and a term independent of the population.
The weights of the states in the Electoral College of the U.S. constitution are
precisely chosen in this fashion.
\item In the limit $a\to 0$, meaning that the groups are almost independent, the constant term in
\eqref{eq:optw} tends to 0, so that $\overline{w}_{\lambda}\to \alpha_{\lambda}$ which
is the result for the simple CBM (see \cite{HOEC}).
\item The sum of the weights \eqref{eq:sumw} is strictly positive and finite, even in the limit
$a\to 1$. This indicates that the choice $\sigma=N$ is reasonable. In fact,
\begin{align}\label{eq:sumw1}
\lim_{a\to 1} \sum_\lambda w_{\lambda}~=~\langle \om_{1}(\rho) \rangle\,.
\end{align}
\end{enumerate}
\end{rem}
\begin{cor}
Under the assumptions of Theorem \ref{thm:asweights}, the minimal democracy deficit $\Delta_{N} $ is asymptotically of the form
\begin{align*}
\Delta_{\infty}~=~D_{1}\,\sum_{\lambda=1}^{M}{\alpha_{\lambda}}^{2}\;+\;D_{2}.
\end{align*}
\end{cor}
\begin{rem}
The constants $D_{1}$ and $D_{2}$ depend on $\mu, \rho, M$ and can be computed from \eqref{eq:Ainv}, \eqref{eq:C1}, and \eqref{eq:C2}.
\end{rem}
\section{Optimal Weights for Tight Correlations}\label{sec:optwtc}
Now we turn to the case of tightly correlated models, i.e.$\!\,$ $a=1$.
Then, in the limit $N\to\infty$, setting $\sigma:=N $, equation \eqref{eq:LES} which describes the critical
points of $\Delta_{N} $ tends to $\tilde{A}w=b$ with
\begin{align*}
\tilde{A}_{\lambda\nu}~=~1 \qquad\text{for all } \lambda,\nu\,.
\end{align*}
The matrix $\tilde{A}$ is degenerate. It has an $(M-1)$-fold degenerate eigenvalue at $0$ and a simple
eigenvalue at $M $.
The democracy deficit $\Delta_{N}$ tends to
\begin{align}
\Delta_{\infty}~&=~\sum_{\lambda=1}^{M}\,\alpha_{\lambda}^{2}\,\big(\langle m_{2}(\rho) \rangle-\langle m_{1}(\rho)^{2} \rangle\big)\;\;+\langle m_{1}(\rho)^{2} \rangle\notag\\
&\quad-\;2\,\langle \om_{1}(\rho) \rangle\,\sum_{\lambda=1}^{M}\,w_{\lambda}\;+\left(\sum_{\lambda=1}^{M}\,w_{\lambda}\right)^{2}.\label{eq:da1}
\end{align}
\eqref{eq:da1} is an equation in $\sum_{\lambda}w_{\lambda}$. The extrema of $\Delta_{\infty}$ are all weights $w_{\lambda}$ such that
\begin{align*}
\sum_{\lambda=1}^{M}\,w_{\lambda}~=~\langle \om_{1}(\rho) \rangle\,.
\end{align*}
This condition is in agreement with \eqref{eq:sumw1}.
\begin{thm}\label{thm:optwtight}
Suppose $a=1$. If
\begin{align*}\sum_{\lambda=1}^{M} w_{\lambda}~ =~ \sum_{\lambda=1}^{M} v_{\lambda},
\end{align*}
then
\begin{align*}
\Delta_{N}(w)~-\Delta_{N}(v)~\to~ 0\qquad\text{as } N\to\infty.
\end{align*}
In particular, any tuple $w$ of weights with $\sum_\lambda w_{\lambda}=\langle \om_{1}(\rho) \rangle$
is close to the minimal democracy deficit in the sense that
\begin{align*}
\Delta_{N}(w)~\to~\min_{v}\; \Delta_{\infty}(v).
\end{align*}
\end{thm}
Theorem \ref{thm:optwtight} implies that for large systems with tight correlation `it doesn't matter'
how the weights are distributed among the groups. This assertion is confirmed by the following observation:
\begin{thm}\label{thm:unani}
If the model CBM($\mu,\rho $) is tightly correlated, then
\begin{align*}
\IP\,\Big(S_{\lambda}>0 \text{ for all } \lambda \quad\text{or }\quad S_{\lambda}<0 \text{ for all } \lambda \Big)~\to~1\quad\text{as }N\to\infty.
\end{align*}
\end{thm}
Thus, in large tightly correlated systems, council votes are almost always unanimous! Consequently,
for $N\to\infty$,
any $w$ with $\sum_\lambda w_{\lambda}>0$ induces the same voting result in the council. This might be surprising at first. As the overall population goes to infinity, the probability of a unanimous council vote goes to 1. Hence, the limit of the optimality condition \eqref{eq:LES} is a linear equation system with an infinity of solutions. More precisely, any set of weights $w_1,\ldots,w_M$ that sum to a fixed positive value given by the limit of \eqref{eq:b} solves \eqref{eq:LES}. As such, the assignation of the voting weights only serves the purpose of appropriately scaling the magnitude of the (unanimous) council vote to bring it in line with $S/\sigma$. However, the constraint on the sum is not binding, as we know that any transformation of a weighted voting system that multiplies all weights by a positive constant while leaving the relative quota untouched is equivalent to the original voting system. Thus, a set of weights which sum to 1 is but a representative of an equivalence class of voting systems. The selection of the optimal weights when a unanimous council vote occurs with high probability is a trivial problem.
\begin{proof}
Set
\begin{align*}
Z_{+}~=~\{ z\in[-1,1]\mid \rho^{z} (0,1]=1 \} \quad \text{and}\quad
Z_{-}~=~\{ z\in[-1,1]\mid \rho^{z} [-1,0)=1\}\,.
\end{align*}
Since the measure $\IP $ is tightly correlated, we have due to Proposition \ref{prop:tight} that
\begin{align*}
Z_{+}\;\cup\; Z_{-}~=[-1,1]\qquad\text{up to a set of $\mu $-measure $0$}\,.
\end{align*}
In particular, $\rho^{z}\not=\delta_{0} $ for $\mu $-almost all $z $, so $\IP(S_{\lambda}=0)\to 0 $
for any $\lambda $.
Thus, it suffices to prove that for any given $\nu\not=\lambda $
\begin{align*}
\IP(S_{\nu}>0, S_{\lambda}<0)\to 0.
\end{align*}
For $t\in (0,1]$, we have
\begin{align*}
P_{t}(S_{\lambda}<0)~\to~ 0;
\end{align*}
thus, for $z\in Z_{+}$,
\begin{align*}
\int\;P_{t}(S_{\lambda}<0)\;\rho^{z}(\textup{d}t)~\to~ 0
\end{align*}
and similarly, for $z\in Z_{-}$,
\begin{align*}
\int\;P_{t}(S_{\nu}>0)\;\rho^{z}(\textup{d}t)~\to~ 0.
\end{align*}
Hence
\begin{align*}
\IP(S_{\nu}>0, S_{\lambda}<0)~\leq~&\int_{Z_{+}}\,\int P_{t}(S_{\lambda}<0)\;\rho^{z}(\textup{d}t)\;\mu(\textup{d}z)
~+~\int_{Z_{-}}\,\int P_{t}(S_{\nu}>0)\;\rho^{z}(\textup{d}t)\;\mu(\textup{d}z) \to~ 0.
\end{align*}
\end{proof}
\section{Specific Models}\label{sec:specmod}
In this section, we analyse some models from Example \ref{ex:1}. In these
examples, we can compute relevant quantities explicitly.
\subsection{Additive Models}\label{sec:additive}
We start with some additive models as in Example \ref{ex:1}.\ref{ex:add} with specific bias measures $ \mu$ and $\rho $.
We recall that for additive models the voting measure $\IP\big(\ux_{1},\ux_{2},\ldots,\ux_{M}\big)$ is given by
\begin{align}\label{eq:addcbm3}
\int \left( \int P_{z+y_{1}}(\ux_{1})\,\rho(\textup{d}y_{1})
\cdots\int P_{z+y_{M}}(\ux_{M})
\,\rho(\textup{d}y_{M}) \right)\;\mu(\textup{d}z).
\end{align}
$\IP$ is indeed a voting measure if both $\mu$ and $\rho $ are symmetric, i.e.\! $\mu [a,b]=\mu [-b,-a] $ and similarly for $\rho $. $\IP$ is sufficiently random except for the (pathological) case $\mu=\frac{1}{2}\big(\delta_{1}+\delta_{-1}\big)$ and $\rho=\delta_{0} $. $\IP $ is tightly correlated if (and only if) for $\mu$-almost all $z$ either $\rho (-z,1]=1$ or $\rho [-1,-z)=1$.
\subsubsection{Uniform Distribution with Weak Global Bias}
In our first example, we take $\mu $ and $\rho $ to be the uniform probability distribution on $[-g,g] $ (for `global' bias) and
$[-\ell,\ell] $ (`local' bias), respectively. We assume first that $g\leq \ell$, indicating that the (average) global
bias is not bigger than the (average) local bias. So the voting measure $\IP(\underline{x}_{1},\ldots,\underline{x}_{M})$ is given by
\begin{align}\label{eq:addcbm}
\frac{1}{2g}\int_{-g}^{+g}\;\left(\,\frac{1}{2\ell}\int_{z-\ell}^{z+\ell} P_{t_{1}}(\underline{x}_{1})\,\textup{d}t_{1}\cdots
\frac{1}{2\ell}\int_{z-\ell}^{z+\ell} P_{t_{M}}(\underline{x}_{M})\,\textup{d}t_{M}\,\right)\,\textup{d}z.
\end{align}
For this specific example, we can explicitly compute the relevant quantities from Definition \ref{def:notation} and Theorem \ref{thm:asweights}. By a straightforward but tedious computation, we obtain:
\begin{align}
&a~=~\frac{1}{3}\,\frac{g^{2}}{\ell^{2}}~\leq~\frac{1}{3},&& \langle m_{1}(\rho)\,d(\rho)
\rangle~=~\frac{1}{3}\,\frac{g^{2}}{\ell},\notag\\
&\langle \om_{1}(\rho)
\rangle~=~\frac{1}{6\ell}\;\big(3\ell^{2}+\,g^{2}\big),
&&\langle \om_{1}(\rho)\rangle- \langle m_{1}(\rho)\,d(\rho)\rangle
~=~\frac{1}{6\ell}\big(3\ell^{2}-g^{2}\big),\notag\\
&\langle m_{1}(\rho)d(\rho)\rangle - a \langle \om_{1}(\rho)\rangle~=~ \frac{g^{2}}{18\ell^{3}}\big(3\ell^{2}-g^{2}\big). \label{eq:addcbm-1}
\end{align}
This gives
\begin{thm}\label{opt_weights_uniform}
For the additive CBM in \eqref{eq:addcbm} with $g\leq\ell$, the optimal weights are
\begin{align}\label{eq:waddgkl}
w_{\lambda}~=~\frac{1}{2}\ell\,\alpha_{\lambda}\;+\;\frac{1}{2}\,\frac{g^{2}\ell}{3\ell^{2}+(M-1)g^{2}}.
\end{align}
\end{thm}
\begin{rem}
\begin{enumerate}
\item If there is no global bias (meaning $g\searrow 0$), we obtain the result for independent groups, i.e.$\!\,$ the weights are proportional to $\alpha_{\lambda}$.
\item The quantity $\langle m_{1}d \rangle-a\langle \om_{1} \rangle$ is non-negative. This is not always the case as we will see in Section \ref{sec:nonneg}.
\end{enumerate}
\end{rem}
\subsubsection{Uniform Distribution with Strong Global Bias}
Now, we turn to the case $\ell\leq g$. In this case, we compute:
\begin{align}
&a~=~1-\frac{2}{3}\,\frac{\ell}{g}~<~1,&& \langle m_{1}(\rho)\,d(\rho)
\rangle~=~\frac{1}{6g}\,\big(3g^{2}-\ell^{2}\big),\notag\\
&\langle \om_{1}(\rho)
\rangle~=~\frac{1}{6g}\;\big(3g^{2}+\,\ell^{2}\big),
&&\langle \om_{1}(\rho)\rangle- \langle m_{1}(\rho)\,d(\rho)\rangle
~=~\frac{1}{3}\frac{\ell^{2}}{g},\notag\\
&\langle m_{1}(\rho)d(\rho)\rangle - a \langle \om_{1}(\rho)\rangle~=~ \frac{\ell}{9g^{2}}\big(3g^{2}-3g\ell +\ell^{2}\big). \label{eq:addcbm-2}
\end{align}
\begin{thm}
For the additive CBM in \eqref{eq:addcbm} with $\ell\leq g$, the optimal weights are
\begin{align}\label{eq:waddlkg}
w_{\lambda}~=~\frac{1}{2}\ell\,\alpha_{\lambda}\;+\;\frac{1}{2}\,\frac{3 g^{2}-3g\ell +\ell^{2}}{3Mg -2(M-1)\ell}.
\end{align}
\end{thm}
\begin{rem}
\begin{enumerate}
\item For the case $g=\ell$, formulae \eqref{eq:waddgkl} and \eqref{eq:waddlkg} agree.
\item In the limit $\ell\to 0 $ we find $a\to 1 $, i.e.$\!\,$ we approach the tightly correlated case. In this case, the weights become constant, independent of the sizes of the groups.
This limit case corresponds to Impartial Anonymous Culture for the \emph{union}.
\end{enumerate}
\end{rem}
\subsubsection{Global Bias Concentrated in Two Points}
We study an additive model for which the global bias may assume the value $+g,-g$ with probability $\frac{1}{2}$ each. The local bias is uniformly distributed on $[-\ell,\ell] $ with $0<g<\ell$.
We just give the final result: the optimal weights are
\begin{align*}
w_{\lambda}~=~\frac{1}{2}\ell\,\alpha_{\lambda}\;+\;\frac{1}{2}\,\frac{\ell g^{2}}{(M-1)g^{2} +\ell^{2}}.
\end{align*}
In this setting, the tightly correlated case is approached in the limit $\ell\searrow g $.
The weights tend to $w_{\lambda}\to \frac{1}{2}g\,\alpha_{\lambda}+ \frac{1}{2}\frac{g}{M}$ as $\ell\searrow g $.
\subsection{Multiplicative Models}\label{sec:mult}
We now analyse the multiplicative models in Example \ref{ex:1}.\ref{ex:mult}. If $\mu\{0\}=0$, the model is tightly correlated if and only if $\supp\,\rho\subset(0,1]$ or $\supp\,\rho\subset[-1,0)$.
Again, we consider uniform distributions on $[\ell_1,\ell_2]$ and $[-g,g]$, respectively, in more detail. The probability of each configuration $\ux_{\lambda}\in\{ -1,1 \}^{N_{\lambda}},\lambda=1,\ldots,M$, is
\begin{align}\label{eq:multcbm}
\frac{1}{2g}\int_{-g}^{+g}\;
\left(\,\frac{1}{\ell_{2}-\ell_{1}}\int_{\ell_{1}}^{\ell_{2}} P_{z y_{1}}(\underline{x}_{1})\,\textup{d}y_{1}\cdots
\frac{1}{\ell_{2}-\ell_{1}}\int_{\ell_{1}}^{\ell_{2}} P_{z y_{M}}(\underline{x}_{M})\,\textup{d}y_{M}\,\right)\,\textup{d}z\,.
\end{align}
So while the global bias measure is uniform on the interval $[-g,g]$, the local bias modifier $\rho$ is uniform on the interval $[\ell_1,\ell_2]$. If $\ell_{1}\geq 0$ (or $\ell_{2}\leq 0 $, which gives the same model class), the model is tightly correlated.
Assuming $\ell_{1}<0<\ell_{2} $, we obtain
\begin{align*}
&a~=~\frac{(\ell_{2}+\ell_{1})^{2}}{(\ell_{2}-\ell_{1})^{2}}~<~1, & \langle m_{1}(\rho)\,d(\rho)
\rangle~=~\frac{g}{4}\,\frac{(\ell_{2}+\ell_{1})^{2}}{(\ell_{2}-\ell_{1})},\notag\\
&\langle \om_{1}(\rho)
\rangle~=~\frac{g}{4}\,\frac{\ell_{2}^{2}+\ell_{1}^{2}}{(\ell_{2}-\ell_{1})},
&\langle \om_{1}(\rho)\rangle- \langle m_{1}(\rho)\,d(\rho)\rangle
~=~-\,\frac{g}{2}\,\frac{\ell_{1}\ell_{2}}{\ell_{2}-\ell_{1}},\notag\\
&\langle m_{1}(\rho)d(\rho)\rangle - a \langle \om_{1}(\rho)\rangle~=~
-\,\frac{g}{2}\,\ell_{1}\ell_{2}\,\frac{(\ell_{2}+\ell_{1})^{2}}{(\ell_{2}-\ell_{1})^{3}}~\geq~0.
\end{align*}
So, for the optimal weights according to Theorem \ref{thm:asweights}, we obtain
\begin{align}
w_{\lambda}~&=~\frac{g}{8}\,(\ell_{2}-\ell_{1})\,\alpha_{\lambda}\;
+\;\frac{g}{8}\,(\ell_{2}-\ell_{1})\,\frac{(\ell_{2}+\ell_{1})^{2}}{(\ell_{2}-\ell_{1})^{2}+
(M-1)(\ell_{2}+\ell_{1})^{2}}, \label{eq:mult1}\\
\intertext{or, equivalently,}\tilde{w}_{\lambda}~&=~\alpha_{\lambda}\;
+\;\frac{(\ell_{2}+\ell_{1})^{2}}{(\ell_{2}-\ell_{1})^{2}+
(M-1)(\ell_{2}+\ell_{1})^{2}}\,.\label{eq:mult2}
\end{align}
For $\ell_{1}\nearrow 0$ approaching the tightly correlated case, we get
\begin{align*}
\tilde{w}_{\lambda}~&=~\alpha_{\lambda}\;
+\;\frac{1}{M}\,.
\end{align*}
Moreover, we observe that the formulae \eqref{eq:mult1} and \eqref{eq:mult2} make sense even in the
tightly correlated case, i.e.$\!\,$ for $\ell_{1}\geq 0$.
Next we turn to the case where $\rho(0,1]=1$ while maintaining the condition \eqref{eq:rhotriv}. Then there are only two possibilities: either the model is tightly correlated and the optimal weights are indeterminate. This is the case if and only if $\mu\{0\}=0$. The complementary case is $0<\mu\{0\}<1$. We can interpret this as the existence of some fraction of the issues which are not subject to any global bias. The multiplicative structure of the local bias means all voters make up their own minds on these issues. We can determine the optimal weights in this case without placing any additional assumptions on the bias measures $\mu$ and $\rho$.
The key observation is that for $\mu$-almost all $z$ the equality $m_{1}(\rho)\,d(\rho) = \om_{1}(\rho)$ holds. The model is not tightly correlated, nor are the voters belonging to different groups independent. So we have $0<a<1$ and
\begin{align*}
w_{\lambda}~&=~\frac{\langle \om_{1}(\rho) \rangle}{1+(M-1)\,a}\quad
\text{or, equivalently,}\quad\overline{w}_{\lambda}~=~\frac{1}{M}\,.
\end{align*}
In conclusion, for this model, the optimal weights have to be chosen equal for all groups $\lambda$, no matter their size $\alpha_\lambda$. As for the intuition behind this result, let us recall the formula for optimal weights in non-tightly correlated models given in Theorem \ref{thm:asweights}: the optimal weights are given by $C_1 \alpha_\lambda+C_2$, with
\begin{align*}
C_{1}~&=~\frac{1}{1-a}\,\Big(\langle \om_{1}(\rho) \rangle-\langle m_{1}(\rho)\, d(\rho) \rangle\Big)\\
\text{and}\quad C_{2}~&=~\frac{1}{1-a}\,\frac{\langle m_{1}(\rho)\,d(\rho) \rangle-a\langle \om_{1}(\rho)\rangle}{1+(M-1)\,a}.
\end{align*}
As mentioned after Definition \ref{def:notation}, $\om_{1}(\rho)$ can be interpreted as a measure of intra-group cohesion, and $m_{1}(\rho)\,d(\rho)$ as a measure of inter-group cohesion. The equality of these two in the present example is intuitively due to the fact that $Z=0$ induces zero cohesion both within each group as well as across group boundaries, and for $Z\neq0$ we have a very strong correlation of all voters due to the assumption $\rho(0,1]=1$ which implies that the sign of each group bias will always be the same as the sign of the global bias. Hence, we have $\langle m_{1}(\rho)\,d(\rho) \rangle = \langle \om_{1}(\rho) \rangle$, and $C_1=0$ implies there is no proportional component to the optimal weights. The constant component $C_2$, on the other hand, does not disappear, because the fraction of issues for which there is independent voting (i.e.\! those for which $Z=0$), induces a non-tight correlation and $a<1$.
This example sheds some light on where the summands in the optimal weight formula in Theorem \ref{thm:asweights} come from: as mentioned previously, the constant component $C_2$ is induced by the correlation between votes belonging to different groups. Now we see that the proportional component is a manifestation of the stronger cohesion within each group when compared to inter-group cohesion.
\section{Non-Negativity of the Weights}\label{sec:nonneg}
In applications on public voting procedures, negative weights would be rather absurd: the consent of
such a voter could decrease the majority margin or even change an `aye' to a `nay'. It seems likely that no group would accept being assigned a negative voting weight. Even if they did, this would not bring about a minimisation of the democracy deficit, since a group with negative weight would face incentives to misrepresent their true preferences. On the other hand, in an
estimation problem, i.e.$\!\,$ for estimating the magnitude of the voting margin, negative weights may make sense. It has been pointed out by an anonymous referee that negative weights may also make sense in an automated preference aggregation setting, when sincere voting can be assumed.
In Theorem \ref{thm:asweights}, we identified the optimal weights $w_{\lambda} $ as
\begin{align}\label{eq:weights2}
w_{\lambda}~=~C_{1}\,\alpha_{\lambda}\;+\;C_{2}\,.
\end{align}
The constant $C_{1} $ is always non-negative. Moreover, in all explicit examples in Section \ref{sec:specmod} the constant $C_{2}$ turned out to be non-negative as well.
In general, the constant $C_{2}$ is non-negative if and only if
\begin{align}\label{eq:nonneg}
a\langle \om_1(\rho) \rangle \leq \langle m_{1}(\rho) \, d(\rho) \rangle \,.
\end{align}
As it turns out, condition \eqref{eq:nonneg} \emph{can} be violated under certain assumptions on
the measures $\mu $ and $\rho^{z} $. Consequently, for small $\alpha_{\lambda} $, equation \eqref{eq:weights2} prescribes negative weights.
\subsection{An Example with Negative Optimal Weights}
To see that \eqref{eq:nonneg} can be violated, we consider an additive model with $\mu=\eh(\delta_{g}+\delta_{-g})$ and $\rho=\frac{1}{4}(\delta_{-\ell_{2}}+\delta_{-\ell_{1}}+\delta_{\ell_{1}}+\delta_{\ell_{2}}) $ and choose $0<\ell_{1}<g<\ell_{2} $ with $g+\ell_{2}\leq 1$.
Then, for the additive model with $\mu $ and $\rho $ we compute:
\begin{align*}
a=\langle d(\rho)^{2} \rangle=\frac{1}{4},&&\langle m_{1}d(\rho) \rangle=\eh g,
&&\langle \om_{1}(\rho)\rangle=\eh g + \eh \ell_{2}.
&\end{align*}
Consequently, the constant term $C_{2}$ in the optimal weight \eqref{eq:weights2} is \emph{negative} if $\ell_{2}>3 g$. In this case, the optimal weight is negative
for small $\alpha_{\lambda}$.
An analogous result holds for uniform distributions both for $\mu $ (around $\pm g$) and for $\rho $
around $\pm\ell_{1}$ and $\pm\ell_{2} $, as long as these six intervals are small enough.
In the remainder of this section, we will focus on the additive model and the problem of negative weights. For simplicity's sake, we will assume for the rest of Section \ref{sec:nonneg} that the support of both $\mu$ and $\rho$ belongs to $[-1/2,1/2]$.
\subsection{Non-Negativity of $w$ in Additive Collective Bias Models with $\mu=\rho$}
In this section, we consider the case where the central bias and the
group modifiers of an additive CBM follow the same distribution. Of course, all bias variables and modifiers
$Z$ and the $Y_{\lambda}$ are still assumed to be independent. So the random variables $Z,Y_1,\ldots,Y_M$ are all i.i.d. As noted in Section \ref{warm-up}, the assumption of identically distributed $Z$ and $Y_\lambda$ reflects that global bias and local bias each have the same influence on the voters, with neither of the two dominating. We will use the notation
\[
r:=\langle m_{1}(\rho)\,d(\rho) \rangle,\quad m:=\langle \om_{1}(\rho) \rangle.
\]
Recall that according to Theorem \ref{thm:asweights}, the optimal weights are proportional to
\[
w_{\lambda}=r-am+\left(1+\left(M-1\right)a\right)\left(m-r\right)\alpha_{\lambda}.
\]
We prove that for this setup the optimal weights can never be negative.
\begin{thm}
\label{thm:mu_rho}If $\mu=\rho$ in an additive CBM, the constant term in
the optimal weights $r-am$ is non-negative and $r-am=0$ holds if
and only if $\mu=\delta_{0}$. Furthermore, $0\leq a\leq1/3$, where
$a=0$ holds if and only if $\mu=\delta_{0}$, and $a=1/3$ if and
only if $\mu$ has no atoms, i.e., for all $x\in\mathbb{R}$, $\mu\left\{ x\right\} =0$.
\end{thm}
This theorem says -- among other things -- that the constant term
in the optimal weights $r-am$ is 0 if and only if $\mu=\rho=\delta_{0}$.
But the latter equality implies that all voters are independent, a
case which we discarded earlier. (Note that if $\mu=\rho=\delta_{0}$,
the optimal weights are \emph{not} proportional to the group sizes.
Instead, the square root law holds and the optimal weights are proportional
to $\sqrt{\alpha_{\lambda}}$.) Hence, by Theorem \ref{thm:mu_rho},
for all $\mu=\rho\neq\delta_{0}$, the optimal weights are the sum
of a positive constant $r-am>0$ and a term proportional to the group
size $\alpha_{\kappa}$.
Under the assumption $\mu=\rho$, we consider $a=\mathbb{E}\left(\chi_{1}\chi_{2}\right)$
as a function of the measure $\mu$. So $a:\mathcal{M}_{\leq1}\left(\left[-1/2,1/2\right]\right)\rightarrow\mathbb{R}_{+}$,
where $\mathcal{M}_{\leq1}\left(\left[-1/2,1/2\right]\right)$ is
the set of all sub-probability measures on $\left[-1/2,1/2\right]$.
We will also write $\mathcal{M}_{1}\left(\left[-1/2,1/2\right]\right)$
for the set of all probability measures. Similarly, $r$ is a function
$r:\mathcal{M}_{\leq1}\left(\left[-1/2,1/2\right]\right)\rightarrow\mathbb{R}_{+}$.
To show the theorem, we consider the cases of discrete and continuous
measures separately first and then show the general case.
\begin{prop}
\label{prop:mu_rho_discr}If $\mu$ is discrete, then we have $0\leq a\left(\mu\right)<1/3$.
The supremum over all discrete measures of $a\left(\mu\right)$ is
$1/3$. Within the class of discrete measures with at most $n$ points
belonging to $\textup{supp }\mu$, we have
\[
a\left(\mu\right)\leq\begin{cases}
\frac{\left(n-2\right)\left(n+2\right)}{3n^{2}}, & n\textup{ even},\\
\frac{\left(n-1\right)\left(n+1\right)}{3n^{2}}, & n\textup{ odd}.
\end{cases}
\]
\end{prop}
For measures $\mu$ with no atoms, we have
\begin{prop}
\label{prop:mu_rho_cont}If $\mu\in\mathcal{M}_{\leq1}\left(\left[-1/2,1/2\right]\right)$
has no atoms, then $a\left(\mu\right)\leq1/3$. If $\mu\in\mathcal{M}_{1}\left(\left[-1/2,1/2\right]\right)$,
then $a\left(\mu\right)=1/3$.
\end{prop}
For the remainder of this article, we express $m$ as the sum of two
terms:
\begin{align*}
m & =E\left|T_{1}\right|=E\left|Z+Y\right|=E\left(\,\textup{sgn\ensuremath{\left(Z+Y\right)}}\ensuremath{\cdot}\left(Z+Y\right)\right)=E\left(\,Z\,\textup{sgn\ensuremath{\left(Z+Y\right)}}\right)+E\left(\,Y\,\textup{sgn\ensuremath{\left(Z+Y\right)}}\right).
\end{align*}
The first of these summands equals $r$. The second one, we will call
$s$ from now on. If $\mu=\rho$, then of course $r=s$, and the term
$r-am$ equals $r\left(1-2a\right)$. For the proof of
these results, we need the following auxiliary lemma:
\begin{lem}
\label{lem:mu_rho_a_r}We can express the magnitudes $a\left(\mu\right)$
and $r\left(\mu\right)$ as
\begin{align*}
a\left(\mu\right) & =2\int_{\left(0,1/2\right]}\left(\mu\left(-z,z\right]\right)^{2}\mu\left(\textup{d}z\right),\quad r\left(\mu\right)=2\int_{\left(0,1/2\right]}z\mu\left(-z,z\right]\mu\left(\textup{d}z\right).
\end{align*}
\end{lem}
\begin{cor}
\label{cor:a_r_0}The terms $a$ and $r$ equal $0$ if and only if $\mu=\delta_{0}$.
\end{cor}
This follows easily from the representation of $a$ and $r$ given
in Lemma \ref{lem:mu_rho_a_r}.
The statements in this section are proved in the appendix.
\subsection{Non-Negativity of $w$ in Additive Collective Bias Models with $\mu\protect\neq\rho$}
In this section, we will not assume the two measures $\mu$ and $\rho$ are equal. The random variables $Z,Y_1,\ldots,Y_M$ are all independent and $Y_1,\ldots,Y_M$ are i.i.d.$\!\,$ copies of a random variable $Y$ that follows a distribution according to $\rho$. As we already know, $r-am<0$ is possible in this case. We will give conditions under which this does not
happen. Analogously to Lemma \ref{lem:mu_rho_a_r}, we have these
representations of $a,r$, and $s$:
\begin{lem}
\label{lem:mu_neq_rho_a_r_s}We can express the magnitudes $a,r$,
and $s$ as
\begin{align*}
a & =2\int_{\left(0,1/2\right]}\left(\rho\left(-z,z\right]\right)^{2}\mu\left(\textup{d}z\right),\quad r=2\int_{\left(0,1/2\right]}z\rho\left(-z,z\right]\mu\left(\textup{d}z\right),\quad s=2\int_{\left(0,1/2\right]}y\mu\left(-y,y\right]\rho\left(\textup{d}y\right).
\end{align*}
\end{lem}
First we note that if the group modifiers override the central bias
almost surely, then the groups are independent (but the voters within
each group are still positively correlated!). In this case, the optimal
weights are proportional to the group sizes.
\begin{prop}
If $\left|Y\right|$ almost surely dominates $\left|Z\right|$, then
we have $a=r=0$ and $r-am=0$.
\end{prop}
This easily follows from Lemma \ref{lem:mu_neq_rho_a_r_s}.
\begin{rem}
If, instead, $\left|Z\right|$ almost surely dominates $\left|Y\right|$,
then the CBM is tightly correlated. We note, however, that in that case any set of weights
is optimal, among them weights proportional to the group sizes.
\end{rem}
Now we turn to first order stochastic dominance which is a weaker
form of the general concept of stochastic dominance.
\begin{defn}
We say that a random variable $X_{1}$ first order stochastically
dominates a random variable $X_{2}$ if, for all $x\in\mathbb{R}$,
$P\left(X_{1}\leq x\right)\leq P\left(X_{2}\leq x\right)$ holds.
We will write $X_{1}\succ X_{2}$ for this relation and FOSD for first
order stochastic dominance.
\end{defn}
This is weaker than almost sure dominance as it is possible to have
$X_{1}$ first order stochastically dominate $X_{2}$ without $X_{1}>X_{2}$
holding almost surely. We have the following sufficient conditions
for the non-negativity of the optimal weights:
\begin{prop}
\label{prop:FOSD_suff}If $\left|Z\right|\succ\left|Y\right|$ and
$a\leq1/2$, then $r-am\geq0$. If $\left|Y\right|\succ\left|Z\right|$
and $s\leq2r$, then $r-am\geq0$.
\end{prop}
The next idea is to assume that the measures $\mu$ and $\rho$ assign
similar probabilities to each event.
\begin{prop}
\label{prop:ribbon}Suppose there are constants $c,C>0$ such that,
for all measurable sets $A$,
\[
c\rho A\leq\mu A\leq C\rho A
\]
holds. Then each of the following two conditions is individually sufficient
for $r-am\geq0$:
\[
\text{1}.\;c\geq\frac{C^{2}}{3-C},C<3,\quad\text{2}.\;C\leq c\left(3c^{2}-1\right).
\]
If we assume additionally that $c=1/C$, then a sufficient condition
for $r-am\geq0$ is given by
\[
a\leq\frac{1}{1+C^{2}}.
\]
\end{prop}
Earlier we saw that if both $\mu$ and $\rho$ are uniform distributions (we will write $\mathcal{U}$ for a uniform distribution)
on symmetric intervals around the origin, $r-am\geq0$ holds. We can
generalise this result as follows:
\begin{prop}
\label{prop:rho_unif_mu_any}Let $\rho=\mathcal{U}\left[-1/2,1/2\right]$
and $\mu\in\mathcal{M}_{1}\left(\left[-1/2,1/2\right]\right)$. Then
$r-am\geq0$ is satisfied and $r=am$ if and only if $\mu=1/2\left(\delta_{-1/2}+\delta_{1/2}\right)$.
\end{prop}
\begin{rem}
Since $\mu=1/2\left(\delta_{-1/2}+\delta_{1/2}\right)$ implies that
$\left|Z\right|$ almost surely dominates $\left|Y\right|$, we can
disregard this case. Thus, this proposition implies that for $\rho=\mathcal{U}\left[-1/2,1/2\right]$
the optimal weights are given by a constant and a proportional part.
We also note here that $\rho$ being uniform on the entire interval
$\left[-1/2,1/2\right]$ is important. For every $0<\gamma<1/2,\rho=\mathcal{U}\left[-\gamma,\gamma\right]$,
there is a $\mu$ such that $r-am$ is negative.
The last result in this section concerns a case where $\rho$ is some
symmetric measure and $\mu$ is a contracted version of $\rho$ onto
some shorter interval $\left[-c/2,c/2\right]$ for some $0<c<1$. Hence, the global bias tends to be weaker than the local bias modifier.
For the rest of this section, assume the following conditions hold
\end{rem}
\begin{ass}\label{def:mu_contraction}
\begin{enumerate}
\item $\rho$ has no atoms.
\item There is a function $g:\left[0,\infty\right)\rightarrow\mathbb{R}$
with the property that $\rho\left(0,xy\right)=g\left(x\right)\rho\left(0,y\right)$
holds for all $x\geq0$ and all $y\in\left[0,1/2\right]$ such that
$xy\leq1/2$.
\item $\mu\left(cA\right)=\rho A$ for some fixed $0<c<1$ and all measurable
$A$.
\end{enumerate}
\end{ass}
The second point is a homogeneity condition. The last point is the
aforementioned contraction property.
Let $F_{\rho}$ be the distribution function of the sub-probability
measure $\rho\,\big|\!\left[0,1/2\right]$, i.e.$\!\,$ $\rho$ constrained to the subspace $[0,1/2]$. Note that due to property
1 above, $\rho\left[0,1/2\right]=1/2$, and hence $F_{\rho}\left(0\right)=0$
and $F_{\rho}\left(1/2\right)=1/2$.
The three properties in Assumptions \ref{def:mu_contraction} already determine that the measures $\rho$
and $\mu$ belong to a two-parameter family indexed by $\left(t,c\right)\in\left(0,\infty\right)\times\left(0,1\right)$.
\begin{lem}
\label{lem:g_F_properties}If the second condition in Assumptions \ref{def:mu_contraction}
is satisfied, then
\begin{enumerate}
\item For all $y\in\left[0,1\right]$, $g\left(y\right)=2F_{\rho}\left(y/2\right)$.
\item $\rho$ has no atoms, unless $\rho=\delta_{0}$.
\item $g$ is multiplicative: for all $x,y\geq0$, $g\left(xy\right)=g\left(x\right)g\left(y\right)$.
\item $F_{\rho}$ has the form $F_{\rho}\left(y\right)=2^{t-1}y^{t}$ for
some fixed $t\geq0$.
\end{enumerate}
\end{lem}
\begin{rem}
If $t=0$ in the last point of the lemma, then $\rho=\mu=\delta_{0}$
and all voters are independent. We avoid this case by specifying the
first condition in Assumptions \ref{def:mu_contraction}.
\end{rem}
Now we state the theorem concerning the sign of the term $r-am$.
\begin{thm}
\label{thm:fam_rho_mu}Let the conditions stated in Assumptions \ref{def:mu_contraction}
hold for $\rho$ and $\mu$. Then, by the last lemma, $F_{\rho}\left(y\right)=2^{t-1}y^{t}$.
If $0<t<1$, then there is a unique $c_{0}\in\left(0,1\right)$ such
that, for all $c\in\left(0,c_{0}\right)$, $r-am$ is negative, and, for
all $c\in\left[c_{0},1\right)$, $r-am\geq0$ with equality if and
only if $c=c_{0}$. If $t\geq1$, then $r>am$.
The critical point $c_{0}$ for the regime $t\in\left(0,1\right)$
satisfies $\lim_{t\nearrow1}c_{0}=0$.
\end{thm}
\begin{rem}
Note that $t=1$ is the case of the uniform distributions $\rho=\mathcal{U}\left[-\gamma,\gamma\right],\mu=\mathcal{U}\left[-\beta,\beta\right]$
with $\gamma=1/2$ and $0<\beta=c/2<1/2$.
\end{rem}
The proofs of these statements can be found in the appendix.
\section{Extensions}\label{sec:extensions}
An obvious extension to the general CBM framework is to allow different
conditional distributions $\rho_{\lambda}^{z}$ for the different groups to
account for more strongly or more weakly correlated groups. More precisely,
\begin{align}
\IP\big(\ux_{1},\ux_{2},\ldots,\ux_{M}\big)~
=~&\int \left( \int P_{t_{1}}(\ux_{1})\,\rho_{1}^{z}(\textup{d}t_{1})
\cdots\int P_{t_{M}}(\ux_{M})
\,\rho_{M}^{z}(\textup{d}t_{M}) \right)\;\mu(\textup{d}z)\label{eq:extCBM}
\end{align}
A large part of the analysis in Sections \ref{sec:demdef} and \ref{sec:asymp} can be done for this more general case as well. In fact, with the definitions \eqref{eq:A}, \eqref{eq:b}, and \eqref{eq:AN}, the optimal weights $w$ for this model again satisfy
\begin{align*}
A_{N}\,w~=~b_{N}.
\end{align*}
In the limit $N\to\infty $, using the same technique as in Section \ref{sec:asymp}, we obtain
\begin{align*}
A\,w~=~b,
\end{align*}
with
\begin{align*}
A_{\lambda\nu}~&=~ \left\{
\begin{array}{cl}
1, & \hbox{if $\lambda=\nu $,} \\
\big\langle\, d(\rho_{\lambda})\, d(\rho_{\nu}) \,\big\rangle , & \hbox{if $\lambda\not=\nu $,}
\end{array}
\right.
\\
b_{\nu}~&=~\big(\langle \om_{1}(\rho_{\nu})\rangle-\langle m_{1}(\rho_{\nu})d(\rho_{\nu}) \rangle\big)\,\alpha_{\nu}\;+\;\sum_{\lambda=1}^{M}\,\big\langle\, m_{1}(\rho_{\lambda})\, d(\rho_{\nu}))\, \big\rangle\,\alpha_{\lambda}.
\end{align*}
As in Proposition \ref{prop:posdef}, it is easy to see, that the matrix $A$ is positive semi-definite. Moreover, we show
\begin{thm}\label{thm:extpos}
The matrix $A$ is positive definite, and hence invertible, if $\big\langle\, d(\rho_{\lambda})^{2} \,\big\rangle<1$ for all but possibly one $\lambda$.
\end{thm}
\begin{proof}
For $x\in\IR^{M}$, we compute
\begin{align}
\big(x,A\,x\big)~&=~\sum_{\nu=1}^{M} \big(1-\big\langle\, d(\rho_{\nu})^{2} \,\big\rangle\big)\,x_{\nu}^{2}\;+\;
\sum_{\nu,\lambda=1}^{M} \big\langle\, d(\rho_{\nu})\, d(\rho_{\lambda}) \,\big\rangle\,x_{\nu}\,x_{\lambda}\notag\\
&=~\sum_{\nu=1}^{M} \big(1-\big\langle\, d(\rho_{\nu})^{2} \,\big\rangle\big)\,x_{\nu}^{2}\;
+\;\left\langle \left(\sum_{\nu=1}^M x_{\nu} d(\rho_{\nu})\right)^{2} \right\rangle\label{eq:posdef}
\end{align}
Both terms in \eqref{eq:posdef} are non-negative. If $\big\langle\, d(\rho_{\nu})^{2} \,\big\rangle<1$ for \emph{all} $\nu $, the first sum in \eqref{eq:posdef} is strictly positive for $x\not=0 $, hence $A$ is positive definite in this case. If all but one $\nu^*$ have this property, the conclusion follows from considering that the first summand is 0 if and only if $x_{\nu^*}$ is the only coordinate of $x$ different than 0. But in that case, the second summand is positive.
\end{proof}
We have two partial converses to Theorem \ref{thm:extpos}.
\begin{prop}
If $|\,\langle \, d(\rho_{\nu})\,d(\rho_{\lambda})\,\rangle\,|=1 $ for some $\nu\not=\lambda $, then
the matrix $A$ is \emph{not} invertible.
\end{prop}
\begin{proof}
If $|\,\langle \, d(\rho_{\nu})\,d(\rho_{\lambda})\,\rangle\,|=1 $, then an application of the
Cauchy-Schwarz inequality shows that $\langle \, d(\rho_{\nu})^{2}\,\rangle = \langle\,d(\rho_{\lambda})^{2}\,\rangle=1 $.
This implies that the matrix
\begin{align*}
\begin{pmatrix} \langle \, d(\rho_{\nu})^{2}\,\rangle & \langle \, d(\rho_{\nu})\,d(\rho_{\lambda})\,\rangle\\
\langle \, d(\rho_{\nu})\,d(\rho_{\lambda})\,\rangle & \langle\,d(\rho_{\lambda})^{2}\,\rangle
\end{pmatrix}
\end{align*}
is not invertible, hence $A $ is not invertible.
\end{proof}
\begin{prop}
Let for $\mu$-almost all $z$ $\textup{sgn }d\left(\rho_{\lambda}\right)=\textup{sgn } d\left(\rho_{\nu}\right), \lambda,\nu=1,\ldots,M$. Then the matrix $A$ being positive definite implies
$\big\langle\,d(\rho_{\lambda})^{2}\,\big\rangle<1$ for all but possibly
one $\lambda$.
\end{prop}
\begin{rem}
Additive CBMs satisfy the condition of all $d\left(\rho_{\lambda}\right)$
having the same sign for $\mu$-almost all $z$ owing to the symmetry
of each measure $\rho_{\lambda}$. Multiplicative CBMs have this property
if we assume a certain asymmetry for each $\rho_{\lambda}$: If, for all
$\lambda$, $\rho_{\lambda}\left(0,1\right]>1/2$, or, for all $\lambda$, $\rho_{\lambda}\left(0,1\right]<1/2$,
then the condition holds. This can be interpreted as all group bias
modifiers tending to reinforce the global bias, or, to the contrary,
all tending to go against the global bias. Note, however, that this
is a far weaker tendency than required by tight correlation.
\end{rem}
\begin{proof}
Assume there are two distinct indices $\lambda_{1}$ and $\lambda_{2}$ such
that $\langle\,d(\rho_{\lambda_{1}})^{2}\,\rangle,\langle\,d(\rho_{\lambda_{2}})^{2}\,\rangle=1$.
We show that $A$ is not positive definite. Define an $x\in\mathbb{R}^{M}$
by setting $x_{\lambda_{1}}=1,x_{\lambda_{2}}=-1$, and all other entries equal
to 0. We calculate
\begin{align*}
\big(x,A\,x\big)~ & =~\big(1-\big\langle\,d(\rho_{\lambda_{1}})^{2}\,\big\rangle\big)\,+\big(1-\big\langle\,d(\rho_{\lambda_{2}})^{2}\,\big\rangle\big)\;+\;\left\langle\big(d(\rho_{\lambda_{1}})-d(\rho_{\lambda_{2}})\big)^{2}\right\rangle\\
& =\left\langle d(\rho_{\lambda_{1}})^{2}\right\rangle -2\left\langle d(\rho_{\lambda_{1}})d(\rho_{\lambda_{2}})\right\rangle +\left\langle d(\rho_{\lambda_{2}})^{2}\right\rangle
\end{align*}
By assumption, $d(\rho_{\lambda_{1}})^{2}$ and $d(\rho_{\lambda_{2}})^{2}$
are 1 $\mu$-almost surely. Hence, $d(\rho_{\lambda_{1}})$ and $d(\rho_{\lambda_{2}})$
are 1 in absolute value and they have the same sign $\mu$-almost
surely. Thus, the term $\big\langle d(\rho_{\lambda_{1}})d(\rho_{\lambda_{2}})\big\rangle$
equals 1.
\end{proof}
We will now consider a scenario in which there are two clusters of
groups -- think of them as parts of the overall population that tend
to vote together. While voters within clusters tend to hold the same
opinion, we will assume that there is antagonism between the two clusters.
Let groups $1,\ldots,M_{1}$ belong to cluster $C_{1}$ and $M_{1}+1,\ldots,M$
to cluster $C_{2}$. We set $M_{2}:=M-M_{1}$. The fraction of the
overall population belonging to groups in each cluster will be called
$\eta_{i}:=\sum_{\lambda\in C_{i}}\alpha_{\lambda},i=1,2$. The conditional
distributions $\rho_{\lambda}^{z}$ are identical within each cluster:
In $C_{i}$, all groups follow $\rho_{i}^{z}$. To obtain antagonistic
behaviour, we will assume that
\begin{equation}
\rho_{1}^{z}=\rho_{2}^{-z}\label{eq:antisymmetry}
\end{equation}
holds for all $z$.
Now we have to distinguish the quantities
\[
r_{ij}:=\left\langle m_{1}\left(\rho_{i}\right)d\left(\rho_{j}\right)\right\rangle ,\quad m_{i}:=\left\langle \overline{m}_{1}\left(\rho_{i}\right)\right\rangle ,\quad i,j=1,2.
\]
However, due to the antisymmetry condition (\ref{eq:antisymmetry}),
we have the following equalities:
\begin{lem}
\label{lem:m,r,i,j}Under the assumptions presented above, we have
\begin{align*}
r :=r_{ii}=-r_{ij}\quad and \quad m :=m_{i}
\end{align*}
for all $i,j=1,2,i\neq j$.
\end{lem}
\begin{proof}
We omit the short calculation that yields the result.
\end{proof}
The covariance matrix $A$ has block form
\[
A=
\left(\begin{array}{cc}
A_{1} & B\\
B^{T} & A_{2}
\end{array}\right),
\]
where the $A_{i}$ are the covariance matrices of the groups belonging
to cluster $i$. They have the form we know from \eqref{eq:Ainf},
i.e. diagonal entries equal 1 and off-diagonal entries $0<a<1$. Of
course, $A_{i}\in\mathbb{R}^{M_{i}\times M_{i}}$. The matrix $B$
holds the covariances between groups of different clusters. Due to
(\ref{eq:antisymmetry}), all entries of $B$ are equal to $-a$.
We invert $A$ and obtain
\[
\left(A^{-1}\right)_{\lambda\nu}=\frac{1}{D}\begin{cases}
1+\left(M-2\right)a, & \lambda=\nu,\\
-a, & \lambda,\nu\in C_{i},\quad i=1,2,\quad\lambda\neq\nu,\\
a, & \lambda\in C_{i},\quad\nu\in C_{j},\quad i,j=1,2,\quad i\neq j,
\end{cases}
\]
where $D=\left(1-a\right)\left(1+\left(M-1\right)a\right)$. Note
that the entries within clusters are identical to those given in \eqref{eq:Ainv} for the model with identical conditional distributions $\rho_{\lambda}^{z}$.
Using Lemma \ref{lem:m,r,i,j}, we calculate the entries $\lambda\in C_i,i=1,2,$
of $b$:
\[
b_{\lambda}=\left(m-r\right)\alpha_{\lambda} + r\left(\eta_i - \eta_j\right).
\]
In the formula above, the index $j$ is the cluster $\lambda$ does \emph{not} belong to.
Now a lengthy but straightforward calculation yields the optimal
weights for each group $\nu$:
\begin{thm}
Let $\lambda$ be a group in cluster $i=1,2$ and let $j$ be the other cluster. Then the optimal weight of group $\lambda$ is given by
\begin{align*}
w_{\lambda} =\left(A^{-1}b\right)_{\lambda} = D_{1}\alpha_{\lambda}+D_{2},
\end{align*}
where the coefficients are
\begin{align*}
D_1 =\frac{m-r}{1-a}\quad and \quad D_2 =\frac{\left(r-am\right)\left(\eta_{i}-\eta_{j}\right)}{\left(1-a\right)\left(1+\left(M-1\right)a\right)}.
\end{align*}
\end{thm}
$D_1$ is equal to the coefficient $C_1$ given in \eqref{eq:C1} for the model with identical conditional distributions. We note that if both clusters have exactly half the overall population,
then $D_{2}$ vanishes, and the optimal weights are proportional to
the population of each group. If the two clusters represent different
proportions of the overall population, then $w_{\lambda}$ is the sum
of a proportional term $D_{1}\alpha_{\lambda}$ and a constant $D_{2}$.
If $\lambda$ belongs to the larger of the two clusters, then $D_{2}$
has the same sign as the coefficient $C_{2}$ in \eqref{eq:C2} in the
identical conditional distribution model, and $D_{2}$ is a rescaled version of $C_2$ by the factor $\left(\eta_{i}-\eta_{j}\right)$.
If $\lambda$ belongs to the smaller of the two clusters, then $D_{2}$
has the opposite sign compared to $C_{2}$ in \eqref{eq:C2} and it is
once again rescaled.
\section{Conclusion}\label{sec:conclusion}
We have defined and analysed a multi-group version of the CBM which allows for correlated voting across group boundaries. This CBM was then applied to the problem of calculating the optimal weights in a two-tier voting system. By the term `optimal weights', we mean those council weights which minimise the democracy deficit, i.e.\! the expected quadratic deviation of the council vote from a hypothetical referendum over all possible issues which can be voted on. The main findings in this paper are:
\begin{itemize}
\item{We determined the asymptotic behaviour of the CBM in Theorem \ref{thm:convthm}. The theorem states that the global bias measure $\mu$ and the group bias measure $\rho$ describe the limiting distribution of the normalised voting margins.}
\item{We distinguished the tightly correlated case from its complement. We characterised tight correlation in Proposition \ref{prop:tight} in terms of the bias measures. Tight correlation means intuitively that there is perfect positive correlation between the different group votes in the council. This leads to non-unique optimal weights, as the assignation of weights does not matter if all groups vote alike anyway. We gave a sufficient condition for non-tight correlation between groups called `sufficient randomness'. This criterion states that all possible council votes occur with positive probability.}
\item{In the non-tightly correlated case, we showed that the optimal council weights are uniquely determined in Theorem \ref{thm:weights}. The optimal weights are given by the sum of a constant term equal for all groups and a summand which is proportional to each group's population as stated in Theorem \ref{thm:asweights}.}
\item{We analysed the optimal weights' properties and showed that there are cases in which these weights are negative for the smallest groups. This is due to the fact that while the coefficient of the group's size is positive, the constant term can have any sign, depending on the bias measures. We gave sufficient conditions for the non-negativity of the optimal weights as well as examples in which the weights are negative in Section \ref{sec:nonneg}.}
\end{itemize}
\section*{Appendix}
\subsection*{Proof of Proposition \ref{prop:mu_rho_discr}}
We prove the claim for $n=2k+1$. The case of even $n$ can be shown
analogously. We prove by induction on $k$ that
\begin{equation}
a\left(\mu\right)\leq\frac{2k\left(2k+2\right)}{3\left(2k+1\right)^{2}}, \label{eq:a_discr_inequ}
\end{equation}
with equality if $\mu$ is chosen to be the uniform distribution
on the $2k+1$ points conforming the support of $\mu$.
\uline{Base case:} Let $k=1$. Then the support of $\mu$ consists
of three points: 0 and two points $-x_{1},x_{1}$ such that $0<x_{1}\leq1/2$.
The measure $\mu$ is given by $\beta_{0}\delta_{0}+\beta_{1}\left(\delta_{-x_{1}}+\delta_{x_{1}}\right)$
and the constants satisfy $\beta_{0}+2\beta_{1}=1$. Set $\beta:=\beta_{1}$.
To show the upper bound (\ref{eq:a_discr_inequ}), we solve the maximisation
problem $\max_{\beta}a\left(\mu\right)$. The first order condition
is
\[
\left(1-\beta\right)^{2}-2\beta\left(1-\beta\right)=0,
\]
which has two solutions: $\beta=1$ and $\beta=1/3$. The second order
condition shows that $\beta=1$ minimises $a\left(\mu\right)$ and
$\beta=1/3$ maximises it. So, for $k=1$, the uniform distribution
maximises $a\left(\mu\right)$ and, for the uniform distribution $\mu_{3}$
on $\left\{ -x_{1},0,x_{1}\right\} $,
\[
a\left(\mu_{3}\right)=\frac{8}{27}=\frac{2k\left(2k+2\right)}{3\left(2k+1\right)^{2}},
\]
and the upper bound (\ref{eq:a_discr_inequ}) holds with equality.
\uline{Induction step:} Assume that for some $k\in\mathbb{N}$
and all sets $\left\{ -x_{k},\ldots,0,\ldots,x_{k}\right\} ,0<x_{1}<\cdots<x_{k}\leq1/2,$
the uniform distribution $\mu_{2k+1}$ maximises $a\left(\mu\right)$
and $a\left(\mu_{2k+1}\right)=\frac{2k\left(2k+2\right)}{3\left(2k+1\right)^{2}}$.
We add another point $1/2\geq x_{k+1}>x_{k}$ (if $x_{k}=1/2$, then
relabel the last two points) with probability $1\geq\eta\geq0$ and
solve the maximisation problem
\[
\max_{\mu,\eta}\;a\left(\left(1-2\eta\right)\mu+\eta\left(\delta_{-x_{k+1}}+\delta_{x_{k+1}}\right)\right),
\]
where $\mu$ is any symmetric probability measure on $\left\{ -x_{k},\ldots,0,\ldots,x_{k}\right\} $.
Set $\nu:=\left(1-2\eta\right)\mu$\\$+\,\eta\left(\delta_{-x_{k+1}}+\delta_{x_{k+1}}\right)$
and we calculate
\begin{align*}
& \qquad a\left(\nu\right)/2=\int_{\left(0,1/2\right]}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right)\\
& =\int_{\left(0,x_{k}\right]}\left(\left(1-2\eta\right)\mu\left(-z,z\right]\right)^{2}\left(1-2\eta\right)\mu\left(\textup{d}z\right)+\int_{\left(0,x_{k}\right]}\left(\left(1-2\eta\right)\mu\left(-z,z\right]\right)^{2}\eta\delta_{x_{k+1}}\left(\textup{d}z\right)\\
& \quad+\int_{\left(x_{k},1/2\right]}\left(\left(1-2\eta\right)+\eta\delta_{x_{k+1}}\left(-z,z\right]\right)^{2}\left(1-2\eta\right)\mu\left(\textup{d}z\right)+\int_{\left(x_{k},1/2\right]}\left(\left(1-2\eta\right)+\eta\delta_{x_{k+1}}\left(-z,z\right]\right)^{2}\eta\delta_{x_{k+1}}\left(\textup{d}z\right).
\end{align*}
The second summand is 0 because $\delta_{x_{k+1}}\left(0,x_{k}\right]=0$.
The third summand is 0 due to $\mu\left(x_{k},1/2\right]=0$. We continue
\begin{align*}
& \qquad a\left(\nu\right)/2=\\
& =\int_{\left(0,x_{k}\right]}\left(\left(1-2\eta\right)\mu\left(-z,z\right]\right)^{2}\left(1-2\eta\right)\mu\left(\textup{d}z\right)+\int_{\left(x_{k},1/2\right]}\left(\left(1-2\eta\right)+\eta\delta_{x_{k+1}}\left(-z,z\right]\right)^{2}\eta\delta_{x_{k+1}}\left(\textup{d}z\right)\\
& =\left(1-2\eta\right)^{3}\int_{\left(0,1/2\right]}\left(\mu\left(-z,z\right]\right)^{2}\mu\left(\textup{d}z\right)+\eta\left(\left(1-2\eta\right)+\eta\delta_{x_{k+1}}\left(-x_{k+1},x_{k+1}\right]\right)^{2}\\
& =\left(1-2\eta\right)^{3}a\left(\mu\right)/2+\eta\left(1-\eta\right)^{2}.
\end{align*}
As we see, $\mu$ and $\eta$ can be chosen independently of each
other to maximise $a\left(\nu\right)$. By assumption, the maximising
$\mu$ is the uniform distribution on $\left\{ -x_{k},\ldots,0,\ldots,x_{k}\right\} $
$\mu_{2k+1}$. Hence,
\begin{align*}
\max_{\nu}\;a\left(\nu\right) & =\max_{\eta}\;\left(1-2\eta\right)^{3}a\left(\mu_{2k+1}\right)+2\eta\left(1-\eta\right)^{2}.
\end{align*}
Since $a\left(\mu_{2k+1}\right)$ is independent of the choice of
$\eta$, the first order condition is
\[
3\left(1-2\eta\right)^{2}a\left(\mu_{2k+1}\right)=\left(1-\eta\right)\left(1-3\eta\right).
\]
The solutions of this quadratic equation are
\[
\eta=\frac{1}{3}\frac{2-3a\left(\mu_{2k+1}\right)}{1-2a\left(\mu_{2k+1}\right)}\pm\frac{1}{3}\sqrt{\left(\frac{2-3a\left(\mu_{2k+1}\right)}{1-2a\left(\mu_{2k+1}\right)}\right)^{2}-\frac{3}{2}\frac{2-3a\left(\mu_{2k+1}\right)}{1-2a\left(\mu_{2k+1}\right)}}.
\]
By substituting $a\left(\mu_{2k+1}\right)=\frac{2k\left(2k+2\right)}{3\left(2k+1\right)^{2}}$,
we see that the root with the negative sign gives a negative $\eta$.
The positive root is $\eta=\frac{1}{2k+3}$. This implies that the
maximising measure $\nu$ on $\left\{ -x_{k+1},\ldots,0,\ldots,x_{k+1}\right\} $
is the uniform distribution $\mu_{2(k+1)+1}$. This concludes the
proof by induction that for finitely many points in the support, the
uniform distribution maximises $a$ and this maximum is given by the
upper bound in (\ref{eq:a_discr_inequ}).
Next we show that for discrete measures with infinite support the
upper bound $1/3$ holds as well. If $\left|\textup{supp }\mu\right|=\infty$,
then $\mu$ is of the form $\beta_{0}\delta_{0}+\sum_{i=1}^{\infty}\beta_{i}\left(\delta_{-x_{i}}+\delta_{x_{i}}\right)$,
where $x_{i}>0$ for all $i\in\mathbb{N}$. Set for each $n\in\mathbb{N}$
$\nu_{n}:=\beta_{0}\delta_{0}+\sum_{i=1}^{n}\beta_{i}\left(\delta_{-x_{i}}+\delta_{x_{i}}\right)$.
To obtain a contradiction, suppose that $a\left(\mu\right)>1/3$ and
set $\tau:=a\left(\mu\right)-1/3>0$. The sequence $\left(a\left(\nu_{n}\right)\right)_{n}$
is monotonically increasing:
\begin{align*}
a\left(\nu_{n}\right) & =2\int_{\left(0,1/2\right]}\left(\nu_{n}\left(-z,z\right]\right)^{2}\nu_{n}\left(\textup{d}z\right)\leq2\int_{\left(0,1/2\right]\cap\left\{ x_{1},\ldots,x_{n}\right\} }\left(\nu_{n+1}\left(-z,z\right]\right)^{2}\nu_{n}\left(\textup{d}z\right)\\
& \quad+2\left(\nu_{n+1}\left(-x_{n+1},x_{n+1}\right]\right)^{2}\beta_{n+1}=a\left(\nu_{n+1}\right).
\end{align*}
For any $\varepsilon>0$, there is some $m\in\mathbb{N}$ such that
for all measurable sets $A\subset\left[-1/2,1/2\right]$ the inequality
$\mu A-\nu_{m}A<\varepsilon$ holds. So
\begin{align*}
a\left(\nu_{m}\right) & =2\int_{\left(0,1/2\right]}\left(\nu_{m}\left(-z,z\right]\right)^{2}\nu_{m}\left(\textup{d}z\right)>2\int_{\left(0,1/2\right]}\left(\mu\left(-z,z\right]-\varepsilon\right)^{2}\left(\mu-\varepsilon\right)\left(\textup{d}z\right)\\
& =a\left(\mu\right)-2\varepsilon\int_{\left(0,1/2\right]}\left(\mu\left(-z,z\right]\right)^{2}\textup{d}z-4\varepsilon\int_{\left(0,1/2\right]}\mu\left(-z,z\right]\mu\left(\textup{d}z\right)\\
& \quad+4\varepsilon^{2}\int_{\left(0,1/2\right]}\mu\left(-z,z\right]\textup{d}z+2\varepsilon^{2}\int_{\left(0,1/2\right]}\mu\left(\textup{d}z\right)-2\varepsilon^{3}\int_{\left(0,1/2\right]}\textup{d}z.
\end{align*}
By letting $\varepsilon$ go to 0, we see that $a\left(\nu_{n}\right)\nearrow a\left(\mu\right)$
and there is an $n\in\mathbb{N}$ such that $a\left(\nu_{n}\right)>a\left(\mu\right)-\tau/2>1/3$.
This is a contradiction because the cardinality $\left|\textup{supp }\nu_{n}\right|$
equals $2n+1$ and therefore $a\left(\nu_{n}\right)\leq1/3$.
Next we note that $\sup a\left(\mu\right)$ over all discrete probability
measures $\mu$ is $1/3$. This is easy to see because of the following
facts:
\begin{lem}
The sequences $\left(\frac{\left(n-2\right)\left(n+2\right)}{3n^{2}}\right)_{n\textup{ even}}$
and $\left(\frac{\left(n-1\right)\left(n+1\right)}{3n^{2}}\right)_{n\textup{ odd}}$
are monotonically increasing and their limit is equal to $1/3$.
\end{lem}
As we have proved, for uniform distributions $\mu_{n}$, $a\left(\mu_{n}\right)$
is equal to one of these expressions depending on the parity of $n$.
From this lemma, it follows that by choosing a discrete uniform distribution
on either an even- or odd-numbered support we can get arbitrarily
close to $1/3$. This concludes the proof of Proposition \ref{prop:mu_rho_discr}.
Next we prove the result for continuous measures.
\subsection*{Proof of Proposition \ref{prop:mu_rho_cont}}
Let $\mu\in\mathcal{M}_{1}\left(\left[-1/2,1/2\right]\right)$ have
no atoms. We show $a\left(\mu\right)=1/3$ by approximating $a\left(\mu\right)/2$
by a sum and then prove that the sum in question is a Riemann sum
of the function $x\mapsto4x^{2}$ on the interval $\left(0,1/2\right)$.
Let $\varepsilon>0$ be given. Then there is a partition $\mathcal{P}_{n}=\left(I_{1},\ldots,I_{n}\right)$
of $\left(0,1/2\right)$ with the property that for all $i=1,\ldots,n$
$\mu I_{i}<\varepsilon$. We assume the intervals are ordered from
left to right. It is possible to choose at most $n\leq\left\lceil 1/\varepsilon\right\rceil $
intervals for the partition $\mathcal{P}_{n}$. Then we define the
upper and lower sum
\begin{align*}
U\left(\mathcal{P}_{n}\right) & :=\sum_{i=1}^{n}\left(\sum_{j=1}^{i}2\mu I_{j}\right)^{2}\mu I_{i},\quad L\left(\mathcal{P}_{n}\right):=\sum_{i=1}^{n}\left(\sum_{j=1}^{i-1}2\mu I_{j}\right)^{2}\mu I_{i}.
\end{align*}
For each summand $i=1,\ldots,n$, we have
\begin{align}
\left|\left(\sum_{j=1}^{i}2\mu I_{j}\right)^{2}\mu I_{i}-\left(\sum_{j=1}^{i-1}2\mu I_{j}\right)^{2}\mu I_{i}\right| & \leq2\left|\sum_{j=1}^{i}2\mu I_{j}-\sum_{j=1}^{i-1}2\mu I_{j}\right|\mu I_{i}=2\cdot2\left(\mu I_{i}\right)^{2}<4\varepsilon^{2}.\label{eq:summand_diff}
\end{align}
In the inequality above, we used that for all $x,y\in\left[0,1\right]$
$\left|x^{2}-y^{2}\right|<2\left|x-y\right|$ holds. Also for all
$i=1,\ldots,n$ and all $z\in I_{i}$
\[
\sum_{j=1}^{i-1}2\mu I_{j}\leq\mu\left(-z,z\right]\leq\sum_{j=1}^{i}2\mu I_{j},
\]
and, therefore,
\begin{equation}
L\left(\mathcal{P}_{n}\right)\leq\int_{\left(0,1/2\right]}\left(\mu\left(-z,z\right]\right)^{2}\mu\left(\textup{d}z\right)\leq U\left(\mathcal{P}_{n}\right).\label{eq:sum_inequ}
\end{equation}
Due to (\ref{eq:summand_diff}) and (\ref{eq:sum_inequ}), we have
\begin{align*}
0 & \leq U\left(\mathcal{P}_{n}\right)-L\left(\mathcal{P}_{n}\right)\leq n\cdot4\varepsilon^{2}\leq\left\lceil 1/\varepsilon\right\rceil \cdot4\varepsilon^{2}\leq4\varepsilon\left(1+\varepsilon\right).
\end{align*}
This shows that the upper and lower sum approximate $a\left(\mu\right)/2$
well as we let the number of intervals in $\mathcal{P}_{n}$ go to
infinity.
The next step is to show $U\left(\mathcal{P}_{n}\right)$ is an upper
Riemann sum of the function $x\mapsto4x^{2}$. For the partition $\mathcal{P}_{n}$,
there is a corresponding partition $\mathcal{Q}_{n}=\left(J_{1},\ldots,J_{n}\right)$
in which the intervals are once again assumed to be ordered from left
to right and for each $i=1,\ldots,n$ the interval lengths $\left|J_{i}\right|$
equal $\mu I_{i}$. We define
\[
R\left(\mathcal{Q}_{n}\right):=\sum_{i=1}^{n}\sup_{x\in J_{i}}4x^{2}\cdot\left|J_{i}\right|.
\]
This is an upper Riemann sum of $x\mapsto4x^{2}$. On the other hand,
we have
\begin{align*}
R\left(\mathcal{Q}_{n}\right) & =\sum_{i=1}^{n}\left(2\sup_{x\in J_{i}}x\right)^{2}\left|J_{i}\right|=\sum_{i=1}^{n}\left(2\sup J_{i}\right)^{2}\left|J_{i}\right|=\sum_{i=1}^{n}\left(2\sum_{j=1}^{i}\left|J_{j}\right|\right)^{2}\left|J_{i}\right|\\
& =\sum_{i=1}^{n}\left(\sum_{j=1}^{i}2\mu I_{j}\right)^{2}\mu I_{i}=U\left(\mathcal{P}_{n}\right)\searrow a\left(\mu\right)/2.
\end{align*}
Since $R\left(\mathcal{Q}_{n}\right)$ is an upper Riemann sum of
$x\mapsto4x^{2}$, $R\left(\mathcal{Q}_{n}\right)\searrow\int_{0}^{1/2}4x^{2}\textup{d}x=1/6$
holds as we let the number of intervals in the partition go to infinity
and we are done.
Finally, we show the case of a general probability measure.
\subsection*{Proof of Theorem \ref{thm:mu_rho}}
Let $\mu\in\mathcal{M}_{1}\left(\left[-1/2,1/2\right]\right)$ . We
can express $\mu$ as the sum of a discrete sub-probability measure
$\delta$ and a sub-probability measure $\gamma$ that has no atoms.
Both $\delta$ and $\gamma$ must satisfy the symmetry condition (\ref{eq:symmetry}).
Therefore, $\delta$ must have the form $\beta_{0}\delta_{0}+\sum_{i=1}^{\infty}\beta_{i}\left(\delta_{-x_{i}}+\delta_{x_{i}}\right)$.
Similarly to the proof of Proposition \ref{prop:mu_rho_discr}, we
truncate the sum to $\delta_{n}=\beta_{0}\delta_{0}+\sum_{i=1}^{n}\beta_{i}\left(\delta_{-x_{i}}+\delta_{x_{i}}\right)$
choosing $n$ large enough for a condition $\delta A-\delta_{n}A<\varepsilon$
to hold for all measurable sets $A$ and proceed with $\delta_{n}$
instead of $\delta$. Set $\nu:=\delta_{n}+\gamma$. Our strategy
is to show that if we remove one pair of the points $-x_{i},x_{i}$
from $\textup{supp }\delta_{n}$ and add the probability mass $2\delta\left\{ x_{i}\right\} $
to $\gamma$ as a uniform distribution on two small intervals around
$-x_{i},x_{i}$, we obtain a new measure $\nu^{\left(0\right)}$ and
we increase $a$: $a\left(\nu\right)<a\left(\nu^{\left(0\right)}\right)$.
So by removing the $2n+1$ points in $\textup{supp }\delta_{n}$ in
pairs (except for the origin where we remove a single point), we obtain
a monotonically increasing finite sequence $a\left(\nu^{\left(i\right)}\right)_{i=0,\ldots,n}$.
After $n+1$ steps, we have a sub-probability measure $\nu^{\left(n\right)}$
with no atoms and the bound $a\left(\nu^{\left(n\right)}\right)\leq1/3$
thus applies.
Let $x=x_{i}$ for some $i\in\left\{ 1,\ldots,n\right\} $ and set
$\alpha:=\delta\left\{ x\right\} >0$. Let $\varepsilon>0$ be given.
Then we choose $\eta>0$ with the properties
\begin{enumerate}
\item $2\left|\int_{\left(0,x-\eta\right]}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right)-\int_{\left(0,x\right)}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right)\right|<\varepsilon$,
\item $\left[x-\eta,x\right)\cap\textup{supp }\delta_{n}=\emptyset$,
\item $\gamma\left(x-\eta,x\right)<\varepsilon$,
\item $\nu\left(-x,x\right)-\nu\left(-\left(x-\eta\right),x-\eta\right]<\varepsilon$.
\end{enumerate}
Next define the sub-probability measure
\[
\pi_{\varepsilon}:=\delta_{n}-\alpha\left(\delta_{-x}+\delta_{x}\right)+\gamma-\gamma\left|\left(x-\eta,x\right)\right.+\alpha\left(\mathcal{U}\left(x-\eta,x\right)+\mathcal{U}\left(-x,-x+\eta\right)\right).
\]
Here $\mathcal{U}$ stands for a uniform distribution. We remove the
points $-x,x$ from $\delta_{n}$ as well as the continuous measure
$\gamma$ on the interval $\left(x-\eta,x\right)$ and add in the
probability mass $2\alpha$ on small intervals close to $-x$ and
$x$, respectively. Note that by property 2 above, $\gamma\left|\left(x-\eta,x\right)\right.=\nu\left|\left(x-\eta,x\right)\right.$.
Also, by 3, $\nu\left[-1/2,1/2\right]-\varepsilon<\pi_{\varepsilon}\left[-1/2,1/2\right]$.
We divide $a\left(\nu\right)$ into four summands
\begin{align*}
a\left(\nu\right) & =2\left[\int_{\left(0,x-\eta\right]}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right)+\int_{\left(x-\eta,x\right)}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right)+\right.\\
& \quad\left.+\left(\nu\left(-x,x\right)+\alpha\right)^{2}\alpha+\int_{\left(x,1/2\right]}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right)\right].
\end{align*}
We define the terms
\begin{align*}
A_{\nu} & :=\int_{\left(0,x-\eta\right]}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right),\quad B_{\nu}:=\int_{\left(x-\eta,x\right]}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right),\quad C_{\nu}:=\int_{\left(x,1/2\right]}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right),
\end{align*}
and, analogously, we define $A_{\pi_{\varepsilon}},B_{\pi_{\varepsilon}},C_{\pi_{\varepsilon}}$
on the same intervals in each case. We note that
\[
B_{\nu}=\int_{\left(x-\eta,x\right)}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right)+\left(\nu\left(-x,x\right)+\alpha\right)^{2}\alpha.
\]
Let $T:=\int_{\left(x-\eta,x\right)}\left(\nu\left(-z,z\right]\right)^{2}\nu\left(\textup{d}z\right)$.
Due to property 1 above, we have
\[
\left|T\right|=\left|B_{\nu}-\left(\nu\left(-x,x\right)+\alpha\right)^{2}\alpha\right|<\varepsilon.
\]
Then we calculate these terms:
\uline{\mbox{$A$}:} $A_{\nu}=A_{\pi_{\varepsilon}}$.
\uline{\mbox{$B$}:}
\begin{align*}
B_{\pi_{\varepsilon}} & =\int_{\left(x-\eta,x\right]}\left(\pi_{\varepsilon}\left(-z,z\right]\right)^{2}\pi_{\varepsilon}\left(\textup{d}z\right)=\int_{\left(x-\eta,x\right]}\left(\nu\left(-\left(x-\eta\right),x-\eta\right]+\frac{\alpha}{\eta}\cdot2\left(z-\left(x-\eta\right)\right)\right)^{2}\frac{\alpha}{\eta}\textup{d}z,
\end{align*}
where we used property 2. We set $y:=\nu\left(-\left(x-\eta\right),x-\eta\right]$.
By a change of variables $u:=y+\frac{2\alpha}{\eta}\left(z-\left(x-\eta\right)\right)$,
we obtain
\begin{align*}
B_{\pi_{\varepsilon}} & =\int_{y}^{y+2\alpha}u^{2}\cdot\frac{1}{2}\textup{d}z=\frac{1}{6}\left(\left(y+2\alpha\right)^{3}-y^{3}\right).
\end{align*}
The inequality $\nu\left(-x,x\right)-\nu\left(-\left(x-\eta\right),x-\eta\right]>-\varepsilon$
holds. We calculate bounds for $B_{\pi_{\varepsilon}}$ in terms of
$B_{\nu}$:
\begin{align*}
B_{\pi_{\varepsilon}}-B_{\nu} & =\frac{1}{6}\left(6\alpha y^{2}+12\alpha^{2}y+8\alpha^{3}\right)-T-\left(\left(\nu\left(-x,x\right)\right)^{2}+2\alpha\nu\left(-x,x\right)+\alpha^{2}\right)^{2}\alpha\\
& =\alpha\left(y^{2}-\left(\nu\left(-x,x\right)\right)^{2}\right)+2\alpha^{2}\left(y-\nu\left(-x,x\right)\right)+\frac{1}{3}\alpha^{3}-T,
\end{align*}
so we obtain the bounds
\[
-\alpha\cdot2\varepsilon-2\alpha^{2}\varepsilon-\varepsilon<B_{\pi_{\varepsilon}}-\left(B_{\nu}+\frac{1}{3}\alpha^{3}\right)<\alpha\cdot2\varepsilon+2\alpha^{2}\varepsilon+\varepsilon,
\]
and hence
\[
-5\varepsilon<B_{\pi_{\varepsilon}}-\left(B_{\nu}+\frac{1}{3}\alpha^{3}\right)<5\varepsilon.
\]
\uline{\mbox{$C$}:} Since $\nu\left|\left(\left[-1/2,-x\right)\cup\left(x,1/2\right]\right)\right.$
is equal to $\pi_{\varepsilon}\left|\left(\left[-1/2,-x\right)\cup\left(x,1/2\right]\right)\right.$,
we have
\[
0\geq C_{\pi_{\varepsilon}}-C_{\nu}>-2\varepsilon\int_{\left(x,1/2\right]}\nu\left(\textup{d}z\right)\geq-2\varepsilon.
\]
In the second step above, we used that, for all $x,y\in\left[0,1\right]$,
$\left|x^{2}-y^{2}\right|<2\left|x-y\right|$ is satisfied.
Putting together the three parts, we obtain the lower bound for
\begin{align*}
a\left(\pi_{\varepsilon}\right)-a\left(\nu\right) & =2\left(A_{\pi_{\varepsilon}}-A_{\nu}+B_{\pi_{\varepsilon}}-B_{\nu}+C_{\pi_{\varepsilon}}-C_{\nu}\right)=2\left(B_{\pi_{\varepsilon}}-B_{\nu}\right)+2\left(C_{\pi_{\varepsilon}}-C_{\nu}\right)\\
& >2\left(\frac{1}{3}\alpha^{3}-5\varepsilon\right)-2\varepsilon=\frac{2}{3}\alpha^{3}-12\varepsilon.
\end{align*}
Similarly, the upper bound is
\begin{align*}
a\left(\pi_{\varepsilon}\right)-a\left(\nu\right) & <2\left(\frac{1}{3}\alpha^{3}+5\varepsilon\right)=\frac{2}{3}\alpha^{3}+10\varepsilon.
\end{align*}
If we let $\varepsilon$ go to 0, we see that $a\left(\pi_{\varepsilon}\right)-a\left(\nu\right)$
goes to $2/3\alpha^{3}$. Hence removing a pair of points from the
discrete measure $\delta_{n}$ and adding the probability mass to
the continuous measure $\gamma$ increases $a$ as claimed.
We have shown that for any probability measure $\mu$ $a\left(\mu\right)\leq1/3$
holds. Since $r\geq0$ holds as can be seen from Lemma \ref{lem:mu_rho_a_r},
the term
\[
r-am=r-a\cdot2r
\]
is non-negative if and only if $r=0$ or $a\leq1/2$. The latter inequality
we have proved holds for all probability measures $\mu$. Corollary
\ref{cor:a_r_0} says that $r=0$ if and only if $\mu=\delta_{0}$.
For all other measures $\mu$, the optimal weights will be composed
of a constant and a proportional part.
\subsection*{Proof of Proposition \ref{prop:FOSD_suff}}
We use the well known characterisation of FOSD in terms of increasing
functions (usually referred to as utility functions in the context
of consumer theory in microeconomics):
\begin{lem}
\label{lem:char_FOSD}We have $\left|Z\right|\succ\left|Y\right|$
if and only if for all increasing functions $u:\left[0,1/2\right]\rightarrow\mathbb{R}$
the inequality $E_{\mu}u\geq E_{\rho}u$ holds.
\end{lem}
We employ the previous lemma to show
\begin{lem}
These two statements hold:
\begin{enumerate}
\item If for all $z\in\left(0,1/2\right]$ $\mu\left[-z,z\right]\leq\rho\left[-z,z\right]$,
then, for all $z\in\left[0,1/2\right]$, $\mu\left(-z,z\right)\leq\rho\left(-z,z\right)$.
\item If for all $z\in\left(0,1/2\right]$ $\mu\left[-z,z\right]\leq\rho\left[-z,z\right]$,
then, for all $z\in\left[0,1/2\right]$, $\mu\left(-z,z\right]\leq\rho\left(-z,z\right]$.
\end{enumerate}
\end{lem}
\begin{proof}
Let $z\in\left(0,1/2\right]$. Then, for all $t<z$, $\mu\left[-t,t\right]\leq\rho\left[-t,t\right]\leq\rho\left(-z,z\right)$.
By letting $t\nearrow z$, we obtain $\mu\left(-z,z\right)\leq\rho\left(-z,z\right)$
due to the continuity of the measure $\mu$, and we have proved the
first assertion. Next we show the second assertion:
\begin{align*}
\mu\left(-z,z\right] & =\mu\left(-z,0\right)+\mu\left\{ 0\right\} +\mu\left(0,z\right]\\
& =\frac{\mu\left(-z,0\right)+\mu\left\{ 0\right\} +\mu\left(0,z\right)}{2}+\frac{\mu\left[-z,0\right)+\mu\left\{ 0\right\} +\mu\left(0,z\right]}{2}\\
& =\frac{\mu\left(-z,z\right)}{2}+\frac{\mu\left[-z,z\right]}{2}\leq\frac{\rho\left(-z,z\right)}{2}+\frac{\rho\left[-z,z\right]}{2}=\rho\left(-z,z\right].
\end{align*}
We used the symmetry of $\mu$ and $\rho$ in steps 2 and 5 above
and the first assertion of the lemma in step 4.
\end{proof}
Now we calculate
\begin{align*}
r & =2\int_{\left(0,1/2\right]}z\rho\left(-z,z\right]\mu\left(\textup{d}z\right)\geq2\int_{\left(0,1/2\right]}y\rho\left(-y,y\right]\rho\left(\textup{d}y\right)\geq2\int_{\left(0,1/2\right]}y\mu\left(-y,y\right]\rho\left(\textup{d}y\right)=s.
\end{align*}
The first inequality is due to Lemma \ref{lem:char_FOSD}: The function
$z\mapsto u\left(z\right):=z\rho\left(-z,z\right]$ is increasing
and hence $E_{\mu}u\geq E_{\rho}u$ holds. The second inequality holds
by the definition of $\left|Z\right|\succ\left|Y\right|$. Therefore,
\[
am=a\left(r+s\right)\leq2ar,
\]
and $a\leq1/2$ is sufficient for $r-am\geq0$.
We turn the second statement of Proposition \ref{prop:FOSD_suff}.
Assume $\left|Y\right|\succ\left|Z\right|$. As we know from Theorem
\ref{thm:mu_rho},
\begin{align*}
a & =2\int_{\left(0,1/2\right]}\left(\rho\left(-z,z\right]\right)^{2}\mu\left(\textup{d}z\right)\leq2\int_{\left(0,1/2\right]}\left(\mu\left(-z,z\right]\right)^{2}\mu\left(\textup{d}z\right)\leq1/3.
\end{align*}
So a sufficient condition for $r-am\geq0$ is $m\leq3r$, which is
equivalent to $s\leq2r$.
\subsection*{Proof of Proposition \ref{prop:ribbon}}
The inequality $r-am\geq0$ we want to show is equivalent to $r\left(1-a\right)\geq as$.
The left hand side of this has a lower bound
\[
r\left(1-a\right)\geq\left(1-a\right)\cdot2\int_{\left(0,1/2\right]}z\rho\left(-z,z\right]c\rho\left(\textup{d}z\right),
\]
whereas the right hand side is bounded above by
\[
as\leq a\cdot2\int_{\left(0,1/2\right]}yC\rho\left(-y,y\right]\rho\left(\textup{d}y\right).
\]
So $c\left(1-a\right)\geq Ca$ is sufficient. This is itself equivalent
to
\begin{equation}
a\leq\frac{c}{c+C}.\label{eq:a_upper_bound_ribbon}
\end{equation}
We find two upper bounds for $a$:
\begin{align}
a & \leq C\cdot2\int_{\left(0,1/2\right]}\left(\rho\left(-z,z\right]\right)^{2}\rho\left(\textup{d}z\right),\label{eq:ribbon_1}\\
a & \leq\frac{1}{c^{2}}\cdot2\int_{\left(0,1/2\right]}\left(\mu\left(-z,z\right]\right)^{2}\mu\left(\textup{d}z\right).\label{eq:ribbon_2}
\end{align}
Theorem \ref{thm:mu_rho} says that both integrals on the right hand
side are bounded above by $1/6$. By stating inequalities of the right
hand side of (\ref{eq:a_upper_bound_ribbon}) and the right hand sides
of (\ref{eq:ribbon_1}) and (\ref{eq:ribbon_2}), respectively, we
obtain the sufficient conditions stated in Proposition \ref{prop:ribbon}:
\begin{align*}
\frac{C}{3}\leq\frac{c}{c+C} & \iff c\geq\frac{C^{2}}{3-C},\\
\frac{1}{3c^{2}}\leq\frac{c}{c+C} & \iff C\leq c\left(3c^{2}-1\right).
\end{align*}
The last claim follows from substituting $c=1/C$ into (\ref{eq:a_upper_bound_ribbon}).
\subsection*{Proof of Proposition \ref{prop:rho_unif_mu_any}}
We first refine Lemma \ref{lem:mu_neq_rho_a_r_s} using that $\rho=\mathcal{U}\left[-1/2,1/2\right]$:
\begin{align*}
a & =4E\left(Z^{2}\right),\quad r=2E\left(Z^{2}\right),\quad s=2\int_{0}^{1/2}y\mu\left(-y,y\right]\textup{d}y.
\end{align*}
So the inequality $r-am\geq0$ is equivalent to
\begin{align}
T\left(\mu\right):=\int_{0}^{1/2}y\mu\left(-y,y\right]\textup{d}y+E\left(Z^{2}\right)\leq\frac{1}{4}.\label{T_def}
\end{align}
The mapping $T:\mathcal{M}_{<\infty}\left(\left[-1/2,1/2\right]\right)\rightarrow\mathbb{R}$
from the set of all finite measures on $\left[-1/2,1/2\right]$ is
linear.
\begin{lem}
\label{lem:T_linear}For all $\nu_{1},\nu_{2}\in\mathcal{M}_{<\infty}\left(\left[-1/2,1/2\right]\right)$
and all $\alpha_{1},\alpha_{2}\in\mathbb{R}$, we have
\[
T\left(\alpha_{1}\nu_{1}+\alpha_{2}\nu_{2}\right)=\alpha_{1}T\left(\nu_{1}\right)+\alpha_{2}T\left(\nu_{2}\right).
\]
\end{lem}
This can be easily verified.
Our strategy to prove Proposition \ref{prop:rho_unif_mu_any} is to
show the result for discrete measures with finite support, and then
use the fact that discrete measures are a dense subset of $\mathcal{M}_{1}\left(\left[-1/2,1/2\right]\right)$.
The proof for discrete measures proceeds by induction on the size
of $\left|\textup{supp }\mu\right|\leq2k+1$.
\uline{Base case:} Let $k=1$. Then the support of $\mu$ consists
of at most three points: 0 and two points $-x_{1},x_{1}$ such that
$0<x_{1}\leq1/2$. The measure $\mu$ is given by $\beta_{0}\delta_{0}+\beta_{1}\left(\delta_{-x_{1}}+\delta_{x_{1}}\right)$
and the constants satisfy $\beta_{0}+2\beta_{1}=1$. Set $\beta:=\beta_{0}$.
We calculate
\begin{align*}
T\left(\mu\right) & =\int_{0}^{1/2}y\mu\left(-y,y\right]\textup{d}y+E\left(Z^{2}\right)=\frac{1}{8}+\frac{x_{1}^{2}}{2}\left(1-\beta\right).
\end{align*}
We see that we can choose the parameter $\beta$ and the point $x_{1}$
independently of each other to maximise $T\left(\mu\right)$. This
maximum is $1/4$ and it is achieved if and only if $\beta=0$ and
$x_{1}=1/2$. This shows the claim for $\left|\textup{supp }\mu\right|\leq2\cdot1+1$.
\uline{Induction step:} Assume for all $\mu\in\mathcal{M}_{1}\left(\left[-1/2,1/2\right]\right),\left|\textup{supp }\mu\right|\leq2k+1$,
the inequality $T\left(\mu\right)\leq1/4$ holds and equality is achieved
if and only if $\mu=1/2\left(\delta_{-1/2}+\delta_{1/2}\right)$.
We show that the claim also holds for all $\nu\in\mathcal{M}_{1}\left(\left[-1/2,1/2\right]\right)$
with $\left|\textup{supp }\nu\right|\leq2\left(k+1\right)+1$. Let
$0<x_{1}<\cdots<x_{k+1}\leq1/2$ be the points of the support of $\nu$.
Then $\nu$ must have the form
\[
\nu=\left(1-\eta\right)\mu+\eta\frac{1}{2}\left(\delta_{-x_{k+1}}+\delta_{x_{k+1}}\right)
\]
for some $0\leq\eta\leq1$. By Lemma \ref{lem:T_linear},
\begin{align*}
T\left(\nu\right) & =\left(1-\eta\right)T\left(\mu\right)+\eta T\left(\frac{1}{2}\left(\delta_{-x_{k+1}}+\delta_{x_{k+1}}\right)\right)\leq\left(1-\eta\right)\frac{1}{4}+\eta\frac{1}{4}=\frac{1}{4}.
\end{align*}
Furthermore, as $\left|\textup{supp }\mu\right|$ and $\left|\textup{supp }\frac{1}{2}\left(\delta_{-x_{k+1}}+\delta_{x_{k+1}}\right)\right|$
are at most $2k+1$ and $1/2\notin\textup{supp }\mu$, equality holds
if and only if $x_{k+1}=1/2$ and $\eta=1$. Hence, the second part
of the claim holds for $\left|\textup{supp }\nu\right|\leq2\left(k+1\right)+1$,
too.
A well known result concerning probability measures is
\begin{thm}
Let $X$ be a separable metric space. Then the set of discrete probability
measures on $X$ is dense in $\mathcal{M}_{1}\left(X\right)$ if we
consider $\mathcal{M}_{1}\left(X\right)$ as a space endowed with
the topology of weak convergence.
\end{thm}
See e.g.$\!\,$ Theorem 6.3 on page 44 in \cite{Pa}. Note that we can even
choose the subset of discrete probability measures with finite support
as a dense subset of $\mathcal{M}_{1}\left(X\right)$. We will now
show that the mapping $T$ is continuous. Let $\left(\mu_{n}\right)$
be a sequence in $\mathcal{M}_{1}\left(\left[-1/2,1/2\right]\right)$
with the limit $\mu\in\mathcal{M}_{1}\left(\left[-1/2,1/2\right]\right)$,
i.e. $\mu_{n}\xrightarrow[n\rightarrow\infty]{w}\mu$. We show that
both summands in the definition \eqref{T_def} of $T\left(\mu_n\right)$ converge.
The sequence of functions $y\mapsto y\mu_{n}\left(-y,y\right]$ is
uniformly bounded in $n$. $\mu_{n}\xrightarrow[n\rightarrow\infty]{w}\mu$
is equivalent to the convergence of the distribution functions. Let
$F_{n}$ be the distribution of $\mu_{n}$ for each $n$ and $F$
the distribution function of $\mu$. Then $F_{n}$ converges to $F$
pointwise on the set $C$ of continuity points of $F$. As $F$ is
monotonic, the complement $C^{c}$ is at most countable and hence
a Lebesgue null set. This means $y\mapsto y\mu_{n}\left(-y,y\right]$
converges almost everywhere on $\left(0,1/2\right]$. By dominated
convergence, the integrals $\int_{0}^{1/2}y\mu_{n}\left(-y,y\right]\textup{d}y$
converge to $\int_{0}^{1/2}y\mu\left(-y,y\right]\textup{d}y$.
The function $z\mapsto z^{2}$ is continuous and bounded on $\left(0,1/2\right]$.
Hence, $\mu_{n}\xrightarrow[n\rightarrow\infty]{w}\mu$ by definition
implies the convergence of $2\int_{\left(0,1/2\right]}z^{2}\mu_{n}\left(\textup{d}z\right)$
to $2\int_{\left(0,1/2\right]}z^{2}\mu\left(\textup{d}z\right)$.
We have previously shown that for all measures $\mu$ with finite
support $T\left(\mu\right)\leq1/4$. If $\mu$ is now any probability
measure, then there is a sequence of finitely supported measures $\mu_{n}$
that converge to $\mu$. As $T$ is continuous, this implies $T\left(\mu\right)\leq1/4$,
and the claim has been proved.
\subsection*{Proof of Theorem \ref{thm:fam_rho_mu}}
We first prove the four claims in Lemma \ref{lem:g_F_properties}
only assuming property 2 in Assumptions \ref{def:mu_contraction}.
\uline{Claim 1:} For all $y\in\left[0,1\right]$ $g\left(y\right)=2F_{\rho}\left(y/2\right)$.
Let $y\in\left[0,1\right]$. Since $F_{\rho}\left(\frac{1}{2}\right)=\frac{1}{2}$,
\[
\frac{g\left(y\right)}{2}=g\left(y\right)F_{\rho}\left(\frac{1}{2}\right)=F_{\rho}\left(\frac{y}{2}\right).
\]
\uline{Claim 2:} $\rho$ has no atoms, unless $\rho=\delta_{0}$.
We will write $f\left(x+\right)$ for the right limit $\lim_{t\searrow x}f\left(t\right)$
and $f\left(x-\right)$ for the left limit $\lim_{t\nearrow x}f\left(t\right)$
of any function $f$ and any $x\in\mathbb{R}$. Suppose $x>0$ is
an atom of $\rho$: $\rho\left\{ x\right\} >0$. Then $F_{\rho}\left(x-\right)<F_{\rho}\left(x\right)$.
Hence, for all $c<1$ $F_{\rho}\left(cx\right)=g\left(c\right)F_{\rho}\left(x\right)$.
Letting $c\nearrow1$, we get $F_{\rho}\left(x-\right)=g\left(1-\right)F_{\rho}\left(x\right)$,
so $0<g\left(1-\right)<1$. Thus we have, for all $y>0$, $F_{\rho}\left(y-\right)=g\left(1-\right)F_{\rho}\left(y\right)$
and $F_{\rho}\left(y-\right)<F_{\rho}\left(y\right)$, and $y$ is
an atom. This is a contradiction, since $\rho$ cannot have uncountably
many atoms. Therefore, $x>0$ cannot be an atom of $\rho$ and the
only possible atom is $0$. We next show that if $\rho\left\{ 0\right\} >0$,
then $\rho\left\{ 0\right\} =1$.
Suppose $0<\eta<1$ and $\rho\left|\left[0,1/2\right]\right.=\eta\delta_{0}+\frac{1-\eta}{2}\nu$,
where $\nu\in\mathcal{M}_{\leq1}\left(\left[0,1/2\right]\right)$
has no atoms. As $\nu$ has no atoms,
\[
\lim_{c\searrow0}F_{\rho}\left(cy\right)=F_{\rho}\left(0\right)
\]
holds for all $y\geq0$. On the other hand,
\[
\lim_{c\searrow0}g\left(c\right)F_{\rho}\left(y\right)=g\left(0+\right)F_{\rho}\left(y\right).
\]
Suppose $p:=g\left(0+\right)>0$. Fix some $b\in\left(0,1\right)$ such that $0<F_{\rho}\left(b\right)<\frac{1}{2}$.
Then $p\left(F_{\rho}\left(\frac{1}{2}\right)-F_{\rho}\left(b\right)\right)>0$ and
\[
\lim_{c\searrow0}F_{\rho}\left(c\cdot\frac{1}{2}\right)=\lim_{c\searrow0}F_{\rho}\left(bc\right)=F_{\rho}\left(0\right)
\]
due to the right continuity of the distribution function $F_{\rho}$.
This implies $F_{\rho}\left(\frac{c}{2}\right)-F_{\rho}\left(bc\right)\rightarrow0$
as $c\searrow0$. But $g\left(c\right)\left(F_{\rho}\left(\frac{1}{2}\right)-F_{\rho}\left(b\right)\right)\rightarrow p>0$.
This is a contradiction and $g\left(0+\right)>0$ must be false. By
the first statement of this lemma, $F_{\rho}\left(0\right)=1/2\,g\left(0\right)\leq1/2\,g\left(0+\right)=0$.
The inequality is due to $g$ being increasing.
\uline{Claim 3:} $g$ is multiplicative: for all $x,y\geq0$, $g\left(xy\right)=g\left(x\right)g\left(y\right)$.
Let $x,y\geq0$. Then we have
\[
\frac{g\left(xy\right)}{2}=g\left(xy\right)F_{\rho}\left(\frac{1}{2}\right)=F_{\rho}\left(\frac{xy}{2}\right)=g\left(x\right)g\left(y\right)F_{\rho}\left(\frac{1}{2}\right)=\frac{g\left(x\right)g\left(y\right)}{2}.
\]
\uline{Claim 4:} $F_{\rho}$ has the form $F_{\rho}\left(y\right)=2^{t-1}y^{t}$
for some fixed $t\geq0$.
Since $g$ is multiplicative by statement 3, the transformation $x\mapsto f\left(x\right):=\ln\left(g\left(e^{x}\right)\right)$
is additive. Due to statement 1, $g$ is increasing. Hence, we can
apply the Cauchy functional condition to conclude that $f$ is linear,
i.e.$\!\,$ there is some $t\in\mathbb{R}$ such that $f\left(x\right)=tx$
for all $x\geq0$. As $f$ is increasing, $t$ must be non-negative.
So
\begin{align*}
\ln\left(g\left(e^{x}\right)\right) & =tx\iff g\left(e^{x}\right)=e^{tx}
\end{align*}
and
\[
g\left(x\right)=g\left(e^{\ln x}\right)=\left(e^{\ln x}\right)^{t}=x^{t}.
\]
Using statement 1, we obtain, for all $y\in\left[0,1/2\right]$,
\[
F_{\rho}\left(y\right)=\frac{g\left(2y\right)}{2}=\frac{\left(2y\right)^{t}}{2}=2^{t-1}y^{t}.
\]
From now on, we assume all three properties in Assumptions \ref{def:mu_contraction}
and show Theorem \ref{thm:fam_rho_mu}. First, we note that the homogeneity
property 2 of the measure $\rho$ is inherited by $\mu$:
\begin{lem}
For all $x\geq0$ and all $y\in\left[0,1/2\right]$ such that $xy\leq1/2$,
we have $\mu\left(0,xy\right)=g\left(x\right)\mu\left(0,y\right)$.
\end{lem}
The proof is straightforward and we thus omit it.
We next calculate an inequality equivalent to $r-am\geq0$:
\begin{align*}
r\left(1-a\right)\geq as & \quad\iff\\
2\int_{\left(0,1/2\right]}z\rho\left(-z,z\right]\mu\left(\textup{d}z\right)\left[1-2\int_{\left(0,1/2\right]}\left(\rho\left(-z,z\right]\right)^{2}\mu\left(\textup{d}z\right)\right]\geq\\
2\int_{\left(0,1/2\right]}\left(\rho\left(-z,z\right]\right)^{2}\mu\left(\textup{d}z\right)\cdot2\int_{\left(0,1/2\right]}y\mu\left(-y,y\right]\rho\left(\textup{d}y\right) & \quad\iff\\
\int_{\left(0,c/2\right)}z\rho\left(0,z\right)\mu\left(\textup{d}z\right)\left[1-8\int_{\left(0,1/2\right)}\left(\mu\left(0,cz\right)\right)^{2}\mu\left(\textup{d}z\right)\right]\geq\\
8\int_{\left(0,1/2\right)}\left(\mu\left(0,cz\right)\right)^{2}\mu\left(\textup{d}z\right)\left[\int_{\left(0,c/2\right)}y\mu\left(0,y\right)\rho\left(\textup{d}y\right)+\frac{1}{2}\int_{\left(c/2,1/2\right)}y\rho\left(\textup{d}y\right)\right],
\end{align*}
where we used the symmetry of $\rho$ and $\mu$, the fact that $\rho$
-- and hence $\mu$ -- has no atoms, and $\mu\left(0,cz\right)=\rho\left(0,z\right)$.
The left hand side of the last inequality above can be expressed as
\[
\int_{\left(0,c/2\right)}z\rho\left(0,z\right)\mu\left(\textup{d}z\right)\left[1-\frac{g\left(c\right)^{2}}{3}\right],
\]
where we applied Theorem \ref{thm:mu_rho}. The right hand side can
be treated similarly:
\[
\frac{g\left(c\right)^{2}}{3}\left[\int_{\left(0,c/2\right)}y\mu\left(0,y\right)\rho\left(\textup{d}y\right)+\frac{1}{2}\int_{\left(c/2,1/2\right)}y\rho\left(\textup{d}y\right)\right].
\]
We note that the inequality
\begin{align}
\left[3-g\left(c\right)^{2}\right]\int_{\left(0,c/2\right)}z\rho\left(0,z\right)\mu\left(\textup{d}z\right) & \geq g\left(c\right)^{2}\left[\int_{\left(0,c/2\right)}y\mu\left(0,y\right)\rho\left(\textup{d}y\right)+\frac{1}{2}\int_{\left(c/2,1/2\right)}y\rho\left(\textup{d}y\right)\right]\label{eq:r_am_equiv_1}
\end{align}
is thus equivalent to our original inequality. Now we show
\begin{lem}
We can switch the measures in the integrals as follows:
\[
\int_{\left(0,c/2\right)}z\rho\left(0,z\right)\mu\left(\textup{d}z\right)=\int_{\left(0,c/2\right)}y\mu\left(0,y\right)\rho\left(\textup{d}y\right).
\]
\end{lem}
\begin{proof}
The proof uses the Lebesgue-Stieltjes versions of the integrals. Let
$F_{\mu}$ be the distribution function of $\mu\left|\left[0,1/2\right]\right.$.
We have
\begin{align*}
\int_{\left(0,c/2\right)}z\rho\left(0,z\right)\mu\left(\textup{d}z\right) & =\int_{\left(0,c/2\right)}z\mu\left(0,cz\right)\mu\left(\textup{d}z\right)=g\left(c\right)\int_{\left(0,c/2\right)}z\mu\left(0,z\right)\textup{d}F_{\mu}\left(z\right)\\
& =\int_{\left(0,c/2\right)}z\mu\left(0,z\right)\textup{d}\left(g\left(c\right)F_{\mu}\left(z\right)\right)=\int_{\left(0,c/2\right)}z\mu\left(0,z\right)\textup{d}F_{\mu}\left(cz\right)\\
& =\int_{\left(0,c/2\right)}z\mu\left(0,z\right)\textup{d}F_{\rho}\left(z\right)=\int_{\left(0,c/2\right)}y\mu\left(0,y\right)\rho\left(\textup{d}y\right)
\end{align*}
by a substitution formula (see e.g. \cite{FaTe}).
\end{proof}
Using this lemma, we can restate (\ref{eq:r_am_equiv_1}) as
\begin{equation}
\left[3-2g\left(c\right)^{2}\right]\int_{\left(0,c/2\right)}z\rho\left(0,z\right)\mu\left(\textup{d}z\right)\geq\frac{g\left(c\right)^{2}}{2}\int_{\left(c/2,1/2\right)}y\rho\left(\textup{d}y\right).\label{eq:r_am_equiv_2}
\end{equation}
Next we prove
\begin{lem}
The following equality holds:
\[
\frac{1}{g\left(c\right)}\int_{\left(0,c/2\right)}z\rho\left(0,z\right)\mu\left(\textup{d}z\right)=c\int_{\left(0,1/2\right)}y\rho\left(0,y\right)\rho\left(\textup{d}y\right).
\]
\end{lem}
\begin{proof}
We calculate
\begin{align*}
\int_{\left(0,1/2\right)}cz\rho\left(0,z\right)\rho\left(\textup{d}z\right) & =\int_{\left(0,1/2\right)}\frac{1}{g\left(c\right)}g\left(c\right)cy\rho\left(0,y\right)\textup{d}F_{\rho}\left(y\right)=\frac{1}{g\left(c\right)}\int_{\left(0,c/2\right)}g\left(c\right)x\rho\left(0,\frac{x}{c}\right)\textup{d}F_{\rho}\left(\frac{x}{c}\right)\\
& =\frac{1}{g\left(c\right)}\int_{\left(0,c/2\right)}x\rho\left(0,x\right)\textup{d}F_{\mu}\left(x\right).
\end{align*}
\end{proof}
Together the last lemma and statement 4 of Lemma \ref{lem:g_F_properties}
imply the inequality (\ref{eq:r_am_equiv_2}) is equivalent to
\[
c^{1-t}\left[3-2c^{2t}\right]\geq\frac{1+2t}{1+t}\left(1-c^{1+t}\right).
\]
We define the function $h$ by setting
\[
h\left(c\right):=c^{1+t}-3\left(1+t\right)c^{1-t}+1+2t.
\]
The original inequality holds if and only if $h\left(c\right)\leq0$.
We calculate the first derivative of $h$,
\[
h'\left(c\right)=\left(1+t\right)c^{t}-3\left(1+t\right)\left(1-t\right)c^{-t}
\]
and the critical point is given by
\begin{equation}
c=x_{0}:=\left(3\left(1-t\right)\right)^{\frac{1}{2t}}\label{eq:x_0},
\end{equation}
which is positive if $t<1$. The second derivative of $h$ is
\[
h''\left(c\right)=\left(1+t\right)tc^{t-1}+3\left(1+t\right)\left(1-t\right)tc^{-\left(1+t\right)}.
\]
The sign of the second derivative is positive for all $c>0$. We also
note that $h\left(0\right)>0$ and $h\left(1\right)<0$. The positive
sign of $h''$ on $\left(0,\infty\right)$ implies that $h'$ is strictly
increasing on $\left[0,\infty\right)$. Furthermore, $h$ is strictly
decreasing on $\left[0,x_{0}\right)$ and strictly increasing on $\left[x_{0},\infty\right)$.
It is clear from (\ref{eq:x_0}) that $x_{0}\geq1$ if and only if
$t\leq2/3$. When this holds, $h$ is strictly decreasing on $\left[0,1\right]$.
For $2/3<t<1$, $h$ is first decreasing and then increasing. However,
as $h\left(1\right)<0$, we have for all $t<1$ a uniquely determined
$c_{0}\in\left(0,x_{0}\wedge1\right)$ such that $h\left(c_{0}\right)=0$
and $c_{0}$ is the only zero of $h$ on the interval $\left[0,1\right]$.
For $t=1$, the claim follows from Proposition \ref{prop:rho_unif_mu_any}.
For $t>1$, we note that $h$ is undefined at $c=0$ but $\lim_{c\searrow0}h\left(c\right)=-\infty$.
As $h\left(1\right)<0$ holds for any value of $t$ and $h$ is continuous,
$h<0$ on $\left(0,1\right)$ is clear. This shows the claim concerning
the sign of $r-am$.
As for the behaviour of the critical $c_{0}$ at which $r=am$, by
inspecting (\ref{eq:x_0}), we see that $\lim_{t\nearrow1}x_{0}=0$
and as $0<c_{0}<x_{0}$, the second claim follows.
|
\subsection*{Acknowledgements}}
\title{Online Policy Gradient for Model Free Learning \\ of Linear Quadratic Regulators with $\sqrt{T}$ Regret}
\author{%
Asaf Cassel%
\thanks{School of Computer Science, Tel Aviv University; \texttt{<EMAIL>}.}
\and
Tomer Koren%
\thanks{School of Computer Science, Tel Aviv University, and Google Research, Tel Aviv; \texttt{<EMAIL>}.}
}
\newcommand{\AC}[1]{\textcolor{cyan}{\bfseries\{AC : #1\}}}
\newcommand{\TK}[1]{\textcolor{magenta}{\bfseries\{TK : #1\}}}
\RequirePackage{mathtools}
\RequirePackage{dsfont}
\RequirePackage{delimset}
\newcommand{\innProd}[2][*]{\brk[a]#1{#2}}
\newcommand{\ceil}[2][*]{\delim\lceil\rceil#1{#2}}
\newcommand{\floor}[2][*]{\delim\lfloor\rfloor#1{#2}}
\newcommand{\indEvent}[2][*]{\mathds{1}{\brk[c]#1{#2}}}
\newcommand{\indFunc}[2][\infty]{\mathds{I}_{\brk[s]{#2,#1}}}
\newcommand{\indFuncAt}[3][\infty]{\indFunc[#1]{#2}\brk*{#3}}
\newcommand{\tr}[2][*]{\mathrm{Tr}\brk*{#2}}
\newcommand{\overbar}[1]{\mkern 2mu\overline{\mkern-2mu#1\mkern-2mu}\mkern 2mu}
\newcommand{^{\mkern-1.5mu\mathsf{T}}}{^{\mkern-1.5mu\mathsf{T}}}
\newcommand{\Oof}[2][*]{O\brk#1{#2}}
\newcommand{\OtildeOf}[2][*]{\smash{\widetilde{O}}\brk#1{#2}}
\newcommand{\EE}[1][]{\mathbb{E}_{#1}}
\newcommand{\EEBrk}[2][*]{\EE\delim{[}{]}#1{#2}}
\newcommand{\RR}[1][]{\mathbb{R}^{#1}}
\newcommand{\ZZ}[1][]{\mathbb{Z}^{#1}}
\newcommand{\ZZplus}[1][]{\mathbb{Z}^{#1}_{+}}
\newcommand{\PP}[2][*]{\mathbb{P}\brk#1{#2}}
\newcommand{\sphere}[1]{\mathcal{S}^{#1}}
\newcommand{\ball}[1]{\mathcal{B}^{#1}}
\newcommand{\gaussDist}[2]{\mathcal{N}\brk0{#1, #2}}
\DeclareMathOperator*{\argmax}{arg\,max}
\DeclareMathOperator*{\argmin}{arg\,min}
\DeclareMathOperator*{\essinf}{ess\,inf}
\DeclareMathOperator*{\esssup}{ess\,sup}
\DeclarePairedDelimiterX\setDef[1]\lbrace\rbrace{\def\;\delimsize\vert\;}#1{\;\delimsize\vert\;}#1}
\newcommand{\seqDef}[3]{\brk[c]*{#1}_{#2}^{#3}}
\def\noindent{\noindent}
\newcommand{\ifrac}[2]{#1/#2}
\newcommand{\minEigOf}[2][*]{\lambda_{\mathrm{min}}\brk#1{#2}}
\newcommand{\maxEigOf}[2][*]{\lambda_{\mathrm{max}}\brk#1{#2}} %
\usepackage{amsthm}
\usepackage{amssymb}
\usepackage{thmtools}
\declaretheoremstyle[
spaceabove=\topsep,
spacebelow=\topsep,
bodyfont=\normalfont\itshape,
]{theorem}
\declaretheorem[style=theorem,name=Theorem]{theorem}
\declaretheoremstyle[
spaceabove=\topsep,
spacebelow=\topsep,
bodyfont=\normalfont,
]{definition}
\let\proof\undefined
\declaretheoremstyle[
spaceabove=\topsep,
spacebelow=\topsep,
%
bodyfont=\normalfont,
notefont=\normalfont\bfseries,
%
notebraces={}{},
%
qed=$\blacksquare$,
%
%
]{proofstyle}
\declaretheorem[style=proofstyle,numbered=no,name=Proof]{proof}
\declaretheorem[style=theorem,name=Lemma]{lemma}
\declaretheorem[style=theorem,name=Corollary]{corollary}
\declaretheorem[style=theorem,name=Claim]{claim}
\declaretheorem[style=theorem,name=Proposition]{proposition}
\declaretheorem[style=theorem,name=Fact]{fact}
\declaretheorem[style=theorem,name=Observation]{observation}
\declaretheorem[style=theorem,name=Conjecture]{conjecture}
\declaretheorem[style=theorem,numbered=no,name=Theorem]{theorem*}
\declaretheorem[style=theorem,numbered=no,name=Lemma]{lemma*}
\declaretheorem[style=theorem,numbered=no,name=Corollary]{corollary*}
\declaretheorem[style=theorem,numbered=no,name=Proposition]{proposition*}
\declaretheorem[style=theorem,numbered=no,name=Claim]{claim*}
\declaretheorem[style=theorem,numbered=no,name=Fact]{fact*}
\declaretheorem[style=theorem,numbered=no,name=Observation]{observation*}
\declaretheorem[style=theorem,numbered=no,name=Conjecture]{conjecture*}
\declaretheorem[style=definition,name=Definition]{definition}
\declaretheorem[style=definition,name=Remark]{remark}
\declaretheorem[style=definition,name=Example]{example}
\declaretheorem[style=definition,name=Question]{question}
\declaretheorem[style=definition,name=Assumption]{assumption}
\declaretheorem[style=definition,numbered=no,name=Definition]{definition*}
\declaretheorem[style=definition,numbered=no,name=Remark]{remark*}
\declaretheorem[style=definition,numbered=no,name=Example]{example*}
\declaretheorem[style=definition,numbered=no,name=Question]{question*}
\declaretheorem[style=definition,numbered=no,name=Assumption]{assumption*}
%
\newcommand{C}{C}
\newcommand{\infCostOf}[1]{C\brk*{#1}}
\newcommand{\mathcal{X}}{\mathcal{X}}
\newcommand{\bar{f}}{\bar{f}}
\newcommand{f_*}{f_*}
\newcommand{f}{f}
\newcommand{\costOf}[1]{f\brk*{#1}}
\newcommand{\nabla\cost}{\nablaf}
\newcommand{\costGradOf}[1]{\nabla\cost\brk*{#1}}
\newcommand{\gradOracle}[1][t]{\hat{g}_{#1}}
\newcommand{\gradOracleOf}[1]{\gradOracle\brk*{#1}}
\newcommand{\gradCorrpution}[1][t]{w_{#1}}
\newcommand{\beta}{\beta}
\newcommand{\alpha}{\alpha}
\newcommand{\mu}{\mu}
\newcommand{D_{0}}{D_{0}}
\newcommand{\pCorrupt}[1][t]{\varepsilon_{#1}}
\newcommand{\pEffCorrupt}[1][t]{\bar{\varepsilon}_{#1}}
\newcommand{\varepsilon_{max}}{\varepsilon_{max}}
\newcommand{G}{G}
\newcommand{\bar{f}}{\bar{f}}
\newcommand{\alpha_0}{\alpha_0}
\newcommand{\alpha_1}{\alpha_1}
\newcommand{\sigma}{\sigma}
\newcommand{\psi}{\psi}
\newcommand{\nu}{\nu}
\newcommand{\nu_0}{\nu_0}
\newcommand{W}{W}
\newcommand{A_{\star}}{A_{\star}}
\newcommand{B_{\star}}{B_{\star}}
\newcommand{K_{\star}}{K_{\star}}
\newcommand{P_{\star}}{P_{\star}}
\newcommand{J_{\star}}{J_{\star}}
\newcommand{J}{J}
\newcommand{\Jof}[2][*]{J\brk#1{#2}}
\newcommand{\regret}[1][T]{R_{#1}}
\newcommand{\ct}[1][t]{c_{#1}}
\newcommand{\dCt}[1][t]{\Delta C_{#1}}
\newcommand{\dCi}[1][i]{\Delta C_{#1}}
\newcommand{\cji}[1][\ind0,\ind1]{c_{#1}}
\newcommand{\cjis}[1][\ind0,\ind1,\ind2]{c_{#1}}
\newcommand{\cjitau}[1][\ind0,\ind1,\tau]{c_{#1}}
\newcommand{\uji}[1][\ind0,\ind1]{U_{#1}}
\newcommand{\mj}[1][\ind0]{m_{#1}}
\newcommand{\rj}[1][\ind0]{r_{#1}}
\newcommand{\gj}[1][\ind0]{\hat{g}_{#1}}
\newcommand{\Kj}[1][\ind0]{K_{#1}}
\newcommand{\Kji}[1][\ind0,\ind1]{K_{#1}}
\newcommand{n}{n}
\newcommand{\uGji}[1][\ind0,\ind1]{\smash{\widetilde{U}_{#1}}}
\newcommand{\Jr}[1][r]{J^{#1}}
\newcommand{\JrOf}[2][\rj]{\Jr[#1]\brk*{#2}}
\newcommand{d_x}{d_x}
\newcommand{d_u}{d_u}
\newcommand{d_{\text{min}}}{d_{\text{min}}}
\newcommand{\mathcal{K}}{\mathcal{K}}
\newcommand{\steadyCov}[1][K]{\Sigma_{#1}}
\newcommand{P}{P}
\newcommand{\costToGoOf}[1][K]{P_{#1}}
\newcommand{\isUpRound}[1][t]{\chi_{#1}}
\newcommand{\hat{f}}{\hat{f}}
\newcommand{\costOracleOf}[1]{\hat{f}\brk*{#1}}
\newcommand{\gradCorrpution}{\gradCorrpution}
\newcommand{\costNoise}[1][t]{\zeta_{#1}}
\newcommand{\hat{g}_m^r}{\hat{g}_m^r}
\newcommand{\gradEstOf}[1]{\hat{g}_m^r\brk*{#1}}
\newcommand{\smoothCost}[1][r]{f^{#1}}
\newcommand{\smoothCostOf}[2][r]{f^{#1}\brk*{#2}}
\newcommand{\smoothCostGrad}[1][r]{\nablaf^{#1}}
\newcommand{\smoothCostGradOf}[2][r]{\nablaf^{#1}\brk*{#2}}
\newcommand{\epsilon_v}{\epsilon_v}
\newcommand{\sigma_v}{\sigma_v}
\newcommand{\nu}{\nu}
\newcommand{\ind}[1]{\ifcase#1 j \or i \else s\fi}
\newcommand{\ui}[1][\ind1]{U_{#1}}
\newcommand{\rho}{\rho}
\newcommand{\hist}[1][t]{\mathcal{H}_{#1}}
\newcommand{\filt}[1][t]{\mathcal{F}_{#1}}
\newcommand{\cropNoise}[1][t]{\tilde{w}_{#1}}
\newcommand{\tilde{W}}{\tilde{W}}
\newcommand{\tilde{\sigma}}{\tilde{\sigma}}
\newcommand{S}{S}
\newcommand{\cropCost}[1][t]{\tilde{c}_{#1}}
\newcommand{\cropJof}[2][*]{\tilde{J}\brk#1{#2}}
\newcommand{\tilde{J}_\star}{\tilde{J}_\star}
\newcommand{\noiseCov}[1][w]{\Sigma_{#1}}
\usepackage{enumitem}
\usepackage[numbers]{natbib}
\bibliographystyle{abbrvnat}
\usepackage{crossreftools}
\pdfstringdefDisableCommands{%
\let\Cref\crtCref
\let\cref\crtcref
}
\usepackage{algorithm}
\usepackage[noend]{algpseudocode}
\algnewcommand{\IfThenElse}[3]{%
\State \algorithmicif\ #1\ \algorithmicthen\ #2\ \algorithmicelse\ #3}
\begin{document}
\maketitle
\begin{abstract}
We consider the task of learning to control a linear dynamical system under fixed quadratic costs, known as the Linear Quadratic Regulator (LQR) problem. While model-free approaches are often favorable in practice, thus far only model-based methods, which rely on costly system identification, have been shown to achieve regret that scales with the optimal dependence on the time horizon~$T$. We present the first model-free algorithm that achieves similar regret guarantees. Our method relies on an efficient policy gradient scheme, and a novel and tighter analysis of the cost of exploration in policy space in this setting.
\end{abstract}
\section{Introduction}
\label{sec:intro}
Model-free, policy gradient algorithms have become a staple of Reinforcement Learning (RL) with both practical successes~\citep{lillicrap2015continuous,haarnoja2018soft}, and strong theoretical guarantees in several settings~\citep{sutton1999policy,silver2014deterministic}.
In this work we study the design and analysis of such algorithms for the adaptive control of Linear Quadratic Regulator (LQR) systems, as seen through the lens of regret minimization~\citep{abbasi2011regret,cohen2019learning,mania2019certainty}. In this continuous state and action reinforcement learning setting, an agent chooses control actions $u_t$ and the system state $x_t$ evolves according to the noisy linear dynamics
\begin{align*}
x_{t+1}
=
A_{\star} x_t
+
B_{\star} u_t
+
w_t
,
\end{align*}
where $A_{\star}$ and $B_{\star}$ are transition matrices and $w_t$ are i.i.d zero-mean noise terms. The cost is a quadratic function of the current state and action, and the regret is measured with respect to the class of linear policies, which are known to be optimal for this setting.
Model-based methods, which perform planning based on a system identification procedure that estimates the transition matrices, have been studied extensively in recent years.
This started with \citet{abbasi2011regret}, which established an $\Oof[0]{\sqrt{T}}$ regret guarantee albeit with a computationally intractable method. More recently, \citet{cohen2019learning,mania2019certainty} complemented this result with computationally efficient methods, and \citet{cassel2020logarithmic,simchowitz2020naive} provided lower bounds, showing that this rate is generally unavoidable; regardless of whether the algorithm is model free or not.
In comparison, the best existing model-free algorithms are policy iteration procedures by \citet{krauth2019finite} and \citet{abbasi2019model} that respectively achieve $\OtildeOf[0]{T^{2/3}}$ and $\OtildeOf[0]{T^{2/3+\epsilon}}$ regret for $\epsilon = \Theta(1/\log{T})$.
Our main result is an efficient (in fact, linear time per step) policy gradient algorithm that achieves $\OtildeOf[0]{\sqrt{T}}$ regret, thus closing the (theoretical) gap between model based and free methods for the LQR model.
An interesting feature of our approach is that while the policies output by the algorithm are clearly state dependent, the tuning of their parameters requires no such access. Instead, we only rely on observations of the incurred cost, similar to bandit models (e.g., \citealp{cassel2020bandit}).
One of the main challenges of regret minimization in LQRs (and more generally, in reinforcement learning) is that it is generally infeasible to change policies as often as one likes.
Roughly, this is due to a burn-in period following a policy change, during which the system converges to a new steady distribution, and typically incurs an additional cost proportional to the change in steady states, which is in turn proportional to the distance between policies.
There are several ways to overcome this impediment.
The simplest is to restrict the number of policy updates and explore directly in the action space via artificial noise (see e.g., \citealp{simchowitz2020naive}).
Another approach by \citet{cohen2019learning} considers a notion of slowly changing policies, however, these can be very prohibitive for exploration in policy space.
Other works (e.g., \citealp{agarwal2019logarithmic}) consider a policy parameterization that converts the problem into online optimization with memory, which also relies on slowly changing policies. This last method is also inherently model-based and thus not adequate for our purpose.
A key technical contribution that we make is to overcome this challenge by exploring directly in policy space. While the idea itself is not new, we provide a novel and tighter analysis that allows us to use larger perturbations, thus reducing the variance of the resulting gradient estimates. We achieve this by showing that the additional cost depends only quadratically on the exploration radius, which is a crucial ingredient for overcoming the $\Oof[0]{T^{2/3}}$ limitation.
The final ingredient of the analysis involves a sensitivity analysis of the gradient descent procedure that uses the estimated gradients. Here again, while similar analyses of gradient methods exist,
we provide a general result that gives appropriate conditions for which the optimization error depends only quadratically on the error in the gradients.
\paragraph{Related work.}
Policy gradient methods in the context of LQR has seen significant interest in recent years. Notably, \citet{fazel2018global} establish its global convergence in the perfect information setting, and give complexity bounds for sample based methods.
Subsequently, \citet{malik2019derivative} improve the sample efficiency but their result holds only with a fixed probability and thus does not seem applicable for our purposes.
\citet{hambly2020policy} also improve the sample efficiency, but in a finite horizon setting.
\citet{mohammadi2020learning} give sample complexity bounds for the continuous-time variant of LQR.
Finally, \citet{tu2019gap} show that a model based method can potentially outperform the sample complexity of policy gradient by factors of the input and output dimensions. While we observe similar performance gaps in our regret bounds, these were not our main focus and may potentially be improved by a more refined analysis.
Moving away from policy gradients, \citet{yang2019global,jin2020analysis,yaghmaie2019using} analyze the convergence and sample complexity of other model free methods such as policy iteration and temporal difference (TD) learning, but they do not include any regret guarantees.
\section{Preliminaries}
\subsection{Setup: Learning in LQR}
We consider the problem of regret minimization in the LQR model.
At each time step $t$, a state $x_t \in \RR[d_x]$ is observed and action $u_t \in \RR[d_u]$ is chosen.
The system evolves according to
\begin{align*}
x_{t+1}
=
A_{\star} x_t + B_{\star} u_t + w_t
,
\quad
(x_0 = 0 ~ \text{w.l.o.g.})
,
\end{align*}
where the state-state $A_{\star} \in \RR[d_x \times d_x]$ and state-action $B_{\star} \in \RR[d_x \times d_u]$ matrices form the transition model and the $w_t$ are bounded, zero mean, i.i.d.~noise terms with a positive definite covariance matrix $\noiseCov \succ 0$. Formally, there exist $\sigma, W > 0$ such that
\begin{align*}
\EE w_t = 0
\quad
,
\norm{w_t} \le W
\quad
,
\noiseCov
=
\EE w_t w_t^{\mkern-1.5mu\mathsf{T}} \succ \sigma^2 I
.
\end{align*}
The bounded noise assumption is made for simplicity of the analysis, and in \cref{sec:gaussian} we show how to accommodate Gaussian noise via a simple reduction to this setting.
At time $t$, the instantaneous cost is
$$
c_t = x_t^{\mkern-1.5mu\mathsf{T}} Q x_t + u_t^{\mkern-1.5mu\mathsf{T}} R u_t,
$$
where $0 \prec Q,R \preceq I$ are positive definite. We note that the upper bound is without loss of generality since multiplying $Q$ and $R$ by a constant factor only re-scales the regret.
A policy of the learner is a potentially time dependent mapping from past history to an action $u \in \RR[d_u]$ to be taken at the current time step.
Classic results in linear control establish that, given the system parameters $A_{\star},B_{\star},Q$ and $R$, a linear transformation of the current state is an optimal policy for the infinite horizon setting.
We thus consider policies of the form $u_t = K x_t$ and define their infinite horizon expected cost,
\begin{align*}
\Jof{K} = \lim_{T \to \infty} \frac{1}{T} \EEBrk{\sum_{t=1}^{T} x_t^{\mkern-1.5mu\mathsf{T}} \brk*{Q + K^{\mkern-1.5mu\mathsf{T}} R K} x_t},
\end{align*}
where the expectation is taken with respect to the random noise variables $w_t$. Let
$
K_{\star}
=
\argmin_K\Jof{K}
$
be a (unique) optimal policy and
$
J_{\star}
=
\Jof{K_{\star}}
$
denote the optimal infinite horizon expected cost, which are both well defined under mild assumptions.%
\footnote{These are valid under standard, very mild stabilizability assumptions (see \citealp{bertsekas1995dynamic}) that hold in our setting.}
We are interested in minimizing the \emph{regret} over~$T$ decision rounds, defined as
\begin{align*}
\regret
=
%
\sum_{t=1}^{T} \brk!{x_t^{\mkern-1.5mu\mathsf{T}} Q x_t + u_t^{\mkern-1.5mu\mathsf{T}} R u_t - J_{\star}}
%
.
\end{align*}
We focus on the setting where the learner does not have a full a-priori description of the transition parameters $A_{\star}$ and $B_{\star}$, and has to learn them while controlling the system and minimizing the regret.
Throughout, we assume that the learner has knowledge of constants $\alpha_0 > 0$ and $\psi \ge 1$ such that
\begin{align*}
%
%
%
%
%
%
\norm{Q^{-1}}, \norm{R^{-1}} \leq \ifrac{1}{\alpha_0}
,
\text{ and }
\norm{B_{\star}} \le \psi
.
\end{align*}
We also assume that there is a known stable (not necessarily optimal) policy $K_0$ and $\nu >0$ such that $\Jof{K_0} \le \frac14\nu$. We note that all of the aforementioned parameters could be easily estimated at the cost of an additive constant regret term by means of a warm-up period. However, recovering the initial control $K_0$ gives a constant that depends exponentially on the problem parameters as shown by \citet{chen2020black,mania2019certainty,cohen2019learning}.
Finally, denote the set of all ``admissable'' controllers
\begin{align*}
\mathcal{K}
=
\brk[c]{K \mid \Jof{K} \le \nu}
.
\end{align*}
By definition, $K_0 \in \mathcal{K}$. As discussed below, over the set $\mathcal{K}$ the LQR cost function $J$ has certain regularity properties that we will use throughout.
\subsection{Smooth Optimization}
\citet{fazel2018global} show that while the objective $\Jof{\cdot}$ is non-convex, it has properties that make it amenable to standard gradient based optimization schemes. We summarize these here as they are used in our analysis.
\begin{definition}[PL-condition]
\label{def:PL}
A function $f : \mathcal{X} \to \RR$ with global minimum $f^*$ is said to be $\mu$-PL if it satisfies the Polyak-Lojasiewicz (PL) inequality with constant $\mu > 0$, given by
\begin{align*}
\mu(f(x) - f^*)
\le
\norm{\nabla f(x)}^2
\quad,
\forall x \in \mathcal{X}
.
\end{align*}
\end{definition}
\begin{definition}[Smoothness]
\label{def:smooth}
A function $f :\RR[d] \to \RR$ is locally $\beta, D_{0}$-\emph{smooth} over $\mathcal{X} \subseteq \RR[d]$ if for any $x \in \mathcal{X}$ and $y \in \RR[d]$ with $\norm{y-x} \le D_{0}$
\begin{align*}
\norm{
\nabla f(x) - \nabla f(y)
}
\le
{\beta}\norm{x - y}
.
\end{align*}
\end{definition}
\begin{definition}[Lipschitz]
\label{def:lipschitz}
A function $f :\RR[d] \to \RR$ is locally $\beta, D_{0}$-\emph{Lipschitz} over $\mathcal{X} \subseteq \RR[d]$ if for any $x \in \mathcal{X}$ and $y \in \RR[d]$ with $\norm{y-x} \le D_{0}$
\begin{align*}
\abs{
f(x) - f(y)
}
\le
{G}\norm{x - y}
.
\end{align*}
\end{definition}
It is well-known that for functions satisfying the above conditions and for sufficiently small step size $\eta$, the gradient descent update rule
\begin{align*}
x_{t+1} = x_t - \eta \nabla f(x_t)
\end{align*}
converges exponentially fast, i.e., there exists $0 \le \rho < 1$ such that
$
{f(x_t) - f^*}
\le
\rho^t (f(x_0) - f^*)
$
(e.g., \citealp{nesterov2003introductory}).
This setting has also been investigated in the absence of a perfect gradient oracle. Here we provide a clean result that shows that the error in the optimization objective is only limited by the \emph{squared error} of any gradient estimate.
Finally, we require the notion of a one point gradient estimate~\citep{flaxman2005online}. Let $f : \mathcal{X} \to \RR$ and define its smoothed version with parameter $r > 0$ as
\begin{align}
\label{eq:smoothedVer}
\smoothCostOf{x}
=
\EE[B]{\costOf{x + r B}}
,
\end{align}
where $B \in \ball{d}$ is a uniform random vector over the Euclidean unit ball.
The following lemma is standard (we include a proof in \cref{sec:TechnicalLemmas} for completeness).
\begin{lemma}
\label{lemma:smoothFuncProperties}
If $f$ is $(D_{0}, \beta)$-\emph{locally smooth} and $r \le D_{0}$, then:
\begin{enumerate}
\item
$
\nabla \smoothCostOf{x}
=
\frac{d}{r} \EE[{\ui[]}]\brk[s]{\costOf{x + r \ui[]} \ui[]}
,
$
where $\ui[] \in \sphere{d}$ is a uniform random vector of the unit sphere;
\item
$
\norm{\nabla \smoothCostOf{x} - \nabla \costOf{x}}
\le
\beta r
,\;
\forall x \in \mathcal{X}
.
$
%
%
%
%
%
%
%
%
%
%
%
\end{enumerate}
\end{lemma}
\subsection{Background on LQR}
It is well-known for the LQR problem that
\begin{align*}
\Jof{K}
=
\tr{\costToGoOf[K] \noiseCov}
=
\tr{(Q + K^{\mkern-1.5mu\mathsf{T}} R K) \steadyCov[K]}
,
\end{align*}
where $\costToGoOf[K], \steadyCov[K]$ are the positive definite solutions to
\begin{align}
\label{eq:Pbellman}
&\costToGoOf[K]
=
Q + K^{\mkern-1.5mu\mathsf{T}} R K + (A_{\star}+B_{\star} K)^{\mkern-1.5mu\mathsf{T}} \costToGoOf[K] (A_{\star}+B_{\star} K)
,
\\
&\steadyCov[K]
=
\noiseCov + (A_{\star}+B_{\star} K) \steadyCov[K] (A_{\star}+B_{\star} K)^{\mkern-1.5mu\mathsf{T}}
.
\end{align}
Another important notion is that of strong stability \citep{cohen2018online}.
This is essentially a quantitative version of classic stability notions in linear control.
\begin{definition}[strong stability]
A matrix $M$ is $(\kappa, \gamma)$-strongly stable (for $\kappa \ge 1$ and $0 < \gamma \le 1$) if there exists matrices $H \succ 0$ and $L$ such that $M = H L H^{-1}$ with $\norm{L} \le 1 - \gamma$ and $\norm{H}\norm{H^{-1}} \le \kappa$. A controller $K$ for is $(\kappa, \gamma)-$strongly stable if $\norm{K} \le \kappa$ and the matrix $A_{\star} +B_{\star} K$ is $(\kappa, \gamma)$-strongly stable.
\end{definition}
The following lemma, due to \citet{cohen2019learning}, relates the infinite horizon cost of a controller to its strong stability parameters.
\begin{lemma}[\citealp{cohen2019learning}, Lemma 18]
\label{lemma:costToStability}
Suppose that $K \in \mathcal{K}$ then $K$ is $(\kappa, \gamma)-$strongly stable with
$
\kappa
=
\sqrt{\ifrac{\nu}{\alpha_0\sigma^2}}
$
and
$
\gamma
=
\ifrac{1}{2\kappa^2}
%
%
.
$
\end{lemma}
The following two lemmas, due to \citet{cohen2018online,cassel2020logarithmic}, show that the state covariance converges exponentially fast, and that the state is bounded as long as controllers are allowed to mix.
\begin{lemma}[\citealp{cohen2018online}, Lemma 3.2]
\label{lemma:steadyState}
Suppose we play some fixed $K \in \mathcal{K}$ starting from some $x_0 \in \RR[d_x]$, then
\begin{align*}
\norm{\EE\brk[s]{x_t x_t^{\mkern-1.5mu\mathsf{T}}} - \steadyCov[K]}
&\le
\kappa^2 e^{-2\gamma t} \norm{x_0 x_0^{\mkern-1.5mu\mathsf{T}} - \steadyCov[K]}
,
\\
\abs{\EE\brk[s]{\ct} - \Jof{K}}
&\le
%
%
\frac{\nu \kappa^2}{\sigma^2} e^{-2\gamma t} \norm{x_0 x_0^{\mkern-1.5mu\mathsf{T}} - \steadyCov[K]}
.
\end{align*}
\end{lemma}
\begin{lemma}[\citealp{cassel2020logarithmic}, Lemma 39]
\label{lemma:stateBound}
Let
$K_1, K_2, \ldots \in \mathcal{K}$. If we play each controller $K_i$ for at least $\tau \ge 2 \kappa^2 \log 2\kappa$ rounds before switching to $K_{i+1}$ then for all $t \ge 1$ we have that
$
\norm{x_t}
\le
6\kappa^4 W
$
and
$
\ct
\le
36 \nu \kappa^{8} W^2 / \sigma^2
.
$
\end{lemma}
The following is a summary of results from \citet{fazel2018global} that describe the main properties of $\steadyCov[K], \costToGoOf[K], \Jof{K}$. See \cref{sec:TechnicalLemmas} for the complete details.
\begin{lemma}[\citealp{fazel2018global}, Lemmas 11, 13, 16, 27 and 28]
\label{lemma:fazel}
Let $K \in \mathcal{K}$ and $K' \in \RR[d_u \times d_x]$ with
%
\begin{align*}
\norm{K - K'}
\le
%
%
\frac{1}{8 \psi \kappa^{3}}
=
D_{0}
,
\end{align*}
then we have that
\begin{enumerate}
\item
$
\tr{\costToGoOf[K]}
\le
\Jof{K} / \sigma^2
;
$
$
\tr{\steadyCov[K]}
\le
\Jof{K} / \alpha_0
;
$
\item
$
\norm{\steadyCov[K] - \steadyCov[K']}
\le
(\ifrac{8 \psi \nu \kappa^3 }{\alpha_0})
\norm{K - K'}
;
$
\item
$
\norm{\costToGoOf[K] - \costToGoOf[K']}
\le
16 \psi \kappa^7
\norm{K - K'}
;
$
\item $J$ satisfies the local Lipschitz condition (\cref{def:lipschitz}) over $\mathcal{K}$ with $D_{0}$ and
$
G
=
\ifrac{4 \psi \nu \kappa^7}{\alpha_0}
;
$
\item $J$ satisfies the local smoothness condition (\cref{def:smooth}) over $\mathcal{K}$ with $D_{0}$ and
$
\beta
=
\ifrac{112 \sqrt{d_x} \nu \psi^2 \kappa^8}{\alpha_0}
;
$
\item $J$ satisfies the PL condition (\cref{def:PL}) with
$
\mu
=
\ifrac{4 \nu}{\kappa^4}
.
$
\end{enumerate}
\end{lemma}
\section{Algorithm and Overview of Analysis}
We are now ready to present our main algorithm for model free regret minimization in LQR. The algorithm, given in \cref{alg:main}, optimizes an underlying controller $\Kj$ over epochs of exponentially increasing duration.
Each epoch consists of sub-epochs, during which a perturbed controller $\Kji$ centered at $\Kj$ is drawn and played for $\tau$ rounds.
At the end of each epoch, the algorithm uses $\cjitau$, which is the cost incurred during the final round of playing the controller $\Kji$,
to construct a gradient estimate which in turn is used to calculate the next underlying controller $\Kj[\ind0+1]$.
Interestingly, we do not make any explicit use of the state observation $x_t$ which is only used implicitly to calculate the control signal, via $u_t = K_t x_t$.
Furthermore, the algorithm makes only $\Oof{d_u d_x}$ computations per time step.
\begin{algorithm}[ht]
\caption{LQR Online Policy Gradient} \label{alg:main}
\begin{algorithmic}[1]
\State \textbf{input:}
initial controller $K_0 \in \mathcal{K}$,
step size $\eta$,
mixing length $\tau$,
parameters $\mu, \rj[0], \mj[0]$
%
%
\For{epoch $\ind0 = 0,1,2,\ldots$}
\State \textbf{set}
$
\rj
=
\rj[0] (1 - \mu\eta/3)^{\ind0 / 2}
%
,
\mj
=
\mj[0] (1 - \mu\eta/3)^{-2 \ind0}
$
\For{$\ind1 = 1, \ldots, m_{\ind0}$}
\State \textbf{draw} $\uGji \in \RR[d_u \times d_x]$ with i.i.d.\ $\gaussDist{0}{1}$ entries
\State \textbf{set}
$\uji = \uGji / \norm{\uGji}_F$
\State \textbf{play} $\Kji = \Kj + \rj \uji$ for $\tau$ rounds
\State \textbf{observe} the cost of the final round $\cjitau$
\EndFor
\State \textbf{calculate}
$
\gj
=
\frac{d_x d_u}{\mj \rj} \sum_{\ind1=1}^{\mj} \cjitau \uji
$
\State \textbf{update}
$
K_{\ind0 + 1}
=
\Kj - \eta \gj$
\EndFor
\end{algorithmic}
\end{algorithm}
Our main result regarding \cref{alg:main} is stated in the following theorem: a high-probability $O(\sqrt{T})$ regret guarantee with a polynomial dependence on the problem parameters.
\begin{theorem}
\label{thm:main}
Let
$
\kappa
=
\sqrt{{\nu}/{\alpha_0 \sigma^2}}
$
and suppose we run \cref{alg:main} with parameters
\begin{align*}
&\eta
=
\frac{\alpha_0}{128 \nu \psi^2 \kappa^{10}}
,\;
\tau
=
2\kappa^2
\log (7\kappa T)
,
\\
&\mu
=
\frac{4\nu}{\kappa^{4}}
,\;
\rj[0]
=
\frac{\alpha_0}{448 \sqrt{d_x} \psi^2 \kappa^{10}}
,
\\
&\sqrt{\mj[0]}
=
\frac{2^{17} d_u d_x^{3/2} \psi^2 \kappa^{20} W^2}{\alpha_0 \sigma^{2}} \sqrt{\log \frac{240 T^4}{\delta}}
,
\end{align*}
then with probability at least $1 - \delta$,
\begin{align*}
\regret
=
\Oof{
\frac{d_u d_x^{3/2} \psi^4 \kappa^{36} W^2}{\alpha_0}
\sqrt{T \tau \log \frac{T}{\delta}}
}
.
%
%
\end{align*}
\end{theorem}
Here we give an overview of the main steps in proving \cref{thm:main}, deferring the details of each step to later sections.
Our first step is analyzing the utility of the policies $\Kj$ computed at the end of each epoch. We show that the regret of each $\Kj$ (over epoch $\ind0$) in terms of its long-term (steady state) cost compared to that of the optimal $K_{\star}$, is controlled by the inverse square-root of the epoch length $\mj$.
\begin{lemma}[exploitation]
\label{lemma:exploitCost}
Under the parameter choices of \cref{thm:main}, for any $\ind0 \ge 0$ we have that with probability at least $1 - \delta/8T^2$,
\begin{align*}
\Jof{\Kj}
-
J_{\star}
=
\Oof{\nu \sqrt{\frac{\mj[0]}{\mj}}}
=
\Oof{
{d_u d_x^{3/2} \psi^2 \kappa^{22} W^2} \sqrt{\frac{1}{\mj} \log \frac{T}{\delta}}
}
%
%
,
\end{align*}
and further that $\Jof{\Kj} \le \nu / 2$.
%
%
%
%
%
%
\end{lemma}
The proof of the lemma is based on a careful analysis of gradient descent with inexact gradients and crucially exploits the PL and local-smoothness properties of the loss~$\Jof{\cdot}$. More details can be found in \cref{sec:optimization}.
The more interesting (and challenging) part of our analysis pertains to controlling the costs associated with exploration, namely, the penalties introduced by the perturbations of the controllers~$\Kj$.
The direct cost of exploration is clear: instead of playing the $\Kj$ intended for exploitation, the algorithm actually follows the perturbed controllers $K_{\ind0, \ind1}$ and thus incurs the differences in long-term costs $\Jof{K_{\ind0, \ind1}} - \Jof{\Kj}$.
Our following lemma bounds the accumulation of these penalties over an epoch $\ind0$; importantly, it shows that while the bound scales linearly with the length of the epoch $\mj$, it has a \emph{quadratic} dependence on the exploration radius $\rj$.
\begin{lemma}[direct exploration cost]
\label{lemma:KjiToKj}
Under the parameter choices of \cref{thm:main}, for any $\ind0 \ge 0$ we have that with probability at least $1 - \delta/4T$,
\begin{align*}
\sum_{\ind1 = 1}^{\mj} \Jof{K_{\ind0, \ind1}} - \Jof{\Kj}
=
\Oof{
\frac{\sqrt{d_x} \nu \psi^2 \kappa^{8}}{\alpha_0} \rj^2 \mj
+
\nu \sqrt{\mj \log \frac{T}{\delta}}
}
%
%
%
%
%
%
.
\end{align*}
\end{lemma}
There are additional, indirect costs associated with exploration however: within each epoch the algorithm switches frequently between different policies, thereby suffering the indirect costs that stem from their ``burn-in'' period. This is precisely what gives rise to the differences between the realized cost $\cjis$ and the long-term cost $\Jof{\Kji}$ of the policy $\Kji$, the cumulative effect of which is bounded in the next lemma.
Here again, note the quadratic dependence on the exploration radius $\rj$ which is essential for obtaining our~$\sqrt{T}$-regret result.
\begin{lemma}[indirect exploration cost]
\label{lemma:lqrSwitch}
%
%
%
%
%
%
%
%
%
%
%
%
%
%
Under the parameter choices of \cref{thm:main}, for any $\ind0 \ge 0$ we have that with probability at least $1 - \delta/4T$,
\begin{align*}
&\sum_{\ind1 = 1}^{\mj}
\sum_{\ind2 = 1}^{\tau}
\brk!{\cjis - \Jof{\Kji}}
%
=
\Oof{
\frac{\nu \kappa^8 W^2}{\sigma^2} \tau \sqrt{\mj \log \frac{T}{\delta}}
+
\frac{d_x \nu \psi^2 \kappa^{10}}{\alpha_0}
\mj \rj^2
}
%
%
%
%
%
%
.
\end{align*}
\end{lemma}
The technical details for \cref{lemma:KjiToKj,lemma:lqrSwitch} are discussed in \cref{sec:explorationCost}.
We now have all the main pieces required for proving our main result.
\begin{proof}[of \cref{thm:main}]
Taking a union bound, we conclude that \cref{lemma:lqrSwitch,lemma:KjiToKj,lemma:exploitCost} hold for all $\ind0 \ge 0$ with probability at least $1 - \delta$.
Now, notice that our choice of parameters is such that
\begin{align*}
\rj^2 \mj
=
\rj[0]^2 \sqrt{\mj[0] \mj}
=
\Oof{
\frac{\sqrt{d_x} d_u \alpha_0 W^2}{\psi^2 \sigma^2}
\sqrt{\mj \log \frac{T}{\delta}}
}
.
\end{align*}
Plugging this back into \cref{lemma:lqrSwitch,lemma:KjiToKj} we get that for all $\ind0$,
\begin{align*}
\sum_{\ind1 = 1}^{\mj}
\sum_{\ind2 = 1}^{\tau}
\brk!{ \cjis \!-\! \Jof{\Kji} }
%
&=
\Oof{
\frac{d_u d_x^{3/2} \nu \kappa^{10} W^2}{\sigma^2} \tau \sqrt{\mj \log \frac{T}{\delta}}
}
,
\\
\tau \sum_{\ind1 = 1}^{\mj} \Jof{K_{\ind0, \ind1}} \!-\! \Jof{\Kj}
&=
\Oof{
\frac{d_u d_x \nu \kappa^{8} W^2}{ \sigma^2}
\tau \sqrt{\mj \log \frac{T}{\delta}}
}
.
\end{align*}
We conclude that the regret during epoch $\ind0$ is bounded as
\begin{align*}
\sum_{\ind1 = 1}^{\mj}
\sum_{\ind2 = 1}^{\tau}
\brk!{ \cjis - J_{\star} }
&=
\brk[s]*{\sum_{\ind1 = 1}^{\mj}
\sum_{\ind2 = 1}^{\tau}
\brk!{ \cjis - \Jof{\Kji} }
}
+
\brk[s]*{\tau \sum_{\ind1 = 1}^{\mj} \Jof{K_{\ind0, \ind1}} - \Jof{\Kj}}
+
\brk[s]{\tau \mj (\Jof{\Kj} - J_{\star})}
\\
&=
\Oof{
{d_u d_x^{3/2} \psi^2 \kappa^{22} W^2} \tau \sqrt{\mj \log \frac{T}{\delta}}
}
,
\end{align*}
where the second step also used the fact that $\nu / \sigma^2 \le \kappa^2$.
Finally, a simple calculation (see \cref{lemma:mjSum}) shows that
\begin{align*}
\sum_{\ind0=0}^{n-1} \sqrt{\mj}
=
\Oof{\frac{1}{\mu \eta}\sqrt{T / \tau}}
=
\Oof{
\frac{\psi^2 \kappa^{14}}{\alpha_0} \sqrt{T / \tau}
}
,
\end{align*}
and thus summing over the regret accumulated in each epoch concludes the proof.
\end{proof}
\section{Optimization Analysis}
\label{sec:optimization}
At its core, \cref{alg:main} is a policy gradient method with $\Kj$ being the prediction after $\ind0$ gradient steps. In this section we analyze the sub-optimality gap of the underlying controllers $\Kj$ culminating in the proof of \cref{lemma:exploitCost}. To achieve this, we first consider a general optimization problem with a corrupted gradient oracle, and show that the optimization rate is limited only by the square of the corruption magnitude. We follow this with an analysis of the LQR gradient estimation from which the overall optimization cost follows readily.
\subsection{Inexact First-Order Optimization}
Let $f : \RR[d] \to \RR$ be a function with global minimum $f_* > -\infty$. Suppose there exists $\bar{f} \in \RR$ such that $f$ is $\mu$-\emph{PL}, $(D_{0}, \beta)$-\emph{locally smooth}, and $(D_{0},G)$-\emph{locally Lipschitz} over the sub-level set
$
\mathcal{X}
=
\brk[c]{x \mid \costOf{x} \le \bar{f}}
.
$
We consider the update rule
\begin{align}
\label{eq:pgUpdateRule}
x_{t+1}
=
x_t - \eta \gradOracle[t]
,
\end{align}
where $\costOf{x_0} \le \bar{f}$, and $\gradOracle[t] \in \RR[d]$ is a corrupted gradient oracle that satisfies
\begin{align}
\label{eq:gradCorruption}
\norm{\gradOracle[t] - \nabla \costOf{x_t}}
\le
\pCorrupt[t]
,
\end{align}
where $\pCorrupt \le \min\brk[c]{G,\sqrt{(\bar{f} - f_*)\mu} / 2}$ is the magnitude of the corruption at step~$t$.
Define the effective corruption up to round $t$ as
\begin{align*}
\pEffCorrupt^2
=
\max_{s \le t}\brk[c]*{
\pCorrupt[s]^2
\brk[s]{1 - (\mu \eta / 3)}^{t-s}
}
,
\end{align*}
and notice that if
$
\pCorrupt[s] \brk[s]{1 - (\mu \eta / 3)}
\le
\pCorrupt[s+1]
$
then $\pEffCorrupt = \pCorrupt$.
The following result shows that this update rule achieves a linear convergence rate up to an accuracy that depends quadratically on the corruptions.
\begin{theorem}[corrupted gradient descent]
\label{thm:corruptedConvergence}
Suppose that
$
\eta
\le
\min\brk[c]{
\ifrac{1}{\beta}
,
\ifrac{4}{\mu}
,
\ifrac{D_{0}}{2G}
.
}
$
Then for all $t \geq 0$,
\begin{align*}
\costOf{x_t} - f_*
\le
\max\brk[c]*{
\frac{4\pEffCorrupt[t-1]^2}{\mu}
,
\brk[s]*{1 - \frac{\mu\eta}{3}}^t \brk*{\costOf{x_0} - f_*}
}
%
%
%
,
\end{align*}
and consequently $x_t \in \mathcal{X}$.
\end{theorem}
\begin{proof}
For ease of notation, denote $\gradCorrpution = \gradOracle - \nabla \costOf{x_t}$.
Now, suppose that $x_t \in \mathcal{X}$, i.e., $\costOf{x_t} \le \bar{f}$. Then we have that
\begin{align*}
\norm{x_{t+1} - x_t}
=
\norm{\gradOracle[t]} \eta
\le
(\norm{\nabla\costOf{x_t}} + \pCorrupt[t]) D_{0} / 2 G
\le
D_{0}
,
\end{align*}
where the second step used our choice of $\eta$ and the third step used the Lipschitz assumption on $x_t$ and the bound on $\pCorrupt[t]$.
We conclude that $x_t, x_{t+1}$ satisfy the conditions for local smoothness and so we have that
\begin{align*}
\costOf{x_{t+1}} - \costOf{x_t}
&\le
\nabla\costOf{x_t}^{\mkern-1.5mu\mathsf{T}} \brk*{x_{t+1} - x_t} + \frac{\beta}{2}\norm{x_{t+1} - x_t}^2
\\
&=
\nabla\costOf{x_t}^{\mkern-1.5mu\mathsf{T}} \brk*{-\eta \gradOracle[t]}
+
\frac{\beta}{2}
\norm{\eta \gradOracle[t]}^2
\\
&=
\eta \brk[s]*{
-
\norm{\nabla\costOf{x_t}}^2
-
\nabla\costOf{x_t}^{\mkern-1.5mu\mathsf{T}} \gradCorrpution
+
\frac{\eta\beta}{2}
\norm{\nabla\costOf{x_t} + \gradCorrpution}^2
}
\\
&=
\eta \brk[s]*{-\brk*{1 - \frac{\eta\beta}{2}}\norm{\nabla\costOf{x_t}}^2 - \brk*{1 - \eta\beta}\nabla\costOf{x_t}^{\mkern-1.5mu\mathsf{T}} \gradCorrpution + \frac{\eta\beta}{2}\norm{\gradCorrpution}^2}
\\
&=
\eta \brk[s]*{- \brk*{1 - \eta\beta}\norm{\frac{1}{2}\nabla\costOf{x_t} + \gradCorrpution}^2 -\brk*{\frac{3}{4} - \frac{\eta\beta}{4}}\norm{\nabla\costOf{x_t}}^2 + \brk*{1 - \frac{\eta\beta}{2}}\norm{\gradCorrpution}^2}
\\
&\le
\eta \brk[s]*{-\brk*{\frac{3}{4} - \frac{\eta\beta}{4}}\norm{\nabla\costOf{x_t}}^2 + \brk*{1 - \frac{\eta\beta}{2}}\norm{\gradCorrpution}^2},
\end{align*}
where the last transition holds by choice of $\eta\beta \le 1$.
Next, using the PL condition and the bound on $\gradCorrpution$ (see \cref{eq:gradCorruption}) we get that
\begin{align*}
\costOf{x_{t+1}} - \costOf{x_t}
&\le
\eta \brk[s]*{
-\mu\brk*{\frac{3}{4} - \frac{\eta\beta}{4}}
\brk*{\costOf{x_t} - f_*}
+
\brk*{1 - \frac{\eta\beta}{2}}\pCorrupt^2}
,
\end{align*}
and adding $\costOf{x_t} - f_*$ to both sides of the equation we get
\begin{align*}
\costOf{x_{t+1}} - f_*
\le
\brk[s]*{1 - \frac{\mu\eta}{4}\brk*{3 - \eta\beta}}\brk*{\costOf{x_t} - f_*} + \brk*{1 - \frac{\eta\beta}{2}}\eta\pCorrupt^2.
\end{align*}
Now, if $\frac{4\pCorrupt^2}{\mu} \le \costOf{x_t} - f_*$ then since $\eta\beta \le 1$ we have that
\begin{align}
\label{eq:thm1res1}
\begin{aligned}
\costOf{x_{t+1}} - f_*
&\le
\brk[s]*{
1
-
\frac{\mu\eta}{4}\brk*{3 - \eta\beta}
+
\frac{\mu\eta}{4}\brk*{1 - \frac{\eta\beta}{2}}
}
\brk*{\costOf{x_t} - f_*}
\\
&=
\brk[s]*{1 - \frac{\mu\eta}{2} \brk*{1 - \frac{\eta\beta}{4}}}\brk*{\costOf{x_t} - f_*}
\\
&\le
\brk[s]*{1 - \frac{\mu\eta}{3}} \brk*{\costOf{x_t} - f_*}
.
\end{aligned}
\end{align}
On the other hand, if $\frac{4\pCorrupt^2}{\mu} \ge \costOf{x_t} - f_* \ge 0$ then we have that
\begin{align}
\label{eq:thm1res2}
\begin{aligned}
\costOf{x_{t+1}} - f_*
&\le
\brk[s]*{\max\brk[c]*{
0
,
\frac{4}{\mu}
-
\eta\brk*{3 - \eta\beta}
}
+
\eta\brk*{1 - \frac{\eta\beta}{2}}
}\pCorrupt^2
\\
&\le
\max\brk[c]*{
\eta
,
\frac{4}{\mu} - \eta\brk*{2 - \eta\beta}
}\pCorrupt^2
\le
\frac{4\pCorrupt^2}{\mu}
,
\end{aligned}
\end{align}
where the last transition holds, again, by our choice of $\eta$.
Combining \cref{eq:thm1res1,eq:thm1res2} we conclude that
\begin{align}
\label{eq:convergenceSingle}
\costOf{x_{t+1}} - f_*
\le
\max\brk[c]*{
\frac{4\pCorrupt^2}{\mu}
,
\brk[s]*{1 - \frac{\mu\eta}{3}} \brk*{\costOf{x_t} - f_*}
}
.
\end{align}
In particular, this implies that
$
\costOf{x_{t+1}}
\le
\max\brk[c]*{
f_* + \frac{4\pCorrupt^2}{\mu}
,
\costOf{x_t}
}
\le
\bar{f}
,
$
and thus $x_{t+1} \in \mathcal{X}$. Since we assume that $x_0 \in \mathcal{X}$, this completes an induction showing that $x_t \in \mathcal{X}$ for all $t \ge 0$. We can thus unroll \cref{eq:convergenceSingle} recursively to obtain the final result.
\end{proof}
\subsection{Gradient Estimation}
The gradient estimate $\gj$ is a batched version of the typical one-point gradient estimator. We bound it in the next lemma using the following inductive idea: if $\Jof{\Kj} \le \nu / 2$, then $\Kji \in \mathcal{K}$ and standard concentration arguments imply that the estimation error is small with high probability and thus \cref{thm:corruptedConvergence} implies that $\Jof{\Kj[\ind0+1]} \le \nu / 2$.
\begin{lemma}[Gradient estimation error]
\label{lemma:GradEstBound}
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
Under the parameter choices of \cref{thm:main}, for any $\ind0 \ge 0$ we have that with probability at least $1 - (\delta / 8T^3)$,
\begin{align*}
\norm{\gj - \nabla\Jof{\Kj}}_F
\le
\frac{\sqrt{\mu \nu}}{4}
\brk*{1 - \frac{\mu \eta}{3}}^{\ind0 / 2}
.
\end{align*}
\end{lemma}
\begin{proof}[of \cref{lemma:GradEstBound}]
Assume that conditioned on the event $\Jof{\Kj[\ind0]'} \le \nu / 2$ for all $\ind0' \le \ind0$, the claim holds with probability at least $1 - \delta / 8 T^4$.
We show by induction that we can peel-off the conditioning by summing the failure probability of each epoch.
Concretely, we show by induction that the claim holds for all $\ind0' \le \ind0$ with probability at least $1 - \ind0 \delta / 8 T^4$. Since the number of epochs is less than $T$ (in fact logarithmic in $T$), this will conclude the proof.
The induction base follows immediately by our conditional assumption and the fact that $\Jof{\Kj[0]} \le \nu / 4$. Now, assume the hypothesis holds up to $\ind0-1$. We show that the conditions of \cref{thm:corruptedConvergence} are satisfied with $\bar{f} = \nu / 2$ up to round $\ind0$, and thus $\Jof{\Kj[\ind0']} \le \nu / 2$ for all $\ind0' \le \ind0$. We can then invoke our conditional assumption and a union bound to conclude the induction step.
We verify the conditions of \cref{thm:corruptedConvergence}. First, the Lipschitz, smoothness, and PL conditions hold by \cref{lemma:fazel}. Next,
notice that by definition $J_{\star} \le \Jof{\Kj[0]} \le \nu / 4$, and so by the induction hypothesis
$
\norm{\gj[\ind0'] - \nabla\Jof{\Kj[\ind0']}}_F
\le
\sqrt{\nu \mu} / 4
\le
\sqrt{(\bar{f} - f_*)\mu} / 2
\le
G
,
$
for all $\ind0' < \ind0$. Finally, noticing that $\kappa^2 > d_x$ it is easy to verify the condition on $\eta$.
It remains to show the conditional claim holds.
The event $\Jof{\Kj[\ind0']} \le \nu / 2$ for all $\ind0' \le \ind0$ essentially implies that the policy gradient scheme did not diverge up to the start of epoch $\ind0$.
Importantly, this event is independent of any randomization during epoch $\ind0$ and thus will not break any i.i.d.~assumptions within the epoch.
Moreover, by \cref{lemma:fazel} and since $\rj[0] \le \nu / 2 G$, this implies that
$
\Jof{\Kji[\ind0', \ind1]} \le \Jof{\Kj} + G \rj \le \nu
,
$
i.e., $\Kji[\ind0', \ind1] \in \mathcal{K}$ for all $\ind1$ and $\ind0' \le \ind0$.
For the remainder of the proof, we implicitly assume that this holds, allowing us to invoke \cref{lemma:steadyState,lemma:stateBound,lemma:fazel}. For ease of notation, we will not specify this explicitly.
Now, let $\Jr$ be the smoothed version of $J$ as in \cref{eq:smoothedVer}.
Since $\rj \le D_{0}$ we can use \cref{lemma:smoothFuncProperties} to get that
\begin{align*}
\norm{\gj - \nabla \Jof{\Kj}}_F
%
&\le
\norm{\gj - \nabla\JrOf{\Kj}}_F
+
\norm{\nabla\JrOf{\Kj} - \nabla\Jof{\Kj}}_F
\\
&\le
\beta \rj
+
\norm{\gj - \nabla\JrOf{\Kj}}_F
,
\end{align*}
Next, we decompose the remaining term using the triangle inequality to get that
\begin{align*}
\norm{\gj - \nabla\JrOf{\Kj}}_F
&=
\norm*{
\frac{1}{\mj} \sum_{\ind1=1}^{\mj} \brk{
\frac{d_x d_u}{\rj} \cjitau \uji
-
\nabla\JrOf{\Kj}
}
}_F
\\
&\le
\norm*{
\frac{1}{\mj} \sum_{\ind1=1}^{\mj} \brk{
\frac{d_x d_u}{\rj} \Jof{\Kji} \uji
-
\nabla\JrOf{\Kj}
}
}_F
+
\norm*{
\frac{1}{\mj} \sum_{\ind1=1}^{\mj} \brk{
\frac{d_x d_u}{\rj} (\cjitau - \Jof{\Kji}) \uji
}
}_F
.
\end{align*}
By \cref{lemma:smoothFuncProperties}, we notice that, conditioned on $\Kj$, the first term is a sum of zero-mean i.i.d random vectors with norm bounded by $2 d_u d_x \nu / \rj$.
We thus invoke \cref{lemma:vectorAzuma} (Vector Azuma) to get that with probability at least $1 - \ifrac{\delta}{16T^4}$
\begin{align*}
\norm*{
\frac{1}{\mj} \sum_{\ind1=1}^{\mj}
\frac{d_x d_u}{\rj} \Jof{\Kji} \uji
-
\nabla\JrOf{\Kj}
}_F
\le
\frac{d_u d_x \nu}{\rj}
\sqrt{\frac{8}{\mj} \log\frac{240 T^4}{\delta}}
.
\end{align*}
Next, denote
$
Z_{\ind1}
=
\frac{d_x d_u}{\rj} (\cjitau - \Jof{\Kji}) \uji
,
$
and notice that the remaining term is exactly
$
\norm{\frac{1}{\mj} \sum_{\ind1 = 1}^{\mj} Z_{\ind1}}_F
.
$
Let $x_{\ind0, \ind1, \tau}$ be the state during the final round of playing controller $\Kji$, and $\filt[\ind1]$ be the filtration adapted to
$
x_{\ind0, 1, \tau}
,
\ldots
,
x_{\ind0, \ind1, \tau}
,
\uji[\ind0, 1]
,
\ldots
,
\uji
.
$
We use \cref{lemma:steadyState} to get that
\begin{align*}
\norm{\EE\brk[s]*{
Z_{\ind1} \mid \filt[\ind1 - 1]
}}_F
%
&\le
\EE\brk[s]*{\norm{\EE\brk[s]*{
Z_{\ind1} \mid \filt[\ind1 - 1], \Kji
}}_F \mid \filt[\ind1 - 1]}
\\
&\le
\frac{d_x d_u}{\rj}
\EE\brk[s]*{
\abs*{
\EE\brk[s]*{\cjitau \mid \filt[\ind1 - 1], \Kji}
-
\Jof{\Kji}
} \;\Big|\; \filt[\ind1 - 1]}
\\
&\le
\frac{d_x d_u \nu \kappa^2}{\rj \sigma^2} e^{-2\gamma \tau}
\EE\brk[s]*{
\norm{x_{\ind0, \ind1 - 1, \tau} x_{\ind0, \ind1 - 1, \tau}^{\mkern-1.5mu\mathsf{T}} - \steadyCov[\Kji]} \;\Big|\; \filt[\ind1 - 1]}
\\
&\le
\frac{37 d_x d_u \nu \kappa^{10} W^2}{\rj \sigma^2} e^{-2\gamma \tau}
\\
&
\le
\frac{d_x d_u \nu \kappa^{8} W^2}{\rj \sigma^2 T^2}
,
\end{align*}
where the last step plugged in the value of $\tau$ and the one before that used \cref{lemma:stateBound,lemma:fazel} to bound $\norm{\steadyCov[\Kji]} \le \nu / \alpha_0 = \kappa^2 \sigma^2$ and $\norm{x_{\ind0, \ind1 - 1, \tau}} \le 6 \kappa^4 W$.
Further using \cref{lemma:stateBound} to bound $\cjitau$, we also get that
\begin{align*}
\norm{
Z_{\ind1}
-
\EE\brk[s]*{
Z_{\ind1}
\mid
\filt[\ind1 - 1]
}
}_F
\le
\norm{Z_{\ind1}}_F
+
\norm{\EE\brk[s]*{Z_{\ind1} \mid \filt[\ind1 - 1]}}_F
\le
\frac{d_x d_u \cjitau}{\rj}
+
\norm{\EE\brk[s]*{Z_{\ind1} \mid \filt[\ind1 - 1]}}_F
\le
\frac{37 d_x d_u \nu \kappa^{8} W^2}{\rj \sigma^2}
.
\end{align*}
Since $Z_{\ind1}$ is $\filt[\ind1]-$measurable we can invoke \cref{lemma:vectorAzuma} (Vector Azuma) to get that with probability at least $1 - \frac{\delta}{16T^4}$,
\begin{align*}
\norm*{\frac{1}{\mj} \sum_{\ind1 = 1}^{\mj} Z_{\ind1}}_F
%
&\le
\frac{1}{\mj}\norm*{\sum_{\ind1 = 1}^{\mj}
Z_{\ind1}
-
\EE\brk[s]*{
Z_{\ind1}
\mid
\filt[\ind1 - 1]
}
}_F
+
\frac{1}{\mj}\sum_{\ind1 = 1}^{\mj} \norm{\EE\brk[s]*{
Z_{\ind1} \mid \filt[\ind1 - 1]
}}_F
\\
&\le
\frac{d_x d_u \nu \kappa^{8} W^2}{\rj \sigma^2}
\brk[s]*{
37 \sqrt{\frac{2}{\mj}\log\frac{240 T^4}{\delta}}
+
\frac{1}{T^2}
}
\\
&\le
\frac{54 d_x d_u \nu \kappa^{8} W^2}{\rj \sigma^2}
\sqrt{\frac{1}{\mj}\log\frac{240 T^4}{\delta}}
.
\end{align*}
Using a union bound and putting everything together, we conclude that with probability at least $1 - (\delta / 8T^4)$,
\begin{align*}
\norm{\gj - \nabla \Jof{\Kj}}_F
%
&\le
\beta \rj
+
\frac{54 d_x d_u \nu \kappa^{8} W^2}{\rj \sigma^2}
\sqrt{\frac{1}{\mj}\log\frac{240 T^4}{\delta}}
\\
&=
\brk[s]*{
\beta \rj[0]
+
\frac{54 d_x d_u \nu \kappa^{8} W^2}{\sigma^2 \rj[0] \mj[0]^{1/2}}
\sqrt{\log\frac{240 T^4}{\delta}}
}
\brk*{1 - \frac{\mu \eta}{3}}^{\ind0 / 2}
\\
&\le
2\beta \rj[0]
\brk*{1 - \frac{\mu \eta}{3}}^{\ind0 / 2}
\\
&\le
\frac{\sqrt{\mu \nu}}{4}
\brk*{1 - \frac{\mu \eta}{3}}^{\ind0 / 2}
,
\end{align*}
where the last steps plugged in the values of $\mu, \beta, \rj[0]$, and $\mj[0]$.
\end{proof}
\subsection{Proof of \cref{lemma:exploitCost}}
\cref{lemma:exploitCost} is a straightforward consequence of the previous results.
\begin{proof}
For $\ind0 = 0$ the claim holds trivially by our assumption that $\Jof{\Kj[0]} \le \nu / 4$. Now, for $\ind0 \ge 1$,
we use a union bound on \cref{lemma:GradEstBound} to get that with probability at least $1 - \delta / 8T^2$
\begin{align*}
\norm{\gj - \nabla\Jof{\Kj}}
\le
\frac{\sqrt{\mu \nu}}{4}
\brk*{1 - \frac{\mu \eta}{3}}^{\ind0 / 2}
,
\qquad
\forall \ind0 \ge 0
.
\end{align*}
Then by \cref{thm:corruptedConvergence} we have that
\begin{align*}
\Jof{\Kj}
\le
J_{\star}
+
\frac{\nu}{4}
\brk*{1 - \frac{\mu \eta}{3}}^{\ind0 - 1}
%
\le
\min\brk[c]*{
\frac{\nu}{2}
,
J_{\star}
+
\frac{\nu}{2}
\brk*{1 - \frac{\mu \eta}{3}}^{\ind0}
}
,
\end{align*}
where the last step used the facts that
$
J_{\star}
\le
\Jof{\Kj[0]}
\le
\nu / 4
$
and
$1 - \mu \eta / 3 \ge 1/2$.
\end{proof}
\section{Exploration Cost Analysis}
\label{sec:explorationCost}
In this section we demonstrate that exploring near a given initial policy does not incur linear regret in the exploration radius (as more straightforward arguments would give), and use this crucial observation for proving \cref{lemma:KjiToKj,lemma:lqrSwitch}.
We begin with \cref{lemma:lqrSwitch}. The main difficulty in the proof is captured by the following basic result,
which roughly shows that the expected cost for transitioning between two i.i.d.~copies of a given random policy scales with the \emph{variance} of the latter. This would in turn give the quadratic dependence on the exploration radius we need.
\begin{lemma}
\label{lemma:singleSwitch}
Let $K \in \mathcal{K}$ be fixed. Suppose $K_1, K_2$ are i.i.d.~random variables such that $\EE K_i = K$, and
$
\norm{K_i - K}_F
\le
r
\le
D_{0}
.
$
If $x_\tau(K_1)$ is the result of playing $K_1$ for $\tau \ge 1$ rounds starting at $x_0 \in \RR[d_x]$, then
\begin{align*}
\EE\brk[s]{
x_\tau(K_1) ^{\mkern-1.5mu\mathsf{T}} (\costToGoOf[K_2] - \costToGoOf[K_1]) x_\tau(K_1)
}
\le
\frac{256 d_x \nu \psi^2 \kappa^{10}}{\alpha_0}
r^2
+
32 d_x \psi \kappa^9 (\norm{x_0}^2 + \kappa^2 \sigma^2) r e^{-2\gamma \tau}
.
\end{align*}
\end{lemma}
\begin{proof}
Notice that the expectation is with respect to both controllers and the $\tau$ noise terms, all of which are jointly independent.
We begin by using \cref{lemma:steadyState,lemma:fazel} to get that
\begin{align*}
{\tr{(\costToGoOf[K_2] - \costToGoOf[K_1]) (\EE\brk[s]{x_\tau(K_1) x_\tau(K_1)^{\mkern-1.5mu\mathsf{T}} \mid K_1} - \steadyCov[K_1])}
}
%
&\le
32 d_x \psi \kappa^7 r \norm{\EE\brk[s]{x_\tau(K_1) x_\tau(K_1)^{\mkern-1.5mu\mathsf{T}} \mid K_1} - \steadyCov[K_1])}
\\
&\le
32 d_x \psi \kappa^9 r
e^{-2\gamma \tau} \norm{x_0 x_0^{\mkern-1.5mu\mathsf{T}} - \steadyCov[K_1]}
\\
&\le
32 d_x \psi \kappa^9 (\norm{x_0}^2 + \kappa^2 \sigma^2) r e^{-2\gamma \tau}
,
\end{align*}
where the last step also used the fact that $\kappa^2 \sigma^2 = \nu / \alpha_0$.
Now, since $\costToGoOf[K_1], \costToGoOf[K_2]$ do not depend on the noise, we can use the law of total expectation to get that
\begin{align*}
\EE\brk[s]{x_\tau(K_1)^{\mkern-1.5mu\mathsf{T}} (\costToGoOf[K_2] - \costToGoOf[K_1]) x_\tau(K_1)}
%
&=
\EE\brk[s]{ \tr{(\costToGoOf[K_2] - \costToGoOf[K_1]) \EE\brk[s]{x_\tau(K_1) x_\tau(K_1)^{\mkern-1.5mu\mathsf{T}} \mid K_1}}
}
\\
&\le
\EE\brk[s]{ \tr{(\costToGoOf[K_2] - \costToGoOf[K_1]) \steadyCov[K_1]}
}
+
4 d_x \alpha_0 \kappa^2 (\norm{x_0}^2 + \kappa^2 \sigma^2) e^{-2\gamma \tau}
.
\end{align*}
To bound the remaining term, notice that since $K_1, K_2$ are i.i.d, we may change their roles without changing the expectation, i.e.,
\begin{align*}
\EE\brk[s]{ \tr{(\costToGoOf[K_2] - \costToGoOf[K_1]) \steadyCov[K_1]}
}
=
\EE\brk[s]{ \tr{(\costToGoOf[K_1] - \costToGoOf[K_2]) \steadyCov[K_2]}
}
,
\end{align*}
we conclude that
\begin{align*}
\EE\brk[s]{ \tr{(\costToGoOf[K_2] - \costToGoOf[K_1]) \steadyCov[K_1]}
}
&=
\frac12 \EE\brk[s]{ \tr{(\costToGoOf[K_2] - \costToGoOf[K_1]) (\steadyCov[K_1] - \steadyCov[K_2])}
}
\\
&\le
\frac{d_x}{2}\norm{\costToGoOf[K_2] - \costToGoOf[K_1]} \norm{\steadyCov[K_2] - \steadyCov[K_1]}
\\
&\le
\frac{256 d_x \nu \psi^2 \kappa^{10}}{\alpha_0}
r^2
,
\end{align*}
where the last step also used \cref{lemma:fazel}.
\end{proof}
\subsection{Proof of \cref{lemma:lqrSwitch}}
\label{sec:lqrSwitchProof}
Before proving \cref{lemma:lqrSwitch} we introduce a few simplifying notations. Since the lemma pertains to a single epoch, we omit its notation $\ind0$ wherever it is clear from context. For example, $\Kji$ will be shortened to $\Kji[\ind1]$ and $x_{\ind0,\ind1,\ind2}$ to $x_{\ind1,\ind2}$. In any case, we reserve the index $\ind0$ for epochs and $\ind1$ for sub-epochs. In this context, we also denote
the gap between realized and idealized costs during sub-epoch $\ind1$ by
\begin{align*}
\dCi
=
\sum_{\ind2=1}^{\tau} (\cjis[\ind1,\ind2] - \Jof{\Kji[\ind1])}
,
\end{align*}
and the filtration $\hist[\ind1]$ adapted to $w_{1, 1}, \ldots, w_{\ind1, \tau-1}, \Kji[1], \ldots, \Kji[\ind1]$.
We note that $\Kji[\ind1]$ and $\dCi$ are $\hist[\ind1]-$measurable.
The following lemma uses \cref{eq:Pbellman} to decompose the cost gap at the various time resolutions.
See proof at the end of this section.
\begin{lemma}
\label{lemma:dCi}
If the epoch initial controller satisfies $\Jof{\Kj} \le \nu / 2$ then
(recall that $\costToGoOf[K]$ is the positive definite solution to \cref{eq:Pbellman}):
\begin{enumerate}
\item
$
\cjis[\ind1, \ind2] - \Jof{\Kji[\ind1]}
=
x_{\ind1,\ind2}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,\ind2} - \EE[w_{\ind1,\ind2}]\brk[s]{x_{\ind1,\ind2+1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,\ind2+1}}
%
;
$
\item
$
\EE\brk[s]{\dCi \mid \hist[\ind1-1]}
%
%
%
=
\EE\brk[s]*{
x_{\ind1, 1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1, 1}
-
x_{\ind1+1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1+1,1}
\mid
\hist[\ind1-1]
}
;
$
\item
$
\sum_{\ind1=1}^{\mj} \EE\brk[s]{\dCi \mid \hist[\ind1-1]}
\le
\EE\brk[s]{x_{1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[1]}] x_{1,1}}
%
%
%
+
\sum_{\ind1=2}^{\mj} \big(
\EE\brk[s]{x_{\ind1, 1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1, 1} \mid \hist[\ind1-1]}
%
%
%
%
-
\EE\brk[s]{x_{\ind1, 1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1-1]}] x_{\ind1,1} \mid \hist[\ind1-2]}
\big)
.
$
\end{enumerate}
\end{lemma}
We are now ready to prove the main lemma of this section.
\begin{proof}[of \cref{lemma:lqrSwitch}]
First, by \cref{lemma:exploitCost}, the event $\Jof{\Kj[\ind0']} \le \nu / 2$ for all $\ind0' \le \ind0$ holds with probability at least $1 - \delta / 8T$.
As in the proof of \cref{lemma:GradEstBound}, we will implicitly assume that this event holds, which will not break any i.i.d assumptions during epoch $\ind0$ and implies that $\Kji[\ind1] \in \mathcal{K}$ for all $1 \le \ind1 \le \mj$.
%
%
%
%
%
%
%
%
%
%
We also use this to invoke \cref{lemma:stateBound,lemma:fazel} to get that for any $1 \le \ind1, \ind1' \le \mj$ and $1 \le \ind2 \le \tau$ we have
$
x_{\ind1,\ind2}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1']}] x_{\ind1,\ind2}
\le
36 \nu \kappa^8 W^2 / \sigma^2
=
\nu_0
.
$
Now, recall that $\dCi$ is $\hist[\ind1]$-measurable and thus $\dCi - \EE\brk[s]{\dCi \mid \hist[\ind1-1]}$ is a martingale difference sequence. Using the first part of \cref{lemma:dCi} we also conclude that each term bounded by $\tau \nu_0$. Applying Azuma's inequality we get that with probability at least $1 - (\delta / 16T)$
\begin{align*}
\sum_{\ind1=1}^{\mj} \dCi
&=
\sum_{\ind1=1}^{\mj} \dCi - \EE\brk[s]{\dCi \mid \hist[\ind1-1]} + \EE\brk[s]{\dCi \mid \hist[\ind1-1]}
\\
&\le
\sqrt{2 \mj \tau^2 \nu_0^2 \log \frac{16 T}{\delta}}
+
\sum_{\ind1=1}^{\mj} \EE\brk[s]{\dCi \mid \hist[\ind1-1]}
.
\end{align*}
Now, recall from \cref{lemma:dCi} that
\begin{align*}
\sum_{\ind1=1}^{\mj}& \EE\brk[s]{\dCi \mid \hist[\ind1-1]}
\\
&\le
\EE\brk[s]{x_{1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[1]}] x_{1,1}}
%
%
+
\sum_{\ind1=2}^{\mj}
\EE\brk[s]{x_{\ind1, 1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1, 1} \mid \hist[\ind1-1]}
-
\EE\brk[s]{x_{\ind1, 1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1-1]}] x_{\ind1,1} \mid \hist[\ind1-2]}
\\
&=
\EE\brk[s]{x_{1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[1]}] x_{1,1}}
%
%
+
\sum_{\ind1=2}^{\mj}
\EE\brk[s]{x_{\ind1, 1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1, 1} \mid \hist[\ind1-1]}
-
\EE\brk[s]{x_{\ind1, 1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1, 1} \mid \hist[\ind1-2]}
%
%
+
\EE\brk[s]{x_{\ind1, 1}^{\mkern-1.5mu\mathsf{T}}
(
\costToGoOf[{\Kji[\ind1]}]
-
\costToGoOf[{\Kji[\ind1-1]}]
)
x_{\ind1, 1} \mid \hist[\ind1-2]}
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
%
.
\end{align*}
The first two terms in the sum form a martingale difference sequence with each term being bound by $\nu_0$. We thus have that with probability at least $1 - \delta/16 T$,
\begin{align*}
\sum_{\ind1=1}^{\mj} \EE\brk[s]{\dCi \mid \hist[\ind1-1]}
%
\le
\nu_0
+
\sqrt{2 \mj \nu_0^2 \log \frac{16 T}{\delta}}
%
%
+
\sum_{\ind1=2}^{\mj}
\EE\brk[s]{x_{\ind1, 1}^{\mkern-1.5mu\mathsf{T}}
(
\costToGoOf[{\Kji[\ind1]}]
-
\costToGoOf[{\Kji[\ind1-1]}]
)
x_{\ind1, 1} \mid \hist[\ind1-2]}
.
\end{align*}
Notice that the summands in remaining term fit the setting of \cref{lemma:singleSwitch} and thus
\begin{align*}
\sum_{\ind1=2}^{\mj}
\EE\brk[s]{x_{\ind1, 1}^{\mkern-1.5mu\mathsf{T}}
(
\costToGoOf[{\Kji[\ind1]}]
-
\costToGoOf[{\Kji[\ind1-1]}]
)
x_{\ind1, 1} \mid \hist[\ind1-2]}
&\le
\frac{256 d_x \nu \psi^2 \kappa^{10}}{\alpha_0}
\rj^2 \mj
+
\sum_{\ind1=1}^{\mj}
32 d_x \psi \kappa^9 (\norm{x_{\ind1, 1}}^2 + \kappa^2 \sigma^2) \rj e^{-2\gamma \tau}
\\
&
\le
\frac{256 d_x \nu \psi^2 \kappa^{10}}{\alpha_0}
\rj^2 \mj
+
\frac{25 d_x \psi \kappa^{15} W^2 \rj \mj}{T^2}
\\
&
\le
\frac{257 d_x \nu \psi^2 \kappa^{10}}{\alpha_0}
\rj^2 \mj
,
\end{align*}
where the second transition plugged in $\tau$ and used \cref{lemma:stateBound} to bound $\norm{x_{\ind1, 1}}$, and the third transition used the fact that $T^{-2} \le \mj^{-2} \le \rj / \mj[0]$.
Plugging in the value of $\nu_0$ and using a union bound, we conclude that with probability at least $1 - \delta / 4 T$,
\begin{align*}
\sum_{\ind1=1}^{\mj} \dCi
%
\le
\frac{144 \nu \kappa^8 W^2}{\sigma^2} \tau \sqrt{\mj \log \frac{16 T}{\delta}}
%
%
+
\frac{257 d_x \nu \psi^2 \kappa^{10}}{\alpha_0}
\rj^2 \mj
,
\end{align*}
as desired.
\end{proof}
\begin{proof}[of \cref{lemma:dCi}]
By our assumption that $\Jof{\Kj} \le \nu / 2$ we have that $\Jof{\Kji[\ind1]} \le \nu$ and thus $\costToGoOf[{\Kji[\ind1]}]$ is well defined.
Now, recall that $x_{\ind1, \ind2+1} = \brk{A_{\star} + B_{\star} \Kji[{\ind1}]} x_{\ind1, \ind2} + w_{\ind1,\ind2}$ and $\Jof{\Kji[\ind1]} = \EE[w_{\ind1,\ind2}]\brk[s]{w_{\ind1,\ind2}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] w_{\ind1,\ind2}}$ where $\costToGoOf[{\Kji[\ind1]}]$ satisfies \cref{eq:Pbellman} with $K=\Kji[\ind1]$. Then we have that
\begin{align*}
\EE[w_{\ind1,\ind2}]\brk[s]{x_{\ind1,\ind2+1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,\ind2+1}}
&=
\EE[w_{\ind1,\ind2}]\brk[s]{((A_{\star} + B_{\star} \Kji[\ind1]) x_{\ind1,\ind2} + w_{\ind1,\ind2})^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] ((A_{\star} + B_{\star} \Kji[\ind1]) x_{\ind1,\ind2} + w_{\ind1,\ind2})} \\
&=
((A_{\star} + B_{\star} \Kji[\ind1]) x_{\ind1,\ind2})^{\mkern-1.5mu\mathsf{T}} \costToGoOf[K_{t}] ((A_{\star} + B_{\star} \Kji[\ind1]) x_{\ind1,\ind2})
+
\EE[w_{\ind1,\ind2}]\brk{w_{\ind1,\ind2}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] w_{\ind1,\ind2}} \\
&=
((A_{\star} + B_{\star} \Kji[\ind1]) x_{\ind1,\ind2})^{\mkern-1.5mu\mathsf{T}} \costToGoOf[K_{t}] ((A_{\star} + B_{\star} \Kji[\ind1]) x_{\ind1,\ind2})
+
\Jof{\Kji[\ind1]}.
\end{align*}
Now, multiplying \cref{eq:Pbellman} by $x_{\ind1,\ind2}$ from both sides we get that
\begin{align*}
x_{\ind1,\ind2}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,\ind2}
&=
x_{\ind1,\ind2}^{\mkern-1.5mu\mathsf{T}} \brk{Q + \Kji[\ind1]^{\mkern-1.5mu\mathsf{T}} R K_t} x_{\ind1,\ind2}
+
((A_{\star} + B_{\star} \Kji[\ind1]) x_{\ind1,\ind2})^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] ((A_{\star} + B_{\star} \Kji[\ind1]) x_{\ind1,\ind2})
\\
&=
\cjis[\ind1,\ind2]
+
\EE[w_{\ind1,\ind2}]\brk[s]{x_{\ind1,\ind2+1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,\ind2+1}}
-
\Jof{\Kji[\ind1]},
\end{align*}
where the second transition plugged in the previous equality. Changing sides concludes the first part of the proof.
For the second part, notice that taking expectation with respect to $w_{\ind1,\ind2}$ is equivalent to conditional expectation with respect to all past epochs and $w_{1,1}, \ldots, w_{\ind1, \ind2-1}, \Kji[1], \ldots, \Kji[\ind1]$ of the current epoch. Since for all $1 \le \ind2 \le \tau$ this contains $\hist[\ind1-1]$, we use the law of total expectation to get that
\begin{align*}
\EE\brk[s]{\cjis[\ind1,\ind2] - \Jof{\Kji[\ind1]} \mid \hist[\ind1-1]}
&=
\EE\brk[s]{x_{\ind1,\ind2}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,\ind2} \mid \hist[\ind1-1]}
-
\EE\brk[s]{\EE[w_{\ind1,\ind2}]\brk[s]{x_{\ind1,\ind2+1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,\ind2+1}} \mid \hist[\ind1-1]}
\\
&=
\EE\brk[s]{x_{\ind1,\ind2}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,\ind2} \mid \hist[\ind1-1]}
-
\EE\brk[s]{x_{\ind1,\ind2+1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,\ind2+1} \mid \hist[\ind1-1]}
.
\end{align*}
Summing over $\ind2$, noticing that the sum is telescopic, and that time $(\ind1, \tau+1)$ is in fact the start of the next sub-epoch, i.e., $(\ind1+1, 1)$, concludes the second part of the proof.
Finally, we sum over $\ind1$ to get that
\begin{align*}
&\sum_{\ind1=1}^{\mj} \EE\brk[s]{\dCi \mid \hist[\ind1-1]}
=
\sum_{\ind1=1}^{\mj} \EE\brk[s]*{x_{\ind1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,1} - x_{\ind1+1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1+1,1}
\mid
\hist[\ind1-1]
}
\\
&=
\EE\brk[s]{x_{1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[1]}] x_{1,1}}
-
\EE\brk[s]{x_{\mj+1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\mj]}] x_{\mj+1,1} \mid \hist[\mj-1]}
+
\sum_{\ind1=2}^{\mj}
\EE\brk[s]{x_{\ind1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,1} \mid \hist[\ind1-1]}
-
\EE\brk[s]{x_{\ind1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1-1]}] x_{\ind1,1} \mid \hist[\ind1-2]}
\\
&\le
\EE\brk[s]{x_{1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[1]}] x_{1,1}}
+
\sum_{\ind1=2}^{\mj}
\EE\brk[s]{x_{\ind1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1]}] x_{\ind1,1} \mid \hist[\ind1-1]}
-
\EE\brk[s]{x_{\ind1,1}^{\mkern-1.5mu\mathsf{T}} \costToGoOf[{\Kji[\ind1-1]}] x_{\ind1,1} \mid \hist[\ind1-2]}
,
\end{align*}
concluding the third part of the proof.
\end{proof}
\subsection{Proof of \cref{lemma:KjiToKj}}
\begin{proof}[of \cref{lemma:KjiToKj}]
By \cref{lemma:exploitCost}, the event $\Jof{\Kj} \le \nu / 2$ occurs with probability at least $1 - \delta / 8T^2$. Similarly to \cref{lemma:lqrSwitch,lemma:GradEstBound}, we implicitly assume that this event holds, which does not break i.i.d assumptions inside the epoch and implies that $\Kji \in \mathcal{K}$ for all $1 \le \ind1 \le \mj$.
Now, notice that
$
\EE\brk[s]{\Kji \mid \Kj}
=
\Kj
.
$
Since $\Kj \in \mathcal{K}$ and $\rj \le D_{0}$, we can invoke the local smoothness of $\Jof{\cdot}$ (see \cref{lemma:fazel}) to get that
\begin{align*}
\EE\brk[s]{\Jof{\Kji} \mid \Kj}
&\le
\Jof{\Kj}
+
\nabla \Jof{\Kj}^{\mkern-1.5mu\mathsf{T}} \EE\brk[s]{\Kji - \Kj \mid \Kj}
%
%
+
\frac12 \beta \EE\brk[s]{\norm{\Kji - \Kj}^2 \mid \Kj}
\\
&=
\Jof{\Kj}
+
\frac12 \beta \rj^2
.
\end{align*}
We thus have that
\begin{align*}
%
\sum_{\ind1 = 1}^{\mj} \Jof{K_{\ind0, \ind1}} - \Jof{\Kj}
%
%
\le
\frac12 \beta \rj^2 \mj
+
\sum_{\ind1 = 1}^{\mj} \Jof{K_{\ind0, \ind1}}
-
\EE\brk[s]{\Jof{\Kji} \mid \Kj}
.
\end{align*}
The remaining term is a sum of zero-mean i.i.d.\ random variables that are bounded by $\nu$. We use Hoeffding's inequality and a union bound to get that with probability at least $1 - \delta / 4 T$
\begin{align*}
\sum_{\ind1 = 1}^{\mj} \Jof{K_{\ind0, \ind1}} - \Jof{\Kj}
\le
\frac12 \beta \rj^2 \mj
+
\nu \sqrt{\frac12 \mj \log \frac{8 T}{\delta}}
,
\end{align*}
and plugging in the value of $\beta$ from \cref{lemma:fazel} concludes the proof.
\end{proof}
\begin{ack}
We thank Nadav Merlis for numerous helpful discussions.
This work was partially supported by the Israeli Science Foundation (ISF) grant 2549/19, by the Len Blavatnik and the Blavatnik Family foundation, and by the Yandex Initiative in Machine Learning.
\end{ack}
|
\section{Introduction}
\label{Introduction}
In the next generation cellular networks, massive machine-type communications (mMTC) will play an essential role for building the massive Internet of Things (IoT) and hence have been identified as one of the three main use cases in 5G and Beyond 5G (B5G) services. The mMTC service has its particular features:
(1) compared with human-to-human (H2H) communications, the number of potential users (devices) in mMTC scenario could reach up to millions \cite{3GPP1808866};
(2) the user activity patterns are very sporadic;
(3) mMTC demonstrates a salient short-packet-transmission property \cite{ShortPacketsBockelmann};
(4) the users are very sensitive to energy consumption \cite{3GPP1905187}.
Let $N$ represent the total number of potential users in a cell and $N_{ac}$ the number of active users during one random access opportunity (RAO). Hence, the system sparsity could be defined as $\lambda = \frac{N_{ac}}{N}$. Even $\lambda$ is normally small in mMTC scenario, $N_{ac}$ is still significantly larger than the \textcolor[rgb]{0,0,0}{size of LTE preamble pool}. Because $N$ is extremely large \cite{Durisi2016Toward}. As a result, the traditional \textcolor[rgb]{0,0,0}{LTE} grant-based four-step random access will encounter a serious preamble collision problem, which dramatically reduces the access success probability and increases the access latency \cite{ShortPacketsBockelmann}. \textcolor[rgb]{0,0,0}{Implementing the synchronization, active user detection, channel estimation, as well as the data recovery in a one-shot joint operation becomes a promising direction, in grant-free (GF) RA \cite{3GPP165174, 3GPP166095, 3GPP164268}. Since GF RA realizes an "arrival-and-go" transmission of payload, it has} attracted significant attentions in recent years.
\textcolor[rgb]{0,0,0}{Obviously, the non-orthogonal multiple access (NOMA) technology directly motivate the concept of GF RA as NOMA allows multiple users to share the same time-frequency resource. Base Station (BS) can distinguish multi-user data through their different signature patterns if the system overload does not exceed a certain level. Hence the cellular users can transmit their data whenever available. A range of GF NOMA schemes were proposed, including power-based GF NOMA \cite{Ding2014On}, spreading-based GF NOMA \cite{Bayesteh2014Blind, Wanwei2018Uplink, Yuan2017Blind }, interleaving-based GF NOMA \cite{3GPP164269}, etc. However, before grant-free transmission, GF NOMA schemes still require necessary overheads to tackle with the synchronization, identification and channel estimation of active users. Some of the proposals even assume that the active users already connect to BS, or the active users and BS know almost everything about each other, such as the number of multiplexing users, their modulation and coding schemes. These facts imply that realizing an idealized GF RA is extremely challenging, and the required overhead may be inevitably large in practice.}
\textcolor[rgb]{0,0,0}{As a result, reducing the overhead for GF RA becomes an important issue. In this spirit, compressive sensing (CS) technology becomes another foundation to enable GF RA \cite{Fazel2013Random, Hong2015Sparsity }. Because CS is capable of recovering the desired signals from far fewer measurements than the total signal dimensions if a certain signal sparsity is guaranteed, CS is normally employed in GF RA schemes to overcome the challenges of user identification and channel estimation \cite{Chen2018Sparse, PartI2018}, or even simultaneously recover the payload data \cite{Senel2018Grant}}.
\textcolor[rgb]{0,0,0}{In the CS based GF RA, each device is assigned a user specific pilot sequence, termed as preamble. According to CS theory, the preamble length for enabling a successful signal reconstruction is impacted by the number of total users, the sparsity, and the type of measurement matrix. This preamble length is also regarded as a dominant overhead metric of CS based GF RA. Therefore, investigating the minimum preamble length (MPL) is important to mitigate high overhead problem. The associated theoretical analysis has been attempted in \cite{DecodingTao2005, Donoho2003}. However, only an asymptotic order of MPL was obtained. Some uncertain parameters were still involved in this asymptotic order, whose value have to be experimentally tested according to specific scenarios. This problem obviously limits the application of these theoretical results. On the other hand, with the increasing number of potential users or active users, the preamble length has to be increased accordingly. Consequently, in future ultra-dense cellular IoT networks ($\textgreater 10^6$ devices$/km^2$) \cite{3GPP1903968}, the high overhead problem may still exist even employing CS based GF RA, which limits the number of affordable users within the same cell and also aggravate the constrained power budget of RA procedure.}
\textcolor[rgb]{0,0,0}{Extending our horizon further}, to address the overload problem in future random access channel (RACH), a range of other methods have been proposed \cite{Sharma2020Toward}. These existing approaches could be categorized into push-based and pull-based. In push-based approaches, the RA requests are \textcolor[rgb]{0,0,0}{triggered from} the device side while in pull-based approaches, the contention is controlled from the \textcolor[rgb]{0,0,0}{BS} side. \textcolor[rgb]{0,0,0}{Among these methods, it is noticeable that}, grouping is an efficient alternative to \textcolor[rgb]{0,0,0}{relax the cellular density, hence facilitates the massive connectivity and reduces the energy consumption of RA procedure.} That is, all users can be grouped according to various metrics including the quality of service (QoS) \cite{QOS31}, the level of received energy\cite{AngleDomain2019}, the maximum tolerable delay, etc. Most of the grouped transmissions fall under the category of the pull-based approach \cite{DN2015}, where every group has its unique group identification (GID) and the users of a group will access the BS if and only if their GID is granted by the paging message sent from the BS \cite{TsoukaneriGroup}. In \cite{Chih2011Energy}, every group further selects one of its members as the group head (GH). The GH will act as a relay node for other members in the same group.In general, grouped random access demonstrated some particular advantages.
However, these pull-based schemes may cause a serious latency problem while the interval between two paging messages that grant the same GID is quite long\cite{WGroup7}.
In this paper, we more focus on solving the active user detection challenge encountered in mMTC scenarios. \textcolor[rgb]{0,0,0}{More specifically, this paper aims at reducing the CS signaling overhead, saving the random access energy consumption, and accommodating significantly more cellular users}. Hence, a \textcolor[rgb]{0,0,0}{layered grouping RA scheme based on dynamic preamble selection} is proposed. The main contributions of this paper are summarized as follows:
\begin{enumerate}
\item A layered grouping network framework is presented, where a cell is divided into several large clusters based on their geographical locations, and each cluster is further partitioned into a number of small groups according to the proposed construction and maintenance algorithms. It is assumed that users are \textcolor[rgb]{0,0,0}{normally} active \textcolor[rgb]{0,0,0}{in units of small groups,} not individually.
\item In \textcolor[rgb]{0,0,0}{this layered grouping cellular architecture}, the initialization, update, as well as group head (GH) selection procedure of every small group are implemented autonomously, where a self-organizing \textcolor[rgb]{0,0,0}{optimum energy consumption (Opt-EC)} based K-means algorithm is designed and employed. The user which is capable of maximizing the energy efficiency of the entire group is selected as the GH.
\item \textcolor[rgb]{0,0,0}{Associated with the layered cellular architecture}, the RA procedure is divided into cluster load estimation phase \textcolor[rgb]{0,0,0}{(namely, phase-I, in which RA operates in a push-based manner)} and active group detection phase \textcolor[rgb]{0,0,0}{(namely, phase-II, in which RA operates in a pull-based manner). The} conventional user ID (UID) based random access is replaced by a unique group ID (GID) based random access. \textcolor[rgb]{0,0,0}{Here, the layered cellular architecture, the two-phase access procedure, as well as the formed groups access entity, allow a BS to connect with much more coexisted users in an energy efficient manner.}
\item Two kinds of preambles, short and long preambles, are employed, where the short preambles are orthogonal and \textcolor[rgb]{0,0,0}{allocated to clusters as their signature, while the long preambles having the cluster-load depended minimum length are non-orthogonal and allocated to group heads as group identity.} The state-of-the-art approximated message passing (AMP) algorithm \cite{Senel2018Grant, PartI2018} is employed for realizing the active group detection.
\item To analyze the overhead problem and the impact of the preamble length in the AMP algorithm, a tight lower bound on the minimum preamble length (MPL) is derived based on the state-evolution method.
\item \textcolor[rgb]{0,0,0}{Based on the preamble categorization and MPL lower bound analysis, a dynamic preamble selection (DPS) strategy is adopted in phase-II, where the required preambles having the cluster-load depended minimum length are dynamically selected.} It is shown that, benefiting from both the hierarchical preamble assignment and \textcolor[rgb]{0,0,0}{the DPS} strategy, the \textcolor[rgb]{0,0,0}{overhead} of the proposed \textcolor[rgb]{0,0,0}{RA} strategy can be significantly \textcolor[rgb]{0,0,0}{reduced}.
\end{enumerate}
The rest of this paper is organized as follows: in Section \ref{sec:System Model}, the layered grouping network framework is proposed. \textcolor[rgb]{0,0,0}{Its} associated constructing and maintaining mechanism is designed, while the two-phase RA scheme is also depicted. Then, critical techniques employed in the proposed \textcolor[rgb]{0,0,0}{two-phase DPS RA strategy} including optimal energy computing (Opt-EC) based K-means grouping algorithm, AMP detection algorithm, and dynamic preamble selection algorithm are discussed in Section \ref{sec:Key}. The tight lower bound on the minimum preamble length is derived in Section \ref{The}. Simulation results are demonstrated and analyzed in Section \ref{sec:simulation}. Finally, the paper is concluded in Section \ref{sec:conclu}.
\section{System Model}
\label{sec:System Model}
\subsection{Network Framework of Layered Grouping }
\label{preaccess}
We consider an mMTC cell having a radius $R$, where the \textcolor[rgb]{0,0,0}{BS} equipped with a single antenna locates in the cell center and a total number of $N$ coexisted users randomly distribute in the cell. It is assumed that all the mMTC users are mainly static, a typical scenario in mMTC applications \cite{Sharma2020Toward}, e.g. the interactions among machines in the industrial automation, the monitoring in smart agriculture, the environment monitoring for public safety, etc.
According to predefined system configurations including cell size, QoS requirement, maximum number of coexisted cellular users, etc. , BS will divide all the cellular users into a number of $K$ clusters. \textcolor[rgb]{0,0,0}{Typically, $K$ could be a small value, e.g. $K = 2,4,8,$ or $16$.} During phase-I of the proposed RA, in order to reduce the overhead, if the number of clusters is relatively small, then every cluster can be distinguished by a very short orthogonal preamble. Hence, the overhead is reduced. On the other hand, in order to facilitate the practical synchronization, it is better that the users in the same cluster experience similar transmission delays, normally located in a geographical area having roughly the same distance to the BS. This could be achieved by estimating the received signal strength (RSS) at the BS \cite{Lau07Enhanced}, \cite{Bhattacharjee19Weather}. Ideally, the entire cellular is divided into a number of $K$ rings and all the users located in the same ring will be assigned to the same cluster. This effect is visualized by the dashed ellipses in Fig.\ref{fig1SystemModle}. In practice, owing to the limited localization accuracy, some edge users may be assigned to their adjacent cluster. But this potential mismatch will not impact the proposed two-phase GF RA scheme, as the proposed RA scheme is capable of adapting to unbalanced cluster loads.
\begin{figure}
\begin{center}
\includegraphics[width=0.9\textwidth]{fig1SystemModel.eps}
\caption{Network topology of the layered grouping in an mMTC cell.}
\label{fig1SystemModle}
\vspace{-0.1cm}
\end{center}
\end{figure}
Then, the users assigned to the $k^{th}$ cluster $c_k$, $k=1, 2, \cdots, K$ will further participate in a number of $M^{(k)}$ groups. These groups are initialized and updated in a self-organizing manner. The $m^{th}$ group in $k^{th}$ cluster is denoted by $g_{k,m}$. The users pertaining to a group is termed as group members and the set of their user ID is denoted by $\mathcal{G}_{k,m}$. The number of group members is termed as the group size and denoted by $|g_{k.m}|$. We would like to constrain the group size by a small value. Because a large group size normally results in longer distances between GH and its group members, which implies less reliable \textcolor[rgb]{0,0,0}{device-to-device (D2D)} links. Actually, it was shown in Fig.3 of \cite{8847234} that packet error rate over D2D links becomes non-negligible after the group size exceeds 20.
Hence we could further reasonably assume that all the group members are close to each other. As a benefit, the internal message exchange among group members can be reliably realized by the \textcolor[rgb]{0,0,0}{D2D} communication technique \textcolor[rgb]{0,0,0}{\cite{8847234}}, which could be interference-free to other groups and hence spectral efficient. Moreover, a particular group member, namely $\dot{u}_{k,m}$ will be selected as the GH of group $g_{k,m}$. If a normal group member $u_{n}$ wants to communicate with the BS, it firstly sends the message to its GH $\dot u_{k,m}$. Then the GH $\dot u_{k,m}$ relays this message to BS, and vice versa. It implies that throughout the GF RA procedure, the GH $\dot u_{k,m}$ will communicate with the BS on behalf of all the members in the same group. The above-mentioned mechanism is visualized by the dotted ellipses in Fig.\ref{fig1SystemModle}, where the GH is equivalently denoted as a relay node.
Finally, a GH $\dot u_{k,m}$ possesses two kinds of access preambles. The first one, namely $\mathbf{s}_k^I$ is only used in phase-I and actually the signature of cluster $c_k$. All cluster signatures are orthogonal to each other, i.e.
$\langle{\mathbf{s}_i^I, \mathbf{s}_j^I\rangle}=\left\{
\begin{aligned}
0,\quad if\ j\neq i\\
1,\quad if\ j=i
\end{aligned}
\right.
$
, where $i,j=1, 2, \cdots, K$. Since $K$ is relatively small, this orthogonality can be easily satisfied, even for a short preamble length. The second one, namely $\mathbf{s}_{k,m}^{II}$ is only used in phase-II and actually the signature of group $g_{k,m}$. Since $M^{(k)}$ is normally a very large number in mMTC scenarios, $\mathbf{s}_{k,m}^{II}$ employed by different groups in the same cluster has to be non-orthogonal sequences for reducing overhead. More details of preamble assignment could be found in Section \ref{section2B} and Section \ref{dpa}.
\subsection{Construction and Maintenance of the Layered Grouping }
\label{section2B}
The construction of clusters can be controlled by BS, where only an approximated distance from a user to BS is required. On the other hand, the formation and update of groups in each cluster may also be implemented in a centralized manner \cite{8847234} \cite{7992937}, where BS controls the selection of GHs and assigns their group members. However, this centralized management requires a range of global information including users' accurate positions, propagations, data rates, battery levels, etc. Aggregating these information from millions devices in the mMTC scenario may become prohibitive. Hence distributed self-organized formation and update of groups are advocated in this paper. Thus, the construction and maintenance procedures of the proposed \textcolor[rgb]{0,0,0}{layered} network framework are designed as follows
\begin{enumerate}
\item While a user (device) $u_{n}, n = 1,2,\ldots, N,$ firstly powers on in the cell and hears the system broadcast information (including the power level of control channel) from BS, its registration process will then start by sending a registration message containing user ID, device type, and a couple of reference signals, to BS in a contention free manner\footnote{Since the number of users simultaneously switching on is normally extremely low, contention-free transmission of registration message could be realized by predefining a small set of specific channel resources.}.
\item Based on the reference signals contained in the registration message, BS is capable of approximating the distance between a user and itself by utilizing RSS aided positioning techniques \cite{Lau07Enhanced}, \cite{Bhattacharjee19Weather} and further assigning $u_{n} $ to an appropriate cluster $c_{k}$. Then, BS assigns the generation method of a pair of preambles $\mathbf{s}^{I}_{k} $ and $\mathbf{s}^{II}_{k,m} $ , the initial preamble lengths, as well as the group size $\left\vert g_{k,m} \right\vert $ to $u_{n}$. These information and a couple of reference signals are contained in the registration response message (RRM).
\item Based on the reference signals in RRM, user $u_{n}$ is capable of estimating the channel from BS, and the associated channel state information (CSI) is denoted by $h_{n,b} $. It is assumed that all the channels are reciprocal and have a relative long coherent time based on the fact that the mMTC devices are mainly static in our application scenarios. According to RRM, user $u_{n}$ becomes aware of its cluster index $k$. Then, user $u_{n}$ will further autonomously select itself as a GH in a probability of $\frac{1}{\left\vert g_{k,m} \right\vert} $ .
\item BS will periodically broadcast a group update opportunity message (GUOM) to all cellular users. Bearing the quasi-static property of our application scenario in mind, the group update period could be generally long, say daily or even weekly for reducing system overhead.
\item Once the cellular users hear GUOM, they will implement the group initialization\footnote{In the case a group has not been created before.} or update\footnote{In the case a group has existed.} procedure in a self-organizing manner and via D2D links. An Opt-EC based K-means grouping algorithm is designed to iteratively improve the grouping relationship and select the energy efficient GHs, which will be elaborated in Section \ref{OPTECKMEANS}. With the aid of this K-means grouping algorithm, the groups in a cluster are constructed and updated.
\item If the role of a user changes (i.e. switches from a normal group member to a GH and vice versa), it will inform BS of its new state. The BS will add or remove the associated group ID from its group list.
\end{enumerate}
The above-mentioned construction and maintenance procedures are illustrated in Fig.\ref{fig2Networking}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{modifiednetworking.eps}
\caption{Construction and maintenance of the layered grouping network framework.}
\label{fig2Networking}
\vspace{-0.1cm}
\end{center}
\end{figure}
\subsection{Two-Phase Random Access}
\label{RA}
As depicted in Fig.\ref{fig3RandomAccess}, we divide the proposed RA procedure into phase-I and phase-II. During phase-I, after receiving the RAO message from the BS, the GHs of all the active groups in the cell will first transmit their cluster preambles $\mathbf{s}_k^{I}$. Bear in mind that $\mathbf{s}_k^{I}$ of a GH $\dot u_{k,m}$ has been specified during the registration process introduced in Section \ref{section2B}. Furthermore, a group $g_{k,m}$ is regarded as an active group if one or more members in this group want to transmit payload data to the BS. Accordingly, the signal received at the BS during phase-I can be written as
\begin{equation}
\label{yc}
\mathbf{y}^I=\sum_{k=1}^K \sum_{m=1}^{M^{(k)}} a_{k,m}\sqrt{P_{k,m}}\mathbf{s}_k^Ih_{k,m\rightarrow b}+\mathbf{\omega},
\end{equation}
where the activity state $a_{k,m}\in\{0,1\}$ indicates the activity of a group $g_{k,m}$. If $g_{k,m}$ is active, $a_{k,m}=1$, and so forth. $h_{k,m\rightarrow b}$ is the CSI of the channel from the GH $\dot u_{k,m}$ to the BS, which simultaneously encapsulates both the large-scale fading and small-scale fading. $\mathbf{\omega}$ is an AWGN vector, whose elements obey i.i.d complex Gaussian distribution having zero mean and variance $\sigma^2$. $P_{k,m}$ is the actual transmit power of the GH $\dot u_{k,m}$, which is given by
\begin{equation}
\label{rho}
P_{k,m}=P\cdot\frac{\beta_{min}}{\beta_{k,m}},
\end{equation}
where $P$ is a common transmit power that can be afforded by all the GHs. The value of $P$ could be assigned to the users during their registration process. $\beta_{k,m}$ is the average large-scale fading coefficient of the channel from the GH $\dot u_{k,m}$ to the BS. It could be continuously updated by testing the reference signals transmitted by the BS, e.g. the reference signals included in the registration response message, the GUO message and the RAO message. $\beta_{min}$ is the minimum value among $\beta_{k,m}, k\in\{1, 2, \cdots, K\}, m\in\{1, 2, \cdots, M^{(k)}\}$, which could be carried in the RAO message. By substituting \eqref{rho} to \eqref{yc}, it is apparent that specifying the actual transmit power $P_{k,m}$ according to \eqref{rho} is equivalent to employing an adaptive power control mechanism. Hence, at the BS, the average power of the signal received from $\dot u_{k,m}$ approximately equals to $P\cdot \beta_{min}$ \cite{Senel2018Grant}.
During a RAO, the number of active groups in a cluster $c_k$ is termed as its cluster load. Hence, based on the received signal $\mathbf{y}^I$, the BS is capable of estimating the cluster load of $c_k$, which is given by
\begin{equation}
\label{load}
\hat{M}_{ac}^{(k)} = \frac{\langle{\mathbf{y}^I,\mathbf{s}_k^I}\rangle}{\sqrt{P\beta_{min}}}.
\end{equation}
Correspondingly, the sparsity of cluster $c_k$ is approximated by
\begin{equation}
\hat{\lambda}_k=\frac{\hat{M}_{ac}^{(k)}}{M^{(k)}}.
\end{equation}
After the cluster load estimation formulated by \eqref{load}, the BS could rank the access priority of different clusters according to their cluster loads $\hat{M}_{ac}^{(k)}$. A higher access priority is assigned to the cluster having a larger cluster load. As indicated by the \textcolor[rgb]{0,0,0}{largest} streams in the middle and bottom of Fig.\ref{fig3RandomAccess}, the GHs in the highest loaded cluster will firstly send their group access preambles and their payload data, respectively. Based on the cluster load estimation, as well as the compressive sensing theorem, the BS will adaptively select the group preamble length of $L_{k} = |\mathbf{s}^{II}_{k,m} |$ for different clusters. After ranking the access priority and selecting the group preamble length, the BS could arrange the access slots of every cluster. Then, the cluster-specific access slots and preamble lengths that will be used in phase-II are broadcasted by the BS. This message is called as the ``Phase-II solution message'' (PSM) in Fig.\ref{fig3RandomAccess}. More details of \textcolor[rgb]{0,0,0}{DPS strategy} is provided in Section \ref{dpa}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{randomaccess.eps}
\caption{Transmission stream of the two-phase grant-free random access procedure.}
\label{fig3RandomAccess}
\vspace{-0.1cm}
\end{center}
\end{figure}
After receiving the phase-II solution messages, the GHs of all the active groups will generate their unique preambles of $\mathbf{s}_{k,m}^{II}$. The associated generation method has been determined in the user registration process\footnote{In this paper, complex gaussian sequences having zero mean and variance $\frac{1}{\left|\mathbf{s}_{k,m}^{II}\right|}$ are employed as $\mathbf{s}_{k,m}^{II}$.} and the preamble length is indicated by the phase-II solution messages. At this moment, the proposed \textcolor[rgb]{0,0,0}{ two-phase} RA procedure starts its phase-II operations. Firstly, the GHs of all the active groups in the same cluster will simultaneously send their group preambles during a specific access slot that has been indicated by the phase-II solution messages, which are illustrated by the shadowed arrows having dashed, solid, and dotted borders in the middle of Fig.\ref{fig3RandomAccess}. During the access slot of cluster $c_k$, the signal received at the BS is given by
\begin{equation}
\label{yg}
\mathbf{y}_k^{II}=\sum_{m=1}^{M^{(k)}}a_{k,m}\sqrt{P_{k}}\mathbf{s}_{k,m}^{II}h_{k,m\rightarrow b}+\mathbf{\omega}_k =\sqrt{P_{k}}\mathbf{S}_{k}^{II}\mathbf{x}_{k}+\mathbf{\omega}_k,
\end{equation}
where $a_{k,m}$ and $h_{k,m \rightarrow b}$ have been defined in \eqref{yc}. $\mathbf{\omega}_k$ is the AWGN vector in the access slot of $c_k$. $P_{k}$ is the standard transmission power of all the GHs in phase-II. Let $L_k$ denote the length of $\mathbf{s}_{k,m}^{II}$, then we have $\mathbf{S}_{k}^{II}=[\mathbf{s}_{k,1}^{II},\mathbf{s}_{k,2}^{II},\cdots,\mathbf{s}_{k,M^{(k)}}^{II}]\in\mathbb{R}^{L_k\times M^{(k)}}$ and $\mathbf{x}_{k}=[x_{k,1},x_{k,2},\cdots,x_{k,M^{(k)}}]^T$, where $x_{k,m}=a_{k,m}h_{k,m\rightarrow b}$.
As illustrated in the middle of Fig.\ref{fig3RandomAccess}, after the active group detection of all the clusters are completed, the BS will broadcast payload data transmission solution (PDTS) message to every cluster. The active group ID detected by the BS and its payload data transmission time slot are carried in the PDTS message. Finally, as observed at the bottom of Fig.\ref{fig3RandomAccess} that the GH of every active group will relay the payload data of their group members to the BS at different time slots as indicated by PDTS messages.
\section{Group Maintenance and Detection Algorithms}
\label{sec:Key}
\subsection{Group Initialization and Update Algorithms}
\label{OPTECKMEANS}
As stated in Section \ref{section2B}, during step 5 of the construction and maintenance procedure of the layered grouping framework, after receiving the group update opportunity message from BS, the existing GHs $\dot{u}_{k,m}, 1\leq m \leq M^{(k)}$ will broadcast grouping messages. Group ID is contained in these grouping messages. A registered neighboring non-GH user $u_{n}$ will select the one from which it hears the strongest grouping message as its intended GH. Then, $u_{n}$ \textcolor[rgb]{0,0,0}{broadcasts} a subscribing message (SM), which carries its user ID, intended group ID. After receiving all the subscribing messages, a GH $\dot{u}_{k,m}$ is capable of determining all of its group members, and then feedbacks this temporary decision of $\mathcal{G}_{k,m}$ to all the group members. The transmission of above-mentioned intra-group signaling messages will rely on dedicated spectrum such as 5905-5925 MHz in 5G NR V2X (PC5 interface) \cite{V2Xstandard}, or unlicensed spectrum technologies, namely D2D outband communications \cite{8847234}. Hence, the impact of these extra overheads on the cellular radio access network (RAN) can be negligible. These intra-group signaling are summarized as lines 3-5 in Algorithm \ref{algK}.
The proposed Opt-EC based K-means grouping algorithm
aims at selecting the best GH, which simultaneously minimizes the energy required by intra-group communications\footnote{Because, during the ensuing payload transmission slot, the group members will first send their data to the GH, then the GH relays all the data to the BS.} and that required by external \textcolor[rgb]{0,0,0}{cellular communications.} In order to realize this optimization, every group member should be aware of all the channel conditions from other group members to itself, namely $\mathcal{H}_n=\{h_{i,n}\ :\ n\in \mathcal{G}_{k,m},\ i=1, 2, \cdots, \left|g_{k,m}\right|,\ i\neq n\}$.
Again, this requirement could be satisfied by exploiting D2D outband communications and the associated overhead is negligible for cellular RAN. The update of $\mathcal{H}_n$ for every group member is summarized as line 6 in Algorithm \ref{algK}.
\begin{algorithm}[t]
\caption{Opt-EC based K-means grouping algorithm}
\label{algK}
\SetKwInput{KwOut}{Initialize}
\KwOut{cluster index $k$.}
\Begin{
\nlset{1}\While{ (Predefined \textcolor[rgb]{0,0,0}{max} iterations has not been reached.)}
{ \lnlset{loopbegin}{2}
\For{($m=1$ \KwTo $M^{(k)}$)}{
\nlset{3}$\dot{u}_{k,m}$ selected in last iteration broadcasts grouping message;
}
\For{($n=1$ to $N$)}{
\For{($m=1$ \KwTo $M^{(k)}$)}{
\nlset{4}$u_n$ attempts to hear from $\dot u_{k,m}$ and estimate $h_{\dot u_{k,m},n}$;
}
\nlset{5}$ u_n \xrightarrow{subscribe \ group} \mathop{\arg\max}\limits_{\dot u_{k,m}\ heard\ by\ u_n}h_{\dot u_{k,m},n}$;
}
\For{$m=1$ to $M^{(k)}$}{
\nlset{6} Update $\mathcal{G}_{k,m}$ and every $\mathcal{H}_n$; \\
\lnlset{loopend}{7}$\dot u_{k,m} = \mathop{\arg\min}\limits_{u_n \in \mathcal{G}_{k,m}} \gamma_{k,m}(u_n)$;
}
}}
\end{algorithm}
Herein, we further assume that the transmit power, packet size and bandwidth of the reference signals contained in the intra-group signaling messages are fixed to $P$, $D$, and $B$, respectively.
Assume that $u_n$, $u_i$ are group members of $g_{k,m}$, i.e. $n,i\in\mathcal{G}_{k,m}$. The achievable error-free data-rate from $u_i$ to $u_n$ could be characterized by \textcolor[rgb]{0,0,0}{$R_{i,n}=B\log_2\left[1+\frac{P\left|h_{i,n}\right|^2}{N_o}\right]$, where $N_o$ denotes the power density of additive Gaussian noise. Accordingly, if we select $u_n$ as the GH $\dot u_{k,m}$, the energy required by intra-group communications in ensuing payload transmission slot could be characterized by $\epsilon^{(n)}_{\rm{inner}} = \sum_{i\in \mathcal{G}_{k,m}\atop i\neq n} P\cdot \frac{D}{R_{i,n}}.$}
\textcolor[rgb]{0,0,0}{Similarly, the achievable error-free data-rate from $u_n$ to BS could be characterized by $R_{n,b}=B\log_2\left[1+\frac{P\left|h_{n,b}\right|^2}{N_o}\right]$. Again, if we select $u_n$ as GH $\dot u_{k,m}$, the energy required by external communications between $u_n$ and BS could be characterized by $\epsilon^{(n)}_{\rm{outer}} = P\cdot \frac{\vert g_{k,m} \vert \cdot D}{R_{n,b}}.$}
\textcolor[rgb]{0,0,0}{Finally, the energy efficiency of selecting $u_n$ as the GH $\dot u_{k,m}$ could be characterized by $\gamma_{k,m}(u_n)=\epsilon^{(n)}_{\rm{inner}}+\epsilon^{(n)}_{\rm{outer}}$, where a smaller value of $\gamma_{k,m}(u_n)$ implies a better energy efficiency. $\gamma_{k,m}(u_n),\ n\in\mathcal{G}_{k,m}$ will be calculated at the group member $u_n$ and then be forwarded to current GH. Hence, by running the Opt-EC based K-means algorithm, the GH of $g_{k,m}$ could be selected according to $\dot u_{k,m}=\mathop{\arg\min}\limits_{u_n\in\mathcal{G}_{k,m}}\gamma_{k,m}(u_n),$ which is summarized as line 7 in Algorithm \ref{algK}.}
The above-mentioned operations can be repeated again among the cellular users for further adjusting the grouping relationships and optimizing GH selections. But, in practice, owing to the limited energy budget, the iterations of Algorithm \ref{algK} has to be terminated within a predefined maximum number.
\subsection{MMSE Denoiser Based AMP Algorithm}
\label{amp}
The classical system model used in compressive sensing is represented as
\begin{equation} \label{ampmodel}
\mathbf{y}=\mathbf{Ax+\omega},
\end{equation}
where $\mathbf{x}$ is the original signal vector having a number of $M$ elements. $\mathbf{A}\in \mathbb{R}^{L\times M}$ is the measurement matrix. $\omega$ is an AWGN vector. \textcolor[rgb]{0,0,0}{Since $M \gg L$}, $\mathbf{y}$ is actually a compressed and corrupted observation of $\mathbf{x}$. As an efficient solution of \textcolor[rgb]{0,0,0}{recovering $\mathbf{x}$ from $\mathbf{y}$}, the approximated message passing (AMP) algorithm is first proposed in \cite{AMP2009}. Its theoretical derivation could be found in \cite{DonohoHow}.
In more detail, the AMP algorithm could be formulated by the following iterative procedures
\begin{equation}\label{eq4}
\mathbf{x}^{t+1} = \eta_t(\mathbf{A}^*\mathbf{z}^t+\mathbf{x}^t,\mathbf{\tau}_t),
\end{equation}
\vspace{-0.3cm}
\begin{equation}\label{eq5}
\mathbf{z}^{t+1} = \mathbf{y}-\mathbf{A}\mathbf{x}^{t+1}+\frac1 \mu \mathbf{z}^{t}\langle \eta^{'}_t(\mathbf{A}^*\mathbf{z}^{t}+\mathbf{x}^{t};\mathbf{\tau}_{t})\rangle,
\end{equation}
\vspace{-0.1cm}
\begin{equation}\label{eq6}
{\tau}_t \approx \frac1 {\sqrt L}\lVert \mathbf{z}^t \rVert _2.
\end{equation}
where $\mathbf{x}$ is initialized to a zero vector, i.e. $\mathbf{x}^0=0$, $\eta_t(\cdot)$ is the soft thresholding function and $t$ \textcolor[rgb]{0,0,0}{is} the index of iteration, $\mathbf{x}^t$ represents the estimation of $\mathbf{x}$ at the $t^{th}$ iteration, $\mathbf{z}^t$ calculates the residual component, $\mathbf{A}^*$ denotes the conjugate transpose of $\mathbf{A}$, $\langle\cdot\rangle$ denotes the average of a vector, $\eta^{'}_t$ is the first derivative of $\eta_t$ with respect to the first argument, and $\mu = \frac{L}{M}$ is the under-sampling ratio. In contrast to the conventional iterative thresholding algorithms, $\eta^{'}_t(\mathbf{A}^*\mathbf{z}^{t}+\mathbf{x}^{t};\mathbf{\tau}^{t})$ is a new component invoked by the AMP algorithm and known as the \textcolor[rgb]{0,0,0}{``Onsager reaction term''}, which is identified as the fundamental improvement of the AMP algorithm.
Furthermore, in \cite{Chen2018Sparse}, the soft thresholding denoiser $\eta_{t}\left( \cdot \right)$ is developed to an MMSE denoiser \textcolor[rgb]{0,0,0}{as follows
\begin{equation}\label{eq7}
\eta_t(\hat{x}_m^t,g_m) = \mathbb{E}[X|\hat{X}^t=\hat{x}_m^t,G=g_m],
\end{equation}
where $X$, \textcolor[rgb]{0,0,0}{$\hat{X}^t$, $\hat{x}^{t}_{m}$, and $g_{m}$ have the same definitions as that in} \cite{Chen2018Sparse}.}
\textcolor[rgb]{0,0,0}{This} MMSE denoiser can employ the large-scale fading coefficient $G$ known at the BS as a priori information of AMP algorithm. Hence it results in a better recovery accuracy. The above-stated MMSE based AMP algorithm is employed to solve the active group detection problem by replacing the classical compressive sensing model given in \eqref{ampmodel} with the group access model given in \eqref{yg}. \textcolor[rgb]{0,0,0}{Accordingly,} the variables $\mathbf{y}$, $\mathbf{A}$, $\mathbf{x}$, $\mathbf{\omega}$ involved in \eqref{ampmodel}$\sim$\eqref{eq6} \textcolor[rgb]{0,0,0}{are} replaced by the variables $\mathbf{y}_k^{II}$, $\mathbf{S}_k^{II}$, $\mathbf{x}_k$, $\mathbf{\omega}_k$ given in \eqref{yg}, respectively. The under-sampling ratio of $\mu$ in \eqref{eq5} \textcolor[rgb]{0,0,0}{is} calculated by $\frac{L_k}{M^{(k)}}$. The number of total elements $M$ and nonzero elements $M_{ac}$ in $\mathbf{x}$ is replaced by that of total groups $M^{(k)}$ and active groups $M^{(k)}_{ac}$ in a cluster $c_{k}$, respectively.
\begin{algorithm}[t]
\caption{Dynamic preamble selection algorithm.}
\label{algD}
\KwIn{$\bm{y}^I$}
\KwOut{$\bm{V}$, $\bm{L}=[L_1, L_2, \cdots, L_K]$ \tcc*[r]{$\bm{V}$ indicates the access slot indices of every cluster, $\bm{L}$ indicates the preamble lengths selected for every cluster.}
}
\SetKwInput{KwOut}{Initialize}
\KwOut{$K$, target $pF$, target $pM$, $\left\lbrace M^{(k)}, k=1, 2, \cdots, K\right\rbrace$,
$\left\lbrace[R_1^{(k)}, R_2^{(k)}], k=1, 2, \cdots, K\right\rbrace$
\tcc*[r]{$R_1^{(k)}, R_2^{(k)}$ are the inner and outer radius of the $k^{th}$ cluster, respectively.}}
\BlankLine
\Begin{
\nlset{1}\For{k=1 \KwTo $K$}{
\nlset{2}$\hat{M}_{ac}^{(k)}=\mathrm{Estimate\_Cluster\_Load}(\bm{y}^I, k)$ \tcc*[r]{according to \eqref{load}.}
\nlset{3}$L_k=\mathrm{Lower\_Bound\_on\_MPL}(\hat{M}_{ac}^{(k)}, M^{(k)}, R_1^{(k)}, R_2^{(k)}, pF, pM)$;
\tcc*[f]{according to \eqref{neweq5}, \eqref{neweq9} and \eqref{Lk3}.}
\nlset{4}$\bm{L}[k]=1.1*L_k$ \tcc*[r]{slightly enlarge $L_k$, see Section \ref{The}.}
\nlset{5}$\hat{\mathbf{M}}_{ac}[k] = \hat{M}_{ac}^{(k)}$;
}
\nlset{6}$\bm{V}=\mathrm{Allocate\_Access\_Slot}(\hat{\mathbf{M}}_{ac}, \bm{L})$
\tcc*[r]{arrange the access priority of every cluster in the descending order of $\hat{M}_{ac}^{(k)}$, then allocate the access slot indices of every cluster according to $\bm{L}$.}
\nlset{7}\KwSty{Return:} $\bm{V}$, $\bm{L}$;
}
\end{algorithm}
\subsection{Dynamic Preamble Selection Algorithm}
\label{dpa}
In the context of our active group \textcolor[rgb]{0,0,0}{detection and according to the CS theorem \cite{DecodingTao2005,Eldar2012Compressed},} to satisfy a certain detection accuracy, the minimum preamble length is related to the number of active groups in a cluster, namely $M_{ac}^{(k)}$ and to the total number of groups in a cluster, namely $M^{(k)}$. Apparently, in practice, the value of $M_{ac}^{(k)}$ and $M^{(k)}$ shall be salient different in different clusters. It implies employing a preamble with inappropriate length will result in either a serious detection inaccuracy or an excessive overhead. Accordingly, a \textcolor[rgb]{0,0,0}{DPS} algorithm shown in algorithm \ref{algD} is designed and employed in phase-II of the random access. \textcolor[rgb]{0,0,0}{The technical challenge of algorithm \ref{algD} occurs at its line 3 that evaluates the MPL. Owing to analyzing complexity and importance, we specifically discuss it in Section \ref{The}.}
\section{Theoretical Analysis on the Minimum Preamble Length}
\label{The}
According to Section \ref{dpa}, finding the MPL required by MMSE based AMP algorithm for achieving a certain data recovery accuracy makes great sense. Similar works have been attempted in \cite{DecodingTao2005} and \cite{Eldar2012Compressed}. However, two deficiencies of the MPL calculation method given in \cite{Eldar2012Compressed} prevent us from applying it in \textcolor[rgb]{0,0,0}{our DPS algorithm:} (1) two constant parameters, namely $C_1$ and $C_2$ are involved, i.e., instead of an exact value, it only provides an asymptotical order; (2) it does not relate to a specific data reconstruction method.
\textcolor[rgb]{0,0,0}{Furthermore}, the authors of \cite{Aksoylar13Sparse} and \cite{Chen17Capacity} attempted to answer this fundamental question from the perspective of classical asymptotic information theoretic analysis. In \cite{Chen17Capacity}, seeking for the MPL is termed as the “minimum user-identification cost” problem. In their Gaussian many-access channel (MnAC) model, the minimum number of channel uses for guaranteeing an error-free random user identification is given by
\begin{equation}\label{equx}
L=\frac{N\cdot H_2(\frac{N_{ac}}{N})}{\frac{1}{2}\log(1+N_{ac}\gamma)},
\end{equation}
where $N$ denotes the total number of cellular users. In contrast, $N_{ac}$ denotes the average number of active cellular users. $\gamma$ denotes the signal-to-noise ratio (SNR) and it is assumed in \cite{Chen17Capacity} that every active user is subject to the same power constraint of $\gamma$. Besides, the entropy function is defined as $H_2(\mathnormal{p})=-\mathnormal{p}\log(\mathnormal{p})-(1-\mathnormal{p})\log(1-\mathnormal{p})$. However, the theorems provided in \cite{Chen17Capacity} are \textcolor[rgb]{0,0,0}{still not} suitable in our scenarios owing to two reasons: (1) the result shown in \eqref{equx} does not relate to any specific active user detection algorithm, hence a significant gap may exist between the MPL required by AMP algorithm and that predicted by \eqref{equx}, as illustrated in Fig.\ref{fig4LBCompares}; (2) only Gaussian noisy channels are considered. However, both large-scale and small-scale fading effects are taken into account in our system for modelling a more practical random access scenario.
In the following, we will provide a tight lower bound on MPL \textcolor[rgb]{0,0,0}{for MMSE-AMP algorithm.}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{Fig4LBcompare.eps}
\caption{MPL versus sparsity. The MPL required by \textcolor[rgb]{0,0,0}{MMSE-AMP} algorithm and that predicted by \cite{Chen17Capacity} are compared. AWGN channels are assumed.}
\label{fig4LBCompares}
\vspace{-0.1cm}
\end{center}
\end{figure}
The state evolution method proposed in \cite{Chen2018Sparse} is employed, where the mean square error (MSE) of data reconstruction is regarded as a state variable. In more detail, at every iteration of MMSE based AMP algorithm, $\hat{X}^t$ in \eqref{eq7} is modeled as a noise corrupted signal. Hence, $\hat{X}^t$ could be formulated as
\begin{equation}\label{neweq0}
\hat{X}^t = X +\tau_t\cdot \omega_0,
\end{equation}
where the random variable $\omega_0$ follows complex Gaussian distribution with zero mean and unit variance. Then, $\tau_t$ is \textcolor[rgb]{0,0,0}{referred to} the variance of the $t^{th}$ estimation $\hat{X}^t$. Particularly, according to \cite{Chen2018Sparse}, $\tau_t$ is given by
\begin{equation}\label{taut2}
\tau_{t+1}^2 = \frac {\sigma^2} {P_kL_k}+ \frac {M^{(k)}} {L_k} MSE(\tau_t),
\end{equation}
where $\sigma^2$ is the variance of background noise $\omega$ involved in \eqref{yc}. The function $MSE(\cdot)$ evaluates the \textcolor[rgb]{0,0,0}{MSE} of its input variable and is specified in \cite{Chen2018Sparse}.
Based on the state evolution method \cite{Chen2018Sparse}, in order to achieve a high data reconstruction accuracy, the recursive reconstruction progress formulated by \eqref{eq4}-\eqref{eq6} has to converge. It means the variance of the AMP estimation, namely $\tau_t$ should constantly decrease to $\frac {\sigma^2} {P_kL_k}$. Hence, the following inequality holds
\begin{equation}\label{taut21}
\tau_{t+1}^2 \leq \tau_t^2, \forall t.
\end{equation}
By substituting \eqref{taut2} into the inequality \eqref{taut21}, a lower bound on preamble length for satisfying the convergence of AMP algorithm is given by
\begin{equation}\label{Lk}
L_k \geq \frac {\frac {\sigma^2} {P_k}+ {M^{(k)}}MSE(\tau_t)} {\tau_t^2}.
\end{equation}
In order to evaluate the performance of AMP algorithm, two metrics are invoked: (1) the probability of missed detection (pM) in cluster $c_k$; (2) the probability of false alarm (pF) in cluster $c_k$. They are defined as
\begin{equation}\label{eqpM}
pM^{(k)} = \frac{\sum_{m=1}^{M^{(k)}}\{\hat{x}_{k,m}< \theta \ \&\ x_{k,m}\neq 0\}}{\sum_{m=1}^{M^{(k)}}\{x_{k,m}\neq 0\}},
\end{equation}
\begin{equation}\label{eqpF}
pF^{(k)} = \frac{\sum_{m=1}^{M^{(k)}}\{\hat{x}_{k,m} \geq \theta\ \&\ x_{k,m}= 0\}}{\sum_{m=1}^{M^{(k)}}\{x_{k,m}= 0\}},
\end{equation}
where $x_{k,m}$ is defined in \eqref{yg}, $\hat{x}_{k,m}$ is the estimation of $x_{k,m}$ given by the AMP based active group detection, and $\theta$ denotes the decision threshold employed by AMP algorithm. While $\hat{x}_{k,m} \geq \theta $, AMP algorithm will regard $g_{k,m}$ as an active group. Then, according to the state evolution method, the $\mathrm{pM}^{\left( k \right)}$ and $\mathrm{pF}^{\left( k \right)}$ \textcolor[rgb]{0,0,0}{that can} be achieved in the $t^{\mathrm{th}}$ AMP iteration could be characterized by
\begin{equation}\label{neweq5}
\left\{
\begin{array}{l}
pF^{(k)} = e^{-\frac{\theta^2}{\tau_t^2}}, \\
pM^{(k)} = \frac 1 M \sum \limits_{m=1}^M(1-e^{-\frac{\theta^2}{\tau_t^2+g_{m}^2}}) = \displaystyle{\int (1-e^{-\frac{\theta^2}{\tau_t^2+g^2}})\cdot P_G^{(k)}(g)\ dg} ,
\end{array}
\right.
\end{equation}
where $P_G^{(k)}(g)$ is the probability density function of the large-scale fading coefficient $g$. The random variable $g$ takes both the path-loss effect and the shadowing effect into account. In more detail, the path-loss effect is modeled as $\alpha+\beta log_{10}(d)$, where $d$ is the distance between a group head and the BS. The shadowing effect follows log-normal distribution with a variance of $\sigma^{2}_{s}$.
In practical applications, we aim at a target performance of $pM$ and $pF$, namely $pM_{obj}$ and $pF_{obj}$, respectively. Then, by substituting the target $pM_{obj}$ and $pF_{obj}$ into \eqref{neweq5}, the appropriate decision threshold $\theta$ and the required variance of $t^{th}$ AMP estimation $\tau_t$ can be determined while given the large-scale fading model. The associated solutions of $\theta$ and $\tau_t$ could be denoted as $\theta_{obj}$ and $\tau_{obj}$, respectively.
Bear the above statements in mind, in order to obtain $\theta_{obj}$ and $\tau_{obj}$, we shall specify the large-scale fading model. According to the proposed layered grouping network framework shown in Fig.\ref{fig1SystemModle}, the users of a cluster uniformly locates in the same ring, whose inner and outer radius are represented by $R_1$ and $R_2$, respectively. Hence the distance between a user and the BS obeys $d \sim [R_1,R_2]$. Accordingly, the probability density function of large-scale fading coefficient could be formulated as
\begin{equation}\label{neweq8}
P_{G}(g) \triangleq a_1g^{-\gamma_1}Q_1(g)-a_2g^{-\gamma_2}Q_2(g),
\end{equation}
where
\begin{equation}\label{neweq9}
\left\{
\begin{array}{lr}
a_{1} = \frac {40} {(R_2-R_1)^2\beta \sqrt{\pi}} exp(\frac{2(\ln 10)^2\sigma_s^2}{\beta^2} - \frac{2 \ln(10)\alpha}{\beta}),\\
a_{2} = \frac {40R1} {(R_2-R_1)^2\beta \sqrt{\pi}} exp(\frac{(\ln 10)^2\sigma_s^2}{2\beta^2} - \frac{\ln(10)\alpha}{\beta}),\\
\gamma_{1} \triangleq \frac {40} {\beta} +1, \gamma_{2}\triangleq \frac {20} {\beta} +1,\\
Q_i(g) = \int_{b\ln g+c_{i1}}^{b\ln g+c_{i2}} exp(-s^2)ds, i\in \{{1,2}\}\\
c_{i2} = \frac{-\alpha - \beta log_{10}(R_1)} {\sqrt{2}\sigma_s} - \frac{20}{i \beta b}, i\in \{{1,2}\}\\
c_{i1} = \frac{-\alpha - \beta log_{10}(R_2)} {\sqrt{2}\sigma_s} - \frac{20}{i \beta b}, i\in \{{1,2}\}\\
b = -\frac {10\sqrt{2}}{\ln (10)\sigma_s}.\\
\end{array}
\right.
\end{equation}
According to the state evolution method, if the AMP algorithm always achieves the target performance of $pM_{obj}$ and $pF_{obj}$, then the following inequality has to be true as long as a sufficient large iteration number $t$ is chosen
\begin{equation}\label{taut3}
\tau_{t+1} \leq \tau_{obj} \leq \tau_t.
\end{equation}
By substituting \eqref{taut2} into the above inequality, it results in
\begin{equation}\label{Lk2}
L_k \geq \frac {\frac {\sigma^2} {P_k}+ {M^{(k)}}MSE(\tau_t)} {\tau_{obj}^2}.
\end{equation}
Then, it is provable that $\mathrm{MSE \left( \cdot \right)}$ is a monotonically increasing function in the region of $g\in\left[0, 100\right]$. Hence we have $MSE(\tau_{obj}) \leq MSE(\tau_t)$ in practical scenarios. It implies replacing $MSE(\tau_t)$ by $MSE(\tau_{obj})$ in \eqref{Lk2} will yield a relaxed lower bound (LB) on MPL, which is given by
\begin{equation}\label{Lk3}
L_k \geq \frac {\frac {\sigma^2} {P_k}+ {M^{(k)}}MSE(\tau_{obj})} {\tau_{obj}^2}.
\end{equation}
For example, in practice, simultaneously achieving $pM_{obj}=0.05$ and $pF_{obj}=0.05$ may be an acceptable active group detection performance. While considering the large-scale fading model given in \eqref{neweq8}-\eqref{neweq9}, the associated solution of \eqref{neweq5} is $\theta_{obj} \approx 8.65 \times 10^{-8}, \tau_{obj} \approx 5 \times 10^{-8} $. Assuming that network configurations including the SNR, the cluster size $M^{(k)}$, as well as the sparsity $\lambda$ are known. Then, by substituting $\tau_{obj} \approx 5 \times 10^{-8} $ into \eqref{Lk3}, we could calculate the exact lower bound on MPL that enables the AMP algorithm to achieve the target active user detection performance.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{Fig5LB1.eps}
\caption{Comparison between simulated MPL and its lower bound with respect to different transmit powers. The practical preamble length required by conventional ``no grouping'' RA scheme \cite{Chen2018Sparse} is provided for comparison.}
\label{fig4Bound1}
\vspace{-0.1cm}
\end{center}
\end{figure}
It is evidenced in Fig.\ref{fig4Bound1} that the lower bounds on MPL \eqref{Lk3} gets quite close to the actual MPLs estimated by the Monte Carlo simulations for different transmit powers, although their discrepancy will be slightly enlarged while increasing the sparsity $\lambda$.
The comparison between simulated MPL and its lower bound with respect to different coverage areas are illustrated in Fig.\ref{fig5Bound2}. Fig.\ref{fig5Bound2} demonstrates that the discrepancy between the simulated MPL and its lower bound will be impacted by different cluster coverages. This phenomenon is due to the fact that the large-scale fading effect will be impacted by the cluster coverage as formulated in \eqref{neweq8} and \eqref{neweq9}. On the condition of having a low sparsity of $\lambda \leq 0.05$, this discrepancy would not exceed $10\%$ of the theoretical lower bound. Therefore, in algorithm \ref{algD}, we first calculate the lower bound on MPL for a specific cluster $c_k$, then the $L_k$ employed by the active groups in $c_k$ will be $10\%$ higher than the lower bound.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{Fig6LB2.eps}
\caption{Comparison between simulated MPL and its lower bound with respect to different cluster coverages. The practical preamble length required by conventional ``no grouping'' RA scheme \cite{Chen2018Sparse} is provided for comparison.}
\label{fig5Bound2}
\vspace{-0.1cm}
\end{center}
\end{figure}
Furthermore, the conventional no grouping GF RA scheme proposed in \cite{Chen2018Sparse}, which also employs the MMSE based AMP algorithm, is shown in Fig.\ref{fig4Bound1} and Fig.\ref{fig5Bound2} as well. It is demonstrated in Fig.\ref{fig4Bound1} and Fig.\ref{fig5Bound2} that the MPL required by the proposed layered grouping based RA scheme is always significantly less than that required by its counterpart in \cite{Chen2018Sparse} regardless of different SNR values, different sparsities, different coverages. For the sake of fair comparison, the total number of cellular users remains the same in Fig.\ref{fig4Bound1} and Fig.\ref{fig5Bound2}.
\section{Simulation Results and Discussions}
\label{sec:simulation}
In this section, the active user detection performance of the proposed \textcolor[rgb]{0,0,0}{two-phase DPS aided} RA scheme is simulated. \textcolor[rgb]{0,0,0}{The obtained results are} compared with the \textcolor[rgb]{0,0,0}{classical CS aided RA} that does not exploit any grouping strategy\textcolor[rgb]{0,0,0}{\cite{Chen2018Sparse} and with the conventional group paging aided RA that does not exploit CS technologies\cite{2016Group}.} Without loss of generality, the group size $|g_{k,m}|$ \textcolor[rgb]{0,0,0}{and the} number of groups in different clusters $M^{(k)}$ \textcolor[rgb]{0,0,0}{are prefixed to constants} regardless of the \textcolor[rgb]{0,0,0}{indices of $k$ and $m$.} Other system parameters are listed in Table.\ref{t2}.
\begin{table}
\begin{center}
\small
\setlength{\tabcolsep}{0.01mm}{
\caption{System Configuration.}
\begin{tabular}{c|c}
\hline
Parameter & Value \\ \hline
Radius of the cell & $1000 m$ \\
Path-loss model & $15.3+37.6\log_{10}(d(m))$ \\
Variance of shadowing $\sigma_s^2 $ & $8$ \\
Background noise $N_o$ & -99dBm \\
Total cellular devices $N$ & $10000, 20000 $ \\
Number of clusters $K$ & \textcolor[rgb]{0,0,0}{$2, 4, 8$} \\
Group size $\vert g_{k,m} \vert$ & $5$ \\
Number of groups per cluster $M^{(k)}$ & \textcolor[rgb]{0,0,0}{$M= 1000, 2000$} \\
Preamble type and length of $\mathbf{s}^{I}_{k}$ & Walsh Seq., $\vert\mathbf{s}^{I}_{k} \vert = 32$\\
TX power $P$ defined in \eqref{rho} & 23dBm \\
Preamble type of $\mathbf{s}^{II}_{k,m}$ & Gaussian Random Seq.\\
TX power $P_k$ defined in \eqref{yg} & 23dBm \\ \\
Cluster coverages of $K=2$& $\left[ {{R_1^{(k)}},{R_2^{(k)}}} \right] = \left\{ {\begin{array}{*{20}{c}}
{\left[ {0,500} \right]\;,\;\;\;k = 1}\\
{\left[ {500,1000} \right],\;\;\;k = 2}
\end{array}} \right.$ \\ \\
Cluster coverages of $K=4$& $\left[ {{R_1^{(k)}},{R_2^{(k)}}} \right] = \left\{ {\begin{array}{*{20}{c}}
{\left[ {0,250} \right]\;,\;\;\;k = 1}\\
{\left[ {250,500} \right],\;\;\;k = 2}\\
{\left[ {500,750} \right],\;\;\;k = 3}\\
{\left[ {750,1000} \right],\;\;\;k = 4}\\
\end{array}} \right.$ \\ \hline
\end{tabular}
\label{t2}}
\end{center}
\end{table}
In Fig.\ref{fig6pFpM}, the $pF$, $pM$ versus transmit power in the \textcolor[rgb]{0,0,0}{two-phase DPS aided} RA is compared with that of \textcolor[rgb]{0,0,0}{CS aided RA} \cite{Chen2018Sparse}, which also employs the MMSE based AMP algorithm. It is a general consensus that $pF=pM$ \textcolor[rgb]{0,0,0}{implies a good performance balance of active user detection can be achieved.} Hence the decision threshold employed by \textcolor[rgb]{0,0,0}{MMSE based AMP} algorithm is adjusted for achieving $pF=pM$.
Comparing the solid line labelled by diamonds with the dashed line labelled by crosses in Fig.\ref{fig6pFpM}, the \textcolor[rgb]{0,0,0}{two-phase DPS aided RA} achieves a dramatic power gain with respect to the conventional \textcolor[rgb]{0,0,0}{CS based RA} \cite{Chen2018Sparse} while using the same preamble length of 400. Comparing the solid line labelled by triangles with the solid line labelled by squares in Fig.\ref{fig6pFpM}, the grouping strategy of $K=4,\ M=1000$ achieves a better performance than that of $K=2,\ M=2000$, even they employ the same preamble length of 800. It implies a \textcolor[rgb]{0,0,0}{RA power} gain is available by adjusting the number of clusters and the number of groups.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.55\textwidth]{Fig7pFpM.eps}
\caption{$\mathrm{pF}$, $\mathrm{pM}$ versus transmit power of \textcolor[rgb]{0,0,0}{the two-phase DPS aided RA}. It is clear that the layered grouping scheme employing a preamble length of only 800 can approach the performance of no grouping scheme \cite{Chen2018Sparse} which employs a very long preamble length of 2600.}
\label{fig6pFpM}
\vspace{-0.1cm}
\end{center}
\end{figure}
Then, we define the probability of successful detection ($pS$) in cluster $c_k$ as
\begin{equation}\label{pS}
pS^{(k)}=\frac{\sum_{m=1}^{M^{(k)}}\{\hat{x}_{k,m} \geq \theta \ \&\ x_{k,m}\neq 0 \}}{\sum_{m=1}^{M^{(k)}}\{x_{k,m}\neq 0\}}.
\end{equation}
According to this definition and equation \eqref{neweq5}, $pF$, $pM$ and $pS$ have following relationships
\begin{equation}\label{neweq12}
\left\{
\begin{array}{l}
pS^{(k)} = 1-pM^{(k)} = 1-\int (1-e^{-\frac{\theta^2}{\tau_t^2+g^2}})\cdot P_G^{(k)}(g)dg,\\
pS^{(k)} = 1-\frac {1-\lambda_k} {\lambda_k}pF^{(k)} =1-\frac {1-\lambda_k} {\lambda_k}e^{-\frac{\theta^2}{\tau_t^2}}.
\end{array}
\right.
\end{equation}
The parameter $\tau_t$ involved in \eqref{neweq12} can be determined according to \eqref{taut2}. \textcolor[rgb]{0,0,0}{Hence, \eqref{neweq12} enable us to theoretically analyze the active group detection performance of the proposed RA strategy.}
\textcolor[rgb]{0,0,0}{As a result, the} $pS$ achieved by preamble length fixed strategy is compared with that achieved by \textcolor[rgb]{0,0,0}{DPS} strategy in Fig.\ref{fig8Dynamic}, \textcolor[rgb]{0,0,0}{where both of them employ two-phase RA framework and the grouping strategy is fixed to $K=4, M= 1000$.}
Observe at Fig.\ref{fig8Dynamic} that while employing a predefined preamble length of $|\mathbf{s}_{k,m}^{II}|=64,\ 128$, the $pS$ of \textcolor[rgb]{0,0,0}{grouped} RA still rapidly drops along with the growth of sparsity. In contrast, benefiting from the \textcolor[rgb]{0,0,0}{DPS aided RA} scheme, the \textcolor[rgb]{0,0,0}{system} is capable of achieving a high $pS$ probability throughout the entire sparsity region. Actually, the \textcolor[rgb]{0,0,0}{ DPS strategy} approaches a similar active group detection performance to a preamble length fixed counterpart having $\vert \mathbf{s}^{II}_{k,m} \vert = 256$. However, the average preamble length required by the \textcolor[rgb]{0,0,0}{DPS} strategy is always less than 210 throughout the entire sparsity region. It \textcolor[rgb]{0,0,0}{means} the \textcolor[rgb]{0,0,0}{DPS} strategy will \textcolor[rgb]{0,0,0}{further save} considerable overhead.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{nFig8DPLS.eps}
\caption{ $\mathrm{pS}$ versus sparsity performance of two-phase \textcolor[rgb]{0,0,0}{DPS aided} RA, where preamble length fixed strategy \textcolor[rgb]{0,0,0}{is compared with.} The preamble length required by dynamic scheme is always less than 210 after averaged over all clusters.}
\label{fig8Dynamic}
\vspace{-0.1cm}
\end{center}
\end{figure}
We would like to further evaluate the energy \textcolor[rgb]{0,0,0}{consumption} of two-phase \textcolor[rgb]{0,0,0}{DPS aided} RA. According to interpretation in Section \ref{sec:System Model}, the transmission energy required for constructing the layered grouping network framework is negligible. Because the group construction only happens once after a GH user is registered. Besides, the group update only happens once during a very long period.
\textcolor[rgb]{0,0,0}{Hence we only focus on the random access energy desired on the active user side in the entire “Phase-I” plus “Phase-II” durations. As shown in Fig.\ref{fig3RandomAccess}, it consists of six parts: (1) energy required by transmitting cluster preambles for the sake of cluster load estimation, i.e. $\epsilon_{\mathrm{pmb-I}}=\frac 1 {N\cdot\lambda} {\sum_{k=1}^{K}\sum_{m=1}^{M^{(k)}}a_{k,m}P_{k,m}|\mathbf{s}_k^I|}$, (2) energy consumed by waiting for the Phase-II solution message, i.e. $\epsilon_{\rm{wait-I}}= P_{\rm{wait}}\cdot \overline{T}_{\rm{wait-I}}$, (3) energy required by processing the received Phase-II solution message, i.e. $\epsilon_{\mathrm{pcs-I}}= P_{\rm{pcs}}\cdot \overline{T}_{\rm{pcs-I}}$, (4) energy required by transmitting access preambles for the sake of active group identification, i.e. $\epsilon_{\mathrm{pmb-II}}=\frac 1 {N\cdot\lambda} {\sum_{k=1}^{K}\sum_{m=1}^{M^{(k)}}a_{k,m}P_{k,m}|\mathbf{s}_k^{II}|}$, (5) energy consumed by the waiting model (as defined in \cite{Sesia2009LTE}) in the entire “Phase-II” duration, i.e. $\epsilon_{\rm{wait-II}}= P_{\rm{wait}}\cdot \overline{T}_{\rm{wait-II}}$, (6) energy required by processing the PDTS message, i.e. $\epsilon_{\mathrm{pcs-II}}= P_{\mathrm{pcs-II}}\cdot \overline{T}_{\mathrm{pcs-II}}$. Hence, the average random access energy per active user could be calculated as:}
\begin{equation}\label{energyTotal}
\textcolor[rgb]{0,0,0}{\overline{\epsilon} = \epsilon_{\rm{pmb-I}} + \epsilon_{\rm{wait-I}} + \epsilon_{\rm{pcs-I}} + \epsilon_{\rm{pmb-II}} + \epsilon_{\rm{wait-II}} + \epsilon_{\rm{pcs-II}}}
\end{equation}
\textcolor[rgb]{0,0,0}{In the above energy parts, $P_{k,m}$ and $P_k$ are defined in \eqref{rho} and \eqref{yg}, respectively. Thus, the values of $P_{\rm{wait}}$, $P_{\rm{pcs}}$, $\overline{T}_{\rm{wait-I}}$ and $\overline{T}_{\rm{wait-II}}$ are specified according to similar parameters given in \cite{2016Group}-\cite{Sesia2009LTE}. Particularly, $\overline{T}_{\rm{wait-II}}$ is the average waiting time of an active group required in the entire “Phase-II” duration, which equals to subtracting length of $|\mathbf{s}_k^{II}|$ and $\overline{T}_{\rm{pcs-II}}$ from duration of “Phase-II” \footnote{The duration of “Phase-II” is determined by system configurations of $N$,$K$,$\lambda$ etc. It can be calculated by experimental method as shown in Fig.\ref{fig12Delay}.}. More specifically, according to LTE standard, up to 839 symbols can be transmitted within a single time slot (i.e. 0.5 ms). Hence, we equivalently employ the number of symbols as our time metric.}
\textcolor[rgb]{0,0,0}{Furthermore, if we replace practical preamble length $|\mathbf{s}_k^{II}|$ used in \eqref{energyTotal} by its theoretical lower bound (LB) specified in \eqref{Lk3}, the LB of $\epsilon_{\rm{pmb-II}}$ is given by}
\begin{equation}\label{energypreambleII}
\textcolor[rgb]{0,0,0}{\epsilon_{\rm{pmb-II}}^* = \frac 1 {N\cdot\lambda} {\sum_{k=1}^{K}\sum_{m=1}^{M^{(k)}}a_{k,m}P_{k}\frac {\frac {\sigma^2} {P_k}+ {M^{(k)}}MSE(\tau_{obj})} {\tau_{obj}^2}}},
\end{equation}
\textcolor[rgb]{0,0,0}{Simultaneously, the length of $\overline{T}_{\rm{wait-II}}$ is also minimized, which results in the minimization of $\epsilon_{\rm{wait-II}}$, namely $\epsilon_{\rm{wait-II}}^*$. Substitute $\epsilon_{\rm{pmb-II}}^*$ and $\epsilon_{\rm{wait-II}}^*$ into \eqref{energyTotal}, we refer to the resultant $\overline{\epsilon}$ as theoretical RA energy of our two-phase DPS aided RA, namely $\overline{\epsilon}^*$.}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{nFig9ECcluster.eps}
\caption{\textcolor[rgb]{0,0,0}{RA energy comparison between the proposed scheme, group-paging RA scheme\cite{2016Group} and no grouping RA scheme\cite{Chen2018Sparse}, where $N =20000$, $\lambda = 0.05$, $K$ increases from 2 to 8.}}
\label{fig9EC}
\vspace{-0.1cm}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{nFig10ECgsize.eps}
\caption{\textcolor[rgb]{0,0,0}{RA energy comparison between the proposed scheme, group-paging RA scheme \cite{2016Group} and no grouping RA scheme \cite{Chen2018Sparse}, where $N =20000$, $\lambda = 0.05$, $\vert g_{k,m} \vert$ increases from 5 to 20.}}
\label{fig10ECgsize}
\vspace{-0.1cm}
\end{center}
\end{figure}
Observe at Fig.\ref{fig9EC} and Fig.\ref{fig10ECgsize} that on the condition of approaching the same $pS \geq 90\%$, the two-phase \textcolor[rgb]{0,0,0}{DPS aided} RA scheme achieves a significant power gain compared with \textcolor[rgb]{0,0,0}{both the} conventional no grouping \textcolor[rgb]{0,0,0}{CS based} RA scheme \cite{Chen2018Sparse} \textcolor[rgb]{0,0,0}{and the conventional no CS aided group paging strategy\cite{2016Group}.} \textcolor[rgb]{0,0,0}{There are three major advantages of the proposed layered grouping RA scheme, i.e. (1) in the proposed system, only GH of an active group has to send access preamble, while in \cite{Chen2018Sparse},\cite{2016Group}, every active user has to send access preamble; (2) GHs also have better channel conditions than other group members owing to the Opt-EC based K-means grouping algorithm; (3) the DPS algorithm given in Algorithm \ref{algD} effectively reduces the preamble length used in phase-II. In more details, the energy consumed per active user is reduced as the number of clusters K is increased, as illustrated in Fig.\ref{fig9EC}, while the associated penalty is the increase of RA delay as shown in Fig \ref{fig12Delay} per active user versus. On the other hand, as can be seen from Fig.\ref{fig10ECgsize}, the energy consumed per active user can be reduced more significantly when the group size is increased. However, it may become impractical if the group size is too big due to the complexity and synchronization requirements.}
\textcolor[rgb]{0,0,0}{The capability of DPS aided RA scheme to support the massive connectivity is depicted in Fig. \ref{fig11Capacity}, which is measured in the maximum number of coexisted users in a cell while fixing the preamble length, the sparsity, the transmit power, as well as the target level of $pF$ and $pM$. As can be observed from Fig. \ref{fig11Capacity}, for the given condition, the proposed RA scheme is capable of supporting more than $10^{5}$ users. In contrast, no grouping strategy \cite{Chen2018Sparse} can only support} \textcolor[rgb]{0,0,0}{approximately 2500 }\textcolor[rgb]{0,0,0}{users for the same amount of physical resources.}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{nFig11capacity.eps}
\caption{\textcolor[rgb]{0,0,0}{Comparison on affordable number of coexisted users in a cell between the proposed RA and no grouping CS aided RA.}}
\label{fig11Capacity}
\vspace{-0.1cm}
\end{center}
\end{figure}
\textcolor[rgb]{0,0,0}{Finally,} the average \textcolor[rgb]{0,0,0}{time} required by an active group head $\dot{u}_{k,m}$ for completing random access procedure is adopted as our delay metric. According to the proposed \textcolor[rgb]{0,0,0}{RA} procedure in Fig.\ref{fig3RandomAccess} \textcolor[rgb]{0,0,0}{and in line with our RA energy analysis,} the \textcolor[rgb]{0,0,0}{RA} delay of $\dot{u}_{k,m}$ consists of \textcolor[rgb]{0,0,0}{six} major components:
(1) The time $\overline{T}_1$ required by all active GHs for transmitting their cluster preambles of $|\bm{s}_k^I|,\; k=1,2,\cdots, K$; \textcolor[rgb]{0,0,0}{(2) the waiting time $\overline{T}_{\rm{wait-I}}$ during "Phase-I"; (3) the time $\overline{T}_{\rm{pcs-I}}$ required for processing phase-II solution message; (4) the time $\overline{T}_{2}$ required by an active GH for transmitting its group access preamble; (5) the waiting time $\overline{T}_{\rm{wait-II}}$ during “Phase-II” and (6) the time $\overline{T}_{\rm{pcs-II}}$ required for processing PDTS message. Again, the number of symbols is employed as the time metric.}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\textwidth]{Fig12delay.eps}
\caption{Access delay versus number of clusters in the proposed two-phase GF RA scheme.}
\label{fig12Delay}
\vspace{-0.1cm}
\end{center}
\end{figure}
The random access delay performance of \textcolor[rgb]{0,0,0}{the proposed DPS aided} RA scheme is demonstrated in Fig.\ref{fig12Delay}, where $\lambda = 0.05,\; N=20000$. Observe at Fig.\ref{fig12Delay} that the \textcolor[rgb]{0,0,0}{proposed} RA scheme imposes a higher random access delay than the conventional no grouping \textcolor[rgb]{0,0,0}{CS based RA scheme, especially when the number of clusters $K$ grows. But, benefiting from the DPS scheme, employing a large group size could slightly mitigate the latency.} Consequently, \textcolor[rgb]{0,0,0}{our proposal} may be more suitable for the \textcolor[rgb]{0,0,0}{mMTC devices which} have a relative higher tolerance of time delay. Fortunately, benefiting from the \textcolor[rgb]{0,0,0}{DPS aided RA} scheme, the random access delay will not linearly increase with respect to the number of clusters.
\section{Conclusions}
\label{sec:conclu}
With \textcolor[rgb]{0,0,0}{the} explosive growth of IoT \textcolor[rgb]{0,0,0}{devices (may approach around 125 billion by 2030 \cite{Sharma2020Toward}),} the mMTC communication will become one of the most important services of forthcoming B5G networks. An extremely large number of devices are accommodated in a single cell. Neighborhood devices are inclined to have a similar communication behavior and QoS requirement. These fundamental characteristics of future mMTC scenario motivate us to propose a two-phase \textcolor[rgb]{0,0,0}{DPS aided} RA scheme. \textcolor[rgb]{0,0,0}{Benefiting from the proposed layered grouping network framework,} instead of a large number of active users directly access the BS, only the GHs of active groups access the BS on behalf of all the active members. The mechanism of constructing and maintaining this layered grouping network framework, as well as the two-phase RA \textcolor[rgb]{0,0,0}{procedure} are carefully designed, especially the Opt-EC based K-means grouping algorithm, the orthogonal sequence based cluster load estimation, the dynamic preamble selection, as well as the MMSE based AMP algorithm. A tight lower bound on MPL required by AMP algorithm for achieving the given detection accuracy is provided. \textcolor[rgb]{0,0,0}{In summery, compared with the existing RA schemes, the proposed DPS aided RA scheme
achieves three major improvements: a) reducing the access overhead as shown in Fig. \ref{fig4Bound1}, Fig. \ref{fig5Bound2} and Fig.\ref{fig8Dynamic}; b) saving the access energy as shown in Fig.\ref{fig6pFpM} and Fig.\ref{fig9EC}; c) increasing the number of coexisted cellular users as shown inFig.\ref{fig11Capacity}, at the price of relatively longer access delay.}
{
\bibliographystyle{ieeetr}
|
\section{Introduction}
Electronic structure methods that can efficiently handle both static and dynamic correlation remain an important area of investigation.
Because there is a wide spectrum of strongly correlated problems, ranging from mildly "quasi"-degenerate scenarios (e.g. in the electronic structure of diradicals \cite{abe2013diradical, nakano2017diradical}) to extensively near-degenerate problems (e.g. in the electronic structure of multi-centre transition metal clusters \cite{kurashige2013entangled, sandeep2014fe2s2, zhendong2019pcluster}) a variety of theoretical strategies have been proposed.
For highly-degenerate problems, it is common to combine dynamic correlation methods with an explicit treatment of a multi-reference state.
The representation of the multi-reference state can range from an exact complete active space (CAS) representation~\cite{das1972new, werner1980quadratically, werner1985second} (for small numbers of orbitals), to density matrix renormalization group (DMRG)~\cite{white1992real, white1992density, white1993density, white1999ab, chan2002highly, legeza2003controlling, chan2004algorithm, legeza2008applications, chan2009density, marti2010density, chan2011density, sharma2012spin, wouters2014density, szalay2015tensor, yanai2015density} (e.g. when almost all orbitals are degenerate), to selected configuration interaction~\cite{ivanic2001identification, huron1973iterative, GarSceGinCafLoo-JCP-18,LooSceBloGarCafJac-JCTC-18,GarSceLooCaf-JCP-17, buenker1974individualized, evangelisti1983convergence, harrison1991approximating, STEINER1994263, Neese2003, ABRAMS2005121, BYTAUTAS200964, evangelista2014driven, knowles2015compressive, schriber2016communication, tubman2016deterministic, liu2016ici, caffarel2016using, Holmes_2016, sandeep2017, LiOttHolShaUmr-JCP-18} and Monte Carlo approximations~\cite{booth2009fermion, cleland2010communications, ghanem2019unbiasing, sabzevari2018improved, neuscamman2016subtractive} for intermediate cases.
On top of these, various flavours of perturbation theory~\cite{andersson1990second, angeli2001introduction, angeli2001n, angeli2002n, neuscamman2010review, yanai2010multireference, kurashige2011second, sharma2014communication, sharma2015multireference, guo2016n, sokolov2017time, sharma2017semistochastic, guo2018communication}, configuration interaction~\cite{lischka1981new, szalay2012multiconfiguration, lischka2001high, saitow2015fully}, and
exponential approximations~\cite{paldus1999adv, kong2009state, jeziorski2010multireference, sharma2015multireference} have been explored.
However, the combination of dynamic correlation
with multi-reference representations is not straightforward and usually leads to added conceptual, implementation, and computational complexity.
For quasi-degenerate problems, an alternative strategy can be used, which incorporates a limited amount of static correlation on top of an existing SR method. This has been particularly popular
in conjunction with SR coupled cluster methods.~\cite{vcivzek1966correlation, vcivzek1969use, vcivzek1971correlation, paldus1972correlation, bartlett1995modern, shavitt2009many} Some examples include
variants of tailored coupled cluster~\cite{kinoshita2005coupled, hino2006tailored, lyakh2011tailored, veis2016coupled, faulstich2019numerical, vitale2020fciqmc} and externally corrected coupled cluster methods.~\cite{paldus1994valence, planelles1994valence, planelles1994valence2, paldus2017externally} The computational cost of such SR static correlation methods is often lower than that of true MR dynamic correlation methods and thus large active spaces become affordable. However, the approximations are limited to problems with only a modest amount of degeneracy.
In this work, we will focus on quasi-degenerate problems, and in particular, we will investigate the externally corrected coupled cluster method.~\cite{paldus1994valence, planelles1994valence, planelles1994valence2, paldus2017externally}
This extracts static correlation from a MR method by using the MR wavefunction as an "external" source of higher order coupled cluster amplitudes. For example, in the ecCCSD approximation, the $T_3$ and $T_4$ amplitudes are extracted from the external source and a new set of $T_1$ and $T_2$ amplitudes are computed in their presence. Should the $T_3$ and $T_4$ amplitudes be exact, then the $T_1$ and $T_2$ amplitudes {and the energy} will be exact. A different, spiritually related, approximation is tailored CCSD~\cite{kinoshita2005coupled, hino2006tailored, lyakh2011tailored}, which has been of renewed interest of late~\cite{veis2016coupled, faulstich2019numerical, vitale2020fciqmc}. Here, instead of higher order cluster amplitudes, the large (active space) $T_1$ and $T_2$ amplitudes are fixed from an external source.
The ecCCSD method has a long history and external sources, ranging from unrestricted Hartree-Fock~\cite{paldus1982cluster, paldus1983bond, paldus1984approximate, piecuch1996approximate} to CASSCF and CASCI~\cite{peris1997single, li1997externally, peris1998externally, stolarczyk1994complete, xu2015externally} and, most recently, full configuration interaction quantum Monte Carlo\cite{deustua2018}, have been used. One of the more successful applications
of ecCCSD is the reduced multireference (RMR) CCSD method\cite{li1997, li2000truncated}, which uses a MRCISD wave function as the external source.
RMR CCSD(T), which incorporates some of the residual dynamic correlation through perturbative triples, has also been studied.\cite{li2006}
Despite the promising performance of RMRCCSD(T) in several studies~\cite{li1998dissociation, li1998singlet, li1999simultaneous, li2000reduced, li2000effect, li2006truncated, li2006singlet, li2007reduced, li2008coupled}, it suffers from two main limitations. First, conventional MRCISD can only be applied for modest sizes of active spaces (typically, up to about $16$ orbitals as limited by the exact CAS treatment). Second, the (T) correction, although not divergent like its single-reference counterpart, still overcorrects the dynamical correlation in the bond-stretched region.\cite{li2008n2}
In this work, we make two modifications to ecCCSD to overcome and ameliorate the above limitations.
First, we utilize variational DMRG and HCI wave functions as external sources for ecCCSD.
This allows for the use of larger quasidegenerate active spaces, of the size typically treated by DMRG and HCI.
Second, we explore the renormalized perturbative triples correction. This has been shown in the single reference setting to ameliorate the overcorrection of standard perturbative triples \cite{kowalski2000mmcc}, without affecting the computational scaling.
We describe the use of DMRG and HCI as external sources for ecCCSD in Section~\ref{sec:ecCCSD}.
It is possible to create a near-exact method by using a near-exact external source.
Since near-exact external sources can be computed by DMRG and HCI for the small systems we employ as
test cases in this paper,
the critical question is not simply the accuracy of the method, but the balance between cost and accuracy.
To this end, we explore a variety of approximate treatments of the external DMRG and HCI sources as discussed in Sec.~\ref{sec:ea}.
The various triples approximations for ecCCSD are discussed in Section \ref{sec:recccsdt}.
Computational details are provided in Section \ref{sec:compdetail} and the accuracy of the renormalized perturbative triples correction is assessed for
three potential energy surfaces (PESs) in Sections \ref{sec:fm}-\ref{sec:fpal}. We characterize the range of quasi-degenerate correlations captured in this work in Section \ref{sec:diag}. Finally we discuss the limitations of this method in Section \ref{sec:limit} and summarize our findings in Section \ref{sec:conclusions}.
\section{Theory}
In this work, we use a reference configuration $| \Phi \rangle$ with the same occupancy as the
Hartree-Fock (HF) determinant.
It is useful to define a projection operator onto the space of $k$-tuply excited configurations relative to the reference; we denote this $Q_k$. The external source is used to provide an important subset of the triply and quadruply excited configurations; the projector onto this subset is denoted $Q_k^{ec}$, and its complement is $Q_k^c$.
Thus
\begin{alignat}{2}
Q_k & = Q^{\rm ec}_k + Q^{\rm c}_k, \ \ & k&=3,4, \label{eq:p34}
\end{alignat}
\subsection{Externally corrected CCSD with DMRG and HCI wave functions} \label{sec:ecCCSD}
In ecCCSD, the coupled cluster operator $T$ is given by
\begin{align}
T = T_1 + T_2 + Q^{\rm ec}_3 T_3 + Q^{\rm ec}_4 T_4. \label{eq:t}
\end{align}
Here, $T_n, \ n=1,2,3,4, $ are the $n$-fold cluster operators.
In this work, we extract $Q^{\rm ec}_k T_k$, $k=3,4$ from the DMRG or the HCI variational wave function.
For the DMRG wave function, the triply and quadruply excited configurations $D_p$ that define $Q^{\rm ec}_k$, $k=3, 4$ are chosen
to be those where the magnitude of the CI coefficient $c_p$ is above a threshold, i.e.
\begin{align}
|c_p| > s \sqrt{\omega}, \label{eq:dmrg_V}
\end{align}
where $s$ is an arbitrary scaling factor and $\omega$ is the largest discarded weight of the density matrix at the maximum bond dimension in two-dot DMRG sweeps (carried out without noise). An efficient algorithm to convert a matrix-product state to CI coefficients above a given threshold is described in Appendix \ref{apdx:mpsci}. For the HCI variational wave function, $D_p$ is included in the projectors $Q^{\rm ec}_k$, $k=3, 4$ using the heatbath algorithm with a threshold $\epsilon$,\cite{Holmes_2016} i.e., it is included if
\begin{align}
| \langle D_p | H | D_q \rangle c_q | > \epsilon, \label{eq:shci_V}
\end{align}
for at least one determinant $D_q$ which is already in the variational space. The extracted CI coefficients are then converted into cluster amplitudes.
The $T_1$ and $T_2$ amplitudes are obtained by solving the ecCCSD equations using fixed $Q^{\rm ec}_3T_3$ and $Q^{\rm ec}_4T_4$,
\begin{equation}
0 = (Q_1 + Q_2) (H_N e^{T_1 + T_2 + Q^{\rm ec}_3 T_3 + Q^{\rm ec}_4 T_4})_C | \Phi \rangle,
\end{equation}
where $H_N$ is the Hamiltonian in normal-ordered form, and the subscript $C$ denotes the connected part of the corresponding operator expression. With the relaxed $T_1$ and $T_2$, the ecCCSD correlation energy of the ground state is obtained as
\begin{align}
\Delta E^{\rm ecCCSD}_0 = \langle \Phi \vert (H_N e^{T_1 + T_2 + Q^{\rm ec}_3 T_3 + Q^{\rm ec}_4 T_4})_C \vert \Phi \rangle.
\end{align}
\subsection{Approximations in the external source} \label{sec:ea}
While DMRG and HCI can, in small molecules, be a source of nearly exact $T_3$ and $T_4$ amplitudes even in the full
orbital space,
this provides no computational advantage as such calculations are more expensive than the subsequent coupled cluster calculation.
Consequently, it is important to balance the cost of the external source calculation and that of the subsequent coupled cluster calculation by making approximations in the external source. This introduces the additional complication that one must ensure that errors introduced into the external source do not lead to unacceptable errors in the final coupled cluster calculation.
In this work, we consider {six} different types of approximate external sources with different sizes of active spaces and different values of parameters summarized in Table \ref{table:apprx_ext}.
Type I uses CASSCF-like external sources in minimal active spaces. Since the minimal active space of F$_2$, i.e., two electrons in two orbitals (2e,2o), does not contain $T_3$ and $T_4$, we perform minimal active space ecCC calculations for only H$_2$O and N$_2$ (with (4e,4o) and (6e,6o), respectively).
These provide amplitudes that are very close to the exact CASSCF amplitudes.
However, these amplitudes lack the relaxation that comes from allowing excitations within a larger space
of orbitals, and of course the amplitudes outside the minimal active space are completely absent.
Type III uses CASCI-like external sources (the orbitals are not optimized to save computer time)
in larger active spaces ((8e,18o), (10e,16o), and (14e,16o) for H$_2$O, N$_2$, and F$_2$, respectively).
In either case, one can potentially introduce bad external amplitudes if the effect of relaxation on the amplitude
upon going to a larger space is large relative to the size of the amplitude (e.g. changes its sign).
Thus, we study also Type II and Type IV external sources which are similar to Type I and Type II external sources respectively
except that they employ an additional threshold to screen out all except the largest $T_3$ and $T_4$ amplitudes.
The absolute values of the $T_3$ and $T_4$ elements {at the most stretched geometry of each molecule} are sorted in a single large vector. The norm of the vector is computed. Only the largest elements of the vector are retained such that
the resulting norm is more than 80\% of the norm of the full vector.
{Along PESs of each molecule, we used the same set of elements of $T_3$ and $T_4$ (but with the appropriate values for each geometry) as the external sources, to maintain the smoothness of the PESs.}
{As discussed in Section~\ref{sec:fpal}, the type-III and type-IV sources improve upon the PESs obtained from the type-I and type-II sources,
but the DMRG calculations to obtain the sources} incur a higher computational cost than the subsequent CC calculations.
To reduce the cost, we have also tried type-V sources, which employ loosely converged DMRG wave functions with small bond dimensions of $M=25,50,$ and $100$.
Finally, type-VI sources employ large thresholds $\epsilon=0.01,$ and $0.003$ to obtain loosely converged HCI wave functions in the full orbital spaces.
This combination has the advantage that it can be considered a black-box method wherein a single parameter $\epsilon$ controls the tradeoff between accuracy and cost.
\begin{table}[htp]
\caption{Six types of approximate external sources. Detailed explanation to be found in the main text. }
\begin{tabular}{cccll}\hline \hline
\multirow{2}{*}{Type}& \multirow{2}{*}{Method}
& \multirow{2}{*}{Active space} & \multicolumn{2}{c}{Cut-off parameters} \\
& & & \multicolumn{1}{c}{Wave function} & \multicolumn{1}{c}{Cluster amplitude} \\ \hline
\multirow{2}{*}{I} & DMRG & \multirow{2}{*}{Minimal${}^a$} & $M=2000$ & $100$\% with $s=0.1$ \\
& HCI & & $\epsilon=10^{-7}$& $100$\% \\ \\
\multirow{2}{*}{II} & DMRG & \multirow{2}{*}{Minimal${}^a$} & $M=2000$ & $80$\% with $s=0.1$ \\
& HCI & & $\epsilon=10^{-7}$& $80$\% \\ \\
III & DMRG & Larger${}^b$ & $M=2000$ & $100$\% with $s=0.1$ \\ \\
IV & DMRG & Larger${}^b$ & $M=2000$ & $80$\% with $s=0.1$ \\ \\
\multirow{2}{*}{V} & \multirow{2}{*}{DMRG}& \multirow{2}{*}{Larger${}^b$} & \multirow{2}{*}{$M=25,50,100$} & $80$\% \ \ for H${}_2$O and N${}_2$, \\
& & & & $100$\% for F${}_2$ with $s=0.01$ \\ \\
VI & HCI& Full${}^c$ & $\epsilon=0.01, 0.003$ & $100$\% \\ \hline \hline
\end{tabular} \label{table:apprx_ext}
\begin{alignat*}{6}
& {}^{a} \mathrm{(4e,4o)} \ & \mathrm {for} \ \mathrm{H}{}_2{\mathrm O}, \ \ & \mathrm{(6e,6o) } \ & \mathrm {for} \ \mathrm{N}{}_2 \ \ & & \nonumber \\
& {}^{b} \mathrm{(8e,18o)} \ & \mathrm {for} \ \mathrm{H}{}_2{\mathrm O}, \ \ & \mathrm{(10e,16o)} \ & \mathrm {for} \ \mathrm{N}{}_2, \ \ & \mathrm {(14e,16o)}\ & \mathrm {for} \ \ \mathrm{F}{}_2 \nonumber \\
& {}^{c} \mathrm{(10e,58o)}\ & \mathrm {for} \ \mathrm{H}{}_2{\mathrm O}, \ \ & \mathrm{(14e,60o)} \ & \mathrm {for} \ \mathrm{N}{}_2, \ \ & \mathrm {(14e,58o)}\ & \mathrm {for} \ \ \mathrm{F}{}_2 \nonumber
\end{alignat*}
\end{table}
\subsection{Perturbative triples corrections} \label{sec:recccsdt}
Expressions for the standard, renormalized, and completely renormalized perturbative triples corrections can be written down in analogy with
their single reference definitions~\cite{kowalski2000mmcc}.
We first define the completely renormalized (CR)-ecCCSD(T) correction.
We use the state $|\Psi^{\rm ecCCSD(T)} \rangle$ defined as
\begin{align}
\vert \Psi^{\rm ecCCSD(T)} \rangle & = ( 1 + T_1 + T_2 + Q^{\rm ec}_3 T_3 + Q^{\rm c}_3 T^{[2]}_3 + Q^{\rm c}_3 Z_3 ) \vert \Phi \rangle, \\
T^{[2]}_3 \vert \Phi \rangle & = R^{(3)}_0 (V_N T_2)_C \vert \Phi \rangle, \\
Z_3 \vert \Phi \rangle & = R^{(3)}_0 V_N T_1 \vert \Phi \rangle.
\end{align}
where $V_N$ is the two-body part of the Hamiltonian in normal-ordered form and $R^{(3)}_0$ denotes the three-body component of the reduced resolvent operator in many-body perturbation theory, given by differences of orbital energies in the denominator.\cite{shavitt2009many}
The resulting formula for the CR-ecCCSD(T) energy correction is
\begin{align}
\delta_0^{\rm CR-ecCCSD(T)} & = \frac{ \langle \Psi^{\rm ecCCSD(T)} | Q^{\rm c}_3 ( H_N e^{T} )_C | \Phi \rangle
}{ \langle \Psi^{\rm ecCCSD(T)} | e^{T} | \Phi \rangle }.
\end{align}
The energy corrections for renormalized (R)-ecCCSD(T) and ecCCSD(T) can be obtained by taking
the lowest-order estimates of the correction and by assuming the denominator to be one, i.e.,
\begin{align}
\delta_0^{\rm R-ecCCSD(T)} & = \frac{ \langle \Psi^{\rm ecCCSD(T)} | Q^{\rm c}_3 (V_N T_2)_C | \Phi \rangle }{ \langle \Psi^{\rm ecCCSD(T)} | e^{T} | \Phi \rangle }, \label{eq:Recc} \\
\delta_0^{\rm ecCCSD(T)} & = \langle \Psi^{\rm ecCCSD(T)} | Q^{\rm c}_3 (V_N T_2)_C | \Phi \rangle.
\end{align}
Unlike perturbative triples without external correction (as in CCSD(T)), the approximate perturbative expression $T_3^{[2]}$ is only evaluated for determinants omitted in the external source. Thus it is not expected to diverge as long as the external source includes all degeneracies. Nonetheless, it can still overestimate the triples correlation. The role of the denominator in the "renormalized" triples approximations is to rescale this correction, which can be expected to
reduce the overestimation.
\section{Computational Details} \label{sec:compdetail}
All CC calculations were performed using cc-pVTZ basis sets.~\cite{dunning1989a}
The ecCC calculations were performed with the six different types of external sources summarized in Table~\ref{table:apprx_ext} and discussed in Section~\ref{sec:ea}. {The CASSCF-like external sources (type I and II) used natural orbitals in the minimal active spaces and CASSCF orbitals for the core and external spaces, while
the CASCI-like external sources (type III-VI) used HF orbitals.
}
All CC calculations were carried out using a local version of PySCF \cite{PYSCF} interfaced with StackBLOCK \cite{garnet2002, garnet2004, ghosh2008, sandeep2012} for DMRG and Arrow \cite{LiOttHolShaUmr-JCP-18, ShaHolJeaAlaUmr-JCTC-17, HolTubUmr-JCTC-16} for HCI.
We used Dice \cite{ShaHolJeaAlaUmr-JCTC-17, HolTubUmr-JCTC-16, smith2017} to get accurate SHCI PESs.
\section{Numerical results}
\subsection{PESs with type-I external sources} \label{sec:fm}
The dissociation PESs of H${}_2$O and N${}_2$, shown in Fig.~\ref{fig:fm}, were obtained by the ecCC methods using the type-I external sources in Table~\ref{table:apprx_ext} (i.e. near exact wave functions in the minimal active spaces and all amplitudes of $T_3$ and $T_4$).
As expected, the ecCC curves with tightly converged HCI and DMRG external sources are almost identical. Thus, we show the PESs of ecCC using only one of these external sources (colored solid lines) and the PESs of CC (colored dotted lines) in Fig. \ref{fig:fm}. These are compared against accurate PESs represented as black lines obtained by SHCI in the full space. The accurate energies from SHCI are given in the Supporting Information. The mean absolute errors (MAE) and the non-parallelity errors (NPE) of the PESs are listed in Table \ref{table:a}.
The CCSD curves (blue dotted lines) have an unphysical {dip due to an inadequate treatment of static correlation.}
This problem is completely eliminated within ecCCSD (blue solid line). However, there are significant errors with respect to the black curve at large distances, giving a (MAE,NPE) of ($24.0, 65.9$) m$E_{\rm H}$ for H${}_2$O and ($64.1, 149.9$) m$E_{\rm H}$ for N${}_2$.
We see that the (T) correction captures much of the missing dynamic correlation, such that the green ecCCSD(T) curves reach an accuracy of 1 mE${}_H$ around the equilibrium geometry. However, when the bond is stretched, the (T) correction overestimates the dynamic correlation, leading to another unphysical dip in the PESs.
Although this overestimation can easily be reduced by increasing the size of the active space for the
small systems treated here, this would be very expensive for large systems.
Alternatively, we can use the renormalized triples formula to damp the (T) correction in Eq. (\ref{eq:Recc}).
R-ecCCSD(T) (the solid red curves) completely removes the unphysical dips in the PESs.
These attain (MAE,NPE) = ($3.9, 23.1$) and ($12.8, 41.4$) m$E_{\rm H}$ for H${}_2$O and N${}_2$, respectively.
\begin{figure}[htp]
\centering
\caption{PESs of H${}_2$O and N${}_2$ in the top and bottom panels, respectively, obtained with the CC methods and the ecCC methods using the type-I external sources. The type-I sources correspond to near exact wave functions from the minimal active space, and use all the amplitudes of $T_3$ and $T_4$. The blue, green, and red solid lines are the ecCCSD, ecCCSD(T), and R-ecCCSD(T) PESs, respectively. These are to be compared with the PESs of CC represented as dotted lines. The black lines are accurate PESs obtained by SHCI.
}
\includegraphics[width=0.7 \textwidth,height=20cm,keepaspectratio]{Figure/rev_naming/full_minimal_active.pdf} \label{fig:fm}
\end{figure}
\subsection{PESs with type-II external sources} \label{sec:pm}
The difference between the type-I and type-II external sources is that the former use all the amplitudes of $T_3$ and $T_4$ while the latter use only a small number of the largest amplitudes, which contribute about 80\% of the total weight of $T_3$ and $T_4$ (see Table~\ref{table:apprx_ext} and Section~\ref{sec:ea}).
For the minimal active space of H${}_2$O, i.e. (4e,4o), the external source contains only one non-zero $T_4$ amplitude, and all elements of $T_3$ are zero to within numerical noise.
Thus, the PESs of ecCC using 80\% of the external amplitudes (type-II) are almost identical to those using 100\% of the external amplitudes (type-I).
On the other hand, for the minimal active space of N${}_2$, i.e. (6e,6o), the external source contains several
large $T_3$ and $T_4$ amplitudes.
At the stretched geometry corresponding to a bond length of $10 a_0$, 80\% of the total $T_3$ and $T_4$ amplitude weight is recovered by the nine largest elements of $T_4$.
Figure \ref{fig:pm} shows that the type-II external sources, although using fewer amplitudes, improve the PESs of ecCCSD and R-ecCCSD(T) relative to the type-I sources. The PES of R-ecCCSD(T) displays a (MAE,NPE) = ($5.4, 13.9$) m$E_{\rm H}$.
One concern with partial use of the amplitudes in the minimal active space is the possibility of divergence in (T).
Assuming a HF occupation of the natural orbitals, the recomputed orbital energies (diagonal parts of the Fock matrix) of the holes and those of the particles are such that the system retains a sizable gap. N${}_2$ at the $10.0a_0$ bond length, for example, has a gap of around $0.1$ Hartrees between the
three occupied orbitals and the three unoccupied orbitals.
This prevents the divergence of the (T) and the renormalized (T) corrections at the bond lengths we considered (although (T) largely overestimates dynamic correlation).
\begin{figure}[htp]
\centering
\caption{PESs of N${}_2$ obtained by the ecCC methods using all the external amplitudes (type-I) and some of the largest external amplitudes (type-II) from the minimal active space (solid and dotted lines, respectively). Other descriptions are the same as in Fig.\ref{fig:fm}}
\includegraphics[width=0.7\textwidth,height=20cm,keepaspectratio]{Figure/rev_naming/partial_minimal_active.pdf} \label{fig:pm}
\end{figure}
\subsection{PESs with type-III and type-IV external sources} \label{sec:fpa}
In addition, we investigated PESs of R-ecCCSD(T) using larger active spaces with all amplitudes (type-III) and 80\% of the amplitudes (type-IV).
We used active spaces of (8e,18o), (10e,16o), and (14e,16o) for H${}_2$O, N${}_2$, and F${}_2$, respectively, and obtained near exact wave functions in the spaces.
The PESs of ecCC using the type-III and type-IV external sources are shown as colored solid and dotted lines, respectively, in Figure \ref{fig:fpa}. The MAE and NPE of the PESs are given in Table \ref{table:a}.
For H${}_2$O and N${}_2$, the resulting PESs of R-ecCCSD(T) with all amplitudes (red solid lines in the top and middle panels) achieve (MAE,NPE) = ($2.9,18.0$) and ($12.1,44.3$) m$E_{\rm H}$, respectively. These are minor improvements compared to the results using the minimal active space external sources.
However, the PESs obtained using only 80\% of the external amplitudes (red dotted lines) are much closer to the black, accurate PES, and reach a (MAE,NPE) = ($1.1,8.1$) and ($4.6,8.4$) m$E_{\rm H}$.
{Similarly to in the minimal active space, 80\% of the amplitude weight in the larger active space corresponds to only one element of $T_4$ for H${}_2$O and the nine largest elements of $T_4$ for N${}_2$. }
When we can find such large elements, the partial usage of the amplitudes clearly has advantages in both accuracy and efficiency. It furthermore points to the importance of accounting for the error from missing relaxation in the external source amplitudes.
Unlike in H${}_2$O and N${}_2$, there are no particularly {large} elements in the $T_3$ and $T_4$ amplitudes of F${}_2$. To truncate the amplitudes to 80\% of their weight, we used the approximately 400 largest elements of $T_3$ and $T_4$ summing to 80\% of the total $T_3$ and $T_4$ weights at the bond length of $5.0$ \AA.
The red solid and dotted lines in the bottom panel of Figure \ref{fig:fpa} show the PESs of R-ecCCSD(T) using all and 80\% of the amplitudes, respectively.
These two PESs are very close to the accurate black curve and reach a (MAE,NPE) = ($0.9,1.4$) and ($1.7,1.6$) m$E_{\rm H}$ for 100\% and 80\% of the amplitudes, respectively.
Although both are accurate, the partial use of the external source in this case (where there is not a small number of large elements) leads to slightly worse accuracy.
\begin{figure}[htp]
\centering
\caption{PESs of H${}_2$O, N${}_2$, and F${}_2$ in the top, middle, and bottom panels, respectively, for the ecCC methods using the larger than minimal active spaces and the near exact external sources. The solid and dotted lines correspond to PESs obtained using all external amplitudes of $T_3$ and $T_4$ (type-III) and only the largest amplitudes of $T_4$ (type-IV), respectively.
The descriptions are otherwise the same as those in Fig.\ref{fig:pm}.}
\includegraphics[width=\textwidth,height=20cm,keepaspectratio]{Figure/rev_naming/additional_active.pdf} \label{fig:fpa}
\end{figure}
\begin{table}[htp]
\centering
\caption{MAE and NPE (m$E_{\rm H}$) of H${}_2$O, N${}_2$, and F${}_2$ PESs in a range of geometries $R \in [0.68, 7.80]$ \AA, $R \in [1.5, 10.0]$ $a_0$, and $R \in [1.14, 5.00]$ \AA, respectively.
ecCC methods using the near exact external sources with all elements of $T_3$ and $T_4$ amplitudes ("100\%") or only the largest elements of the $T_4$ amplitudes ("80\%") from the external source.
}
\begin{tabular}{lcccccccccc}\hline \hline
\multirow{2}{*}{Method} & \multirow{2}{*}{$T_3$, $T_4$}
& \multicolumn{3}{c}{H${}_2$O} & \multicolumn{3}{c}{N${}_2$} & \multicolumn{3}{c}{F${}_2$} \\ \cline{3-11}
& & Active & MAE & NPE & Active & MAE & NPE & Active & MAE & NPE \\ \hline
CCSD & & & 15.9 & 30.4 & & 32.4 & 102.3 & & 34.8 & 57.7 \\ \hline
\multirow{4}{*}{ecCCSD} & 100\% & \multirow{2}{*}{4e,4o} & 24.0 & 65.9 & \multirow{2}{*}{6e,6o} & 64.1 & 149.9 & \multirow{2}{*}{2e,2o} & \multirow{2}{*}{a} & \multirow{2}{*}{a} \\
& 80\% & & 24.0 & 65.9 & & 57.5 & 114.5 & & & \\ \cline{2-11}
& 100\% & \multirow{2}{*}{8e,18o} & 17.6 & 49.9 & \multirow{2}{*}{10e,16o} & 48.4 & 113.0 & \multirow{2}{*}{14e,16o} & 25.2 & 41.4 \\
& 80\% & & 15.6 & 39.4 & & 41.4 & 74.0 & & 27.0 & 40.9 \\ \hline
R-CCSD(T) & & & 6.8 & 43.0 & & 32.1 & 162.1 & & 6.2 & 10.6 \\ \hline
\multirow{4}{*}{R-ecCCSD(T)}& 100\% & \multirow{2}{*}{4e,4o} & 3.9 & 23.1 & \multirow{2}{*}{6e,6o} & 12.8 & 41.4 & \multirow{2}{*}{2e,2o} & \multirow{2}{*}{a} & \multirow{2}{*}{a}\\
& 80\% & & 3.9 & 23.1 && 5.4 & 13.9 &&&\\ \cline{2-11}
& 100\% & \multirow{2}{*}{8e,18o} & 2.9 & 18.0 & \multirow{2}{*}{10e,16o} & 12.1 & 44.3 & \multirow{2}{*}{14e,16o} & 0.9 & 1.4 \\
& 80\% & & 1.1 & 8.1 & & 4.6 & 8.4 & & 1.7 & 1.6 \\ \hline \hline
\end{tabular} \label{table:a}
${}^{\rm a}$ no $T_3$ and $T_4$ in the minimal active space of F${}_2$
\end{table}
\subsection{PESs with type-V external sources} \label{sec:fpal}
In the previous section, we showed that the use of larger active spaces significantly improves the ecCC PESs.
However, obtaining tightly converged DMRG wave functions in large active spaces requires more
CPU time than the subsequent CC calculation.
For example, optimizing an external wave function with 16 orbitals requires around a few minutes of CPU time for one DMRG sweep with $M=2000$.
Although a few minutes is not prohibitively large in many applications, it is large compared to the subsequent ecCC calculation which only takes tens of seconds, at least for the small molecules considered here. In addition, the fact that in some cases only a partial use of the amplitudes led to better results in the last section suggests that it is not a good use of computational time to tightly converge the external source.
We thus now consider loosely converged DMRG sources (type-V) in the larger active spaces. We re-computed PESs of R-ecCCSD(T) shown in Figure \ref{fig:fpa} using a DMRG source with bond dimensions 25, 50, and 100. We used the truncation to 80\% amplitude weight for H${}_2$O and N${}_2$, while we used all the external amplitudes for F${}_2$.
Table \ref{table:loose} shows the corresponding MAE and NPE in the same range of geometries in Table \ref{table:a}.
For all cases, when we reduced the bond dimension to $M=100$, the MAE increased by $0.1 \sim 0.5$ m$E_H$, and the NPE increased by $0.3 \sim 1.0$ m$E_H$, compared to using $M=2000$. However for $M=100$, one DMRG sweep with 16 orbitals took only a few seconds of CPU time at the 10$a_0$ bond length of N${}_2$, giving a better computational balance between the DMRG calculation and the subsequent ecCC calculations. When we further reduced the bond dimension to $M=25$, the MAE increased by $0.9$ m$E_H$ and the NPE increased by $0.6$ and $2.7$ m$E_H$ for H${}_2$O and N${}_2$.
In the case of F${}_2$, which does not have a small number of large $T_3$ and $T_4$ elements, the MAE and NPE increased more, by $3.4$ and $5.4$ m$E_H$.
\subsection{PESs with type-VI external sources} \label{sec:vi}
Type VI sources use the full orbital space and employ the single HCI parameter, $\epsilon$, to select the large $T_3$ and $T_4$ amplitudes.
They do not require a choice for the CAS space, and we did not further threshold the $T_3$ and $T_4$ amplitudes.
However, for systems where the molecule has a low-spin ground state, but the dissociated
fragments have high-spin ground states, the HCI wave functions at stretched geometries
are strongly spin-contaminated when $\epsilon$ is large, even when time-reversal symmetry
is employed to reduce the spin contamination. (This type of spin-contamination is avoided in the small bond dimension DMRG calculations via the full use of spin symmetry).
Although this has little effect on the HCI energy, it makes the amplitudes unsuitable for ecCC calculations.
In the second part of Table \ref{table:loose} the MAE and NPE from type-VI sources using
$\epsilon= 0.01$ and $0.003$ are shown. These are evaluated over a smaller range of
geometries than those used in the first part of Table \ref{table:loose} for the reason mentioned
above.
Since N$_2$ is a singlet that dissociates into atoms that are quartets, it has particularly
large errors. The errors improve upon going from $\epsilon=0.01$ to $\epsilon=0.003$, in particular
the PEC has an unphysical dip at $\epsilon=0.01$ which disappears at $\epsilon=0.003$, but
the errors are still substantial.
Similarly to how using all the amplitudes could lead to larger errors than only using some of the amplitudes in the previous sections, the errors incurred from the type VI sources here emphasize that the quality of the external amplitudes cannot be judged solely from the energy obtained by the external method.
\begin{table}[htp]
\centering
\caption{MAE and NPE (m$E_{\rm H}$) of R-ecCCSD(T) PESs for H${}_2$O, N${}_2$, and F${}_2$ using the approximate DMRG and HCI sources with a small bond dimension ($M$) or large threshold ($\epsilon$).
Detailed explanation to be found in the main text.}
\begin{tabular}{lccccccccc}\hline \hline
\multicolumn{1}{c}{\multirow{2}{*}{$M$}} & \multicolumn{3}{c}{H${}_2$O} & \multicolumn{3}{c}{N${}_2$} & \multicolumn{3}{c}{F${}_2$} \\ \cline{2-10}
& Active & MAE ${}^a$ & NPE ${}^a$ & Active & MAE ${}^a$ & NPE ${}^a$ & Active & MAE ${}^a$ & NPE ${}^a$ \\ \hline
25 & \multirow{4}{*}{(8e,18o)} & 3.8 & 8.7 & \multirow{4}{*}{(10e,16o)} & 5.5 & 11.1 & \multirow{4}{*}{(14e,16o)} & 4.3 & 6.8 \\
50 & & 3.4 & 8.2 & & 5.2 & 9.6 & & 3.5 & 3.4 \\
100 & & 3.0 & 8.4 & & 4.9 & 9.4 & & 1.4 & 2.1 \\
2000 & & 2.9 & 8.1 & & 4.6 & 8.4 & & 0.9 & 1.4 \\ \hline
\multicolumn{1}{c}{\multirow{2}{*}{$\epsilon$}} & \multicolumn{3}{c}{H${}_2$O} & \multicolumn{3}{c}{N${}_2$} & \multicolumn{3}{c}{F${}_2$} \\ \cline{2-10}
& Active & MAE ${}^b$ & NPE ${}^b$ & Active & MAE ${}^b$ & NPE ${}^b$ & Active & MAE ${}^b$ & NPE ${}^b$ \\ \hline
$0.01$ & \multirow{2}{*}{(10e,58o)} & 1.6 & 7.1 & \multirow{2}{*}{(14e,60o)} & 8.1 & 61.5 & \multirow{2}{*}{(14e,58o)}
& 3.6 & 4.7 \\
$0.003$ & & 1.3 & 4.4 & & 3.6 & 15.8 & & 3.0 & 5.8 \\ \hline \hline
\end{tabular} \label{table:loose}
${}^a$ $R \in [0.68, 7.80]$ \AA, $[1.5, 10.0]$ $a_0$, and $[1.14, 5.00]$ \AA $\ $ for H${}_2$O, N${}_2$, and F${}_2$ PESs, respectively. \\
${}^b$ $R \in [0.68, 3.41]$ \AA, $[1.5, 5.4]$ $a_0$, and $[1.14, 2.80]$ \AA $\ $ for H${}_2$O, N${}_2$, and F${}_2$ PESs, respectively.
\end{table}
\subsection{Error analysis} \label{sec:diag}
\begin{figure}[htp]
\centering
\caption{Absolute errors of R-CCSD(T) and ecCC methods on a log scale plotted against the $D_2$ diagnostic. Detailed explanation in the main text. }
\includegraphics[width=\textwidth,height=\textheight,keepaspectratio]{Figure/rev_naming/diag_d2.pdf} \label{fig:diag}
\end{figure}
In this section, we present the errors of R-ecCCSD(T) for the systems in this work and analyze them using the well known CC error diagnostic $D_2$, defined by the matrix 2-norm of the $T_2$ amplitudes.~{\cite{nielsen1999double}} The magnitude of $D_2$ can be used to distinguish between the SR and MR character of the different geometries on the PESs.
Organic molecules are sometimes considered to have MR character when $D_2$ is larger than $0.18$.~{\cite{nielsen1999double}}
Figure \ref{fig:diag} shows the absolute errors on a log scale versus the $D_2$ diagnostic.
Each symbol represents a geometry on the PESs of one of the molecules, H${}_2$O, N${}_2$, and F${}_2$.
Square symbols denote R-CCSD(T) and circle symbols denote the various externally corrected theories, using the type-V and the type-VI external sources.
For $D_2$ ranging from $0$ to $1.7$, the absolute errors of R-ecCCSD(T) (red circles) are less than $0.015$ E${}_H$ and most of the errors are smaller than those of R-CCSD(T) (black squares) and ecCCSD (blue circles).
The absolute errors of R-ecCCSD(T) are less than those of ecCCSD(T) (green circles) in the MR region where $D_2 > 0.4$, while they are {mostly} greater than those of ecCCSD(T) in the range $0.0 < D_2 < 0.4$.
Overall, it is clear to see that R-ecCCSD(T) offers the most balanced treatment of errors across a wide range of SR and MR character. However, for very weakly correlated systems, the original (T) correction is slightly more accurate than the renormalized (T) correction.
\subsection{Limitations} \label{sec:limit}
Although the work above shows that it is possible to obtain quantitative accuracy across the full potential energy surface, at a reasonable computational cost, using the combination of external sources and the renormalized (T) correction, SR ecCC approaches have a fundamental limitation when the CI coefficient of the reference configuration (the HF configuration in this work) in the external source is exactly or numerically zero.
This leads to overflow due to the assumption of intermediate normalization.
The smallest reference coefficient value we encountered in this work was $0.2$ at the stretched $10a_0$ geometry of N${}_2$.
{Although this coefficient is not numerically zero,}
it is difficult to converge the ecCCSD energy. We extracted initial amplitudes of $T_1$ and $T_2$ from the external source and then iteratively converged the ecCCSD energy using a damping parameter of $0.05$ to update the amplitudes. (We did not use the direct inversion in the iterative subspace (DIIS) algorithm).
At this geometry, the energy could be converged monotonically up to a threshold of $10^{-5}$ Hartrees, although the norm of the amplitudes could not be converged, and we simply used amplitudes from the last iteration in the set of monotonically decreasing energies.
\section{Conclusion} \label{sec:conclusions}
In this work, we explored externally corrected coupled cluster with a renormalized triples correction (R-ecCCSD(T)) using DMRG and HCI and external sources. The critical question is how to best balance the accuracy and cost of computing the external source with the cost of the overall method. To this end, we considered multiple types of external sources: "exact" external sources, where the DMRG and HCI wavefunctions were tightly converged, and "approximate" external sources where they were loosely converged.
We also considered both "full" usage of the $T_3$ and $T_4$ amplitudes,
and "partial" usage, wherein we retained only the largest elements.
For all systems considered here, we found that R-ecCCSD(T) can significantly improve on the description of the potential energy surface given either by the external source alone or CC alone. For example, the unphysical dips in the PES of H${}_2$O and N${}_2$ in the bond-stretched region can be completely eliminated.
The use of approximate external sources, possibly with truncation to only the {large} $T_3$ and $T_4$ amplitudes, appears to be a practical way to balance the cost of the external calculation and the coupled cluster calculation in small molecules.
Using the $D_2$ diagnostic to characterize the different points on the potential energy surfaces, we find that R-ecCCSD(T) gives absolute errors of less than 15 m$E_H$ in the range of $D_2$ from $0.0$ to $1.7$. In fact, the errors of R-ecCCSD(T) are less than those of ecCCSD and R-CCSD(T) in almost all cases, except when $D_2$ is very small, where the renormalized (T) correction appears to be slightly less accurate than the simple (T) correction.
There are several interesting questions remaining which lie beyond what we have considered in this work. For example, while R-ecCCSD(T) appears quite stable up to large values of the $D_2$ diagnostic, what is the largest amount of multi-reference character which can be handled? Here, the difficulty in solving the CC equations, and the divergence of the amplitudes reflecting the problems of intermediate normalization, cannot be ignored. In addition, in the realm of quasidegenerate problems, we can ask whether other non-iterative corrections such as the "completely renormalized" triples and quadruples corrections,~\cite{kowalski2000mmcc} corresponding to CR-ecCCSD(T) and CR-ecCCSD(T,Q), would further improve on the present R-ecCCSD(T) method. Finally, the current approximation, with its modest computational requirements on the external source, is applicable to the same scale of systems that can be handled by single reference coupled cluster methods. Thus understanding the performance of this method in larger correlated systems is of interest.
\textit{Note}: As this work was being prepared for submission, we were made aware of a related recent submission to the arxiv~\cite{magoulas2021externally}, that also discusses selected configuration interaction as an external source and perturbative triples corrections in the context of externally corrected coupled cluster methods.
\begin{acknowledgement}
Work by SL, HZ and GKC was supported by the US National Science Foundation via Award CHE-1655333. GKC is a Simons Investigator.
SS was supported by the US National Science Foundation grant CHE-1800584 and the Sloan research fellowship.
CJU was supported in part by the AFOSR under grant FA9550-18-1-0095.
\end{acknowledgement}
\begin{suppinfo}
\end{suppinfo}
\clearpage
|
\section{Introduction}
Exclusive lepto-production of vector mesons $e p\to e'\gamma^* p \to e'p'V$ is a versatile process that can address a wide spectrum of key questions about the proton structure. Depending on the $\gamma^*p$
center-of-mass energy $W$, photon virtuality $Q^2$ and meson species, the spacetime picture of the reaction looks different and requires different theoretical frameworks. At high energy in the large-$Q^2$ region, and for the longitudinally polarized virtual photon, a QCD factorization theorem \cite{Collins:1996fb} dictates that the scattering amplitude can be written in terms of the generalized parton distribution (GPD), the meson distribution amplitude (DA) and the perturbatively calculable hard part. Even in cases where such a rigorous treatment is not available, many phenomenologically successful models exist. Generally speaking, light vector mesons probe the up and down quark contents of the proton, whereas in the Regge-regime $W\to \infty$, or for heavy vector mesons such as quarkonia ($J/\psi,\Upsilon,...$), the process is primarily sensitive to the gluonic content of the proton.
The $\phi$-meson (a bound state of $s\bar{s}$) has a somewhat unique status in this context because its mass $m_\phi=1.02$ GeV, being roughly equal to the proton mass $m_N=0.94$ GeV, is neither light nor heavy. In the literature, $\phi$-production is often discussed on a similar footing as quarkonium production. Namely, since the proton does not contain valence $s$-quarks, the $\phi$-proton interaction proceeds mostly via gluon exchanges. However, the proton contains a small but non-negligible fraction of strange sea quarks already on nonperturbative scales of $\sim 1 \, {\rm GeV}^2$. It is then a nontrivial question whether the gluon exchanges, suppressed by the QCD coupling $\alpha_s$, always dominate over the $s$-quark exchanges.
In this paper, we take a fresh look at the lepto-production of $\phi$-mesons $e p \to e'p'\phi$ in the high-$Q^2$ region {\it near threshold}, namely when $W$ is very low and barely enough to produce a $\phi$-meson $W\gtrsim m_\phi+m_N= 1.96$ GeV. Measurements in the threshold region have been performed at LEPS \cite{Mibe:2005er,Chang:2010dg,Hiraiwa:2017xcs} for photo-production $Q^2\approx 0$ and at the Jefferson laboratory (JLab) for both photo- and lepto-productions $4\gtrsim Q^2 \gtrsim 0$ \cite{Santoro:2008ai,Qian:2010rr,Dey:2014tfa}. One of the main motivations of these experiments was to study the nature of gluon (or Pomeron) exchanges. Our motivation however is drastically different, so let us quickly provide relevant background information:
It has been argued
that the near-threshold photo-production of a heavy quarkonium (see
\cite{Ali:2019lzf} for an ongoing experiment) is sensitive to the gluon gravitational form factors $\langle P'|T_g^{\mu\nu}|P\rangle$ where $T_g^{\mu\nu}$ is the gluon part of the QCD energy momentum tensor \cite{Frankfurt:2002ka,Hatta:2018ina,Hatta:2019lxo} (see also \cite{Xie:2019soz,Mamo:2019mka}). In lepto-production at high-$Q^2$, this connection can be cleanly established by using the operator product expansion (OPE) \cite{Boussarie:2020vmu}. The analysis \cite{Frankfurt:2002ka} of $J/\psi$ photo-production over a wide range of $W$ indicates that the $t$-dependence of these form factors (or that of the gluon GPDs at high energy $W$) is weaker than what one would expect from the $Q^2$-dependence of the electromagnetic form factors. This suggests that the gluon fields in the nucleon are more compact and localized than the charge distribution.
Moreover, it has been demonstrated \cite{Hatta:2018ina,Hatta:2019lxo,Boussarie:2020vmu} that the shape of the differential cross section $d\sigma/dt$ is sensitive to the gluon D-term $D_g(t)$ which, after Fourier transforming to the coordinate space, can be interpreted as an internal force (or `pressure') exerted by gluons inside the proton \cite{Polyakov:2018zvc} (see however, \cite{Freese:2021czn}). The experimental determination of the parameter $D_g(t=0)$ is complementary to the ongoing effort \cite{Burkert:2018bqq,Kumericki:2019ddg,Dutrieux:2021nlz} to extract the $u,d$-quark contributions to the D-term $D_{u,d}$ from the present and future deeply virtual Compton scattering (DVCS) experiments. Together they constitute the total D-term
\begin{eqnarray}
D(0)=D_u(0)+D_d(0) + D_s(0) + D_g(0)+\cdots. \label{mis}
\end{eqnarray}
Only the sum is conserved (renormalization-group invariant) and represents a fundamental constant of the proton.
The strangeness contribution to the D-term $D_s$ has received very little, if any, attention in the literature so far. However, a large-$N_c$ argument \cite{Goeke:2001tz} suggests the approximate flavor independence of the D-term $D_u\approx D_d$ (nontrivial because $u$-quarks are more abundant than $d$-quarks in the proton) which by extension implies that $D_s$ could be comparable to $D_{u,d}$ at least in the flavor SU(3) symmetric limit. It is thus a potentially important piece of the sum rule (\ref{mis}) when the precision study of the D-term becomes possible in future.
The purpose of this paper is to argue that the lepto-production of $\phi$-mesons near threshold is governed by the strangeness gravitational form factors including $D_s$. By doing so, we basically postulate that the $\phi$-meson couples more strongly to the $s\bar{s}$ content of the proton than to the gluon content at least near the threshold at high-$Q^2$, contrary to the prevailing view in the literature.
Also, by using the local version of the OPE following Ref.~\cite{Boussarie:2020vmu}, we do not work in the collinear factorization framework \cite{Collins:1996fb} whose applicability to the threshold region does not seem likely. In the appendix, we briefly discuss the connection between the two approaches. Other approaches to $\phi$-meson photo- or lepto-production near threshold can be found in \cite{Laget:2000gj,Titov:2003bk,Strakovsky:2020uqs}.
In Section II, we collect general formulas for the cross section of lepto-production. In Section III, we describe our model of the scattering amplitude inspired by the OPE and the strangeness gravitational form factors. The numerical results for the differential cross section $d\sigma/dt$ are presented in Section IV for the kinematics relevant to the JLab and the future electron-ion collider (EIC).
\section{Exclusive $\phi$-meson leptoproduction}
Consider exclusive $\phi$-meson production cross section in electron-proton scattering $ep\to e'\gamma^* p\to e'p'\phi$. The $ep$ and $\gamma^*p$ center-of-mass energies are denoted by $s_{ep}=(\ell+P)^2$ and $W^2=(q+P)^2$, respectively. The outgoing $\phi$ has momentum $k^\mu$ with $k^2=m_\phi^2$. The cross section can be written as
\begin{eqnarray}
\frac{d\sigma}{dWdQ^2}&=& \frac{\alpha_{em}^2}{4\pi} \frac{1}{16 (P\cdot \ell)^2 Q^4 P_{cm}}\int \frac{d\phi_\ell}{2\pi} L_{\mu\nu}\int dt \,\frac{1}{2}\sum_{spin}\langle P|J_{em}^\mu(-q)|P'\phi\rangle\langle P'\phi|J_{em}^\nu(q)|P\rangle ,
\end{eqnarray}
where $J_{em}$ is the electromagnetic current operator, $m_N$ is the proton mass and $P_{cm}$ is the proton momentum in the $\gamma^*p$ center-of-mass frame
\begin{eqnarray}
P_{cm}=\frac{\sqrt{W^4-2W^2(m_N^2-Q^2)+(m_N^2+Q^2)^2}}{2W}.
\end{eqnarray}
For simplicity, we average over the outgoing lepton angle $\phi_\ell$, but its dependence can be restored \cite{Arens:1996xw} if need arises.
We can then write, in a frame where the virtual photon is in the $+z$ direction,
\begin{eqnarray}
\int \frac{d\phi_\ell}{2\pi} L^{\mu\nu}=\frac{2Q^2}{1-\epsilon} \left(\frac{1}{2}g^{\mu\nu}_\perp +\epsilon \varepsilon^\mu_L \varepsilon^\nu_L \right) \label{diehl}
\end{eqnarray}
where $g_\perp^{ij}=\delta^{ij}$ and
\begin{eqnarray}
\varepsilon_L^\mu(q) = \frac{1}{Q\sqrt{1+\gamma^2}} \left(q^\mu+ \frac{Q^2}{P\cdot q}P^\mu\right).
\end{eqnarray}
is the polarization vector of the longitudinal virtual photon.
$\epsilon$ is the longitudinal-to-transverse photon flux ratio
\begin{eqnarray}
\epsilon=\frac{1-y-\frac{y^2\gamma^2}{4}}{1-y+\frac{y^2}{2}+\frac{y^2\gamma^2}{4}}, \qquad \frac{1}{1-\epsilon}= \frac{2}{y^2}\frac{1-y+\frac{y^2}{2}+\frac{y^2\gamma^2}{4}}{1+\gamma^2}
\end{eqnarray}
where
\begin{eqnarray}
y\equiv \frac{P\cdot q}{P\cdot \ell}, \qquad \gamma\equiv\frac{2x_Bm_N}{Q} = \frac{m_NQ}{P\cdot q}
\end{eqnarray}
The parameter $\gamma$ accounts for the target mass correction which should be included in near-threshold production.
The variable $y$ can be eliminated in favor of $s_{ep}$ using the relation
\begin{eqnarray}
W^2=y(s_{ep}-m_N^2)+m_N^2-Q^2. \label{y}
\end{eqnarray}
In the hadronic part, let us define
\begin{eqnarray}
\frac{1}{2}\sum_{spin} \langle P|J_{em}^\mu(-q)|P'\phi\rangle \langle P'\phi|J_{em}^\nu(q)|P\rangle \equiv - ( {\cal M}_{\rho}^{\ \mu})^*{\cal M}^{\rho\nu},
\end{eqnarray}
where the minus sign is from the vector meson polarization sum $\sum \varepsilon_V^\rho \varepsilon_V^{\rho'}=-g^{\rho\rho'}+\frac{k^\rho k^{\rho'}}{m_\phi^2}$.
(The amplitude ${\cal M}$ satisfies the conditions $k_\rho {\cal M}^{\rho\mu}= {\cal M}^{\rho\mu}q_\mu=0$.)
Contracting with the lepton tensor (\ref{diehl}), we get
\begin{eqnarray}
-\left(\frac{1}{2}g^{\mu\nu}_\perp +\epsilon \varepsilon^\mu_L \varepsilon^\nu_L \right) {\cal M}^*_{\rho\mu}{\cal M}^\rho_{\ \nu} &=& -\left(\frac{1}{2}g^{\mu\nu}_\perp +\frac{\epsilon}{1+\gamma^2}\frac{Q^2}{(P\cdot q)^2}P^\mu P^\nu \right) {\cal M}^*_{\rho\mu}{\cal M}^\rho_{\ \nu} \nonumber \\
&=& \left(\frac{1}{2}g^{\mu\nu} -\frac{\left(\frac{1}{2}+\epsilon\right)Q^2}{(1+\gamma^2)(P\cdot q)^2}P^\mu P^\nu\right) {\cal M}^*_{\rho\mu}{\cal M}^\rho_{\ \nu} ,
\label{need}
\end{eqnarray}
where we used
\begin{eqnarray}
g_\perp^{\mu\nu}=-g^{\mu\nu} + \varepsilon_L^\mu \varepsilon_L^\nu -\frac{q^\mu q^\nu}{Q^2} .
\end{eqnarray}
We thus arrive at
\begin{eqnarray}
\frac{d\sigma}{dWdQ^2}=\frac{\alpha_{em}}{4\pi}\frac{W(W^2-m_N^2)}{(P\cdot \ell)^2 Q^2(1-\epsilon)} \int dt \frac{d\sigma}{dt}, \label{epcross}
\end{eqnarray}
with\footnote{ In Ref.~\cite{Boussarie:2020vmu}, the authors calculated the spin-averaged cross section $g^{\mu\nu} {\cal M}^*_{\rho\mu}{\cal M}^\rho_{\ \nu}$. To directly compare with the lepto-production data, one should rather use the formula (\ref{need}). }
\begin{eqnarray}
\frac{d\sigma}{dt}= \frac{\alpha_{em}}{8(W^2-m_N^2)WP_{cm}} \left(\frac{1}{2}g^{\mu\nu} -\frac{\left(\frac{1}{2}+\epsilon\right)Q^2}{(1+\gamma^2)(P\cdot q)^2}P^\mu P^\nu\right) {\cal M}^*_{\rho\mu}{\cal M}^\rho_{\ \nu} . \label{dt}
\end{eqnarray}
Note that the $\gamma^*p$ cross section (\ref{dt}) depends on the $ep$ center-of-mass energy $s_{ep}$ because $\epsilon$ depends on $y$, and $y$ depends on $s_{ep}$ as in (\ref{y}).
Experimentalists at the JLab have measured $\frac{d\sigma}{dW dQ^2}$ \cite{Santoro:2008ai}, and from the data they have reconstructed the differential cross section (\ref{dt}) and the total cross section $\sigma(W,Q^2)=\int dt\, \frac{d\sigma}{dt}$. In the next section we present a model for the scattering amplitude ${\cal M}$.
\section{Description of the model}
Our model for the matrix element $\langle P'\phi(k)|J_{em}^\nu(q)|P\rangle$ has been inspired by the recently developed new approach to the near-threshold production of heavy quarkonia such as $J/\psi$ and $\Upsilon$ \cite{Boussarie:2020vmu}.
The main steps of \cite{Boussarie:2020vmu} are summarized as follows. One first relates the $J/\psi$ production amplitude $\langle P'J/\psi(k)|J_{em}^\nu(q)|P\rangle$ to the correlation function of the charm current operator $J_c^\mu = \bar{c}\gamma^\mu c$
\begin{eqnarray}
\int dxdy e^{ik\cdot x -iq\cdot y}\langle P'|{\rm T}\{J_c^\mu(x)J_c^\nu (y)\} |P\rangle,
\end{eqnarray}
in a slightly off-shell kinematics $k^2\neq m_{J/\psi}^2$. One then
performs the operator product expansion (OPE) in the regime
$Q^2\gg |t|,m_{J/\psi}^2$ and picks up gluon bilinear operators $\sim FF$.
The off-forward matrix elements $\langle P'|FF|P\rangle$ are parameterized by the gluon gravitational form factors. These include the gluon momentum fraction $A_g$ (the second moment of the gluon PDF), the gluon D-term $D_g$ and the gluon condensate (trace anomaly) $\langle F^{\mu\nu}F_{\mu\nu}\rangle$.
When adapting this approach to $\phi$-production, we recognize a few important differences. First, we need to keep $s$-quark bilinear operators $\sim \bar{s}s$ rather than gluon bilinears. A quick way to estimate their relative importance in the present approach is to compare the momentum fractions $A_{s+\bar{s}}$ and $\frac{\alpha_s}{2\pi}A_g$. Taking $A_{s+\bar{s}}\approx 0.04$, $A_g\approx 0.4$ \cite{Maguire:2017ypu} and $\alpha_s=0.2\sim 0.3$ for example, we see that the strange sea quarks are more important than gluons, though not by a large margin. In the appendix, we give a slightly improved argument and show that the $s$-quark contribution gets an additional factor $\sim 2$. Second, the condition $Q^2\gg |t|$ is more difficult to satisfy.\footnote{This condition is needed in order to ensure that the large momentum $Q$ does not flow into the nucleon vertex so that one can perform the $JJ$ OPE. } For example, the momentum transfer at the threshold is
\begin{eqnarray}
|t_{th}|=\frac{m_N(m_V^2+Q^2)}{m_N+m_V}.
\end{eqnarray}
When $Q^2\gg m_V^2$, $|t_{th}|$ is a larger fraction of $Q^2$ in $\phi$-production $m_V=m_\phi\approx m_N$ than in $J/\psi$-production $m_V=m_{J/\psi}\approx 3m_N$. As one goes away (but not too far away) from the threshold, the region $Q^2\gg |t|$ does exist. In principle, our predictions are limited to such regions, though in practice they can be smoothly extrapolated to $|t|>Q^2$ as long as $|t|$ is not too small.
We now perform the OPE. A simple calculation shows
\begin{eqnarray}
{\cal A}_s^{\mu\nu}&\equiv &i\int d^dr e^{ir\cdot q} \Bigl(\bar{s}(0)\gamma^\mu S(0,-r)\gamma^\nu s(-r) + \bar{s}(-r)\gamma^\nu S(-r,0)\gamma^\mu s(0) \Bigr) \nonumber \\
&=& \frac{-4}{(q^2-m_s^2)^2} \left( q^2g^{\mu\alpha}g^{\nu\beta}-g^{\mu\alpha}q^\nu q^\beta -g^{\nu\beta}q^\mu q^\alpha + g^{\mu\nu}q^\alpha q^\beta \right)T_{\alpha\beta}^s + \cdots \nonumber \\
&\to& \frac{-4}{(q^2-m_s^2)^2}
(q\cdot k g^{\mu\alpha}g^{\nu\beta} -g^{\mu\alpha}k^\nu q^\beta -g^{\nu\beta} q^\mu k^\alpha + k^\alpha q^\beta g^{\mu\nu})T^s_{\alpha\beta}+\cdots,
\label{last}
\end{eqnarray}
where
\begin{eqnarray}
S(0,-r) = \int \frac{d^d q}{(2\pi)^d} e^{-iq\cdot r} \frac{i(\Slash q+m_s)}{q^2-m_s^2},
\end{eqnarray}
is the $s$-quark propagator and
\begin{eqnarray}
T^s_{\alpha\beta}= i\bar{s}\gamma_{(\alpha} \overleftrightarrow{D}_{\beta)} s. \label{em}
\end{eqnarray}
is the $s$-quark contribution to the energy momentum tensor ($\overleftrightarrow{D}\equiv \frac{D-\overleftarrow{D}}{2}$). We neglect the $s$-quark mass $m_s$ whenever it appears in the numerator. However, we keep it in the denominator to regularize the divergence $\frac{1}{q^2-m_s^2}\sim \frac{1}{Q^2+m_s^2}$ just in case we may want to extrapolate our results to smaller $Q^2$ values in future applications.
In the last line of (\ref{last}), we have implemented minimal modifications to make ${\cal A}_s^{\mu\nu}$ transverse with respect to both $q^\nu$ and $k^\mu$ as required by gauge invariance. While this is {\it ad hoc} at the present level of discussion, we anticipate that total derivative/higher twist operators restore gauge invariance, similarly to what happens in deeply virtual Compton scattering (DVCS) \cite{Anikin:2000em}
Of course, even after restricting ourselves to quark bilinears, there are other operators that can contribute to (\ref{last}). Potentially important operators include the axial vector operator $\bar{s}\gamma_\mu \gamma_5 s$ and the $s$-quark twist-two operators with higher spins. The former is related to the (small) $s$-quark helicity contribution $\Delta s$ to the proton spin in the forward limit. Its off-forward matrix element is basically unknown. The latter are discussed in the appendix where it is found that, unlike in the quarkonium case \cite{Boussarie:2020vmu}, the twist-two, higher-spin operators are not negligible for the present problem. To mimic their effect, we introduce an overall phenomenological factor of 2.5 in (\ref{last}).
To evaluate the matrix element $\langle P'|{\cal A}_s^{\mu\nu}|P\rangle$, we use the following parameterization of the gravitational form factors \cite{Kobzarev:1962wt,Ji:1996ek}
\begin{eqnarray}
\langle P'|T_s^{\alpha\beta}|P\rangle = \bar{u}(P')\left[A_s(t)\gamma^{(\alpha}\bar{P}^{\beta)}+B_s(t)\frac{\bar{P}^{(\alpha}i\sigma^{\beta)\lambda}\Delta_\lambda}{2m_N} + D_s(t)\frac{\Delta^\alpha \Delta^\beta-g^{\alpha\beta}\Delta^2}{4m_N} + \bar{C}_s (t)m_N g^{\alpha\beta}\right]u(P), \label{mom}
\end{eqnarray}
where $\bar{P}=\frac{P+P'}{2}$, $\Delta^\mu=P'^\mu - P^\mu$ and $t=\Delta^2$.
$D(t)/4$ is often denoted by $C(t)$ in the literature. We neglect $B_s$ following the empirical observation that the flavor-singlet $B_{u+d}$ is unusually small (see, e.g., \cite{Hagler:2003jd}).
We further set $\bar{C}_s=-\frac{1}{4}A_s$ assuming that the trace anomaly is insignificant in the strangeness sector. [However, this point may be improved as was done for gluons in \cite{Hatta:2019lxo,Boussarie:2020vmu}.] For the remaining form factors, we employ the dipole and tripole ansatze suggested by the perturbative counting rules at large-$t$ \cite{Tanaka:2018wea,Tong:2021ctu}
\begin{eqnarray}
A_{s}(t)=\frac{A_{s}(0)}{(1-t/m_A^2)^2}, \qquad D_s(t)
= \frac{D_s(0)}{(1-t/m_D^2)^3}, \label{an}
\end{eqnarray}
with $A_s(0)=A_{s+\bar{s}}=0.04$ as mentioned above.
We use the same effective masses $m_A=1.13$ GeV and $m_D=0.76$ GeV as for the gluon gravitational form factors from lattice QCD \cite{Shanahan:2018nnv}. This is reasonable given that $s$-quarks in the nucleon are generated by the gluon splitting $g\to s\bar{s}$.\footnote{In \cite{Frankfurt:2002ka} and more recently in \cite{Wang:2021dis}, the authors fitted the $\phi$-meson photo- and lepto-production data using the form $d\sigma/dt \propto A^2(t)\propto 1/(1-t/m_A^2)^4$ and found that the mass parameter $m_A$ is consistent with the $J/\psi$ case. This partially supports our procedure. } The value $D_s(0)$ is our main object of interest, and is treated here as a free parameter. As mentioned in the introduction, even though $s$-quarks are much less abundant in the proton $A_s\ll A_{u,d}$, a large-$N_c$ argument suggests that the D-terms are `flavor blind' $D_s\sim D_u\approx D_d$ \cite{Goeke:2001tz}.\footnote{The prediction $|D_u+D_d|\gg |D_u-D_d|$ from large-$N_c$ QCD is supported by the lattice simulations \cite{Hagler:2003jd}. Interestingly, and in contrast, the $B$-form factor is dominantly a flavor nonsinglet quantity $|B_u-B_d|\gg |B_u+B_d|$ as already mentioned.
} In the flavor SU(3) limit, and at asymptotically large scales, the relation ($C_F=\frac{N_c^2-1}{2N_c}=4/3$)
\begin{eqnarray}
D_s(0) \approx \frac{1}{4C_F} D_g(0),
\end{eqnarray}
together with the recent lattice result for the gluon D-term $D_g(0)\approx -7.2$ \cite{Shanahan:2018nnv} gives $D_s(0)\approx -1.35$. We thus vary the parameter in the range $0>D_s(0)>-1.35$, with a particular interest in the possibility that $|D_s|$ is of order unity. Note that this makes $\phi$-production rather special, compared to light or heavy meson productions. If $A_s\ll A_{u,d,g}$ but $D_s\sim D_{u,d}$, the effect of the D-term will be particularly large in the strangeness sector.
Finally, the proportionality constant between ${\cal M}^{\mu\nu}$ and $\langle P'|{\cal A}_s^{\mu\nu}|P\rangle$ can be determined similarly to the $J/\psi$ case (see Eqs.~(48,49) of \cite{Boussarie:2020vmu}). Using the
$\phi$-meson mass $m_\phi=$1.02 GeV and its leptonic decay width $\Gamma_{e^+e^-}=1.27$ keV,
we find
\begin{eqnarray}
\frac{e_s^4e^2m_\phi^4}{g_{\gamma\phi }^2 } = \frac{4\pi e_s^4 \alpha^2_{em}m_\phi}{3\Gamma_{e^+e^-}} \approx 2.21, \label{overall}
\end{eqnarray}
where $e_s=-1/3$ and $g_{\gamma \phi}$ is the decay constant.
There is actually an uncertainty of order unity in the overall normalization of the amplitude as mentioned in \cite{Boussarie:2020vmu} (apart from the factor of 2.5 mentioned above).
This can be fixed by fitting to the total cross section data. Then the shape of $d\sigma/dt$ is the prediction of our model.
\section{Numerical results and discussions}
We now present our numerical results for the JLab kinematics with a 6 GeV electron beam ($\sqrt{s_{ep}}\approx 3.5$ GeV). The experimental data \cite{Santoro:2008ai} have been taken in the range
\begin{eqnarray}
1.4<Q^2<3.8\, {\rm GeV}^2, \qquad |t-t_{min}|<3.6\, {\rm GeV}^2, \qquad 2<W<3\, {\rm GeV}, \label{mix}
\end{eqnarray}
where $t_{min}$ is the kinematical lower limit of $t$ which depends on $Q^2$ and $W$. Admittedly, even the maximal value $Q^2=3.8$ GeV$^2$ is not large from a perturbative QCD point of view.
However, (\ref{mix}) is the only kinematical window where the lepto-production data exist. Our model actually provides smooth curves for observables in the above range of $Q^2$.
Fig.~\ref{1} shows $d\sigma/dt$ at $W=2.5$ GeV, $Q^2=3.8$ GeV$^2$.
We chose $m_s= 100$ MeV for the current $s$-quark mass.\footnote{ While this choice is natural in the present framework, it leads to a too steep rise of the cross section $\sigma(Q^2)$ as $Q^2$ is decreased towards $1$ GeV$^2$. Of course, our approach breaks down in this limit, but it is still possible to get a better $Q^2$ behavior in the low-$Q^2$ region by switching to the constituent $s$-quark mass $m_s=m_\phi/2\approx 500$ MeV, or perhaps even $m_s\to m_\phi$
as in the vector meson dominance (VMD) model. } The four curves correspond to different values of the $D$-term, $D_s=0,-0.4,-0.7,-1.3$ in descending order.
We see that if the D-term is large enough, it causes a flattening or even a bump in the $|t|$-distribution in the small-$t$ region. This is due to the explicit factors of $\Delta^\mu$ ($t=\Delta^2$) multiplying $D_s$ in (\ref{mom}) which tend to shift the peak of the $t$-distribution to larger values. Unfortunately, we cannot directly compare our result with the JLab data. The relevant plot, Fig.~18 of Ref.~\cite{Santoro:2008ai} is a mixture of data from different values of $Q^2$ in the range (\ref{mix}). For a meaningful comparison, $d\sigma/dt$ should be plotted for a fixed (large) value of $Q^2$, and there should be enough data points in the most interesting region $|t|\lesssim 1$ GeV$^2$. By the same reason, we cannot adjust the overall normalization of the amplitude mentioned at the end of the previous section. Incidentally, we note that in this kinematics the cross section is dominated by the contribution from the transversely polarized photon, namely, the part proportional to $g_\perp^{\mu\nu}$ in (\ref{need}).
\begin{figure}
\includegraphics[width=0.6\linewidth]{dsigma}
\caption{Differential cross section $d\sigma/dt$ in units of nb/GeV$^2$ as a function of $|t|$ (in GeV$^2$). $W=2.5$ GeV, $Q^2=3.8$ GeV$^2$. The four curves correspond to $D_s=0,-0.4,-0.7,-1.3$ from top to bottom.
}
\label{1}
\end{figure}
For illustration, in Fig.~\ref{2} we show the result with $Q^2=20$ GeV$^2$ and $W=2.5$ GeV, having in mind the kinematics of the Electron-Ion Colliders (EICs) in the U.S. \cite{Accardi:2012qut,Proceedings:2020eah} and in China \cite{Anderle:2021wcy}. We chose $\sqrt{s_{ep}}=30$ GeV for definiteness, but the dependence on $s_{ep}$ is very weak as it only enters the parameter $\epsilon$ in (\ref{dt}) and $\epsilon \approx 1$ in the present kinematics. (Of course the $ep$ cross section (\ref{epcross}) strongly depends on $s_{ep}$.) The contribution from the longitudinally polarized photon (the part proportional to $\epsilon \varepsilon_L^\mu \varepsilon_L^\nu$ in (\ref{need})) is now comparable to the transverse part.
Again the impact of the D-term is noticeable, but the bump has almost disappeared and we only see a flattening of the curve in the extreme case $D_s=-1.3$. The reason is simple. The cross section schematically has the form
\begin{eqnarray}
\frac{d\sigma}{dt} \sim \frac{f(t)}{(1-t/m^2)^{a}},
\end{eqnarray}
where $f(t)$ is a low-order polynomial in $t$ and $a=4,5,6$. (See (\ref{an}). The amplitude squared is a linear combination of $A_s^2(t),A_s(t)D_s(t)$ and $D_s^2(t)$.) The $t$-dependence of $f(t)$ comes from the D-term and gamma matrix traces involving nucleon spinors (\ref{mom}). Clearly, $f(t)$ can affect the shape of $d\sigma/dt$ only when $|t|< m^2\sim 1$ GeV$^2$. Beyond that, one simply has the power law $d\sigma/dt \sim 1/t^{c}$ with $c<4$. In Fig.~\ref{1}, $|t_{min}|<1$ GeV, and this is why we see more interesting structures. As $Q^2$ gets larger, so does $|t_{min}|$ and the structure disappears. \\
In conclusion, we have proposed a new model of $\phi$-meson lepto-production near threshold. In our model, the cross section is solely determined by the strangeness gravitational form factors, similarly to the $J/\psi$ case where it is determined by the gluon counterparts \cite{Boussarie:2020vmu}. Of particular interest is the value of $D_s$, the strangeness contribution to the proton D-term. While $D_s$ is ignored in most literature, an argument based on the large-$N_c$ QCD suggests that it may actually be comparable to $D_{u,d}$ \cite{Goeke:2001tz}. If this is the case, we predict a flattening or possibly a bump in the $t$-distribution of $d\sigma/dt$ in the small-$t$ region. It is very interesting to test this scenario by re-analyzing the JLab data \cite{Santoro:2008ai} or conducting new experiments focusing on the $|t|<1$ GeV$^2$ region.
There are number of directions for improvement. As already mentioned, operators other than the energy momentum tensors should be included as much as possible, although this will unavoidably introduce more parameters in the model. We have argued in the appendix that the contribution from the twist-two, higher spin operators is small, but this needs to be checked. Also the renormalization group evolution of the form factors should be taken into account if in future one can measure this process over a broad range in $Q^2$ such as at the EICs in the U.S. and in China \cite{Accardi:2012qut,Anderle:2021wcy}.
\begin{figure}
\includegraphics[width=0.6\linewidth]{dsigma2}
\caption{Differential cross section $d\sigma/dt$ in units of pb/GeV$^2$ as a function of $|t|$ (in GeV$^2$). $W=2.5$ GeV, $Q^2=20$ GeV$^2$, $\sqrt{s_{ep}}=30$ GeV. The four curves correspond to $D_s=0,-0.4,-0.7,-1.3$ from top to bottom.
}
\label{2}
\end{figure}
\section*{Acknowledgments}
Y.~H. thanks the Yukawa Institute for Theoretical Physics for hospitality.
This work is supported by the U.S. Department of Energy, Office of
Science, Office of Nuclear Physics, under contracts No. DE- SC0012704, DE-SC-0002145
and DE-FG02-93ER40771.
It is also supported in part by Laboratory Directed Research and Development (LDRD)
funds from Brookhaven Science Associates.
|
\section{Introduction}
The transmission of quantum states between devices is crucial for many quantum network protocols.
In the near-term, quantum memory limitations will restrict quantum networks to ``prepare and measure'' functionality \cite{wehner2018quantum}, which allows for quantum communication between separated parties but requires measurement immediately upon reception.
Prepare and measure scenarios exhibit quantum advantages for tasks that involve distributed information processing \cite{BuhrmanComplexity2010} or establishing nonlocal correlations which cannot be reproduced by bounded classical communication and shared randomness \cite{Vicente2017shared}. These nonlocal correlations lead to quantum advantages in random access codes \cite{ambainis2008quantum,tavakoli2015}, randomness expansion \cite{li2011semi}, device self-testing \cite{tavakoli2018}, semi-device-independent key distribution \cite{pawlowski2011semi}, and dimensionality witnessing \cite{brunner2008testing,hendrych2012experimental}.
The general communication process is depicted in Fig. \ref{Fig:simulation}(a) with Alice (the sender) and Bob (the receiver) connected by some quantum channel $\mathcal{N}^{A\to B}$. Alice encodes a classical input $x\in\mathcal{X}$ into a quantum state $\rho_x$ and sends it through the channel to Bob, who then measures the output using a positive-operator valued measure (POVM) $\{\Pi_y\}_{y\in\mathcal{Y}}$ to obtain a classical message $y\in\mathcal{Y}$.
The induced classical channel, denoted by $\mathbf{P}_\mathcal{N}$, has transition probabilities
\begin{equation}
\label{Eq:channel-induce}
P_{\mathcal{N}}(y|x)=\textrm{Tr}\Big[\Pi_y\mathcal{N}\big(\rho_x\big)\Big].
\end{equation}
\begin{figure}[b]
\centering
\includegraphics[width=.48\textwidth]{figures/figure1.png}
\caption{\linespread{1}\selectfont{\small
A general classical communication process. We represent classical information as blue double lines, quantum information as black solid lines, and shared randomness as dotted red lines. (a)
A classical channel $\mathbf{P}_{\mathcal{N}}$ is generated from a quantum channel $\mathcal{N}$ via Eq. \eqref{Eq:channel-induce}.
A classical-quantum encoder $\Psi$ maps the classical input $x\in\mathcal{X}$ into a quantum state $\rho_x$.
A quantum-classical decoder $\Pi$ implements POVM $\{\Pi_y\}_{y\in\mathcal{Y}}$
. (b) Channel $\mathbf{P}_{\mathcal{N}}$ is simulated using shared randomness and a noiseless classical channel via Eq. \eqref{Eq:d-simulate}.
Alice encodes input $x$ into classical message $m$ with probability $T_\lambda(m|x)$ while Bob decodes message $m$ into output $y$ with probability $R_\lambda(y|m)$.
The protocol is coordinated using a shared random value $\lambda$ drawn from sample space $\Lambda$ with probability $q(\lambda)$.
}}
\label{Fig:simulation}
\end{figure}
\noindent A famous result by Holevo implies that the communication capacity of
$\mathbf{P}_\mathcal{N}$ is limited by $\log_2d$, where $d$ is the input Hilbert space dimension of $\mathcal{N}$ \cite{holevo1973bounds}; hence a noiseless classical channel transmitting $d$ messages has a capacity no less than $\mathbf{P}_\mathcal{N}$.
However, channel capacity is just one figure of merit, and there may be other features of a $\mathbf{P}_{\mathcal{N}}$ that do not readily admit a classical simulation.
The strongest form of simulation is an exact replication of the transition probabilities $P_{\mathcal{N}}(y|x)$ for any set of states $\{\rho_x\}_{x\in\mathcal{X}}$ and POVM $\{\Pi_y\}_{y\in\mathcal{Y}}$.
This problem falls in the domain of zero-error quantum information theory \cite{Korner-1998a, Duan-2009a, Cubitt-2011b, Cubitt-2011a, Duan-2016a}, which considers the classical and quantum resources needed to perfectly simulate a given channel.
Unlike the capacity, a zero-error simulation of $\mathbf{P}_\mathcal{N}$ typically requires additional communication beyond the input dimension of $\mathcal{N}$.
For example, a noiseless qubit channel $\textrm{id}_2$ can generate channels $\mathbf{P}_{\textrm{id}_2}$ that cannot be faithfully simulated using a one bit of classical communication \cite{Vicente2017shared}.
The simulation question becomes more interesting if ``static'' resources are used for the channel simulation \cite{Devetak-2004a, Devetak-2008a}, in addition to the ``dynamic'' resource of noiseless classical communication. For example, shared randomness is a relatively inexpensive classical resource that Alice and Bob can use to coordinate their encoding and decoding maps used in the simulation protocol shown in Fig. \ref{Fig:simulation}(b). Using shared randomness, a channel can be exactly simulated with a forward noiseless communication rate that asymptotically approaches the channel capacity; a fact known as the Classical Reverse Shannon Theorem \cite{Bennett2002_rev_shannon}. More powerful static resources such as shared entanglement or non-signaling correlations could also be considered \cite{Cubitt-2011a, Wang-2020a, Fang-2020a}.
While the Classical Reverse Shannon Theorem describes many-copy channel simulation, this work focuses on zero-error channel simulation in the single-copy case. The minimum amount of classical communication (with unlimited shared randomness) needed to perfectly simulate every classical channel $\mathbf{P}_{\mathcal{N}}$ having the form of Eq. \eqref{Eq:channel-induce} is known as the signaling dimension of $\mathcal{N}$ \cite{Dall'Arno_no_hypersignaling_2017}.
Significant progress in understanding the signaling dimension was made by Frenkel and Weiner who showed that every $d$-dimensional quantum channel requires no more than $d$ classical messages to perfectly simulate \cite{Frenkel2015}.
This result is a ``fine-grained'' version of Holevo's Theorem for channel capacity mentioned above.
However, the Frenkel-Weiner bound is not tight in general.
For example, consider the completely depolarizing channel on $d$ dimensions, $\mathcal{D}(\rho)=\mathbb{I}/d$.
For any choice of inputs $\{\rho_x\}_x$ and POVM $\{\Pi_y\}_y$, the Frenkel-Weiner protocol yields a simulation of $\mathbf{P}_\mathcal{D}$ that uses a forward transmission of $d$ messages. However, this is clearly not optimal since $\mathbf{P}_{\mathcal{D}}$ can be reproduced with no forward communication whatsoever; Bob just samples from the distribution $P(y)=\textrm{Tr}[\Pi_y]/d$.
A fundamental problem is then to understand when a noisy classical channel sending $d$ messages from Alice to Bob actually requires $d$ noiseless classical messages for zero-error simulation. As a main result of this paper, we provide a family of simple tests that determine when this amount of communication is needed. In other words, we characterize the conditions in which the simulation protocol of Frenkel and Weiner is optimal for the purposes of sending $d$ messages over a $d$-dimensional quantum channel.
This work pursues a device-independent certification of signaling dimension similar to previous approaches used for the device-independent dimensionality testing of classical and quantum devices \cite{Gallego-2010a, Dall'Arno-2012a, Ahrens-2012a, brunner2013dim_test, Dall'Arno-2017b}.
Specifically, we obtain Bell inequalities that stipulate necessary conditions on the signaling dimension of $\mathcal{N}$ in terms of the probabilities $P_\mathcal{N}(y|x)$, with no assumptions made about the quantum states $\{\rho_x\}_x$, POVM $\{\Pi_y\}_y$, or channel $\mathcal{N}$ \cite{Dall'Arno-2017a}.
Complementary results have been obtained by Dall'Arno \textit{et al.} who approached the simulation problem from the quantum side and characterized the set of channels $\mathbf{P}_\mathcal{N}$ that can be obtained using binary encodings for special types of quantum channels $\mathcal{N}$ \cite{Dall'Arno-2017a}.
In this paper, we compute a wide range of Bell inequalities using the adjacency decomposition technique \cite{christof2001decomposition}, recovering prior results of Frenkel and Weiner \cite{Frenkel2015} and generalizing work by Heinosaari and Kerppo \cite{Heinosaari2019}. For certain cases we prove that these inequalities are complete, i.e. providing both necessary and sufficient conditions for signaling dimension.
As a further application, we compute bounds for the signaling dimension of partial replacer channels.
Proofs for our main results are found in the Appendix while our supporting software is found on Github \cite{SignalingDimension.jl}.
\section{Signaling Polytopes}
We begin our investigation by reviewing the structure of channels that use noiseless classical communication and shared randomness. Let $\mathcal{P}^{n\to n'}$ denote the family of channels having input set $\mathcal{X}=[n]:=\{1,\cdots,n\}$ and output set $\mathcal{Y}=[n']$.
A channel $\mathbf{P} \in \mathcal{P}^{n \to n'}$ is represented by an $n'\times n$ column stochastic matrix, and we thus identify $\mathcal{P}^{n\to n'}$ as a subset of $\mathbb{R}^{n'\times n}$, the set of all $n'\times n$ real matrices. In general we refer to a column (or row) of a matrix as being stochastic if its elements are non-negative and sum to unity, and a column (resp. row) stochastic matrix has only stochastic columns (resp. rows). The elements of a real matrix $\mathbf{G}\in \mathbb{R}^{n'\times n}$ are denoted by $G_{y,x}$, while those of a column stochastic matrix $\mathbf{P}\in\mathcal{P}^{n\to n'}$ are denoted by $P(y|x)$ to reflect their status as conditional probabilities. The Euclidean inner product between $\mathbf{G},\mathbf{P}\in\mathbb{R}^{n'\times n}$ is expressed as $\langle\mathbf{G},\mathbf{P}\rangle:=\sum_{x,y}G_{y,x}P(y|x)$, and for any $\mathbf{G}\in\mathbb{R}^{n'\times n}$ and $\gamma\in\mathbb{R}$, we let the tuple $(\mathbf{G},\gamma)$ denote the linear inequality $\langle\mathbf{G},\mathbf{P}\rangle\leq \gamma$.
Consider now a scenario in which Alice and Bob have access to a noiseless channel capable of sending $d$ messages. They can use this channel to simulate a noisy channel by applying pre- and post-processing maps.
If they coordinate these maps using a shared random variable $\lambda$ with probability mass function $q(\lambda)$, then they can simulate any channel $\mathbf{P}$ that decomposes as
\begin{equation}
\label{Eq:d-simulate}
P(y|x)=\sum_\lambda q(\lambda)\sum_{m\in[d]}R_\lambda(y|m)T_\lambda(m|x),
\end{equation}
\noindent where $m\in[d]$ is the message sent from Alice to Bob and $T(m|x)$ (resp. $R(y|m)$) is an element of Alice's encoder $\mathbf{T}\in\mathcal{P}^{n\to d}$ (resp. Bob's decoder $\mathbf{R}\in\mathcal{P}^{d\to n'}$).
\begin{definition} \label{Def-signaling-polytope}
For given positive integers $n$, $n'$, and $d$, the set of all channels satisfying Eq. \eqref{Eq:d-simulate} constitute the \textit{signaling polytope}, denoted by $\mathcal{C}_d^{n\to n'}$.
\end{definition}
\noindent The signaling polytope $\mathcal{C}_d^{n\to n'}$ is a convex polytope of dimension $n(n'-1)$ whose vertices $\mathbf{V}\in\mathcal{P}^{n \to n'}$ have 0/1 matrix elements and $\rank(\mathbf{V}) \leq d$. We define $\langle\mathbf{G},\mathbf{P}\rangle\leq\gamma$ as a Bell inequality for $\mathcal{C}_d^{n\to n'}$ if $\mathcal{C}_d^{n\to n'}\subset\{\mathbf{P}\in\mathcal{P}^{n \to n'}\;|\;\langle \mathbf{G},\mathbf{P}\rangle\leq \gamma\}$, and it is a ``tight'' Bell inequality if the equation $\langle\mathbf{G},\mathbf{P}\rangle=\gamma$ is also solved by $n(n'-1)$ affinely independent vertices. When the latter holds, the solution space to $\langle\mathbf{G},\mathbf{P}\rangle=\gamma$ is called a facet of $\mathcal{C}_d^{n\to n'}$.
The Weyl-Minkowski Theorem ensures that a complete set of tight Bell inequalities $ \{\langle\mathbf{G}_k,\mathbf{P}\rangle\leq\gamma_k\}_{k=1}^r$ exists such that $\mathbf{P}\in \mathcal{C}_d^{n\to n'}$ \textit{iff} it satisfies all inequalities in this set \cite{Ziegler-2012a}.
Additional details about signaling polytopes are found in Appendix \ref{Appendix-polytope-structure}.
Having introduced signaling polytopes, we can now define the signaling dimension of a channel.
This terminology is adopted from recent work by Dall'Arno \textit{et al.} \cite{Dall'Arno_no_hypersignaling_2017} who defined the signaling dimension of a system in generalized probability theories;
an analogous quantity without shared randomness has also been studied by Heinosaari \textit{et al.} \cite{Heinosaari-2020a}. In what follows, we assume that $\mathcal{N}:\mathcal{S}(A)\to\mathcal{S}(B)$ is a completely positive trace-preserving (CPTP) map, with $\mathcal{S}(A)$ denoting the set of density operators (i.e. trace-one positive operators) on system $A$, and similarly for $\mathcal{S}(B)$.
\begin{definition}\label{def-signaling-dimension}
Let $\mathcal{P}_{\mathcal{N}}^{ n \to n'}$ be the set of all classical channels $\mathbf{P}_{\mathcal{N}}\in\mathcal{P}^{n \to n'}$ generated from
$\mathcal{N}$ via Eq. \eqref{Eq:channel-induce}. The $n\to n'$ signaling dimension of $\mathcal{N}$, denoted by $\kappa^{n\to n'}(\mathcal{N})$, is the smallest $d$ such that $\mathcal{P}_{\mathcal{N}}^{ n \to n'}\subset\mathcal{C}_d^{n\to n'}$. The signaling dimension of a channel $\mathcal{N}$, denoted by $\kappa(\mathcal{N})$, is the smallest $d$ such that $\mathcal{P}_{\mathcal{N}}^{n\to n'}\!\!\subset\mathcal{C}_d^{n\to n'}$ for all $ n,n'$.
\end{definition}
For any channel $\mathcal{N}$, a trivial upper bound on the $n \to n'$ signaling dimension is given by
\begin{equation}
\label{Eq:classical-cardinality-bound}
\kappa^{n\to n'}(\mathcal{N})\leq\min\{n,n'\}.
\end{equation}
\noindent Indeed, when this bound is attained, Alice and Bob can simulate any $\mathbf{P}\in\mathcal{P}^{n \to n'}$: either Alice applies channel $\mathbf{P}$ on her input and sends the output to Bob, or she sends the input to Bob and he applies $\mathbf{P}$ on his end.
In Theorem \ref{Thm:main1} we provide necessary and sufficient conditions for when this trivial upper bound is attained.
For a quantum channel $\mathcal{N}$, the trivial upper bound is
\begin{equation}
\label{Eq:bound-dimension}
\kappa(\mathcal{N})\leq\min\{d_A,d_B\},
\end{equation}
\noindent where $d_A$ and $d_B$ are the Hilbert space dimensions of Alice and Bob's systems. This bound is a direct consequence of Frenkel and Weiner's result \cite{Frenkel2015}, which can be restated in our terminology as $\kappa({\text{id}_d}) =d$, where $\text{id}_d$ is the noiseless channel on a $d$-dimensional quantum system.
To prove Eq. \eqref{Eq:bound-dimension}, Alice can either send the states $\{\rho_x\}_x$ to Bob who then performs the POVM $\{\mathcal{N}^\dagger(\Pi_y)\}_y$, or she can send the states $\{\mathcal{N}(\rho_x)\}_x$ to Bob who then perfoms the POVM $\{\Pi_y\}_y$. Here $\mathcal{N}^\dagger$ denotes the adjoint map of $\mathcal{N}$.
Another relationship we observe is
\begin{equation}
\label{Eq:bound-carethedory}
\kappa^{n\to n'}(\mathcal{N})=\kappa^{n\to d_B^2}(\mathcal{N})
\qquad \forall \; n'\geq d_B^2.
\end{equation}
This follows from Carath\'{e}odory's Theorem \cite{Barvinok-2002a}, which implies that every POVM on a $d_B$-dimensional system can be expressed as a convex combination of POVMs with no more than $d_B^2$ outcomes \cite{Davies-1978a}. Since shared randomness is free, Alice and Bob can always restrict their attention to POVMs with no more than $d_B^2$ outcomes for the purposes of simulating any channel in $\mathcal{P}_\mathcal{N}^{n\to n'}$ when $n'\geq d_B^2$.
The notion of signaling dimension also applies to noisy classical channels. A classical channel from set $\mathcal{X}$ to $\mathcal{Y}$ can be represented by a CPTP map $\mathcal{N}:\mathcal{S}(\mathbb{C}^{|\mathcal{X}|})\to\mathcal{S}(\mathbb{C}^{|\mathcal{Y}|})$ that completely dephases its input and output in fixed orthonormal bases $\{\ket{x}\}_{x\in\mathcal{X}}$ and $\{\ket{y}\}_{y\in\mathcal{Y}}$, respectively. The transition probabilities of $\mathcal{N}$ are then given by
Eq. \eqref{Eq:channel-induce} as $P_{\mathcal{N}}(y|x) = \textrm{Tr}\big[\op{y}{y}\mathcal{N}\big(\op{x}{x}\big)\big]$.
The channel $\mathcal{N}$ can be used to generate another channel $\ol{\mathcal{N}}$ with input and output alphabets $\ol{\mathcal{X}}$ and $\ol{\mathcal{Y}}$ by performing a pre-processing map $\mathbf{T}:\ol{\mathcal{X}}\to\mathcal{X}$ and post-processing map $\mathbf{R}:\mathcal{Y}\to\ol{\mathcal{Y}}$, thereby yielding the channel $\mathbf{P}_{\ol{\mathcal{N}}}=\mathbf{R}\mathbf{P}_\mathcal{N}\mathbf{T}$. When this relationship holds, $\mathbf{P}_{\ol{\mathcal{N}}}$ is said to be \textit{ultraweakly majorized} by $\mathbf{P}_\mathcal{N}$ \cite{Heinosaari2019, Heinosaari-2020a}, and the signaling dimension of $\mathbf{P}_{\ol{\mathcal{N}}}$ is no greater than that of $\mathbf{P}_\mathcal{N}$ \cite{Cubitt-2011a}.
In practice, the channel connecting Alice and Bob may be unknown or not fully characterized. This is the case in most experimental settings where unpredictable noise affects the encoded quantum states. In such scenarios it is desirable to ascertain certain properties of the channel without having to perform full channel tomography, a procedure that requires trust in the state preparation device on Alice's end and the measurement device on Bob's side. A device-independent approach infers properties of the channel by analyzing the observed input-output classical correlations $P(y|x)$ obtained as sample averages over many uses of the memoryless channel \cite{Dall'Arno-2017a}. The Bell inequalities introduced in the next section can be used to certify the signaling dimension of the channel: if the correlations $P(y|x)$ are shown to violate a Bell inequality of $\mathcal{C}_d^{n\to n'}$, then the signaling dimension $\kappa^{n\to n'}(\mathcal{N})> d$.
If these correlations arise from some untrusted quantum channel $\mathcal{N}^{A\to B}$, by Eq. \eqref{Eq:bound-dimension} it then follows that $\min\{d_A,d_B\}>d$. Hence a device-independent certification of signaling dimension leads to a device-independent certification of the physical input/output Hilbert spaces of the channel connecting Alice and Bob.
\section{Bell Inequalities for Signaling Polytopes}
In this section we discuss Bell inequalities for signaling polytopes. Since signaling polytopes are invariant under the relabelling of inputs and outputs, all discussed inequalities describe a family of inequalities where each element is obtained by a permutation of the inputs and/or outputs.
Additionally, a Bell inequality for one signaling polytope can be lifted to a polytope having more inputs and/or outputs \cite{Pironio2005,Rosset2014} (see Fig. \ref{Fig:input-output-lifting}). Formally, a Bell inequality $\langle\mathbf{G},\mathbf{P}\rangle\leq\gamma$ is said to be input lifted to $\langle\mathbf{G}'',\mathbf{P}\rangle\leq\gamma$ if $\mathbf{G}''\in\mathbb{R}^{n'\times m}$ is obtained from $\mathbf{G}\in\mathbb{R}^{n'\times n}$ by padding it with $(m-n)$ all-zero columns. On the other hand, a Bell inequality $\langle\mathbf{G},\mathbf{P}\rangle\leq\gamma$ is said to be output lifted to $\langle\mathbf{G}',\mathbf{P}\rangle\leq\gamma$ if $\mathbf{G}'\in \mathbb{R}^{m'\times n}$ is obtained from $\mathbf{G}\in\mathbb{R}^{n'\times n}$ by copying rows; \textit{i.e.}, there exists a surjective function $f:[m']\to [n']$ such that $G'_{y,x}=G_{f(y),x}$ for all $y\in[m']$ and $x\in[n]$.
Note that $m'>n'$ and $m>n$ in these examples.
\begin{figure}[t]
\centering
\begin{tabular}{c c c c}
(a) & $\quad\mathbf{G} = \begin{bmatrix}
1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \\
\end{bmatrix}$ & $\to$ & $\mathbf{G}'' = \begin{bmatrix}
1 & 0 & 0 & 0 \\ 0 & 1 & 0 & 0 \\ 0 & 0 & 1 & 0 \\
\end{bmatrix}$ \\
\hfill\\
(b) & $\quad\mathbf{G} = \begin{bmatrix}
1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \\
\end{bmatrix}$ & $\to$ & $\mathbf{G}' = \begin{bmatrix}
1 & 0 & 0 \\ 1 & 0 & 0 \\ 0 & 1 & 0 \\ 0 & 0 & 1 \\
\end{bmatrix}$
\end{tabular}
\caption{\linespread{1}\selectfont{\small
(a) Input and (b) output liftings of $\mathbf{G}=\mathbb{I}_3$.
}}
\label{Fig:input-output-lifting}
\end{figure}
To obtain polytope facets, it is typical to first enumerate the vertices, then use a transformation technique such as Fourier-Motzkin elimination to derive the facets \cite{Ziegler-2012a}.
Software such as PORTA \cite{PORTA, XPORTA.jl} assists in this computation, but the large number of vertices leads to impractical run times.
To improve efficiency, we utilize the adjacency decomposition technique which heavily exploits the permutation symmetry of signaling polytopes \cite{ christof2001decomposition} (see Appendix \ref{Appendix-Adjacency}).
Our software and computed facets are publicly available on Github \cite{SignalingDimension.jl} while a catalog of general tight Bell inequalities is provided in Appendix \ref{Appendix-facet-classes}.
We now turn to a specific family of Bell inequalities motivated by our computational results.
\subsection{Ambiguous Guessing Games}
For $k\in[0,n']$ and $d\leq \min\{n,n'\}$, let $\mathbf{G}^{n, n'}_{k,d}$ be any $n'\times n$ matrix such that (i) $k$ rows are stochastic with 0/1 elements, and (ii) the remaining $(n'-k)$ rows have $1/(n-d+1)$ in each column. As explained below, it will be helpful to refer to rows of type (i) as ``guessing rows'' and rows of type (ii) as ``ambiguous rows.''
For example, if $n=n'=6$, $k=5$, and $d=2$, then up to a permutation of rows and columns we have
\begin{equation}
\mathbf{G}^{6,6}_{5,2} = \begin{bmatrix}
1 & 0 & 0 & 0 & 0 & 0 \\
1 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 \\
\tfrac{1}{5} & \tfrac{1}{5} & \tfrac{1}{5} & \tfrac{1}{5} & \tfrac{1}{5} & \tfrac{1}{5} \\
\end{bmatrix}.
\end{equation}
\noindent
For any channel $\mathbf{P}\in\mathcal{C}_d^{n\to n'}$, the Bell inequality
\begin{align}
\label{Eq:Ambiguous}
\langle\mathbf{G}^{n, n'}_{k,d},\mathbf{P}\rangle\leq d
\end{align}
\noindent is satisfied. To prove this bound, suppose without loss of generality that the first $k$ rows of $\mathbf{G}^{n, n'}_{k,d}$ are guessing rows.
Let $\mathbf{V}$ be any vertex of $\mathcal{C}_d^{n\to n'}$ where $t$ of its first $k$ rows are nonzero. If $t=d$, then clearly Eq. \eqref{Eq:Ambiguous} holds. Otherwise, if $t<d$, then $\langle\mathbf{G}^{n, n'}_{k,d},\mathbf{V}\rangle\leq t+(n-t)/(n-d+1)\leq d$, where the last inequality follows after some algebraic manipulation.
Equation \eqref{Eq:Ambiguous} can be interpreted as the score of a guessing game that Bob plays with Alice. Suppose that Alice chooses a channel input $x\in[n]$ with uniform probability and sends it through a channel $\mathbf{P}$. Based on the channel output $y$, Bob guesses the value of $x$. Formally, Bob computes $\hat{x}=f(y)$ for some guessing function $f$, and if $\hat{x}=x$ then he receives one point.
In this game, Bob may also declare Alice's input as being ambiguous or indistinguishable, meaning that $f:[d]\to [n]\cup\{?\}$ with $``?"$ denoting Bob's declaration of the ambiguous input. However, whenever Bob declares $``?"$ he only receives $1/(n-d+1)$ points.
Then, Eq. \eqref{Eq:Ambiguous} says that whenever $\mathbf{P}\in\mathcal{C}_d^{n\to n'}$ Bob's average score is bounded by $\frac{d}{n}$.
Note, there is a one-to-one correspondence between each $\mathbf{G}^{n,n'}_{k,d}$ and the particular guessing function $f$ that Bob performs.
If $y$ labels a guessing row of $\mathbf{G}^{n,n'}_{k,d}$, then $f(y)=\hat{x}$, where $\hat{x}$ labels the only nonzero column of row $y$. On the other hand, if $y$ labels an ambiguous row, then $f(y)=``?"$.
We define the $(k,d)$-\textit{ambiguous polytope} $\mathcal{A}_{k,d}^{n\to n'}$ as the collection of all channels $\mathbf{P}\in\mathcal{P}^{n\to n'}$ satisfying Eq. \eqref{Eq:Ambiguous} for every $\mathbf{G}^{n,n'}_{k,d}$.
Naturally, $\mathcal{C}_d^{n \to n'}\subset \mathcal{A}_{k,d}^{n \to n'}$ for all $k\in[0,n']$, therefore, if $\mathbf{P}\notin\mathcal{A}_{k,d}^{n \to n'}$, then $\mathbf{P}\notin \mathcal{C}_d^{n \to n'}$.
Based on the discussion of the previous paragraph, it is easy to decide membership of $\mathcal{A}_{k,d}^{n\to n'}$.
\begin{proposition}
\label{Prop:ambiguous-condition}
A channel $\mathbf{P}\in\mathcal{P}^{n\to n'}$ belongs to $\mathcal{A}_{k,d}^{n\to n'}$ \textit{iff}
\begin{equation}
\label{Eq:Prop-ambiguous-form}
\max_{\pi\in\mathsf{S}_{n'}}\sum_{i=1}^k\Vert \mathbf{r}_{\pi(i)}\Vert_\infty+\frac{1}{n-d+1}\sum_{i=k+1}^{n'}\Vert \mathbf{r}_{\pi(i)}\Vert_1\leq d,
\end{equation}
\noindent where the maximization is taken over all permutations on $[n']$, $\mathbf{r}_i$ denotes the $i^{th}$ row of $\mathbf{P}$, $\Vert\mathbf{r}_i\Vert_\infty$ is the largest element in $\mathbf{r}_i$, and $\Vert\mathbf{r}_i\Vert_1$ is the row sum of $\mathbf{r}_i$.
\end{proposition}
The maximization on the LHS of Eq. \eqref{Eq:Prop-ambiguous-form} can be performed efficiently using the following procedure. For each row $\mathbf{r}_i$ we assign a pair $(a_i, b_i)$ where $a_i=\Vert\mathbf{r}_i\Vert_\infty$ and $b_i=\frac{1}{n-d+1}\Vert\mathbf{r}_i\Vert_1$. Define $\delta i=a_i - b_i$, and relabel the rows of $\mathbf{P}$ in non-increasing order of the $\delta_i$.
Then according to this sorting, we have an ambiguous guessing game score of $\sum_{i=1}^k a_i +\sum_{i=k+1}^{n'} b_i$,
which we claim attains the maximum on the LHS of Eq. \eqref{Eq:Prop-ambiguous-form}. Indeed, for any other row permutation $\pi$, the guessing game score is given by
\begin{align}
&\sum_{\substack{i\in\{1,\cdots,k\}\\\pi(i)\in\{1,\cdots,k\}}} a_i+\sum_{\substack{i\in\{1,\cdots,k\}\\\pi(i)\in\{k+1,\cdots,n'\}}} b_i\notag\\
+&\sum_{\substack{i\in\{k+1,\cdots,n'\}\\\pi(i)\in\{1,\cdots,k\}}} a_i+\sum_{\substack{i\in\{k+1,\cdots,n'\}\\\pi(i)\in\{k+1,\cdots,n'\}}}b_i.\end{align}
Hence the difference in these two scores is
\begin{align}
\sum_{\substack{i\in\{1,\cdots,k\}\\\pi(i)\in\{k+1,\cdots,n'\}}} (a_i-b_i)-\sum_{\substack{i\in\{k+1,\cdots,n'\}\\\pi(i)\in\{1,\cdots,k\}}} (a_i-b_i)\geq 0,
\end{align}
where the inequality follows from the fact that we have ordered the indices in non-increasing order of $(a_i-b_i)$, and the number of terms in each summation is the same since $\pi$ is a bijection.
A special case of the ambiguous guessing games arises when $k=n'$. Then up to a normalization factor $\frac{1}{n}$, we interpret the LHS of Eq. \eqref{Eq:Prop-ambiguous-form} as the success probability when Bob performs maximum likelihood estimation of Alice's input value $x$ given his outcome $y$ (i.e. he chooses the value $x$ that maximizes $P(y|x)$). We hence define $\mathcal{M}_d^{n\to n'}:=\mathcal{A}_{n',d}^{n\to n'}$ as the \textit{maximum likelihood (ML) estimation polytope}. Using Proposition \ref{Prop:ambiguous-condition} we see that
\begin{equation}
\mathbf{P}\in \mathcal{M}_d^{n\to n'}\quad\Leftrightarrow\quad \sum_{y=1}^{n'}\max_{x\in[n]}P(y|x)\leq d.
\end{equation}
An important question is whether the ambiguous guessing Bell inequalities of Eq. \eqref{Eq:Ambiguous} are tight for a signaling polytope $\mathcal{C}_d^{n\to n'}$. In general this will not be case. For instance, $\langle\mathbf{G}^{n, n'}_{k,d},\mathbf{P}\rangle\leq d$ is trivially satisfied whenever $k=0$. Nevertheless, in many cases we can establish tightness of these inequalities. A demonstration of the following facts is carried out in Appendix \ref{proof-of-prop-ambiguous-guessing-facets}.
\begin{proposition}
\label{prop-ambiguous-guessing-facets}
\begin{enumerate}
\item[]
\item[(i)] For $\min\{n,n'\} > d > 1$ and $k=n'$, Eq. \eqref{Eq:Ambiguous} is a tight Bell inequality of $\mathcal{C}_d^{n\to n'}$ \textit{iff} $\mathbf{G}^{n,n'}_{k,d}$ can be obtained by performing input/output liftings and row/column permutations on an $m\times m$ identity matrix $\mathbb{I}_m$, with $\min\{n,n'\} \geq m > d$.
\item[(ii)] For $n'>k\geq n>d>1$, Eq. \eqref{Eq:Ambiguous} is a tight Bell inequality of $\mathcal{C}_d^{n\to n'}$ \textit{iff} $\mathbf{G}^{n,n'}_{k,d}$ can be obtained from the $(n+1)\times n$ matrix
\begin{equation}
\label{Eq:ambiguous-canonical}
\begin{bmatrix}&&\mathbb{I}_n&&\\\tfrac{1}{n-d+1}&\cdots&\tfrac{1}{n-d+1}&\cdots \tfrac{1}{n-d+1}\end{bmatrix}
\end{equation}
by performing output liftings and row/column permutations.
\end{enumerate}
\end{proposition}
\noindent Note that the input/output liftings are used to manipulate the identity matrix $\mathbb{I}_m$ and the matrix of Eq. \eqref{Eq:ambiguous-canonical} into an $n'\times n$ matrix $\mathbf{G}_{k,d}^{n,n'}$. The tight Bell inequalities described in Proposition \ref{prop-ambiguous-guessing-facets}(i) completely characterize the ML polytope $\mathcal{M}_d^{n\to n'}$. For this reason, we refer to any $\mathbf{G}^{n,n'}_{k,d}$ satisfying the conditions of Proposition \ref{prop-ambiguous-guessing-facets}(i) as a maximum likelihood (ML) facet (see Appendix \ref{Appendix-ml-facets}). Likewise, we refer to any $\mathbf{G}^{n,n'}_{k,d}$ satisfying the conditions of Proposition \ref{prop-ambiguous-guessing-facets}(ii) as an ambiguous guessing facet (see Appendix \ref{Appendix-ambiguous-guessing-facets}).
\subsection{Complete Sets of Bell Inequalities}
In general, we are unable to identify the complete set of tight Bell inequalities that bound each signaling polytope $\mathcal{C}_d^{n \to n'}$.
However, we analytically solve the problem in special cases.
\begin{theorem}
\label{Thm:main1}
Let $n$ and $n'$ be arbitrary integers.
\begin{enumerate}
\item[(i)] If $d=n'-1$, then $\mathcal{C}^{n\to n'}_d=\mathcal{M}_d^{n\to n'}$.
\item[(ii)] If $d=n-1$, then $\mathcal{C}_d^{n \to n'} = \displaystyle\bigcap_{k=n}^{n'} \mathcal{A}_{k,d}^{n\to n'}$.
\end{enumerate}
\end{theorem}
\noindent In other words, to decide whether a channel can be simulated by an amount of classical messages strictly less than the input/output alphabets, it suffices to consider the ambiguous guessing games. Moreover, by Eq. \eqref{Eq:Prop-ambiguous-form} it is simple to check if these conditions are satisfied for a given channel $\mathbf{P}$. A proof of Theorem \ref{Thm:main1} is found in Appendix \ref{appendix-proof-thm-main1}.
\begin{figure}[b]
\centering
\begin{tabular}{c c}
(a) $2 \geq \begin{bmatrix}
1 & 0 & 0 & 0 & 0 & 0 \\
1 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 \\
\end{bmatrix}$ & (b) $2 \geq \begin{bmatrix}
1 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 0 \\
\end{bmatrix}$ \\
\hfill \\
(c) $3 \geq \begin{bmatrix}
1 & 1 & 0 & 0 & 0 & 0 \\
1 & 0 & 1 & 0 & 0 & 0 \\
0 & 1 & 1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 & 0 \\
\end{bmatrix}$ & (d) $5 \geq \begin{bmatrix}
1 & 1 & 1 & 0 & 0 & 0 \\
1 & 0 & 0 & 1 & 1 & 0 \\
0 & 1 & 0 & 1 & 0 & 1 \\
0 & 0 & 1 & 0 & 1 & 1 \\
\end{bmatrix}$ \\
\hfill \\
(e) $4 \geq \begin{bmatrix}
2 & 0 & 0 & 0 & 0 & 0 \\
0 & 2 & 0 & 0 & 0 & 0 \\
0 & 0 & 2 & 0 & 0 & 0 \\
1 & 1 & 1 & 0 & 0 & 0 \\
\end{bmatrix}$ & (f) $4 \geq \begin{bmatrix}
2 & 0 & 0 & 0 & 0 & 0 \\
0 & 2 & 0 & 0 & 0 & 0 \\
0 & 0 & 1 & 1 & 0 & 0 \\
1 & 1 & 1 & 0 & 0 & 0 \\
\end{bmatrix}$\\
\hfill \\
(g) $4 \geq \begin{bmatrix}
2 & 0 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 1 & 0 & 0 \\
0 & 0 & 1 & 0 & 1 & 0 \\
1 & 1 & 1 & 0 & 0 & 0 \\
\end{bmatrix}$ & (h) $4 \geq \begin{bmatrix}
1 & 0 & 0 & 1 & 0 & 0 \\
0 & 1 & 0 & 0 & 1 & 0 \\
0 & 0 & 1 & 0 & 0 & 1 \\
1 & 1 & 1 & 0 & 0 & 0 \\
\end{bmatrix}$
\end{tabular}
\caption{\linespread{1}\selectfont{\small Generator facets for the $\mathcal{C}_2^{6 \to 4}$ signaling polytope.
Each inequality is expressed as $\gamma \geq \mathbf{G}$ where the inner product $\langle \mathbf{G} , \mathbf{P}\rangle$ is implied.
(a) ML facet input/output lifted from $\mathcal{C}_2^{3,3}$.
(b) ML facet output lifted from $\mathcal{C}_2^{4 \to 4}$.
(c) Anti-guessing facet output lifted from $\mathcal{C}_2^{4 \to 4}$.
(d) $k$-guessing facet of $\mathcal{C}_2^{6 \to 4}$. (e) Ambiguous guessing facet output lifted from $\mathcal{C}_2^{3 \to 4}$.
(f-h) Rescalings of the $\mathcal{C}_2^{3 \to 4}$ ambiguous guessing facet output lifted to $\mathcal{C}_2^{6 \to 4}$.
General forms of these tight Bell inequalities are derived in Appendix \ref{Appendix-facet-classes}.
}}
\label{Fig:6-2-4_facets}
\end{figure}
We also characterize the $\mathcal{C}^{n\to 4}_2$ signaling polytope. As an application, this case can be used to understand the classical simulation cost of performing Bell measurements on a two-qubit system, since this process induces a classical channel with four outputs.
\begin{theorem}
\label{Thm:main2}
For any integer $n$, a channel $\mathbf{P}\in \mathcal{P}^{n\to 4}$ belongs to $\mathcal{C}^{n\to 4}_2$ \textit{iff} it satisfies the eight Bell inequalities depicted in Fig. \ref{Fig:6-2-4_facets} and all their input/output permutations.
\end{theorem}
\noindent Remarkably, this result shows that no new facet classes for $\mathcal{C}^{n\to 4}_2$ are found when $n > 6$.
Consequently, to demonstrate that a channel $\mathbf{P}\in \mathcal{P}^{n\to 4}$ requires more than one bit for simulation, it suffices to consider input sets of size no greater than six. For $n<6$, the facet classes of $\mathcal{C}^{n\to 4}_2$ are given by the facets in Fig. \ref{Fig:6-2-4_facets} having $(6-n)$ all-zero columns.
We conjecture that in general, no more than $\binom{n'}{d}$ inputs are needed to certify that a channel $\mathbf{P}\in\mathcal{P}^{n\to n'}$ has a signaling dimension larger than $d$. A proof of Theorem \ref{Thm:main2} is found in Appendix \ref{appendix-proof-of-thm-2}.
\subsection{The Signaling Dimension of Replacer Channels}\label{section-replacer-channel}
In the device-independent scenario, Alice and Bob make minimal assumptions about the channel $\mathcal{N}^{A\to B}$ connecting them; they simply try to lower bound the dimensions of $\mathcal{N}$ using input-output correlations $P_{\mathcal{N}}(y|x)$. Applying the results of the previous section, if $\langle\mathbf{G},\mathbf{P}\rangle\leq \gamma$ is a Bell inequality for $\mathcal{C}_d^{n\to n'}$ and
\begin{equation}\label{eq:quantum-optimization}
\max_{\{\rho_x\}_x,\{\Pi_y\}_y} \sum_{y,x} G_{y,x} \textrm{Tr}\Big[\Pi_y \mathcal{N}\big(\rho_x\big)\Big]>\gamma,
\end{equation}
\noindent then $\min\{d_A,d_B\}\geq \kappa(\mathcal{N})>d$. Eq. \eqref{eq:quantum-optimization} describes a conic optimization problem that can be analytically solved only in special cases \cite{cv-2021a}. Hence deciding whether a given quantum channel can violate a particular Bell inequality is typically quite challenging.
Despite this general difficulty, we nevertheless establish bounds for the signaling dimension of partial replacer channels.
A $d$-dimensional partial replacer channel has the form
\begin{equation} \label{Eq:replacer-channel}
\mathcal{R}_{\mu}(X)= \mu X+(1-\mu)\textrm{Tr}[X]\sigma,
\end{equation}
\noindent where $1 \geq \mu \geq 0$ and $\sigma$ is some fixed density matrix. The partial depolarizing channel $\mathcal{D}_{\mu}$ corresponds to $\sigma$ being the maximally mixed state whereas the partial erasure channel $\mathcal{E}_\mu$ corresponds to $\sigma$ being an erasure flag $\op{E}{E}$ with $\ket{E}$ being orthogonal to $\{\ket{1},\cdots,\ket{d}\}$.
\begin{theorem} \label{thm-replacer-bounds}
The signaling dimension of a partial replacer channel is bounded by
\begin{equation} \label{Eq:replacer-bounds}
\lceil \mu d+1-\mu\rceil\leq\kappa(\mathcal{R}_\mu)\leq\min\{d,\lceil\mu d+1\rceil\}.
\end{equation}
Moreover, for the partial erasure channel, the upper bound is tight for all $\mu\in[0,1]$.
\end{theorem}
\begin{proof}
We first prove the upper bound in Eq. \eqref{Eq:replacer-bounds}. The trivial bound $\kappa(\mathcal{R}_\mu)\leq d$ was already observed in Eq. \eqref{Eq:bound-dimension}. To show that $\kappa(\mathcal{R}_\mu)\leq \lceil\mu d+1\rceil$, let $\{\rho_x\}_x$ be any collection of inputs and $\{\Pi_y\}_y$ a POVM. Then
\begin{equation}
\label{Eq:Depolarizing-classical-channel}
P_{\mathcal{R}_\mu}(y|x)=\mu P(y|x)+(1-\mu)S(y),
\end{equation}
where $P(y|x)=\textrm{Tr}[\Pi_y\rho_x]$ and $S(y)=\textrm{Tr}[\Pi_y\sigma]$. From Ref. \cite{Frenkel2015}, we know that $P(b|x)$ can be decomposed like Eq. \eqref{Eq:d-simulate}. Substituting this into Eq. \eqref{Eq:Depolarizing-classical-channel} yields
\begin{align}
P_{\mathcal{R}_\mu}(u|x)&=\sum_\lambda q(\lambda)\sum_{m=1}^dR_\lambda(m|x)\notag\\
&\qquad\times[\mu T_\lambda(y|m)+(1-\mu)S(y)].
\end{align}
For $r=\lceil\mu d +1\rceil$, let $\nu$ be a random variable uniformly distributed over $\{\binom{d}{r-1}\}$, which is the collection of all subsets of $[d]$ having size $r-1$. For a given $\lambda$, $\nu$, and input $x$, Alice performs the channel $\mathbf{T}_\lambda$. If $m\in\nu$, Alice sends message $m'=m$; otherwise, Alice sends message $m'=0$. Upon receiving $m'$, Bob does the following: if $m'\not=0$ he performs channel $\mathbf{R}_\lambda$ with probability $\tfrac{\mu d}{r-1}$ and samples from distribution $S(y)$ with probability $1-\tfrac{\mu d}{r-1}$; if $m'\not=0$ he samples from $S(y)$ with probability one. Since $Pr\{m\in\nu\}=\frac{r-1}{d}$, this protocol faithfully simulates $P_{\mathcal{R}_\mu}$. To establish the lower bound in Eq. \eqref{Eq:Depolarizing-classical-channel}, suppose that Alice sends orthogonal states $\{\ket{1},\cdots,\ket{d}\}$
and Bob measures in the same basis. Then
\begin{equation}
\sum_{i=1}^d\bra{i}\mathcal{R}_\mu(\op{i}{i})\ket{i}=d\mu+(1-\mu),
\end{equation}
which will violate Eq. \eqref{Eq:Ambiguous} for the ML polytope $\mathcal{M}^{d\to d}_{r}$ whenever $r< \mu d+(1-\mu)$. Hence any zero-error simulation will require at least $\lceil \mu d+1-\mu\rceil$ classical messages. For the erasure channel, this lower bound can be tightened by considering the score for other ambiguous games, as detailed in Appendix \ref{Appendix-replacer}.
\end{proof}
\section*{Discussion}
In this work, we have presented the signaling dimension of a channel as its classical simulation cost.
In doing so, we have advanced a device-independent framework for certifying the signaling dimension of a quantum channel as well as its input/output dimensions. While this work focuses on communication systems, our framework also applies to computation and memory tasks.
The family of ambiguous guessing games includes the maximum likelihood facets, which say that $\sum_{y=1}^{n'}\max_{x\in[n]}P(y|x)\leq d$ for all $\mathbf{P}\in\mathcal{C}_d^{n\to n'}$. Since the results of Frenkel and Weiner imply that $\mathcal{P}_{\mathcal{N}}^{n\to n'}\subset \mathcal{C}_d^{n\to n'}$ whenever $d \geq \min\{d_A,d_B\}$ for channel $\mathcal{N}^{A\to B}$ \cite{Frenkel2015}, it follows that
\begin{equation}
\max_{\substack{\{\rho_x\}_{x\in[n]}\\\{\Pi_y\}_{y\in[n]}}} \sum_{x=1}^n \textrm{Tr}\Big[\Pi_x \mathcal{N}\big(\rho_x\big)\Big]\leq d,
\end{equation}
an observation also made in Ref. \cite{brunner2013dim_test}. Despite the simplicity of this bound, in general it is too loose to certify the input/output Hilbert space dimensions of a channel. For example, consider the $50:50$ erasure channel $\mathcal{E}_{1/2}$ acting on a $d_A=3$ system.
It can be verified that $\mathcal{P}_{\mathcal{E}_{1/2}}^{n\to n'}\subset \mathcal{M}_2^{n\to n'}$, \textit{i.e.} $\sum_x\textrm{Tr}\big[\Pi_x \mathcal{E}_{1/2}(\rho_x)\big]\leq 2$ for all $\{\rho_x\}_x$ and $\{\Pi_y\}_y$.
Hence maximum likelihood estimation yields the lower bound $\kappa(\mathcal{E}_{1/2}) \geq 2$.
On the other hand, the classical channel
\begin{equation}
\mathbf{P_{\mathcal{E}_{1/2}}} = \begin{bmatrix}
0.5 & 0 & 0 \\
0 & 0.5 & 0 \\
0 & 0 & 0.5 \\
0.5 & 0.5 & 0.5 \\
\end{bmatrix}
\end{equation}
\noindent generated by orthonormal input states $\{\ket{1},\ket{2},\ket{3}\}$ and a measurement in the orthonormal basis $\{\ket{1},\ket{2},\ket{3},\ket{E}\}$ violates Eq. \eqref{Eq:Prop-ambiguous-form} for the $\mathcal{A}_{3,2}^{3\to 4}$ ambiguous polytope.
Hence $\mathbf{P}_{\mathcal{E}_{1/2}} \notin \mathcal{A}_{3,2}^{3\to 4}$ and it follows that $\kappa^{3 \to 4}(\mathcal{E}_{1/2}) \geq 3$.
Therefore, the ambiguous guessing game certifies the qutrit nature of the input space whereas maximum likelihood estimation does not.
Our results can be extended in two key directions.
First, our characterization of the signaling polytope is incomplete.
Novel Bell inequalities, lifting rules, and complete sets of facets can be derived beyond those discussed in this work.
Such results would help improve the signaling dimension bounds and the efficiency of computing Bell inequalities.
Second, the signaling dimension specifies the classical cost of simulating a quantum channel, but not the protocol that achieves the classical simulation.
Such a simulation protocol would apply broadly across the field of quantum information science and technology.
\subsection*{Supporting Software}
This work is supported by SignalingDimension.jl \cite{SignalingDimension.jl}.
This software package includes our signaling polytope computations, numerical facet verification, and signaling dimension certification examples.
SignalingDimension.jl is publicly available on Github and written in the Julia programming language \cite{Julia-2017}.
The software is documented, tested, and reproducible on a laptop computer.
The interested reader should review the software documentation as it elucidates many details of our work.
\vspace{12pt}
\subsection*{Acknowledgements}
We thank Marius Junge for enlightening discussions during the preparation of this paper. We acknowledge NSF Award \# 2016136 for supporting this work.
|
\section*{Supplementary Information}
\section{Specifying the inverse problem}
\label{sec:inverse-problem}
We define the target phase behavior by specifying the homogeneous phase volume fractions $\vec\phi^{(0)}$, as well as the composition, $x_i^{(\alpha)}$, for each enriched component $i$ in each target phase $\alpha$.
The compositions of the depleted components in each target phase are not specified; however, we require that all the depleted components comprise a volume fraction within the target phase that is less than $1/M_\alpha$, where $M_\alpha$ is the number of enriched components in the $\alpha$ phase.
Therefore, we arrive at the first constraint in the main text,
\begin{equation}
\label{eq:constraint-1}
x_i^{(\alpha)} \le \frac{\zeta}{NM_\alpha}\text{ if species }i\text{ is depleted in phase }\alpha,
\end{equation}
where $\zeta < 1$ is a constant that we specify when solving the convex optimization problem (see \appref{sec:convex-optimization}).
The second constraint in the main text derives from the conditions for phase coexistence in the mean-field model.
Phase coexistence requires that the dimensionless chemical potentials of all components,
\begin{equation}
\mu_i = \frac{\partial F}{\partial \phi_i} = \log \phi_i - \log(1 - \phi_{\text{T}}) - 2\sum_j \chi_{ij} \phi_j,
\end{equation}
and the dimensionless osmotic pressure,
\begin{equation}
\Pi = \sum_i \phi_i \mu_i - F = -\log(1 - \phi_{\text{T}}) - \sum_{i,j} \chi_{ij} \phi_i \phi_j,
\end{equation}
be equal across all phases.
We set $\mu_i^{(\alpha)} = \mu_i^{(0)}$ and $\Pi^{(\alpha)} = \Pi^{(0)}$, and then eliminate the terms involving $\phi_{\text{T}}^{(\alpha)}$ from the equations for the chemical potentials.
Approximating $\phi_{\text{T}}^{(\alpha)} \approx 1$ for each condensed target phase allows us to write the second constraint in the main text as
\begin{equation}
\label{eq:constraint-2}
\sum_{i,j}\left[x_i^{(\alpha)} x_j^{(\alpha)} \bm{1}_i^{(\alpha)} \bm{1}_j^{(\alpha)}\! - 2\delta_{ik} (x_j^{(\alpha)} \bm{1}_j^{(\alpha)}\! - \phi_j^{(0)}) - \phi_i^{(0)}\phi_j^{(0)}\right]\chi_{ij} = -\log\frac{x_k^{(\alpha)}}{\phi_k^{(0)}},
\end{equation}
where ${\bm{1}_i^{(\alpha)} = 1}$ if species $i$ is enriched in phase $\alpha$ and is zero otherwise; as a result of this approximation, the osmotic pressures of the various phases are not exactly equal, and we obtain a grand potential landscape in which the condensed phases are supersaturated relative to the homogeneous phase.
We note that the supersaturation can be adjusted by tuning the volume fractions in the homogeneous phase, $\vec\phi^{(0)}$.
We obtain the third constraint in the main text by eliminating the term $1/(1-\phi_{\text{T}}^{(\alpha)})$ from $\partial^2 F/\partial \phi_i \partial \phi_j |_\alpha$ by again using the approximation $\Pi^{(\alpha)} \simeq \Pi^{(0)}$:
\begin{equation}
\label{eq:constraint-3}
\left.\frac{\partial^2 F}{\partial \phi_i \partial \phi_j}\right|_\alpha = \frac{\delta_{ij}}{\phi_i^{(\alpha)}} + w(\bm{\chi}) - 2\chi_{ij} \succ 0.
\end{equation}
As noted in the main text, $w(\bm{\chi})$ has a negligible effect on the minimum eigenvalue of $\partial^2 F/\partial \phi_i \partial \phi_j |_\alpha$.
We therefore approximate $w(\bm{\chi})$ by a constant expression when solving the optimization problem,
\begin{equation}
w(\bm{\chi}) \simeq \frac{\exp\left[\langle\chi\rangle_{\text{target}} - \big(\phi_{\text{T}}^{(0)}\big)^2 \langle\chi\rangle_{\text{homo}}\right]}{1 - \phi_{\text{T}}^{(0)}},
\end{equation}
where, assuming the enriched components comprise equal volume fractions in all target phases, we have defined the expected averages of $\chi$ in the target, ${\langle\chi\rangle_{\text{target}} \equiv [\log(N / M \phi_{\text{T}}^{(0)}) + (2\phi_{\text{T}}^{(0)} - (\phi_{\text{T}}^{(0)})^2)(N-M)\log(\zeta/N) / 2N] / (1 - \phi_{\text{T}}^{(0)})^2}$, and homogeneous phases, ${\langle\chi\rangle_{\text{homo}} \equiv \langle\chi\rangle_{\text{target}} + (N - M)\log(\zeta/N) / 2N}$.
With these approximations, the inverse problem constraints \eqref{eq:constraint-1}, \eqref{eq:constraint-2}, and \eqref{eq:constraint-3} are linear in $\bm{\chi}$ and only involve quantities specified in the problem definition, i.e., $\vec\phi^{(0)}$ and the compositions $\{x_i^{(\alpha)}\}$ of the enriched components.
\section{Solving the convex optimization problems}
\label{sec:convex-optimization}
When performing numerical tests using the mean-field model, we generate each target phase by randomly choosing $M$ of the $N$ species.
We repeat this procedure to determine which species are enriched in each of the $n$ target phases, and we verify that the targets generated in this way are in fact unique.
The calculations presented in the main text assume an equimolar homogeneous phase with $\phi_{\text{T}}^{(0)} = 0.25$.
When searching for the optimal $\bm{\chi}$ matrix within the SDP and QP formulations, we seek to minimize the objective functions
\begin{equation}
\label{eq:LSDP}
\mathcal{L}_{\text{SDP}} \equiv -\min_{\alpha=0,\ldots,n}\left\{\min_{i=1,\ldots,N}\left[\lambda_i^{(\alpha)}(\bm{\chi})\right]\right\} + \kappa\text{Var}_{i<j}(\bm{\chi})
\end{equation}
or
\begin{equation}
\label{eq:LQP}
\mathcal{L}_{\text{QP}} \equiv \frac{(\phi_{\text{T}}^{(0)})^2}{N} \text{Var}_{i<j}(\bm{\chi}) + \sum_{\alpha=1}^n\frac{1}{M}\text{Var}_{i<j}^{(\alpha)}(\bm{\chi})
\end{equation}
subject to the constraints
\begin{eqnarray}
\sum_{i,j}\left[x_i^{(\alpha)}\!x_j^{(\alpha)} \bm{1}_i^{(\alpha)} \bm{1}_j^{(\alpha)}\! - 2\delta_{ik} (x_j^{(\alpha)} \bm{1}_j^{(\alpha)}\! - \phi_j^{(0)}) - \phi_i^{(0)}\!\phi_j^{(0)}\right]\chi_{ij} &=& -\log\frac{x_k^{(\alpha)}}{\phi_k^{(0)}} \;\;\text{if species }k\text{ is enriched in phase }\alpha, \\
\sum_{i,j}\left[x_i^{(\alpha)}\!x_j^{(\alpha)} \bm{1}_i^{(\alpha)} \bm{1}_j^{(\alpha)}\! - 2\delta_{ik} (x_j^{(\alpha)} \bm{1}_j^{(\alpha)}\! - \phi_j^{(0)}) - \phi_i^{(0)}\!\phi_j^{(0)}\right]\chi_{ij} &\ge& -\log\frac{\zeta}{NM\phi_k^{(0)}} \;\;\text{if species }k\text{ is depleted in phase }\alpha.\qquad
\end{eqnarray}
We obtain the QP objective function by writing the stability matrix, $\partial^2 F/\partial\phi_i\partial\phi_j$, in the form
\begin{equation}
\left.\frac{\partial^2 F}{\partial \phi_i \partial \phi_j}\right|_\alpha =
\left[\begin{array}{@{}c|c@{}}
A & B^\top \\
\hline
B & C^{\phantom\top} \end{array}\right],
\end{equation}
where the submatrix $A$ represents the $M \times M$ block of enriched components in the $\alpha$ phase.
The stability condition \eqref{eq:constraint-3} is equivalent to ${(\partial^2F) / C \equiv A - B^\top C^{-1} B \succ 0}$.
Furthermore, because the submatrix $C$ pertains to the components that are depleted in the $\alpha$ phase, the inverse of $C$ has eigenvalues clustered near $x_i^{(\alpha)} \phi_{\text{T}}^{(\alpha)} \lesssim \zeta / N M_\alpha$.
Finally, by treating $(\partial^2 F)/C$ as a symmetric random matrix, we can seek to maximize its lowest eigenvalue by minimizing the variance of its off-diagonal elements~\cite{wigner1967random,furedi1981eigenvalues}.
This random-matrix approximation is a reasonable assumption if the target-phase compositions are uncorrelated.
The prefactors in \eqref{eq:LQP} are determined from the scaling behavior of the lowest eigenvalue, $\lambda_{\text{min}}$, of $\partial^2 F/\partial\phi_i\partial\phi_j$, assuming that this matrix can be treated as a symmetric random matrix~\cite{furedi1981eigenvalues}:
\begin{equation}
\lambda_{\text{min}}^{(0)} \simeq \frac{N}{\phi_{\text{T}}^{(0)}} - 2\sqrt{N\text{Var}_{i<j}(\bm{\chi})}
\qquad\text{and}\qquad
\lambda_{\text{min}}^{(\alpha)} \simeq M - 2\sqrt{M\text{Var}_{i<j}^{(\alpha)}(\bm{\chi})}\;\text{ for }\alpha=1,\ldots,n.
\end{equation}
Thus, if $\zeta$ is sufficiently small, the mean-field inverse problem has a solution if ${\mathcal{L}_{\text{QP}} \lesssim (n+1)/4}$, in which case the $\bm{\chi}$ matrix that minimizes $\mathcal{L}_{\text{QP}}$ is a good approximation of the $\bm{\chi}$ matrix that minimizes $\mathcal{L}_{\text{SDP}}$.
In the results presented in the main text, the depleted-component partition-coefficient constant $\zeta$ and the regularization constant $\kappa$ are chosen to be $10^{-3}$ and $10^{-5}\!$, respectively.
Numerical optimization is performed using the SCS solver~\cite{odonoghue2016conic} within the CVXPY framework~\cite{diamond2016cvxpy}.
\section{Predicting the thresholding transition}
\label{sec:thresholding}
We now elaborate on the random graph arguments discussed in the main text.
The probability that a pair of components $i$ and $j$ are both enriched in any one of the target phases is
\begin{equation}
p_{\text{edge}} = 1 - \left[1 - \frac{M(M - 1)}{N(N-1)}\right]^n \simeq \frac{nM^2}{N^2}.
\end{equation}
The expectation value for the number of $(M+1)$-sized cliques that contain an $M$-sized target phase in the random graph is
\begin{equation}
\label{eq:EK}
\mathbb{E} \le n (N-M) p_{\text{edge}}^{M}.
\end{equation}
With high probability, a random graph does not contain such a clique if $\mathbb{E} < 1$; thus, we obtain a threshold for $n$,
\begin{equation}
\label{eq:n_threshold}
n^* \sim N^{(2M-1)/(M+1)},
\end{equation}
when $N \gg M$.
As discussed in the main text, we expect that $n_{1/2}$ will scale according to \eqref{eq:n_threshold} in the mean-field model.
We note that the threshold for the appearance of a random $(M+1)$-sized clique in an Erd\H{o}s--R\'{e}nyi random graph $G(N,p_{\text{edge}})$ with uncorrelated edges scales as~\cite{frieze2016introduction}
\begin{equation}
n^*_{K_{M+1}} \sim N^{2(M-1)/M}.
\end{equation}
Our predicted scaling for $n_{1/2}$ grows more slowly with respect to $N$ than $n^*_{K_{M+1}}$ when $N \gg M$ because \eqref{eq:EK} accounts for the fact that $M$-sized cliques exist in the graph by design.
Thus, we expect to encounter $(M+1)$-sized cliques that directly destabilize target phases before we encounter completely random cliques of size $M+1$ or greater.
This expectation is borne out by the close agreement between our predicted scaling, \eqref{eq:n_threshold}, and the empirical scaling exponents determined from power law fits to the data in \figref{fig:2}b,c in the main text.
\section{Performing stability and growth simulations}
\label{sec:simulations}
We perform grand-canonical Monte Carlo (GCMC) simulations of stability and growth using a three-dimensional cubic lattice model.
As described in the main text, we obtain the interaction matrix, $\bm{\epsilon}$, from the mean-field solution by rescaling $\bm{\chi}$.
We then set the chemical potential, $\mu$, to be the same for all components; $\mu$ is chosen such that the target condensed phases are supersaturated relative to the homogeneous phase, and the chemical potential of a vacancy is zero.
Simulations are then carried out by proposing particle exchanges at every lattice site with equal probability and accepting these moves according to the standard Metropolis criterion~\cite{frenkel2001understanding}.
When running stability simulations, we initialize a $10 \times 10 \times 10$ cubic lattice with periodic boundary conditions in each one of the target phases, including the homogeneous phase.
This is accomplished by randomly assigning particles to lattice sites such that the lattice composition is consistent with the target composition, $\vec x^{(\alpha)}$.
We then evolve the lattice via GCMC simulation, keeping track of the order parameter $\cos\theta^{(\alpha)}$.
By construction, $\cos\theta^{(\alpha)} = 1$ at the start of each simulation.
However, $\cos\theta^{(\alpha)}$ is found to decrease markedly whenever the initial phase is unstable with respect to composition fluctuations; we define $\cos\theta^{(\alpha)} \ge 0.95$ to be the threshold for a ``successful'' stability simulation.
By running these simulations for only 100 Monte Carlo sweeps, we ascertain whether the free-energy barriers surrounding each one of the target phases are all higher than a few $kT$.
We note that when target phases are found to be unstable, the simulation usually ends up in a condensed-phase free-energy basin that is enriched in more than $M$ components.
Such chimeric phases tend to have greater compositional entropy and are thus lower in free energy than any of the $M$-component target phases.
When running growth simulations, we initialize a $100 \times 10 \times 10$ periodic cubic lattice with a $10 \times 10 \times 10$ seed of the target phase surrounded by the metastable, homogeneous gas phase.
The initial configurations of both the seed and the gas phase are prepared in the same manner as the stability simulations described above.
When sufficiently supersaturated relative to the gas phase, the seed grows by advancing both interfaces with the gas phase at a constant velocity parallel to the long axis of the simulation box.
Simulations are halted once the condensed phase occupies at least 90\% of the simulation box.
At this point, we classify a simulation as ``successful'' if the composition of the lattice occupied by the condensed phase matches the composition of the target phase using the same order parameter and threshold as above, i.e., if $\cos\theta^{(\alpha)} \ge 0.95$.
\section{Calculating free-energy landscapes}
\label{sec:landscapes}
We perform Wang--Landau simulations~\cite{wang2001efficient} to calculate the grand-potential free-energy surfaces between pairs of target phases in the lattice model, as shown in \figref{fig:4}a in the main text.
Following the method introduced in~\cite{jacobs2013predicting}, we define an order parameter $\Delta\phi_{\alpha\beta}$ for a pair of phases $\alpha$ and $\beta$,
\begin{equation}
\Delta\phi_{\alpha\beta}(\vec\phi) \equiv (\vec\phi - \vec\xi_{\alpha\beta}) \cdot \hat\nu_{\alpha\beta},
\end{equation}
where ${\vec\xi_{\alpha\beta} \equiv (\vec\phi^{(\alpha)} + \vec\phi^{(\beta)}) / 2}$ and ${\hat\nu_{\alpha\beta} \equiv (\vec\phi^{(\beta)} - \vec\phi^{(\alpha)}) / |\vec\phi^{(\beta)} - \vec\phi^{(\alpha)}|}$.
To constrain the simulation to explore the phase space near phases $\alpha$ and $\beta$, we add a harmonic potential in directions of concentration space orthogonal to $\vec\xi_{\alpha\beta}$,
\begin{equation}
U_{\alpha\beta}(\vec\phi) \equiv k_\perp \big|(\vec\phi - \vec\xi_{\alpha\beta}) - [(\vec\phi - \vec\xi_{\alpha\beta}) \cdot \hat\nu_{\alpha\beta}] \hat\nu_{\alpha\beta}\big|^2.
\end{equation}
To improve the efficiency of the simulation, particle exchanges from a lattice site occupied by a particle (or a vacancy) of type $i$ to a particle (or a vacancy) of type $j$ are proposed with probability
\begin{equation}
p_{\text{gen}}(i \rightarrow j) = \begin{cases}
0.5 &\text{if }j\text{ is a vacancy} \\
(0.5 - 0.01) / M_\beta &\text{if }j\text{ is enriched in phase }\beta \\
0.01 / (N - M_\beta) &\text{if }j\text{ is depleted in phase }\beta
\end{cases}
\end{equation}
if $\alpha$ is the homogeneous phase and $\beta$ is a condensed target phase, or with probability
\begin{equation}
p_{\text{gen}}(i \rightarrow j) = \begin{cases}
(1 - 0.01) / M_{\alpha\beta} &\text{if }j\text{ is enriched either in phase }\alpha\text{ or in phase }\beta \\
0.01 / (N + 1 - M_{\alpha\beta}) &\text{otherwise, including if }j\text{ is a vacancy},
\end{cases}
\end{equation}
where $M_{\alpha\beta}$ is the number of components that are enriched either in phase $\alpha$ or in phase $\beta$, if $\alpha$ and $\beta$ are both condensed target phases.
The Metropolis acceptance criterion is appropriately modified to account for these non-uniform generation probabilities.
The free-energy landscapes presented in the main text are obtained using the combined potential, ${\mathcal{H}_{\text{LG}} + U_{\alpha\beta}}$, on an $L \times L \times L$ periodic lattice with ${L = 6}$ and ${k_\perp = 10^4}$.
\figref{fig:4}a shows landscapes for 1,000 randomly selected pairs of target phases and for all 250 homogeneous-target phase pairs.
The barrier heights, $\Omega^\ddagger$, are computed by finding the maximum on the computed free-energy surface between phases $\alpha$ and $\beta$ at coexistence between the two phases.
\end{document}
|
\section{Introduction}
In the notes of this invited talk, we would like to give the reader a short introduction to Interpretable Machine Learning (IML) from the perspective of Formal Concept Analysis (FCA), which can be considered as a mathematical framework for concept learning, Frequent Itemset Mining (FIM) and Association Rule Mining (ARM).
Among the variety of concept learning methods, we selected the rule-based JSM-method named after J.S.~Mill in its FCA formulation. Another possible candidate is Version Spaces. To stress the difference between concept learning paradigm and formal concept we used concept-based learning term in case of usage of FCA as a mathematical tool and language.
We assume, that interpretation by means of game-theoretic attribute ranking is also important in an unsupervised setting as well, and demonstrate its usage via attribution of stability indices of formal concepts (concept stability is also known as the robustness of closed itemset in the FIM community).
Being a convenient language for JSM-method (hypotheses learning) and Frequent Itemset Mining, its direct application to large datasets is possible only under a reasonable assumption on the number of attributes or data sparseness. Direct computation of the Shapley value for a given attribute also requires enumeration of almost all attribute subsets in the intent of a particular object or concept. One of the possibilities to cope with the data volume is approximate computations, while another one lies in the reduction of the number of attributes or their grouping by similarity.
The paper is organised as follows. Section~\ref{sec:RelWork} observes several closely related studies and useful sources on FCA and its applications. Section~\ref{sec:SupLearn} is devoted to concept-based learning where formal intents are used as classification hypotheses and specially tailored Shapley value helps to figure out contributions of attributes in those hypotheses when a particular (e.g., unseen) object is examined. Section~\ref{sec:UnsupLearn} shows that the Shapley value approach can be used for attribution to stability (or robustness) of formal concepts, thus we are able to rank single attributes of formal intents (closed itemsets) in an unsupervised setting. Section~\ref{sec:AttrSimRed} sheds light on the prospects of usage attribute-based similarity of concepts and attribute reduction for possibly large datasets (formal contexts). Section~\ref{sec:Concl} concludes the paper.
\section{Related Work}\label{sec:RelWork}
Formal Concept Analysis is an applied branch of modern Lattice Theory suitable for knowledge representation and data analysis in various domains~\cite{Ganter:1999}. We refer the reader to a modern textbook on FCA with a focus on attribute exploration and knowledge extraction~\cite{Ganter:2016}, surveys on FCA models and techniques for knowledge processing and representation~\cite{Kuznetsov:2013,Poelmans:2013b} as well as on their applications~\cite{Poelmans:2013a}. Some of the examples in subsequent sections are also taken from a tutorial on FCA and its applications~\cite{Ignatov:2014a}.
Since we deal with interpretable machine learning, we first need to establish basic machine learning terminology in FCA terms. In the basic case, our data are Boolean object-attribute matrices or formal contexts, which are not necessarily labeled w.r.t. a certain target attribute. Objects can be grouped into clusters (concept extents) by their common attributes, while attributes compose a cluster (concept intent) if they belong to a certain subset of objects. The pairs of subsets of objects and attributes form the so-called formal concepts, i.e. maximal submatrices (w.r.t. of rows and attribute permutations) of an input context full of ones in its Boolean representation. Those concepts form hierarchies or concept lattices (Galois lattices), which provide convenient means of visualisation and navigation and enables usage of suprema and infima for incomparable concepts.
The connection between well-known concept learning techniques (for example, Version Spaces, and decision tree induction) from machine learning and FCA was well established in~\cite{Ganter:2003,Kuznetsov:2004}. Thus Version Spaces studied by T.~Mitchell~\cite{Mitchell:1977} also provides hierarchical means for hypotheses learning and elimination, where hypotheses are also represented as conjunctions of attributes describing the target concept. Moreover, concept lattices can be used for searching for globally optimal decision trees in the domains where we should not care about the trade-off between time spent for the training phase and reached accuracy (e.g., medical diagnostics) but should rather focus on all valid paths in the global search space~\cite{Belohlavek:2009,Kashnitsky:2016}.
In case we deal with unsupervised learning, concept lattices can be considered as a variant of hierarchical clustering where one has the advantage to use multiple inheritance in both bottom-up and top-down directions~\cite{Carpineto:96,Valtchev:2000,Stumme:2001,Bocharov:2016}. Another fruitful property of formal concepts allows one not only to receive a cluster of objects without any clue why they are similar but to reveal objects' similarity in terms of their common attributes. This property allows considering a formal concept as bicluster~\cite{Mirkin:1996,Ignatov:2012,Kaytoue:2014}, i.e. a biset of two clusters of objects and attributes, respectively.
Another connection between FCA and Frequent Itemset Mining is known for years~\cite{Pasquier:1999,Lakhal:2005}. In the latter discipline, transactions of attributes are mined to find items frequently bought together~\cite{Agrawal:93}. The so-called closed itemsets are used to cope with a huge number of frequent itemsets for large input transaction bases (or contexts), and their definition coincides with the definition of concept intents (under the choice of constraint on the concept extent size or itemset support). Moreover, attribute dependencies in the form of implications and partial implications~\cite{Luxenburger:1991} are known as association rules, which appeared later in data mining as well~\cite{Agrawal:93}\footnote{One of the earlier precursors of association rules can be also found in~\cite{Hajek:1966} under the name of ``almost true implications''}.
This is not a coincidence that we discuss data mining, while stressed interpretability and machine learning in the title. Historically, data mining was formulated as a step of the Knowledge Discovery in Databases process that is ``the nontrivial process of identifying valid, novel, potentially useful, and ultimately understandable patterns in data.''~\cite{Fayyad:96}. While understandable patterns are a must for data mining, in machine learning and AI in general, this property should be instantiated as something extra, which is demanded by analysts to ease decision making as the adjectives explainable (AI) and interpretable (ML) suggest~\cite{Molnar:2019}.
To have a first but quite comprehensive reading on interpretable ML we suggest a freely available book~\cite{Molnar:2019}, where the author states that ``Interpretable Machine Learning refers to methods and models that make the behaviour and predictions of machine learning systems understandable to humans''.
The definition of interpretability may vary from the degree to which a human can understand the cause of a decision to the degree to which a human can consistently predict the model's result.
The taxonomy of IML methods has several aspects. For example, models can be roughly divided into \emph{intrinsic} and \emph{post hoc} ones. The former include simpler models like short rules or sparse linear models, while among the latter black-box techniques with post hoc processing after their training can be found. Some researchers consistently show that in case of the necessity to have interpretable models, one should not use post hoc techniques for black-box models but trust naturally interpretable models~\cite{Rudin:2019}. Another aspect is the universality of the method, the two extremes are \emph{model-specific} (the method is valid for only one type of models) and or \emph{model-agnostic} (all models can be interpreted with the method). There is one more important aspect, whether the method is suitable for the explanation of the model's predictions for a concrete object (\emph{local method}) or it provides an interpretable picture for the entire model (\emph{global method}). Recent views on state-of-the-art techniques and practices can be found in~\cite{Caruana:2020,Kaur:2020}.
FCA provides interpretable patterns a priori since it deals with such understandable patterns as sets of attributes to describe both classes (by means of classification rules or implications) and clusters (e.g., concept intents). However, FCA theory does not suggest the (numeric) importance of separate attributes. Here, a popular approach based on Shapley value from Cooperative Game Theory~\cite{Shapley:1953} recently adopted by the IML community may help~\cite{Strumbelj:2014,Lundberg:2017,Kadyrov:2019}.
The main idea of Shapley value based approaches in ML for ranking separate attributes is based on the following consideration: each attribute is considered as a player in a specific game-related to classification or regression problem and attributes are able to form (winning) coalitions. The importance of such a player (attribute) is computed over all possible coalitions by a combinatorial formula taking into account the number of winning coalitions where without this attribute the winning state is not reachable.
One of the recent popular implementations is SHAP library~\cite{Lundberg:2017}, which however cannot be directly applied to our concept-based learning cases: JSM-hypotheses and stability indices. The former technique assumes that unseen objects can be left without classification or classified contradictory when for an examined object there is no hypothesis for any class or there are at least two hypotheses from different classes~\cite{Finn:1983,Kuznetsov:1990}. This might be an especially important property for such domains as medicine and finance where wrong decisions may lead to regrettable outcomes. We can figure out what are the attributes of the contradictory hypotheses we have but which attributes have the largest positive or negative impact on the classification is still unclear without external means. The latter case of stability indices, which were originally proposed for ranking JSM-hypotheses by their robustness to the deletion of object subsets from the input contexts (similarly to cross-validation)~\cite{Kuznetsov:1991,Kuznetsov:2007a}, is considered in an unsupervised setting. Here, supervised interpretable techniques like SHAP are not directly applicable. To fill the gap we formulated two corresponding games with specific valuation functions used in the Shapley value computations.
Mapping of the two proposed approaches onto the taxonomy of IML methods says that in the case of JSM-hypotheses it is an intrinsic model, but applying Shapley values on top of it is post hoc. At the same time, this concrete variant is rather model-specific since it requires customisation. This one is local since it explains the classification of a single object. As for attribution of concept stability, this one is definitely post hoc, model-specific, and if each pattern (concept) is considered separately this one is rather local but since the whole set of stable concepts can be attributed it might be considered as a global one as well.
It is important to note that one of the stability indices was rediscovered in the Data Mining community and known under the name of the robustness of closed itemsets~\cite{Tatti:2011,Kuznetsov:2018} (where each transaction/object is kept with probability $\alpha=0.5$). So, the proposed approach also allows attribution of closed itemsets.
Classification and direct computation of Shapley values afterwards might be unfeasible for large sets of attributes~\cite{Caruana:2020}. So, we may think of approximate ways to compute Shapley values~\cite{Strumbelj:2014} or pay attention to attribute selection, clarification, and reduction known in the FCA community. We would like to draw the reader's attention to scale coarsening as feature selection tools~\cite{Ganter:2008} and a comparative survey on FCA-based attribute reduction techniques~\cite{Konecny:2017,Konecny:2018}. However, we prefer to concentrate on attribute aggregation by similarity~\footnote{Similarity between concepts is discussed in~\cite{Eklund:2012}} as an attribute reduction technique which will not allow us to leave out semantically meaningful attributes even if they are highly-correlated and redundant in terms of extra complexity paid for their processing otherwise.
The last note on related works, which is unavoidable when we talk about IML, is the relation to Deep Learning (DL) where black-box models predominate~\cite{Shrikumar:2017}. According to the textbook~\cite{Goodfellow:2016}, ``Deep Learning is a form of machine learning that enables computer to learn from experience and understand the world in terms of a hierarchy of concepts.'' The authors also admit that there is no need for a human computer operator to formally specify all the knowledge that the computer needs and obtained hierarchy of concepts allows the computer to learn complicated concepts by building them out of simpler ones. The price of making those concepts intelligible for the computer but not necessary for a human is paid by specially devised IML techniques in addition to DL models.
Since FCA operates with concept hierarchies and is extensively used in human-centric applications~\cite{Poelmans:2013a}, the question ``What can FCA do for DL?'' is open. For example, in~\cite{Rudolph:2007} closure operators on finite sets of attributes were encoded by a three-layered feed-forward neural network, while in~\cite{Kuznetsov:2017} the authors were performing neural architecture search based on concept lattices to avoid overfitting and increase the model interpretability.
\section{Supervised Learning: From Hypotheses to Attribute Importance}\label{sec:SupLearn}
In this section, we discuss how interpretable concept-based learning for JSM-method can be achieved with Shapley Values following our previous study on the problem~\cite{Ignatov:2020a}.
Let us start with a piece of history of inductive reasoning. In XIX century, John Stuart Mill proposed several schemes of inductive reasoning. Let us consider, for example, the Method of Agreement~\cite{Mill:1843}:
``If two or more instances of the phenomenon under investigation have only one circumstance in common, ... [it] is the cause (or effect) of the given phenomenon.''
The JSM-method (after J.S. Mill) of hypotheses generation proposed by Viktor K. Finn in the late 1970s is an attempt to describe induction in purely deductive form~\cite{Finn:1983}.
This new formulation was introduced in terms of many-valued many-sorted extension of the First Order Predicate Logic~\cite{Kuznetsov:2005}.
This formal logical treatment allowed usage of the JSM-method as a machine learning technique~\cite{Kuznetsov:1991}. While further algebraic redefinitions of the logical predicates to express similarity of objects as an algebraic operation allowed the formulation of JSM-method as a classification technique in terms of formal concepts~\cite{Kuznetsov:1996,Kuznetsov:2005}.
\subsection{JSM-hypotheses in FCA}
In FCA, a formal concept consists of an extent and an intent. The intent is formed by all attributes that describe the concept, and the extent contains all objects belonging to the concept. In FCA, the JSM-method is known as rule-based learning from positive and negative examples with rules in the form ``concept intent $\to$ class''.
Let a formal context $\mathbb{K} :=(G, M, I)$ be our universe, where the binary relation $I \subseteq G \times M$ describes if an object $g\in G$ has an attribute $m\in M$. For $A\subseteq G$ and $B\subseteq M$ the derivation (or Galois) operators are defined by:
\[ A' =\{\,m\in M \mid \forall a\in A\, aIm\,\} \text{ and } B' =\{\,g\in G \mid \forall b\in B\, gIb\,\} . \]
A (formal) concept is a pair $(A,B)$ with $A\subseteq G$, $B\subseteq M$ such that $A'=B$ and $B'=A$. We call $B$ its intent and $A$ its extent. An implication of the form $H\to m$ holds if all objects having the attributes in $H$ also have the attribute $m$, i.e. $H' \subseteq m'$.
The set of all concepts of a given context $\context$ is denoted by $\BGMI$; the concepts are ordered by the ``to be a more general concept'' relation as follows: $(A,B) \geq (C,D) \iff C \subseteq A \ (\mbox{equivalently } B \subseteq D)$.
The set of all formal concepts $\BGMI$ together with the introduced relation form the \emph{concept lattice}, which line diagram is useful for visual representation and navigation through the concept sets.
Let $w\notin M$ be a \emph{target attribute}, then $w$ partitions $G$ into three subsets:
\begin{itemize}
\item \emph{positive examples}: $G_+\subseteq G$ of objects known to satisfy $w$,
\item \emph{negative examples}: $G_-\subseteq G$ of objects known not to have $w$,
\item \emph{undetermined examples}: $G_{\tau}\subseteq G$ of objects for which it
remains unknown whether they have the target attribute or do not have it.
\end{itemize}
This partition gives rise to three subcontexts $\context_\varepsilon:=(G_\varepsilon,M,I_\varepsilon)$ with $ \varepsilon \in \{-, +, \tau\}$.
\begin{itemize}
\item The \emph{positive context} $\context_+$ and the \emph{negative context} $\context_-$ form the training set called by \emph{learning context}:
$$\context_\pm = (G_+ \cup G_-, M \cup \{w\}, I_+ \cup I_- \cup G_+\times \{w\}).$$
\item The subcontext $\context_\tau$ is called the \emph{undetermined context} and is used to predict the class of not yet classified objects.
\end{itemize}
The whole \emph{classification context} is the context
$$\context_c = (G_+ \cup G_- \cup G_\tau, M \cup \{w\}, I_+ \cup I_- \cup I_\tau \cup G_+\times \{w\}).$$ The derivation operators in the subcontexts $\context_\varepsilon$ are denoted by $(\cdot)^+$ $(\cdot)^-$, and $(\cdot)^\tau$, respectively. The goal is to classify the objects in $G_{\tau}$ with respect to $w$.
To do so let us form the positive and negative hypotheses as follows.
A \emph{positive hypothesis} $H\subseteq M$ ($H\neq \emptyset$)
is a intent of $\context_+$ that is not contained in the intent of a
negative example; i.e. $H^{++} = H$ and $H'\subseteq G_+\cup G_{\tau}$ ($H \to w$).
A \emph{negative hypothesis} $H\subseteq M$ ($H\neq \emptyset$)
is an intent of $\context_-$ that is not contained in the intent of a
positive example; i.e. $H^{--} = H$ and $H'\subseteq G_-\cup G_{\tau}$ ($H \to \overline{w}$).
An intent of $\context_+$ that is contained in the intent of a negative example is called a \emph{falsified (+)-generalisation}. A \emph{falsified (-)-generalisation\emph} is defined in a similar way.
To illustrate these notions we use the credit scoring context in Table~\ref{tbl:scor}~\cite{Ignatov:2014b}. Note that we use \emph{nominal scaling} to transform many-valued context to one-valued context~\cite{Ganter:1999} with the following attributes, $Ma$, $F$ (for two genders), $Y$, $MI$, $O$ (for young, middle, and old values of the two-valued attribute Age , resp.), $HE$, $Sp$, $SE$ (for higher, special, and secondary education, resp.), $Hi$, $L$, $A$ (for high, low, and average salary, resp.), and $w$ and $\overline{w}$ for the two-valued attribute Target.
\begin{table}[htbp]
\centering
\caption{Many-valued classification context for credit scoring}\label{tbl:scor}
\begin{tabular}{|c||c|c|c|c|c|}
\hline
G / M & Gender & Age & Education & Salary & Target\\
\hline
\hline
1 & Ma & young & higher & high & $+$\\
2 & F & middle & special & high & $+$\\
3 & F & middle & higher & average & $+$\\
4 & Ma & old & higher & high & $+$\\
\hline
5 & Ma & young & higher & low & $-$\\
6 & F & middle & secondary & average &$-$\\
7 & F & old & special & average &$-$\\
8 & Ma & old & secondary & low &$-$\\
\hline
9 & F & young & special & high & $\tau$\\
10 & F & old & higher & average & $\tau$\\
11 & Ma & middle & secondary & low & $\tau$\\
12 & Ma & old & secondary & high & $\tau$\\
\hline
\end{tabular}
\end{table}
\begin{figure}[h]
\begin{minipage}[h]{0.45\linewidth}
\centering
\begin{picture}(150,160)
\unitlength 0.20mm
\begin{diagram}{270}{260}
\Node{1}{155}{252}
\Node{2}{157}{17}
\Node{3}{84}{148}
\Node{4}{75}{80}
\Node{5}{137}{83}
\Node{6}{152}{202}
\Node{7}{245}{145}
\Node{8}{260.0}{80.0}
\Node{9}{198}{80.0}
\Node{10}{220.0}{200.0}
\Edge{6}{1}
\Edge{7}{10}
\Edge{3}{1}
\Edge{4}{3}
\Edge{5}{10}
\Edge{2}{4}
\Edge{4}{6}
\Edge{2}{9}
\Edge{8}{7}
\Edge{9}{7}
\Edge{2}{5}
\Edge{5}{3}
\Edge{10}{1}
\Edge{7}{6}
\Edge{2}{8}
\NoDots
\centerObjbox{4}{0}{13}{$g_3$}
\rightObjbox{5}{10}{10}{$g_2$}
\centerObjbox{8}{0}{10}{$g_1$}
\centerObjbox{9}{0}{13}{$g_4$}
\leftAttbox{2}{25}{0}{L,SE}
\leftAttbox{3}{0}{10}{F, Mi}
\leftAttbox{4}{0}{10}{A}
\centerAttbox{5}{0}{13}{Sp}
\rightAttbox{6}{3}{10}{HE}
\rightAttbox{7}{0}{10}{Ma}
\centerAttbox{8}{3}{10}{Y}
\centerAttbox{9}{0}{13}{O}
\centerAttbox{10}{0}{10}{HS}
\CircleSize{11}
\end{diagram}
\put(-195,80){\color{tomato}{\circle*{10}}}%
\put(-50,200){\color{tomato}{\circle*{10}}}%
\put(-133,83){\color{lightgray}{\circle*{10}}}%
\put(-72,80){\color{lightgray}{\circle*{10}}}%
\put(-10,80){\color{lightgray}{\circle*{10}}}%
\put(-25,145){\color{lightgray}{\circle*{10}}}%
\end{picture}
\end{minipage}
\hspace{2mm}
\begin{minipage}[h]{0.45\linewidth}
\begin{picture}(100,100)
\unitlength 0.20mm
\begin{diagram}{210}{200}
\Node{1}{160.0}{200}
\Node{2}{158.0}{11.0}
\Node{3}{33.0}{150.0}
\Node{4}{115.0}{80.0}
\Node{5}{30.0}{80.0}
\Node{6}{270.0}{145.0}
\Node{7}{200.0}{80.0}
\Node{8}{270.0}{80.0}
\Node{9}{127.0}{142.0}
\Node{10}{195.0}{140.0}
\Edge{2}{5}
\Edge{4}{10}
\Edge{7}{6}
\Edge{7}{9}
\Edge{2}{7}
\Edge{8}{10}
\Edge{9}{1}
\Edge{5}{3}
\Edge{6}{1}
\Edge{3}{1}
\Edge{10}{1}
\Edge{2}{4}
\Edge{8}{6}
\Edge{2}{8}
\Edge{4}{3}
\Edge{4}{9}
\NoDots
\centerObjbox{4}{-5}{13}{$g_8$}
\centerObjbox{5}{0}{13}{$g_5$}
\centerObjbox{7}{5}{13}{$g_6$}
\centerObjbox{8}{0}{13}{$g_7$}
\centerAttbox{2}{0}{-20}{HS}
\centerAttbox{3}{0}{13}{L,Ma}
\centerAttbox{5}{25}{8}{HE,Y}
\centerAttbox{6}{0}{13}{A,F}
\centerAttbox{7}{0}{13}{Mi}
\centerAttbox{8}{10}{10}{Sp}
\centerAttbox{9}{-3}{13}{Se}
\centerAttbox{10}{3}{13}{O}
\CircleSize{11}
\end{diagram}
\put(-177,150){\color{lightseagreen}{\circle*{10}}}%
\put(-83,142){\color{lightseagreen}{\circle*{10}}}%
\put(60,80){\color{lightseagreen}{\circle*{10}}}%
\put(-10,80){\color{lightgray}{\circle*{10}}}%
\put(-95,80){\color{lightgray}{\circle*{10}}}%
\put(-180,80){\color{lightgray}{\circle*{10}}}%
\end{picture}
\end{minipage}
\caption{The line diagrams of the lattice of positive context (left) and the lattice of negative context (right).}
\label{fig:posneglatts}
\end{figure}
For example, the intent of the red node labelled by the attribute $A$ in the left line diagram (Fig.~\ref{fig:posneglatts}), is $\{A, Mi, F, HE \}$, and this is not contained in the intent of any node labelled by the objects $g_5, g_6, g_7$, and $g_8$. So we believe in the rule $H\to w$. Note that the colours of the nodes in Fig.~\ref{fig:posneglatts} represent different types of hypotheses: the red ones correspond to minimal hypotheses (cf. the definition below), the see green nodes correspond to negative hypotheses, while light grey nodes correspond to non-minimal positive and negative hypotheses for the left and the right line diagrams, respectively.
The undetermined examples $g_\tau$ from $G_{\tau}$ are classified according to the following rules:
\begin{itemize}
\item If $g^{\tau}_\tau$ contains a positive, but no negative hypothesis, then $g_\tau$ is \emph{classified positively}.
\item If $g^{\tau}_\tau$ contains a negative, but no positive hypothesis, then $g_\tau$ is \emph{classified negatively}.
\item If $g^{\tau}_\tau$ contains both negative and positive hypotheses, or if $g^{\tau}_\tau$
does not contain any hypothesis, then this object classification is \emph{contradictory} or \emph{undetermined}, respectively.
\end{itemize}
To perform classification by the aforementioned rules, it is enough to have only \emph{minimal hypotheses} (w.r.t. $\subseteq$) of both signs.
Let $\mathcal{H}_+$ (resp. $\mathcal{H}_-$) be the set of minimal positive (resp. minimal negative) hypotheses. Then,
\begin{small}
\begin{align*}
\mathcal{H}_+ = \big\{ \{F, Mi, HE, A\}, \{HS\} \big\} \text{ and }
\mathcal{H}_- = \big\{ \{F, O, Sp, A\}, \{Se\}, \{Ma, L\} \big\} .
\end{align*}
\end{small}
We proceed to classify the four undetermined objects below.
\begin{itemize}
\item $g_9'=\{F, Y, Sp, HS\}$ contains the positive hypothesis $\{HS\}$, and no negative hypothesis. Thus, $g_9$ is classified positively.
\item $g_{10}'=\{F, O, HE, A\}$ does not contain neither positive nor negative hypotheses. Hence, $g_{10}$ remains undetermined.
\item $g_{11}'=\{Ma, Mi, Se, L\}$ contains two negative hypotheses: $\{Se\}$ and $\{Ma, L\}$, and no positive hypothesis. Therefore, $g_{11}$ is classified negatively.
\item $g_{12}'=\{Ma, O, Se, HS\}$ contains the negative hypothesis $\{Se\}$ and the positive hypothesis $\{HS\}$, which implies that $g_{12}$ remains undetermined.
\end{itemize}
Even though we have a clear explanation of why a certain object belongs to one of the classes in terms of contained positive and negative hypotheses, the following question arises: Do all attributes play the same role in the classification of certain examples? If the answer is no, then one more question appears: How can we rank attributes with respect to their importance in classifying examples, for example, $g_{11}$ with attributes $ Ma, Mi, Se$, and $L$? Game Theory offers several indices for such comparison: e.g., the Shapley value and the Banzhaf index. For the present contribution, we concentrate on the use of Shapley values.
\subsection{Shapley values and JSM-hypotheses}
To answer the question ``What are the most important attributes for classification of a particular object?'' in our case, we follow to basic recipe studied in \cite{Strumbelj:2014,Lundberg:2017,Molnar:2019}.
To compute the Shapley value for an example $x$ and an attribute $m$, one needs to define $f_x(S)$, the expected value of the model prediction conditioned on a subset $S$ of the input attributes.
\begin{equation}\label{eq:shapley}
\phi_m = \sum_{S \subseteq M \setminus \{m\}} \frac{|S|!(|M| - |S| -1)!}{|M|!} \left ( f_x(S \cup \{m\}) - f_x(S) \right ),
\end{equation}
where $M$ is the set of all input attributes and $S$ a certain coalition of players, i.e. set of attributes.
Let $\context_c = (G_+ \cup G_- \cup G_\tau, M \cup \{w\}, I_+ \cup I_- \cup I_\tau \cup G_+\times \{w\})$ be our classification context, and $\mathcal{H_+}$ (resp. $\mathcal{H_-}$) the set of minimal positive (resp. negative) hypotheses of $\context_c$.
Since we deal with hypotheses (i.e. sets of attributes) rather than compute the expected value of the model's prediction, we can define a valuation function $v$ directly. For $g \in G$, the \emph{Shapley value of an attribute} $m \in g'$:
\begin{equation}\label{eq:shapley_hyp}
\varphi_m (g) = \sum_{S \subseteq g' \setminus \{m\}} \frac{|S|!(|g'| - |S| -1)!}{|g'|!} \left ( v(S \cup \{m\}) - v(S) \right ),
\end{equation}
where
$$v(S)=\begin{cases}
1, & \exists H_+ \in \mathcal{H_+}: H_+ \subseteq S \mbox{ and } \forall H_- \in \mathcal{H_-}: H_- \not\subseteq S, \\
-1, & \exists H_- \in \mathcal{H_-}: H_- \subseteq S \mbox{ and } \forall H_+ \in \mathcal{H_+}: H_+ \not\subseteq S \\
0, & \text{otherwise}
\end{cases} $$
The Shapley value $\varphi_{m}(g)$ is set to 0 for every $m \in M\setminus g'$. The Shapley vector for a given object $g$ is denoted by $\varPhi(g)$. To differentiate between the value in cases when $m \in M\setminus g'$ and $m \in g'$, we will use decimal separator as follows, 0 and 0.0, respectively.
For the credit scoring context, the minimal positive and the negative hypotheses are
\begin{align*}
\mathcal{H}_+ &= \{\{F, Mi, HE, A\}, \{HS\}\}; \
\mathcal{H}_- = \{\{F, O, Sp, A\}, \{Se\}, \{M, L\}\}.
\end{align*}
The Shapley values for JSM-hypotheses have been computed with our freely available Python scripts\footnote{\url{https://github.com/dimachine/Shap4JSM}} for the objects in $G_\tau$:
\begin{itemize}
\item $g_9'=\{F, Y, Sp, HS\} \supseteq \{HS\}$, and $g_9$ is classified positively. $\varphi_\text{HS}(g_9)=1$ and
and its Shapley vector is
$\varPhi(g_9)=(0, 0.0, 0.0, 0, 0, 0, 0.0, 0, 1.0, 0, 0) \mbox{ .}$
\item $g_{10}'=\{F, O, HE, A\}$ and $g_{10}$ remains undetermined. Its Shapley vector is $\varPhi(g_{10})=(0, 0.0, 0, 0, 0.0, 0.0, 0, 0, 0, 0.0, 0) \mbox{ .}$
\item $g_{11}'=\{Ma, Mi, Se, L\}\supseteq \{Se\},\ \{Ma, L\}$. Its Shapley vector is \\
${\varPhi(g_{11})=(-1/6, 0, 0, 0.0, 0, 0, 0, -2/3, 0, 0, -1/6)} \mbox{ .}$
\item $g_{12}'=\{Ma, O, Se, HS\}\supseteq \{HS\},\ \{Se\}$. $\varphi_\text{Se}(g_{12})=-1$, $\varphi_\text{HS}(g_{12})=1$. Its Shapley vector is $\varPhi(g_{12})=(0.0, 0, 0, 0, 0.0, 0, 0, -1.0, 1.0, 0, 0) \mbox{ .}$
\end{itemize}
Let us examine example $g_{11}$. Its attribute $Mi$ has zero importance according to the Shapley value approach since it is not in any contained hypothesis used for the negative classification. The most important attribute is $Se$, which is alone two times more important than the attributes $Ma$ and $L$ together. It is so, since the attribute $Se$, which is the single attribute of the negative hypothesis $\{Se\}$, forms more winning coalitions $S \cup\{Se\}$ with $v(S \cup \{Se\}) - v(S)=1$ than $Ma$ and $L$, i.e. six vs. two. Thus, $\{Se\}\uparrow \setminus \{Ma, L\} \uparrow = \{\{Se\}, \{Ma, Se\}, \{Mi, Se\}, \{Se, L\}, \{Mi, Se, L\},$ $\{Ma, Mi, Se\}\} $\footnote{$S\uparrow$ is the up-set of $S$ in the Boolean lattice $(\mathcal{P}\{Ma, Mi, Se, L\},\subseteq)$} are such winning coalitions for $Se$, while $\{Ma, L\}$, $\{Ma, Mi, L\}$, are those for $Ma$ and $L$.
The following properties hold:
\begin{theorem}[\cite{Ignatov:2020a}]
The Shapley value, $\varphi_m(g)$, of an attribute $m$ for the JSM-classification of an object $g$, fulfils the following properties:
\begin{enumerate}
\item $\sum\limits_{m \in g'}\varphi_m(g)=1$ if $g$ is classified positively;
\item $\sum\limits_{m \in g'}\varphi_m(g)=-1$ if $g$ is classified negatively.
\item $\sum\limits_{m \in g'}\varphi_m(g)=0$ if $g$ is classified contradictory or undetermined.
\end{enumerate}
\end{theorem}
The last theorem expresses the so-called \emph{efficiency property} or axiom~\cite{Shapley:1953}, where it is stated that the sum of Shapley values of all players in a game is equal to the total pay-off of their coalition, i.e. $v(g')$ in our case.
It is easy to check $\varphi_m(g)=0$ for every $m \in g'$ that does not belong to at least one positive or negative hypothesis contained in $g'$. Moreover, in this case for any $S \subseteq g'\setminus \{m\}$ it also follows $v(S) = v(S \cup \{m\})$ and these attributes are called \emph{null} or \emph{dummy players}~\cite{Shapley:1953}.
We also performed experiments on the Zoo dataset\footnote{\url{https://archive.ics.uci.edu/ml/datasets/zoo}}, which includes 101 examples (animals) and their 17 attributes along with the target attribute (7 classes of animals).
The attributes are binary except for the number of legs, which can be scaled nominally and treated as categorical.
We consider a binary classification problem where \emph{birds} is our positive class, while all the rest form the negative class.
There are 19 positive examples (birds) and 80 negative examples since we left out two examples for our testing set, namely, chicken and warm.
The hypotheses are
$\mathcal{H}_+ = \big\{\{ feathers, eggs, backbone, breathes, legs_2, tail \}\big\}$ and
\begin{small}
\begin{align*}
\mathcal{H}_-= &\big\{\{venomous\},
\{eggs, aquatic, predator, legs_5\},
\{legs_0\},
\{eggs, legs_6\},\\
& \{predator, legs_8\},
\{hair, breathes \},
\{milk, backbone, breathes \},
\{legs_4 \},\\
& \{toothed, backbone\}
\big\}.
\end{align*}
\end{small}
The intent $aardvark'=\{hair, milk, predator, toothed, backbone, breathes, legs_4, catsize\}$ contains four negative hypotheses and no positive one.
\footnotesize
The Shapley vector for the \textbf{aardvark} example i
$$(-0.1,
0,
0,
-0.0167,
0,
0,
0.0,
-0.1,
-0.133,
-0.133,
0,
0,
-0.517,
0,
0,
0,
0,
0,
0,
0,
0.0) \mbox{ .}$$
\textbf{Backbone}, \textbf{breathes}, and \textbf{four legs} are the most important attributes with
values -0.517, -0.133, and -0.133, respectively, while \textbf{catsize} is not important in terms of Shapley value.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.6\linewidth]{pics/aardvark}
\caption{The Shapley vector diagram for the \textbf{aardvark} example} \label{fig:lollipop}
\end{figure}
A useful interpretation of classification results could be an explanation for true positive or true negative cases. However, in the case of our test set both examples, \textbf{chicken} and \textbf{warm}, are classified correctly as bird and non-bird, respectively. Let us have a look at their Shapley vectors. Our test objects have the following intents:
$$chicken'=\{feathers, eggs, airborne, backbone, breathes, legs_2, tail, domestic\}$$
and
$$warm'=\{eggs, breathes, legs_0\} .$$
\begin{figure}[ht!]
\centering
\begin{minipage}[h]{0.49\linewidth}
\centering
\includegraphics[width=1\linewidth]{pics/chicken}
\end{minipage}
\begin{minipage}[h]{0.49\linewidth}
\centering
\includegraphics[width=1\linewidth]{pics/warm}
\end{minipage}
\caption{The Shapley vector diagram for the \textbf{chicken} (left) and \textbf{warm} (right) examples} \label{fig:lollipop}
\end{figure}
Thus, for the \textbf{chicken} example all six attributes that belong to the single positive hypothesis have equal Shapley values, i.e. 1/6. The attributes \textbf{airborne} and \textbf{domestic} have zero importance. The \textbf{warm} example has only one attribute with non-zero importance, i.e. the absence of legs with importance -1. It is so since the only negative hypothesis, $\{legs_0\}$, is contained in the object intent.
\section{Unsupervised Learning: Contribution to Stability and Robustness}\label{sec:UnsupLearn}
(Intensional) stability indices were introduced to rank the concepts (intents) by their robustness under objects deletion and provide evidence of the non-random nature of concepts~\cite{Roth:2006}. The extensional stability index is defined as the proportion of intent subsets generating this intent; it shows the robustness of the concept extent under attributes deletion~\cite{Roth:2006}. Our goal here is to find out whether all attributes play the same role in the stability indices. To measure the importance of an attribute for a concept intent, we compare generators with this attribute to those without it. In this section, we demonstrate how Shapley values can be used to assess attribute importance for concept stability.
\subsection{Stability indices of a concept}
Let $\mathbb{K} :=(G, M, I)$ be a formal context. For any closed subset $X$ of attributes or objects, we denote by $\mathrm{gen}(X)$ the set of generating subsets of $X$. The \emph{extensional stability index}~\cite{Roth:2006} of a concept $(A,B)$ is
\begin{align*}
\sigma_e(A,B) &
= \frac{|\{Y \subseteq B \mid Y''=B\}|}{2^{|B|}} = \frac{|\mathrm{gen}(B)|}{2^{|B|}}.
\end{align*}
We can also restrict to generating subsets of equal size. The extensional stability index of the $k$-th level of $(A,B)$ is
$$J_k(A,B):=|\{Y \subseteq B\mid |Y|=k, Y''=B\}| \Big/\binom{|B|}{k}.$$
\subsection{Shapley vectors of intents for concept stability}
Let $(A,B)$ be a concept of $(G,M,I)$ and $m\in B$. We define an indicator function by
$$v(Y)= 1 \mbox{ if } Y''=B \text{ and } \ Y\neq \emptyset, \mbox{ and } v(Y) = 0 \text{ otherwise}. $$
Using the indicator $v$, the Shapley value of $m \in B$ for the stability index of the concept $(A,B)$ is defined by:
\begin{equation}\label{eq:shapley_stab}
\varphi_m(A,B):=\frac{1}{|B|}\sum\limits_{Y\subseteq B\setminus \{m\}} \frac{1}{\binom{|B|-1}{|Y|}}\Big(v(Y \cup \{m\})-v(Y)\Big) .
\end{equation}
The Shapley vector of $(A,B)$ is then $\left(\varphi_m(A,B)\right)_{m\in B}$.
An equivalent formulation is given using upper sets of minimal generators~\cite{Ignatov:2020b}. In fact, for $m \in X_m \in \mathrm{mingen}(B)$ and $m \notin X_{\overline{m}}\in \mathrm{mingen}(B)$, we have
\begin{equation*}\label{eqn:shap-stab_0}
\varphi_m(A,B)=\frac{1}{|B|}\sum\limits_{D\sqcup\{m\} \in \bigcup X_m\uparrow\setminus \bigcup X_{\overline{m}}\uparrow} \frac{1}{\binom{|B|-1}{|D|}},
\end{equation*}
where $\sqcup$ denotes the disjoint union, $X_m$ and $X_{\overline{m}}$ the minimal generators of $B$ with and without $m$, respectively.
\begin{figure}[htbp]
\centering
\includegraphics[width=.25\linewidth]{pics/BooleanLattice.pdf} \hspace*{2cm}
\includegraphics[width=.25\linewidth]{pics/Mingen.pdf}
\caption{Computing Shapley vectors for concept stability}
\end{figure}
To compute $\varphi_m$, additional simplifications are useful:
\begin{theorem}[\cite{Ignatov:2020b}]
Let $(A,B)$ be a concept and $m \in B$.
\begin{itemize}
\item[(i)] $\varphi_m(A,B)=\sum\limits_{k=1}^{|B|} \frac{J_k(A,B)}{k} - \sum\limits_{D \subseteq B\setminus\{m\}}\frac{1}{|D|\binom{|B|-1}{|D|}}v(D)$.
\item[(ii)] If $m \in X_m \in \mathrm{mingen}(B)$ and $Y \subseteq B\setminus \{m\}$ with $(A,B)\prec (Y',Y)$ then
\begin{equation*}
\varphi_m(A,B)=\frac{1}{|B|}\sum\limits_{D \in \bigcup [X_m\setminus{\{m\}},Y]} \frac{1}{\binom{|B|-1}{|D|}}.
\end{equation*}
\item[(iii)] If $m \in X \in \mathrm{mingen}(B)$ and $|\mathrm{mingen}(B)|=1$, then
\begin{equation}\label{eqn:shap-stab-mingen}
\varphi_m(A,B)=\sum\limits_{k=1}^{|B|} \frac{J_k(A,B)}{k}=\frac{1}{|X|} \ .
\end{equation}
\end{itemize}
\end{theorem}
To illustrate the importance of attributes in concept stability, we consider the
the fruits context~\cite{Kuznetsov:2004}, where we extract the subcontext with the first four objects (Table~\ref{tab:fruits}).
\begin{table}[]
\caption{A many-valued context of fruits}
\label{tab:fruits}
\begin{center}
\begin{tabular}{|c|c|cccc|}
\hline
& G $\setminus$ M & color & f{i}rm & smooth & form \\
\hline
1& apple & yellow & no& yes & round \\
2& grapefruit & yellow & no& no & round \\
3& kiwi & green & no& no& oval\\
4& plum & blue & no& yes& oval \\
\hline
\end{tabular}
\end{center}
\end{table}
After scaling we get the binary context and its concept lattice diagram (Fig.~\ref{fig:fruits_latt}).
\begin{figure}
\centering
\begin{minipage}{0.49\linewidth}
\begin{cxt}%
\cxtName{Fruits}%
\att{w}
\att{y}
\att{g}
\att{b}
\att{f}
\att{$\bar{f}$}
\att{s}
\att{$\bar{s}$}
\att{r}
\att{$\bar{r}$}
\obj{.x...xx.x.}{1 apple}
\obj{.x...x.xx.}{2 grapefruit}
\obj{..x..x.x.x}{3 kiwi}
\obj{...x.xx..x}{4 plum}
\end{cxt}
\end{minipage}
\begin{minipage}{0.49\linewidth}
\begin{picture}(100,130)
\unitlength 0.20mm
\begin{diagram}{270}{200}
\Node{1}{160.0}{200}
\Node{2}{160.0}{20}
\Node{3}{60.0}{140}
\Node{4}{60.0}{80}
\Node{5}{120.0}{80}
\Node{6}{260.0}{140}
\Node{7}{200.0}{80}
\Node{8}{260.0}{80}
\Node{9}{120.0}{140}
\Node{10}{200.0}{140}
\Edge{9}{1}
\Edge{4}{9}
\Edge{10}{1}
\Edge{4}{3}
\Edge{2}{5}
\Edge{8}{6}
\Edge{3}{1}
\Edge{7}{6}
\Edge{2}{4}
\Edge{5}{3}
\Edge{5}{10}
\Edge{2}{8}
\Edge{7}{9}
\Edge{8}{10}
\Edge{6}{1}
\Edge{2}{7}
\leftObjbox{4}{0}{13}{$g_3$}
\centerObjbox{5}{0}{13}{$g_4$}
\centerObjbox{7}{0}{13}{$g_2$}
\rightObjbox{8}{0}{13}{$g_1$}
\centerAttbox{1}{0}{13}{$\bar{f}$}
\centerAttbox{2}{0}{20}{f,w}
\centerAttbox{3}{0}{13}{$\bar{r}$}
\leftAttbox{4}{-10}{13}{g}
\centerAttbox{5}{0}{13}{b}
\centerAttbox{6}{0}{13}{r,y}
\centerAttbox{9}{0}{13}{$\bar{s}$}
\centerAttbox{10}{0}{13}{s}
\CircleSize{10}
\end{diagram}
\end{picture}
\end{minipage}
\caption{A scaled fruits context and the line diagram of its concept lattice}
\label{fig:fruits_latt}
\end{figure}
\noindent
For each concept, the stability index $\sigma_e$ and its Shapley vector $\phi$ are computed.
\begin{table}[]
\caption{The concepts of fruits context and their stability indices along with Shapley vectors}
\label{tab:my_label}
\centering
\setlength{\tabcolsep}{5pt}
\renewcommand{\arraystretch}{1.2}
\begin{tabular}{|c|l|c|c|}
\hline
Concepts & $\sigma_e$ & $\Phi$
\\ \hline
$(\{4\}, \{b, \bar{f}, s, \bar{r}\})$ & 0.625 & (2/3, 0.0, 1/6, 1/6)
\\ \hline
$(\{3\}, \{g, \bar{f}, \bar{s}, \bar{r}\})$ & 0.625 &
(2/3, 0.0, 1/6, 1/6)
\\ \hline
$(\{3, 4\}, \{\bar{f}, \bar{r}\})$ & 0.5 & (0.0, 1.0)
\\ \hline
$(\{2\}, \{y, \bar{f}, \bar{s}, r\})$ & 0.375 & (1/6, 0.0, 2/3, 1/6)
\\ \hline
$(\{2, 3\}, \{\bar{f}, \bar{s} \})$ & 0.5 & (0.0, 1.0)
\\ \hline
$(\{1\}, \{y, \bar{f}, s, r\})$ & 0.375 & (1/6, 0.0, 2/3, 1/6)
\\ \hline
$(\{1, 4\}, \{\bar{f}, s\})$ & 0.5 & (0.0, 1.0)
\\ \hline
$(\{1, 2\}, \{y, \bar{f}, r\})$ & 0.75 & (0.5, 0.0, 0.5)
\\ \hline
$(\{1, 2, 3, 4\}, \{\bar{f}\})$ & 1 & (0.0)
\\ \hline
\end{tabular}
\medskip
\begin{tabular}{|c|}
\hline
$\sigma_e(\emptyset, \{w, y, g, b, f, \bar{f}, s, \bar{s}, r, \bar{r}\})=0.955$\\
\hline
$\Phi=(0.256, 0.069, 0.093, 0.093, 0.260, 0.0, 0.052, 0.052, 0.069, 0.052)$ \\
\hline
\end{tabular}
\end{table}
For the Zoo dataset we obtain 357 concepts in total. The top-3 most stable are $c_1, c_2, c_3$ with extent stability indices:
$\sigma_e(G,\emptyset)=1$, $\sigma_e(\emptyset,M)=0.997$,
$\sigma_e(A, A')=0.625$, respectively, where \\
$A' = \{feathers, eggs, backbone, breathes, legs_2, tail\}$ and
\noindent $A=\{11, 16, 20, 21, 23, 33, 37, 41, 43, 56, 57, 58, 59, 71, 78, 79, 83, 87, 95, 100\} \ .$
\begin{footnotesize}
\begin{figure}[ht!]
\begin{minipage}[b]{0.5\linewidth}
\includegraphics[width=1\linewidth]{pics/Shap_0}
\end{minipage
\begin{minipage}[b]{0.5\linewidth}
\includegraphics[width=1\linewidth]{pics/Shap_174}
\end{minipage}
\caption{The Shapley vector for concept $c_2=(\emptyset,M)$ (left) and $c_3$ (right)}
\end{figure}
The most important attributes are \textbf{six legs}, \textbf{eight legs}, \textbf{five legs}, \textbf{feathers}, and \textbf{four legs} for $c_2$, and
\textbf{feathers}, \textbf{eggs}, and \textbf{two legs} for $c_3$, w.r.t. to the Shapley vectors.
\end{footnotesize}
The demo is available on GitHub\footnote{\url{https://github.com/dimachine/ShapStab/}}.
Shapley values provide a tool for assessing the attribute importance of stable concepts. Comparison with other (not only Game-theoretic) techniques for local interpretability is desirable. We believe that the attribute importance can be lifted at the context level, via an aggregation, and by then offer a possibility for attribute reduction, similar to the principal component analysis (PCA) method.
\section{Attribute Similarity and Reduction}\label{sec:AttrSimRed}
Computation of attribute importance could lead to ranking the attributes of the context, and by then classifying the attributes with respect to their global importance, similar to principal component analysis. Therefore cutting off at a certain threshold could lead to attribute reduction in the context. Other methods leading to attributes reduction are based on their granularity, an ontology or an is-a taxonomy, by using coarser attributes. Less coarse attributes are then put together by going up in the taxonomy and are considered to be similar. In the present section, we briefly discuss the effect of putting attributes together on the resulting concept lattice. Doing this leads to the reduction of the number of attributes, but not always in the reduction of the number of concepts.
Before considering such compound attributes, we would like to draw the readers' attention to types of data weeding that often overlooked outside of the FCA community~\cite{Priss:2011,Konecny:2017,Konecny:2018}, namely, clarification and reduction.
\def\neu#1{{\bf #1}}
\def\mathbb{K}{\mathbb{K}}
\def\mathbb{N}{\mathbb{N}}
\def\mathbb{R}{\mathbb{R}}
\def\mathop{\mbox{\rm J}}{\mathop{\mbox{\rm J}}}
\def\mathop{\mbox{\rm I}}{\mathop{\mbox{\rm I}}}
\def\cal#1{\mathcal{#1}}
\def\frak#1{\mathfrak{#1}}
\def\mathop{\mbox{\rm Ext}}{\mathop{\mbox{\rm Ext}}}
\subsection{Clarification and reduction}
A context $(G, M, I)$ is called \emph{clarified}~\cite{Ganter:1999}, if for any objects $g,h \in G$ from $g'=h'$ it always follows that $g=h$ and, similarly, $m'=n'$ implies $m=n$ for all $m,n \in M$. A clarification consists in removing duplicated lines and columns from the context.
This context manipulation does not alter the structure of the concept lattice, though objects with the same intents and attributes with the same extents are merged, respectively.
The structure of the concept lattice remains unchanged in case of removal of \emph{reducible attributes} and \emph{reducible objects}~\cite{Ganter:1999}; An attribute $m$ is reducible if it is a combination of other attributes, i.e. $m'=Y'$ for some $Y \subseteq M$ with $m \not \in Y$. Similarly, an object $g$ is reducible if $g'=X'$ for some $X \subseteq G$ with $g \not \in X$.
For example, full rows ($g'=M$) and full columns ($m'=G$) are always reducible.
However, if our aim is a subsequent interpretation of patterns, we may wish to keep attributes (e.g. in aggregated form), rather than leaving them out before knowing their importance.
\subsection{Generalised attributes}\label{sec:gen attributes}
As we know, FCA is used for conceptual clustering and helps discover patterns in terms of clusters and rules. However, the number of patterns can explode with the size of an input context. Since the main goal is to maintain a friendly overview of the discovered patterns, several approaches have been investigated to reduce the number of attributes without loss of much information~\cite{Priss:2011,Konecny:2017}. One of these suggestions consists in using is-a taxonomies. Given a taxonomy on attributes, how can we use it to discover \emph{generalised} patterns in the form of clusters and rules? If there is no taxonomy, can we (interactively) design one? We will discuss different scenarios of grouping attributes or objects, and the need of designing similarity measures for these purposes in the FCA setting.
To the best of our knowledge the problem of mining generalised association rules was first introduced around 1995 in~\cite{Srikant:95,Srikant:97}, and rephrased as follows: Given a large database of transactions, where each transaction consists of a set of items, and a taxonomy (is-a hierarchy) on the items, the goal is to find associations between items at any level of the taxonomy. For example, with a taxonomy that says that \texttt{jackets} is-a \texttt{outerwear} and \texttt{outerwear} is-a \texttt{clothes}, we may infer a rule that ``people who buy \texttt{outerwear} tend to buy \texttt{shoes}''. This rule may hold even if rules that ``people who buy \texttt{jackets} tend to buy \texttt{shoes}'', and ``people who buy \texttt{clothes} tend to buy \texttt{shoes}'' do not hold. (See Fig.~\ref{fig:genattr_setting})
\begin{figure}[ht]
\begin{subfigure}[b]{0.5\linewidth}
\centering
\caption{Database $\mathcal D$}
\label{fig7:a}
\begin{tabular}{cl}
\hline
Transaction & Items bought \\
\hline
100 & Shirt \\
200 & Jacket, Hiking Boots \\
300 & Ski Pants, Hiking Boots \\
400 & Shoes \\
500 & Shoes \\
600 & Jacket \\
\hline
\end{tabular}
\vspace{4ex}
\end{subfigure
\begin{subfigure}[b]{0.5\linewidth}
\begin{minipage}[h]{0.5\linewidth}
\resizebox{1.9\linewidth}{!}{%
\Tree [.Clothes [.Outwear [.Jackets ] [.{Ski Pants} ] ]
[.Shirts ]]
\Tree [.Footwear [.Shoes ] [.{Hiking Boots} ]]
}\hfill
\end{minipage}
\caption{Taxonomy $\mathcal{T}$}
\label{fig7:b}
\vspace{4ex}
\end{subfigure}
\begin{subfigure}[b]{0.5\linewidth}
\centering
\caption{Frequent itemsets}
\label{fig7:c}
\begin{tabular}{lc}
\hline
Itemset & Support \\
\hline
Jacket & 2 \\
Outwear & 3 \\
Clothes & 4 \\
Shoes & 2 \\
Hiking Boots & 2 \\
Footwear & 4 \\
Outwear, Hiking Boots & 2\\
Clothes, Hiking Boots & 2\\
Outwear, Footwear & 2\\
Clothes, Footwear & 2\\
\hline
\end{tabular}
\end{subfigure
\begin{subfigure}[b]{0.5\linewidth}
\centering
\caption{Association rules}
\label{fig7:d}
\begin{tabular}{lcc}
\hline
Rule & Support & Confidence \\
\hline
Outwear $\to$ Hiking Boots & 1/3 & 2/3\\
Outwear $\to$ Footwear & 1/3 & 2/3\\
Hiking Boots $\to$ Outwear & 1/3 & 1 \\
Hiking Boots $\to$ Clothes & 1/3 & 1 \\
\hline
\end{tabular}
\end{subfigure}
\caption{A database of transactions, taxonomies and extracted rules~\cite{Srikant:95,Srikant:97}}
\label{fig:genattr_setting}
\end{figure}
A generalised association rule is a (partial) implication $X \to Y$, where $X, Y$ are disjoint itemsets and no item in $Y$ is a generalisation of any item in $X$~\cite{Srikant:95,Srikant:97}. We adopt the following notation:
$\mathcal{I}=\{i_1,i_2, \cdots, i_m \}$ is a set of items and $\mathcal{D} = \{t_1, t_2, \cdots, t_n\}$ a set of transactions. Each transaction $t \in \mathcal{D}$ is a subset of items $\mathcal{I}$.
Let $\mathcal{T}$ be a set of taxonomies (i.e directed acyclic graph on items and generalised items).
We denote by $(\mathcal{T},\le)$ its transitive closure. The elements of $\mathcal{T}$ are called ``general items''. A transaction $t$ \emph{supports} an item $x$ (resp. a general item $y$) if $x$ is in $t$ (resp. $y$ is a generalisation of an item $x$ in $t$).
A set of transactions $T$ \emph{supports} an itemset $X\subseteq \mathcal{I}$ if $T$ supports every item in $X$.
In FCA setting, we build a generalised context $(\mathcal{D},\mathcal{I}\cup \mathcal{T}, I)$, where the set of objects, $\mathcal{D}$,
is the set of transactions (strictly speaking transaction-ID), and the set of attributes, $M=\mathcal{I}\cup \mathcal{T}$, contains all items ($\mathcal{I}$) and general items ($\mathcal{T}$). The incidence relation $I \subseteq \mathcal{D} \times M$ is defined by
\[
tIm \iff
\begin{cases}
m\in\mathcal{I} \text{ and } m\in t \\
m\in\mathcal{T} \text{ and } \exists n\in\mathcal{I}, n\in t \text{ and } n\le m.
\end{cases}
\]
Below is the context associated to the example on Figure~\ref{fig:genattr_setting}.
\begin{center}
\setlength{\tabcolsep}{3pt}
\begin{tabular}{|c|c|c|c|c|c||c|c|c|}
\hline
& Shirt & Jacket & Hiking Boots & Ski Pants & Shoes & Outerwear & Clothes & Footwear \\ \hline
100 & $\times$ &&&&&&$\times$& \\ \hline
200 & & $\times$ & $\times$ &&&$\times$&$\times$&$\times$ \\ \hline
300 & & & $\times$ & $\times$ &&$\times$&$\times$&$\times$ \\ \hline
400 &&&& & $\times$ &&&$\times$\\ \hline
500 &&&& & $\times$ &&&$\times$ \\ \hline
600 && $\times$ &&&& $\times$ & $\times$ & \\ \hline
\end{tabular}
\end{center}
The basic interestingness measures for a generalised rule $X \to Y$ are support and confidence (see association rules in~Fig.~\ref{fig:genattr_setting} (d)). Its \emph{support} $supp(X \to Y)$ is defined as $\frac{|(X \cup Y)'|}{|\mathcal D|}$, while its \emph{confidence} $conf(X \to Y)$ is $\frac{|(X \cup Y)'|}{|X'|}$.
For some applications, it would make sense to work only with the subcontext $(\mathcal{D},\mathcal{T}, I \cap \mathcal{D} \times \mathcal{T})$ instead of $(\mathcal{D},\mathcal{I}\cup \mathcal{T}, I)$, for example if the goal is to reduce the number of attributes, concepts or rules. Sometimes, there is no taxon available to suggest that considered attributes should be put together. However, we can extend the used taxonomy, i.e. put some attributes together in a proper taxon, and decide when an object satisfies the grouped attributes.
\subsection{Generalising scenarios}
Let $\mathbb{K}:=(G,M,I)$ be a context. The attributes of $\mathbb{K}$ can be grouped to form another set of attributes, namely $S$, whose elements are called \textbf{generalised} attributes. For example, in basket market analysis, items (products) can be generalised into product lines and then product categories, and even customers may be generalised to groups according to specific features (e.g., income, education). This replaces $(G,M,I)$ with a context $(G,S,J)$ where $S$ can be seen as an index set such that $\{m_s\mid s\in S\}$ covers $M$.
How to define the incidence relation $J$, is domain dependent. Let us consider several cases below~\cite{Kwuida:2009,Kwuida:2012,Kwuida:2014}:
\begin{itemize}
\item[$(\exists)$] $gJs:\iff \exists m\in s,\, g\mathop{\mbox{\rm I}} m$. \ When companies are described by the locations of their branches then cities can be grouped to regions or states.
A company $g$ operates in a state $s$ if $g$ has a branch in a city $m$ which is in $s$.
\item[$(\forall)$] $gJs:\iff \forall m\in s,\, g\mathop{\mbox{\rm I}} m$. \ For exams with several components (e.g. written, oral, and thesis), we might require students to pass all components in order to succeed.
\item[$(\alpha\%)$] $gJs:\iff \frac{|\{m\in s\ \mid\ g I m\}|}{|s|}\geq\alpha_s$ with $\alpha_s$ a threshold. \ In the case of exams discussed above, we could require students to pass just some parts, defined by a threshold.
\end{itemize}
Similarly, objects can also be put together to get ``generalised objects''. In \cite{Prediger:2003} the author described \emph{general on objects} as
classes of individual objects that are considered to be extents of concepts
of a formal context. In that paper, different contexts with general objects are defined and their conceptual structure and relation to other contexts is analysed with FCA methods. Generalisation on both objects and attributes can be carried out with the combinations below, with $A \subseteq G$ and $B \subseteq M$:
\begin{enumerate}
\item $AJB$ iff $\exists a\in A$, $\exists b\in B$ such that $a\mathop{\mbox{\rm I}} b$ (i.e. some objects from $A$ are in relation with some attributes in $B$);
\item $AJB$ iff $\forall a\in A$, $\forall b\in B\ a\mathop{\mbox{\rm I}} b$ (i.e. each object in $A$ has all attributes in $B$);
\item $AJB$ iff $\forall a\in A$, $\exists b\in B$ such that $a\mathop{\mbox{\rm I}} b$ (i.e. each object in $A$ has at least one attribute in $B$);
\item $AJB$ iff $\exists b\in B$ such that $\forall a\in A$ $a\mathop{\mbox{\rm I}} b$ (i.e. an attribute in $B$ is satisfied by all objects of $A$);
\item $AJB$ iff $\forall b\in B$, $\exists a\in A$ such that $a\mathop{\mbox{\rm I}} b$ ( i.e. each property in $B$ is satisfied by an object of $A$);
\item $AJB$ iff $\exists a\in A$ such that $\forall b\in B\ a\mathop{\mbox{\rm I}} b$ (i.e. an object in $A$ has all attributes in $B$);
\item $AJB$ iff $\frac{\left|\{a\in A\mid \frac{|\{b\in B\mid a\mathop{\mbox{\rm I}} b\}|}{|B|}\geq \beta_{B}\} \right|}{|A|}\geq\alpha_A$ (i.e. at least $\alpha_{A}\%$ of objects in $A$ have each at least $\beta_{B}\%$ of the attributes in $B$);
\item $AJB$ iff $\frac{\left|\left\{b\in B\mid
\frac{|\{a\in A\mid a\mathop{\mbox{\rm I}} b\}|}{|A|}\geq
\alpha_{A}\right\} \right|}{|B|}\geq\beta_B$ (i.e. at least $\beta_{B}\%$ of attributes in $B$ belong altogether to at least $\alpha_{A}\%$ of objects in the group $A$);
\item $AJB$ iff $\frac{|A\times B\cap I|}{|A\times B|}\geq \alpha$ (i.e. the density of the rectangle $A\times B$ is at least $\alpha$).
\end{enumerate}
\subsection{Generalisation and extracted patterns}
After analysing several generalisation cases, including simultaneous generalisations on both objects and attributes as above, the next step is to look at the extracted patterns. From contexts, knowledge is usually extracted in terms of clusters and rules. When dealing with generalised attributes or objects, we coin the term ``generalised'' to all patterns extracted. An immediate task is to compare knowledge gained after generalising with those from the initial context.
New and interesting rules as seen in Figure~\ref{fig:genattr_setting} can be discovered~\cite{Srikant:95,Srikant:97}.
Experiments have shown that the number of extracted patterns quite often decreases. Formal investigations are been carried out to compare these numbers. For $\forall$-generalisations, the number of concepts does not increase~\cite{Kwuida:2014}. But for $\exists$-generalisations, the size can actually increase~\cite{Kwuida:2009,Kwuida:2012,Kwuida:2014,Kwuida:2017,Kwuida:2020}.
In~\cite{Belohlavek:2014} the authors propose a method to control the structure of concept lattices derived from Boolean data by specifying granularity levels of attributes. Here a taxonomy is already available, given by the granularity of the attributes. They suggest that granularity levels should be chosen by a user based on his expertise and experimentation with the data. If the resulting formal concepts are too specific and there is a large number of them, the user can choose to use a coarser level of granularity. The resulting formal concepts are then less specific and can be seen as resulting from a zoom-out. Similarly, one may perform a zoom-in to obtain finer, more specific formal concepts. Through all these precautions, the number of concepts can still increase when attributes are coarser: ``The issue of when attribute coarsening results in an increase in the number of formal concepts needs a further examination, as well as the possibility of informing automatically a user who is selecting a new level of granularity that the new level results in an increase in the number of concepts.''~\cite{Belohlavek:2014}
In~\cite{Kwuida:2020} a more succinct analysis of $\exists$-generalisations presents a family of contexts where generalising two attributes results in an exponential increase in the number of concepts. An example of such context is given in the Table~\ref{tab:S6_7} (left).
\begin{table}[ht]
\caption{A formal context (left) and its $\exists$-generalisation that puts $m_1$ and $m_2$ together. The number of concepts increases from 48 to 64, i.e. by 16.}
\label{tab:S6_7}
\centering
\begin{tabular}{|l|l|l|l|l|l|l|l|l|} \hline
& $1$ & $2$ & $3$ & $4$ & $5$ & $6$ & $m_{1}$ & $m_{2}$ \\ \hline
$1$ & & $\times $ & $\times $ & $\times $ &$\times$ & $\times$ & $\times$ & \\ \hline
$2$ & $\times $ & & $\times $ & $\times $ & $\times $ &$\times$ & $\times$ & $\times$ \\ \hline
$3$ & $\times $ & $\times$ & & $\times $ & $\times $ &$\times$ & $\times$ & $\times$ \\ \hline
$4$ & $\times $ & $\times$ & $\times$ & & $\times $ &$\times$ & $\times$ & $\times$ \\ \hline
$5$ & $\times $ & $\times$ & $\times$ & $\times $ & & $\times$ & $\times$ & $\times$ \\ \hline
$6$ & $\times $ & $\times$ & $\times$ & $\times $ &$\times$ & & & $\times$ \\ \hline
$g_{1}$ & $\times $ & $\times $ & $\times$ & $\times $ & $\times$ & $\times$ & & \\ \hline
\end{tabular}
\quad $\Longrightarrow$\quad
\begin{tabular}{|l|l|l|l|l|l|l|l|} \hline
& $1$ & $2$ & $3$ & $4$ & $5$ & $6$ & $m_{12}$ \\ \hline
$1$ & & $\times $ & $\times $ & $\times $ &$\times$ & $\times$ & $\times$ \\ \hline
$2$ & $\times $ & & $\times $ & $\times $ & $\times $ &$\times$ & $\times$ \\ \hline
$3$ & $\times $ & $\times$ & & $\times $ & $\times $ &$\times$ & $\times$ \\ \hline
$4$ & $\times $ & $\times$ & $\times$ & & $\times $ &$\times$ & $\times$ \\ \hline
$5$ & $\times $ & $\times$ & $\times$ & $\times $ & & $\times$ & $\times$ \\ \hline
$6$ & $\times $ & $\times$ & $\times$ & $\times $ &$\times$ & & $\times$ \\ \hline
$g_{1}$ & $\times $ & $\times $ & $\times$ & $\times $ & $\times$ & $\times$ & \\ \hline
\end{tabular}
\end{table}
Putting together some attributes does not always reduce the number of extracted patterns. It's therefore interesting to get measures that suggest which attributes can be put together, in the absence of a taxonomy. The goal would be to not increase the size of extracted patterns.
\subsection{Similarity and existential generalisations}
This section presents investigations on the use of certain similarity measures in generalising attributes.
A \emph{similarity measure} on a set $M$ of attributes is a function $S:M\times M\to \mathbb{R}$ such that for all $m_1,m_2$ in $M$,
\begin{itemize}
\item[(i)] $S(m_1,m_2)\geq 0$, \hfill \textit{positivity}
\item[(ii)] $S(m_1,m_2)=S(m_2,m_1)$ \hfill \textit{symmetry}
\item[(iii)] $S(m_1,m_1) \geq S(m_1,m_2)$ \hfill \textit{maximality}
\end{itemize}
We say that $S$ is \emph{compatible} with generalising attributes if whenever $m_1, m_2$ are more similar than $m_3,m_4$, then putting $m_1,m_2$ together should not lead to more concepts than putting $m_3,m_4$ together does.
To give the formula for some known similarity measures that could be of interest in FCA setting, we adopt the following notation for $m_1,m_2$ attributes in $\mathbb{K}$:
\[a=|m_1'\cap m_2'|,\quad d=|m_1'\Delta m_2'|, \quad b=|m_1'\setminus m_2'|,\quad c=|m_2'\setminus m_1'|.\]
\begin{table}[htbp]
\caption{Some similarity measures relevant in FCA}
\label{tab:similarities}
\centering
\begin{small}
\renewcommand{\arraystretch}{2.5} \tabcolsep=4pt
\begin{tabular}{|lr||lr|} \hline
Name & Formula & Name & Formula \\ \hline
Jaccard (Jc)& $\dfrac{a}{a+b+c}$ & Sneath/Sokal (SS$_1$) & $\dfrac{2(a+d)}{2(a+d)+b+c}$ \\
Dice (Di) & $\dfrac{2a}{2a+b+c}$ & Sneath/Sokal (SS$_2$) & $\dfrac{0.5a}{0.5a+b+c}$ \\
Sorensen (So) & $\dfrac{4a}{4a+b+c}$ & Sokal/Michener (SM) & $\dfrac{a+d}{a+d+b+c}$\\
Anderberg (An) & $\dfrac{8a}{8a+b+c}$ & Rogers/Tanimoto (RT) & $\dfrac{0.5(a+d)}{0.5(a+d)+b+c}$ \\
Orchiai (Or) & $\dfrac{a}{\sqrt{(a+b)(a+c}}$ & Russels/Rao (RR) & $\dfrac{a}{a+d+b+c}$ \\
Kulczynski (Ku) & $\dfrac{0.5a}{a+b}+\dfrac{0.5a}{a+c}$ & Yule/Kendall (YK) & $\dfrac{ad}{ad+bc}$ \\ \hline
\end{tabular}
\end{small}
\end{table}
\noindent
For the context left in Table~\ref{tab:S6_7}, we have computed $S(m_1,x),\ x=1,\ldots , 6, m_2$. Although $m_1$ is more similar to $m_2$ than any attribute $i<6$, putting $m_1$ and $m_2$ together increases the number of concepts. Note that putting $m_1$ and $6$ together is equivalent to removing $m_1$ from the context, and thus, reduces the number of concepts.
\begin{table}[ht]
\caption{The values of considered similarity measures $S(m,i)$}\label{tab:sim_val}
\def\arraystretch{1.2}\tabcolsep=4pt
\centering
\begin{tabular}{|c r r r r r r r r r r r|}
\hline
& Jc & Di & So & An & SS$_2$ & Ku & Or & SM & RT & SS$_1$ & RR \\
\hline
$i\in S_5$ & 0.57 & 0.80 & 0.89 & 0.94 & 0.50 & 0.80 & 0.80 & 0.71 & 0.56 & 0.83 & 0.57 \\
$i=6$ & 0.83 & 0.91 & 0.95 & 0.97 & 0.71 & 0.92 & 0.91 & 0.75 & 0.75 & 0.92 & 0.71 \\
$i=m_2$ & 0.67 & 0.80 & 0.89 & 0.94 & 0.50 & 0.80 & 0.80 & 0.71 & 0.56 & 0.83 & 0.57 \\ \hline
\end{tabular}
\end{table}
Let $\mathbb{K}$ be a context $(G,M,I)$ with $a,b\in M$ and
$\mathbb{K}_{00}$ be its subcontext without $a,b$. Below, $\mathop{\mbox{\rm Ext}}(\mathbb{K}_{00}$) means all the extents of concepts of the context $\mathbb{K}_{00}$. In order to describe the increase in the number of concepts after putting $a,b$ together, we set
\begin{align*}
\mathcal{H}(a)&:=
\left\{A\cap a^\prime \mid A\in\mathop{\mbox{\rm Ext}}(\mathbb{K}_{00}) \text{ and } A\cap a^\prime \notin\mathop{\mbox{\rm Ext}}(\mathbb{K}_{00})\right\} \\
\mathcal{H}(b)&:=
\left\{A\cap b^\prime \mid A\in\mathop{\mbox{\rm Ext}}(\mathbb{K}_{00}) \text{ and } A\cap b^\prime \notin\mathop{\mbox{\rm Ext}}(\mathbb{K}_{00})\right\} \\
\mathcal{H}(a\cup b)&:=
\left\{A\cap (a^\prime\cup b^\prime)\mid A\in\mathop{\mbox{\rm Ext}}(\mathbb{K}_{00}),\ A\cap(a^\prime\cup b^\prime)\notin\mathop{\mbox{\rm Ext}}(\mathbb{K}_{00})\right\} \\
\mathcal{H}(a\cap b)&:=
\left\{A\cap (a^\prime\cap b^\prime) \mid A\in\mathop{\mbox{\rm Ext}}(\mathbb{K}_{00}),\ A\cap(a^\prime\cap b^\prime)\notin\mathop{\mbox{\rm Ext}}(\mathbb{K}_{00})\right\}.
\end{align*}
\noindent
The following proposition shows that the increase can be exponential.
\begin{theorem}[\cite{Kwuida:2020}]
Let $(G,M,\mathop{\mbox{\rm I}})$ be an attribute reduced context and $a, b$ be two attributes such that their generalisation $s=a\cup b$ increases the size of the concept lattice. Then
$|\frak{B}(G,M,I)| = |\frak{B}(G,M\setminus\{a,b\},I \subseteq G \times M\setminus\{a,b\} )|+|\cal{H}(a,b)|$, with
\begin{align*}
|\cal{H}(a,b)|&=|\cal{H}(a)\cup \cal{H}(b)\cup \cal{H}(a\cap b)| \le 2^{|a'|+|b'|}-2^{|a'|}-2^{|b'|}+1.
\end{align*}
This upper bound can be reached.
\end{theorem}
The difference $|\cal{H}(a,b)|$ is then used to define a compatible similarity measure. We set
$\psi(a,b) := |\cal{H}(a\cup b)|-|\cal{H}(a,b)|$,\quad
$\delta(a,b):=
\begin{cases}
1 & \text{ if \ }\psi (a,b)\leq 0 \\
0 & \text{ else}%
\end{cases}%
$, and define
$S(a,b):=\dfrac{1+\delta(a,b)}{2}-\dfrac{|\psi(a,b)|}{2n_0}$ with $n_0=\max\{\psi(x,y)\mid x,y\in M\}$. Then
\begin{theorem}[\cite{Kuitche:2018}]
$S$ is a similarity measure compatible with the generalisation.
\[S(a,b)\ge \frac{1}{2} \iff \psi(a,b)\le 0\]
\end{theorem}
\section{Conclusion}\label{sec:Concl}
The first two parts contain a concise summary of the usage of Shapley values from Cooperative Game Theory for interpretable concept-based learning in the FCA playground with its connection to Data Mining formulations. We omitted results related to algorithms and their computational complexity since they deserve a separate detailed treatment.
The lessons drawn from the ranking attributes in JSM classification hypotheses and those in the intents of stable concepts show that valuation functions should be customised and are not necessarily zero-one-valued.
This is an argument towards that of Shapley values approach requires specification depending on the model (or type of patterns) and thus only conditionally is model-agnostic. The other lesson is about the usage of Shapley values for pattern attribution concerning their contribution interestingness measures like stability or robustness.
The third part is devoted to attribute aggregation by similarity, which may help to apply interpretable techniques to larger sets of attributes or bring additional aspects to interpretability with the help of domain taxonomies.
The desirable property of similarity measures to provide compatible generalisation helps to reduce the number of output concepts or JSM-hypotheses as well.
The connection between attribute similarity measures and Shapley interaction values~\cite{Lundberg:2017}, when the interaction of two or more attributes on the model prediction is studied, is also of interest.
In addition to algorithmic issues, we would like to mention two more directions of future studies. The first one lies in the interpretability by means of Boolean matrix factorisation (decomposition), which was used for dimensionality reduction with explainable Boolean factors (formal concepts)~\cite{Belohlavek:2010} or interpretable ``taste communities'' identification in collaborative filtering~\cite{Ignatov:2014b}. In this case, we are transitioned from the importance of attributes to attribution of factors. The second one is a closely related aspect to interpretability called fairness~\cite{Alves:2020}, where, for example, certain attributes of individuals should not influence much to the model prediction (disability, ethnicity, gender, etc.).
\subsubsection*{Acknowledgements.} The study was implemented in the framework of the Basic Research Program at the National Research University Higher School of Economics and funded by the Russian Academic Excellence Project '5-100'. The second author was also supported by Russian Science Foundation under grant 17-11-01276 at St. Petersburg Department of Steklov Mathematical Institute of Russian Academy of Sciences, Russia. The second author would like to thank Fuad Aleskerov, Alexei Zakharov, and Shlomo Weber for the inspirational lectures on Collective Choice and Voting Theory.
\bibliographystyle{splncs04}
|
\section{Introduction}
Let $\Omega$ be a bounded domain in $\mathbb C$ with $0\in \partial \Omega$.
The famous Wiener's criterion states that $0$ is a regular point if and only if
$$
\sum_{k=1}^\infty \frac{k}{\log [1/\mathcal C_l( A_k\cap \Omega^c )]}=\infty,
$$
where $
A_k:=\{z\in \mathbb C:\frac1{2^{k+1}}\le |z|\le \frac1{2^k}\}.
$ and $\mathcal C_l(\cdot)$ stands for the standard logarithmic capacity (relative to $\infty$).
It is also known that the regularity of $0$ implies that the Bergman kernel is exhaustive at $0$ (cf. \cite{Ohsawa}) and the Bergman metric is complete at $0$ (cf. \cite{BP}, \cite{Herbort}). Zwonek \cite{Zwonek} showed that the Bergman kernel is exhaustive at $0$ if and only if
$$
\sum_{k=1}^\infty \frac{2^{2k}}{\log [1/\mathcal C_l (A_k(z)\cap \Omega^c)]}\rightarrow \infty\ \ \ (z\rightarrow 0),
$$
where $A_k(z):=\{w\in \mathbb C:\frac1{2^{k+1}}\le |w-z|\le \frac1{2^k}\}$.
On the other hand, a similar characterization for the Bergman completeness at $0$ is still missing, although some partial results exist (cf. \cite{PZ}, \cite{Jucha}). Only recently, interesting lower and upper bounds of the Bergman kernel $K_\Omega(z)$ for planar domains were obtained in terms of the logarithmic capacity of $\Omega^c$ relative to $z$ (cf. \cite{BlockiSuita}, \cite{BZ}).
In the spirit of Wiener's criterion, we give the following
\begin{theorem}\label{th:CapBerg}
Let $\Omega$ be a bounded pseudoconvex domain in $\mathbb C^n$. Suppose there are constants $\alpha>1$ and $\varepsilon>0$ such that for any $z\in \Omega$, there exists a complex line $\mathcal L_z\ni z$ such that
\begin{equation}\label{eq:CB_1}
\mathcal C_l\left(\mathcal L_z\cap B(z,\alpha \delta_\Omega(z))\cap \Omega^c\right)\ge \varepsilon\, \delta_\Omega(z),
\end{equation}
where $\delta_\Omega(z):=d(z,\partial \Omega)$ and $B(z,r)=\{w\in \mathbb C^n:|w-z|<r\}$. Then
\begin{equation}\label{eq:CB_2}
K_\Omega(z) \gtrsim \delta_\Omega(z)^{-2},\ \ \ z\in \Omega.
\end{equation}
\end{theorem}
It is well-known that \eqref{eq:CB_2} holds for any bounded pseudoconvex domain with Lipschitz boundary (cf. \cite{OT}). As a consequence of Theorem \ref{th:CapBerg}, we obtain the following
\begin{theorem}\label{th:KernelEstimate}
Let $\Omega$ be a bounded pseudoconvex domain in $\mathbb C^n$. Then \eqref{eq:CB_2} holds under one of the following conditions:
\begin{enumerate}
\item[$(a)$] $\partial\Omega$ is $C^0$, i.e., it can be written locally as the graph of a continuous function.
\item[$(b)$] There are constants $\varepsilon,r_0>0$ such that
\begin{equation}\label{eq:CB_3}
|B(\zeta,r)\cap \Omega^c|\ge \varepsilon |B(\zeta,r)|,\ \ \ \zeta\in \partial \Omega,\,0<r\le r_0,
\end{equation}
where $|\cdot|$ stands for the volume.
\item[$(c)$] $\Omega$ is fat $($i.e., $\overline{\Omega}^\circ=\Omega)$ and\/ $\overline{\Omega}$ is $\mathcal O(U)-$convex, where $U$ is a neighborhood of\/ $\overline{\Omega}$.
\item[$(d)$] $\Omega$ is strongly hyperconvex.
\end{enumerate}
\end{theorem}
Here are a few remarks. Recall that the $\mathcal O(U)-$convex hull of a compact set $E\subset U$ is given by
$$
\widehat{E}_{\mathcal O(U)}:=\left\{z\in U: |f(z)|\le \sup_E |f| \ \text{for all}\ f\in \mathcal O(U)\right\}.
$$
We say that $E$ is $\mathcal O(U)-$convex if $E= \widehat{E}_{\mathcal O(U)}$; in particular, if $U=\mathbb C^n$ then $E$ is called polynomially convex.
Recall also that a bounded domain $\Omega\subset \mathbb C^n$ is called strongly hyperconvex if there exists a continuous negative psh function $\rho$ defined in a neighborhood $U$ of $\overline{\Omega}$ such that $\Omega=\{z\in U:\rho(z)<0\}$. Based on the work of Zwonek \cite{Zwonek}, Pflug-Zwonek \cite{PZ} proved that for any $\varepsilon>0$ there exists a constant $C_\varepsilon>0$ such that $K_\Omega(z)\ge C_\varepsilon\, \delta_\Omega(z)^{-2+\varepsilon}$.
Note that conditions of type \eqref{eq:CB_3} are popular in second order elliptic boundary problems $($cf. \cite{Kenig}, Chapter 1,\,\S\,1$)$, and spectrum theory of the Laplacian $($cf. \cite{Lieb},\,\cite{Davis}$)$.
We also have the following analogous result which does not follow from Theorem \ref{th:CapBerg}.
\begin{proposition}\label{prop:Runge}
If\/ $\Omega\subset \mathbb C^n$ is a bounded pseudoconvex Runge domain, then \eqref{eq:CB_2} holds.
\end{proposition}
\begin{remark}
It is known that every star domain in $\mathbb C^n$ is a Runge domain $($cf. \cite{Almer} or \cite{Kasimi}$)$.
\end{remark}
The Bergman kernel estimate \eqref{eq:CB_2} has an unexpected application to holomorphic motions. Following \cite{MSS}, we define a holomorphic motion of a set $E\subset \mathbb C$
as a mapping
$
F:\mathbb D \times E \rightarrow {\mathbb C},
$
which enjoys the following properties:
\begin{enumerate}
\item $F(0,z)=z$ for all $z\in
E$.
\item For every fixed $\lambda\in \mathbb D$, the mapping
$F(\lambda,\cdot):E\rightarrow {\mathbb C}$ is an injection.
\item For
every fixed $z\in E$, the mapping $F(\cdot,z):\mathbb D\rightarrow {\mathbb
C}$ is holomorphic.
\end{enumerate}
Let $F:\mathbb D\times {\mathbb C}\rightarrow
{\mathbb C}$ be a holomorphic motion of\/ $\mathbb C$\/ fixing $\infty$. Define quasi-bidiscs as follows
$$
\Gamma_r(F):=\left\{(\lambda,F(\lambda,z)): \lambda\in \mathbb D_r,z\in \mathbb D\right\},\ \ \ 0<r<1,
$$
where $\mathbb D_r=\{z\in \mathbb C:|z|<r\}$.
We obtain a geometric property of quasi-bidiscs as follows.
\begin{theorem}\label{th:HM}
Let $\delta_\lambda(w)$ denote the boundary distance of the slice $F(\lambda,\mathbb D)$, $\lambda\in \mathbb D$. Then for any $0<r'<r<1$, there exists a constant $C>0$ such that
\begin{equation}\label{eq:HM_1}
\delta_\lambda(w)
\le C \delta_{ \Gamma_r(F)}(\lambda,w)
\end{equation}
whenever $|\lambda|\le r'$.
\end{theorem}
\begin{remark}
The advantage of Theorem \ref{th:HM} is that the analysis in terms of $\delta_\lambda$ is easier on $\Gamma_r(F)$ $($compare \cite{ChenH-index}, p. 1450--1451$)$.
\end{remark}
The argument in Zwonek \cite{Zwonek} yields the following result closely related to Theorem \ref{th:CapBerg}.
\begin{theorem}\label{th:CapBerg_2}
Let $\Omega$ be a bounded pseudoconvex domain in $\mathbb C^n$. Suppose there are constants $\alpha>1$, $\beta\ge 1$ and $\varepsilon>0$ such that for any $z\in \Omega$, there exists a complex line $\mathcal L_z\ni z$ such that
\begin{equation}\label{eq:CB3}
\mathcal C_l\left(\mathcal L_z\cap B(z,\alpha \delta_\Omega(z))\cap \Omega^c\right)\ge \varepsilon\, \delta_\Omega(z)^\beta.
\end{equation}
Then
\begin{equation}\label{eq:CB4}
K_\Omega(z) \gtrsim \delta_\Omega(z)^{-2}|\log \delta_\Omega(z)|^{-1}
\end{equation}
for all $z$ sufficiently close to $\partial \Omega$.
\end{theorem}
\begin{corollary}\label{cor:VolBerg_2}
Let $\Omega$ be a bounded pseudoconvex domain in $\mathbb C^n$. Suppose there are constants $\varepsilon,r_0>0$ and $\beta> 1$, such that
\begin{equation}\label{eq:CB5}
|B(\zeta,r)\cap \Omega^c|\ge \varepsilon |B(\zeta,r)|^\beta,\ \ \ \zeta\in \partial \Omega,\,0<r\le r_0,
\end{equation}
then \eqref{eq:CB4} holds.
\end{corollary}
For the Bergman kernel, the cerebrated Ohsawa-Takegoshi extension theorem \cite{OT} serves as a bridge passing from one-dimensional estimates to high dimensions. Unfortunately, such a powerful tool is not available for estimating other objects like the pluricomplex Green function or the Bergman metric/distance. Thus we have to focus on bounded domains $\Omega\subset \mathbb C$ in the sequel.
Motivated by the work of Carleson-Totik \cite{CarlesonTotik}, we give the following
\begin{definition}
Let $\varepsilon>0$ and $0<\lambda< 1$ be fixed. For every $a\in \partial \Omega$ we set
\begin{eqnarray*}
K_t(a) & := & \overline{\mathbb D_{t}(a)} - \Omega;\ \ \ \mathbb D_{t}(a):=\{z:|z-a|<t\} \\
\mathcal N_a(\varepsilon,\lambda) & := & \left\{n\in \mathbb Z^+: \mathcal C_l\left(K_{\lambda^n}(a)\right)\ge \varepsilon \lambda^{ n}\right\}\\
\mathcal N_a^n(\varepsilon,\lambda) &:=&\mathcal N_a(\varepsilon,\lambda)\cap \{1,2,\cdots,n\}.
\end{eqnarray*}
We define the $(\varepsilon,\lambda)-$capacity density of $\partial \Omega$ at $a$ by
$$
\mathcal D_a(\varepsilon,\lambda):=\liminf_{n\rightarrow\infty} \frac{|\mathcal N_a^n(\varepsilon,\lambda)|}n.
$$
We define the weak and strong $(\varepsilon,\lambda)-$capacity density of $\partial \Omega$ by
$$
\mathcal D_W(\varepsilon,\lambda) := \liminf_{n\rightarrow\infty} \frac{\inf_{a\in \partial \Omega}|\mathcal N_a^n(\varepsilon,\lambda)|}n
$$
and
\begin{eqnarray*}
\mathcal D_S (\varepsilon,\lambda) := \liminf_{n\rightarrow\infty} \frac{|\bigcap_{a\in \partial \Omega}\mathcal N_a^n(\varepsilon,\lambda)|}n
\end{eqnarray*}
respectively.
\end{definition}
It is easy to see that $ \mathcal D_W (\varepsilon,\lambda)\ge \mathcal D_S (\varepsilon,\lambda)$, and
$\mathcal C_l(K_{t}(a))\ge \varepsilon t$, $\forall\,t>0$, implies $ \mathcal D_a(\varepsilon,\lambda)=1$. Recall that $\partial \Omega$ is said to be\/ {\it uniformly perfect} if $\inf_{a\in \partial \Omega} \mathcal C_l(K_{t}(a))\ge \varepsilon t$, $\forall\,t>0$ (cf. \cite{Pommerenke}). Thus if $\partial \Omega$ is uniformly perfect then $\mathcal D_W (\varepsilon,\lambda)=\mathcal D_S (\varepsilon,\lambda)=1$ for some $\varepsilon>0$.
On the other hand, it was pointed out in \cite{CarlesonTotik} that the domain
$$
\Omega= \mathbb D-\{0\}-\bigcup _{k=1}^\infty \left[2^{-2^{2k+1}},2^{-2^{2k}}\right]
$$
satisfies $D_W (\varepsilon,1/2)>0$ for some $\varepsilon>0$ while
$\partial \Omega$ is non-uniformly perfect. Actually, one may verify that $D_S (\varepsilon,1/2)>0$.
Let $z_0$ be a fixed point in $\Omega$. Carleson-Totik proved the following quantitative analogue of the sufficient part of Wiener's criterion.
\begin{theorem}[cf. \cite{CarlesonTotik}]\label{th:CT}
If\/ $ \mathcal D_W(\varepsilon,\lambda)>0$\/ for some $\varepsilon,\lambda$, then there exists $\beta>0$ such that the Green function $g_\Omega$ satisfies
\begin{equation}\label{eq:CT}
-g_\Omega(z,z_0)\lesssim \delta_\Omega(z)^\beta
\end{equation}
for all $z$ sufficiently close to $\partial \Omega$.
\end{theorem}
\begin{remark}
It is remarkable that the converse of Theorem \ref{th:CT} holds under the additional condition that $\Omega$ contains a fixed size cone with vertex at any $a\in \partial\Omega$
$($cf. \cite{CarlesonTotik}$)$.
\end{remark}
Since the original proof in \cite{CarlesonTotik} is rather technical, it seems worthwhile to give a more transparent approach. Actually, Theorem \ref{th:CT} turns out to be a simple consequence of a quantitative estimate of the capacity potential given in Theorem \ref{th:basic}, which can be applied in more general context.
\begin{definition}
For $\varepsilon>0$, $0<\lambda< 1$ and $\gamma>1$ we set
\begin{eqnarray*}
\mathcal N_a(\varepsilon,\lambda,\gamma) &:=& \left\{n\in \mathbb Z^+: \mathcal C_l(K_{\lambda^{n}}(a))\ge \varepsilon \lambda^{\gamma n}\right\}\\
\mathcal N_a^n (\varepsilon,\lambda,\gamma) &:=& \mathcal N_a(\varepsilon,\lambda,\gamma)\cap \{1,2,\cdots,n\}. \end{eqnarray*}
We define the $(\varepsilon,\lambda,\gamma)-$capacity density of $\partial \Omega$ at $a$ by
$$
\mathcal D_a(\varepsilon,\lambda,\gamma):=\liminf_{n\rightarrow\infty} \frac{\sum_{k\in \mathcal N_a^n(\varepsilon,\lambda,\gamma)} k^{-1}}{\log n}
$$
and the weak and strong $(\varepsilon,\lambda,\gamma)-$capacity densities of $\partial \Omega$ by
$$
\mathcal D_W(\varepsilon,\lambda,\gamma):=\liminf_{n\rightarrow\infty} \frac{\inf_{a\in \partial \Omega} \sum_{k\in \mathcal N_a^n(\varepsilon,\lambda,\gamma)} k^{-1} }{\log n}
$$
and
$$
\mathcal D_S(\varepsilon,\lambda,\gamma):=\liminf_{n\rightarrow\infty} \frac{ \sum_{k\in \bigcap _{a\in \partial \Omega}\mathcal N_a^n(\varepsilon,\lambda,\gamma)} k^{-1} }{\log n}
$$
respectively.
\end{definition}
Note that $ \mathcal D_W(\varepsilon,\lambda,\gamma)\ge \mathcal D_S(\varepsilon,\lambda,\gamma)$. If $\mathcal C_l(K_t(a))\ge \varepsilon t^\gamma$, $\forall\,t>0$, for some $\varepsilon>0$ and $\gamma>1$, then
$
\mathcal D_a(\varepsilon,\lambda,\gamma)=1
$
in view of the following well-known formula
$$
\lim_{n\rightarrow \infty} \left( 1+\frac12+\cdots+\frac1n - \log n \right) = \text{Euler constant}.
$$
\begin{theorem}\label{th:logGreen}
\begin{enumerate}
\item[$(1)$] If\/ $ \mathcal D_W(\varepsilon,\lambda,\gamma)>0$\/ for some $\varepsilon,\lambda,\gamma$, then there exists $\beta>0$ such that
\begin{equation}\label{eq:LG_1}
-g_\Omega (z,z_0) \lesssim (-\log\delta_\Omega(z))^{-\beta}
\end{equation}
for all $z$ sufficiently close to $\partial \Omega$.
\item[$(2)$]
Suppose\/ $\inf_{a\in \partial \Omega} \mathcal C_l(K_t(a))\ge \varepsilon t^\gamma$ for some $\varepsilon>0$ and $\gamma>1$. For every $\tau<\frac1{\gamma-1}$ there exists $C>0$ such that
\begin{equation}\label{eq:LG_2}
-g_\Omega (z,z_0) \le C (-\log \delta_\Omega(z))^{-\tau}
\end{equation}
for all $z$ sufficiently close to $\partial \Omega$.
\end{enumerate}
\end{theorem}
We also obtain the following lower bounds for the Bergman distance $d_B$.
\begin{theorem}\label{th:LowDist}
\begin{enumerate}
\item[$(1)$] If\/ $ \mathcal D_S(\varepsilon,\lambda)>0$\/ for some $\varepsilon,\lambda$, then
\begin{equation}\label{eq:1.1}
d_B(z_0,z) \gtrsim |\log \delta_\Omega(z)|
\end{equation}
for all $z$ sufficiently close to $\partial \Omega$.
\item[$(2)$] If\/ $ \mathcal D_W(\varepsilon,\lambda)>0$\/ for some $\varepsilon,\lambda$, then
\begin{equation}\label{eq:1.2}
d_B(z_0,z) \gtrsim \frac{|\log \delta_\Omega(z)|}{\log|\log \delta_\Omega(z)|}
\end{equation}
for all $z$ sufficiently close to $\partial \Omega$.
\item[$(3)$] If\/ $ \mathcal D_W(\varepsilon,\lambda,\gamma)>0$\/ for some $\varepsilon,\lambda,\gamma$, then
\begin{equation}\label{eq:1.3}
d_B(z_0,z) \gtrsim \log \log |\log \delta_\Omega(z)|
\end{equation}
for all $z$ sufficiently close to $\partial \Omega$.
\end{enumerate}
\end{theorem}
\begin{remark}
$(1)$ In \cite{ChenEssay}, \eqref{eq:1.1} was verified by a different method in case $\partial \Omega$ is uniformly perfect.
$(2)$ Estimate of type \eqref{eq:1.2} was first obtained by B\l ocki \cite{BlockiGreen} for bounded pseudoconvex domains with Lipschitz boundaries in $\mathbb C^n$ $($see also \cite{ChenH-index} and \cite{DO} for related results$)$.
$(3)$ We conjecture that \eqref{eq:1.3} can be improved to $ d_B(z_0,z) \gtrsim \log |\log \delta_\Omega(z)|$.
\end{remark}
By now it becomes a standard method for lower bounds of Bergman functions by estimating the set
$
\{g_\Omega(\cdot,z)\le -c\}
$
for $c>0$, where $g_\Omega(\cdot,z)$ stands for the (pluricomplex) Green function with pole at $z$. Namely, one has
\begin{equation}\label{eq:BG_1}
K_\Omega(z) \gtrsim_c | \{g_\Omega(\cdot,z)\le -c\}|^{-1},
\end{equation}
\begin{equation}\label{eq:BG_2}
d_B(z,z')\gtrsim_c 1\ \ \text{whenever}\ \ \{g_\Omega(\cdot,z)\le -c\}\cap \{g_\Omega(\cdot,z')\le -c\}=\emptyset,
\end{equation}
where $A\gtrsim_c B$ means $A\ge CB$ for some constant $C=C(c)>0$ (cf. \cite{BlockiGreen}, \cite{BP}, \cite{BZ}, \cite{Chen99,CZ,ChenEssay,ChenH-index}, \cite{DO}, \cite{Herbort}, \cite{Ohsawa}). In the proofs of Theorem \ref{th:CapBerg} and Theorem \ref{th:LowDist}, we use \eqref{eq:BG_1} and \eqref{eq:BG_2} respectively.
\section{Capacities}
In this section we shall review different notions of capacities and present some basic properties of them.
Let $\Omega$ be a bounded domain in $\mathbb C$ and $K\subset \Omega$ a compact (non-polor) set in $\Omega$. We define the Dirichlet capacity $\mathcal C_d(K,\Omega)$ of $K$ relative to $\Omega$ by
\begin{equation}\label{eq:C_1}
\mathcal C_d(K,\Omega) = \inf_{\phi\in \mathcal L(K,\Omega)} \int_\Omega |\nabla \phi|^2
\end{equation}
where $\mathcal L(K,\Omega)$ is the set of all locally Lipschitz functions $\phi$ on $\Omega$ with a compact support in $\overline{\Omega}$
such that $0\le \phi\le 1$ and $\phi|_K=1$. If $\Omega = \mathbb C$, then we write $\mathcal C_d(K)$ for $\mathcal C_d(K,\Omega)$.
By the definition we have
\begin{equation}\label{eq:C_2}
K_1\subseteq K_2\ \ \text{and}\ \ \Omega_1\supseteq \Omega_2\ \ \Rightarrow\ \ \mathcal C_d(K_1,\Omega_1)\le \mathcal C_d(K_2,\Omega_2).
\end{equation}
In view of Dirichlet's principle, the infimum in \eqref{eq:C_1} is attained at the function $\phi_{\min}$ which is exactly the Perron solution to the following (generalized) Dirichlet problem in $\Omega\backslash K$:
\begin{equation}\label{eq:C_3}
\Delta u =0;\ \ u =0\ \text{n.e.\ on\ } \partial \Omega;\ \ u=1\ \text{n.e.\ on\ } \partial K.
\end{equation}
We call $\phi_{\min}$ the\/ {\it capacity potential}\/ of $K$ relative to $\Omega$. In case $\partial \Omega$ and $\partial K$ are both $C^2-$smooth, integration by parts gives
\begin{eqnarray}\label{eq:C_4}
\mathcal C_d(K,\Omega) & = & \int_\Omega |\nabla \phi_{\min}|^2 = \int_{\Omega\backslash K} |\nabla \phi_{\min}|^2\nonumber\\
& = & -\int_{\Omega\backslash K} \phi_{\min}\,\Delta \phi_{\min} +\int_{\partial (\Omega\backslash K)} \phi_{\min} \frac{\partial \phi_{\min}}{\partial \nu} d\sigma\nonumber\\
& = & \int_{\partial K} \frac{\partial \phi_{\min}}{\partial \nu} d\sigma=: - {\text{flux}}_{\partial K}\, \phi_{\min}
\end{eqnarray}
where $\nu$ is the outward unit normal vector fields on $\partial (\Omega\backslash K)$. By the Hopf lemma we conclude that $\partial \phi_{\min}/\partial \nu\ge 0$ holds on $\partial K$.
Let $g_\Omega(z,w)$ be the (negative) Green function on $\Omega$. Let $z\in \Omega\backslash K$ be given. Since $\Delta g_\Omega(\cdot,z) =2\pi \delta_z$, where $\delta_z$ stands for the Dirac measure at $z$, we infer from Green's formula that
\begin{eqnarray}\label{eq:C_4'}
2\pi \phi_{\min}(z) & = & \int_{\Omega\backslash K} \phi_{\min}\,\Delta g_\Omega(\cdot,z)= \int_{\Omega\backslash K} g_\Omega(\cdot,z)\,\Delta \phi_{\min}\nonumber\\
&& + \int_{\partial (\Omega\backslash K)} \phi_{\min}\,\frac{\partial g_\Omega(\cdot,z)}{\partial \nu} d\sigma
- \int_{\partial (\Omega\backslash K)} g_\Omega(\cdot,z)\,\frac{\partial \phi_{\min}}{\partial \nu} d\sigma\nonumber\\
& = & \int_{\partial K} \frac{\partial g_\Omega(\cdot,z)}{\partial \nu} d\sigma - \int_{\partial K} g_\Omega(\cdot,z)\,\frac{\partial \phi_{\min}}{\partial \nu} d\sigma\nonumber\\
& = & - \int_{\partial K} g_\Omega(\cdot,z)\,\frac{\partial \phi_{\min}}{\partial \nu} d\sigma
\end{eqnarray}
because $g_\Omega(\cdot,z)$ is harmonic on $K$. This equality combined with \eqref{eq:C_4} gives the following fundamental inequality which connects the capacity, Green's function and the capacity potential:
\begin{equation}\label{eq:C_4''}
\frac{\mathcal C_d(K,\Omega)}{2\pi}\, \inf_{\partial K}(-g_\Omega(\cdot,z))\le \phi_{\min}(z)\le \frac{\mathcal C_d(K,\Omega)}{2\pi}\,\sup_{\partial K}(-g_\Omega(\cdot,z)),\ \ z\in \Omega\backslash K.
\end{equation}
Since $\Omega\backslash K$ can be exhausted by bounded domains with smooth boundaries, we conclude by passing to a standard limit process that the same inequality holds for every compact set $K$.
For a finite Borel measure $\mu$ on $\mathbb C$ whose support is contained in $K$ we define its Green potential relative to $\Omega$ by
$$
p_{\mu}(z) = \int_\Omega g_\Omega(z,w) d\mu(w), \ \ z\in \Omega.
$$
Clearly, $p_\mu$ is negative, subharmonic on $\Omega$, harmonic on $\Omega\backslash K$, and satisfies $p_\mu(z)=0$ n.e. on $ \partial \Omega$. Given $\phi\in C^\infty_0(\Omega)$ we have
\begin{eqnarray*}
\int_\Omega p_\mu \Delta \phi dV & = & \int_\Omega \left[\int_\Omega g_\Omega(z,w) d\mu(w) \right] \Delta\phi(z)dV(z)\\
& = & \int_\Omega \left[\int_\Omega g_\Omega(z,w) \Delta\phi(z)dV(z) \right] d\mu(w) \ \ (\text{Fubini's theorem})\\
& = & \int_\Omega \left[\int_\Omega \Delta g_\Omega(z,w) \phi(z)dV(z) \right] d\mu(w)\ \ (\text{Green's formula})\\
& = & \int_\Omega 2\pi \phi(w) d\mu(w).
\end{eqnarray*}
Thus we obtain
\begin{equation}\label{eq:C_5}
\Delta p_\mu =2\pi \mu
\end{equation}
in the sense of distributions.
The Green energy $I(\mu)$ of $\mu$ is given by
$$
I(\mu):= \int_\Omega p_\mu d\mu = \int_\Omega \int_\Omega g_\Omega(z,w) d\mu(z)d\mu(w).
$$
By \eqref{eq:C_5} we have
\begin{equation}\label{eq:C_6}
I(\mu) = \frac1{2\pi} \int_\Omega p_\mu\, \Delta p_\mu=-\frac1{2\pi} \int_\Omega |\nabla p_\mu|^2.
\end{equation}
Every compact set $K$ has an\/ {\it equilibrium measure} $\mu_{\max}$, which maximizes $I(\mu)$ among all Borel probability measures $\mu$ on $K$. A fundamental theorem of Frostman states that
\begin{enumerate}
\item[$(1)$] $p_{\mu_{\max}}\ge I(\mu_{\max})$ on $\Omega$;
\item[$(2)$] $p_{\mu_{\max}}=I(\mu_{\max})$ on $K\backslash E$ for some $F_\sigma$ polor set $E\subset \partial K$.
\end{enumerate}
By the uniqueness of the solution of the (generalized) Dirichlet problem we have
\begin{equation}\label{eq:C_7}
\phi_{\min}= p_{\mu_{\max}}/ I(\mu_{\max}).
\end{equation}
We define the Green capacity $\mathcal C_{g}(K,\Omega)$ of $K$ relative to $\Omega$ by
$$
\mathcal C_{g}(K,\Omega):= e^{I(\mu_{\max})}.
$$
It follows from \eqref{eq:C_6} and \eqref{eq:C_7} that
\begin{equation}\label{eq:C_8}
\frac{\mathcal C_d(K,\Omega)}{2\pi}=-\frac1{\log \mathcal C_g(K,\Omega)}.
\end{equation}
Analogously, we may define the logarithmic capacity $\mathcal C_l(K)$ of $K$ by
$$
\log \mathcal C_l(K):=\sup_\mu \int_\mathbb C \int_\mathbb C \log |z-w| d\mu(z) d\mu(w)
$$
where the supremum is taken over all Borel propability measures $\mu$ on $\mathbb C$ whose support is contained in $K$. Let $R$ be the diameter of $\Omega$ and set $d=d(K,\Omega)$. Since
$$
\log |z-w|/R \le g_\Omega(z,w)\le \log |z-w|/d,\ \ z,w\in K,
$$
we have
\begin{equation}\label{eq:C_9}
\log \mathcal C_l(K)-\log R \le \log \mathcal C_g(K,\Omega)\le \log \mathcal C_l(K)-\log d.
\end{equation}
\section{Estimates of the capacity potential}
We first give a basic lemma as follows.
\begin{lemma}\label{lm:basic}
Let $\Omega$ be a bounded domain in $\mathbb C$ with $0\in \partial \Omega$. Let $0\le h\le 1$ be a harmonic function on $\Omega$ such that $h=0$ n.e. on $\partial \Omega \cap \mathbb D_{r_0}$ for some $r_0<1$. For all $0<\alpha< 1/16$ and $r\le \alpha r_0$ we have
\begin{equation}\label{eq:G_0}
\sup_{\Omega\cap \mathbb D_r} h \le \exp\left[-\frac{\log 1/(16\alpha)}{\log 1/\alpha} \int_{r}^{\alpha r_0} \left({t\log \frac{ t/\alpha}{2 \mathcal C_l(K_{ t})}} \right)^{-1} dt \right]
\end{equation}
where $K_t:=\overline{\mathbb D}_t-\Omega$.
\end{lemma}
\begin{proof}
The idea of the proof comes from \cite{GW} (see also \cite{ChenHolder}). Let $\mathbb D$ be the unit disc.
For $t<r_0$ and $|z|=t$ we have
\begin{equation}\label{eq:G_1}
\sup_{\partial K_{\alpha t}}(-g_{\mathbb D}(\cdot,z))\le \log 2+ \sup_{\partial K_{\alpha t}}(-\log|\cdot-z|) \le \log 2-\log |t-\alpha t|\le \log 4/t.
\end{equation}
Let $\phi_{\alpha t}$ be the capacity potential of $K_{\alpha t}$ relative to $\mathbb D$. By \eqref{eq:C_4''} and \eqref{eq:G_1} we have
$$
\phi_{\alpha t}(z)\le (\log 4/t)\cdot\frac{\mathcal C_d(K_{\alpha t},\mathbb D)}{2\pi}\ \ \ \ \text{for}\ \ |z|=t.
$$
It follows that for $z\in \Omega\cap \partial \mathbb D_t$,
\begin{equation}\label{eq:G_2}
(1-\phi_{\alpha t}(z)) \sup_{\Omega\cap \mathbb D_t} h \ge \left[1-\log \frac4t \cdot\frac{\mathcal C_d(K_{\alpha t},\mathbb D)}{2\pi}\right]h(z),
\end{equation}
while the same inequality holds for $z\in \partial \Omega\cap \mathbb D_t$, because $\lim_{z\rightarrow \zeta}h=0$ for n.e. $\zeta\in \partial \Omega\cap \mathbb D_t$. By the (generalized) maximum principle, \eqref{eq:G_2} holds on $\Omega\cap \mathbb D_t$. On the other hand, since for $|z|= \alpha t$ we have
\begin{equation}\label{eq:G_3}
\inf_{\partial K_{\alpha t}}(-g_{\mathbb D}(\cdot,z)) \ge \log 1/2+ \inf_{\partial K_{\alpha t}}(-\log|\cdot-z|) \ge \log 1/2-\log (2\alpha t) = \log \frac1{4\alpha t},
\end{equation}
it follows from \eqref{eq:C_4''} that
\begin{equation}\label{eq:G_4}
\phi_{\alpha t}(z)\ge \log \frac1{4\alpha t}\cdot \frac{\mathcal C_d(K_{\alpha t},\mathbb D)}{2\pi}\ \ \ \ \text{for}\ \ |z|=\alpha t.
\end{equation}
Substituting \eqref{eq:G_4} into \eqref{eq:G_2}, we have
\begin{eqnarray}\label{eq:G_5}
h(z) & \le & \sup_{\Omega\cap \mathbb D_t} h \cdot \frac{1-\log \frac1{4\alpha t}\cdot\frac{\mathcal C_d(K_{\alpha t},\mathbb D)}{2\pi}}{1-\log \frac4t \cdot\frac{\mathcal C_d(K_{\alpha t},\mathbb D)}{2\pi}}\nonumber\\
&\le & \sup_{\Omega\cap \mathbb D_t} h \left(1+\frac{\log (16\alpha)\cdot\frac{\mathcal C_d(K_{\alpha t},\mathbb D)}{2\pi}}{1-\log \frac4t \cdot\frac{\mathcal C_d(K_{\alpha t},\mathbb D)}{2\pi}}\right)
\end{eqnarray}
for $z\in \Omega\cap \mathbb D_{\alpha t}$. Set $M(t):=\sup_{\Omega\cap \mathbb D_t} h$. It follows from \eqref{eq:G_5} and \eqref{eq:C_8} that
\begin{eqnarray*}
\frac{\log M(t)}t-\frac{\log M(\alpha t)}t & \ge & \log\frac1{16\alpha}\left({t\log \frac{t}{4 \mathcal C_g(K_{\alpha t},\mathbb D)}} \right)^{-1}\\
& \ge & \log\frac1{16\alpha} \left({t\log \frac{t}{2 \mathcal C_l(K_{\alpha t})}} \right)^{-1}\ \ \ \ (\text{by\ } \eqref{eq:C_9}).
\end{eqnarray*}
Integration from $r/\alpha$ to $r_0$ gives
\begin{eqnarray*}
&& \log\frac1{16\alpha} \int_{r/\alpha}^{r_0} \left[{t\log \frac{t}{2 \mathcal C_l(K_{\alpha t})}} \right]^{-1} dt\\
& \le & \int_{r/\alpha}^{r_0} \frac{\log M(t)}t dt- \int_{r/\alpha}^{r_0} \frac{\log M(\alpha t)}t dt\\
& = & \int_{r/\alpha}^{r_0} \frac{\log M(t)}t dt- \int_{r}^{\alpha r_0} \frac{\log M( t)}t dt\\
& \le & \int_{\alpha r_0}^{r_0} \frac{\log M(t)}t dt -\int_r^{r/\alpha} \frac{\log M(t)}t dt\\
&\le & (\log M(r_0)-\log M(r)) \log 1/\alpha,
\end{eqnarray*}
because $M(t)$ is nondecreasing. Thus \eqref{eq:G_0} holds because $M(r_0)\le 1$.
\end{proof}
\begin{theorem}\label{th:basic}
Fix a compact set $E$ in $\Omega$ with $\mathcal C_l(E)>0$. Let $\phi_{E}$ be the capacity potential of $E$ relative to $\Omega$. Set $d=d(E,\partial \Omega)$. Then for all $0<\alpha<1/16$
\begin{equation}\label{eq:G_6}
\sup_{\Omega\cap \mathbb D_r} \phi_{E} \le \exp\left[-\frac{\log 1/(16\alpha)}{\log 1/\alpha} \int_{r}^{\alpha d} \left({t\log \frac{t/\alpha}{2 \mathcal C_l(K_{t})}} \right)^{-1} dt \right].
\end{equation}
\end{theorem}
\begin{proof}
The solution of the (generalized) Dirichlet problem gives $\lim_{z\rightarrow \zeta}\phi_E(z)=0$ for n.e. $\zeta\in \partial \Omega$, and the (generalized) maximum principle gives $0\le \phi_E\le 1$. Thus Lemma \ref{lm:basic} applies.
\end{proof}
\section{Growth of the Green function}
\begin{proof}[Proof of Theorem \ref{th:CT}]
Since $ \mathcal D_W(\varepsilon,\lambda)>0$, there exist $c>0$ and $n_0\in \mathbb Z^+$ such that
$$
|\mathcal N_a^n(\varepsilon,\lambda)|\ge c n,\ \ \ \forall\,n\ge n_0\ \text{and}\ \ a\in \partial \Omega.
$$
Since $\mathcal N_a^n(\varepsilon,\lambda)$ is decreasing in $\varepsilon$, we may assume that $\varepsilon$ is as small as we want.
Note that for $n\gg N\gg 1$
\begin{eqnarray}\label{eq:G_8'}
\int_{\lambda^n}^{\lambda^N} \left({t\log \frac{t/\alpha}{2 \mathcal C_l(K_{t}(a))}} \right)^{-1} dt & \ge & \sum_{k\in \mathcal N_a^n(\varepsilon,\lambda)\backslash \mathcal N_a^N(\varepsilon,\lambda)}
\int_{\lambda^{k}}^{\lambda^{k-1}} \left({t\log \frac{t/\alpha}{2 \mathcal C_l(K_{t}(a))}} \right)^{-1} dt\nonumber\\
& \ge & \sum_{k\in \mathcal N_a^n(\varepsilon,\lambda)\backslash \mathcal N_a^N(\varepsilon,\lambda)} \left(\log\frac1{2\lambda\varepsilon\alpha}\right)^{-1} \int_{\lambda^{k}}^{\lambda^{k-1}} \frac{dt}t \nonumber \\
& = &
\log1/\lambda \cdot \left(\log\frac1{2\lambda\varepsilon\alpha}\right)^{-1} |\mathcal N_a^n(\varepsilon,\lambda)\backslash \mathcal N_a^N(\varepsilon,\lambda)|\nonumber\\
&\ge & \log1/\lambda \cdot \left(\log\frac1{2\lambda\varepsilon\alpha}\right)^{-1}\cdot\frac{cn}2\nonumber\\
& = & \frac{c}2\cdot \left(\log\frac1{2\lambda\varepsilon\alpha}\right)^{-1}\cdot \log 1/\lambda^n.
\end{eqnarray}
Since for every $z$ there exists $n\in \mathbb Z^+$ such that $\lambda^n\le |z-a|\le \lambda^{n-1}$, it follows from \eqref{eq:G_6} and \eqref{eq:G_8'} that
$$
\phi_E(z)\lesssim |z-a|^\beta
$$
for suitable constant $\beta>0$ which is independent of $a$. Since $-g_\Omega(z,z_0)\asymp \phi_E(z)$ for all $z$ sufficiently close to $\partial \Omega$, we conclude that \eqref{eq:CT} holds.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:logGreen}]
$(1)$ Since $ \mathcal D_W(\varepsilon,\lambda,\gamma)>0$, there exist $c>0$ and $n_0\in \mathbb Z^+$ such that
$$
\sum_{k\in \mathcal N_a^n(\varepsilon,\lambda,\gamma)} k^{-1} \ge c \log n,\ \ \ \forall\,n\ge n_0\ \ \text{and}\ \ a\in \partial \Omega.
$$
Note that for $n\gg N\gg 1$,
\begin{eqnarray*}
\int_{\lambda^n}^{\lambda^N} \left({t\log \frac{t/\alpha}{2\mathcal C_l(K_{t}(a))}} \right)^{-1} dt
& \ge & \sum_{k\in \mathcal N_a^n(\varepsilon,\lambda,\gamma)\backslash \mathcal N_a^N(\varepsilon,\lambda,\gamma)}
\int_{\lambda^{k}}^{\lambda^{k-1}} \left({t\log \frac{t/\alpha}{2\mathcal C_l(K_{t}(a))}} \right)^{-1} dt\\
& \ge & \sum_{k\in \mathcal N_a^n(\varepsilon,\lambda,\gamma)\backslash \mathcal N_a^N(\varepsilon,\lambda,\gamma)} \left(\log\frac{\lambda^{(1-\gamma)k-1}}{2 \varepsilon\alpha}\right)^{-1} \int_{\lambda^{k}}^{\lambda^{k-1}} \frac{dt}t \\
& \gtrsim & \sum_{k\in \mathcal N_a^n(\varepsilon,\lambda,\gamma)\backslash \mathcal N_a^N(\varepsilon,\lambda,\gamma)} k^{-1}\\
& \gtrsim & \log n
\end{eqnarray*}
where the implicit constants are independent of $a$.
This combined with \eqref{eq:G_6} gives
$$
\phi_E(z)\lesssim (-\log |z-a|)^{-\beta}
$$
for some constant $\beta>0$ independent of $a$, which in turn implies
\eqref{eq:LG_1}.
$(2)$
By \eqref{eq:G_6} we have for every $a\in \partial \Omega$
\begin{eqnarray*}
\sup_{\Omega\cap \mathbb D_r(a)} \phi_{E} & \le & \exp\left[-\frac{\log 1/(16\alpha)}{\log 1/\alpha} \int_{r}^{\alpha d} \left({t\log \frac{t/\alpha}{2\mathcal C_l(K_{t}(a))}} \right)^{-1} dt \right]\\
& \le & \exp\left[-\frac{\log 1/(16\alpha)}{\log 1/\alpha} \int_{r}^{\alpha d} \frac{dt} {t((\gamma-1)\log 1/t+ \log 1/(2\alpha \varepsilon) )} \right]\\
& \le & {\rm const}_\tau\, (-\log r)^\tau
\end{eqnarray*}
provided $\alpha$ sufficiently small, from which the assertion follows.
\end{proof}
\section{Lower bounds of the Bergman kernel}
\subsection{Proof of Theorem \ref{th:CapBerg}}
\begin{lemma}[cf. \cite{Grigoryan}]
Let $\Omega$ be a bounded domain in $\mathbb C$ and $U$ a relatively compact open set in $\Omega$. For every $w\in U$ we have
\begin{equation}\label{eq:G_12}
\min_{\partial U} (-g_\Omega(\cdot,w))\le \frac{2\pi}{\mathcal C_d\left(\overline{U},\Omega\right)} \le \max_{\partial U} (-g_\Omega(\cdot,w)).
\end{equation}
\end{lemma}
\begin{proof}
Since $g_\Omega(\cdot,w)$ is harmonic on $\Omega\backslash \overline{U}$ and vanishes n.e on $\partial \Omega$, it follows from the maximum principle that
$$
\sup_{\Omega\backslash U} (-g_\Omega(\cdot,w)) = \max_{\partial U} (-g_\Omega(\cdot,w))\ \ \text{and}\ \ \inf_{ \overline{U}} (-g_\Omega(\cdot,w)) = \min_{\partial U} (-g_\Omega(\cdot,w)).
$$
Then we have
$$
\left\{-g_\Omega(\cdot,w)\ge \max_{\partial U} (-g_\Omega(\cdot,w))\right\} \subset \overline{U}\subset \left\{-g_\Omega(\cdot,w)\ge \min_{\partial U} (-g_\Omega(\cdot,w))\right\}.
$$
Set $F_c=\{-g_\Omega(\cdot,w)\ge c\}$. It suffices to show
$$
\mathcal C_d(F_c,\Omega) =2\pi/c.
$$
Indeed, the function $\phi_c:=-c^{-1} g_\Omega(\cdot,w)$ is the capacity potential of $F_c$ relative to $\Omega$. Thus we have
$$
\mathcal C_d(F_c,\Omega) = -\text{flux}_{\partial F_c}\, \phi_c= -\text{flux}_{\partial \Omega}\, \phi_c = c^{-1} \text{flux}_{\partial \Omega}\, g_\Omega(\cdot,w)=2\pi/c.
$$
where the second and last equalities follow from Green's formula.
\end{proof}
\begin{lemma}\label{lm:Green_2}
Let $\Omega$ be a bounded domain in $\mathbb C$ with $0\in \partial \Omega$. Let $\beta> \alpha>0$. Suppose $\mathcal C_l(K_r)\ge \varepsilon r$ for some $\varepsilon,r>0$. There exists a positive number $c$ depending only on $\alpha,\beta,\varepsilon$ such that for every point $w$ with $|w|=\beta r$ and
$\mathbb D_{2\alpha r}(w)\subset \Omega$ we have
\begin{equation}\label{eq:G_13}
\{ g_\Omega(\cdot,w)\le -c\}\subset \mathbb D_{\alpha r}(w).
\end{equation}
\end{lemma}
\begin{proof}
By \eqref{eq:G_12} and Harnack's inequality it suffices to show
\begin{equation}\label{eq:G_14}
\mathcal C_d\left(\overline{\mathbb D_{\alpha r}(w)},\Omega\right)\ge c'
\end{equation}
for some positive constant $c'$ depending only on $\alpha,\beta,\varepsilon$. By the definition we see that
\begin{eqnarray}\label{eq:G_15}
\mathcal C_d\left(\overline{\mathbb D_{\alpha r}(w)},\Omega\right) & = & \mathcal C_d\left(\Omega^c,\mathbb C_\infty-\overline{\mathbb D_{\alpha r}(w)}\right)\nonumber\\
& \ge & \mathcal C_d\left(K_r,\mathbb C_\infty-\overline{\mathbb D_{\alpha r}(w)}\right),
\end{eqnarray}
where $\mathbb C_\infty$ denotes the Riemann sphere.
Let us consider the conformal map
$$
T: \mathbb C_\infty-\overline{\mathbb D_{\alpha r}(w)}\rightarrow \mathbb D,\ \ \ z\mapsto
\frac{\alpha r}{z-w}.
$$
Since the Dirichlet energy is invariant under conformal maps, it follows that
\begin{equation}\label{eq:G_16}
\mathcal C_d\left(K_r,\mathbb C_\infty-\overline{\mathbb D_{\alpha r}(w)}\right) = \mathcal C_d(T(K_r),\mathbb D).
\end{equation}
Since
$$
K_r\subset \mathbb D_{(1+\beta) r}(w)-{\mathbb D_{2\alpha r}(w)},
$$
we have
$$
T(K_r)\subset \overline{\mathbb D}_{1/2}-\mathbb D_{\alpha/(1+\beta)},
$$
so that
\begin{eqnarray}\label{eq:G_17}
\mathcal C_d(T(K_r),\mathbb D) =-\frac{2\pi}{\log \mathcal C_g(T(K_r),\mathbb D)} \ge \frac{2\pi}{\log 2-\log \mathcal C_l(T(K_r))}. \end{eqnarray}
Since
$$
|T^{-1}(z_1)-T^{-1}(z_2)|=\frac{\alpha r}{|z_1z_2|}\cdot |z_1-z_2|\le \frac{(1+\beta)^2r}{\alpha }\cdot |z_1-z_2|
$$
for all $z_1,z_2\in T(K_r)$, we infer from Theorem 5.3.1 in \cite{Ransford} that
$$
\mathcal C_l(T(K_r))\ge \frac{\alpha}{(1+\beta)^2r}\cdot \mathcal C_l(K_r) \ge \frac{\alpha \varepsilon}{(1+\beta)^2}.
$$
This combined with $\eqref{eq:G_15} \sim \eqref{eq:G_17}$ gives \eqref{eq:G_14}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:CapBerg}]
Take $z^\ast\in \mathcal L_z\cap B(z,\alpha\delta_\Omega(z)) \cap \partial \Omega$. Since
$$
B(z,\alpha\delta_\Omega(z))\subset B(z^\ast,2\alpha \delta_\Omega(z)),
$$
we have
$$
\mathcal C_l\left( \mathcal L_z \cap B(z^\ast,2\alpha \delta_\Omega(z))\cap \Omega^c\right)\ge \frac{\varepsilon}{2\alpha}\cdot (2\alpha\delta_\Omega(z)).
$$
By Lemma \ref{lm:Green_2}, there exists a constant $c=c(\alpha,\varepsilon)>0$ such that
$$
\{g_{\mathcal L_z\cap \Omega}(\cdot,z)\le -c\}\subset \mathcal L_z\cap B(z,\delta_\Omega(z)/2).
$$
Hence it follows from \eqref{eq:BG_1} that
$$
K_{\mathcal L_z\cap \Omega}(z)\gtrsim |\{g_{\mathcal L_z\cap \Omega}(\cdot,z)\le -c\} |^{-1}\gtrsim \delta_\Omega(z)^{-2},
$$
where the implicit constants depend only on $\alpha,\varepsilon$. This together with the Ohsawa-Takegoshi extension theorem yield \eqref{eq:CB_2}.
\end{proof}
\subsection{Proof of Theorem \ref{th:KernelEstimate}}
\begin{proof}[Proof of\/ $(b)$]
Take $z^\ast\in \partial \Omega$ such that $|z-z^\ast| = \delta_\Omega(z)$. Since $$
B(z^\ast,\delta_\Omega(z))\subset B(z,2\delta_\Omega(z)),
$$
it follows from \eqref{eq:CB_3} that
$$
|B(z,2\delta_\Omega(z))\cap \Omega^c|\ge C_n\varepsilon \delta_\Omega(z)^{2n},
$$
where $C_n$ stands for a constant depending only on $n$. We claim that there exists a real line $l_z\ni z$, such that
\begin{equation}\label{eq:kl_3'}
| l_z\cap B(z,2\delta_\Omega(z))\cap \Omega^c|_1 \ge \varepsilon' \delta_\Omega(z),
\end{equation}
where $\varepsilon'=\varepsilon'(n,\varepsilon)>0$, $|\cdot|_1$ stands for the $1-$dimensional Lebesgue measure. Indeed, if we denote $E:=B(z,2\delta_\Omega(z))\cap \Omega^c$ and $E_\zeta:=E\cap l_{z,\zeta}$, where $\zeta\in \mathbb C^n$, $l_{z,\zeta}=\{z+t\zeta:t\in \mathbb R\}$, then
\begin{eqnarray}\label{eq:Vol}
|E| & = & \int_{\mathbb \zeta\in \mathbb S^{2n-1}} \int_{r: z+ r\zeta\in E} r^{2n-1}dr d\sigma(\zeta) \\
& \le & 2^{2n-1} \delta_\Omega(z)^{2n-1} \int_{\mathbb \zeta\in \mathbb S^{2n-1}} |E_\zeta|_1 d\sigma(\zeta),\nonumber
\end{eqnarray}
where $\mathbb S^{2n-1}$ is the unit sphere in $\mathbb C^n$ and $d\sigma$ is the surface element. Thus
$$
\int_{\mathbb \zeta\in \mathbb S^{2n-1}} |E_\zeta|_1 d\sigma(\zeta)\ge C_n 2^{1-2n}\varepsilon \delta_\Omega(z),
$$
so that there exists at least one point $\zeta_0\in \mathbb S^{2n-1}$, such that
$$
|E_{\zeta_0}|_1\ge \frac{C_n\varepsilon}{2^{2n}|\mathbb S^{2n-1}|} \cdot \delta_\Omega(z).
$$
Take $l_z=l_{z,\zeta_0}$ and $\varepsilon'= \frac{C_n\varepsilon}{2^{2n}|\mathbb S^{2n-1}|}$, we get \eqref{eq:kl_3'}.
Let $\mathcal L_z$ be the complex line determined by $l_z$. Then
$$
\mathcal C_l(\mathcal L_z\cap B(z,2\delta_\Omega(z))\cap \Omega^c)\ge \frac{|E_{\zeta_0}|_1}4\ge \frac{\varepsilon'}4\cdot\delta_\Omega(z),
$$
where the first inequality follows from Theorem 5.3.2 in \cite{Ransford}. Thus Theorem \ref{th:CapBerg} applies.
\end{proof}
\begin{proof}[Proof of\/ $(d)$]
By Theorem \ref{th:CapBerg}, it suffices to verify that there exists a complex line $\mathcal L_z\ni z$, such that
\begin{equation}\label{eq:kl_4}
\mathcal C_l(\mathcal L_z\cap B(z,2\delta_\Omega(z))\cap \Omega^c) \ge \delta_\Omega(z)/4.
\end{equation}
This follows from an argument of Pflug-Zwonek \cite{PZ05}: Take
$c_0>0$, so that $\{\rho<c_0\}\subset\subset U$, where $\rho$ is defined on an open set $U\supset \overline{\Omega}$. Take $z^\ast\in \partial \Omega\cap \partial B(z,\delta_\Omega(z))$ and let $\mathcal L_z$ be the complex line determined by $z$ and $z^\ast$. We claim that
\begin{equation}\label{eq:S-hyper}
\mathcal L_z\cap \partial B(z,r)\cap \Omega^c\neq \emptyset, \ \ \ \delta_\Omega(z)<r<d_0-\delta_\Omega(z),
\end{equation}
where $d_0=d(\overline{\Omega},\{\rho=c_0\})$.
Indeed, if $ \mathcal L_z\cap \partial B(z,r)\cap \Omega^c= \emptyset$, i.e.,
$$
\mathcal L_z\cap \partial B(z,r)\subset \Omega.
$$
then the maximum principle and continuity of $\rho$ yield
$$
\rho(z^\ast)\le \max\{\rho(w):w\in \mathcal L_z\cap \partial B(z,r)\}<0 =\rho(z^\ast),
$$
this is a contradiction!
Since the circle projection $T:w\mapsto |w|$ satisfies $|T(w)-T(w')|\le |w-w'|$, it follows from Theorem 5.3.1 and Theorem 5.3.2 in \cite{Ransford} that \eqref{eq:S-hyper} gives \eqref{eq:kl_4}.
\end{proof}
\begin{proof}[Proof of\/ $(c)$]
Given $0<\varepsilon\ll 1$, define
$$
\Omega_\varepsilon:=\left\{z\in U: d(z,\overline{\Omega})< \varepsilon\right\}.
$$
Since $\overline{\Omega}^\circ =\Omega$, it is not difficult to show that $\delta_{\Omega_\varepsilon}(z)\rightarrow \delta_\Omega(z)$ uniformly on $\overline{\Omega}$ as $\varepsilon\rightarrow 0$. Fix a domain $V$ with $\overline{\Omega}\subset V\subset\subset U$.
Since $\overline{\Omega}$ is $\mathcal O(U)-$convex, there exist a finite number of holomorphic functions $f_1,\cdots,f_m$ on $U$, where $m=m_\varepsilon$, such that
$$
\sup_{\overline{\Omega}}\max_{1\le j\le m} |f_j|<1\ \ \text{and}\ \ \inf_{\overline{V}\backslash \Omega_\varepsilon} \max_{1\le j\le m} |f_j|\ge 1.
$$
Clearly, $\rho:=\max_{1\le j\le m} |f_j|-1$ is a continuous psh function on $U$, and we have
$$
\overline{\Omega}\subset \{z\in V: \rho(z)<0\}=:\Omega'\subset \Omega_\varepsilon.
$$
Note that $\Omega$ is pseudoconvex since $\overline{\Omega}$ is the limit of a decreasing family of analytic polyhedrons.
Let $z\in \Omega$ be a fixed point which is sufficiently close to $\partial \Omega$. Take $\varepsilon\ll 1$ such that $\delta_{\Omega_\varepsilon}(z)\le 3\delta_\Omega(z)/2$. Then we have
$$
\delta_{\Omega'}(z)\le \delta_{\Omega_\varepsilon}(z)\le 3\delta_\Omega(z)/2.
$$
Take $z^\ast\in \partial \Omega'\cap \partial B(z,\delta_{\Omega'}(z))$ and let $\mathcal L_z$ be the complex line determined by $z$ and $z^\ast$. From the proof of $(d)$, we know that
$$
\mathcal L_z\cap \partial B(z,r)\cap (\mathbb C^n\setminus \Omega')\neq \emptyset, \ \ \ \delta_{\Omega'}(z)<r<r_0,
$$
where $r_0>0$ is suitable constant. Since $\delta_{\Omega'}(z)\le \frac32\, \delta_\Omega(z)$, we conclude that
$$
\mathcal C_l(\mathcal L_z\cap B(z,2\delta_\Omega(z))\cap \Omega^c) \ge \mathcal C_l(\mathcal L_z\cap B(z,2\delta_\Omega(z))\cap (\mathbb C^n\setminus \Omega')) \ge \frac14\cdot \frac{\delta_\Omega(z)}2
$$
in view of the circular projection.
Thus Theorem \ref{th:CapBerg} applies.
\end{proof}
\begin{proof}[Proof of\/ $(a)$]
Let $z_0\in \partial \Omega$. After a complex affine transformation, we may assume that $z_0=0$ and there exist a ball in $\mathbb C^{n-1}\times \mathbb R$ given by
$$
B'_{r_0}:=\left\{(z',\mathrm{Re}\,z_n)\in \mathbb C^{n-1} \times \mathbb R: |z'|^2+(\mathrm{Re}\,z_n)^2<r_0^2\right\},
$$
and a continuous real-valued function $g$ on $B'_{r_0}$ such that
$$
\Omega\cap \left\{B'_{r_0}\times (-r_0,r_0)\right\}=\left\{z\in B'_{r_0}\times (-r_0,r_0): \mathrm{Im}\,z_n < g(z',\mathrm{Re}\,z_n) \right\}.
$$
Given $0\le t\le \varepsilon_0\ll r_0$, define
$$
\Omega_t:=\left\{z\in B'_{\frac{r_0}2+t}\times (-r_0/2-t,r_0): \mathrm{Im}\,z_n < g(z',\mathrm{Re}\,z_n) +t \right\}.
$$
Note that $\Omega_0=\Omega\cap \left\{B'_{\frac{r_0}2}\times (-r_0/2,r_0)\right\}$ and $\{\Omega_t\}_{t\in [0,\varepsilon_0]}$ forms an increasing $1-$parameter family of bounded pseudoconvex\/ {\it fat}\/ domains.
By $(c)$ and the well-known localization principle of the Bergman kernel, it suffices to show that the closure $\overline{\Omega}_0$ of $\Omega_0$ is $\mathcal O(\Omega_{\varepsilon_0})-$convex. Given $w\in \Omega_{\varepsilon_0}\setminus \overline{\Omega}_0$, take $0<t_1<\varepsilon_0$ such that $w\in \partial \Omega_{t_1}$.
Take $\gamma>0$ and $t_1<t_2<\varepsilon_0$ such that $\min_{\overline{\Omega}_0}\delta_{\Omega_{t_1}}>\gamma$ and $\delta_{\Omega_{t_2}}(w)=\gamma/2$. We are going to construct a holomorphic function $f$ on $\Omega_{t_2}$ satisfying $|f(w)|>1$ and $|f|<1$ on $\overline{\Omega}_0$. The argument is essentially standard, but we include the detail here for the sake of completeness. Define $\psi:=-\log \delta_{\Omega_{t_2}}+\log \gamma$. Clearly, $\psi$ is a locally Lipschitz continuous psh exhaustion function on $\Omega_{t_2}$ satisfying
$$
\overline{\Omega}_0\subset \{\psi<0\}\ \ \text{and}\ \ \psi(w)=\log 2.
$$
Let $\kappa$ be a convex nondecreasing function on $\mathbb R$ with $\kappa=0$ on $(-\infty,0]$ and $\kappa',\kappa''>0$ on $(0,\infty)$.
Take a smooth cut-off function $\chi$ on $\mathbb R$ such that $\chi=0$ on $[\frac14 \log 2,\frac13 \log 2]\cup [3\log 2,\infty)$ and $\chi=1$ on $(-\infty,0]\cup [\frac12 \log 2,2\log 2]$.
Define $h=\frac12$ on $\{\psi\le \frac14 \log 2\}$ and $h=2$ on $\{\frac13 \log 2\le \psi \le 3\log 2\}$. Then $\chi(\psi) h$ defines a Lipschitz continuous function with compact support in $\Omega_{t_2}$. By H\"ormander's $L^2-$estimates (cf. \cite{Hormander}), there exists a solution of $\bar{\partial}u=\bar{\partial}(\chi(\psi) h)$ such that
\begin{eqnarray*}
\int_{\Omega_{t_2}} |u|^2 e^{-\kappa(\psi)-2n\log |z-w|} & \lesssim & \int_{\Omega_{t_2}} |h|^2 |\chi'(\psi)|^2 |\bar{\partial}\psi|^2 e^{-\kappa(\psi)-2n\log |z-w|}\\
& \lesssim & \int_{\{0<\psi<\frac12 \log 2\}\cup \{2\log 2<\psi<3\log 2\}} e^{-\kappa(\psi)} |\bar{\partial}\psi|^2,
\end{eqnarray*}
where the implicit constants are independent of the choice of $\kappa$, which can be arbitrarily small provided that $\kappa$ increases on $(0,\infty)$ sufficiently rapidly. Thus $f:=\chi(\psi)h-u$ is a holomorphic function on $\Omega_{t_2}$ which satisfies $f(w)=h(w)=2$ and since $f-h=-u$ is holomorphic on $\{\psi<0\}$, it follows the mean-value inequality that
\begin{eqnarray*}
\sup_{\overline{\Omega}_0} |f-h|^2 & \lesssim &
\int_{\{\psi<0\}} |u|^2 e^{-2n\log |z-w|}\\
& \lesssim & \int_{\Omega_{t_2}} |u|^2 e^{-\kappa(\psi)-2n\log |z-w|},
\end{eqnarray*}
which can be arbitrarily small provided that $\kappa$ increases on $(0,\infty)$ sufficiently rapidly. Thus $|f|<1$ on $\overline{\Omega}_0$.
Finally, since $\{\Omega_t\}_{t\in [0,\varepsilon_0]}$ forms an increasing $1-$parameter family of bounded pseudoconvex domains, it follows from a classical result of Docquier-Grauert \cite{DG} that $(\Omega_{t_2},\Omega_{\varepsilon_0})$ forms a Runge pair, i.e., every holomorphic function on $\Omega_{t_2}$ can be approximated locally uniformly on $\Omega_{t_2}$ by holomorphic functions on $\Omega_{\varepsilon_0}$. From this we obtain a holomorphic function $\widetilde{f}$ on $\Omega_{\varepsilon_0}$ satisfying $|\widetilde{f}(w)|>1$ and $|\widetilde{f}|<1$ on $\overline{\Omega}_0$. Thus $\overline{\Omega}_0$ is $\mathcal O(\Omega_{\varepsilon_0})-$convex.
\end{proof}
\subsection{Proof of Proposition \ref{prop:Runge}}
We first give an elementary lemma as follows.
\begin{lemma}\label{lm:sc}
If\/ $D\subsetneq \mathbb C$ is simply-connected, then
$$
K_D(z)\ge \frac1{16\pi}\cdot \delta_D(z)^{-2},\ \ \ z\in D.
$$
\end{lemma}
\begin{proof}
Given $z\in D$, consider the Riemann mapping $F:D\rightarrow \mathbb D$ satisfying
$$
F(z)=0,\,F'(z)>0.
$$
It is a standard consequence of the Kobe one-quarter theorem that
$\delta_D(z)\ge \frac1{4|F'(z)|}$ (cf. \cite{CG}, Theorem 1.4). Thus
$$
K_D(z)=K_{\mathbb D}(0) |F'(z)|^2\ge \frac1{16 \pi}\cdot \delta_D(z)^{-2}.
$$
\end{proof}
\begin{proof}[Proof of Proposition \ref{prop:Runge}]
Note that the Runge domain $\Omega$ enjoys the following geometric property: for every complex line $\mathcal L$, all components of $\Omega\cap L$ are simply-connected. To see this, let $D$ be a component of $\Omega\cap \mathcal L$. Given $f\in \mathcal O(D)$, the holomorphic function on $\Omega\cap \mathcal L$ which equals $f$ on $D$ and $0$ on other components can be extended to some $\widetilde{f}\in \mathcal O(\Omega)$, for $\Omega$ is pseudoconvex.
As $\Omega$ is Runge, there exists a sequence of complex polynomials $\{P_j\}$ converging locally uniformly to $\widetilde{f}$ on $\Omega$. Thus the polynomials $\left\{P_j|_{\mathcal L}\right\}$ converge locally uniformly on $D$ to $f$, so that $D$ has to be simply-connected.
Given $z\in \Omega$, take $z^\ast\in \partial \Omega$ so that $|z-z^\ast|=\delta_\Omega(z)$. Let $\mathcal L_z$ be the complex line determined by $z,z^\ast$. Let $D_z$ be the component of $\Omega\cap \mathcal L_z$ which contains $z$. Clearly, we have
\begin{equation}\label{eq:sc}
\delta_{D_z}(z)=\delta_\Omega(z).
\end{equation}
Set $f=K_{D_z}(\cdot,z)/\sqrt{K_{D_z}(z)}$ on $D_z$ and $f=0$ on other components of $\Omega\cap \mathcal L_z$. By the Ohsawa-Takegoshi extension theorem, there is a holomorphic function $\widetilde{f}$ on $\Omega$ such that $\widetilde{f}|_{\Omega\cap \mathcal L_z}=f$ and
$$
\int_\Omega |\widetilde{f}|^2 \lesssim \int_{\Omega\cap \mathcal L_z} |f|^2 \lesssim 1,
$$
where the implicit constants depend only on the diameter of $\Omega$. Thus
$$
K_\Omega(z)\ge \frac{|\widetilde{f}(z)|^2}{\int_\Omega |\widetilde{f}|^2} \gtrsim K_{D_z}(z) \gtrsim \delta_{D_z}(z)^{-2}
$$
in view of Lemma \ref{lm:sc}. This together with \eqref{eq:sc} conclude the proof.
\end{proof}
\subsection{Proofs of Theorem \ref{th:CapBerg_2} and Corollary \ref{cor:VolBerg_2}}
The following lemma is essentially implicit in \cite{Zwonek}.
\begin{lemma}\label{lm:CapBerg_2}
Let $\Omega$ be a bounded domain in $\mathbb C$ and $\alpha>1$. Then for any $z\in \Omega$ with $\alpha \delta_\Omega(z)\le 1/2$,
\begin{equation}\label{eq:Z_1}
K_\Omega(z)\gtrsim (\alpha\delta_\Omega(z))^{-2}(-\log \mathcal C_l(B(z,\alpha\delta_\Omega(z))\cap \Omega^c))^{-1},
\end{equation}
where the implicit constant depends only on $\mathrm{diam\,}\Omega$.
\end{lemma}
\begin{proof}
Given $z\in \Omega$, set $E:=\overline{B(z,\alpha\delta_\Omega(z))}\cap \Omega^c$. Without loss of generality, we assume $\mathcal C_l(E)>0$. Dividing $\mathbb C$ into three trisection angles with vertices at $z$, one may divide $E$ into three parts $E_1,E_2,E_3$. We may assume $\mathcal C_l(E_1)\ge \mathcal C_l(E_2)\ge \mathcal C_l(E_3)$. Then
\begin{equation}\label{eq:Z_2}
\frac1{-\log \mathcal C_l(E)}\le \frac1{-\log \mathcal C_l(E_1)}+ \frac1{-\log \mathcal C_l(E_2)} + \frac1{-\log \mathcal C_l(E_3)}\le \frac3{-\log \mathcal C_l(E_1)}.
\end{equation}
Let $\mu_{E_1}$ denote the equilibrium measure of $E_1$. Consider the following function
$$
f_{E_1}(\zeta):=\int_{E_1} \frac{d\mu_{E_1}(w)}{\zeta-w},\ \ \ \zeta\in \mathbb C\setminus E_1.
$$
Then $f_{E_1}$ is holomorphic on $\mathbb C\setminus E_1\supset \Omega$ and satisfies
$$
\int_\Omega |f_{E_1}|^2 \lesssim {-\log \mathcal C_l(E_1)} \lesssim {-\log \mathcal C_l(E)},
$$
where the implicit constant depends only on $\mathrm{diam\,}\Omega$ (cf. \cite{Zwonek}, Lemma 3). On the other hand, since $|z-w|\le \alpha \delta_\Omega(z)$ for all $w\in E_1$ and $E_1$ is contained in an angle with vertex $z$ and of the apex angle $2\pi/3$, we conclude that there exists a numerical constant $C_0>0$ such that
$$
| f_{E_1}(z) | \ge C_0 \int_{E_1} \frac{d\mu_{E_1}(w)}{|z-w|}\ge \frac{C_0}{\alpha \delta_\Omega(z)},
$$
for $d\mu_{E_1}$ is a probability measure. Thus
$$
K_\Omega(z)\ge \frac{|f_{E_1}(z)|^2}{\|f\|_{L^2(\Omega)}^2}\gtrsim (\alpha \delta_\Omega(z))^{-2}(-\log \mathcal C_l(E))^{-1}.
$$
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:CapBerg_2}]
The conclusion follows directly from the Ohsawa-Takegoshi extension theorem and Lemma \ref{lm:CapBerg_2}.
\end{proof}
\begin{proof}[Proof of Corollary \ref{cor:VolBerg_2}]
The argument is a mimic of the proof of Theorem \ref{th:KernelEstimate}/$(b)$. Given $z\in \Omega$, take $z^\ast\in \partial \Omega$ such that $|z-z^\ast| = \delta_\Omega(z)$. Since $B(z^\ast,\delta_\Omega(z))\subset B(z,2\delta_\Omega(z))$, it follows from \eqref{eq:CB5} that
\begin{equation}\label{eq:Z_3}
|B(z,2\delta_\Omega(z))\cap \Omega^c|\ge \varepsilon (C_n \delta_\Omega(z))^{2n\beta}.
\end{equation}
We claim that there exists a real line $l_z\ni z$, such that
\begin{equation}\label{eq:Z_4}
| l_z\cap B(z,2\delta_\Omega(z))\cap \Omega^c|_1 \ge \varepsilon' \delta_\Omega(z)^{2n\beta-2n+1}
\end{equation}
for suitable $\varepsilon'=\varepsilon'(n,\varepsilon)>0$. Indeed, if we denote $E:=B(z,2\delta_\Omega(z))\cap \Omega^c$ and $E_\zeta:=E\cap l_{z,\zeta}$, then \eqref{eq:Vol} together with \eqref{eq:Z_3} yield
$$
\int_{\mathbb \zeta\in \mathbb S^{2n-1}} |E_\zeta|_1 d\sigma(\zeta)\ge C_n^{2n\beta} 2^{1-2n}\varepsilon \delta_\Omega(z)^{2n\beta-2n+1},
$$
so that there exists at least one point $\zeta_0\in \mathbb S^{2n-1}$, such that
$$
|E_{\zeta_0}|_1\ge \frac{C_n^{2n\beta}\varepsilon}{2^{2n}|\mathbb S^{2n-1}|} \cdot \delta_\Omega(z)^{2n\beta-2n+1}.
$$
Take $l_z=l_{z,\zeta_0}$ and $\varepsilon'= \frac{C_n^{2n\beta}\varepsilon}{2^{2n}|\mathbb S^{2n-1}|}$, we get \eqref{eq:Z_4}.
Let $\mathcal L_z$ be the complex line determined by $l_z$. Then
$$
\mathcal C_l(\mathcal L_z\cap B(z,2\delta_\Omega(z))\cap \Omega^c)\ge \frac{|E_{\zeta_0}|_1}4\ge \frac{\varepsilon'}4\cdot\delta_\Omega(z)^{2n\beta-2n+1}.
$$
Thus Theorem \ref{th:CapBerg_2} applies.
\end{proof}
\section{Lower bounds of the Bergman distance}
\begin{proposition}\label{prop:Green_2}
Suppose\/ $ \mathcal D_S(\varepsilon,\lambda)>0$\/ for some $\varepsilon,\lambda$. There exists $c\gg 1$ such that for every $k\in \bigcap_{a\in \partial \Omega}\mathcal N_a^n(\varepsilon,\lambda)$ and every $w$ with $\lambda^{k-1/3}\le \delta_\Omega(w)\le \lambda^{k-1/2}$,
\begin{equation}\label{eq:G_18}
\{ g_\Omega(\cdot,w)\le -c\}\subset \left\{ \lambda^k< \delta_\Omega<\lambda^{k-1}\right\}.
\end{equation}
\end{proposition}
\begin{proof}
Take $a(w)\in \partial \Omega$ such that $|w-a(w)|=\delta_\Omega(w)$. Note that
$$
\mathcal C_l(K_{\lambda^k}(a(w)))\ge \varepsilon \lambda^k.
$$
Thus Lemma \ref{lm:Green_2} applies.
\end{proof}
\begin{proposition}\label{prop:Green_1}
If\/ $ \mathcal D_W(\varepsilon,\lambda,\gamma)>0$\/ for some $\varepsilon,\lambda,\gamma$, then there exists $c\gg 1$ such that
\begin{equation}\label{eq:G_10}
\{ g_\Omega(\cdot,w)\le -1\}\subset \left\{c^{-1}\phi_E(w)^{\frac{1+\beta}\beta}< \phi_E < c\,\phi_E(w)^{\frac{\beta}{1+\beta}}\right\}
\end{equation}
where $\beta$ is given as \eqref{eq:LG_1}.
\end{proposition}
\begin{proof}
We shall first adopt a trick from \cite{CG}. Consider two points $z,w$ with $|z-w|$, $\delta_\Omega(z)$ and $\delta_\Omega(w)$ are sufficiently small. We want to show
\begin{equation}\label{eq:G_11}
|\phi_E(z)-\phi_E(w)|\le c_0 (-\log |z-w|)^{-\beta}
\end{equation}
for some numerical constant $c_0>0$. Without loss of generality, we assume $\delta_\Omega(w)\ge \delta_\Omega(z)$. If $|z-w|\ge \delta_\Omega(w)/2$, this follows directly from \eqref{eq:LG_1}. Since $\phi_E$ is a positive harmonic function on $\mathbb D_{\delta_\Omega(w)}(w)$, we see that if $|z-w|\le \delta_\Omega(w)/2$ then
\begin{eqnarray*}
|\phi_E(z)-\phi_E(w)| & \le & \sup_{\mathbb D_{\delta_\Omega(w)/2}(w)} |\nabla \phi_E|\, |z-w|\\
& \le & c_1 \delta(w)^{-1} (-\log \delta_\Omega(w))^{-\beta} |z-w|\ \ \ (\text{by\ } \eqref{eq:LG_1})\\
& \le & c_1 (2|z-w|)^{-1} (-\log (2|z-w|))^{-\beta} |z-w| \\
& \le & c_0 (-\log|z-w|)^{-\beta}.
\end{eqnarray*}
The remaining argument is standard. Let $R$ be the diameter of $\Omega$. By (\ref{eq:G_11}) we conclude that if $\phi_E(z)=\phi_E(w)/2$ then
$$
\log \frac{|z-w|}R \ge -\left(\frac{2c_0}{\phi_E(w)}\right)^{1/\beta}-\log R\ge -c_2 \phi_E(w)^{-1/\beta}
$$
Since $-\phi_E$ is subharmonic on $\Omega$, it follows that
$$
\psi(z):=\left\{
\begin{array}{ll}
\log\frac{|z-w|}{R} & \text{ if\ } \phi_E(z)\ge \phi_E(w)/2\\
\max\left\{\log\frac{|z-w|}{R},- 2c_2 \phi_E(w)^{-1-1/\beta} \phi_E(z)\right\} & \text{otherwise}.
\end{array}
\right.
$$
is a well-defined negative subharmonic function on $\Omega$ with a logarithmic pole at $w$, and if $\phi_E(z)\le \phi_E(w)/2$ then we have
$$
g_\Omega(z,w)\ge \psi(z)\ge - 2c_2 \phi_E(w)^{-1-1/\beta} \phi_E(z),
$$
so that
$$
\{g_\Omega(\cdot,w)\le -1\}\cap \{\phi_E\le \phi_E(w)/2\}\subset \left\{\phi_E\ge (2c_2)^{-1}\phi_E(w)^{1+1/\beta}\right\}.
$$
Since $\{\phi_E\ge \phi_E (w)/2\} \subset \{\phi_E> c^{-1}\phi_E(w)^{1+1/\beta}\}$ if $c\gg 1$, we have
$$
\{g_\Omega(\cdot,w)\le -1\} \subset \left\{\phi_E > c^{-1}\phi_E(w)^{1+1/\beta}\right\}.
$$
By the symmetry of $g_\Omega$, we immediately get
$$
\{g_\Omega(\cdot,w)\le -1\} \subset \left\{\phi_E < (c\phi_E(w))^{\frac{\beta}{1+\beta}}\right\}.
$$
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:LowDist}]
$(1)$ Let $c$ be as Proposition \ref{prop:Green_2}. Let $z$ be sufficiently close to $\partial \Omega$. Take $n\in \mathbb Z^+$ such that $\lambda^n\le \delta_\Omega(z)\le \lambda^{n-1}$. Write
$$
\bigcap_{a\in \partial \Omega}\mathcal N_a^n(\varepsilon,\lambda)=\{k_1<k_2<\cdots<k_{m_n}\}.
$$
We may choose a Bergman geodesic jointing $z_0$ to $z$, and a finite number of points on this geodesic with the following order
$$
z_0 \rightarrow z_{k_1}\rightarrow z_{k_2}\rightarrow \cdots \rightarrow z,
$$
such that
$$
\lambda^{k_j-1/3}\le \delta_\Omega(z_{k_j})\le \lambda^{k_j-1/2}.
$$
By Proposition \ref{prop:Green_2} we have
$$
\{g_\Omega(\cdot,z_{k_j})\le -c\}\cap \{g_\Omega(\cdot,z_{k_{j+1}})\le -c\}=\emptyset
$$
so that $d_B(z_{k_j},z_{k_{j+1}})\ge c_1>0$ for all $j$, in view of \eqref{eq:BG_2}.
Since $m_n\gtrsim n$, we have
$$
d_B(z_0,z) \ge \sum_j d_B(z_{k_j},z_{k_{j+1}})\gtrsim n \gtrsim |\log \delta_\Omega(z)|.
$$
$(2)$ The assertion follows directly from Theorem \ref{th:CT} and Corollary 1.8 in \cite{ChenH-index}.
$(3)$ Let $c$ be as Proposition \ref{prop:Green_1}. Let $z$ be sufficiently close to $\partial \Omega$. We may choose a Bergman geodesic jointing $z_0$ to $z$, and a finite number of points $\{z_k\}_{k=1}^m$ on this geodesic with the following order
$$
z_0 \rightarrow z_1\rightarrow z_2\rightarrow \cdots \rightarrow z_m \rightarrow z,
$$
where
$$
c\,\phi_E(z_{k+1})^{\frac{\beta}{1+\beta}}=c^{-1}\phi_E(z_{k})^{\frac{1+\beta}\beta}
$$
and
$$
c^{-1} \phi_E(z_m)^{\frac{1+\beta}\beta}\le \phi_E(z)\le c\phi_E(z_{m})^{\frac{\beta}{1+\beta}}.
$$
By Proposition \ref{prop:Green_1} we have
$$
\{g_\Omega(\cdot,z_k)\le -1\}\cap \{g_\Omega(\cdot,z_{k+1})\le -1\}=\emptyset
$$
so that $d_B(z_k,z_{k+1})\ge c_1>0$ for all $k$.
Note that
\begin{eqnarray*}
\log \phi_E(z_0) & = & \left(\frac{\beta}{1+\beta}\right)^2 \log \phi_E(z_1)+ \frac{\beta}{1+\beta}\, \log c^2=\cdots \\
& = & \left(\frac{\beta}{1+\beta}\right)^{2m} \log \phi_E(z_m) + \frac{\beta}{1+\beta}\,\frac{1-\left(\frac{\beta}{1+\beta}\right)^{2m}}{1-\left(\frac{\beta}{1+\beta}\right)^2}\, \log c^2.
\end{eqnarray*}
Thus we have
$$
m\asymp \log |\log \phi_E(z_m)|\asymp \log |\log \phi_E(z)| \gtrsim \log\log |\log \delta_\Omega(z)|,
$$
so that
$$
d_B(z_0,z) \ge \sum_{k=1}^{m-1} d_B(z_k,z_{k+1})\ge c_1(m-1) \gtrsim \log\log |\log \delta_\Omega(z)|.
$$
\end{proof}
\section{Proof of Theorem \ref{th:HM}}
Recall the following basic properties of quasiconformal (q.c.) mappings (cf. \cite{Ahlfors}) and holomorphic motions (cf. cf. \cite{MSS}, \cite{BersRoyden}, \cite{Slodkowski}, \cite{ST}).
\begin{enumerate}
\item (Circular Distortion Theorem). If $f$ is a $K-$q.c. mapping of $\mathbb C$ fixing $\infty$, then
$$
|f(z_2)-f(z_0)|\le C_K\, |f(z_1)-f(z_0)|,\ \ \ \forall\,z_1,z_2\in \partial \mathbb D_r(z_0),\,\forall\,r>0.
$$
\item (H\"older continuity). If $f$ is a $K-$q.c. mapping of $\mathbb C$ fixing $\infty$, then
$$
C_K^{-1}|z-z'|^K \le |f(z)-f(z')|\le C_K|z-z'|^{1/K}.
$$
\item ($\lambda-$Lemma). Every holomorphic motion $F:\mathbb D\times
E\rightarrow \mathbb C$ admits an extension to a holomorphic motion $F:\mathbb D\times \mathbb C\rightarrow
\mathbb C$ such that
$F(\lambda,\cdot):\mathbb C\rightarrow \mathbb C$ is a
quasiconformal self-homeomorphism of dilatation not exceeding $
\frac{1+|\lambda|}{1-|\lambda|}$, $\forall\,\lambda\in \mathbb D$.
\end{enumerate}
\begin{proposition}\label{prop:HM}
Given $0<r<1$, there exists a constant $C>0$ such that
$$
K_{ \Gamma_r(F)}((\lambda,w))\ge C \delta_{ \Gamma_r(F)}(\lambda,w)^{-2}.
$$
\end{proposition}
\begin{proof}
Clearly, $\Gamma_r(F)$ is a fat domain. On the other hand, $\overline{\Gamma_r(F)}$ is polynomially convex (cf. \cite{Slodkowski}, Proposition 3.3). Thus Theorem \ref{th:KernelEstimate}/$(c)$ applies.
\end{proof}
\begin{proof}[Proof of Theorem \ref{th:HM}]
We shall borrow some techniques from \cite{CZ2}. Given $z^\ast\in
\mathbb D$, take $z_\ast\in \mathbb D^c$ with $ z_\ast/ z^\ast\in \mathbb R^+$ and
$\frac{z^\ast+z_\ast}2\in \partial
\mathbb D$. It follows the Circular Distortion Theorem that
\begin{eqnarray*}
\max_{|z-z^\ast|=1-|z^\ast|}|F(\lambda,z)-F(\lambda,z^\ast)| & \le &
C_r
\min_{|z-z^\ast|=1-|z^\ast|}|F(\lambda,z)-F(\lambda,z^\ast)|\\
\max_{|z-z_\ast|=1-|z^\ast|}|F(\lambda,z)-F(\lambda,z_\ast)| & \le &
C_r
\min_{|z-z_\ast|=1-|z^\ast|}|F(\lambda,z)-F(\lambda,z_\ast)|
\end{eqnarray*}
and
\begin{eqnarray*}
&& \max_{|z-\frac{z^\ast+z_\ast}2|=1-|z^\ast|}\left|F(\lambda,z)-F\left(\lambda,\frac{z^\ast+z_\ast}2\right)\right|\\
& \le & C_r
\min_{|z-\frac{z^\ast+z_\ast}2|=1-|z^\ast|}\left|F(\lambda,z)-F\left(\lambda,\frac{z^\ast+z_\ast}2\right)\right|
\end{eqnarray*}
hold for all $\lambda\in \overline{\mathbb D}_{r}$ and a (generic)
constant $C_r>0$. It follows that if $C_r\gg 1$ then
\begin{equation}\label{eq:HM_5}
\bigcup_{\lambda\in \mathbb D_{r}}\left\{(\lambda,w)\in {\mathbb
C}^2:|w-F(\lambda,z^\ast)|<C_r^{-1} \left|F\left(\lambda,z_\ast\right)-F(\lambda,z^\ast)\right|
\,\right\}\subset \Gamma_r(F)
\end{equation}
and
$$
h(\lambda,w):=\frac{F(\lambda,z^\ast)-F(\lambda,z_\ast)}{C_r[w-F(\lambda,z_\ast)]}
$$
gives a holomorphic mapping from $\Gamma_r(F)$ into $\mathbb D$. Given
$(\lambda^\ast,F(\lambda^\ast,z^\ast))\in \Gamma_{r'}(F)$,
define a holomorphic embedding as follows
\begin{eqnarray*}
\iota: && \Gamma_r(F)
\rightarrow \mathbb D_r\times \mathbb D\\
&& (\lambda,w) \mapsto (\lambda,w'):=
\left(\lambda,
\frac{h(\lambda,w)-h(\lambda,F(\lambda,z^\ast))}2\right).
\end{eqnarray*}
Since
\begin{eqnarray*}
h(\lambda,w)-h(\lambda,F(\lambda,z^\ast)) & = &
-C_r^{-1} \frac{w-F(\lambda,z^\ast)}{w-F(\lambda,z_\ast)},
\end{eqnarray*}
there exists a constant $s=s(r)<1$ such that
\begin{equation}\label{eq:inclusion}
\mathbb D_r\times \mathbb D\supset \iota (\Gamma_r(F))\supset \mathbb D_r \times \mathbb D_s.
\end{equation}
Set $w^\ast:=F(\lambda^\ast,z^\ast)$. Note that
\begin{eqnarray*}
K_{\Gamma_r(F)}((\lambda^\ast,w^\ast)) & = & K_{\iota(\Gamma_r(F))}( \iota(\lambda^\ast,w^\ast) ) | \partial w'/\partial w|^2\\
& = & K_{\iota(\Gamma_r(F))}( (\lambda^\ast,0)) \, C_r^{-2} | F(\lambda^\ast,z^\ast)-F(\lambda^\ast,z_\ast)|^{-2} \\
&\asymp &
\left|F(\lambda^\ast,z^\ast)-F\left(\lambda^\ast,z_\ast\right)\right|^{-2}
\end{eqnarray*}
in view of \eqref{eq:inclusion}, where the implicit constants depend only $r,r'$. This combined with Proposition \ref{prop:HM} gives
$$
\left|F(\lambda^\ast,z^\ast)-F\left(\lambda^\ast,z_\ast \right)\right|
\lesssim \delta_{ \Gamma_r(F)}(\lambda^\ast,w^\ast)
$$
whenever $\delta_{ \Gamma_r(F)}(\lambda^\ast,w^\ast)\le 1/2$.
Since
$$
\left|F(\lambda^\ast,z^\ast)-F\left(\lambda^\ast,z_\ast\right)\right|\ge \delta_{\lambda^\ast}(w^\ast),
$$
we conclude the proof.
\end{proof}
|
\section{Introduction}
Set
\[
x^{\otimes m} = \underbrace{x \otimes x \otimes \dots \otimes x}_{\text{$m$ times}},
\]
where $\otimes$ denotes the Kronecker product~\cite[Section~11.4]{Handbook}, and $e = [1,\,1\,\dots,\,1]^T \in \mathbb{R}^n$, the vector of all ones. Let $R \in \mathbb{R}^{n\times n^m}$ be a non-negative matrix such that $e^T R = (e^T)^{\otimes m}$, and $v\in\mathbb{R}^n$ be a \emph{stochastic} vector, i.e., a non-negative vector with $e^T v = 1$.
The \emph{Multilinear Pagerank problem} consists in finding a stochastic solution $x\in\mathbb{R}^n$ to the equation
\begin{equation} \label{mlpr}
x = f_\alpha(x), \quad f_\alpha(x) = \alpha R(x^{\otimes m}) + (1-\alpha) v,
\end{equation}
for a certain $m\in \mathbb{N}, m \geq 2$ and a given value $\alpha \in (0,1)$.
Equation~\eqref{mlpr} is a generalization of the equation behind the well-known Pagerank model~\cite{Pageetal98} (to which it reduces for $m=1$), and has been introduced in~\cite{DBLP:journals/corr/GleichLY14} as a simplified version of a higher-order Markov chain model with memory. The equation itself has a probabilistic interpretation that was suggested in~\cite{BenGL17}.
Problem~\eqref{mlpr} can be reduced to the computation of Z-eigenvalues of tensors~\cite{lim2005singular,NgQZ09,Qi07}, so some theory and algorithms for that problem can also be applied here.
Various algorithms have been suggested to compute solutions of this equation; see e.g.~\cite{CipRT20, DBLP:journals/corr/GleichLY14,MeiP}. Among the simplest choices we have the fixed-point iteration
\[
x_{k+1} = f_\alpha(x_k), \quad k=0,1,2,\dots,
\]
or the Newton--Raphson method on the function $f_\alpha(x)-x$, i.e.,
\begin{equation} \label{newton}
x_{k+1} = x_k - (\alpha P_x-I)^{-1} (f_{\alpha}(x_k) - x_k), \quad k=0,1,2,\dots.
\end{equation}
with $\alpha P_x$ the Jacobian matrix of $f_\alpha$.
It is generally recognized that the problem is easier for small $\alpha$, especially for $\alpha \leq \frac1m$. For values of $\alpha$ approaching $1$, its numerical solution is more complicated, and the solution may be non-unique. Sufficient conditions for the uniqueness of solutions have been proposed in literature~\cite{Uniqueness,LiLNV17}; the simplest (but weakest) of them is $\alpha \leq \frac1m$.
In view of this property, various of the algorithms proposed fall in the setting of extrapolation methods, where the a sequence of solutions $x^{(1)}, x^{(2)}, x^{(3)}, \dots$ associated to increasing values of $\alpha^{(1)} < \alpha^{(2)} < \alpha^{(3)} < \alpha^{(k)}$ is computed, and used to obtain an initial guess for a further solution with parameter $\alpha^{(k+1)} > \alpha^{(k)}$. In this way, one can start by solving problems in the `easier regime' $\alpha \approx \frac1m$, and at each step use the previously computed values of $x$ to get a sufficiently accurate initial value.
However, as acknowledged in~\cite{MeiP}, this idea can fail spectacularly, because in some cases the solution of the problem is not given by a continuous function $x(\alpha)$. An example where this is well visible is obtained with the $6\times 6^2$ example \texttt{R6\_3} in the dataset of~\cite{DBLP:journals/corr/GleichLY14} for $\alpha = 0.99$: see Figure~\ref{fig:homoplot} for a plot that shows the behaviour of one entry of $x$ with respect to $\alpha$.
\begin{comment}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{homoplot}
\caption{\protect{The first entry $x_1$ of all solution(s) vs. the parameter $\alpha$ for the example \texttt{R6\_3} in~\cite{DBLP:journals/corr/GleichLY14}. There are three distinct solutions in $\mathcal{Z}_1$ for $\alpha \in $ (approximately), and one solution in the rest of $[0,1]$. This figure originally appeared in~\cite{MeiP}.}}
\label{fig:homoplot}
\end{figure}
\end{comment}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{R6_3_plot.png}
\caption{\protect{The first entry $x_1$ of all solution(s) vs. the parameter $\alpha$ for the example \texttt{R6\_3} in~\cite{DBLP:journals/corr/GleichLY14}. There are three distinct stochastic solutions for $\alpha \in [\alpha_1,\alpha_2] \approx [0.9749,\,0.9899]$, and one solution in the rest of $[0,1]$.}}
\label{fig:homoplot}
\end{figure}
When one attempts to compute solutions for increasing values of $\alpha$, they end up tracking the solution with the largest value of $x_1$ in the figure. When one first surpasses $\alpha = \alpha_2$, this solution is then worthless as an initial value, and the method basically needs to restart without a useful initial guess. In addition, according to the informal description in~\cite[Fig.~9]{DBLP:journals/corr/GleichLY14}, for $\alpha$ slightly larger than $\alpha_2$ the common algorithms are slowed down by the presence of a so-called \emph{pseudo-solution}; they spend many iterations in which the iterates wander about in the set of stochastic vectors before convergence kicks in.
In this paper, we suggest an improved strategy to deal with these problematic examples. We consider the zero set of
\[
H(x,\alpha) := f_\alpha(x) - x = \alpha R( x^{\otimes m}) + (1-\alpha) v - x, \quad H : \mathbb{R}^n \times \mathbb{R} \to \mathbb{R}^n
\]
as a curve in $\mathbb{R}^{n+1}$, and use extrapolation techniques in the family of \emph{continuation algorithms}~\cite{Allgower:2003:INC:945750} to compute points following this curve. More precisely, the algorithms we use belong to the family of~\emph{predictor-corrector} continuation algorithms.
In cases such as the one in Figure~\ref{fig:homoplot}, the curve makes an \emph{S-bend}, and we cannot consider it anymore (locally) as a parametrized curve with the coordinate $\alpha$ as a parameter: instead, we compute points following the curve, using (implicitly) an arc-length parametrization. Hence, once it reaches $\alpha_2$, our algorithm reduces the value of $\alpha$ tracking the bottom part of the curve in Figure~\ref{fig:homoplot}, reaches $\alpha_1$ again and then increases $\alpha$ a second time.
We give a theoretical contribution, proving that there is indeed a connected solution curve that allows us to track the solution correctly up to $\alpha=1$ in problems without singular points, and then we present a practical algorithm that allows us to solve multilinear Pagerank problem with higher reliability than the existing methods.
The paper is structured as follows. In Section~\ref{sec:structure}, we recall various results on the properties of solutions, focusing on uniqueness in particular. The results are stated for generic $m$, unlike other references which focus on $m=2$. In Section~\ref{sec:curve}, we focus more closely on the geometrical structure of the solution set as a curve in $\mathbb{R}^{n+1}$. We continue with a general introduction on predictor-corrector methods in Section~\ref{sec:continuation}, and then we present the specific variant of the method that we use for this problem with all its algorithmic details in Section~\ref{sec:thiscontinuation}. In Section~\ref{sec:experiments} we present numerical experiments that prove the effectiveness of this method, and we end with some conclusions in Section~\ref{sec:conclusions}.
\section{Structure of solutions} \label{sec:structure}
We recall various results on the structure of the solution set of~\eqref{mlpr}. Most of these results appear already in previous works (see e.g.~\cite{MeiP}), mostly for $m=2$, but we present them here for completeness in the case of a general $m$.
\begin{thm} \label{theorem 10}
Let $x\in \mathbb{R}^n$ a nonnegative solution of equation~\eqref{mlpr} with $\alpha \in (0;1)$, then $e^T x =1$ or $e^Tx=c_{\alpha}$, where $c_\alpha$ is the other positive solution (besides $z=1$) of the equation $g(z)=0$, with $g(z) := \alpha z^m -z +(1-\alpha)$, and satisfies
\begin{equation} \label{calpha}
\begin{cases}
c_{\alpha}>1 & \text{if } \alpha<\frac{1}{m},\\
c_{\alpha}=1 & \text{if }\alpha=\frac{1}{m},\\
c_{\alpha}<1 & \text{if } \alpha>\frac{1}{m}.\\
\end{cases}
\end{equation}
\end{thm}
\begin{proof}
Let $x$ be a nonnegative solution of~\eqref{mlpr}.
Multiplying on the left by $e^T$, we get
\[
e^T x = \alpha e^T R(x^{\otimes m}) + (1-\alpha)e^Tv = \alpha (e^T)^{\otimes m} x^{\otimes m} + (1-\alpha) = \alpha (e^T x)^m + (1-\alpha).
\]
Hence $g(e^Tx) = 0$. It remains to prove that the only positive solutions of $g(z)=0$ are $1$ and another one $c_\alpha$ satisfying~\eqref{calpha}.
By studying the sign of $g'(z)$, one sees that $g(z)$ has a global minimum
\[
z_* = \sqrt[m-1]{\frac{1}{\alpha m}},
\]
for which
\[
\begin{cases}
z_*>1 & \text{if } \alpha<\frac{1}{m},\\
z_*=1 & \text{if }\alpha=\frac{1}{m},\\
z_*<1 & \text{if } \alpha>\frac{1}{m}.\\
\end{cases}
\]
In particular, $g(z_*) \leq g(1) = 0$. The function $g(z)$ is decreasing in $(0,z_*)$ and increasing in $(z_*,\infty)$, hence (unless $z_*=1$) it has two intersections with the $x$-axis, one in $(0,z_*)$ and one in $(z_*,\infty)$.
\end{proof}
One can show that solutions with $e^Tx=1$ and $e^Tx = c_\alpha$ do indeed exist. For any $c \geq 0$, let
\begin{equation} \label{Zc}
\mathcal{Z}_c := \{x \in \mathbb{R}^n \colon e^T x = c, \, x \geq 0 \},
\end{equation}
the set of non-negative vectors with fixed entry sum $c$ (in particular, $\mathcal{Z}_1$ is the set of stochastic vectors).
\begin{thm} \label{existence}
Equation~\ref{mlpr} has at least one solution in $\mathcal{Z}_{c_{\alpha}}$, and one in $\mathcal{Z}_1$.
\end{thm}
\begin{proof}
The set $\mathcal{Z}_{c_\alpha}$ is a simplex, and in particular it is compact and convex. Moreover,
$f_\alpha(\mathcal{Z}_{c_\alpha})\subset \mathcal{Z}_{c_\alpha}$; indeed $\forall y\in\mathcal{Z}_{c_{\alpha}}$
\begin{align*}
e^T f_\alpha(y) & = \alpha e^T R (y^{\otimes m})+(1-\alpha) e^T v\\
& = \alpha (e^T y)^m + 1-\alpha\\
& = \alpha c_\alpha^m + 1-\alpha= c_\alpha,
\end{align*}
where we have used the fact that $g(c_\alpha)=0$ in the last equality.
Hence by the Brouwer fixed-point theorem~\cite{Brouwer} we can conclude that $f$ has a fixed point in $\mathcal{Z}_{c_\alpha}$, and this fixed point is a solution of~\eqref{mlpr}.
The same proof works after replacing $c_\alpha$ with $1$, since the only property that we have used is that $g(c_\alpha)=0$.
\end{proof}
Moreover, there is a unique solution which achieves the smaller among the possible two values of $e^Tx$. This unique solution is called \emph{minimal solution} in~\cite{MeiP} and other literature on similar matrix equations.
\begin{thm}
For each fixed $\alpha \in [0,1)$, Equation~\eqref{mlpr} has a unique solution in $\mathcal{Z}_{\min(1,c_\alpha)}$.
\end{thm}
\begin{proof}
Consider the iteration
\begin{equation} \label{fixedpoint}
x_0 = 0, \quad x_{k+1}=f_\alpha(x_k), \quad k=0,1,2,\dots .
\end{equation}
We can show by induction that $x_{k+1}\geq x_k$: the base step $k=0$ is obvious, and
\[
x_{k+1}-x_k= f_\alpha(x_k) - f_\alpha(x_{k-1}) = \alpha R(x_k^{\otimes m} -x_{k-1}^{\otimes m}) \geq 0.
\]
Let $y$ be a solution of~\eqref{mlpr} with $e^Ty = \min(1,c_\alpha)$.
Similarly, we can show by induction that $x_k \leq y$. The base step $k=0$ is again obvious, and
\[
y - x_{k+1} = f_\alpha(y) - f_\alpha(x_k) = \alpha R(y^{\otimes m} - x_k^{\otimes m}) \geq 0.
\]
So the sequence is weakly increasing and bounded, hence it converges to a limit $x_\infty := \lim_{k\to\infty} x_k$.
Passing to the limit, we see that $x_\infty$ is a solution of~\eqref{mlpr}, hence
\[
e^T x_\infty \geq \min(1,c_\alpha) = e^T y.
\]
Again passing to the limit, we see that $x_\infty \leq y$. These two properties together imply that $y = x_\infty$. Thus we have proved that $y=x_\infty$ for each solution $y \in \mathcal{Z}_{\min(1,c_\alpha)}$.
\end{proof}
The following observation seems to be novel.
\begin{thm}
Let $v$ be a strictly positive vector and let $0 \leq \alpha< 1$.
Any nonnegative solution of~\eqref{mlpr} is strictly positive. \label{Positive}
\end{thm}
\begin{proof}
We have
\[
x_i = e_i^T x = \alpha e_i^T R(x^{\otimes m})+(1-\alpha)e_i^T v =\alpha R_i(x^{\otimes m})+(1-\alpha)v_i > 0,
\]
since $\alpha R_i(x^{\otimes m})$ and $(1-\alpha)v_i > 0$.
\end{proof}
\begin{cor} \label{cor:solsincube}
In the previous setting, $x\in (0,1)^n$.
\begin{proof}
Since every component of $x$ is strictly positive and the vector is stochastic every component must also be less then 1.
\end{proof}
\end{cor}
\section{The curve of solutions} \label{sec:curve}
In this section, we argue that the stochastic solutions of~\eqref{mlpr} form a smooth curve (under some regularity assumptions).
We denote with $P_x \in \mathbb{R}^{n\times n}$ the Jacobian matrix of $Rx^{\otimes m}$ with respect to $x$, i.e.,
\begin{align} \label{Px}
P_x = R\frac{\partial x^{\otimes m}}{\partial x} = R(I \otimes x \otimes \dots \otimes x + x \otimes I \otimes x \otimes \dots \otimes x + \dots + x \otimes \dots \otimes x \otimes I).
\end{align}
The Jacobian matrix of $H(x,\alpha)$ is
\begin{equation} \label{Jac}
J_H[x,\alpha] = \begin{bmatrix}
\alpha P_x - I & R(x^{\otimes m}) - v
\end{bmatrix} \in \mathbb{R}^{n\times (n+1)},
\end{equation}
where the first $n\times n$ block column $\frac{\partial H}{\partial x}[x,\alpha] = P_x - I$ contains derivatives with respect to $x$ and the second contains derivatives with respect to $\alpha$.
We let
\[
\Delta_c := \{x \in \mathbb{R}^n : e^Tx = c\}.
\]
Note that this definition differs from that of $\mathcal{Z}_c$ in that we do not require that $x\geq 0$.
We start from a result proving the surjectivity of $\frac{\partial H}{\partial x}[x,\alpha]$ under some assumptions.
\begin{lemma}
Let $(x,\alpha)$ be such that $H(x,\alpha) = 0$, and $x \in \mathcal{Z}_1$.
\begin{enumerate}
\item For $\alpha < \frac1m$, the matrix $\frac{\partial H}{\partial x}[x,\alpha] = \alpha P_x - I$ has full rank $n$.
\item For $\alpha = \frac1m$, the matrix $\frac{\partial H}{\partial x}[x,\alpha] = \alpha P_x - I$ has rank $n-1$, and $e^T$ is the unique row vector (up to multiples) such that $e^T \frac{\partial H}{\partial x}[x,\alpha]$.
\item For $\alpha = \frac1m$, the matrix $\frac{\partial H}{\partial x}[x,\alpha] = \alpha P_x - I$ is surjective as a linear map from $\Delta_0$ to $\Delta_0$.
\end{enumerate}
\end{lemma}
\begin{proof}
We prove the three points.
\begin{enumerate}
\item Note that the matrix $P_x$ is nonnegative, and that
\begin{align}
e^T P_x & = (e^T)^{\otimes m} (I \otimes x \otimes \dots \otimes x + x \otimes I \otimes x \otimes \dots \otimes x + \dots + x \otimes \dots \otimes x \otimes I) \nonumber \\
&= e^T \otimes e^Tx \otimes \dots \otimes e^Tx + e^Tx \otimes e^T \otimes e^Tx \otimes \dots \otimes e^Tx + \dots + e^Tx \otimes \dots \otimes e^Tx \otimes e^T \nonumber \\
& = m(e^Tx)^{m-1} e^T. \label{Pxstochasticity}
\end{align}
Hence $e^T \alpha P_x < e^T$ when $\alpha < \frac1m$ and $e^Tx=1$. Thus $\alpha P_x$ is a sub-stochastic matrix, and $\alpha P_x - I$ is invertible by the Perron-Frobenius theorem \cite[Section~10.4]{Handbook}.
\item Again thanks to the relation~\eqref{Pxstochasticity}, the matrix $\alpha P_x$ is (column-)stochastic when $\alpha = \frac1m$ and $e^Tx=1$; hence $e^T (\alpha P_x - I) = 0$. It remains to prove that the rank of $\frac{1}{m} P_x - I$ is exactly $n-1$. Suppose, on the contrary, that it is $n-2$ or lower. Hence the eigenvalue 1 has geometric multiplicity at least $2$ in the column-stochastic matrix $\frac{1}{m} P_x$. It follows from the Perron-Frobenius theory of stochastic matrices (see e.g. \cite[Section~10.4, Fact 1(g)]{Handbook}) that $\frac{1}{m} P_x$ has at least two distinct ergodic classes: that is, there exist two disjoint subsets of indices $E_1, E_2$ such that one can write (reordering the matrix entries and setting $E_3 = (E_1 \cup E_2)^c)$
\[
P_x =
\begin{blockarray}{cccc}
& E_1 & E_2 & E_3\\
\begin{block}{c[ccc]}
E_1 & P_{E_1E_1} & 0 & P_{E_1E_3}\\
E_2 & 0 & P_{E_2E_2} & P_{E_2E_3}\\
E_3 & 0 & 0 & P_{E_3E_3}\\
\end{block}
\end{blockarray},
\]
with $\frac{1}{m} P_{E_1E_1}$ and $\frac{1}{m} P_{E_2E_2}$ irreducible and stochastic.
Expanding the formula~\eqref{Px}, one sees that the entries of $P_x$ are computed according to
\begin{multline*}
P_{i,j} = \sum_{k_1,k_2,k_{m-1}=1}^n \left(R_{i, (j,k_1,k_2,\dots, k_{m-1})} + R_{i, (k_1,j,k_2,\dots, k_{m-1})} + \dots \right. \\
\left. + R_{i, (k_1,k_2,\dots, k_{m-1},j)} \right) x_{k_1}x_{k_2}\dotsm x_{k_{m-1}},
\end{multline*}
where we have used tuples $(j_1,j_2,\dots,j_m) \in \{1,2,\dots,n\}^m$ as indices for the columns of $R$. Let us consider a tuple $(j_1,j_2,\dots,j_m)$ such that $j_1\in E_1$ and $j_2\in E_2$. Since $x>0$ and $R \geq 0$, we must have
$R_{i,(j_1,j_2,\dots,j_m)}=0$ for each $i\in E_1$, because otherwise the entry $P_{i,j_2}$ in $P_{E_1E_2}$ would be strictly positive. Analogously, $R_{i,(j_1,j_2,\dots,j_m)}=0$ for each $i \in E_1 \cup E_3$, since otherwise the entry $P_{i,j_1}$ in $P_{E_2E_1}$ or $P_{E_3E_1}$ would be strictly positive. It follows that $R_{i,(j_1,j_2,\dots,j_m)}=0$ for all $i$. However, this implies
\[
(e^T R)_{(j_1,j_2,\dots,j_m)} = \sum_{i=1}^n R_{i,(j_1,j_2,\dots,j_m)} = 0,
\]
which contradicts $e^T R = (e^T)^{\otimes m}$.
\item It is sufficient to prove that the map $w \mapsto (\alpha P_x - I)w$ has trivial kernel in $\Delta_0$, i.e., $(\alpha P_x - I)w \neq 0$ for all $w \in \Delta_0 \setminus \{0\}$.
By the Perron--Frobenius theorem, there is a nonnegative vector $z\neq 0$ such that $\alpha P_x z = z$, i.e., $(\alpha P_x - I)z = 0$. Since we know that $\alpha P_x - I$ has rank $n-1$, it follows that $\ker (\alpha P_x - I) = \operatorname{span}(z)$. Since $z$ is nonnegative, $\operatorname{span}(z) \cap \Delta_0 = \{0\}$, hence $\alpha P_x - I$ has trivial kernel in $\Delta_0$. \qedhere
\end{enumerate}
\end{proof}
\begin{thm} \label{thm:curve}
Let $\mathcal{S}=\left\{ (x,\alpha) \mid H(x,\alpha) =0,\, x\in \mathcal{Z}_1\right\}$ and $\mathcal{S}_{I}=\left\{(x,\alpha)\in \mathcal{S} \mid \alpha \in I\right\}$.
\begin{enumerate}
\item $\mathcal{S}_{[0,\frac{1}{m}]}$ is a smooth curve (with boundary).
\end{enumerate}
If moreover $J_H[x,\alpha]$ is surjective as a linear map from $\Delta_0 \times \mathbb{R}$ to $\Delta_0$ for each point $(x,\alpha) \in \mathcal{S}$ with $\alpha \leq 1$, then we have that
\begin{enumerate}[resume]
\item For some $\varepsilon > 0$,
$\mathcal{S}_{[-\varepsilon,1+\varepsilon)}$ is union of disjoint smooth curves (with boundary).
\item The curve $\mathcal{S}^*$ containing $\mathcal{S}_{[0,\frac{1}{m}]}$ reaches the hyperplane $\{\alpha=1\}$.
\end{enumerate}
\end{thm}
\begin{proof}
We prove the three points.
\begin{enumerate}
\item From lemma 1.1 we know that $\frac{\partial H}{\partial x}[x,\alpha]$ is surjective for $0\leq\alpha\leq\frac{1}{m}$; since surjectivity of the Jacobian is an open condition, there exists $\varepsilon > 0$ such that $J_H[x,\alpha]$ is surjective for each point $(x,\alpha)$ with $-\varepsilon <\alpha < 1+\varepsilon$. \\
This allows us to apply the Regular Level Set Theorem~\cite[Section~9.3]{Manifold} to
$$H_{\big|\Delta_1\times \left(-\varepsilon,\frac1m+\varepsilon\right)}:\Delta_1\times \left(-\varepsilon,\frac1m+\varepsilon\right)\to \Delta_0,$$ obtaining that $H_{\big|\Delta_1\times \left[0,\frac1m\right]}^{-1}(0)$ is union of smooth curves with boundary.
By theorem 3, for each $\alpha\in \left[0,\frac{1}{m}\right]$ there is only one solution, hence there is only one curve in the union and it coincides with $\mathcal{S}_{[0,\frac1m]}$.
\item We argue as in the previous point: since surjectivity of the Jacobian is an open condition, there exists $\varepsilon > 0$ such that $J_H[x,\alpha]$ is surjective for each point $(x,\alpha)$ with $\alpha < 1+\varepsilon$. We can now apply the Regular Level Set Theorem to the whole $\Delta_1\times \left[0,1+\varepsilon\right)$.
\item Let $\mathcal{C}$ denote the hypercube $\left\{(x,\alpha)\in\mathbb{R}^n\times [0,1] \hspace{1mm}\big|\hspace{1mm} 0\leq x_i\leq 1 \right\}$. The intersection $\mathcal{C} \cap \mathcal{S}^* $ is a connected compact 1-manifold with boundary $\partial C \cap \mathcal{S}^*$; such a manifold is diffeomorphic to $\mathbb{S}^1$ or to a closed segment.
It cannot be diffeomorphic to $\mathbb{S}^1$ since otherwise there would be two different branches of $\mathcal{S}^*$ starting from $(v,0)$ (the solution at $\alpha = 0$), contradicting the uniqueness of the solution for $\alpha < \frac1m$.
Being diffeomorphic to a segment, the intersection has two boundary points that must have $\alpha = 0$ or $\alpha = 1$ by Corollary~\ref{cor:solsincube}. Since there is only one solution with $\alpha = 0$, there must be one with $\alpha = 1$. \qedhere
\begin{comment}
Note that its codomain of $H$ is effectively $\Delta_0$, indeed
\begin{align*}
e^T H(x,\alpha) &= \alpha e^T R (x^{\otimes m}) + e^T(1-\alpha)v - e^T x \\
&= \alpha (e^T)^{\otimes m} x^{\otimes m}+ (1-\alpha) -e^T x
\\ &= \alpha (e^T x)^m +(1-\alpha)e^T v -(e^T x) = 0.
\end{align*}
\end{comment}
\end{enumerate}
\end{proof}
\begin{rem}
We could not find an explicit example in which $J_H[x,\alpha]$ is not surjective as a linear map from $\Delta_0 \times \mathbb{R}$ to $\Delta_0$ for each point $(x,\alpha) \in \mathcal{S}$ with $\alpha \leq 1$. We conjecture that this condition always hold, but we do not have a proof.
\end{rem}
\section{Numerical continuation methods} \label{sec:continuation}
Numerical continuation methods~\cite{Allgower:2003:INC:945750} are a family of numerical methods to trace solution curves defined as the solution set of a system of nonlinear equations of codimension 1, i.e., $G(y)=0$, where $G: \mathbb{R}^{n+1}\to \mathbb{R}^n$.
These methods are often used to compute solutions in a precise region of the curve following the implicit defined curve numerically: one computes an initial solution in a part of the curve where it is easier, and then uses it as an initial step to compute iteratively a sequence of points on the curve, each close to the previous one, with the goal of arriving to a solution in a different part of the curve.
We focus on prediction-correction methods~\cite[Chapter~2]{Allgower:2003:INC:945750}, which are a class of continuation methods structured in two parts:
\begin{enumerate}
\item (the \emph{predictor step}) Given a starting point $y_k$ on the curve and a step-size $h_k$, we compute a new point $\hat{y}_k$ which is close to the curve (for instance, on its tangent line) and at distance $h_k$ from $y_k$.
\item (the \emph{corrector step}) We seek a point $y_{k+1}$ on the curve near $\hat{y}_k$ through an iterative method.
\begin{figure}[h]
\centering
\begin{tikzpicture}
\draw [thick]
plot [smooth] coordinates { (-2.2,0.99) (-1.1, 1.6) (-0.2,1.46) (1,1.2) (2.1,1.51) (3,2)}
(-2.2,0.97) node[blue] {\textbullet};
\draw [<-] (-0.38,1.58) -- (-0.22,2.11) [blue, line width=0.2mm];
\draw [<-] (2,1.4) -- (2.09,0.99) [blue, line width=0.2mm];
\draw
(-2.4,1.2) node{$y_k$};
\draw plot (-2.2,1) -- (-0.2,2.2)[dashed,red];
\draw (-0.4,1.51) -- (2.1,0.9)[dashed,red];
\draw (-0.2,2.17) node[red] {\textbullet}
(-0.4,1.49) node[blue] {\textbullet}
(-0.4,1.29) node{$y_{k+1}$}
(-0.2,2.5) node{$\hat{y}_k$};
\draw
(2.1,0.89) node[red] {\textbullet}
(2.0,1.46) node[blue] {\textbullet}
(1.8,1.73) node{$y_{k+2}$}
(2.2,0.65) node{$\hat{y}_{k+1}$};
\end{tikzpicture}
\caption{Illustration of predictor step (red dashed lines) and corrector step (blue lines).}
\end{figure}
\end{enumerate}
In order to start the iteration, it is necessary to compute an initial point $y_0$ such that $G(y_0)=0$ and then to alternate the two steps above.
A first example is the strategy suggested in~\cite{MeiP}, which can be interpreted in this framework by setting
\[
G:\begin{bmatrix}
x \\
\alpha
\end{bmatrix}
\mapsto H(x,\alpha).
\]
As initial step, we choose an initial value $\alpha^{(0)}$ that is close to $\frac1m$, and compute a corresponding $x^{(0)}$ such that $H(x^{(0)},\alpha^{(0)}) = 0$ via either the fixed-point iteration~\eqref{fixedpoint}, or a more efficient method (for instance, the Newton method). Then, at each step $k$,
\begin{enumerate}
\item (predictor) we choose $\alpha^{(k)} = \alpha^{(k-1)}+h_k$, for a certain step-size $h_k$ (hence $\alpha^{(0)} < \alpha^{(1)} < \dots < \alpha^{(k-1)} < \alpha^{(k)}$ always form an increasing sequence), and use $x^{(0)}, \dots, x^{(k-1)}$ to compute a vector $\hat{x}$ such that $H(\hat{x}, \alpha^{(k)}) \approx 0$. Several variants were tested~\cite{MeiP}, for instance, linear extrapolation, or a Taylor expansion around $x^{(k-1)}$.
\item (corrector) we use $\hat{x}$ as an initial value for a fixed-point iteration such as~\eqref{fixedpoint} or the Newton method to compute $x^{(k)}$ such that $H(x^{(k)}, \alpha^{(k)})$.
\end{enumerate}
In this setting, at each step we fix a value of $\alpha^{(k)}$ arbitrarily, and then compute the corresponding $x^{(k)}$. Hence essentially we use the variable $\alpha$ to parametrize the curve $\mathcal{S}^*$. This strategy works well in most cases, but runs into trouble when the curve can not be written locally as the graph of a function $x(\alpha)$, i.e., when $\frac{\partial H(x,\alpha)}{\partial x}$ is singular. Figure~\ref{fig:homoplot} is a typical example of this pathological behavior.
Hence, we replace the two steps with more general versions that arise from considering $\mathcal{S}^*$ as a generic curve in $\mathbb{R}^{n+1}$, without singling out $\alpha$ as a preferred parametrization coordinate.
\section{Continuation methods for this problem} \label{sec:thiscontinuation}
We now present the Predictor-Corrector-Newton method, the actual algorithm that we use to solve the Multilinear PageRank problem.
\begin{comment}
\begin{enumerate}
\item (the \emph{predictor step}) We choose a new pair $(\hat{x}, \hat{\alpha}) = (x^{(k-1)} + \delta x, \alpha^{(k-1)} + \delta \alpha)$ such that $H(\hat{x}, \hat{\alpha}) \approx 0$, by taking a point on the tangent line to the solution curve $\mathcal{S}^*$ in $(x^{(k-1)}, \alpha^{(k-1)})$ at a suitable distance, i.e.,
\[
\norm*{\begin{bmatrix}
\delta x\\
\delta \alpha
\end{bmatrix}} = h,
\]
for a suitable step-length $h$.
We need to make sure that we move on the tangent line in the correct direction to go towards new portions of $\mathcal{S}^*$: to do this, we impose
\[
\left\langle
\begin{bmatrix}
x^{(k-2)}\\
\alpha^{(k-2)}
\end{bmatrix}
-
\begin{bmatrix}
x^{(k-1)}\\
\alpha^{(k-1)}
\end{bmatrix}
,\,
\begin{bmatrix}
\delta x\\
\delta \alpha
\end{bmatrix}
\right\rangle \geq 0.
\]
\item (the \emph{corrector step}) We seek a point on the curve near $(\hat{x}, \hat{\alpha})$, i.e., a solution of $H(x,\alpha) = 0$, with a version of Newton's method for undetermined nonlinear systems~\cite[Section~3.3]{Allgower:2003:INC:945750}
\begin{equation} \label{undernewton}
\begin{bmatrix}
x_{k+1}\\
\alpha_{k+1}
\end{bmatrix}
=
\begin{bmatrix}
x_{k}\\
\alpha_{k}
\end{bmatrix}
-
JH[x_k,\alpha_k]^+ H(x_k,\alpha_k), \quad
\begin{bmatrix}
x_{0}\\
\alpha_{0}
\end{bmatrix}
=
\begin{bmatrix}
\hat{x}\\
\hat{\alpha}
\end{bmatrix}.
\end{equation}
Here $JH[x_k,\alpha_k]^+$ denotes the Moore--Penrose pseudo-inverse of the matrix $JH[x_k,\alpha_k] \in \mathbb{R}^{n\times (n+1)}$. This iteration, essentially, is obtained by building a first-order Taylor expansion $L(x,\alpha)$ of the function $H(x,\alpha)$ in
$\begin{bsmallmatrix}
x_{k}\\
\alpha_{k}
\end{bsmallmatrix}$, and taking as
$\begin{bsmallmatrix}
x_{k+1}\\
\alpha_{k+1}
\end{bsmallmatrix}$
the minimum-norm solution of $L(x,\alpha) = 0$.
\end{enumerate}
\end{comment}
\subsection{Predictor step}
Following~\cite{Allgower:2003:INC:945750}, we use the tangent line to $\mathcal{S}^*$ as a predictor. Given a starting point $(x^{(k)},\alpha^{(k)})$ and a step size $\tau_k$, we look for a point $(\hat{x}, \hat{\alpha})$ on the tangent line to the solution curve in $(x^{(k)},\alpha^{(k)})$ at distance $\tau_k$ from the starting point.
By the implicit function theorem (see e.g.~\cite[Appendix~B]{Manifold}), the tangent direction is the kernel of the Jacobian matrix $JH[x^{(k)},\alpha^{(k)}]$.
We compute it through a $QR$ decomposition
\[
JH[x^{(k)},\alpha^{(k)}]^T = [Q\quad q_{n+1}] \begin{bmatrix} R\\ 0 \end{bmatrix}.
\]
Notice that $q_{n+1}\in \mathbb{R}^n$ spans the kernel of $JH[x^{(k)},\alpha^{(k)}]$ (assuming that $JH[x^{(k)},\alpha^{(k)}]$ has full rank).
We also need to make sure that we move on the tangent line in the correct direction to go towards new portions of the curve: to do this, we change the sign of $q_{n+1}$, if needed, so that
\[
\left\langle
\begin{bmatrix}
x^{(k-1})\\
\alpha^{(k-1)}
\end{bmatrix}
-
\begin{bmatrix}
x^{(k)}\\
\alpha^{(k)}
\end{bmatrix}
,\,
q_{n+1}
\right\rangle \geq 0.
\]
Then we can set
\[
\begin{bmatrix}
\hat{x}\\
\hat{\alpha}
\end{bmatrix}
=
\begin{bmatrix}
x^{(k)}\\
\alpha^{(k)}
\end{bmatrix} + q_{n+1} \tau_k.
\]
We point out that $QR$ factorization can be prohibitively expensive when $n$ is large; in this case, it may be preferable to replace this strategy with linear extrapolation, that is,
\begin{equation}
(\hat{x}, \hat{\alpha}) = (x^{(k)},\alpha^{(k)}) + \tau_k \frac{(x^{(k)},\alpha^{(k)})-(x^{(k-1)},\alpha^{(k-1)})}{\norm{(x^{(k)},\alpha^{(k)})-(x^{(k-1)},\alpha^{(k-1)})}}.
\end{equation}
\subsection{Corrector step}
The bulk of our algorithm is the corrector step, where we use Newton's method for underdetermined systems
\begin{equation} \label{pcnewton}
(x_{j+1},\alpha_{j+1}) = (x_j,\alpha_j)-JH[x_j,\alpha_j]^+H(x_j,\alpha_j), \quad j = 0,1,\dots
\end{equation} to compute a solution of $H(x,\alpha)=0$ close to $(\hat{x},\hat{\alpha})$.
Every iteration requires the computation of $JH[x_j,\alpha_j]^+$, which we compute, again, with QR factorization: if $JH[x_j,\alpha_j]^T = [Q \quad q_{n+1}] \begin{bmatrix}
R\\
0
\end{bmatrix}$, then
\[
JH[x_j,\alpha_j]^+ = [Q \quad q_{n+1}] \begin{bmatrix} R^{-T} \\ 0 \end{bmatrix}.
\]
\subsection{Choice of step size}
We choose an adaptive step size $\tau_k$ with a simplified version of the approach in~\cite[Section~6.1]{Allgower:2003:INC:945750}.
Let $\hat{\delta}$ be the distance between the predicted value $\hat{y}$ and the curve $\mathcal{S}^*$. This distance is approximately equal to the square of the Newton correction at the \emph{first} corrector step $j=1$, up to second-order terms, and it can also be shown to be proportional to the square of the predictor step-length, $\tau_k^2$, i.e.,
\begin{equation} \label{Crelation}
\hat{\delta} \approx \norm{JH[x_1,\alpha_1]^+ H(x_1,\alpha_1)}_1 \approx C \tau_k^2
\end{equation}
for a given constant $C$.
We would like this distance to keep close to a prescribed ``nominal distance'' $\delta$ at each iteration. To this purpose, we compute $C$ from~\eqref{Crelation} (replacing the rightmost $\approx$ with an equal sign), and then choose the next step-size $\tau_{k+1}$ so that $C\tau_{k+1}^2 = \delta$. In addition, some safeguards are taken so that the step-size does not change by more than a factor 2 at each step; see Algorithm~\ref{algo:pc} in the following for details.
\subsection{Final step}
Since the solution curve $\mathcal{S}^*$ reaches $\alpha=1$ (Theorem~\ref{thm:curve}), if the continuation method is successful it will eventually obtain an iteration where $\alpha^{(k)} < \alpha \leq \alpha^{(k+1)}$, where $\alpha > \alpha^{(0)}$ is the desired target value for which we wish to compute a solution $x$ to~\eqref{mlpr}.
When this happens, we have not reached yet our goal of computing a solution $x$ to~\eqref{mlpr} with the desired target value of the parameter $\alpha$, but we only have two nearby points $(x^{(k)}, \alpha^{(k)})$ and $(x^{(k+1)}, \alpha^{(k+1)})$ on the solution curve $\mathcal{S}^*$. As a final step, we compute $\hat{x}\in\mathbb{R}^n$ such that $(\hat{x}, \alpha)$ is on the segment between these two points (linear interpolation), and use it as the starting point for a final round of the Newton--Raphson method~\eqref{newton} with fixed $\alpha$. We expect this final round to have very fast convergence, since we have identified a suitable starting point near to the solution curve.
Full pseudocode for the algorithm is presented as Algorithm~\ref{algo:pc}.
\begin{algorithm}
\caption{Predictor-Corrector-Newton method} \label{algo:pc}
\vspace{2mm}
\textbf{Parameters} \hspace{2mm} $tol$ \hfill tolerance;\\
\phantom{\textbf{Parameters}} \hspace{2mm} $\alpha_0$ \hfill initial $\alpha$ for curve-following;\\
\phantom{\textbf{Parameters}} \hspace{2mm} $\tau_0$\hfill initial step-size (adaptive);\\
\phantom{\textbf{Parameters}} \hspace{2mm} $\delta $ \hfill nominal distance to curve.\\
\\
\textbf{Subroutines} \hspace{0mm} $\textsf{newton}(x_0,\alpha)$ \hfill Newton's method~\eqref{newton} with fixed $\alpha$. \\
\textbf{Input} \hspace{12mm} $R$, $\alpha$, $v$ \hfill as described above. \\
\textbf{Output} \hspace{9mm} $x$ \hfill a stochastic solution to \eqref{mlpr}.\\
\\
\phantom{\textbf{compute}} \hspace{6mm} $x_0 \leftarrow \textsf{newton}(v, \alpha_0)$\\
\phantom{\textbf{Put}} \hspace{16mm}$y_0\leftarrow(x_0,\alpha_0)$;\\
\phantom{\textbf{Put}} \hspace{16mm}$t_0 \leftarrow e_{n+1}$; \hfill initial direction for curve-following;\\
\phantom{\textbf{Put}} \hspace{16mm}$k \leftarrow 0$; \\%\hfill number of outer iterations;\\
\phantom{while while wr} \textbf{while} $\alpha_k<\alpha$ \hfill \phantom{a}\\
\phantom{while}\hspace{20mm} $t_{k+1}\leftarrow $ kernel of $JH[y_k]$ (normalized s.t. $\norm{t_{k+1}}=1$);\\
\phantom{while}\hspace{20mm} \textbf{if} $\langle t_k, t_{k+1} \rangle<0$\\
\phantom{while}\hspace{25mm} $t_{k+1}\leftarrow -t_{k+1}$; \hfill ensures $t_{k+1}$ ``points forward'';\\
\phantom{while}\hspace{20mm} \textbf{end}\\
\phantom{while}\hspace{20mm} $\hat{y}_{k+1} \leftarrow y_k+t_{k+1}\tau_k$; \hfill predictor step;\\
\phantom{while}\hspace{20mm} $(\hat{x}_{k+1}, \hat{\alpha}_{k+1}) \leftarrow \hat{y}_{k+1}$;\\
\phantom{while}\hspace{20mm} $j \leftarrow 0$;\\
\phantom{while}\hspace{20mm} \textbf{while} $\norm{H(\hat{y}_{k+1})}_1>tol$ \hfill corrector loop;\\
\phantom{while}\hspace{25mm} $j \leftarrow j+1;$\\
\phantom{while}\hspace{25mm} $d\leftarrow -JH[\hat{y}_{k+1}]^+H(\hat{y}_{k+1})$;\\
\phantom{while}\hspace{25mm} \textbf{if} $j=1$\\
\phantom{while}\hspace{29mm} $f\leftarrow \sqrt{\frac{\norm{d}_1}{\delta}}$; \hfill deceleration factor as in \cite[Sec.~6.1]{Allgower:2003:INC:945750};\\
\phantom{while}\hspace{29mm} \textbf{if} $f>2$ \hfill predictor step too large; \\
\phantom{while}\hspace{33mm} $\tau_k\leftarrow \frac{\tau_k}{2}$;\\
\phantom{while}\hspace{33mm} \textbf{back} to predictor step; \\
\phantom{while}\hspace{29mm} \textbf{end}\\
\phantom{while}\hspace{25mm} \textbf{end} \\ \phantom{while}\hspace{25mm} $\hat{y}_{k+1}\leftarrow \hat{y}_{k+1}+d; $\\
\phantom{while}\hspace{25mm} \textbf{end}\\
\phantom{while}\hspace{20mm} \textbf{end}\\
\phantom{while}\hspace{20mm} $y_{k+1}\leftarrow \hat{y}_{k+1}$;\\
\phantom{while}\hspace{20mm} $(x_{k+1}, \alpha_{k+1})\leftarrow y_{k+1}$;\\
\phantom{while}\hspace{20mm} $f\leftarrow \max{(f, \frac12)}$;\\
\phantom{while}\hspace{20mm} $\tau_{k+1}\leftarrow \frac{\tau_k}{f}$; \hfill choose next step-size;\\
\phantom{while}\hspace{20mm} $\tau_{k+1}\leftarrow \min{(\tau_{k+1},5 \tau_0)}$;\\
\phantom{while}\hspace{20mm} $k \leftarrow k+1$;\\
\phantom{while while wr} \textbf{end}\\
\phantom{while while wr} $\eta \leftarrow \frac{\alpha-\alpha_{k-1}}{\alpha_k-\alpha_{k-1}}$; \hfill linear interpolation;\\
\phantom{while while wr} $\hat{y}\leftarrow y_{k-1}+\eta(y_k-y_{k-1}); $\\
\phantom{while while wr} $(\hat{x},\hat{\alpha}) \leftarrow \hat{y}$; \hfill it must hold that $\hat{\alpha}=\alpha$;\\
\phantom{while while wr} $x\leftarrow \textsf{newton}(\hat{x}, \alpha)$.
\vspace{2mm}
\end{algorithm}
\newpage
\section{Numerical experiments} \label{sec:experiments}
We compare the Predictor-Corrector-Newton algorithm (PC-N, Algorithm~\ref{algo:pc}) with two different methods: the Newton method (N) with $x_0=(1-\alpha)v$ and the Perron-Newton method (PN-IMP). These algorithms are implemented as described in ~\cite{DBLP:journals/corr/GleichLY14} and ~\cite{MeiP} respectively, and have been chosen because they were the better performers among the algorithms considered in those two papers.
We investigate their performance on two sets of problems: the benchmark set used in ~\cite{DBLP:journals/corr/GleichLY14}, and a set of $100\,000$ random matrices $R\in\mathbb{R}^{5 \times 5^2}$. Each of the latter ones has been created starting from the zero matrix and then, for each $j$, setting $R_{ij}=1$, where $i\in \{1,2,3,4,5\}$ is chosen at random (uniformly and independently).
To better compare iteration counts for algorithms that are based on nested iterations, in the following we define one `iteration' to be an inner iteration, that is, anything that requires a matrix factorization on a dense linear algebra subproblem:
\begin{center}
\begin{tabular}{ll}
\toprule
Algorithm & What is counted as one iteration\\
\midrule
N & a Newton iteration; \\
PN-IMP & a Newton or a Perron-Newton iteration;\\
PC-N & a Newton iteration, or a predictor step, or a corrector step.\\
\bottomrule
\end{tabular}
\end{center}
In all the experiments we have set the tolerance $tol=\sqrt{\mathbf{u}}$, where $\mathbf{u}$ is the machine precision, the maximum number of iterations $maxit=10\,000$ and $v=\frac{1}{n}e$. In PC-N, we used $\tau_0=0.01$ and $\delta = 0.1$.
The experiments have been run on Matlab R2020b on a computer equipped with a 64-bit Intel core i7-4700MQ processor.
Note that neither the number of iterations nor the CPU time are particularly indicative, alone, of the true performance to expect from the algorithms on large tensors: indeed, the iterations in the different methods amount to different quantities of work, and the CPU time may scale differently for each algorithm when one switches to linear algebra primitives better suited to large-scale problems. Rather, we focus here on the reliability of the algorithms, i.e., the number of problems that they can solve successfully.
In the first experiment, we ran each algorithm on every tensor of the benchmark set, checking the number of iterations and the CPU times in seconds required for the solution of the problem. Performance profiles (see e.g.~\cite[Section~22.4]{matlabguide} for an introduction to this type of plot) are reported in Figures~\ref{fig:perfprof1}--\ref{fig:perfprof3}.
\begin{figure}
\includegraphics[width=12cm]{gleich_performance_plot_090.png}
\caption{Performance profiles for the 29 benchmark tensors and $\alpha= 0.90$.} \label{fig:perfprof1}
\end{figure}
\begin{figure}
\includegraphics[width=12cm]{gleich_performance_plot_095.png}
\caption{Performance profiles for the 29 benchmark tensors and $\alpha= 0.95$.}
\end{figure}
\begin{figure}
\includegraphics[width=12cm]{gleich_performance_plot_099.png}
\caption{Performance profiles for the 29 benchmark tensors and $\alpha= 0.99$.} \label{fig:perfprof3}
\end{figure}
The performance profile shows that the Newton method is always the fastest method, when it works, but it fails on various problems (6 out of 29 for $\alpha=0.99$). The new method is about a factor 5 slower, but has much higher reliability, failing on none of the problems. PN-IMP fails on one of the problems, and it is slower in terms of CPU time: this is to be expected, since eigenvalue computations are generally slower than QR factorizations and linear system solutions.
In the second experiment, we counted the number of failures of each of the three algorithms for different values of $\alpha$. The results are reported in Table~\ref{tab:failures}.
\begin{table}
\centering
\begin{tabular}{cccc}
\toprule
$\alpha$ & N & PN-IMP & PC-N \\
\midrule
0.90 & 102 & 22 & 0 \\
0.95 & 317 & 47 & 0 \\
0.99 & 693 & 136 & 0 \\
\bottomrule
\end{tabular}
\caption{Number of failures recorded on 100k random sparse tensors of size $5$.} \label{tab:failures}
\end{table}
\section{Conclusions} \label{sec:conclusions}
The experimental results confirm that the suggested method, based on the combination of predictor-corrector continuation methods and Newton's method for underdetermined systems, is an effective and reliable method to solve multilinear Pagerank problems, even in cases that are problematic for most other methods. The theoretical analysis performed confirms that the solution curve always reaches a valid solution, at least in the case where the solution curve has no singular points. It remains to consider how this method scales to larger problems.
\paragraph{Acknowledgments} The authors are grateful to Beatrice Meini for several useful discussions on this topic.
|
\section{introduction}\label{Sec:1}
Let $d\ge 1$. We consider the wave operator
\[
e^{it\sqrt{\mathsmaller{-}\Delta}} f(x) = \frac{1}{(2\pi)^d} \int_{\mathbb R^d} e^{i (x\cdot \xi + t |\xi|)} \widehat f(\xi) d\xi.
\]
In this note we are mainly concerned with the pointwise behavior of $e^{it\sqrt{\mathsmaller{-}\Delta}} f$ as $t\to 0$, from which we can deduce the
pointwise convergence to the initial data of the solution $u$ to the Cauchy problem:
\[
\partial_t^2 u - \Delta u =0 , \quad u(\cdot,0)= u_0, \quad \partial_t u(\cdot,0)= u_1,
\]
where $(u_0,u_1) \in H^s(\mathbb R^d) \times H^{s-1}(\mathbb R^d) $.
Here, $ H^s(\mathbb R^d)$ denotes the inhomogeneous Sobolev space of order $s$ which is equipped with the norm $\|f\|_{ H^s} = \| (1+|\cdot|^2)^{\frac s 2} \widehat f\|_{L^2} $.
It is well-known that $e^{it\sqrt{\mathsmaller{-}\Delta}} f$ converges to $f$ almost everywhere as $t\rightarrow 0$ for all $f \in H^s(\mathbb R^d)$ if $s>1/2$ while the convergence fails in general if $s\le \frac12$.
The convergence follows from the maximal estimate
\begin{equation}\label{maximal}
\| \sup_{0<t<1} |e^{it\sqrt{\mathsmaller{-}\Delta}} f |\|_{L^2(\mathbb R^d)} \lesssim \|f\|_{ H^s(\mathbb R^d) }
\end{equation}
for $s>1/2$, which is an easy consequence of the Sobolev imbedding and Plancherel's theorem. (See for exmaple \cite{Cowling}.)
The estimate fails for $s\le 1/2$ (see \cite{Walther}) and almost everywhere convergenece also fails for $s\le 1/2$. This can be shown using Stein's maximal theorem \cite{Stein2}, by which almost everywhere convergence for all $f\in H^{s}$ implies $H^{s}$--$L^{2,\infty}$ bound on the maximal operator $\sup_{0<t<1} |e^{it\sqrt{\mathsmaller{-}\Delta}} f |$.
On the other hand, it is not difficult to see the $H^{1/2}$--$L^{2,\infty}$ estimate fails by a construction (see Lemma \ref{weak-type} in Appendix).
We study a more refined pointwise behavior
of $e^{it\sqrt{\mathsmaller{-}\Delta}} f$ by considering the Hausdorff dimension of the divergence set
\[
\mathfrak D_d (f) : = \{x: e^{it\sqrt{\mathsmaller{-}\Delta}} f(x) \not\to f(x) ~\text{as}~ t\rightarrow 0 \}.
\]
Dimension of $\mathfrak D_d (f)$ naturally relates to the regularity exponent $s$ for which $f\in H^s$. Almost everywhere convergence only tells the divergence set is of Lebesgue measure zero.
However, if the initial datum $f$ possesses an additional regularity, one may expect that the set $\mathfrak D_d$ has
smaller dimension. A natural question is how big the (Hausdorff) dimension of $\mathfrak D_d$ can be depending the regularity of $f$.
To study the question, let us set
\begin{equation*}
\mathcal D_d( s ) := \sup_{ f \in H^s(\mathbb R^d)} \dim_H \mathfrak D_d (f) .
\end{equation*}
The problem of determining $\mathcal D_d(s)$ was first considered by Barcel\'o, Bennett, Carbery and Rogers \cite{BBCR}
for the more general dispersive equations and the result was later extended by Luc\`{a} and Rogers \cite{LR}.
Since $|e^{it\sqrt{\mathsmaller{-}\Delta}} f | \lesssim \|f\|_{H^s}$ for $s>d/2$, it follows
$\mathcal D_d(s) =0$ for $s >d/2$. On the other hand, $\mathcal D_d(s)=d$ if $s\le 1/2$ since the $H^{1/2}$--$L^{2,\infty}$ bound \eqref{maximal} generally fails.
Thus it is sufficient to consider
\[1/2< s\le d/2.\]
So, there is nothing interesting if $d=1$. For $d\ge 2$, the problem is no longer trivial. When $d=2$ Barcel\'o et. al. \cite{BBCR} obtained the optimal result: $\mathcal D_2(s) =2-2s$ for $3/4<s\le 1$ and $\mathcal D_2(s) =4-4s$ for $1/2<s\le 3/4$. In higher dimensions the following was conjectured to be true ({\cite[Theorem 5.1]{BBCR}}).
\begin{conj}
\label{conjecture}
For $d\ge 3$,
\begin{align}\label{Dds}
\mathcal D_d(s) & =
\begin{cases}
\, d-2s, \quad & 1\le s \le \frac d2,\\
\, d+2-4s, \quad & \frac 12<s \le 1.
\end{cases}
\end{align}
\end{conj}
\vspace{-6pt}
The lower bound was verified by considering specific functions and measures which show the sharpness of maximal estimate against the measures. As is typical with the problems of similar nature, proving the upper bound is more difficult.
Let $\mathbb B^d(x,\rho)\subset \mathbb R^d$ denote the closed ball centered at $x$ of radius $\rho$.
For $\alpha \in(0,d]$, we denote by $\mathfrak M^d(\alpha)$ the collection of
non-negative Borel measures $\mu$ supported in $\mathbb B^d(0,1)$ such that
\[
\mu ( \mathbb B^d (x, \rho)) \le C_\mu \rho^\alpha
\]
for all $x\in \mathbb R^d$ and $\rho>0$ with a constant $C_\mu$.
If $\mu \in \mathfrak M^d(\alpha)$, we define
\[
\langle \mu\rangle_\alpha = \sup_{x\in \mathbb R^{d},\, \rho>0} \rho^{-\alpha}
\mu( \mathbb B^d (x,\rho) ).
\]
An approach to obtain the upper bound on $\mathcal D_d(s)$ is to exploit the decay order $\beta=\beta(\alpha)$ of the estimate for $L^2$ average of $\widehat \mu$ over the sphere\,:
\begin{equation}\label{scale-beta2}
\int_{\S^{d-1}} | \widehat {\mu}(\lambda\theta) |^2 d\theta \lesssim \lambda^{-\beta} \|\mu\| \langle\mu\rangle_\alpha, \quad \mu\in \mathfrak M^d(\alpha),
\end{equation}
where $\| \mu\| $ is the total variation of $\mu$.
The estimate \eqref{scale-beta2} was previously studied by various authors (\cite{Wolff2, Erdogan, Erdogan2}) being related to the Falconer distance set problem, and further improvements were recently obtained
(\cite{LR, DZ}). These improvements rely on sophisticated argument which has been developed in the study of Fourier restriction problem.
An upper bound on $\mathcal D_d(s)$ can be obtained using \eqref{scale-beta2}. In fact, it was shown in \cite{BBCR} that $\mathcal D_d(s) \le \alpha$ if \eqref{scale-beta2} holds with $\beta >d-2s$.
Thus, making use of the best estimate so far (see \cite{Mattila, DZ}),
we have, for $d\ge 3$,
\begin{align}\label{div-set}
\mathcal D_d(s) \le
\begin{cases}
\, d-2s, ~ & ~ \hfill \frac{d+1}{4} < s \le \frac{d}2, \\
\, \frac{d^2-2ds } {d-1}, ~ & ~ \hfill \frac 12<s \le \frac{d+1}4.
\end{cases}
\end{align}
The bound in the case of $\frac{d+1}{4} < s \le \frac{d}2$ coincides with the sharp upper bound of $\beta= \alpha$ for $0<\alpha\le \frac{d-1}{2}$ (see \cite{Mattila}).
However, when $\frac12 <s \le \frac{d+1}{4}$ and $d\ge 4$, the current approach based on \eqref{scale-beta2} does not seem to be efficient enough to prove the optimal upper bound for $\mathcal D_d(s)$.
In fact, considering the known upper bound on $\beta$ (for example see \cite{Du}) for which \eqref{scale-beta2} holds,
the conjectured optimal upper bound $\mathcal D_d(s)=d+2-4s$ which corresponds to the estimate \eqref{scale-beta2} with $\beta = \frac{d+\alpha-2}{2}$ can not be achieved by the aforementioned implication from \eqref{scale-beta2}.
In this paper we take an alternative approach which relies on the fractal Strichartz estimate with respect to a measure (see \eqref{fracS} below), which was previously studied by some authors (see \cite{Wolff3, Erdogan, CHL, Harris2, Rogers}).
Via the approach we prove the conjecture \eqref{Dds} when $d=3$ and improve the previously known results (see \eqref{div-set}) for higher dimensions $d\ge4$ and $\frac 12<s \le \frac{d+1}4$.
\begin{thm}\label{divergence}
Let $d \ge 3$. Then
\begin{align*}
\mathcal D_d(s) \le
\begin{cases}
\ d-2s, & \quad \frac {d+1}4 \le s \le \frac{d}2,
\\
\ \frac{3d+1}{2}-4s , & \quad \frac {d}4 \le s \le \frac{d+1}4,
\\
\ \frac{d^2-d-1}{d-2}-\frac{2(d-1)s}{d-2}, &\quad
\hfill \frac12 < s \le
\frac{d}4.
\end{cases}
\end{align*}
\end{thm}
In Figure \ref{div-pic} we compare the result in Theorem \ref{divergence}, the conjectured optimal bound \eqref{Dds} and the previously known result \eqref{div-set}.
\begin{figure}
\begin{tikzpicture}[scale=1.3]
\draw [<->] (0,6.5) node (yaxis) [left] {\small $\mathcal D_d(s)$}|- (3.5,0) node ( xaxis) [right] {$s$} ;
\draw (0,6)--(3,6)--(3,0);
\draw [gray, thick, opacity=.8] (1.75, 2.5)--(3,0);
\draw [gray, opacity=.8, thick] (0.5,6) --(1.75,3);
\draw [line width=.3mm] (0.5,6)--(1.5, 3.5)--(1.75,2.5);
\draw [dashed, line width=.3mm] (0.5,6)--(1,4)--(1.75,2.5);
\fill (1.75,2.5) circle (.04cm);
\fill [gray, opacity=.8] (1.75,3) circle (.04cm);
\draw (0.5,0) node [below] {\footnotesize $\frac 12$};
\draw (1.5,0) node [below] {\footnotesize $\frac d4$};
\draw (1.75,0) node [below] {\footnotesize $\frac {d+1}4$};
\draw (1,0) node [below] {\footnotesize $1$};
\draw (3,0) node [below] {\footnotesize $\frac d2$};
\draw [gray, dashed, opacity=.8] (0.5,0)--(0.5,6);
\draw [gray, dashed, opacity=.8] (0,4)--(1,4)--(1,0);
\draw [gray, dashed, opacity=.8] (0,3.5)--(1.5,3.5)--(1.5,0);
\draw [gray, dashed, opacity=.8] (0,3)--(1.75,3)--(1.75,0);
\draw [gray, dashed, opacity=.8] (0,2.5)--(1.75,2.5);
\draw (0,0) node [left] {\small $0$};
\draw (0,2.5) node [left] {\small $\frac {d-1}2$};
\draw (0,3) node [left] {\small $\frac d2$};
\draw (0,3.5) node [left] {\small $\frac {d+1}2$};
\draw (0,4) node [left] {\footnotesize $d-2$};
\draw (0,6) node [left] {\footnotesize $d$};
\draw [gray, thick, opacity=.8] (3.85,5.3)--(4.1,5.3) node [right] {\footnotesize Previous results \eqref{div-set}};
\draw [line width=.3mm] (3.85,5)--(4.1,5) node [right] {\footnotesize Theorem \ref{divergence}};
\draw [dashed, line width=.3mm] (3.85,4.7)--(4.1,4.7) node [right] {\footnotesize Conjecture \ref{conjecture}};
\end{tikzpicture}
\caption{Dimension of the divergence set, $d\ge 6$} \label{div-pic}
\end{figure}
\subsection{Maximal estimates with respect to general measures.}
To examine the size of the divergence set,
we consider the maximal estimate
\begin{equation}\label{suptR}
\| \sup_{0<t <1} |e^{it\sqrt{\mathsmaller{-}\Delta}} f | \|_{L^2 (d\mu)} \lesssim \langle \mu \rangle_\alpha^{\frac12} \|f\|_{H^s(\mathbb R^d)}.
\end{equation}
It was shown in \cite{BBCR} that
the estimate \eqref{suptR} holds only if
$ s\ge \max ( \frac{d-\alpha}{2},~ \frac{d+2-\alpha}{4})$.
(See also Lemma \ref{nece} in Appendix A).
In fact, the lower bound of $\mathcal D_d(s)$ was obtained by proving the failure of \eqref{suptR} for $s< \max ( \frac{d-\alpha}{2},~ \frac{d+2-\alpha}{4})$ which is combined with Frostman's lemma.
For $s\in (1/2, d/2]$ let us define $\alpha_\ast(s)$ to be the infimum of $\alpha$ such that the estimate \eqref{suptR} holds.
Then, by a standard argument (see, for example, \cite[Appendix B]{BBCR} or \cite[Section 17]{Mattila2}), it follows
$
\mathcal D_d( s ) \le \, \alpha_\ast(s) .
$
Thus Theorem \ref{divergence} is a consequence of the following.
\begin{thm}\label{thm:sup}
Let $d\ge3$ and $\mu \in \mathfrak M^{d}(\alpha)$.
Then we have \eqref{suptR}
if
\begin{align}\label{s22}
s> s(\alpha,d):=
\begin{cases}
\, \frac{d-\alpha}2,
& \ \hfill 0 < \alpha \le \frac{d-1}2,
\\
\, \frac{3d+1}8-\frac \alpha 4,
& \
\frac{d-1}2 < \alpha \le \frac{d+1}2, \\
\, \frac{d-\alpha}2+\frac{\alpha-1}{2(d-1)},
& \ \hfill
\frac{d+1}2 < \alpha \le d .
\end{cases}
\end{align}
\end{thm}
Especially $s( \alpha,3)= \frac{5-\alpha}{4}$ for $\alpha\in (1, 3] $, which is sharp.
Concerning the proof of Theorem \ref{thm:sup}, our key observation is that the maximal estimates with respect to an $\alpha$-dimensional measure in $\mathbb R^d$ can be deduced from a family of estimates with respect to $\alpha$-dimensional measures in $\mathbb R^{d+1}$. This can be shown by the Kolmogorov-Seliverstov-Plessner linearization argument.
Thus the Sobolev exponent $s$ in \eqref{suptR} is closely related to the exponent $\gamma=\gamma(\alpha)$ for which the {\sl fractal Strichartz estimate}
\begin{equation}\label{fracS}
\| e^{it\sqrt{-\Delta}} f\|_{L^2(d\nu)}
\lesssim \langle\nu\rangle_\alpha^{\frac 12} \|f\|_{H^\gamma(\mathbb R^d)}
\end{equation}
holds for any $\nu \in \mathfrak M^{d+1}(\alpha)$.
For $d\ge2$, this type of estimates was studied in connection with geometric measure theory (e.g., \cite{Wolff3, Erdogan, O, OO, CHL, Harris2, Rogers}).
In particular, the sharp exponent $\gamma$ was established
by Wolff \cite{Wolff3}, Erdogan \cite{Erdogan} when $d=2$ and by Cho and two of the authors \cite{CHL} when $d=3$.
We discuss the estimate further in Section 2 before the proof of Theorem \ref{divergence}.
\subsection{\texorpdfstring{$L^2$}{L2} estimate for \texorpdfstring{$e^{it \sqrt{-\Delta}}$}{} with a product measure}
The average decay estimate
\eqref{scale-beta2}
has been of interest in relation to not only the Falconer distance set problem but also its refined variant: the pinned distance set problem.
We refer the reader to \cite{DZ, Liu, GIOW, Liu3, DIOWZ} for the most recent progresses related to the distance set problems.
By the argument due to Mattila \cite{Mattila}, it can be shown that if $E$ is a Borel set, then
\begin{equation}\label{dist}
\dim E >\alpha \quad \Rightarrow \quad |\{ |x-y| :x,y\in E \}| >0
\end{equation}
provided that \eqref{scale-beta2}
holds and $\alpha +\beta >d$.
On the other hand, Rogers \cite{Rogers} proposed an approach to Falconer's distance set problem which is based on, instead of \eqref{scale-beta2}, a Strichartz type estimate for $e^{it\sqrt{\mathsmaller{-}\Delta}} f$ with respect to fractal measures. Using the Riesz representation theorem, he showed that \eqref{dist} holds for $d-1 <2\alpha < d+1$ if the estimate
\begin{equation}\label{st}
\| e^{it\sqrt{\mathsmaller{-}\Delta}} P_\lambda f \|_{L^2(\mathbb R^d\times I; d\mu dt)} \lesssim \lambda^{\frac{d-1}{2} - \frac{\gamma}{2}} \|f\|_{L^2(\mathbb R^d)}, \quad \mu\in \mathfrak M^d(\alpha)
\end{equation}
holds true for $ \gamma > d - \alpha$.
Here, $I=(1,2)$ and $P_\lambda$ is the standard Littlewood-Paley projection operator which is given by $\widehat{P_\lambda f}(\xi) =\psi(|\xi|/\lambda)\widehat f(\xi)$ where $\psi\in \mathrm C_c^\infty((2^{-1}, 2))$.
Interestingly the estimates \eqref{scale-beta2} and \eqref{st} turn out to be essentially equivalent to each other.
\begin{thm}\label{equivlem2}
Let $\lambda \gg 1$ and $\mu \in \mathfrak M^d(\alpha)$ for $0< \alpha\le d$. Then \eqref{st} holds for $\gamma<\beta$ if the estimate \eqref{scale-beta2} holds.
Conversely, we have the estimate \eqref{scale-beta2} with $\beta = \gamma $ if \eqref{st} holds.
\end{thm}
Recently, Liu \cite{Liu} deduced the similar implication as \eqref{dist} for the pinned distance set: there exists $x\in E$ such that $D_x(E)=\{|x-y|: y\in E\}$ has positive Lebesgue measure if the estimate \eqref{scale-beta2} holds and $\alpha + \beta >d$. The estimate
\begin{equation}\label{sphe}
\| f\ast \sigma_t \|_{L^2(\mathbb B^d(0,1)\times I;d\mu dt)} \lesssim \|f\|_{H^{ - \frac{\beta}{2}+\epsilon} }
\end{equation}
played a crucial role in Liu's argument where $\sigma_t$ is the normalized surface measure on the sphere $\{x:|x|=t\}$ and $I=(1,2)$. Using an $L^2$ identity, he showed that \eqref{scale-beta2} implies
\eqref{sphe} for any $\epsilon>0$.
The estimate \eqref{sphe} is an easy consequence of \eqref{st} since
$\widehat{d\sigma_t}=c_+ e^{it|\xi|}a_+(t\xi)+c_- e^{-it|\xi|}a_-(t\xi)$ with $a_\pm(t\xi)=O(|t\xi|^{-\frac{d-1}2})$ when $|\xi|\ge 1$.
Using Theorem \ref{equivlem2} and the argument due to Mattila \cite{Mattila}, one can deduce the results in Liu \cite{Liu} and Rogers \cite{Rogers}.
\section{Proof of Theorem \ref{thm:sup} and Theorem \ref{equivlem2}} \label{Sec:2}
To prove Theorem \ref{thm:sup},
we use the fractal Strichartz estimate \eqref{fracS}.
Concerning the Lebesgue measure the estimate \eqref{fracS} is well understood but for the general measure $\nu$ the sharp regularity is not known when $d\ge 4$.
When $d=2$, Wolff \cite{Wolff3} obtained the estimate \eqref{fracS}
with $\alpha <1$ and the sharp estimates for $\alpha\ge1$ was settled by Erdo\u{g}an \cite{Erdogan}. Those results were extended to higher dimensions ($d \ge 3$) by Cho and two of the authors \cite{CHL}, where, especially for $d=3$, the sharp estimate was proved by using the sharp bilinear restriction estimate for the cone.
Recently, Harris \cite{Harris2} improved the bound when $d\ge 4$ and $\frac{d+1}{2}<\alpha \le d$.
These results can be summarized as follows.
\begin{thm}[\cite{CHL, Harris2}]\label{3d-frac}
Let $d\ge3$. For $0<\alpha < d+1$, let $\nu \in \mathfrak M^{d+1}(\alpha)$.
Then \eqref{fracS}
holds with $\gamma > \gamma(\alpha,d)$ where
\begin{align}\label{s2}
\gamma(\alpha,d)=
\begin{cases}
\frac{d-\alpha}2,
\quad & \quad \hfill
0 < \alpha \le \frac{d-1}2, \\
\frac{3d+1}8-\frac \alpha 4,
\quad & \quad \hfill
\frac{d-1}2 < \alpha \le \frac{d+1}2, \\
\frac{d-\alpha}2+\frac{\alpha-1}{2(d-1)},
\quad & \quad \hfill
\frac{d+1}2 < \alpha \le d, \\
\frac{d+1-\alpha}2,
\quad & \quad \hfill
d < \alpha \le d+1.
\end{cases}
\end{align}
\end{thm}
Now we prove Theorem \ref{thm:sup} making use of Theorem \ref{3d-frac}.
\begin{proof}[Proof of Theorem \ref{thm:sup}]
In order to prove Theorem \ref{thm:sup}, by Kolmogorov-Seliverstov-Plessner linearization argument, it suffices to show that
\begin{equation}\label{suptR1}
\| e^{i\mathbf t(\cdot) \sqrt{-\Delta}} f \|_{L^2 (d\mu)} \leq C \langle \mu \rangle_\alpha^{\frac12} \|f\|_{H^s(\mathbb R^d)}
\end{equation}
for any measurable function $\mathbf t:\mathbb B^d(0,1)\to (0,1)$ with $C$ independent of $\mathbf t$.
Let us define a linear functional $\ell$ by
\[
\ell(F) = \int F(x, \mathbf t(x)) d\mu(x)
\]
for any continuous function $F$ in $\mathrm C_0(\mathbb R^{d+1})$.
By the Riesz representation theorem it follows that there is a unique Radon measure $\nu$ on $\mathbb R^{d+1}$ such that
\[
\ell(F) = \int F(y,s) d\nu(y,s) = \int F(x, \mathbf t(x)) d\mu(x) .
\]
Clearly $\nu$ belongs to $\mathfrak M^{d+1}(\alpha )$ and
$\langle \nu \rangle_\alpha \le \langle \mu \rangle_\alpha$.\footnote{It is clear that $|(x-y,\mathbf t(x)-s)| \le r $ implies $|x-y| \le r$ and
\[\nu \big(\mathbb B^{d+1}((y,s),r) \big) =\int \chi_{\mathbb B^{d+1}((y,s),r) } \big(x,\mathbf t(x) \big)\,d\mu(x)
\le \int \chi_{\mathbb B^d(y,r)}(x)\,d\mu(x) \le \langle \mu\rangle_\alpha r^\alpha. \]
}
Since $\| e^{i\mathbf t(\cdot)\sqrt{-\Delta}}f \|_{L^2 ( d\mu)} = \| e^{it\sqrt{\mathsmaller{-}\Delta}} f\|_{L^2 (d\nu)}$,
applying Theorem \ref{3d-frac}, we see that \eqref{suptR} holds if \eqref{s22} is satisfied.
\end{proof}
We now proceed to prove Theorem \ref{equivlem2}.
The relation of \eqref{scale-beta2} and \eqref{st} is basically due to the following lemma.
\begin{lem} \label{equivlem}
Let $q \ge 2$ and $\lambda \gg 1$.
For $\mu \in \mathfrak M^d(\alpha)$, suppose that
\begin{equation}\label{sp-2q}
\Big(\int |\widehat F|^q d\mu \Big)^{\frac1q} \lesssim \lambda^{s}
\|F\|_2
\end{equation}
whenever $\supp F \subset \lambda \mathbb S^{d-1}+O(1)$. Then we have
\begin{equation}\label{co-2q}
\Big( \int_1^2 \int | \widehat G |^q d\mu dt \Big)^{\frac1q} \lesssim
\lambda^{s+\frac 12-\frac 1q}
\|G\|_2
\end{equation}
whenever $\supp G \subset \lambda \Gamma^d+O(1)$. Here $\Gamma^d:=\{ (\xi,|\xi|) \in \mathbb R^{d+1}: 2^{-1} \le |\xi| \le 2\}$.
In particular, when $q=2$, the estimates \eqref{sp-2q} and \eqref{co-2q} are equivalent.
\end{lem}
\begin{proof}
Assume that \eqref{sp-2q} holds.
Let $\psi$ be a smooth function such that
$\widehat \psi\gtrsim 1$ on $[1,2]$ and $\psi$ is supported in $[-1,1]$. Then it follows
\begin{equation}\label{Q1}
\int_1^2 \int \big|\widehat{G}(x,t) \big|^q d\mu(x) dt \lesssim \int \int_1^2 \big|\widehat{G} (x,t) \widehat \psi(t) \big|^q dt d\mu(x).
\end{equation}
We consider the inverse Fourier transform in $t$ and
observe that $\big( \widehat G(x, \cdot) \widehat \psi(\cdot) \big)^\vee(\tau) = \mathcal F_\xi G(x,\cdot) \ast \psi (\tau)$.
Here $\mathcal F_\xi$ denotes the Fourier transform in $\xi$, so that
$\widehat G(x,t) = \int \mathcal F_\xi G(x,\tau) e^{i t\tau} d\tau$.
For each $x$, it is easy to see that $\mathcal F_\xi G(x ,\cdot) \ast\psi$ is supported in $[\lambda /c,c\lambda ]$ for some $c >1$.
Thus, by Bernstein's inequality, we obtain
\begin{equation}\label{Q2}
\int \int_1^2 \big|\widehat{G} (x,t) \widehat \psi(t) \big|^q dt d\mu(x)
\lesssim \lambda^{(\frac 12-\frac 1q)q}
\int \Big( \int_1^2 \big| \widehat G(x,t) \widehat \psi(t) \big|^2\,dt \Big)^{\frac q2}\,d\mu(x).
\end{equation}
By Plancherel's theorem in $t$ and Minkowski's inequality, we
have
\begin{equation}\label{Q3}
\begin{aligned}
\int \! \Big(\! \int_1^2 \big| \widehat G(x,t) \widehat \psi(t) \big|^2\,dt \Big)^{\frac q2}\,d\mu (x)
= & \int\! \Big( \! \int |\mathcal F_\xi G(x ,\cdot) \ast\psi(\tau) |^2 d\tau \Big)^{\frac q2} d\mu (x) \\
\lesssim &
\Big(\!\int\! \Big(\! \int \big|\mathcal F_\xi G(x ,\cdot) \ast\psi(\tau) \big|^q d\mu (x)
\Big)^{\frac 2q} d\tau \Big)^{\frac q2}.
\end{aligned}
\end{equation}
Note that
\[
\mathcal F_\xi G(x ,\cdot) \ast\psi(\tau) =\mathcal F_\xi \Big(\int {G} (\xi,\rho) \psi(\tau-\rho) d\rho\Big).
\]
Since $\psi$ is supported in $[-1,1]$ and $\supp G\subset \lambda \Gamma^{d}+O(1)$, we see that for each fixed $\tau$, $\int {G} (\xi,\rho) \psi(\tau-\rho) d\rho$
is supported in the set $||\xi|-\tau|=O(1)$ and $\tau\sim \lambda$.
Thus, by \eqref{sp-2q} it follows that, for each $\tau$,
\[
\int \big|\mathcal F_\xi G(x ,\cdot) \ast\psi(\tau) \big|^q d\mu\lesssim \lambda^{s q}\Big\|\int {G} (\xi,\rho) \psi(\tau-\rho) d\rho\Big\|^q_{L^2_\xi}.
\]
Combining this with \eqref{Q1}, \eqref{Q2} and \eqref{Q3},
by Minkowski's and Young's convolution inequality we have
\[
\Big( \int_1^2 \int \big|\widehat{G} \big|^q d\mu dt \Big)^{\frac 1q}
\lesssim \lambda^{s+\frac 12-\frac 1q}
\Big(\int \Big\|\int {G} (\xi,\rho) \psi(\tau-\rho) d\rho\Big\|^2_{L^2_\xi} d\tau \Big)^{\frac 12} \lesssim \lambda^{s+\frac 12-\frac 1q} \|G\|_2 .\]
Therefore we get \eqref{co-2q}.
Now let $q =2$.
Let $F$ be a function with $\supp F\subset \lambda \mathbb S^{d-1}+O(1)$ and consider
\[ G(\xi,\tau)=F(\xi) \chi_{[\lambda, \lambda+1]}(\tau) . \]
Then, since $ \int_1^2| \widehat{\chi_{[\lambda, \lambda+1]}}(t) |^2dt\gtrsim 1$ independent of $\lambda$, by Fubini's theorem it is clear
\[ \int \big|\widehat{F} \big|^2 d\mu \lesssim \int_1^2 \int \big|\widehat{G} \big|^2 d\mu dt . \]
By \eqref{co-2q} with $q=2$ and Plancherel's theorem, it follows that
\[ \int \big|\widehat{F} \big|^2 d\mu \lesssim \lambda^{2s} \|G\|^2_2 = \lambda^{2s} \|F\|_2^2. \]
Hence \eqref{sp-2q} and \eqref{co-2q} are equivalent when $q=2$.
\end{proof}
Using this lemma, we prove Theorem \ref{equivlem2}.
\begin{proof}[Proof of Theorem \ref{equivlem2}]
Since $\mu$ has compact support, by duality and the uncertainty principle, one can easily see that \eqref{scale-beta2} is equivalent to
\begin{equation}\label{sphere1}
\big(\int | \widehat{F} | d\mu \big)^2 \lesssim \lambda^{d-1-\beta} \|\mu\|\langle \mu\rangle_\alpha \|F\|^2_2,
\end{equation}
when $\supp F \subset \lambda \mathbb S^{d-1}+O(1)$ and $\mu \in \mathfrak M^d(\alpha)$.
By the argument in \cite{BBCR}
the estimate can be strengthened to
\begin{equation}\label{weak22}
\| \widehat F \|_{L^{2,\infty}(d\nu)}\lesssim \lambda^{(d-1-\beta)/2} \langle \nu\rangle_\alpha^{1/2} \|F\|_2
\end{equation} for $\nu \in \mathfrak M^d(\alpha)$. To see this, we consider a measure $d \nu_E = \nu(E)^{-1} \chi_E d\nu$ for $\nu \in \mathfrak M^d(\alpha)$ and a Borel set $E$. Then $\nu_E \in \mathfrak M^d(\alpha)$, $\|\nu_E\|\le1$, and $\langle \nu_E \rangle_\alpha \le \nu(E)^{-1} \langle \nu \rangle_\alpha$. Applying \eqref{sphere1} with $\nu_E$, we see that $ {\nu(E)}^{-1/2} \int_E |\widehat F| d\nu = \nu(E)^{1/2} \int_E | \widehat F | d\nu_E \lesssim \lambda^\frac{d-1-\beta}{2} \langle\nu\rangle_\alpha^{\frac12} \|F\|_2$. So, we get \eqref{weak22} by taking $E= \{ x: |\widehat F| > \omega \}$.
Interpolating \eqref{weak22} with the trivial estimate $\| \widehat F \|_{L^{\infty}(d\nu)}\lesssim \lambda^{\frac{d-1}2} \|F\|_2$ we obtain $L^2\to L^q(d\nu)$ estimate. Since $\supp \nu\subset \mathbb B^d(0,1)$, taking $q$ arbitrarily close to $2$,
we get
\begin{align}\label{sphere}
\int \big|\widehat{F} \big|^2 d\nu \lesssim \lambda^{d-1-\gamma} \langle \nu\rangle_\alpha \|F\|^2_2
\end{align}
for $\gamma<\beta$.
By Lemma \ref{equivlem}, we can see that \eqref{sphere} is equivalent to
\begin{align}\label{cone}
\int_1^2 \int \big|\widehat{G} \big|^2 d\nu dt \lesssim \lambda^{d-1-\gamma} \langle \nu\rangle_\alpha \|G\|^2_2,
\end{align}
where $\supp G \subset \lambda \Gamma^d+O(1)$.
Let $\varphi\in \mathcal S(\mathbb R^d)$ such that $\varphi\ge 1$ on $\mathbb B^d(0,1)$ and $\supp \widehat \varphi\subset \mathbb B^d(0,1)$.
Then the space time Fourier transform of $\varphi e^{it\sqrt{\mathsmaller{-}\Delta}} P_\lambda f$ is supported in $\lambda\Gamma^d+O(1)$. By \eqref{cone} and Plancherel's theorem
we get
\begin{align}\label{ext-cone}
\int_1^2 \int \Big|\varphi (x)e^{it\sqrt{\mathsmaller{-}\Delta}} P_\lambda f(x) \Big|^2 d\nu dt\lesssim \lambda^{d-1-\gamma} \langle \nu\rangle_\alpha \|f\|_2^2,
\end{align}
which gives \eqref{st} with $\gamma <\beta $ as desired.
To show the converse implication we note that \eqref{st} implies \eqref{cone}.
It can be seen easily by Plancherel's theorem and the standard slicing argument decomposing the conic neighborhood into a family of cones.
Thus by Lemma \ref{equivlem} we have \eqref{sphere} which clearly implies
$\big|\int \widehat{F} d\mu \big|^2 \lesssim \lambda^{d-1-\gamma} \|\mu\|\langle \mu\rangle_\alpha \|F\|_2^2$. Therefore by duality and the uncertainty principle we get \eqref{scale-beta2} with $\beta=\gamma$.
\end{proof}
|
\section{Additional Experiments}\label{app:add_exp}
In this section, we further investigate the effectiveness of the proposed {\sffamily{DRFA}} algorithm. To do so, we use the Adult and Shakespeare datasets.
\noindent \textbf{Experiments on Adult dataset.}~The Adult dataset contains census data, with the target of predicting whether the income is greater or less than $\$50K$. The data has $14$ features from age, race, gender, among others. It has $32561$ samples for training distributed across different groups of sensitive features. One of these sensitive features is gender, which has two groups of ``male'' and ``female''. The other sensitive feature we will use is the race, where it has $5$ groups of ``black'', ``white'', ``Asian-Pac-Islander'', ``Amer-Indian-Eskimo'', and ``other''. We can distribute data among nodes based on the value of these features, hence make it heterogeneously distributed.
For the first experiment, we distribute the training data across $10$ nodes, 5 of which contain only data from the female group and the other $5$ have the male group's data. Since the size of different groups' data is not equal, the data distribution is unbalanced among nodes. Figure~\ref{comp_base_adult9} compares {\sffamily{DRFA}} with AFL~\cite{mohri2019agnostic}, q-FedAvg~\cite{li2019fair}, and FedAvg~\cite{mcmahan2017communication} on the Adult dataset, where the data is distributed among the nodes based on the gender feature. We use logistic regression as the loss function, the learning rate is set to $0.1$ and batch size is $50$ for all algorithms, $\gamma$ is set to $0.2$ for both {\sffamily{DRFA}} and AFL, and $q=0.5$ is tuned for the best results for q-FedAvg. The worst distribution or node accuracy during the communication rounds shows that {\sffamily{DRFA}} can achieve the same level of worst accuracy with a far fewer number of communication rounds, and hence, less overall wall-clock time. However, AFL computational cost is less than that of {\sffamily{DRFA}}. Between each communication rounds {\sffamily{DRFA}}, q-FedAvg and FedAvg have $10$ update steps. FedAvg after the same number of communications as AFL still cannot reach the same level of worst accuracy. Figure~\ref{fig:comp_base_adult9_stdev} shows the standard deviation of accuracy among different nodes as a measure for the fairness of algorithms. It can be inferred that {\sffamily{DRFA}} efficiently decreases the variance with a much fewer number of communication rounds with respect to other algorithms.
\begin{figure*}[h!]
\centering
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/adult/worst_acc_comm_compare_adult_comm_round_base.pdf}
\label{fig:comp_base_adult9_comm_round}
}
\hfill
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/adult/worst_acc_comm_compare_adult_iteration_base.pdf}
\label{fig:comp_base_adult9_iteration}
}
\hfill
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/adult/acc_stdev_comm_compare_adult_comm_round_base.pdf}
\label{fig:comp_base_adult9_stdev}
}
\caption[]{Comparing the worst distribution accuracy on {\sffamily{DRFA}}, AFL, q-FedAvg, and FedAVG on the Adult dataset. We have $10$ nodes, and data is distributed among them based on the gender feature. The loss function is logistic regression. {\sffamily{DRFA}} needs a fewer number of communications to reach the same worst distribution accuracy than the AFL and q-FedAvg algorithms. Also, {\sffamily{DRFA}} efficiently decreases the variance of the performance of different clients.}
\label{comp_base_adult9
\end{figure*}
Next, we distribute the Adult data among clients based on the ``race'' feature, which has $5$ different groups. Again the size of data among these groups is not equal and makes the distribution unbalanced. We distribute the data among $10$ nodes, where every node has only data from one group of the race feature. For this experiment, we use a nonconvex loss function, where the model is a multilayer perceptron (MLP) with $2$ hidden layers, each with $50$ neurons. The first layer has $14$ and the last layer has $2$ neurons. The learning rate is set to $0.1$ and batch size is $50$ for all algorithms, the $\gamma$ is set to $0.2$ for {\sffamily{DRFA}} and AFL, and the $q$ parameter in q-FedAvg is tuned for $0.5$. Figure~\ref{comp_base_adult8} shows the results of this experiment, where again, {\sffamily{DRFA}} can achieve the same worst-case accuracy with a much fewer number of communications than AFL and q-FedAvg. In this experiment, with the same number of local iterations, AFL still cannot reach to the {\sffamily{DRFA}} performance. In addition, the variance on the performance of different clients in Figure~\ref{fig:comp_base_adult8_stdev} suggests that {\sffamily{DRFA}} is more successful than q-FedAvg to balance the performance of clients.
\begin{figure*}[t!]
\centering
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/adult/worst_acc_comm_compare_adult-8_comm_round_base_mlp.pdf}
\label{fig:comp_base_adult8_comm_round}
}
\hfill
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/adult/worst_acc_comm_compare_adult-8_iteration_base_mlp.pdf}
\label{fig:comp_base_adult8_iteration}
}
\hfill
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/adult/acc_stdev_comm_compare_adult_8_comm_round_base.pdf}
\label{fig:comp_base_adult8_stdev}
}
\caption[]{Comparing the worst distribution accuracy on {\sffamily{DRFA}}, AFL, q-FedAvg, and FedAvg with the Adult dataset. We have $10$ nodes, and data is distributed among them based on the race feature. The model is an MLP with $2$ hidden layers, each with $50$ neurons and a cross-entropy loss function. {\sffamily{DRFA}} needs a fewer number of communications to reach the same worst distribution accuracy than the AFL and q-FedAvg algorithms. Moreover, {\sffamily{DRFA}} is more efficient in reducing the performance variance among different clients than q-FedAvg. }
\label{comp_base_adult8
\end{figure*}
\noindent \textbf{Experiments on Shakespeare dataset.}~Now, we run the same experiments on the Shakespeare dataset. This dataset contains the scripts from different Shakespeare's plays divided based on the character in each play. The task is to predict the next character in the text, providing the preceding characters. For this experiment, we use $100$ clients' data to train our RNN model. The RNN model comprises an embedding layer from $86$ characters to $50$, followed by a layer of GRU~\cite{cho2014learning} with $50$ units. The output is going through a fully connected layer with an output size of $86$ and a cross-entropy loss function. We use the batch size of $2$ with $50$ characters in each batch. The learning rate is optimized to $0.8$ for the FedAvg and used for all algorithms. The $\gamma$ is tuned to the $0.01$ for AFL and {\sffamily{DRFA}}, and $q=0.1$ is the best for the q-FedAvg. Figure~\ref{comp_base_shaks} shows the results of this experiment on the Shakespeare dataset. It can be seen that {\sffamily{DRFA}} and FedAvg can reach to the same worst distribution accuracy compared to AFL and q-FedAvg. The reason that FedAvg is working very well in this particular dataset is that the distribution of data based on the characters in the plays does not make it heterogeneous. In settings close to homogeneous distribution, FedAvg can achieve the best results, with {\sffamily{DRFA}} having a slight advantage over that.
\begin{figure*}[t!]
\centering
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/shakespeare/worst_acc_comm_compare_shakespeare_comm_round_base_rnn.pdf}
\label{fig:comp_base_shaks_comm_round}
}
\hfill
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/shakespeare/worst_acc_comm_compare_shakespeare_iteration_base_rnn.pdf}
\label{fig:comp_base_shaks_iteration}
}
\hfill
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/shakespeare/acc_comm_compare_shakespeare_comm_round_base.pdf}
\label{fig:comp_base_shaks_avg}
}
\caption[]{ Comparing different algorithms on training an RNN on Shakespeare dataset using $100$ clients. {\sffamily{DRFA}} and FedAvg outperform the other two algorithms in terms of communication efficiency, however, AFL can achieve the same level with lower computation cost. In the average performance, AFL requires much more communication to reach to the same level as FedAvg and {\sffamily{DRFA}}. }
\label{comp_base_shaks
\end{figure*}
\section{Formal Convergence Theory for Alternative Algorithm in Regularized Case}\label{app: alternative algorithm}
Here, we will present the formal convergence theory of the algorithm we described in Section~\ref{sec:alternative algorithm}, where we use full batch gradient ascent to update $\boldsymbol{\lambda}^{(s)}$. To do so, the server sends the current global model $\bar{\bm{w}}^{(s)}$ to all clients and each client evaluates the global model on its local data shards and send $f_i(\bar{\bm{w}}^{(s)})$ back to the server. Then the server can compute the full gradient over dual parameter $\bm{\lambda}$ and take a gradient ascent (GA) step to update it. The algorithm is named {\sffamily{DRFA-GA}} and described in Algorithm~\ref{alg:5}. We note that {\sffamily{DRFA-GA}} can be considered as communication-efficient variant of AFL, but without sampling clients to evaluate the gradient at dual parameter. We conduct the convergence analysis on the setting where the regularized term is strongly-concave in $\boldsymbol{\lambda}$, and loss function is strongly-convex and nonconvex but satisfying Polyak-Łojasiewicz (PL) condition in $\bm{w}$. So, our theory includes strongly-convex-strongly-concave and nonconvex (PL condition)-strongly-concave cases.
\noindent \textbf{Strongly-Convex-Strongly-Concave case.~} We start by stating the convergence rate when the individual local objectives are strongly convex and the regularizer $g(\bm{\lambda})$ is strongly concave in $\bm{\lambda}$, making the global objective $F (\boldsymbol{w},\boldsymbol{\lambda}) := \sum_{i=1}^N \lambda_i f_i(\boldsymbol{w}) +g(\boldsymbol{\lambda})$ also strongly concave in $\bm{\lambda}$.
\begin{theorem}\label{theorem2}
Let each local function $f_i$ be $\mu$-strongly convex, and global function $F$ is $\mu$-strongly concave in $\boldsymbol{\lambda}$. Under Assumptions~\ref{assumption: smoothness}, \ref{assumption: bounded gradient},\ref{assumption: bounded domain},\ref{assumption: bounded variance}, if we optimize (\ref{regularized loss}) using the {\sffamily{DRFA-GA}} (Algorithm~\ref{alg:5}) with synchronization gap $\tau$, choosing learning rates as $\eta = \frac{4\log T}{\mu T}$ and $\gamma = \frac{1}{ L}$ and $T \geq \frac{16\alpha \log T}{\mu}$, where $\alpha=\kappa L+L$, using the averaging scheme $\hat{\boldsymbol{w}} = \frac{2}{mT} \sum_{t=T/2}^T \sum_{i\in{\mathcal{D}^{(\floor*{\frac{t}{\tau}})}}} \boldsymbol{w}^{(t)}_i$ we have:
\begin{equation*}
\begin{aligned} \mathbb{E}[\Phi(\hat{\boldsymbol{w}})-\Phi(\boldsymbol{w}^*)] = \Tilde{O}\left( \frac{\mu D_{\mathcal{W}}^2}{T} + \frac{\kappa^2L\tau D_{\Lambda}^2 }{T}
+ \frac{\sigma_w^2+ G_w^2}{\mu mT} + \frac{\kappa^2\tau^2(\sigma_w^2 + \Gamma)}{\mu T^2} + \frac{\kappa^6\tau^2G_w^2}{\mu T^2}\right),
\end{aligned}
\end{equation*}
where $\kappa = L/\mu$, and $\bm{w}^*$ is the minimizer of $\Phi$.
\begin{proof}
The proof is given in Section~\ref{sec:scsc}.
\end{proof}
\end{theorem}
\begin{corollary}\label{Coro: 2}
Continuing with Theorem~\ref{theorem2}, if we choose $\tau = \sqrt{T/m}$, we recover the rate:
\begin{align*}
\mathbb{E}[\Phi(\hat{\boldsymbol{w}})-\Phi(\boldsymbol{w}^*)] = \Tilde{O}\left( \frac{\kappa^2L D_{\Lambda}^2 }{ \sqrt{mT}} +\frac{\mu D_{\mathcal{W}}^2}{T} +\frac{\kappa^2 (\sigma_w^2 + \Gamma)+\kappa^6G_w^2}{\mu mT } \right).
\end{align*}
\end{corollary}
Here we obtain $\tilde{O}\left(\frac{\tau}{T}\right)$ rate in Theorem~\ref{theorem2}. If we choose $\tau = 1$, which is fully synchronized SGD, then we recover the same rate $\tilde{O}\left(\frac{1}{T}\right)$ as in vanilla agnostic federated learning~\cite{mohri2019agnostic}. If we choose $\tau$ to be $O(\sqrt{T/m})$, we recover the rate $ \Tilde{O}\left( \frac{1}{\sqrt{mT}} + \frac{1}{mT} \right)$, which can achieve linear speedup with respect to number of sampled workers. The dependency on gradient dissimilarity $\Gamma$ shows that the data heterogeneity will slow down the rate, but will not impact the dominating term.
\begin{algorithm}[t]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\caption{Distributionally Robust Federated Averaging: Gradient Ascent ({\sffamily{DRFA-GA}})}
\label{alg:5}
\begin{algorithmic}[1]
\REQUIRE $N$ clients , synchronization gap $\tau$, total number of iterations $T$, $S=T/\tau$, learning rates $\eta$, $\gamma$, sampling size $m$, initial model $\bar{\boldsymbol{w}}^{(0)}$ and initial $\boldsymbol{\lambda}^{(0)}$.
\ENSURE Final solutions $\hat{\boldsymbol{w}} = \frac{2}{mT} \sum_{t=T/2}^{T}\sum_{i\in{\mathcal{D}^{(\floor{\frac{t}{\tau}})}}} \boldsymbol{w}^{(t)}_i$, $\hat{\boldsymbol{\lambda}} = \frac{1}{S}\sum_{s=0}^{S-1} \boldsymbol{\lambda}^{(s)}$, or (2) $\bm{w}^T$, $\boldsymbol{\lambda}^S$.
\FOR{ $s = 0$ to $S-1$ }
{\STATE {Server \textbf{samples} $\mathcal{D}^{(s)} \subset [N]$ according to $\boldsymbol{\lambda}^{(s)}$ with size of $m$}\\
\STATE Server \textbf{broadcasts} $\bar{\boldsymbol{w}}^{(s)}$ to all clients $i \in
\mathcal{D}^{(s)}$}\\[5pt]
\algemph{antiquewhite}{1}{
\FOR{clients $i \in \mathcal{D}^{(s)}$ \textbf{parallel}}
\STATE Client \textbf{sets} $\boldsymbol{w}_i^{(s\tau)} = \bar{\bm{w}}^{(s)}$
\FOR{$t = s\tau,\ldots,(s+1)\tau-1 $}
\STATE {$\bm{w}^{(t+1)}_i = \prod_{\mathcal{W}}\left(\bm{w}^{(t)}_i - \eta \nabla f_i(\bm{w}^{(t)}_i;\xi^{(t)}_i)\right) $}\\[-8pt]
\ENDFOR
\ENDFOR
\STATE {Client $i \in \mathcal{D}^{(s)}$ \textbf{sends} $\bm{w}^{((s+1)\tau)}_i$ back to the server}}
\algemph{blizzardblue}{1} {
\STATE Server sends $\bar{\bm{w}}^{(s)}$ to all clients \hfill{\mycommfont{{// Update $\boldsymbol{\lambda}$}}}
\STATE Each client $i \in [N]$ evaluates $\bar{\bm{w}}^{(s)}$ on its local data and sends $f_i(\bar{\bm{w}}^{(s)})$ back to server
\STATE {Server updates $\boldsymbol{\lambda}^{(s+1)}= \prod_{\Lambda}\left(\boldsymbol{\lambda}^{(s)}+\gamma \nabla_{\boldsymbol{\lambda}} F\left(\bar{\bm{w}}^{(s)},\boldsymbol{\lambda}^{(s)}\right) \right)$}}
\STATE {Server \textbf{computes} $\bar{\bm{w}}^{(s+1)} = \frac{1}{m} \sum_{i\in \mathcal{D}^{(s)}} \bm{w}^{((s+1)\tau)}_i$ }\\
\ENDFOR
\end{algorithmic}
\end{algorithm}
\noindent\textbf{Nonconvex (PL condition)-Strongly-Concave Setting.} We provide the convergence analysis under the condition where $F$ is nonconvex but satisfies PL condition in $\bm{w}$, and strongly concave in $\boldsymbol{\lambda}$. In the constraint problem, to prove the convergence, we have to consider a generalization of PL condition~\cite{karimi2016linear} as formally stated below.
\begin{definition}
[($\mu$,$\eta$)-generalized Polyak-Łojasiewicz (PL)] \label{assumption: PL} The global objective function $F(\cdot, \boldsymbol{\lambda})$ is differentiable and satisfies the ($\mu$,$\eta$)-generalized Polyak-Łojasiewicz condition with constant $\mu$ if the following holds: $$\frac{1}{2\eta^2}\left\|\bm{w}-\prod_{\mathcal{W}}\left(\bm{w}-\eta \nabla_{\boldsymbol{w}} F(\boldsymbol{w},\boldsymbol{\lambda})\right) \right\|_2^2 \geq \mu(F(\boldsymbol{w},\boldsymbol{\lambda}) - \min_{\bm{w}'\in\mathcal{W}}F(\boldsymbol{w}',\boldsymbol{\lambda})), \forall \boldsymbol{\lambda} \in \Lambda$$.
\end{definition}
\begin{remark}
When the constraint is absent, it reduces to vanilla PL condition~\cite{karimi2016linear}. The similar generalization of PL condition is also mentioned in~\cite{karimi2016linear}, where they introduce a variant of PL condition to prove the convergence of proximal gradient method. Also we will show that, if $F$ satisfies $\mu$-PL condition in $\bm{w}$, $\Phi(\bm{w})$ also satisfies $\mu$-PL condition.
\end{remark}
We now proceed to provide the global convergence of $\Phi$ in this setting.
\begin{theorem}
\label{thm3}
Let global function $F$ satisfy ($\mu$,$\eta$)-generalized PL condition in $\bm{w}$ and $\mu$-strongly-concave in $\boldsymbol{\lambda}$. Under Assumptions~\ref{assumption: smoothness},\ref{assumption: bounded gradient},\ref{assumption: bounded domain},\ref{assumption: bounded variance}, if we optimize (\ref{regularized loss}) using the {\sffamily{DRFA-GA}} (Algorithm~\ref{alg:5}) with synchronization gap $\tau$, choosing learning rates $\eta = \frac{ 4\log T}{\mu T}$, $\gamma = \frac{1}{ L}$ and $m \geq T$, with the total iterations satisfying $T\geq \frac{8\alpha \log T}{\mu}$ where $\alpha = L+\kappa L$, $\kappa = \frac{L}{\mu}$, we have:
\begin{align*}
\mathbb{E}\left[\Phi(\bm{w}^{(T)}) - \Phi(\bm{w}^{*})\right] \nonumber
& \leq O\left(\frac{\Phi(\bm{w} ^{(0)}) - \Phi( \bm{w} ^*)}{T}\right)+
\Tilde{O}\left(\frac{\sigma_w^2+G_w^2}{\mu T}\right) + \Tilde{O}\left(\frac{\kappa^2 L \tau D_\Lambda^2}{T}\right) \nonumber\\
& \quad + \Tilde{O}\left(\frac{\kappa^6\tau^2 G_w^2}{\mu T^2}\right)+ \Tilde{O}\left(\frac{\kappa^2\tau^2 (\sigma_w^2+\Gamma)}{\mu T^2}\right)\nonumber.
\end{align*}
where $\bm{w}^* \in \arg\min_{\bm{w} \in \mathcal{W}}\Phi(\bm{w})$.
\end{theorem}
\begin{proof}
The proof is given in Section~\ref{sec:ncsc}.
\end{proof}
\begin{corollary}
Continuing with Theorem~\ref{thm3}, if we choose $\tau = \sqrt{T/m}$, we recover the rate:
\begin{align*}
\mathbb{E}[\Phi(\hat{\boldsymbol{w}})-\Phi(\boldsymbol{w}^*)] = \Tilde{O}\left( \frac{\kappa^2L D_{\Lambda}^2 }{ \sqrt{T}} +\frac{\Phi(\bm{w}^{(0)}) - \Phi(\bm{w}^{*})}{T} +\frac{\kappa^2 (\sigma_w^2 + \Gamma)+\kappa^6G_w^2}{\mu T } \right).
\end{align*}
\end{corollary}
We obtain $\tilde{O}\left(\frac{ \tau}{T}\right)$ convergence rate here, slightly worse than that of strongly-convex-strongly-concave case. We also get linear speedup in the number of sampled workers if properly choose $\tau$. The best known result of non-distributionally robust version of FedAvg on PL condition is $O(\frac{1}{T})$ \cite{haddadpour2019convergence}, with $O(T^{1/3})$ communication rounds. It turns out that we trade some convergence rate to guarantee a worst case performance. We would like to mention that, here we require $m$, the number of sampled clients to be a large number, which is the imperfection of our analysis. However, we would note that, this is similar to the analysis in~\cite{ghadimi2016mini} for projected SGD on constrained nonconvex minimization problems, where it is required to employ growing mini-batch sizes with iterations to guarantee convergence to a first-order stationary point (i.e., imposing a constraint on minibatch size based on target accuracy $\epsilon$ which plays a similar rule to $m$ in our case).
\section{Proof of Convergence of {\sffamily{DRFA}} for Convex Losses (Theorem~\ref{theorem1})} \label{sec: proof DRFA convex}
In this section we will present the proof of Theorem~\ref{theorem1}, which states the convergence of {\sffamily{DRFA}} in convex-linear setting.
\subsection{Preliminary}
Before delving into the proof, let us introduce some useful variables and lemmas for ease of analysis. We define a virtual sequence $\{ \bm{w}^{(t)}\}_{t=1}^T$ that will be used in our proof, and we also define some intermediate variables:
\begin{equation}
\begin{aligned}
\bm{w}^{(t)} &= \frac{1}{m}\sum_{i\in{\mathcal{D}^{(\floor{\frac{t}{\tau}})}}} \bm{w}^{(t)}_i, && \text{(average model of selected devices)}
\\ \Bar{\bm{u}}^{(t)} &= \frac{1}{m}\sum_{i\in \mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_i(\bm{w} ^{(t)}_i), && \text{(average full gradient of selected devices)} \\
\bm{u}^{(t)} &= \frac{1}{m}\sum_{i\in{\mathcal{D}^{(\floor{\frac{t}{\tau}})}}}\nabla f_i( \bm{w} ^{(t)}_i;\xi^{(t)}_i) && \text{(average stochastic gradient of selected devices)} \label{vs} \\
\Bar{\bm{v}}^{(t)} &= \nabla_{\boldsymbol{\lambda}} F(\bm{w} ^{(t)},\boldsymbol{\lambda}) = \left[f_1(\bm{w}^{(t)}), \ldots,f_N(\bm{w} ^{(t)})\right] && \text{(full gradient w.r.t. dual)}\\
\bar{\Delta}_{s} &= \sum_{t=s\tau+1}^{(s+1)\tau} \gamma \Bar{\bm{v}}^{(t)}, \\
\Delta_{s} &= \tau \gamma \bm{v}, && \text{(see below)}\\
\delta^{(t)} &= \frac{1}{m} \sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \left\|\bm{w}^{(t)}_i - \bm{w}^{(t)}\right\|^2, \nonumber
\end{aligned}
\end{equation}
where $\bm{v} \in \mathbb{R}^N$ is the stochastic gradient for dual variable generated by Algorithm~\ref{alg:1} for updating $\boldsymbol{\lambda}$, such that ${v}_i = f_i(\bm{w}^{(t')};\xi_i)$ for $i \in \mathcal{U} \subset [N]$ where $\xi_i$ is stochastic minibatch sampled from $i$th local data shard, and $t'$ is the snapshot index sampled from $s\tau+1$ to $(s+1)\tau$.
\subsection{Overview of the Proof}
The proof techniques consist of analyzing the one-step progress for the virtual iterates $\bm{w} ^{(t+1)}$ and $\boldsymbol{\lambda}^{(s+1)}$, however periodic decoupled updating along with sampling makes the analysis more involved compared to fully synchronous primal-dual schemes for minimax optimization. Let us start from analyzing one iteration on $\bm{w}$. From the updating rule we can show that
\begin{align}
\mathbb{E}\|\bm{w} ^{(t+1)} - \bm{w} \|^2 &\leq \mathbb{E}\|\bm{w} ^{(t)}- \bm{w} \|^2 -2\eta\mathbb{E}\left[F(\bm{w} ^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) - F(\bm{w} ,\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] \nonumber\\
& \quad + L\eta \mathbb{E}\left[\delta^{(t)}\right] + \eta^2\mathbb{E}\|\Bar{\bm{u}}^{(t)} - \bm{u}^{(t)}\|^2 + \eta^2G_{w}^2.\nonumber
\end{align}
Note that, similar to analysis of local SGD, e.g., ~\cite{stich2018local}, the key question is how to bound the deviation $\delta^{(t)}$ between local and (virtual) averaged model. By the definition of gradient dissimilarity, we establish that:
\begin{align}
\frac{1}{T}\sum_{t=0}^{T}\mathbb{E}\left[\delta^{(t)}\right] = 10\eta^2\tau^2 \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right).\nonumber
\end{align}
It turns out the deviation can be upper bounded by variance of stochastic graident, and the gradient dissimilarity. The latter term controls how heterogenous the local component functions are, and it becomes zero when all local functions are identical, which means we are doing minibatch SGD on the same objective function in parallel.
Now we switch to the one iteration analysis on $\boldsymbol{\lambda}$:
\begin{align}
\mathbb{E} \|\boldsymbol{\lambda}^{(s+1)} - \boldsymbol{\lambda} \|^2 &\leq \mathbb{E}\|\boldsymbol{\lambda}^{(s)}- \boldsymbol{\lambda} \|^2 \nonumber\\ &\quad -\sum_{t=s\tau+1}^{(s+1)\tau } \mathbb{E}[2\gamma(F(\bm{w} ^{(t)},\boldsymbol{\lambda}^{(s)})-F(\bm{w} ^{(t)},\boldsymbol{\lambda}))] + \mathbb{E}\|\bar{\Delta}_{s}\|^2 + \mathbb{E}\|\Delta_{s} - \bar{\Delta}_{s}\|^2.\nonumber
\end{align}
It suffices to bound the variance of $\Delta_s$. Using the identity of independent variables we can prove:
\begin{equation}
\mathbb{E}[\|\Delta_{s} - \bar{\Delta}_{s}\|^2] \leq \gamma^2\tau^2\frac{\sigma_{\lambda}^2}{m}. \nonumber
\end{equation}
It shows that the variance depends quadratically on $\tau$\footnote{This dependency is very heavy, and one open question is to see if we employ a variance reduction scheme to loosen this dependency.}, and can achieve linear speed up with respect to the number of sampled workers. Putting all pieces together, and doing the telescoping sum will yield the result in Theorem~\ref{theorem1}.
\subsection{Proof of Technical Lemmas}
In this section we are going to present some technical lemmas that will be used in the proof of Theorem~\ref{theorem1}.
\begin{lemma}
\label{lemma: bounded variance of w}
The stochastic gradient $\bm{u}^{(t)}$ is unbiased, and its variance is bounded, which implies:
\begin{equation}
\begin{aligned}
\mathbb{E}_{\xi_i^{(t)},\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\left[\bm{u}^{(t)}\right] &= \mathbb{E}_{\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\left[\Bar{\bm{u}}^{(t)}\right] = \mathbb{E} \left[ \sum_{i=1}^N \lambda^{(\floor{\frac{t}{\tau}})}_i \nabla f_i(\bm{w}^{(t)}_i)\right] , \nonumber \\ \mathbb{E}\left[\|\bm{u}^{(t)} - \Bar{\bm{u}}^{(t)}\|^2\right] &= \frac{\sigma^2_{w}}{m}. \nonumber
\end{aligned}
\end{equation}
\begin{proof}
The unbiasedness is due to the fact that we sample the clients according to $\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}$. The variance term is due to the identity $\mathrm{Var}(\sum_{i=1}^m \bm{X}_i) = \sum_{i=1}^m \mathrm{Var}(\bm{X}_i)$.
\end{proof}
\end{lemma}
\begin{lemma}
\label{lemma: bounded variance of lambda}
The stochastic gradient at $\boldsymbol{\lambda}$ generated by Algorithm~\ref{alg:1} is unbiased, and its variance is bounded, which implies:
\begin{equation}
\mathbb{E}[\Delta_{s}] = \bar{\Delta}_{s}, \quad \quad \mathbb{E}[\|\Delta_{s} - \bar{\Delta}_{s}\|^2] \leq \gamma^2\tau^2\frac{\sigma_{\lambda}^2}{m}.\label{l10}
\end{equation}
\begin{proof}
The unbiasedness is due to we sample the workers uniformly. The variance term is due to the identity $\mathrm{Var}(\sum_{i=1}^m \bm{X}_i) = \sum_{i=1}^m \mathrm{Var}(\bm{X}_i)$.
\end{proof}
\end{lemma}
\begin{lemma}[One Iteration Primal Analysis]
\label{lemma: one iteration w}
For {\sffamily{DRFA}}, under the same conditions as in Theorem~\ref{theorem1}, for all $\bm{w} \in \mathcal{W}$, the following holds:
\begin{align}
\mathbb{E}\|\bm{w} ^{(t+1)} - \bm{w} \|^2 &\leq \mathbb{E}\|\bm{w} ^{(t)}- \bm{w} \|^2 -2\eta\mathbb{E}\left[F(\bm{w} ^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) - F(\bm{w} ,\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] \nonumber\\
& \quad + L\eta \mathbb{E}\left[\delta^{(t)}\right] + \eta^2\mathbb{E}\|\Bar{\bm{u}}^{(t)} - \bm{u}^{(t)}\|^2 + \eta^2G_{w}^2.\nonumber
\end{align}
\end{lemma}
\begin{proof}
From the updating rule we have:
\begin{align}
\mathbb{E}\|\bm{w}^{(t+1)} - \bm{w} \|^2 &= \mathbb{E}\left\|\prod_{\mathcal{W}}(\bm{w}^{(t)}- \eta \bm{u}^{(t)}) - \bm{w} \right\|^2 \leq \mathbb{E}\|\bm{w}^{(t)} - \eta \Bar{\bm{u}}^{(t)} - \bm{w} \|^2 + \eta^2 \mathbb{E}\|\Bar{\bm{u}}^{(t)} - \bm{u}^{(t)}\|^2 \nonumber\\
& \leq \mathbb{E}\|\bm{w} ^{(t)}- \bm{w} ^*\|^2 + \underbrace{\mathbb{E}[- 2\eta \langle \Bar{\bm{u}}^{(t)},\bm{w} ^{(t)}- \bm{w} ^* \rangle]}_{T_1} + \underbrace{\eta^2 \mathbb{E}\|\Bar{\bm{u}}^{(t)}\|^2}_{T_2} + \mathbb{E}\|\Bar{\bm{u}}^{(t)} - \bm{u}^{(t)}\|^2 \label{l20}
\end{align}
We are going to bound $T_1$ first:
\begin{align}
T_1 &= \mathbb{E}_{\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\left[ \frac{1}{m} \sum_{i\in \mathcal{D}^{(\floor{\frac{t}{\tau}})}}\left[- 2\eta\left \langle\nabla f_i(\bm{w}^{(t)}_i), \bm{w} ^{(t)}- \bm{w}^{(t)}_i \right\rangle - 2\eta \left\langle \nabla f_i(\bm{w} ^{(t)}_i), \bm{w}^{(t)}_i - \bm{w} ^* \right\rangle\right]\right] \label{l21} \\
&\leq \mathbb{E}_{\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\left[2\eta \frac{1}{m}\sum_{i\in \mathcal{D}^{(\floor{\frac{t}{\tau}})}} \left[f_i(\bm{w} ^{(t)}_i) - f_i(\bm{w} ^{(t)}) + \frac{L}{2}\|\bm{w} ^{(t)} - \bm{w}^{(t)}_i \|^2 + f_i(\bm{w}) - f_i(\bm{w} ^{(t)}_i) \right]\right] \label{l22} \\
& = -2\eta \mathbb{E}\left[\sum_{i=1}^N \lambda^{(\floor{\frac{t}{\tau}})}_i f_i(\bm{w} ^{(t)}) - \lambda^{(\floor{\frac{t}{\tau}})}_i f_i(\bm{w})\right] + L\eta \mathbb{E}\left[\delta^{(t)}\right]\nonumber\\
& = -2\eta\mathbb{E}\left[F(\bm{w} ^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) - F(\bm{w} ,\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] + L\eta \mathbb{E}\left[\delta^{(t)}\right],\nonumber
\end{align}
where from~(\ref{l21}) to~(\ref{l22}) we use the smoothness and convexity properties.
We then turn to bounding $T_2$ as follows:
\begin{align}
T_2 = \eta^2 \mathbb{E}\left\|\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_i(\bm{w} ^{(t)}_i)\right\|^2 \leq \eta^2 \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \mathbb{E}\left\| \nabla f_i(\bm{w} ^{(t)}_i)\right\|^2 \leq \eta^2 G_{w}^2.\nonumber
\end{align}
Plugging $T_1$ and $T_2$ back to~(\ref{l20}) gives:
\begin{align}
\mathbb{E}\|\bm{w} ^{(t+1)} - \bm{w} \|^2 & \leq \mathbb{E}\|\bm{w} ^{(t)}- \bm{w} \|^2 -2\eta\mathbb{E}\left[F(\bm{w} ^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) - F(\bm{w} ,\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] \nonumber\\
& \quad + L\eta \mathbb{E}\left[\delta^{(t)}\right] + \eta^2 \mathbb{E}\|\Bar{\bm{u}}^{(t)} - \bm{u}^{(t)}\|^2 + \eta^2G_{w}^2,\nonumber
\end{align}
thus concluding the proof.
\end{proof}
The following lemma bounds the deviation between local models and (virtual) global average model over sampled devices over $T$ iterations. We note that the following result is general and will be used in all variants.
\begin{lemma}[Bounded Squared Deviation]
\label{lemma: deviation}
For {\sffamily{DRFA}}, {\sffamily{DRFA-Prox}} and {\sffamily{DRFA-GA}} algorithms, the expected average squared norm distance of local models $\bm{w} ^{(t)}_i, i \in \mathcal{D}^{(\floor{\frac{t}{\tau}})}$ and $\bm{w} ^{(t)}$ is bounded as follows:
\begin{align}
\frac{1}{T}\sum_{t=0}^{T}\mathbb{E}\left[\delta^{(t)}\right] \leq 10\eta^2\tau^2 \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right).\nonumber
\end{align}
where expectation is taken over sampling of devices at each iteration.
\end{lemma}
\begin{proof}
Consider $s\tau \leq t\leq (s+1)\tau$. Recall that, we only perform the averaging based on a uniformly sampled subset of workers $\mathcal{D}^{(\floor{\frac{t}{\tau}})}$ of $[N]$. Following the updating rule we have:
{\begin{align}
\mathbb{E}[\delta^{(t)}] &= \mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\|\bm{w} ^{(t)}_i- \bm{w}^{(t)}\|^2\right] \nonumber\\
& \leq \mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\mathbb{E}\left\|\bm{w}^{(s\tau)}- \sum_{r=s\tau}^{t -1} \eta \nabla f_i(\bm{w} ^{(r)}_i;\xi_i^{(r)}) -\left(\bm{w}^{(s\tau)}- \frac{1}{m}\sum_{i'\in\mathcal{D}}\sum_{r=s\tau}^{t -1}\eta \nabla f_{i'}(\bm{w} ^{(r)}_{i'};\xi_{i'}^{(r)}) \right) \right\|^2\right]\nonumber\\
& = \mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \left\| \sum_{r=s\tau}^{t-1}\eta \nabla f_i(\bm{w} ^{(r)}_i;\xi_i^{(r)}) - \frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\sum_{r=s\tau}^{t-1}\eta \nabla f_{i'}(\bm{w} ^{(r)}_{i'};\xi_{i'}^{(r)})\right\|^2 \right]\nonumber\\
& \leq \mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\eta^2\tau\sum_{r=s\tau}^{(s+1)\tau} \left\| \nabla f_i(\bm{w} ^{(r)}_i;\xi_i^{(r)}) - \frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w} ^{(r)}_{i'};\xi_{i'}^{(r)})\right\|^2\right]\nonumber \\
& = \eta^2\tau\mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\sum_{r=s\tau}^{(s+1)\tau} \left\| \nabla f_i(\bm{w} ^{(r)}_i;\xi_i^{(r)}) -\nabla f_i(\bm{w}^{(r)}_i )+ \nabla f_i(\bm{w}^{(r)}_i )-\nabla f_i(\bm{w}^{(r)} ) \vphantom{-\frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w}^{(r)}_{i'})} \right.\right. \nonumber\\
& \qquad \qquad\left. \left. +\nabla f_i(\bm{w}^{(r)})-\frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w}^{(r)})
+ \frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w}^{(r)}) \right.\right. \nonumber\\
&\qquad\qquad \left.\left.-\frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w}^{(r)}_{i'})+\frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w}^{(r)}_{i'})- \frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w} ^{(r)}_{i'};\xi_{i'}^{(r)})\right\|^2\right] \label{eq: deviation 1}
\end{align}}
Applying Jensen's inequality to split the norm yields:
{\begin{align}
\mathbb{E}[\delta^{(t)}] &\leq 5\eta^2\tau\sum_{r=s\tau}^{(s+1)\tau} \left(\sigma_w^2 + L^2\mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\left\|\bm{w}^{(r)}_i - \bm{w}^{(r)} \right\|^2\right]+ L^2\mathbb{E}\left[\frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\left\|\bm{w}^{(r)}_{i'} - \bm{w}^{(r)} \right\|^2\right] \right. \nonumber\\
& \qquad \qquad \left. +\mathbb{E}\left[\frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\left\|\nabla f_{i}(\bm{w}^{(r)}) - \nabla f_{i'}(\bm{w} ^{(r)}) \right\|^2\right]+\frac{\sigma_w^2}{m} \right) \label{eq: deviation 2}\\
& \leq 5\eta^2\tau \sum_{r=s\tau}^{(s+1)\tau}\left(\sigma^2_w + 2L^2\mathbb{E}[\delta^{(r)}] + \Gamma +\frac{\sigma^2_w}{m} \right) \label{eq: deviation 3},
\end{align}}
where from (\ref{eq: deviation 1}) to (\ref{eq: deviation 2}) we use the Jensen's inequality.
Now we sum (\ref{eq: deviation 3}) over $t = s\tau$ to $(s+1)\tau$ to get:
\begin{align}
\sum_{t=s\tau}^{(s+1)\tau}\mathbb{E}[\delta^{(t)}] & \leq 5\eta^2\tau \sum_{t=s\tau}^{(s+1)\tau}\sum_{r=s\tau}^{(s+1)\tau}\left(\sigma^2_w + 2L^2\mathbb{E}[\delta^{(r)}] + \Gamma +\frac{\sigma^2_w}{m} \right) \nonumber\\
& = 5\eta^2\tau^2 \sum_{r=s\tau}^{(s+1)\tau}\left(\sigma_w^2 + 2\mathbb{E}[\delta^{(r)}] + \Gamma +\frac{\sigma_w^2}{m} \right).\nonumber
\end{align}
Re-arranging the terms and using the fact $1-10\eta^2\tau^2 L^2 \geq \frac{1}{2}$ yields:
\begin{align}
\sum_{t=s\tau}^{(s+1)\tau}\mathbb{E}[\delta^{(t)}] \leq 10\eta^2\tau^2 \sum_{r=s\tau}^{(s+1)\tau}\left(\sigma_w^2 + \Gamma +\frac{\sigma_w^2}{m} \right).\nonumber
\end{align}
Summing over communication steps $s=0$ to $S-1$, and dividing both sides by $T=S\tau$ yields:
\begin{align}
\frac{1}{T}\sum_{t=0}^{T} \mathbb{E}[\delta^{(t)}] \leq 10\eta^2\tau^2 \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right),\nonumber
\end{align}
as desired.
\end{proof}
\begin{lemma}[Bounded Norm Deviation]
\label{lemma: deviation2}
For {\sffamily{DRFA}}, {\sffamily{DRFA-Prox}} and {\sffamily{DRFA-GA}}, $\forall i \in \mathcal{D}^{(\floor{\frac{t}{\tau}})}$, the norm distance between $\bm{w} ^{(t)}$ and $\bm{w} ^{(t)}_i$ is bounded as follows:
\begin{align}
\frac{1}{T}\sum_{t=0}^{T}\mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\left\|\bm{w} ^{(t)}_i- \bm{w}^{(t)}\right\|\right] \leq 2\eta \tau \left(\sigma_w +\frac{\sigma_w }{m} + \sqrt{\Gamma} \right).\nonumber
\end{align}
\end{lemma}
\begin{proof}
Similar to what we did in Lemma~\ref{lemma: deviation}, we assume $s\tau \leq t\leq (s+1)\tau$. Again, we only apply the averaging based on a uniformly sampled subset of workers $\mathcal{D}^{(\floor{\frac{t}{\tau}})}$ of $[N]$. From the updating rule we have:
{\begin{align}
& \mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\|\bm{w} ^{(t)}_i- \bm{w}^{(t)}\|\right] \nonumber\\
& = \mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \left\|\bm{w}^{(s\tau)}- \sum_{r=s\tau}^{t-1} \eta \nabla f_i(\bm{w} ^{(r)}_i;\xi_i^{(r)})-\left(\bm{w}^{(s\tau)}- \frac{1}{m}\sum_{i'\in\mathcal{D}}\sum_{r=s\tau}^{t-1}\eta \nabla f_{i'}(\bm{w} ^{(r)}_{i'};\xi_{i'}^{(r)}) \right) \right\|\right]\nonumber \\
& = \mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\mathbb{E}\left\| \sum_{r=s\tau}^{t-1}\eta \nabla f_i(\bm{w} ^{(r)}_i;\xi_i^{(r)}) - \frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\sum_{r=s\tau}^{t-1}\eta \nabla f_{i'}(\bm{w} ^{(r)}_{i'};\xi_{i'}^{(r)})\right\|\right] \nonumber \\
& \leq \mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\eta \sum_{r=s\tau}^{(s+1)\tau}\mathbb{E}\left\| \nabla f_i(\bm{w} ^{(r)}_i;\xi_i^{(r)}) - \frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w} ^{(r)}_{i'};\xi_{i'}^{(r)})\right\|\right]\nonumber \\
& = \eta \mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\sum_{r=s\tau}^{(s+1)\tau} \left\| \nabla f_i(\bm{w} ^{(r)}_i;\xi_i^{(r)}) -\nabla f_i(\bm{w}^{(r)}_i )+ \nabla f_i(\bm{w}^{(r)}_i )-\nabla f_i(\bm{w}^{(r)} ) +\nabla f_i(\bm{w}^{(r)}) \vphantom{\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}} \right.\right.\nonumber\\
& \quad \left. \left. -\frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w}^{(r)}) + \frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w}^{(r)}) -\frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w}^{(r)}_{i'}) \right.\right. \nonumber\\
& \quad \left.\left. +\frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w}^{(r)}_{i'})- \frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_{i'}(\bm{w} ^{(r)}_{i'};\xi_{i'}^{(r)})\right\|\right] \nonumber
\end{align}}
Applying the triangular inequality to split the norm yields:
{ \begin{align}
\mathbb{E}\left[\frac{1}{m} \vphantom{\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}}\right.&\left.\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\|\bm{w} ^{(t)}_i- \bm{w}^{(t)}\|\right] \nonumber\\
& \leq \eta \mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \sum_{r=s\tau}^{(s+1)\tau}\left(\sigma_w + L \left\|\bm{w}^{(r)}_i - \bm{w}^{(r)} \right\|\right.\right.\nonumber\\
&\quad \left. \left. + \frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}L \left\|\bm{w}^{(r)}_{i'} - \bm{w}^{(r)} \right\| +\frac{1}{m}\sum_{i'\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \left\|\nabla f_{i}(\bm{w} ^{(r)}) - \nabla f_{i'}(\bm{w}^{(r)}) \right\|+\frac{\sigma_w}{m} \right)\right] \nonumber\\
& = \eta \sum_{r=s\tau}^{(s+1)\tau}\left(\sigma_w + 2L\mathbb{E}\left[\frac{1}{m}\sum_{i'\in\mathcal{D}^{(r)}}\mathbb{E}\|\bm{w} ^{(r)}_{i'}- \bm{w}^{(r)}\|\right] + \sqrt{\Gamma} +\frac{\sigma_w}{m} \right). \label{eq: deviation2 3}
\end{align}}
Now summing (\ref{eq: deviation2 3}) over $t = s\tau$ to $(s+1)\tau$ gives:
\begin{align}
&\sum_{t=s\tau}^{(s+1)\tau}\mathbb{E}\left[\frac{1}{m}\right. \left.\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\|\bm{w} ^{(t)}_{i}- \bm{w}^{(t)}\|\right] \nonumber\\
& \leq \eta \sum_{t=s\tau}^{(s+1)\tau}\sum_{r=s\tau}^{(s+1)\tau}\left(\sigma_w + 2L\mathbb{E}\left[\frac{1}{m}\sum_{i'\in\mathcal{D}^{(r)}}\|\bm{w} ^{(r)}_{i'}- \bm{w}^{(r)}\|\right] + \sqrt{\Gamma} +\frac{\sigma_w}{m} \right) \nonumber\\
& = \eta\tau \sum_{r=s\tau}^{(s+1)\tau}\left(\sigma_w + 2L\mathbb{E}\left[\frac{1}{m}\sum_{i'\in\mathcal{D}^{(r)}}\|\bm{w} ^{(r)}_{i'}- \bm{w}^{(r)}\|\right] + \sqrt{\Gamma} +\frac{\sigma_w}{m} \right).\nonumber
\end{align}
Re-arranging the terms and using the fact $1-2\eta \tau L \geq \frac{1}{2}$ yields:
\begin{align}
\sum_{t=s\tau}^{(s+1)\tau}\mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\|\bm{w} ^{(t)}_{i}- \bm{w}^{(t)}\|\right] \leq 2 \eta \tau \sum_{r=s\tau}^{(s+1)\tau}\left(\sigma_w + \sqrt{\Gamma} +\frac{\sigma_w }{m} \right).\nonumber
\end{align}
Summing over $s=0$ to $S-1$, and dividing both sides by $T=S\tau$ yields:
\begin{align}
\frac{1}{T}\sum_{t=0}^{T}\mathbb{E}\left[\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}}\|\bm{w} ^{(t)}_{i}- \bm{w}^{(t)}\|\right] \leq 2\eta \tau \left(\sigma_w +\frac{\sigma_w }{m} + \sqrt{\Gamma} \right),\nonumber
\end{align}
which concludes the proof.
\end{proof}
\begin{lemma}[One Iteration Dual Analysis]
\label{lemma: one iteration lambda}
For {\sffamily{DRFA}}, under the assumption of Theorem~\ref{theorem1}, the following holds true for any $\boldsymbol{\lambda} \in \Lambda$:
\begin{align}
\mathbb{E} \|\boldsymbol{\lambda}^{(s+1)} - \boldsymbol{\lambda} \|^2 &\leq \mathbb{E}\|\boldsymbol{\lambda}^{(s)}- \boldsymbol{\lambda} \|^2\nonumber\\ &\quad -\sum_{t=s\tau+1}^{(s+1)\tau} \mathbb{E}[2\gamma(F(\bm{w} ^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})-F(\bm{w} ^{(t)},\boldsymbol{\lambda}))] + \mathbb{E}\|\bar{\Delta}_{t}\|^2 + \mathbb{E}\|\Delta_{t} - \bar{\Delta}_{t}\|^2. \nonumber
\end{align}
\end{lemma}
\begin{proof}
According to the updating rule for $\boldsymbol{\lambda}$ and the fact $F$ is linear in $\boldsymbol{\lambda}$ we have:
\begin{equation}
\begin{aligned}
\mathbb{E}\left\|\boldsymbol{\lambda}^{(s+1)} - \boldsymbol{\lambda}\right\|^2&= \mathbb{E}\left\|\prod_{\Lambda}(\boldsymbol{\lambda}^{(s)}+ \Delta_{s})- \boldsymbol{\lambda} \right\|^2\nonumber\\
&\leq \mathbb{E}\left\|\boldsymbol{\lambda}^{(s)}- \boldsymbol{\lambda} + \Delta_{s}\right\|^2\nonumber\\
& = \mathbb{E}\left\|\boldsymbol{\lambda}^{(s)}- \boldsymbol{\lambda} + \bar{\Delta}_{s}\right\|^2 + \mathbb{E}\left\|\Delta_{s} - \bar{\Delta}_{s}\right\|^2\nonumber\\
& = \mathbb{E}\|\boldsymbol{\lambda}^{(s)}- \boldsymbol{\lambda} \|^2 + \mathbb{E}\left[2\left\langle \bar{\Delta}_{s}, \boldsymbol{\lambda}^{(s)}- \boldsymbol{\lambda}\right\rangle\right] + \mathbb{E}\|\bar{\Delta}_{s}\|^2 +\mathbb{E} \|\Delta_{s} - \bar{\Delta}_{s}\|^2\nonumber\\
& = \mathbb{E}\|\boldsymbol{\lambda}^{(s)}- \boldsymbol{\lambda} \|^2 \nonumber\\
& \quad + 2\gamma \sum_{t= s\tau+1}^{(s+1)\tau}\mathbb{E}\left[\left\langle \nabla_{\boldsymbol{\lambda}} F(\bm{w}^{(t)},\boldsymbol{\lambda}^{(s)}), \boldsymbol{\lambda}^{(s)}- \boldsymbol{\lambda}\right\rangle\right] + \mathbb{E}\|\bar{\Delta}_{s}\|^2 + \mathbb{E}\|\Delta_{s} - \bar{\Delta}_{s}\|^2\nonumber\\
&=\|\boldsymbol{\lambda}^{(s)}- \boldsymbol{\lambda}\|^2 \nonumber \\
& \quad -2\gamma \sum_{t= s\tau+1}^{(s+1)\tau} \mathbb{E}\left[F(\bm{w} ^{(t)},\boldsymbol{\lambda})-F(\bm{w} ^{(t)},\boldsymbol{\lambda}^{(s)}))\right] + \mathbb{E}\|\bar{\Delta}_{s}\|^2 + \mathbb{E}\|\Delta_{s} - \bar{\Delta}_{s}\|^2 \label{l41},
\end{aligned}
\end{equation}
as desired.
\end{proof}
\subsection{Proof for Theorem~\ref{theorem1}}
\begin{proof}
Equipped with above results, we are now turn to proving the Theorem~\ref{theorem1}. We start by noting that $\forall \bm{w} \in \mathcal{W}$, $\forall \boldsymbol{\lambda} \in \Lambda$, according the convexity of global objective w.r.t. $\bm{w} $ and its linearity in terms of $\boldsymbol{\lambda}$ we have:
\begin{align}
&\mathbb{E}[F(\hat{\bm{w} },\boldsymbol{\lambda} ) - \mathbb{E}[F(\bm{w} ,\hat{\boldsymbol{\lambda}})]\nonumber\\
& \leq \frac{1}{T}\sum_{t=1}^T \left\{ \mathbb{E}\left[F( \bm{w}^{(t)},\boldsymbol{\lambda} )\right] - \mathbb{E}\left[F(\bm{w} ,\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] \right\}\nonumber\\
& \leq \frac{1}{T}\sum_{t=1}^T\left \{ \mathbb{E}\left[F( \bm{w}^{(t)},\boldsymbol{\lambda} )\right] -\mathbb{E}\left[F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] +\mathbb{E}\left[ F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] - \mathbb{E}\left[F(\bm{w},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] \right\}\nonumber\\
& \leq \frac{1}{T}\sum_{s=0}^{S-1} \sum_{t=s\tau+1}^{(s+1)\tau}\mathbb{E}\{F( \bm{w}^{(t)},\boldsymbol{\lambda} ) -F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(s)})\} \label{eq: thm1 1}\\
& \quad +\frac{1}{T}\sum_{t=1}^T\mathbb{E}\{ F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) - F(\bm{w} ,\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) \} \label{eq: thm1 2},
\end{align}
To bound the term in (\ref{eq: thm1 1}), pluggin Lemma~\ref{lemma: bounded variance of lambda} into Lemma~\ref{lemma: one iteration lambda}, we have:
\begin{align}
\frac{1}{T}\sum_{s=0}^{S-1} \sum_{t=s\tau+1}^{(s+1)\tau} \mathbb{E}(F(\bm{w} ^{(t)},\boldsymbol{\lambda})-F(\bm{w} ^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})) &\leq \frac{1}{2\gamma T}\|\boldsymbol{\lambda}^{(0)}- \boldsymbol{\lambda}\|^2 + \frac{\gamma\tau }{2}G_{\lambda}^2+ \frac{\gamma\tau\sigma_{ \lambda}^2 }{2m} \nonumber\\
& \leq \frac{D_{\Lambda}^2}{2\gamma T} + \frac{\gamma \tau G_{\boldsymbol{\lambda}}^2}{2} + \frac{\gamma\tau\sigma_{ \lambda}^2 }{2m}.\nonumber
\end{align}
To bound the term in (\ref{eq: thm1 2}), we plug Lemma~\ref{lemma: bounded variance of w} into Lemma~\ref{lemma: one iteration w} and apply the telescoping sum from $t = 1$ to $T$ to get:
\begin{align}
&\frac{1}{T}\sum_{t=1}^T \mathbb{E}(F(\bm{w} ^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})-F(\bm{w} ,\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}))\nonumber\\
& \leq \frac{1}{2T\eta}\mathbb{E}\|\ \bm{w}^{(0)}- \bm{w} \|^2 + 5L\eta^2\tau^2 \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right) + \frac{\eta G_w^2}{2} + \frac{\eta \sigma_{w }^2 }{2m}\nonumber\\
& \leq \frac{D_{\mathcal{W}}^2}{2T\eta} + 5L\eta^2\tau^2 \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right) + \frac{\eta G_w^2}{2} + \frac{\eta \sigma_{w }^2 }{2m}.\nonumber
\end{align}
Putting pieces together, and taking max over dual $\boldsymbol{\lambda}$, min over primal $\bm{w}$ yields:
\begin{align}
&\min_{\bm{w}\in \mathcal{W}}\max_{\boldsymbol{\lambda}\in \Lambda} \mathbb{E}[F(\hat{\bm{w} },\boldsymbol{\lambda} ) - \mathbb{E}[F(\bm{w} ,\hat{\boldsymbol{\lambda}})]\nonumber\\
& \leq \frac{D_{\mathcal{W}}^2}{2T\eta} + 5L\eta^2\tau^2 \left(\sigma_{w}^2+\frac{\sigma_{w}^2}{m} + \Gamma \right) + \frac{\eta G_w^2}{2} + \frac{\eta \sigma_{w }^2 }{2m} + \frac{D_{\Lambda}^2}{2\gamma T} + \frac{\gamma \tau G_{\boldsymbol{\lambda}}^2}{2} + \frac{\gamma\tau\sigma_{ \lambda}^2 }{2m}.\nonumber
\end{align}
Plugging in $\tau = \frac{T^{1/4}}{\sqrt{m}}$, $\eta = \frac{1}{4L \sqrt{T}}$, and $\gamma = \frac{1}{T^{5/8}}$, we conclude the proof by getting:
\begin{equation}
\begin{aligned}
\max_{\boldsymbol{\lambda}\in \Lambda}\mathbb{E}[F(\hat{\boldsymbol{w}},\boldsymbol{\lambda} )] -\min_{\bm{w}\in\mathcal{W}} \mathbb{E}[F(\boldsymbol{w} ,\hat{\boldsymbol{\lambda}} )] \leq O\Big{(}&\frac{D_{\mathcal{W}}^2+G_{w}^2}{\sqrt{T}} +\frac{D_{\Lambda}^2}{T^{3/8}} \nonumber. \\&+\frac{G_{\lambda}^2}{m^{1/2}T^{3/8}} +\frac{\sigma_{\lambda}^2}{m^{3/2}T^{3/8}}+ \frac{\sigma_w^2+\Gamma}{m\sqrt{T} }\Big{)},
\end{aligned}
\end{equation}
as desired.
\end{proof}
\section{Proof of Convergence of {\sffamily{DRFA}} for Nonconvex Losses (Theorem~\ref{thm:nonconvex-linear})}\label{sec: proof DRFA nonconvex}
This section is devoted to the proof of Theorem~\ref{thm:nonconvex-linear}).
\subsection{Overview of Proofs}
Inspired by the techniques in~\cite{lin2019gradient} for analyzing the behavior of stochastic gradient descent ascent (SGDA) algorithm on nonconvex-concave objectives, we consider the Moreau Envelope of $\Phi$:
\begin{align}
\Phi_{p} (\bm{x}) := \min_{\bm{w}\in \mathcal{W}} \left\{ \Phi (\bm{w}) + \frac{1}{2p}\|\bm{w}-\bm{x}\|\right\}.\nonumber
\end{align}
We first examine the one iteration dynamic of {\sffamily{DRFA}}:
\begin{align}
\mathbb{E}[\Phi_{1/2L} (\bm{w}^{(t)})]
& \leq \mathbb{E}[\Phi_{1/2L} (\bm{w}^{(t-1)})] + 2\eta D_{\mathcal{W}}L^2 \mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \left\| \bm{w}^{(t-1)}_i - \bm{w}^{(t-1)} \right\| \right] \nonumber \\
& 2\eta L \left( \mathbb{E}[\Phi(\bm{w}^{(t-1)})]- \mathbb{E}[F(\bm{w}^{(t-1)}, \boldsymbol{\lambda}^{\floor{\frac{t-1}{\tau}}})]\right) - \frac{\eta}{4}\mathbb{E}\left[\left\|\nabla \Phi_{1/2L} (\bm{w}^{(t-1)})\right\|^2\right].\nonumber
\end{align}
We already know how to bound $\mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \left\| \bm{w}^{(t-1)}_i - \bm{w}^{(t-1)} \right\| \right]$ in Lemma~\ref{lemma: deviation2}. Then the key is to bound $ \mathbb{E}[\Phi(\bm{w}^{(t-1)})]- \mathbb{E}[F(\bm{w}^{(t-1)}, \boldsymbol{\lambda}^{(\floor{\frac{t-1}{\tau}})})]$. Indeed this term characterizes how far the current dual variable drifts from the optimal dual variable $\boldsymbol{\lambda}^*(\bm{w}^{(t-1)})$. Then by examining the dynamic of dual variable we have $\forall \boldsymbol{\lambda} \in \Lambda$:
{\begin{align}
&\sum_{t=(s-1)\tau+1}^{s\tau}\left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \right.\left. \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s-1)} )\right]\right) \nonumber\\
&\leq \sum_{t=(s-1)\tau+1}^{s\tau}\left(\mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{t}) )\right]- \mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda} )\right] \right) \nonumber \\ &\quad+ \gamma \tau^2\frac{\sigma_{\lambda}^2}{m}+ \gamma \tau^2 G_\lambda^2+ \frac{1}{2\gamma}\left(\mathbb{E}\left[\left\|\boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right] - \mathbb{E}\left[\left\|\boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s)}\right\|^2 \right]\right).\nonumber
\end{align}}
The above inequality makes it possible to replace $\boldsymbol{\lambda}$ with $\boldsymbol{\lambda}^*$, and doing the telescoping sum so that the last term cancels up. However, in the minimax problem, the optimal dual variable changes every time when we update primal variable. Thus, we divide $S$ global stages into $\sqrt{S}$ groups, and applying the telescoping sum within one group, by setting $\boldsymbol{\lambda}=\boldsymbol{\lambda}^*(\bm{w}^{c\sqrt{S}\tau})$ at $c$th stage.
\subsection{Proof of Useful Lemmas}
Before presenting the proof of Theorem~\ref{thm:nonconvex-linear}, let us introduce the following useful lemmas.
\begin{lemma}[One iteration analysis]\label{lm: nonconvex lm1}
For {\sffamily{DRFA}}, under the assumptions of Theorem~\ref{thm:nonconvex-linear}, the following statement holds:
\begin{align}
\mathbb{E}[\Phi_{1/2L} & (\bm{w}^{(t)})] \leq \mathbb{E}[\Phi_{1/2L} (\bm{w}^{(t-1)})] + 2\eta D_{\mathcal{W}}L^2 \mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \left\| \bm{w}^{(t-1)}_i - \bm{w}^{(t-1)} \right\| \right] \nonumber \\
&+ 2\eta L \left( \mathbb{E}[\Phi(\bm{w}^{(t-1)})]- \mathbb{E}[F(\bm{w}^{(t-1)}, \boldsymbol{\lambda}^{(\floor{\frac{t-1}{\tau})}})]\right) - \frac{\eta}{4}\mathbb{E}\left[\left\|\nabla \Phi_{1/2L} (\bm{w}^{(t-1)})\right\|^2\right].\nonumber
\end{align}
\begin{proof}
Define $\tilde{\bm{w}}^{(t)} = \min_{\bm{w}\in\mathcal{W}} \Phi(\bm{w}) + L\|\bm{w}-\bm{w}^{(t)}\|^2$, the by the definition of $\Phi_{1/2L}$ we have:
\begin{align}
\Phi_{1/2L} (\bm{w}^{(t)}) \leq \Phi (\tilde{\bm{w}}^{(t-1)}) + L\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t)}\|^2.\label{eq: nonconvex lm1 1}
\end{align}
Meanwhile according to updating rule we have:
\begin{align}
&\mathbb{E}\left[\left\| \tilde{\bm{w}}^{(t-1)} -\bm{w}^{(t)}\right\|^2\right] \nonumber\\
&= \mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\prod_{\mathcal{W}}\left(\bm{w}^{(t-1)} - \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \nabla_x f_i(\bm{w}^{(t-1)}_i;\xi_i^{(t-1)})\right)\right\|^2\right]\nonumber\\
&\leq \mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right] + \mathbb{E}\left[\left\|\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \nabla_x f_i(\bm{w}^{(t-1)}_i;\xi_i^{(t-1)}) \right\|^2\right] \nonumber\\
& \quad+ 2\eta\mathbb{E}\left[\left\langle \tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)}, \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \nabla_x f_i(\bm{w}^{(t-1)}_i )\right\rangle\right]\nonumber.
\end{align}
Applying Cauchy inequality to the last inner product term yields:
\begin{align}
&\mathbb{E}\left[\left\| \tilde{\bm{w}}^{(t-1)} -\bm{w}^{(t)}\right\|^2\right] \nonumber\\
& \leq \mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right] + \eta^2(G_w^2+\sigma_w^2) + 2\eta\left\langle \tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)}, \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \nabla_x f_i(\bm{w}^{(t-1)} )\right\rangle \nonumber\\
&\quad+ \eta \mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\| \right]\mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \left\|\nabla_x f_i(\bm{w}^{(t-1)}_i) - \nabla_x f_i(\bm{w}^{(t-1)} )\right\| \right] \nonumber\\
& \leq \mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right] + \eta^2(G_w^2+\sigma_w^2) + \eta D_{\mathcal{W}}L\mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \left\| \bm{w}^{(t-1)}_i - \bm{w}^{(t-1)} \right\| \right] \nonumber\\
&\quad + 2\eta\mathbb{E}\left[\left\langle \tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)}, \nabla_x F(\bm{w}^{(t-1)},\boldsymbol{\lambda}^{\floor{\frac{t-1}{\tau}}} )\right\rangle\right].\label{eq: nonconvex lm1 2}
\end{align}
According to smoothness of $F$ we obtain:
\begin{align}
&\mathbb{E}\left[\left\langle \tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)}, \nabla_x F(\bm{w}^{(t-1)},\boldsymbol{\lambda}^{\floor{\frac{t-1}{\tau}}} )\right\rangle\right] \nonumber\\
&\leq \mathbb{E}\left[F(\tilde{\bm{w}}^{(t-1)},\boldsymbol{\lambda}^{\floor{\frac{t-1}{\tau}}})\right] - \mathbb{E}\left[F( \bm{w}^{(t-1)},\boldsymbol{\lambda}^{\floor{\frac{t-1}{\tau}}})\right] + \frac{L}{2}\mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right] \nonumber\\
& \leq \mathbb{E}\left[\Phi(\tilde{\bm{w}}^{(t-1)} )\right]- \mathbb{E}\left[F( \bm{w}^{(t-1)},\boldsymbol{\lambda}^{\floor{\frac{t-1}{\tau}}})\right]+ \frac{L}{2}\mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right] \nonumber\\
& \leq \underbrace{ \mathbb{E}\left[\Phi(\tilde{\bm{w}}^{(t-1)} )\right]+ L\mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right]}_{\leq \mathbb{E}\left[\Phi( \bm{w}^{(t-1)} )\right]+ L\mathbb{E}\left[\left\|\bm{w}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right]} - \mathbb{E}\left[F( \bm{w}^{(t-1)},\boldsymbol{\lambda}^{\floor{\frac{t-1}{\tau}}})\right] - \frac{L}{2}\mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right] \nonumber\\
& \leq \mathbb{E}\left[\Phi( \bm{w}^{(t-1)} )\right] - \mathbb{E}\left[F( \bm{w}^{(t-1)},\boldsymbol{\lambda}^{\floor{\frac{t-1}{\tau}}})\right] - \frac{L}{2}\mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right].\label{eq: nonconvex lm1 3}
\end{align}
Plugging (\ref{eq: nonconvex lm1 2}) and (\ref{eq: nonconvex lm1 3}) into (\ref{eq: nonconvex lm1 1}) yields:
\begin{align}
\Phi_{1/2L} (\bm{w}^{(t)}) &\leq \Phi (\tilde{\bm{w}}^{(t-1)}) + L \mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right] \nonumber\\
& \quad +L \eta^2(G_w^2+\sigma_w^2) + \eta D_{\mathcal{W}}L^2 \mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \left\| \bm{w}^{(t-1)}_i - \bm{w}^{(t-1)} \right\| \right] \nonumber \\
& \quad + 2L\eta\left( \mathbb{E}\left[\Phi( \bm{w}^{(t-1)} )\right] - \mathbb{E}\left[F( \bm{w}^{(t-1)},\boldsymbol{\lambda}^{\floor{\frac{t-1}{\tau}}})\right] - \frac{L}{2}\mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right]\right)\nonumber\\
& \leq \Phi_{1/2L} (\bm{w}^{(t-1)}) + L \mathbb{E}\left[\left\|\tilde{\bm{w}}^{(t-1)}-\bm{w}^{(t-1)} \right\|^2\right] \nonumber\\
& \quad +L \eta^2(G_w^2+\sigma_w^2) + \eta D_{\mathcal{W}}L^2 \mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t-1}{\tau}})}} \left\| \bm{w}^{(t-1)}_i - \bm{w}^{(t-1)} \right\| \right] \nonumber\\
& \quad+ 2L\eta\left( \mathbb{E}\left[\Phi( \bm{w}^{(t-1)} )\right] - \mathbb{E}\left[F( \bm{w}^{(t-1)},\boldsymbol{\lambda}^{\floor{\frac{t-1}{\tau}}})\right] \right) - \frac{\eta}{4} \mathbb{E}\left[\left\|\nabla \Phi_{1/2L} (\bm{w}^{(t-1)}) \right\|^2\right],\nonumber
\end{align}
where we use the result from Lemma 2.2 in~\cite{davis2019stochastic}, i.e, $\nabla \Phi_{1/2L}(\bm{w}) = 2L (\bm{w}-\Tilde{\bm{w}})$.
\end{proof}
\end{lemma}
\begin{lemma}\label{lm: nonconvex lm2}
For {\sffamily{DRFA}}, $\forall \bm{\lambda} \in \Lambda$, under the same conditions as in Theorem~\ref{thm:nonconvex-linear}, the following statement holds true:
{\begin{align}
\sum_{t=(s-1)\tau+1}^{s\tau}\left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right] \right. & \left. - \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s-1)} )\right]\right) \nonumber\\
&\leq \sum_{t=(s-1)\tau+1}^{s\tau}\left(\mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{t}) )\right]- \mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda} )\right] \right) \nonumber \\
&\quad + \gamma \tau^2\frac{\sigma_{\lambda}^2}{m} +\gamma \tau^2{G_{\lambda}^2}+ \frac{1}{2\gamma}\left(\mathbb{E}\left[\left\|\boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right] - \mathbb{E}\left[\left\|\boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s)}\right\|^2 \right]\right)\nonumber.
\end{align}}
\begin{proof}
$\forall \boldsymbol{\lambda} \in \Lambda$, according to updating rule for $\boldsymbol{\lambda}^{(s-1)}$, we have:
\begin{align}
\left\langle \boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s)}, \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^{(s-1)}- \Delta_{s-1} \right \rangle \geq 0.\nonumber
\end{align}
Taking expectation on both sides, and doing some algebraic manipulation yields:
\begin{align}
\mathbb{E}\left[\left\| \vphantom{\lambda^{(s)}}\right.\right. & \left.\left. \boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s)}\right\|^2 \right] \nonumber \\ &\leq 2 \mathbb{E}\left[\left\langle \boldsymbol{\lambda}^{(s-1)}-\boldsymbol{\lambda}, \Delta_{s-1} \right \rangle\right] + 2 \mathbb{E}\left[\left\langle \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^{(s-1)}, \Delta_{s-1} \right \rangle\right] \nonumber\\
& \quad +\mathbb{E}\left[\left\|\boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right] - \mathbb{E}\left[\left\|\boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right] \nonumber\\
& \leq 2 \mathbb{E}\left[\left\langle \boldsymbol{\lambda}^{(s-1)}-\boldsymbol{\lambda}, \bar{\Delta}_{s-1} \right \rangle\right] + 2 \mathbb{E}\left[\left\langle \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^{(s-1)}, \bar{\Delta}_{s-1} \right \rangle\right]\nonumber\\
& \quad+ 2 \mathbb{E}\left[\left\langle \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^{(s-1)}, \Delta_{s-1} - \bar{\Delta}_{s-1} \right \rangle\right]+ \mathbb{E}\left[\left\|\boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right] - \mathbb{E}\left[\left\|\boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right]\nonumber.
\end{align}
Applying the Cauchy-Schwartz and aritmetic
mean-geometric mean inequality: $2\langle \bm{p},\bm{q} \rangle \leq 2\|\bm{p}\| \|\bm{q}\|\leq \frac{1}{2}\|\bm{p}\|^2+2 \|\bm{q}\|^2$, we have:
\begin{align}
\mathbb{E}\left[\left\| \vphantom{\lambda^{(s)}}\right.\right. & \left.\left. \boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s)}\right\|^2 \right] \nonumber \\
& \leq 2 \gamma \mathbb{E}\left[ \sum_{t=(s-1)\tau+1}^{s\tau}F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s-1)} )- F(\bm{w}^{(t)}, \boldsymbol{\lambda} )\right] + \mathbb{E}\left[\left\|\boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right] \nonumber\\
& \quad+ \mathbb{E}\left[\frac{1}{2}\left\| \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^{(s-1)}\right \|^2 + 2\left\|\Delta_{s-1} - \bar{\Delta}_{s-1} \right \|^2\right]+ \mathbb{E}\left[\frac{1}{2}\left\| \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^{(s-1)}\right \|^2 + 2\left\| \bar{\Delta}_{s-1} \right \|^2\right] \nonumber \\ & \quad - \mathbb{E}\left[\left\|\boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right] \nonumber\\
& \leq 2 \gamma \mathbb{E}\left[ \sum_{t=(s-1)\tau+1}^{s\tau}F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s-1)} )- F(\bm{w}^{(t)}, \boldsymbol{\lambda} )\right] + \gamma^2\tau^2\frac{\sigma_{\lambda}^2}{m}+ \gamma^2\tau^2{G_{\lambda}^2} + \mathbb{E}\left[\left\|\boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right].\nonumber
\end{align}
By adding $\sum_{t=(s-1)\tau+1}^{s\tau}F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{(t)}))$ on both sides and re-arranging the terms we have:
{\begin{align}
&\sum_{t=(s-1)\tau+1}^{s\tau}\left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right] - \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s-1)} )\right]\right) \nonumber\\
&\leq \sum_{t=(s-1)\tau+1}^{s\tau}\left(\mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{t}) )\right]- \mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda} )\right] \right) + \gamma \tau^2\frac{\sigma_{\lambda}^2}{m}+ \gamma \tau^2{G_{\lambda}^2} \nonumber\\
& \quad + \frac{1}{2\gamma}\left(\mathbb{E}\left[\left\|\boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right] - \mathbb{E}\left[\left\|\boldsymbol{\lambda}-\boldsymbol{\lambda}^{(s)}\right\|^2 \right]\right). \nonumber
\end{align}}
\end{proof}
\end{lemma}
\begin{lemma}\label{lm: nonconvex lm3}
For {\sffamily{DRFA}}, under the assumptions in Theorem~\ref{thm:nonconvex-linear}, the following statement holds true:
\begin{align}
\frac{1}{T} \sum_{t=1}^{T}\left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})} )\right]\right) \leq 2\sqrt{S} \tau \eta G_w \sqrt{G_w^2+\sigma_w^2}+ \gamma \tau\frac{\sigma_{\lambda}^2}{m}+ \gamma \tau {G_{\lambda}^2} + \frac{ D_{\Lambda}^2}{2\sqrt{S}\tau\gamma} \nonumber
\end{align}
\begin{proof}
Without loss of generality we assume $\sqrt{S}$ is an integer, so we can equally divide index $0$ to $S-1$ into $\sqrt{S}$ groups. Then we have:
{\begin{align}
\frac{1}{T} \sum_{t=1}^{T}&\left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})} )\right]\right) \nonumber\\
& = \frac{1}{T} \sum_{c=0}^{\sqrt{S}-1}\left[ \sum_{s=c\sqrt{S}+1}^{(c+1)\sqrt{S}}\sum_{t=(s-1)\tau+1}^{s\tau} \left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s-1)} )\right]\right)\right]. \label{eq: nonconvex lm3 0}
\end{align}}
Now we and examine one group. Plugging in Lemma~\ref{lm: nonconvex lm2} and letting $\boldsymbol{\lambda} = \boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau})$ yields:
\begin{align}
&\sum_{s=c\sqrt{S}+1}^{(c+1)\sqrt{S}}\sum_{t=(s-1)\tau+1}^{s\tau} \left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s-1)} )\right]\right) \nonumber\\
&\quad\leq \sum_{s=c\sqrt{S}+1}^{(c+1)\sqrt{S}} \sum_{t=(s-1)\tau+1}^{s\tau}\left(\mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{t}) )\right]- \mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau}) )\right] \right) \nonumber\\
&\qquad + \gamma \tau^2\frac{\sqrt{S}\sigma_{\lambda}^2}{m}+ \gamma \tau^2{\sqrt{S}G_{\lambda}^2}+ \frac{1}{2\gamma}\sum_{s=c\sqrt{S}+1}^{(c+1)\sqrt{S}}\left(\mathbb{E}\left[\left\|\boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau})-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right] - \mathbb{E}\left[\left\|\boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau})-\boldsymbol{\lambda}^{(s)}\right\|^2 \right]\right) \nonumber\\
& \quad \leq \sum_{s=c\sqrt{S}+1}^{(c+1)\sqrt{S}} \sum_{t=(s-1)\tau+1}^{s\tau} \left(\mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{t}) )\right]-\mathbb{E}\left[ F(\bm{w}^{((c+1)\sqrt{S}\tau)}, \boldsymbol{\lambda}^*(\bm{w}^{t}) )\right] \right. \nonumber\\
&\qquad \left. +\mathbb{E}\left[ F(\bm{w}^{((c+1)\sqrt{S}\tau)}, \boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau}) )\right]- \mathbb{E}\left[ F(\bm{w}^{(t)},\boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau}))\right] \right) \nonumber\\
&\qquad + \gamma \tau^2\frac{\sqrt{S}\sigma_{\lambda}^2}{m}+ \gamma \tau^2{\sqrt{S}G_{\lambda}^2}+ \frac{1}{2\gamma}\sum_{s=c\sqrt{S}+1}^{(c+1)\sqrt{S}}\left(\mathbb{E}\left[\left\|\boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau})-\boldsymbol{\lambda}^{(s-1)}\right\|^2 \right] - \mathbb{E}\left[\left\|\boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau})-\boldsymbol{\lambda}^{(s)}\right\|^2 \right]\right) \label{eq: nonconvex lm3 1}\\
&\quad\leq \sum_{s=c\sqrt{S}+1}^{(c+1)\sqrt{S}} \sum_{t=(s-1)\tau+1}^{s\tau}(2\sqrt{S}\tau \eta G_w \sqrt{G_w^2+\sigma_w^2})+ \gamma \tau\frac{\sqrt{S}\sigma_{\lambda}^2}{m}+ \gamma \tau{\sqrt{S}G_{\lambda}^2} + \frac{D_{\Lambda}^2}{2\gamma}\label{eq: nonconvex lm3 2}\\
&\quad \leq 2S\tau^2 \eta G_w \sqrt{G_w^2+\sigma_w^2}+ \gamma \tau^2\frac{\sqrt{S}\sigma_{\lambda}^2}{m}+ \gamma \tau^2{\sqrt{S}G_{\lambda}^2} + \frac{D_{\Lambda}^2}{2\gamma},\label{eq: nonconvex lm3 3}
\end{align}
where from (\ref{eq: nonconvex lm3 1}) to (\ref{eq: nonconvex lm3 2}) we use the $G_w$-Lipschitz property of $F(\cdot,\boldsymbol{\lambda})$ so that $F(\bm{w}^{t_1}, \boldsymbol{\lambda}) - F(\bm{w}^{t_2}, \boldsymbol{\lambda}) \leq G_w \|\bm{w}^{t_1} - \bm{w}^{t_2}\|$.
Now plugging (\ref{eq: nonconvex lm3 3}) back to (\ref{eq: nonconvex lm3 0}) yields:
\begin{align}
\frac{1}{T} \sum_{t=1}^{T}\left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)} )\right]\right) & \leq \frac{1}{T} 2\sqrt{S}S\tau^2 \eta G_w \sqrt{G_w^2+\sigma_w^2}+ \gamma \tau\frac{\sigma_{\lambda}^2}{m} + \gamma \tau{ G_{\lambda}^2}+ \frac{\sqrt{S}D_{\Lambda}^2}{2T\gamma}\nonumber\\
& \leq 2\sqrt{S} \tau \eta G_w \sqrt{G_w^2+\sigma_w^2}+ \gamma \tau\frac{\sigma_{\lambda}^2}{m}+ \gamma \tau{ G_{\lambda}^2} + \frac{ D_{\Lambda}^2}{2\sqrt{S}\tau\gamma}.\nonumber
\end{align}
\end{proof}
\end{lemma}
\subsection{Proof of Theorem~\ref{thm:nonconvex-linear}}
Now we proceed to the formal proof of Theorem~\ref{thm:nonconvex-linear}. Re-arranging terms in Lemma~\ref{lm: nonconvex lm1}, summing over $t = 1$ to $T$, and dividing by $T$ yields:
\begin{align}
\frac{1}{T}\sum_{t=1}^T \mathbb{E}\left[\left\|\nabla \Phi_{1/2L}\right.\right. & \left.\left. (\bm{w}^{(t)})\right\|^2\right] \nonumber\\
&\leq \frac{4}{\eta T}\mathbb{E}[\Phi_{1/2L} (\bm{w}^{(0)})] + \frac{1}{2T}\sum_{t=1}^T D_{\mathcal{W}}L^2 \mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \left\| \bm{w}^{(t)}_i - \bm{w}^{(t)} \right\| \right] \nonumber \\
& \quad + L \frac{1}{2T}\sum_{t=1}^T\left( \mathbb{E}[\Phi(\bm{w}^{(t)})]- \mathbb{E}[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{\floor{\frac{t}{\tau}}})]\right).\nonumber
\end{align}
Plugging in Lemma~\ref{lemma: deviation2} and \ref{lm: nonconvex lm3}
yields:
\begin{align}
\frac{1}{T}\sum_{t=1}^T \mathbb{E}\left[\left\|\nabla \Phi_{1/2L}(\bm{w}^{(t)})\right\|^2\right]
&\leq \frac{4}{\eta T}\mathbb{E}[\Phi_{1/2L} (\bm{w}^{(0)})] + \eta \tau D_{\mathcal{W}}L^2 \left(\sigma_w +\frac{\sigma_w }{m} + \sqrt{\Gamma} \right). \nonumber \\
& \quad + \frac{L}{2 } \left( 2\sqrt{S} \tau \eta G_w \sqrt{G_w^2+\sigma_w^2}+ \gamma \tau\frac{\sigma_{\lambda}^2}{m}+ \gamma \tau{ G_{\lambda}^2} + \frac{ D_{\Lambda}^2}{2\sqrt{S}\tau\gamma}\right)\nonumber\\
& \leq \frac{4}{\eta T}\mathbb{E}[\Phi_{1/2L} (\bm{w}^{(0)})] + \eta \tau D_{\mathcal{W}}L^2 \left(\sigma_w +\frac{\sigma_w }{m} + \sqrt{\Gamma} \right) \nonumber \\
& \quad + \sqrt{S} \tau \eta G_w L\sqrt{G_w^2+\sigma_w^2}+ \gamma \tau\frac{\sigma_{\lambda}^2L}{2m}+ \gamma \tau\frac{ G_{\lambda}^2 L}{2} + \frac{ D_{\Lambda}^2L}{4\sqrt{S}\tau\gamma}.\nonumber
\end{align}
Plugging in $\eta = \frac{1}{4LT^{3/4}}$ , $\gamma = \frac{1}{T^{1/2}}$ and $\tau = T^{1/4}$ we recover the convergence rate as cliamed:
\begin{align}
\frac{1}{T}\sum_{t=1}^T \mathbb{E}\left[\left\|\nabla \Phi_{1/2L} (\bm{w}^{(t)})\right\|^2\right]
& \leq \frac{4}{ T^{1/4}}\mathbb{E}[\Phi_{1/2L} (\bm{w}^{(0)})] + \frac{L^2}{T^{1/2}}\left(\sigma_w +\frac{\sigma_w }{m} + \sqrt{\Gamma} \right) \nonumber \\
& \quad + \frac{1}{T^{1/8}}G_w L\sqrt{G_w^2+\sigma_w^2}+ \frac{\sigma_{\lambda}^2L}{2m T^{1/4}}+ \frac{G_{\lambda}^2L}{2 T^{1/4}} + \frac{ D_{\Lambda}^2L}{4 T^{1/8}}, \nonumber
\end{align}
which concludes the proof.
\qed
\section{Proof of Convergence of {\sffamily{DRFA-Prox}}}
This section is devoted to the proof of convergence of {\sffamily{DRFA-Prox}} algorithm in both convex and nonconvex settings.
\subsection{Convex Setting}\label{sec: proof DRFA-Prox convex}
In this section we are going to provide the proof of Theorem~\ref{thm: regularized convex-linear}, the convergence of {\sffamily{DRFA-Prox}} on convex losses, i.e., global objective $F$ is convex in $\bm{w}$. Let us first introduce a key lemma:
\begin{lemma}\label{lemma: DRFA-prox convex-concave one iteration}
For {\sffamily{DRFA-Prox}}, $\forall \boldsymbol{\lambda} \in \Lambda$, and for any $s$ such that $0 \leq s\leq \frac{T}{\tau}-1$ we have:
\begin{align}
&\sum_{t=s\tau+1}^{(s+1)\tau}\left( \mathbb{E}\left[ F(\bm{w}^{(t)},\boldsymbol{\lambda} )\right] - \mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)} )\right] \right) \nonumber\\
&\leq - \frac{1}{2\gamma}\mathbb{E}\|\boldsymbol{\lambda}^{(s+1)} -\boldsymbol{\lambda}\|^2 + \frac{1}{2\gamma} \mathbb{E}[\| \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}\|^2]+ \frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]\nonumber\\
& \quad +\tau^2 \gamma G_w(G_w + \sqrt{G_w^2 +G_\lambda^2+ \sigma_\lambda^2})+\tau^2 \gamma G_\lambda^2\nonumber
\end{align}
\end{lemma}
\begin{proof}
Recall that to update $\boldsymbol{\lambda}^{(s)}$, we sampled a index $t'$ from $s\tau+1$ to $(s+1)\tau$, and obtain the averaged model $\bm{w}^{(t')}$. Now, consider iterations from $s\tau+1$ to $(s+1)\tau$.
Define following function:
\begin{align}
\Psi(\bm{u}) &= \tau f(\bm{w}^{(t')},\bm{y}) + \tau g(\bm{u}) - \frac{1}{2\gamma}\|\bm{y}+ \Delta_s - \bm{u}\|^2 \nonumber \\
&= \tau f(\bm{w}^{(t')},\bm{y})+ \tau g(\bm{u}) - \frac{1}{2\gamma}\|\bm{y}+\bar{\Delta}_s - \bm{u}\|^2 - \frac{1}{2\gamma}\| \bar{\Delta}_s - \Delta_s\|^2 \\
& \quad + \frac{1}{\gamma}\langle \bar{\Delta}_s - \Delta_s,\bm{y} +\bar{\Delta}_s - \bm{u} \rangle\nonumber.
\end{align}
By taking the expectation on both side, we get:
\begin{align}
&\mathbb{E}[\Psi(\bm{u})]\nonumber\\
&= \mathbb{E}[\tau f(\bm{w}^{(t')},\bm{y})] + \frac{1}{\gamma}\mathbb{E}[\left \langle \bar{\Delta}_s , \bm{u}-\bm{y} \right\rangle] + \mathbb{E}[\tau g(\bm{u})] - \frac{1}{2\gamma}\mathbb{E}\|\bm{u}-\bm{y}\|^2 - \frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s - \Delta_s\|^2 - \frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s \|^2 \nonumber \\
&= \mathbb{E}\left[\sum_{t=s\tau+1}^{(s+1)\tau} F(\bm{w}^{(t)},\bm{u})\right] - \frac{1}{2\gamma}\mathbb{E}\|\bm{u}-\bm{y}\|^2 - \frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s - \Delta_s\|^2-\frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s \|^2 \nonumber
\end{align}
where we used the fact that $\mathbb{E}[\tau f(\bm{w}^{(t')},\bm{y})] = \mathbb{E}\left[\sum_{t=s\tau+1}^{(s+1)\tau} f(\bm{w}^{(t)},\bm{y})\right]$ and $\frac{1}{\gamma}\mathbb{E}[\left \langle \Delta_s , \bm{u}-\bm{y} \right\rangle] =\sum_{t=s\tau+1}^{(s+1)\tau}\mathbb{E}\left[ f(\bm{w}^{(t)},\bm{u}) - f(\bm{w}^{(t)},\bm{y})\right] $.
Define the operator:
\begin{align}
T_{g}(\boldsymbol{y} ) := \arg \max_{\bm{u}\in \Lambda} \left\{\tau g(\bm{u}) - \frac{1}{2\gamma}\|\boldsymbol{y} +\Delta_s - \bm{u} \|^2\right\}
\end{align}
Since $\Psi(\bm{u})$ is $\frac{1}{2\gamma}$-strongly concave, and $T_g(\bm{y})$ is the maximizer of $\Psi(\bm{u})$, we have:
\begin{align}
\mathbb{E}[\Psi(T_g(\bm{y}) )] - \mathbb{E}[\Psi(\bm{u})] \geq \frac{1}{2\gamma}\mathbb{E}\|T_g(\bm{y})-\bm{u}\|^2 \nonumber
\end{align}
Notice that:
\begin{align}
\mathbb{E}[\Psi(T_g(\bm{y}) )] = \mathbb{E}\left[\sum_{t=s\tau+1}^{(s+1)\tau} F(\bm{w}^{(t)},T_g(\bm{y}) )\right] - \frac{1}{2\gamma} \mathbb{E}[\|T_g(\bm{y}) -\bm{y}\|^2] - \frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]-\frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s \|^2 \nonumber
\end{align}
So we know that $ \mathbb{E}\left[\sum_{t=s\tau+1}^{(s+1)\tau} F(\bm{w}^{(t)},T_g(\bm{y}) )\right]\geq \mathbb{E}[\Psi(T_g(\bm{y}) )] $, and hence:
\begin{align}
\mathbb{E}\left[\sum_{t=s\tau+1}^{(s+1)\tau} F(\bm{w}^{(t)},T_g(\bm{y}) )\right] - \mathbb{E}[\Psi(\bm{u})]\geq \mathbb{E}[\Psi(T_g(\bm{y}) )] - \mathbb{E}[\Psi(\bm{u})] \geq \frac{1}{2\gamma}\mathbb{E}\|T_g(\bm{y})-\bm{u}\|^2 \nonumber
\end{align}
Plugging in $\mathbb{E}[\Psi(\bm{u})]$ results in:
{\begin{align}
\mathbb{E}&\left[\sum_{t=s\tau+1}^{(s+1)\tau} F(\bm{w}^{(t)},T_g(\bm{y}) )\right] - \mathbb{E}[\Psi(\bm{u})]\nonumber \\
&= \mathbb{E}\left[\sum_{t=s\tau+1}^{(s+1)\tau} F(\bm{w}^{(t)},T_g(\bm{y}) )\right] -\left(\mathbb{E}\left[\sum_{t=s\tau+1}^{(s+1)\tau} F(\bm{w}^{(t)},\bm{u} )\right] - \frac{1}{2\gamma} \mathbb{E}[\|\bm{u} -\bm{y}\|^2] - \frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2] - \frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s \|^2 \right)\nonumber\\
&\geq \frac{1}{2\gamma}\mathbb{E}\|T_g(\bm{y})-\bm{u}\|^2.\nonumber
\end{align}}
Re-arranging the terms yields:
\begin{align}
\mathbb{E}\left[\sum_{t=s\tau+1}^{(s+1)\tau} F(\bm{w}^{(t)},\bm{u} )\right] & - \mathbb{E}\left[\sum_{t=s\tau+1}^{(s+1)\tau} F(\bm{w}^{(t)},T_g(\bm{y}) )\right] \nonumber\\
&\leq - \frac{1}{2\gamma}\mathbb{E}\|T_g(\bm{y})-\bm{u}\|^2 + \frac{1}{2\gamma} \mathbb{E}[\| \bm{y}-\bm{u}\|^2]+ \frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]+\frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s \|^2 \label{eq: proof DRFA-Prox convex eq2} .
\end{align}
Let $\bm{u} = \boldsymbol{\lambda}$, $\bm{y} = \boldsymbol{\lambda}^{(s)}$, then we have:
\begin{align}
&\sum_{t=s\tau+1}^{(s+1)\tau}\left( \mathbb{E}\left[ F(\bm{w}^{(t)},\boldsymbol{\lambda} )\right] - \mathbb{E}\left[ F(\bm{w}^{(t)},T_g(\boldsymbol{\lambda}^{(s)}) )\right] \right) \nonumber\\
&\leq - \frac{1}{2\gamma}\mathbb{E}\|T_g( \boldsymbol{\lambda}^{(s)})-\boldsymbol{\lambda}\|^2 + \frac{1}{2\gamma} \mathbb{E}[\| \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}\|^2]+ \frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]+\frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s \|^2 .\nonumber
\end{align}
Since $T_g(\boldsymbol{\lambda}^{(s)}) = \boldsymbol{\lambda}^{(s+1)}$, we have:
\begin{align}
&\sum_{t=s\tau+1}^{(s+1)\tau}\left( \mathbb{E}\left[ F(\bm{w}^{(t)},\boldsymbol{\lambda} )\right] - \mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)} )\right] \right) \nonumber\\
&\leq - \frac{1}{2\gamma}\mathbb{E}\|T_g( \boldsymbol{\lambda}^{(s)})-\boldsymbol{\lambda}\|^2 + \frac{1}{2\gamma} \mathbb{E}[\| \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}\|^2]+ \frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]+\frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s \|^2 \nonumber\\
& \quad +\underbrace{\sum_{t=s\tau+1}^{(s+1)\tau}\left( \mathbb{E}\left[ F(\bm{w}^{(t)},\boldsymbol{\lambda}^{(s+1)} )\right] - \mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)} )\right] \right)}_{T_1}.\nonumber
\end{align}
Now our remaining task is to bound $T_1$. By the Lipschitz property of $F$, we have the following upper bound for $T_1$:
\begin{align}
T_1 \leq \tau G_w\mathbb{E}\|\boldsymbol{\lambda}^{(s+1)}- \boldsymbol{\lambda}^{(s)}\|\label{eq: proof DRFA-Prox convex eq3}.
\end{align}
Then, by plugging $\bm{u} = \boldsymbol{\lambda}^{(s)}$, $\bm{y} = \boldsymbol{\lambda}^{(s)}$ into (\ref{eq: proof DRFA-Prox convex eq2}), we have the following lower bound:
\begin{align}
\frac{1}{2\gamma}\mathbb{E}\|\boldsymbol{\lambda}^{(s+1)} - \boldsymbol{\lambda}^{(s)}\|^2 -\frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2] -\frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s \|^2 \leq T_1 \label{eq: proof DRFA-Prox convex eq4}.
\end{align}
Combining (\ref{eq: proof DRFA-Prox convex eq3}) and (\ref{eq: proof DRFA-Prox convex eq4}) we have:
\begin{align}
\frac{1}{2\gamma}\mathbb{E}\|\boldsymbol{\lambda}^{(s+1)} - \boldsymbol{\lambda}^{(s)}\|^2 &-\frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]-\frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s \|^2 \nonumber\\
& \leq \tau G_w\mathbb{E}\|\boldsymbol{\lambda}^{(s+1)} - \boldsymbol{\lambda}^{(s)}\|\leq \tau G_w\sqrt{\mathbb{E}\|\boldsymbol{\lambda}^{(s+1)} - \boldsymbol{\lambda}^{(s)}\|^2}\label{eq: proof DRFA-Prox convex eq5}.
\end{align}
Let $X = \sqrt{\mathbb{E}\|\boldsymbol{\lambda}^{(s+1)} - \boldsymbol{\lambda}^{(s)}\|^2}$, $A = \frac{1}{2\gamma}$, $B = -\tau G_w$ and $C =- \frac{1}{2\gamma}\mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]-\frac{1}{2\gamma}\mathbb{E}\|\bar{\Delta}_s \|^2 $, then we can re-formulate (\ref{eq: proof DRFA-Prox convex eq5}) as:
\begin{align}
AX^2 + BX + C \leq 0.
\end{align}
Obviously $A\geq 0$. According to the root of quadratic equation, we know that:
\begin{align}
X \leq \frac{-B+\sqrt{B^2-4AC}}{2A} &= \frac{\tau G_w + \sqrt{G_w^2\tau^2 + \frac{1}{\gamma^2}(\mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]+ \mathbb{E}\|\bar{\Delta}_s \|^2) }}{1/\gamma} \nonumber \\
&\leq \tau \gamma \left(G_w + \sqrt{G_w^2+G_\lambda^2 + \sigma_\lambda^2}\right).\nonumber
\end{align}
Hence, we have
\begin{align}
T_1 \leq \tau G_w \mathbb{E}\|\boldsymbol{\lambda}^{(s+1)}- \boldsymbol{\lambda}^{(s)}\| \leq \tau^2 \gamma G_w\left(G_w + \sqrt{G_w^2+G_\lambda^2 + \sigma_\lambda^2}\right),\nonumber
\end{align}
which concludes the proof.
\end{proof}
\noindent\textbf{Proof of Theorem~\ref{thm: regularized convex-linear}.}~We start the proof by noting that $\forall \bm{w} \in \mathcal{W}$, $\forall \boldsymbol{\lambda} \in \Lambda$, according the convexity in $\bm{w} $ and concavity in $\boldsymbol{\lambda}$, we have:
\begin{align}
&\mathbb{E}[F(\hat{\bm{w} },\boldsymbol{\lambda} ) - \mathbb{E}[F(\bm{w} ,\hat{\boldsymbol{\lambda}})]\nonumber\\
& \leq \frac{1}{T}\sum_{t=1}^T \left\{ \mathbb{E}\left[F( \bm{w}^{(t)},\boldsymbol{\lambda} )\right] - \mathbb{E}\left[F(\bm{w} ,\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] \right\}\nonumber\\
& \leq \frac{1}{T}\sum_{t=1}^T\left \{ \mathbb{E}\left[F( \bm{w}^{(t)},\boldsymbol{\lambda} )\right] -\mathbb{E}\left[F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] +\mathbb{E}\left[ F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] - \mathbb{E}\left[F(\bm{w},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right] \right\}\nonumber\\
& \leq \frac{1}{T}\sum_{s=0}^{S-1} \sum_{t=s\tau+1}^{(s+1)\tau}\mathbb{E}[F( \bm{w}^{(t)},\boldsymbol{\lambda} ) -F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(s)})] +\frac{1}{T}\sum_{t=1}^T\mathbb{E}[ F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) - F(\bm{w} ,\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) ]. \label{eq: proof DRFA-Prox convex eq1}
\end{align}
To bound the first term in (\ref{eq: proof DRFA-Prox convex eq1}), plugging Lemma~\ref{lemma: bounded variance of lambda} into Lemma~\ref{lemma: DRFA-prox convex-concave one iteration}, and summing over $s=0$ to $S-1$ where $S=T/\tau$, and dividing both sides with $T$ yields:
\begin{align}
\frac{1}{T}\sum_{s=0}^{S-1}\sum_{t=s\tau+1}^{(s+1)\tau}\left\{ \right. & \left.\mathbb{E}\left[ F(\bm{w}^{(t)},\boldsymbol{\lambda} )\right]- \mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)} )\right]\right\} \nonumber\\
& \leq \frac{1}{2\gamma T} D_{\Lambda}^2+ \frac{1}{2\gamma\tau} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]+\tau \gamma G_w(G_w + \sqrt{G_w^2+G_\lambda^2 + \sigma_\lambda^2})+\gamma \tau G_\lambda^2 \nonumber\\
&\leq \frac{1}{2\gamma T} D_{\Lambda}^2+ \frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]+\tau \gamma (G_w + \sqrt{G_w^2+G_\lambda^2 + \sigma_\lambda^2})+\gamma \tau G_\lambda^2\nonumber\\
& \leq \frac{D_{\Lambda}^2}{2\gamma T} + \frac{\gamma \tau\sigma_{ \lambda}^2 }{2m}+\tau \gamma G_w(G_w + \sqrt{G_w^2 +G_\lambda^2+ \sigma_\lambda^2})+\gamma \tau G_\lambda^2.\nonumber
\end{align}
To bound the second term in (\ref{eq: proof DRFA-Prox convex eq1}), we plug Lemma~\ref{lemma: bounded variance of w} and Lemma~\ref{lemma: deviation} into Lemma~\ref{lemma: one iteration w} and apply the telescoping sum from $t = 1$ to $T$ to get:
\begin{align}
\frac{1}{T}\sum_{t=1}^T \mathbb{E}[F(\bm{w} ^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) & -F(\bm{w} ,\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})]\nonumber\\
& \leq \frac{1}{2T\eta}\mathbb{E}\|\ \bm{w}^{(0)}- \bm{w} \|^2 + 5L\eta^2\tau^2 \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right) + \frac{\eta G_w^2}{2} + \frac{\eta \sigma_{w }^2 }{2m}\nonumber\\
& \leq \frac{D_{\mathcal{W}}^2}{2T\eta} + 5L\eta^2\tau^2 \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right) + \frac{\eta G_w^2}{2} + \frac{\eta \sigma_{w }^2 }{2m},\nonumber
\end{align}
So that we can conclude:
\begin{align}
\mathbb{E}[F(\hat{\bm{w} },\boldsymbol{\lambda} ) - \mathbb{E}[F(\bm{w} ,\hat{\boldsymbol{\lambda}})]
& \leq \frac{D_{\mathcal{W}}^2}{2T\eta} + 5L\eta^2\tau^2 \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right) + \frac{\eta G_w^2}{2} + \frac{\eta \sigma_{w }^2 }{2m} + \frac{D_{\Lambda}^2}{2\gamma T} \nonumber\\
& \quad +\gamma\tau G_\lambda^2+ \frac{\gamma \tau\sigma_{ \lambda}^2 }{2m}+\tau \gamma G_w(G_w + \sqrt{G_w^2 + G_\lambda^2+\sigma_\lambda^2}).\nonumber
\end{align}
Since the RHS does not depend on $\bm{w}$ and $\boldsymbol{\lambda}$, we can maximize over $\boldsymbol{\lambda}$ and minimize over $\bm{w}$ on both sides:
\begin{align}
&\min_{\bm{w}\in \mathcal{W}}\max_{\boldsymbol{\lambda}\in \Lambda} \mathbb{E}[F(\hat{\bm{w} },\boldsymbol{\lambda} ) - \mathbb{E}[F(\bm{w} ,\hat{\boldsymbol{\lambda}})]\nonumber\\
& \leq \frac{D_{\mathcal{W}}^2}{2T\eta} + 5L\eta^2\tau^2 \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right) + \frac{\eta G_w^2}{2} + \frac{\eta \sigma_{w }^2 }{2m} + \frac{D_{\Lambda}^2}{2\gamma T} \nonumber\\
& \quad +\gamma\tau G_\lambda^2+ \frac{\gamma \tau\sigma_{ \lambda}^2 }{2m}+\tau \gamma G_w\left(G_w + \sqrt{G_w^2 + G_\lambda^2+\sigma_\lambda^2}\right).\nonumber
\end{align}
Plugging in $\tau = \frac{T^{1/4}}{\sqrt{m}}$, $\eta = \frac{1}{4L \sqrt{T}}$, and $\gamma = \frac{1}{T^{5/8}}$, we get:
\begin{equation}
\begin{aligned}
\max_{\boldsymbol{\lambda}\in \Lambda}\mathbb{E}[F(\hat{\boldsymbol{w}},\boldsymbol{\lambda} )] -\min_{\bm{w}\in\mathcal{W}} \mathbb{E}[F(\boldsymbol{w} ,\hat{\boldsymbol{\lambda}} )] \leq O\Big{(}\frac{D_{\mathcal{W}}^2+G_{w}^2}{\sqrt{T}} +\frac{D_{\Lambda}^2+G_w^2}{T^{3/8}}+\frac{G_\lambda^2 }{m^{1/2}T^{3/8}} +\frac{\sigma_{\lambda}^2}{m^{3/2}T^{3/8}}+ \frac{\sigma_w^2+\Gamma}{m \sqrt{T} }\Big{)},\nonumber
\end{aligned}
\end{equation}
thus concluding the proof.
\subsection{Nonconvex Setting} \label{sec: proof DRFA-Prox nonconvex}
In this section we are going to prove Theorem~\ref{thm:regularized nonconvex-linear}. The whole framework is similar to the proof of Theorem~\ref{thm: regularized convex-linear}, but to bound $\mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})} )\right]$ term, we employ different technique for proximal method. The following lemma characterize the bound of $\mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})} )\right]$:
\begin{lemma}\label{lemma: DRFA-prox nonconvex dual iteration}
For {\sffamily{DRFA-Prox}}, under Theorem~\ref{thm:regularized nonconvex-linear}'s assumption, the following statement holds true:
\begin{align}
&\frac{1}{T} \sum_{t=1}^{T} \mathbb{E}\left[\Phi(\bm{w}^{(t)} )- F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})} )\right] \nonumber\\
&\leq 2\sqrt{S}\tau \eta G_w \sqrt{G_w^2+\sigma_w^2}+ \gamma \tau \frac{ \sigma_{\lambda}^2}{2m}+ \gamma \tau \frac{ G_{\lambda}^2}{2} + \frac{D_{\Lambda}^2}{2\sqrt{S}\tau }+\tau \gamma G_w\left(G_w + \sqrt{G_w^2 +G_\lambda^2+ \sigma_\lambda^2}\right). \nonumber
\end{align}
\begin{proof}
We recall that in Lemma~\ref{lemma: DRFA-prox convex-concave one iteration}, we have:
{\begin{align}
&\sum_{t=s\tau+1}^{(s+1)\tau} \left(\mathbb{E}\left[ F(\bm{w}^{(t)},\boldsymbol{\lambda} )\right] - \mathbb{E}\left[ F(\bm{w}^{(t)},\boldsymbol{\lambda}^{(s)} )\right]\right)\nonumber\\
&\leq - \frac{1}{2\gamma}\mathbb{E}\| \boldsymbol{\lambda}^{(s+1)}-\boldsymbol{\lambda}\|^2 + \frac{1}{2\gamma} \mathbb{E}[\| \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}\|^2]+ \frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]+\frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s \|^2] \nonumber \\ &\quad +\tau^2 \gamma \left(G_w + \sqrt{G_w^2+G_\lambda^2 + \sigma_\lambda^2}\right).\nonumber
\end{align}}
Adding $\sum_{t=s\tau+1}^{(s+1)\tau} \mathbb{E}\left[ \Phi(\bm{w}^{(t)})\right] $ to both sides, and re-arranging the terms give:
{\begin{align}
&\sum_{t=s\tau+1}^{(s+1)\tau} \left(\mathbb{E}\left[ \Phi(\bm{w}^{(t)})\right] \right. \left. - \mathbb{E}\left[ F( \bm{w} ^{(t)},\boldsymbol{\lambda}^{(s)} )\right] \right) \nonumber \\ &\leq \sum_{t=s\tau+1}^{(s)\tau} \left(\mathbb{E}\left[ \Phi(\bm{w}^{(t)})\right] -\mathbb{E}\left[ F( \bm{w}^{(t)},\boldsymbol{\lambda} )\right]\right) - \frac{1}{2\gamma}\mathbb{E}\| \boldsymbol{\lambda}^{(s+1)}-\boldsymbol{\lambda}\|^2\nonumber\\
& \quad + \frac{1}{2\gamma} \mathbb{E}[\| \boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}\|^2]+ \frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s- \Delta_s\|^2]+\frac{1}{2\gamma} \mathbb{E}[\|\bar{\Delta}_s \|^2] +\tau^2 \gamma G_w \left(G_w + \sqrt{G_w^2+G_\lambda^2 + \sigma_\lambda^2}\right).\nonumber
\end{align}}
Then, we follow the same procedure as in Lemma~\ref{lm: nonconvex lm3}. Without loss of generality we assume $\sqrt{S}$ is an integer, so we can equally divide index $0$ to $S-1$ into $\sqrt{S}$ groups. Then we examine one block by summing $s$ from $s = c\sqrt{S}$ to $(c+1)\sqrt{S}-1$, and set $\boldsymbol{\lambda} = \boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau})$:
\begin{align}
\sum_{s=c\sqrt{S}}^{(c+1)\sqrt{S}-1}&\sum_{t=s\tau+1}^{(s+1)\tau} \left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)} )\right]\right) \nonumber\\
&\leq \sum_{s=c\sqrt{S}}^{(c+1)\sqrt{S}-1} \sum_{t=s\tau+1}^{(s+1)\tau} \left(\mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{t}) )\right]- \mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau}) )\right] \right) \nonumber\\
& \quad +\sqrt{S}\tau^2 \gamma G_w (G_w + \sqrt{G_w^2+G_\lambda^2 + \sigma_\lambda^2}) + \gamma \tau^2 \frac{\sqrt{S}\sigma_{\lambda}^2}{2m} + \gamma \tau^2 \frac{\sqrt{S}G_{\lambda}^2}{2} \nonumber\\
& \quad + \frac{1}{2\gamma }\sum_{s=c\sqrt{S}}^{(c+1)\sqrt{S}-1}\left(\mathbb{E}\left[\left\|\boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau})-\boldsymbol{\lambda}^{(s)}\right\|^2 \right] - \mathbb{E}\left[\left\|\boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau})-\boldsymbol{\lambda}^{(s+1)}\right\|^2 \right]\right)\nonumber
\end{align}
Adding and subtracting $\mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{t}) )\right]- \mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau}) )\right] $ yields:
\begin{align}
\sum_{s=c\sqrt{S}}^{(c+1)\sqrt{S}-1}&\sum_{t=s\tau+1}^{(s+1)\tau} \left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)} )\right]\right) \nonumber\\
&\leq \sum_{s=c\sqrt{S}}^{(c+1)\sqrt{S}-1} \sum_{t=s\tau+1}^{(s+1)\tau} \left(\mathbb{E}\left[ F(\bm{w}^{(t)}, \boldsymbol{\lambda}^*(\bm{w}^{t}) )\right]-\mathbb{E}\left[ F(\bm{w}^{((c+1)\sqrt{S}\tau)}, \boldsymbol{\lambda}^*(\bm{w}^{t}) )\right] \right.\nonumber\\
&\quad \left.+\mathbb{E}\left[ F(\bm{w}^{((c+1)\sqrt{S}\tau)}, \boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau}) )\right]- \mathbb{E}\left[ F(\bm{w}^{(t)},\boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau}))\right] \right) \nonumber\\
& \quad+ \gamma \tau^2\frac{\sqrt{S}\sigma_{\lambda}^2}{2m}+ \gamma \tau^2 \frac{\sqrt{S}G_{\lambda}^2}{2} +\sqrt{S}\tau^2 \gamma G_w(G_w + \sqrt{G_w^2 +G_\lambda^2+ \sigma_\lambda^2})\nonumber \\
&\quad + \frac{1}{2\gamma }\sum_{s=c\sqrt{S}+1}^{(c+1)\sqrt{S}}\left(\mathbb{E}\left[\left\|\boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau})-\boldsymbol{\lambda}^{(s)}\right\|^2 \right] - \mathbb{E}\left[\left\|\boldsymbol{\lambda}^*(\bm{w}^{(c+1)\sqrt{S}\tau})-\boldsymbol{\lambda}^{(s+1)}\right\|^2 \right]\right)\nonumber\\
&\leq \sum_{s=c\sqrt{S}}^{(c+1)\sqrt{S}-1} \sum_{t=s\tau+1}^{(s+1)\tau} (2\sqrt{S}\tau\eta G_w \sqrt{G_w^2+\sigma_w^2})+ \gamma \tau^2\frac{\sqrt{S}\sigma_{\lambda}^2}{2m}+ \gamma \tau^2 \frac{\sqrt{S}G_{\lambda}^2}{2} + \frac{D_{\Lambda}^2}{2\gamma }\nonumber\\
&\quad +\sqrt{S}\tau^2 \gamma G_w \left(G_w + \sqrt{G_w^2 +G_\lambda^2+ \sigma_\lambda^2}\right)\nonumber\\
& \leq 2S\tau^2 \eta G_w \sqrt{G_w^2+\sigma_w^2}+ \gamma \tau^2\frac{\sqrt{S} \sigma_{\lambda}^2}{2m}+ \gamma \tau^2 \frac{\sqrt{S}G_{\lambda}^2}{2} + \frac{ D_{\Lambda}^2}{2\gamma } \nonumber \\
& \quad +\sqrt{S}\tau^2 \gamma G_w\left(G_w + \sqrt{G_w^2+G_\lambda^2 + \sigma_\lambda^2}\right). \nonumber
\end{align}
So we can conclude that:
\begin{align}
&\sum_{s=c\sqrt{S}}^{(c+1)\sqrt{S}-1}\sum_{t=s\tau+1}^{(s+1)\tau} \left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)} )\right]\right) \nonumber\\
&\leq 2S\tau^2 \eta G_w \sqrt{G_w^2+\sigma_w^2}+ \gamma \tau^2\frac{\sqrt{S}\sigma_{\lambda}^2}{2m} + \gamma \tau^2\frac{\sqrt{S}G_{\lambda}^2}{2} \frac{D_{\Lambda}^2}{2\gamma } +\sqrt{S}\tau^2 \gamma G_w\left(G_w + \sqrt{G_w^2 +G_\lambda^2+ \sigma_\lambda^2}\right)\nonumber
\end{align}
Summing above inequality over $c$ from $0$ to $\sqrt{S}-1$, and dividing both sides by $T$ gives
\begin{align}
\frac{1}{T}\sum_{s=0}^{S-1}&\sum_{t=s\tau+1}^{(s+1)\tau} \left( \mathbb{E}\left[\Phi(\bm{w}^{(t)} )\right]- \mathbb{E}\left[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)} )\right]\right) \nonumber\\
&\leq 2\sqrt{S}\tau \eta G_w \sqrt{G_w^2+\sigma_w^2}+ \gamma \tau \frac{ \sigma_{\lambda}^2}{2m}+ \gamma \tau \frac{ G_{\lambda}^2}{2 } + \frac{D_{\Lambda}^2}{2\sqrt{S}\tau \gamma }+\tau \gamma G_w\left(G_w + \sqrt{G_w^2 +G_\lambda^2+ \sigma_\lambda^2}\right), \nonumber
\end{align}
which concludes the proof.
\end{proof}
\end{lemma}
\noindent\textbf{Proof of Theorem~\ref{thm:regularized nonconvex-linear}.}~Now we proceed to the formal proof of Theorem~\ref{thm:regularized nonconvex-linear}. Re-arranging terms in Lemma~\ref{lm: nonconvex lm1}, summing over $t = 1$ to $T$, and dividing by $T$ yields:
{
\begin{align}
&\frac{1}{T}\sum_{t=1}^T \mathbb{E}\left[\left\|\nabla \Phi_{1/2L} \right.\right. \left.\left. (\bm{w}^{(t)})\right\|^2\right] \nonumber\\
&\leq \frac{4}{\eta T}\mathbb{E}[\Phi_{1/2L} (\bm{w}^{(0)})] + \frac{1}{2T}\sum_{t=1}^T D_{\mathcal{W}}L^2 \mathbb{E}\left[ \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \left\| \bm{w}^{(t)}_i - \bm{w}^{(t)} \right\| \right] \nonumber \\
& \quad + L \frac{1}{2T}\sum_{t=1}^T\left( \mathbb{E}[\Phi(\bm{w}^{(t)})]- \mathbb{E}[F(\bm{w}^{(t)}, \boldsymbol{\lambda}^{\floor{\frac{t}{\tau}}})]\right).\nonumber
\end{align}}
Plugging in Lemmas~\ref{lemma: deviation2} and \ref{lemma: DRFA-prox nonconvex dual iteration}
yields:
\begin{align}
\frac{1}{T}\sum_{t=1}^T &\mathbb{E}\left[\left\|\nabla \Phi_{1/2L} (\bm{w}^{(t)})\right\|^2\right] \nonumber\\
&\leq \frac{4}{\eta T}\mathbb{E}[\Phi_{1/2L} (\bm{w}^{(0)})] + \eta \tau D_{\mathcal{W}}L^2 \left(\sigma_w +\frac{\sigma_w }{m} + \sqrt{\Gamma} \right). \nonumber \\
&\quad + \frac{L}{2 } \left( 2\sqrt{S}\tau \eta G_w \sqrt{G_w^2+\sigma_w^2}+ \gamma \tau \frac{ \sigma_{\lambda}^2}{2m}+ \gamma \tau \frac{ G_{\lambda}^2}{2} + \frac{D_{\Lambda}^2}{2\sqrt{S}\tau }+\tau\gamma G_w(G_w + \sqrt{G_w^2+G_\lambda^2 + \sigma_\lambda^2})\right)\nonumber\\
& \leq \frac{4}{\eta T}\mathbb{E}[\Phi_{1/2L} (\bm{w}^{(0)})] + \eta \tau D_{\mathcal{W}}L^2 \left(\sigma_w +\frac{\sigma_w }{m} + \sqrt{\Gamma} \right) \nonumber \\
&\quad + \sqrt{S} \tau \eta G_w L\sqrt{G_w^2+\sigma_w^2}+ \gamma \tau\frac{\sigma_{\lambda}^2L}{4m}+ \gamma \tau \frac{ G_{\lambda}^2 L}{4} + \frac{ D_{\Lambda}^2L}{4\sqrt{S}\gamma\tau}+\frac{\tau \gamma L G_w(G_w + \sqrt{G_w^2 +G_\lambda^2+ \sigma_\lambda^2})}{2}.\nonumber
\end{align}
Plugging in $\eta = \frac{1}{4LT^{3/4}}$ , $\gamma = \frac{1}{T^{1/2}}$ and $\tau = T^{1/4}$ we recover the stated convergence rate as:
\begin{align}
\frac{1}{T}\sum_{t=1}^T &\mathbb{E}\left[\left\|\nabla \Phi_{1/2L} \vphantom{\left\|\bm{w}^{(t)}\right\|^2} (\bm{w}^{(t)})\right\|^2\right] \nonumber\\
& \leq \frac{4}{ T^{1/4}}\mathbb{E}[\Phi_{1/2L} (\bm{w}^{(0)})] + \frac{L^2}{T^{1/2}}\left(\sigma_w +\frac{\sigma_w }{m} + \sqrt{\Gamma} \right) \nonumber \\
&+ \frac{1}{T^{1/8}}G_w L\sqrt{G_w^2+\sigma_w^2}+ \frac{\sigma_{\lambda}^2L}{4m T^{1/4}}+ \frac{G_{\lambda}^2L}{4 T^{1/4}} + \frac{ D_{\Lambda}^2L}{4 T^{1/8}}+\frac{ L G_w(G_w + \sqrt{G_w^2 +G_\lambda^2+ \sigma_\lambda^2})}{2T^{1/4}}. \nonumber
\end{align}
\qed
\section{Proof of Convergence of {\sffamily{DRFA-GA}} in Strongly-Convex-Strongly-Concave Setting}\label{sec:scsc}
In this section we proceed to the proof of the convergence in strongly-convex-strongly-concave setting (Theorem~\ref{theorem2}). In this section we abuse the notation and use the following definition for $\bar{\bm{u}}_t$:
\begin{equation}
\bar{\bm{u}}_t = \sum_{i=1}^N \lambda_i^{(\floor{\frac{t}{\tau}})} \nabla f_i(\bm{w}_i^{(t)}).\nonumber
\end{equation}
\subsection{Overview of the Proof}
We again start with the dynamic of one iteration:
\begin{align}
\mathbb{E}\left\|\bm{w} ^{(t+1)} - \bm{w} ^*\right\|^2 &\leq \left (1-\frac{\mu\eta}{2}\right) \mathbb{E}\left\|\bm{w} ^{(t)}- \bm{w} ^*\right\|^2 -\eta\mathbb{E}\left[\Phi( \bm{w}^{(t)})-\Phi(\bm{w} ^*)\right]\nonumber\\
& \quad + \eta^2\frac{2\sigma_w^2+4G_w^2}{m} 4L^2\left(\eta^2+\frac{\eta}{\mu}\right) \mathbb{E}\left[\delta^{(t)} \right] \nonumber\\
& \quad +4\left(\frac{\eta}{\mu}+\eta^2\right)\mathbb{E}\|\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})- \nabla\Phi( \bm{w}^{(t)})\|^2.\nonumber
\end{align}
In addition to the local-global deviation, in this case we also have a new term $\|\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})- \nabla\Phi( \bm{w}^{(t)})\|^2$. Recall that $\nabla\Phi( \bm{w}^{(t)})$ is the gradient evaluated at $\boldsymbol{\lambda}^*(\bm{w}^{(t)})$. A straightforward approach is to use the smoothness of $\Phi$, to convert the difference between gradient to the difference between $\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}$ and $\boldsymbol{\lambda}^*(\bm{w}^{(t)})$. By examining the dynamic of $\boldsymbol{\lambda}$, we can prove that:
\begin{align}
\mathbb{E}\left\|\boldsymbol{\lambda}^*( \bm{w}^{(t)}) - \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})} \right\|^2\leq 2\left(1-\frac{1}{2\kappa}\right)^{(\floor{\frac{t}{\tau}})}\mathbb{E}\left\|\boldsymbol{\lambda}^{(0)}- \boldsymbol{\lambda}^*(\bm{w} ^{(0)})\right\|^2 + 2(4\kappa^2+1)\kappa^2 \tau^2 \eta^2 G_{w}^2.\nonumber
\end{align}
Putting these pieces together, and unrolling the recursion will conclude the proof.
\subsection{Proof of Technical Lemmas}
\begin{lemma} [~\citet{lin2019gradient}. Properties of $\Phi (\cdot)$ and $\boldsymbol{\lambda}^*(\cdot)$] \label{danskin}
If $F(\cdot, \boldsymbol{\lambda})$ is $L$-smooth function and $F(\boldsymbol{w},\cdot)$ is $\mu$-strongly-concave, $L$-smooth function, let $\kappa = \frac{L}{\mu}$, then $\Phi (\boldsymbol{w})$ is $\alpha$-smooth function where $\alpha = L + \kappa L$ and $\boldsymbol{\lambda}^*(\boldsymbol{w})$ is $\kappa$-Lipschitz. Also $\nabla \Phi(\boldsymbol{w}) = \nabla_{\boldsymbol{w}}F(\boldsymbol{w},\boldsymbol{\lambda}^*(\boldsymbol{w}
))$.
\end{lemma}
\begin{lemma}
\label{lemma: one iteration SCSC}
For {\sffamily{DRFA-GA}}, under Theorem~\ref{theorem2}'s assumptions, the following holds true:
\begin{align}
\mathbb{E}\left\|\bm{w} ^{(t+1)} - \bm{w} ^*\right\|^2 &\leq \left (1-\frac{\mu\eta}{2}\right) \mathbb{E}\left\|\bm{w} ^{(t)}- \bm{w} ^*\right\|^2 -\eta\mathbb{E}\left[\Phi( \bm{w}^{(t)})-\Phi(\bm{w} ^*)\right]\nonumber\\
& \quad + \eta^2\frac{2\sigma_w^2+4G_w^2}{m} 4L^2\left(\eta^2+\frac{\eta}{\mu}\right) \mathbb{E}\left[\delta^{(t)} \right] \\ & \quad +4\left(\frac{\eta}{\mu}+\eta^2\right)\mathbb{E}\|\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})- \nabla\Phi( \bm{w}^{(t)})\|^2. \nonumber
\end{align}
\end{lemma}
\begin{proof}
According to Lemma B2 in~\cite{lin2020near}, if $F(\cdot,\boldsymbol{\lambda})$ is $\mu$-strongly-convex, then $\Phi(\cdot)$ is also $\mu$-strongly-convex. Noting this, from the strong convexity and the updating rule we have:
\begin{align}
\mathbb{E} \|\bm{w} ^{(t+1)}& - \bm{w} ^*\|^2 \nonumber\\
&= \mathbb{E}\left\|\prod_{\mathcal{W}}\left(\bm{w} ^{(t)}- \eta \bm{u}^{(t)}\right)- \bm{w} ^*\right\|^2 \leq \mathbb{E}\|\bm{w} ^{(t)}- \eta \Bar{\bm{u}}^{(t)} - \bm{w} ^*\|^2 + \eta^2 \mathbb{E}\|\Bar{\bm{u}}^{(t)} - \bm{u}^{(t)}\|^2 \nonumber\\
& = \mathbb{E}\|\bm{w} ^{(t)}- \bm{w} ^*\|^2 + \underbrace{\mathbb{E}[- 2\eta \langle \Bar{\bm{u}}^{(t)},\bm{w} ^{(t)}- \bm{w} ^* \rangle]}_{T_1} + \underbrace{\eta^2 \mathbb{E}\|\Bar{\bm{u}}^{(t)}\|^2}_{T_2} + \eta^2 \mathbb{E}\|\Bar{\bm{u}}^{(t)} - \bm{u}^{(t)}\|^2 \label{l71}
\end{align}
First we are to bound the variance $\mathbb{E}\|\Bar{\bm{u}}^{(t)} - \bm{u}^{(t)}\|^2$:
\begin{align}
\mathbb{E}\|\Bar{\bm{u}}^{(t)} - \bm{u}^{(t)}\|^2 & = \mathbb{E}\left\|\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_i(\bm{w}_i^{(t)}) - \bar{\bm{u}}^{(t)}\right\|^2\nonumber\\
&= \mathbb{E}\left\|\frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_i(\bm{w}_i^{(t)};\xi_i^{(t)}) - \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \bar{\bm{u}}^{(t)}\right\|^2 \leq \frac{2\sigma_w^2 + 4G_w^2}{m},\nonumber
\end{align}
where we use the fact $Var(\sum_{i=1}^m \bm{X}_i) = \sum_{i=1}^m Var(\bm{X}_i)$ for independent variables $\bm{X}_i, i=1, \ldots, m$, and $Var(\nabla f_i(\bm{w}_i^{(t)};\xi_i^{(t)}) ) = \mathbb{E}\left\| \nabla f_i(\bm{w}_i^{(t)};\xi_i^{(t)}) - \bar{\bm{u}}^{(t)}\right\|^2 \leq 2\left\| \nabla f_i(\bm{w}_i^{(t)};\xi_i^{(t)})- \nabla f_i(\bm{w}_i^{(t)} )\right\|^2+2\left\|\nabla f_i(\bm{w}_i^{(t)} )- \bar{\bm{u}}^{(t)}\right\|^2\leq 2\sigma_w^2 + 4G_w^2$.
Then we switch to bound $T_1$:
\begin{align}
T_1 &= 2\eta\mathbb{E}\left[-\left\langle \nabla \Phi( \bm{w}^{(t)}), \bm{w}^{(t)} - \bm{w} ^* \right\rangle + \left\langle \nabla \Phi( \bm{w}^{(t)}) - \Bar{\bm{u}}^{(t)}, \bm{w}^{(t)} - \bm{w} ^* \right\rangle\right] \nonumber\\
& \leq 2\eta\mathbb{E}\left[-(\Phi( \bm{w}^{(t)})- \Phi(\bm{w} ^*))-\frac{\mu}{2}\| \bm{w}^{(t)} - \bm{w} ^*\|^2+ \frac{1}{\mu}\| \nabla
\Phi( \bm{w}^{(t)})-\Bar{\bm{u}}^{(t)}\|^2+ \frac{\mu}{4}\| \bm{w}^{(t)} - \bm{w} ^* \|^2\right] \nonumber\\
& \leq \mathbb{E}\left[-2\eta(\Phi( \bm{w}^{(t)})- \Phi(\bm{w} ^*))-\frac{\mu\eta}{2}\| \bm{w}^{(t)} - \bm{w} ^*\|^2+ \frac{2\eta}{\mu}\| \nabla
\Phi( \bm{w}^{(t)})-\Bar{\bm{u}}^{(t)}\|^2\right] \nonumber\\
& \leq \mathbb{E}\left[-2\eta(\Phi( \bm{w}^{(t)})- \Phi(\bm{w} ^*))-\frac{\mu\eta}{2}\| \bm{w}^{(t)} - \bm{w} ^*\|^2+ \frac{4\eta}{\mu}\left\| \nabla
\Phi( \bm{w}^{(t)})-\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right\|^2 \right. \nonumber\\
& \left. \quad + \frac{4\eta}{\mu}\| \nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})-\Bar{\bm{u}}^{(t)}\|^2\right]\nonumber\\
& \leq \mathbb{E}\left[-2\eta(\Phi( \bm{w}^{(t)})- \Phi(\bm{w} ^*))-\frac{\mu\eta}{2}\| \bm{w}^{(t)} - \bm{w} ^*\|^2+ \frac{4\eta}{\mu}\left\| \nabla
\Phi( \bm{w}^{(t)})-\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right\|^2 \right. \nonumber\\
& \left.\quad + \frac{4L^2\eta}{\mu}\sum_{k=1}^K \lambda^{(\floor{\frac{t}{\tau}})}_i\| \bm{w}^{(t)}-\bm{w} ^{(t)}_i\|^2\right],\nonumber
\end{align}
where in the second step we use the arithmetic and geometric inequality and the strong convexity of $\Phi(\cdot)$; and at the last step we use the smoothness, the convexity of $\|\cdot\|^2$ and Jensen's inequality.
Then, we can bound $T_2$ as:
\begin{align}
T_2 &\leq \eta^2\mathbb{E}\left[4\left\|\Bar{\bm{u}}^{(t)} - \nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right\|^2 + 4\|\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})- \nabla\Phi( \bm{w}^{(t)})\|^2 \right.\nonumber\\
& \qquad \qquad \qquad\left. +2 \left\|\nabla\Phi( \bm{w}^{(t)})\right\|^2\right] \nonumber\\
& \leq \eta^2\mathbb{E}\left[4\left\|\Bar{\bm{u}}^{(t)} - \nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right\|^2 + 4\left\|\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})- \nabla\Phi( \bm{w}^{(t)})\right\|^2 \right. \nonumber\\
& \qquad\qquad\qquad \left. + 4\alpha (\Phi( \bm{w}^{(t)})-\Phi(\bm{w} ^*)) \vphantom{\left\|\nabla\Phi( \bm{w}^{(t)})\right\|^2} \right]\nonumber\\
& \leq \eta^2\mathbb{E}\left[4L^2\sum_{i=1}^N \lambda^{(\floor{\frac{t}{\tau}})}_i\| \bm{w}^{(t)}-\bm{w} ^{(t)}_i\|^2 + 4\|\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})- \nabla\Phi( \bm{w}^{(t)})\|^2 \right.\nonumber\\
& \qquad\qquad\qquad \left. + 4\alpha(\Phi(\bm{w} _t)-\Phi(\bm{w} ^*)) \vphantom{\left\|\nabla\Phi( \bm{w}^{(t)})\right\|^2} \right ]\nonumber\\
& \leq \eta^2\mathbb{E}\left[4L^2 \frac{1}{m}\sum_{i\in\mathcal{D}^{(\floor{\frac{t}{\tau}})}} \| \bm{w}^{(t)}-\bm{w} ^{(t)}_i\|^2 + 4\|\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})- \nabla\Phi( \bm{w}^{(t)})\|^2 \right.\nonumber\\
& \qquad\qquad\qquad \left. + 4\alpha(\Phi(\bm{w} _t)-\Phi(\bm{w} ^*)) \vphantom{\left\|\nabla\Phi( \bm{w}^{(t)})\right\|^2} \right].\nonumber
\end{align}
Plugging $T_1$ and $T_2$ back to (\ref{l71}) results in:
\begin{align}
\mathbb{E}\left\|\bm{w} ^{(t+1)} - \bm{w} ^*\right\|^2 &\leq \left (1-\frac{\mu\eta}{2}\right) \mathbb{E}\left\|\bm{w} ^{(t)}- \bm{w} ^*\right\|^2 +(4\alpha \eta^2-2\eta)\mathbb{E}\left[\Phi( \bm{w}^{(t)})-\Phi(\bm{w} ^*)\right] \nonumber \\
& \quad + \eta^2\frac{2\sigma_w^2+4G_w^2}{m} + 4L^2\left(\eta^2+\frac{\eta}{\mu}\right) \mathbb{E}\left[\delta^{(t)} \right] \nonumber\\
& \quad +4\left(\frac{\eta}{\mu}+\eta^2\right)\mathbb{E}\|\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})- \nabla\Phi( \bm{w}^{(t)})\|^2 \label{l70}.
\end{align}
By choosing $\eta < \frac{1}{4\alpha}$, it holds that $(4\alpha \eta^2 - 2\eta) < -\eta$, therefore we conclude the proof.
\end{proof}
\begin{lemma}[Decreasing Optimal Gap of $\boldsymbol{\lambda}$]
\label{lemma: optimal gap}
For {\sffamily{DRFA-GA}}, if $F(\bm{w} ,\cdot)$ is $\mu$-strongly-concave, choosing $\gamma = \frac{1}{L}$, the optimality gap of $\boldsymbol{\lambda}$ is decreasing by the following recursive relation:
\begin{align}
\mathbb{E}\left\|\boldsymbol{\lambda}^*( \bm{w}^{(t)}) - \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})} \right\|^2\leq 2\left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}}\mathbb{E}\left\|\boldsymbol{\lambda}^{(0)}- \boldsymbol{\lambda}^*(\bm{w} ^{(0)})\right\|^2 + 2(4\kappa^2+1)\kappa^2 \tau^2 \eta^2 G_{w}^2.\nonumber
\end{align}
\end{lemma}
\emph{Proof:}
Assume $s\tau+1 \leq t\leq (s+1)\tau$. By the Jensen's inequality:
\begin{align}
\mathbb{E}\|\boldsymbol{\lambda}^*( \bm{w}^{(t)}) - \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})} \|^2 &\leq 2 \mathbb{E}\|\boldsymbol{\lambda}^*(\bm{w}^{(t)})- \boldsymbol{\lambda}^*( \bm{w} ^{(s\tau)}) \|^2+2 \mathbb{E}\| \boldsymbol{\lambda}^*( \bm{w} ^{(s\tau)})- \boldsymbol{\lambda}^{(s)} \|^2. \nonumber
\end{align}
Firstly we are going to bound $\mathbb{E}\|\boldsymbol{\lambda}^*(\bm{w}^{(t)})- \boldsymbol{\lambda}^*( \bm{w}^{(s\tau)}) \|^2$. We use the $\kappa$-Lipschitz property of $\boldsymbol{\lambda}^*(\cdot)$:
\begin{align}
\mathbb{E}\left\|\boldsymbol{\lambda}^*(\bm{w}^{(t)})- \boldsymbol{\lambda}^*\left( \bm{w} ^{(s\tau)}\right) \right\|^2 \leq \kappa^2 \mathbb{E}\|\bm{w} ^{(t)} - \bm{w}^{(s\tau)}\|^2 \leq \kappa^2\tau^2\eta^2 G_{w}^2.\nonumber
\end{align}
Then we switch to bound $\mathbb{E}\|\boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^*( \bm{w}^{(s\tau)}) \|^2$. We apply the Jensen's inequality first to get:
\begin{align}
\mathbb{E}\left\|\boldsymbol{\lambda}^{(s)} \right. \left. -\boldsymbol{\lambda}^*( \bm{w}^{(s\tau)})\right\|^2 & \leq \left(1+\frac{1}{2(\kappa-1)}\right)\mathbb{E}\left\|\boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right\|^2\nonumber\\ &\quad +\left(1+2(\kappa-1)\right) \mathbb{E}\left\|\boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right) - \boldsymbol{\lambda}^*\left( \bm{w}^{(s\tau)}\right)\right\|^2 \nonumber\\
& \leq \left(1+\frac{1}{2(\kappa-1)}\right)\mathbb{E}\left\|\boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right\|^2 + 2 \kappa^3 \tau^2 \eta^2G_{w}^2, \label{eq: lambda gap eq 1}
\end{align}
where we use the fact that $\boldsymbol{\lambda}^*(\cdot)$ is $\kappa$-Lipschitz.\\
To bound $\mathbb{E} \left\|\boldsymbol{\lambda}^{(s)}-\boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right\|^2$, by the updating rule of $\boldsymbol{\lambda}$ and the $\mu$-strongly-concavity of $F(\bm{w} ,\cdot)$ we have:
\begin{align}
\mathbb{E} \left\|\boldsymbol{\lambda}^{(s)} \right. & \left. -\boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right\|^2 \nonumber\\ & \leq \mathbb{E}\left\|\boldsymbol{\lambda}^{(s-1)} - \boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right\|^2 +\gamma^2 \left\|\nabla_{\boldsymbol{\lambda}} F\left(\bm{w}^{((s-1)\tau)}, \boldsymbol{\lambda}^{(s-1)}\right)\right\|^2 \nonumber\\
& \quad +2\gamma \left \langle \nabla_{\boldsymbol{\lambda}} F\left(\bm{w}^{((s-1)\tau)}, \boldsymbol{\lambda}^{(s-1)}\right), \boldsymbol{\lambda}^{(s-1)} - \boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right \rangle \nonumber\\
& \leq (1-\mu\gamma) \mathbb{E}\left\|\boldsymbol{\lambda}^{(s-1)} - \boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right\|^2 \nonumber\\
& \quad + \underbrace{(2\gamma^2L - 2\gamma )}_{\leq 0} \left[F\left(\bm{w}^{((s-1)\tau)}, \boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right)-F\left(\bm{w}^{((s-1)\tau)}, \boldsymbol{\lambda}^{(s-1)}\right) \right]\nonumber \\
& \leq \left(1-\frac{1}{\kappa}\right) \mathbb{E}\left\|\boldsymbol{\lambda}^{(s-1)} - \boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right\|^2, \label{eq: lambda gap eq 2}
\end{align}
where we used the smoothness property of $F(\bm{w} ,\cdot)$:
\begin{align}
{\small \left\|\nabla_{\boldsymbol{\lambda}} F\left(\bm{w}^{((s-1)\tau)}, \boldsymbol{\lambda}^{(s-1)}\right)\right\|^2 \leq 2L \left( F\left(\bm{w}^{((s-1)\tau)}, \boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right) - F\left(\bm{w}^{((s-1)\tau)}, \boldsymbol{\lambda}^{(s-1)}\right) \right).}\nonumber
\end{align}
Plugging (\ref{eq: lambda gap eq 2}) into (\ref{eq: lambda gap eq 1}) yields:
\begin{align}
\mathbb{E}\left\|\boldsymbol{\lambda}^{(s)} \right. & \left.-\boldsymbol{\lambda}^*( \bm{w}^{(s\tau)})\right\|^2 \nonumber \\
& \leq \left(1+\frac{1}{2(\kappa-1)}\right)\left(1-\frac{1}{\kappa}\right) \mathbb{E}\left\|\boldsymbol{\lambda}^{(s-1)} - \boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right\|^2 + 2 \kappa^3 \tau^2 \eta^2G_{w}^2 \nonumber\\
& \leq \left(1-\frac{1}{2 \kappa }\right) \mathbb{E}\left\|\boldsymbol{\lambda}^{(s-1)} - \boldsymbol{\lambda}^*\left( \bm{w}^{((s-1)\tau)}\right)\right\|^2 + 2 \kappa^3 \tau^2 \eta^2G_{w}^2.\nonumber
\end{align}
Applying the recursion on the above relation gives:
\begin{align}
\mathbb{E}\|\boldsymbol{\lambda}^{(s)} -\boldsymbol{\lambda}^*( \bm{w}^{(s\tau)})\|^2 \leq \left(1-\frac{1}{2\kappa}\right)^s\mathbb{E}\left\|\boldsymbol{\lambda}^{{0}}- \boldsymbol{\lambda}^*(\bm{w}^{(0)})\right\|^2 + 4 \kappa^4 \tau^2 \eta^2G_{w}^2.\nonumber
\end{align}
Putting these pieces together concludes the proof:
\begin{align}
\mathbb{E}\left\|\boldsymbol{\lambda}^*( \bm{w}^{(t)}) - \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})} \right\|^2\leq 2\left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}}\mathbb{E}\left\|\boldsymbol{\lambda}_{{0}}- \boldsymbol{\lambda}^*(\bm{w} ^{(0)})\right\|^2 + 2(4\kappa^2+1)\kappa^2 \tau^2 \eta^2 G_{w}^2.\nonumber
\end{align}
\qed
\begin{lemma}
\label{lemma: sum of series}
For $\eta \mu \leq 1$, $\kappa >1$,$\tau\geq 1$, the following inequalities holds:
\begin{align}
\sum_{t=0}^T \left(1-\frac{1}{2}\eta\mu\right)^t\left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}} \leq \frac{2\kappa\tau}{1- \frac{1}{2}\eta\mu}, \nonumber\\
\sum_{t=0}^T \left(1-\frac{1}{4}\eta\mu\right)^t\left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}} \leq \frac{2\kappa\tau}{1- \frac{1}{4}\eta\mu}. \nonumber
\end{align}
\end{lemma}
\begin{proof}
\begin{align}
\sum_{t=0}^T (1-\frac{1}{2}\eta\mu)^t(1-\frac{1}{2\kappa})^{\floor{\frac{t}{\tau}}} &= \sum_{s=0}^{S-1}\sum_{t=1}^{\tau} (1-\frac{1}{2}\eta\mu)^{s\tau+t}(1-\frac{1}{2\kappa})^{s} \nonumber \\
& \leq \sum_{s=0}^{S-1}(1-\frac{1}{2\kappa})^{s} \sum_{t=1}^{\tau} \left(1-\frac{1}{2}\eta\mu\right)^{s\tau+t} \nonumber\\
& \leq 2 \sum_{s=0}^{S-1}(1-\frac{1}{2\kappa})^{s} \frac{\left(1-\frac{1}{2}\eta\mu\right)^{s\tau}(1 - \left(1-\frac{1}{2}\eta\mu\right)^{\tau})}{\eta\mu} \nonumber\\
& = \frac{2(1 - \left(1-\frac{1}{2}\eta\mu\right)^{\tau})}{\eta\mu} \sum_{s=0}^{S-1}\left(1-\frac{1}{2\kappa}\right)^{s} \left(1-\frac{1}{2}\eta\mu\right)^{s\tau} \nonumber\\
& \leq \frac{2(1 - \left(1-\frac{1}{2}\eta\mu\right)^{\tau})}{\eta\mu} \sum_{s=0}^{S-1}\left(1-\frac{1}{2\kappa}\right)^{s} \left(1-\frac{1}{2}\eta\mu\right)^{s} \label{lemma: sum of series3}\\
&\leq \frac{2\tau \ln \frac{1}{\left(1-\frac{1}{2}\eta\mu\right)}}{\eta\mu} \frac{1}{1-\left(1-\frac{1}{2\kappa}\right)\left(1-\frac{1}{2}\eta\mu\right)} \label{lemma: sum of series2}\\
&\leq \frac{2\tau \ln \frac{1}{\left(1-\frac{1}{2}\eta\mu\right)}}{\left(\frac{\eta\mu}{2\kappa} +(\frac{1}{2}-\frac{1}{4\kappa})\eta^2\mu^2\right)} \leq \frac{4\kappa \tau }{\eta\mu } \left(\frac{1}{1- \frac{1}{2}\eta\mu}-1\right) \label{lemma: sum of series0}\\
& \leq \frac{2\kappa\tau}{\eta\mu } \left( \frac{\eta\mu}{1-\frac{1}{2} \eta\mu} \right) = \frac{2\kappa\tau}{1- \frac{1}{2}\eta\mu}, \label{lemma: sum of series1}
\end{align}
where from (\ref{lemma: sum of series3}) to (\ref{lemma: sum of series2}) we use the inequality $1-a^x \leq x\ln\frac{1}{a}$, and from (\ref{lemma: sum of series0}) to (\ref{lemma: sum of series1}) we use the inequality $\ln x \leq x-1$.
Similarly, for the second statement:
\begin{align}
\sum_{t=0}^T (1-\frac{1}{4}\eta\mu)^t(1-\frac{1}{2\kappa})^{\floor{\frac{t}{\tau}}} &= \sum_{s=0}^{S-1}\sum_{t=1}^{\tau} \left(1-\frac{1}{4}\eta\mu\right)^{s\tau+t}(1-\frac{1}{2\kappa})^{s} \nonumber\\
& \leq \sum_{s=0}^{S-1}(1-\frac{1}{2\kappa})^{s} \sum_{t=1}^{\tau} \left(1-\frac{1}{4}\eta\mu\right)^{s\tau+t} \nonumber\\
& \leq 2 \sum_{s=0}^{S-1}(1-\frac{1}{2\kappa})^{s} \frac{\left(1-\frac{1}{4}\eta\mu\right)^{s\tau}(1 - \left(1-\frac{1}{4}\eta\mu\right)^{\tau})}{\eta\mu} \nonumber\\
& = \frac{2(1 - \left(1-\frac{1}{4}\eta\mu\right)^{\tau})}{\eta\mu} \sum_{s=0}^{S-1}\left(1-\frac{1}{2\kappa}\right)^{s} \left(1-\frac{1}{4}\eta\mu\right)^{s\tau} \nonumber\\
& \leq \frac{2(1 - \left(1-\frac{1}{4}\eta\mu\right)^{\tau})}{\eta\mu} \sum_{s=0}^{S-1}\left(1-\frac{1}{2\kappa}\right)^{s} \left(1-\frac{1}{4}\eta\mu\right)^{s} \nonumber\\
&\leq \frac{2\tau \ln \frac{1}{\left(1-\frac{1}{4}\eta\mu\right)}}{\eta\mu} \frac{1}{1-\left(1-\frac{1}{2\kappa}\right)\left(1-\frac{1}{4}\eta\mu\right)} \nonumber\\
&\leq \frac{2\tau \ln \frac{1}{\left(1-\frac{1}{4}\eta\mu\right)}}{\left(\frac{\eta\mu}{2\kappa} +(\frac{1}{4}-\frac{1}{8\kappa})\eta^2\mu^2\right)} \leq \frac{4\kappa \tau }{\eta\mu } \left(\frac{1}{1- \frac{1}{4}\eta\mu}-1\right)\nonumber \\
& \leq \frac{2\kappa\tau}{\eta\mu } \left( \frac{\eta\mu}{1-\frac{1}{4} \eta\mu} \right) = \frac{2\kappa\tau}{1- \frac{1}{4}\eta\mu} \nonumber.
\end{align}
\end{proof}
\subsection{Proof of Theorem~\ref{theorem2}}
Now we proceed to the proof of Theorem~\ref{theorem2}. According to Lemma~\ref{lemma: one iteration SCSC} we have:
\begin{align}
\mathbb{E}\left\|\bm{w} ^{(t+1)} - \bm{w} ^*\right\|^2 &\leq \left (1-\frac{\mu\eta}{2}\right) \mathbb{E}\left\|\bm{w} ^{(t)}- \bm{w} ^*\right\|^2 -\eta\mathbb{E}\left[\Phi( \bm{w}^{(t)})-\Phi(\bm{w} ^*)\right]+ \eta^2\frac{2\sigma_w^2+4G_w^2}{m}\nonumber\\
& \quad + 4L^2\left(\eta^2+\frac{\eta}{\mu}\right) \mathbb{E}\left[\delta^{(t)} \right] +4\left(\frac{\eta}{\mu}+\eta^2\right)\mathbb{E}\left\|\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})- \nabla\Phi( \bm{w}^{(t)})\right\|^2 \nonumber\\
&\leq \left (1-\frac{\mu\eta}{2}\right) \mathbb{E}\left\|\bm{w} ^{(t)}- \bm{w} ^*\right\|^2 -\eta\mathbb{E}\left[\Phi( \bm{w}^{(t)})-\Phi(\bm{w} ^*)\right]+ \eta^2\frac{2\sigma_w^2+4G_w^2}{m}\nonumber\\
& \quad + 4L^2\left(\eta^2+\frac{\eta}{\mu}\right) \mathbb{E}\left[\delta^{(t)} \right] +4\left(\frac{\eta}{\mu}+\eta^2\right) L^2 \mathbb{E}\left\| \boldsymbol{\lambda}^*(\bm{w} ^{(t)})- \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}\right\|^2,\nonumber
\end{align}
where we use the smoothness of $F$ at the last step to substitute $\|\nabla_{\bm{w} }F(\bm{w} ^{(t)}, \boldsymbol{\lambda}^*(\bm{w} ^{(t)})) - \nabla_{\bm{w} }F(\bm{w} ^{(t)}, \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\|^2$:
\begin{align}
\left\|\nabla_{\bm{w} }F(\bm{w} ^{(t)}, \boldsymbol{\lambda}^*(\bm{w} ^{(t)})) - \nabla_{\bm{w} }F(\bm{w} ^{(t)}, \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})\right\|^2 \leq L^2 \left\| \boldsymbol{\lambda}^*(\bm{w} ^{(t)})- \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}\right\|^2.\nonumber
\end{align}
Then plugging in Lemma~\ref{lemma: optimal gap} yields:
\begin{align}
\mathbb{E}\|\bm{w} ^{(t+1)} - \bm{w} ^*\|^2
&\leq \left (1-\frac{\mu\eta}{2}\right) \mathbb{E}\left\|\bm{w} ^{(t)}- \bm{w} ^*\right\|^2 -\eta\mathbb{E}\left[\Phi( \bm{w}^{(t)})-\Phi(\bm{w} ^*)\right]+ \eta^2\frac{2\sigma_w^2+4G_w^2}{m}\nonumber\\
&\quad + 4L^2\left(\eta^2+\frac{\eta}{\mu}\right) \mathbb{E}\left[\delta^{(t)} \right] \nonumber\\
& \quad +8\left(\frac{\eta}{\mu}+\eta^2\right)L^2 \left(\left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}}\mathbb{E}\|\boldsymbol{\lambda}^{(0)}- \boldsymbol{\lambda}^*(\bm{w} ^{(0)})\|^2 + \kappa^2\tau^2\eta^2 G_{w}^2 \left(4\kappa^2 +1 \right)\right). \label{mainrecursion}
\end{align}
Unrolling the recursion yields:
\begin{align}
&\mathbb{E}\|\bm{w} ^{(T)} - \bm{w} ^*\|^2 \nonumber\\
& \leq \left(1-\frac{1}{2}\mu \eta\right)^T\mathbb{E}\|\bm{w} ^{(0)}- \bm{w} ^*\|^2 +\sum_{t=1}^{T}\left(1-\frac{1}{2}\mu \eta\right)^t \left[8L^2\kappa^2\tau^2\eta^2 G_{w}^2\left(\frac{\eta}{\mu}+\eta^2\right) \left(4\kappa^2 +1 \right)\right] \nonumber\\
& \quad + \sum_{t=1}^{T}\left(1-\frac{1}{2}\mu \eta\right)^t \left[\eta^2\frac{2\sigma_w^2+4G_w^2}{m}+4L^2\left(\eta^2+\frac{\eta}{\mu}\right) \mathbb{E}\left[\delta^{(t)} \right] \right] \nonumber\\
&\quad + 8\left(\frac{\eta}{\mu}+\eta^2\right)L^2\mathbb{E}\|\boldsymbol{\lambda}^{(0)}- \boldsymbol{\lambda}^*(\bm{w} ^{(0)})\|^2 \sum_{t=1}^{T}\left(1-\frac{1}{2}\mu \eta\right)^t\left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}} \label{t21}\\
& \leq \exp\left(-\frac{1}{2}\mu \eta T\right)D_{\mathcal{W}}^2+ \eta \frac{ 4\sigma_w^2+8G_w^2 }{\mu m}+8L^2\left(\frac{\eta}{\mu}+\frac{1}{\mu^2}\right)\sum_{t=0}^{T}\mathbb{E}\left[\delta^{(t)} \right] \nonumber\\
&\quad + 16L^2\kappa^2\tau^2\eta^2 G_{w}^2\left(\frac{\eta}{\mu}+\frac{1}{\mu^2}\right) \left( 4\kappa^2 +1 \right) + 16L^2 \left( \frac{\kappa\tau }{1- \frac{1}{2}\eta\mu} \right)\left(\frac{\eta}{\mu}+\eta^2\right)D_{\Lambda}^2, \label{t20}
\end{align}
where we used the result from Lemma~\ref{lemma: sum of series} from (\ref{t21}) to (\ref{t20}). Now, we simplify (\ref{mainrecursion}) by applying the telescoping sum on (\ref{mainrecursion}) for $t=\frac{T}{2}$ to $T$:
\begin{align}
&\frac{2}{T}\sum_{t=T/2}^T\mathbb{E}\left[\Phi(\bm{w} ^{(t)})-\Phi(\bm{w} ^*)\right] \nonumber\\
& \leq \frac{2}{\eta T}\mathbb{E}\| \bm{w}^{(T/2)}- \bm{w}^*\|^2 + \eta \frac{ 2\sigma_w^2+4G_w^2}{m} + 4L^2\left(\eta +\frac{1}{\mu}\right) \frac{2}{T}\sum_{t=T/2}^T \mathbb{E}\left[\delta^{(t)} \right] \nonumber\\
& \quad +8\left(\frac{1}{\mu}+\eta\right)L^2D_{\Lambda}^2 \frac{2}{T}\sum_{t=T/2}^T\left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}} +8\left(\frac{1}{\mu}+\eta\right) \kappa^2\tau^2\eta^2 L^2G_{w}^2 \left(4\kappa^2 +1 \right) \nonumber\\
& \leq \frac{2}{\eta T}\mathbb{E}\| \bm{w}^{(T/2)}- \bm{w} ^*\|^2 + \eta \frac{ 2\sigma_w^2+4G_w^2}{m} + 80\eta^2\tau^2L^2\left(\eta +\frac{1}{\mu}\right) \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right)\nonumber\\
& \quad +16\left(\frac{1}{\mu}+\eta\right)L^2 O\left(\frac{ \tau \exp(-\mu\eta T/4 \tau) }{T}D_{\Lambda}^2 \right) +8\left(\frac{1}{\mu}+\eta\right) \kappa^2\tau^2\eta^2 L^2G_{w}^2 \left(4\kappa^2 +1 \right) \nonumber\\
& \leq \frac{2}{\eta T}\mathbb{E}\| \bm{w}^{(T/2)}- \bm{w} ^*\|^2 + \eta \frac{ 2\sigma_w^2+4G_w^2}{m} + 80\eta^2\tau^2L^2\left(\eta +\frac{1}{\mu}\right) \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right)\nonumber\\
& \quad +16\left(\frac{1}{\mu}+\eta\right)L^2 O\left(\frac{ \tau \exp(-\mu\eta T/4 \tau) }{T}D_{\Lambda}^2 \right) +8\left(\frac{1}{\mu}+\eta\right) \kappa^2\tau^2\eta^2 L^2G_{w}^2 \left(4\kappa^2 +1 \right).\nonumber
\end{align}
Plugging in (\ref{t20}) yields:
\begin{align}
\frac{2}{T} & \sum_{t=T/2}^T \mathbb{E}\left[\Phi(\bm{w} ^{(t)})-\Phi(\bm{w} ^*)\right] \nonumber\\
& \leq \frac{2}{\eta T}\left( \exp\left(-\frac{1}{4}\mu \eta T\right)D_{\mathcal{W}}^2+ \eta \frac{ 4\sigma_w^2+8G_w^2 }{\mu m}+8L^2\left(\frac{\eta}{\mu}+\frac{1}{\mu^2}\right)\sum_{t=0}^{T}\mathbb{E}\left[\delta^{(t)} \right]\right)\nonumber \\
& \quad+ \frac{2}{\eta T}\left( 16L^2\kappa^2\tau^2\eta^2 G_{w}^2\left(\frac{\eta}{\mu}+\frac{1}{\mu^2}\right) \left( 4\kappa^2 +1 \right) +16L^2 \left( \frac{\kappa\tau }{1- \frac{1}{2}\eta\mu} \right)\left(\frac{\eta}{\mu}+\eta^2\right)D_{\Lambda}^2\right) \nonumber\\
&\quad+ \eta \frac{ 2\sigma_w^2+4G_w^2}{m} + 80\eta^2\tau^2L^2\left(\eta +\frac{1}{\mu}\right) \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right)\nonumber\\
& \quad +16\left(\frac{1}{\mu}+\eta\right)L^2 O\left(\frac{ \tau \exp(-\mu\eta T/4 \tau) }{T}D_{\Lambda}^2 \right) +8\left(\frac{1}{\mu}+\eta\right) \kappa^2\tau^2\eta^2 L^2G_{w}^2 \left(4\kappa^2 +1 \right).\nonumber
\end{align}
Combining the terms yields:
\begin{align}
&\frac{2}{T}\sum_{t=T/2}^T\mathbb{E}\left[\Phi(\bm{w} ^{(t)})-\Phi(\bm{w} ^*)\right] \nonumber\\
& \leq \frac{2}{\eta T} \exp\left(-\frac{1}{4}\mu \eta T\right)D_\mathcal{W}^2 + 16\left(\frac{1}{\mu}+\eta\right)L^2 O\left(\frac{ \tau \exp(-\mu\eta T/4 \tau) }{T}D_{\Lambda}^2 \right)\nonumber\\
& \quad+ \left(\frac{4}{\mu T}+ \eta\right)\frac{ 2\sigma_w^2+4G_w^2 }{m} + \left( 1 +\frac{2}{\mu \eta T}\right)80\eta^2\tau^2L^2\left(\eta +\frac{1}{\mu}\right)\left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right)\nonumber\\
& \quad+ \left( \frac{4}{\mu\eta T} + 1\right)8L^2\kappa^2\tau^2\eta^2 G_{w}^2\left( 4\kappa^2 +1 \right) \left( \eta +\frac{1}{\mu }\right)\nonumber \\
& \quad+ \frac{32L^2}{T} \left( \frac{\kappa\tau }{1- \frac{1}{2}\eta\mu} \right)\left(\frac{1}{\mu}+\eta \right)D_{\Lambda}^2.\nonumber
\end{align}
And finally, plugging in $\eta = \frac{4 \log T}{\mu T}$ and using the fact that $\Phi(\frac{2}{T}\sum_{t=T/2}^T \bm{w} ^{(t)}) \leq \frac{2}{T}\sum_{t=T/2}^T \Phi(\bm{w} ^{(t)})$ yields:
\begin{align}
&\mathbb{E}[\Phi(\hat{\bm{w} })-\Phi(\bm{w} ^*)]\nonumber\\
&\leq \frac{\mu D_{\mathcal{W}}^2}{2 T\log T} + 16\left(\frac{1}{\mu}+ \frac{4 \log T}{\mu T}\right)L^2 O\left(\frac{ \tau }{T^{(1+1/\tau)}}D_{\Lambda}^2 \right)\nonumber\\
& \quad + \left(\frac{4}{\mu T}+ \frac{4 \log T}{\mu T}\right)\frac{ 2\sigma_w^2+4G_w^2 }{m} + \left( 1 +\frac{2}{\mu \eta T}\right) \frac{1280\kappa^2 \tau^2 \log^2 T }{ T^2} \left(\eta +\frac{1}{\mu}\right)\left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right)\nonumber\\
& \quad + \left( \frac{1}{\log T} + 1\right) \frac{ 8\kappa^4 \tau^2 \log^2 T }{ T^2} G_{w}^2\left( 4\kappa^2 +1 \right) \left( \frac{4 \log T}{\mu T} +\frac{1}{\mu }\right)\nonumber \\
& \quad + \frac{32L^2}{T} \left( \frac{\kappa\tau }{1- \frac{2 \log T}{ T}} \right)\left(\frac{1}{\mu}+\frac{4 \log T}{\mu T} \right)D_{\Lambda}^2 \nonumber\\
&\leq \Tilde{O}\left( \frac{\mu D_{\mathcal{W}}^2 }{ T } \right) + O\left(\frac{\kappa L \tau D_{\Lambda}^2 }{T^{(1+1/\tau)}} \right)+ \Tilde{O}\left(\frac{ \sigma_w^2+ G_w^2 }{\mu mT}\right) + O\left( \frac{\kappa^2\tau^2(\sigma_w^2 + \Gamma)}{\mu T^2}\right) \nonumber\\
& \quad + \Tilde{O}\left(\frac{\kappa^2L\tau D_{\Lambda}^2 }{T}\right)+ \Tilde{O}\left(\frac{\kappa^6\tau^2G_w^2}{\mu T^2}\right).\nonumber
\end{align}
\qed
\section{Proof of Convergence of {\sffamily{DRFA-GA}} in Nonconvex (PL Condition)-Strongly-Concave Setting}\label{sec:ncsc}
\subsection{Overview of Proofs}
In this section we will present formal proofs in nonconvex (PL condition)-strongly-concave setting (Theorem~\ref{thm3}). The main idea is similar to strongly-convex-strongly-concave case: we start from one iteration analysis, and plug in the upper bound of $\delta^{(t)}$ and $\|\nabla_{\bm{w} }F( \bm{w}^{(t)},\boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})})- \nabla\Phi( \bm{w}^{(t)})\|^2$.
However, a careful analysis need to be employed in order to deal with projected SGD in constrained nonconvex optimization problem. We employ the technique used in~\cite{ghadimi2016mini}, where they advocate to study the following quantity:
\begin{align}
P_{\mathcal{W}}(\bm{w}, \bm{g},\eta) = \frac{1}{\eta}\left[\bm{w}-\prod_{\mathcal{W}} \left ( \bm{w}- \eta \bm{g} \right)\right]\nonumber.
\end{align}
If we plug in $\bm{w} = \bm{w}^{(t)}$, $\bm{g} = \bm{u}^{(t)} = \frac{1}{m}\sum_{i\in \mathcal{D}^{(\floor{\frac{t}{\tau}})}} \nabla f_i ( \bm{w}_i^{(t)};\xi_i^t)$, then
\begin{align}
P_{\mathcal{W}}(\bm{w}^{(t)}, \bm{u}^{(t)},\eta) = \frac{1}{\eta} \left[\bm{w}^{(t)} - \prod_{\mathcal{W}}\left (\bm{w}^{(t)} - \eta \bm{u}^{(t)} \right)\right]\nonumber.
\end{align}
characterize the difference between iterates $\bm{w}^{(t+1)}$ and $\bm{w}^{(t)}$. A trivial property of operator $ P_{\mathcal{W}}$ is contraction mapping, which follows the property of projection:
\begin{align}
\left\|P_{\mathcal{W}}(\bm{w},\bm{g}_1,\eta) - P_{\mathcal{W}}(\bm{w}, \bm{g}_2,\eta)\right\|^2 \leq \left\|\bm{g}_1 - \bm{g}_2\right\|^2.\nonumber
\end{align}
The significant property of operator $P_{\mathcal{W}}$ is given by the following lemma:
\begin{lemma}[Property of Projection, \cite{ghadimi2016mini} Lemma 1]\label{lemma: projection1}
For all $\bm{w} \in \mathcal{W} \subset \mathbb{R}^d$, $\bm{g} \in \mathbb{R}^d$ and $\eta > 0$, we have:
\begin{align}
\left \langle \bm{g}, P_{\mathcal{W}}(\bm{w}, \bm{g},\eta) \right\rangle \geq \left\|P_{\mathcal{W}}(\bm{w}, \bm{g},\eta)\right\|^2.\nonumber
\end{align}
\end{lemma}
The above lemma establishes a lower bound for the inner product $ \left \langle \bm{g}, P_{\mathcal{W}}(\bm{y},\bm{g},\eta) \right\rangle$, and will play a significant role in our analysis.
\subsection{Proof of Technical Lemmas}
\begin{lemma}
If $F(\cdot,\boldsymbol{\lambda})$ satisfies $\mu$-generalized PL condition, then $\Phi(\cdot)$ also satisfies $\mu$-generalized PL condition.
\end{lemma}
\begin{proof}
Let $\bm{w}^* \in \arg \min_{\bm{w}\in \mathcal{W}} \Phi(\bm{w})$. Since $F(\cdot,\boldsymbol{\lambda})$ satisfies $\mu$-generalized PL condition, we have for any $\bm{w}\in\mathcal{W}$:
\begin{align}
\frac{1}{2\eta^2}\left\|\bm{w} -\prod_{\mathcal{W}}\left( \bm{w}-\eta\nabla_{\bm{w} } F(\bm{w} ,\boldsymbol{\lambda}^*(\bm{w} ))\right) \right \|^2 & \geq \mu(F(\bm{w} ,\boldsymbol{\lambda}^*(\bm{w} ) -\min_{\bm{w}'\in\mathcal{W}} F(\bm{w}',\boldsymbol{\lambda}^*(\bm{w} )) \nonumber \\
&\geq \mu(F(\bm{w} ,\boldsymbol{\lambda}^*(\bm{w} ) - F(\bm{w} ^*,\boldsymbol{\lambda}^*(\bm{w} )) \nonumber\\ &\geq \mu(F(\bm{w} ,\boldsymbol{\lambda}^*(\bm{w} ) - F(\bm{w}^*,\boldsymbol{\lambda}^*(\bm{w}^*)) .\nonumber
\end{align}
which immediately implies $\frac{1}{2\eta^2}\|\bm{w} -\prod_{\mathcal{W}}\left( \bm{w}-\eta \nabla \Phi(\bm{w} )\right)\|^2 \geq \mu( \Phi(\bm{w} ) - \Phi(\bm{w} ^*))$ as desired.
\end{proof}
\begin{lemma}
\label{lemma: one iteration PLSC}
For {\sffamily{DRFA-GA}}, under Theorem~\ref{thm3}'s assumptions, we have:
\begin{align}
\mathbb{E}\left[\Phi(\bm{w} ^{(t+1)}) - \Phi(\bm{w} ^{*}) \right] &\leq \left(1-\frac{\mu \eta}{4}\right) \mathbb{E}\left[\Phi(\bm{w} ^{(t)}) - \Phi(\bm{w} ^{*}) \right] \nonumber \\
& \quad + \frac{3\eta}{2 } \mathbb{E}\left\|\sum_{i=1}^N \lambda^{(\floor{\frac{t}{\tau}})}_i \nabla f_i(\bm{w} ^{(t)}_i) - \nabla \Phi ( \bm{w}^{(t)}) \right\|^2 + 3\eta\frac{2\sigma_w^2+4G_w^2}{2m},
\end{align}
where $\alpha=L + \kappa L$
\end{lemma}
\begin{proof}
Define the following quantities:
\begin{align}
& \bm{u}_t = \frac{1}{m}\sum_{i\in D^t} \nabla f_i(\bm{w}_i^{(t)};\xi_i^t), \bar{\bm{u}}_t = \sum_{i=1}^N \lambda_i^{(\floor{\frac{t}{\tau}})} \nabla f_i(\bm{w}_i^{(t)}).\nonumber \\
&\Tilde{R}^{(t)} = P_{\mathcal{W}}(\bm{w}^t, \bm{u}_t ,\eta) =\bm{w}^{(t)} -\frac{1}{\eta} \prod_{\mathcal{W}} \left ( \bm{w}^{(t)}-\eta \bm{u}_t \right) \nonumber\\
& {R}^{(t)} = P_{\mathcal{W}}(\bm{w}^t, \bar{\bm{u}}_t,\eta) =\bm{w}^{(t)}- \frac{1}{\eta}\prod_{\mathcal{W}} \left ( \bm{w}^{(t)}- \eta \bar{\bm{u}}_t \right) \nonumber\\
&\hat{R}^{(t)} = P_{\mathcal{W}}(\bm{w}^t,\Phi(\bm{w}^{(t)}) ,\eta) =\bm{w}^{(t)}- \frac{1}{\eta}\prod_{\mathcal{W}} \left ( \bm{w}^{(t)}- \eta \nabla \Phi(\bm{w}^{(t)}) \right) \nonumber.
\end{align}
By the $\alpha$-smoothness of $\Phi$ and the updating rule of $\bm{w}$ we have:
\begin{align}
\mathbb{E}[\Phi( \bm{w}^{(t+1)} )] - \mathbb{E}[\Phi(\bm{w}^{(t)} )] &\leq \frac{ \alpha}{2}\mathbb{E}\left[\left\|\bm{w}^{(t+1)}-\bm{w}^{(t)}\right\|^2\right] + \left \langle \nabla \Phi(\bm{w}^{(t)} ), \bm{w}^{(t+1)}-\bm{w}^{(t)} \right\rangle \nonumber\\
&\leq \frac{\eta^2 \alpha}{2}\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2\right] - \eta \mathbb{E}\left[\left \langle \nabla \Phi(\bm{w}^{(t)} ), \Tilde{R}^{(t)} \right\rangle\right] \nonumber\\
&\leq \frac{\eta^2 \alpha}{2}\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2\right]- \eta \mathbb{E}\left[\left \langle \bm{u}_t, P_{\mathcal{W}}(\bm{y}^t, \bm{u}_t,\eta) \right\rangle\right] \nonumber\\
& \quad - \eta\mathbb{E}\left[ \left \langle \nabla \Phi(\bm{w}^{(t)} )-\bm{u}_t, \Tilde{R}^{(t)} \right\rangle\right] \nonumber.
\end{align}
According to Lemma~\ref{lemma: projection1}, we can bound the first dot product term in the last inequality by $\|\tilde{R}^{(t)}\|^2$, so then we have:
\begin{align}
\mathbb{E}[\Phi( \bm{w}^{(t+1)} )] &- \mathbb{E}[\Phi(\bm{w}^{(t)} )] \nonumber\\
&\leq \frac{\eta^2 \alpha}{2}\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2 \right] - \eta \mathbb{E}\left[\|\Tilde{R}^{(t)}\|^2\right] - \eta \mathbb{E}\left[\left \langle \nabla \Phi(\bm{w}^{(t)} )-\bm{u}_t, \Tilde{R}^{(t)} \right\rangle \right] \nonumber\\
&\leq -\left(\eta-\frac{\eta^2 \alpha}{2}\right)\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2 \right] - \eta \mathbb{E}\left[\left \langle \nabla \Phi(\bm{w}^{(t)} )-\bm{u}_t, \Tilde{R}^{(t)} \right\rangle \right] \nonumber\\
&\leq -\left(\eta-\frac{\eta^2 \alpha}{2}\right)\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2 \right] + \frac{\eta}{2} \mathbb{E}\left[\left \| \nabla \Phi(\bm{w}^{(t)} )-\bm{u}_t\right\|^2+ \left\| \Tilde{R}^{(t)} \right\|^2 \right] \nonumber\\
&\leq \underbrace{-\left(\frac{\eta}{2}-\frac{\eta^2 \alpha}{2}\right)}_{\leq -\frac{1}{4}\eta}\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2 \right] + \eta \mathbb{E}\left[\left \| \nabla \Phi(\bm{w}^{(t)} )-\bar{\bm{u}}_t\right\|^2+ \left \|\bar{\bm{u}}_t- \bm{u}_t\right\|^2 \right] \nonumber\\
&\leq -\frac{1}{4}\eta\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2 \right] + \eta \mathbb{E}\left[\left \| \nabla \Phi(\bm{w}^{(t)} )-\bar{\bm{u}}_t\right\|^2\right] + \frac{\eta (2\sigma_w^2+4G_w^2)}{m}
\label{eq: ncnc lm2 eq1}.
\end{align}
Notice that:
\begin{align}
\mathbb{E}\left[\left\| \hat{R}^{(t)}\right\|^2 \right] &\leq 2\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2 \right] + 2\mathbb{E}\left[\left\| \hat{R}^{(t)}- \Tilde{R}^{(t)}\right\|^2 \right]\nonumber\\
&\leq 2\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2 \right] + 4\mathbb{E}\left[\left\| \hat{R}^{(t)}- {R}^{(t)}\right\|^2 \right]+ 4\mathbb{E}\left[\left\| {R}^{(t)}- \Tilde{R}^{(t)}\right\|^2 \right]\nonumber\\
&\leq 2\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2 \right] + 4\mathbb{E}\left[\left\| \hat{R}^{(t)}- {R}^{(t)}\right\|^2 \right]+ 4\mathbb{E}\left[\left\| \bm{u}^{(t)}- \bar{\bm{u}}^{(t)}\right\|^2 \right]\nonumber\\
&\leq 2\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2 \right] + 4 \mathbb{E}\left[\left \| \nabla \Phi(\bm{w}^{(t)} )-\bar{\bm{u}}_t\right\|^2\right]+ \frac{4\eta (2\sigma_w^2+4G_w^2)}{m} \label{eq: ncnc lm2 eq2}.
\end{align}
Thus, plugging (\ref{eq: ncnc lm2 eq2}) into (\ref{eq: ncnc lm2 eq1}) to substitute $\mathbb{E}\left[\left\| \Tilde{R}^{(t)}\right\|^2 \right]$ yields:
\begin{align}
\mathbb{E}[\Phi( \bm{w}^{(t+1)} )] &- \mathbb{E}[\Phi(\bm{w}^{(t)} )] \nonumber\\
&\leq -\frac{1}{8}\eta \mathbb{E}\left[\left\| \hat{R}^{(t)}\right\|^2\right] + \frac{1}{2}\eta \mathbb{E}\left[\left \| \nabla \Phi(\bm{w}^{(t)} )-\bar{\bm{u}}_t\right\|^2\right] + \frac{\eta (2\sigma_w^2+4G_w^2)}{2m} \nonumber\\
& \quad +\eta \mathbb{E}\left[\left \| \nabla \Phi(\bm{w}^{(t)} )-\bar{\bm{u}}_t\right\|^2\right] + \frac{\eta (2\sigma_w^2+4G_w^2)}{m} \nonumber\\
&\leq -\frac{1}{8}\eta \mathbb{E}\left[\left\| \hat{R}^{(t)}\right\|^2\right] + \frac{3}{2}\eta \mathbb{E}\left[\left \| \nabla \Phi(\bm{w}^{(t)} )-\bar{\bm{u}}_t\right\|^2\right] + \frac{3\eta (2\sigma_w^2+4G_w^2)}{2m}
\label{eq: ncnc lm2 eq3}.
\end{align}
Plugging in the generalized PL-condition:
\begin{align}
\frac{1}{\eta^2}\mathbb{E}\left[\left\|\prod_{\mathcal{W}}\left(\bm{w}^{(t)}- \eta\nabla \Phi(\bm{w}^{(t)}) \right) - \bm{w}^{(t)}\right\|^2\right] = \mathbb{E}\left[\left\| \hat{R}^{(t)}\right\|^2\right] \geq 2\mu \left( \mathbb{E}[\Phi( \bm{w}^{t} )] - \mathbb{E}[\Phi( \bm{w}^{*} )] \right)\nonumber
\end{align} into (\ref{eq: ncnc lm2 eq3}) yields:
\begin{align}
\mathbb{E}\left[\Phi(\bm{w} ^{(t+1)}) - \Phi(\bm{w} ^{*}) \right] &\leq \left(1-\frac{\mu \eta}{4}\right) \mathbb{E}\left[\Phi(\bm{w} ^{(t)}) - \Phi(\bm{w} ^{*}) \right] \nonumber \\
& \quad + \frac{3\eta}{2 } \mathbb{E}\left\|\sum_{i=1}^N \lambda^{(\floor{\frac{t}{\tau}})}_i \nabla f_i(\bm{w} ^{(t)}_i) - \nabla \Phi ( \bm{w}^{(t)}) \right\|^2 + 3\eta\frac{2\sigma_w^2+4G_w^2}{2m}. \nonumber
\end{align}
\end{proof}
\subsection{Proof for Theorem~\ref{thm3}}
Now we proceed to the proof of Theorem~\ref{thm3}. According to Lemma~\ref{lemma: one iteration PLSC} we have:
\begin{align}
\mathbb{E}\left[\Phi(\bm{w} ^{(t+1)}) - \Phi(\bm{w} ^{*}) \right] &\leq \left(1-\frac{\mu \eta}{4}\right) \mathbb{E}\left[\Phi(\bm{w} ^{(t)}) - \Phi(\bm{w} ^{*}) \right] \nonumber \\
& \quad + \frac{3\eta}{2 } \underbrace{\mathbb{E}\left\|\sum_{i=1}^N \lambda^{(\floor{\frac{t}{\tau}})}_i \nabla f_i(\bm{w} ^{(t)}_i) - \nabla \Phi ( \bm{w}^{(t)}) \right\|^2}_{T_1} + 3\eta\frac{2\sigma_w^2+4G_w^2}{2m}.\nonumber
\end{align}
Now, we bound the term $T_1$ in above as:
\begin{align}
T_1 &\leq 2\mathbb{E}\left\| \nabla_{\bm{w} }\Phi(\bm{w} ^{(t)}) -\sum_{i=1}^N\lambda^{(\floor{\frac{t}{\tau}})}_i\nabla_{\bm{w} }f_i(\bm{w} ^{(t)}) \right\|^2\nonumber\\
& \quad + 2\mathbb{E}\left\|\sum_{i=1}^N\lambda^{(\floor{\frac{t}{\tau}})}_i\nabla_{\bm{w} }f_i(\bm{w} ^{(t)}) -\sum_{i=1}^N\lambda^{(\floor{\frac{t}{\tau}})}_i\nabla f_i(\bm{w} ^{(t)}_i) \right\|^2 \nonumber \\
&\leq 2\mathbb{E}\left\| \nabla_{\bm{w} }F(\bm{w} ^{(t)}, \boldsymbol{\lambda}^*(\bm{w} ^{(t)})) - \nabla_{\bm{w} }F(\bm{w} ^{(t)}, \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) \right\|^2 \nonumber \\
& \quad + 2\sum_{i=1}^N\lambda^{(\floor{\frac{t}{\tau}})}_i\mathbb{E}\left\|\nabla_{\bm{w} }f_i(\bm{w} ^{(t)}) -\nabla f_i(\bm{w} ^{(t)}_i) \right\|^2 \nonumber\\
&\leq 2 L^2 \mathbb{E}\left\| \boldsymbol{\lambda}^*(\bm{w} ^{(t)})- \boldsymbol{\lambda}^{(\floor{\frac{t}{\tau}})}) \right\|^2 + 2 L^2\mathbb{E}\left[\delta^{(t)}\right]\nonumber\\
& \leq 2L^2\left(2\left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}}\mathbb{E}\|\boldsymbol{\lambda}^{(0)}- \boldsymbol{\lambda}^*(\bm{w} ^{(0)})\|^2 + 2\kappa^2\tau^2\eta^2 G_{w}^2 \left(4\kappa^2 +1 \right)\right) + 2 L^2\mathbb{E}\left[\delta^{(t)}\right], \nonumber
\end{align}
where we plug in the Lemma~\ref{lemma: optimal gap}.
Plugging $T_1$ back yields:
\begin{align}
&\mathbb{E}\left[\Phi(\bm{w}^{(t+1)}) - \Phi( \bm{w} ^*)\right] \nonumber\\
& \leq \left( 1- \frac{1}{4}\mu\eta\right)\mathbb{E}\left[\Phi(\bm{w}^{(t)}) - \Phi( \bm{w}^*)\right] + 3 \eta \frac{2\sigma_w^2+4G_w^2}{2m} \nonumber\\
& \quad + \frac{3\eta}{2}\left(4L^2\left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}} \mathbb{E}\|\boldsymbol{\lambda}^*(\bm{w}^{(0)}) - \boldsymbol{\lambda}^{(0)} \|^2+4L^2\kappa^2\tau^2\eta^2 G_{w}^2 \left(4\kappa^2 +1 \right)+2 L^2\mathbb{E}\left[\delta^{(t)}\right]\right) \nonumber\\
& \leq \left( 1- \frac{1}{4}\mu\eta\right)\mathbb{E}\left[\Phi(\bm{w}^{(t)}) - \Phi(\bm{w}^*)\right] + 3\eta \frac{2\sigma_w^2+4G_w^2}{2m} \nonumber\\
& \quad + 6\eta L^2 \left(\left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}} \mathbb{E}\|\boldsymbol{\lambda}^*( \bm{w} ^{(0)}) - \boldsymbol{\lambda}^{(0)} \|^2\right) \nonumber\\
& \quad + \frac{3\eta}{2}\left(4L^2\kappa^2\tau^2\eta^2 G_{w}^2 \left(4\kappa^2 +1 \right)+2 L^2\mathbb{E}\left[\delta^{(t)}\right]\right). \nonumber
\end{align}
Unrolling the recursion yields
\begin{align}
&\mathbb{E}\left[\Phi(\bm{w} ^{(T)}) - \Phi( \bm{w} ^*)\right] \nonumber\\
& \leq \left( 1-\frac{1}{4} \mu\eta\right)^T\mathbb{E}\left[\Phi(\bm{w} ^{(0)}) - \Phi( \bm{w} ^*) \right] + \sum_{t=0}^{T} \left( 1-\frac{1}{4} \mu\eta\right)^t 3 \eta \frac{2\sigma_w^2+4G_w^2}{2m} \nonumber\\
& \quad + 6\eta L^2 \mathbb{E}\|\boldsymbol{\lambda}^*( \bm{w}^{(0)}) - \boldsymbol{\lambda}_0 \|^2\sum_{t=0}^{T} \left [\left( 1- \frac{1}{2}\mu\eta\right)^t \left(1-\frac{1}{2\kappa}\right)^{\floor{\frac{t}{\tau}}} \right] \nonumber\\
& \quad + \frac{3}{2}\eta\left(\sum_{t=0}^{T}\left( 1- \frac{1}{4}\mu\eta\right)^t 4L^2\kappa^2\tau^2\eta^2 G_{w}^2 \left(4\kappa^2 +1 \right)+2 L^2\sum_{t=0}^{T}\left( 1- \frac{1}{4}\mu\eta\right)^t \mathbb{E}\left[\delta^{(t)}\right] \right) \nonumber \\
& \leq \exp \left( - \frac{\mu\eta T}{4}\right)\mathbb{E}\left[\Phi(\bm{w}^{(0)}) - \Phi( \bm{w} ^*)\right] + 12 \frac{2\sigma_w^2+4G_w^2}{2\mu m} \nonumber\\
& \quad + 6\eta L^2 \mathbb{E}\|\boldsymbol{\lambda}^*( \bm{w}^{(0)}) - \boldsymbol{\lambda}^{(0)} \|^2 \left(\frac{2\kappa\tau}{1- \frac{1}{4}\eta\mu} \right) \nonumber\\
& \quad + \frac{6}{\mu}\left(4L^2\kappa^2\tau^2\eta^2 G_{w}^2 \left(4\kappa^2 +1 \right) \right) + 3\eta L^2\left( 10\eta^2\tau^2 \left(\sigma_w^2+\frac{\sigma_w^2}{m} + \Gamma \right)\right)T \nonumber,
\end{align}
where we use the result of Lemmas~\ref{lemma: deviation} and \ref{lemma: sum of series}. Plugging in $\eta = \frac{ 4\log T}{\mu T}$, and $m \geq T$, we have:
\begin{align}
\Phi(\bm{w} ^{(t)}) - \Phi( \bm{w} ^*) &\leq O\left(\frac{\Phi(\bm{w} ^{(0)}) - \Phi( \bm{w} ^*)}{T}\right)+
\Tilde{O}\left(\frac{\sigma_w^2+G_w^2}{\mu T}\right) + \Tilde{O}\left(\frac{\kappa^2 L \tau D_\Lambda^2}{T}\right) \nonumber\\
& \quad + \Tilde{O}\left(\frac{\kappa^6\tau^2 G_w^2}{\mu T^2}\right)+ \Tilde{O}\left(\frac{\kappa^2\tau^2 (\sigma_w^2+\Gamma)}{\mu T^2}\right)\nonumber,
\end{align}
thus concluding the proof.
\qed
\section*{Broader Impact}
This work advocates a distributionally robust algorithm for federated learning. The algorithmic solution is designed to preserve the privacy of users, while training a high quality model. The proposed algorithm tries to minimize the maximum loss among worst case distribution over clients' data. Hence, we can ensure that even if the data distribution among users is highly heterogeneous, the trained model is reasonably good for everyone, and not benefiting only a group of clients. This will ensure the fairness in training a global model with respect to every user, and it is vitally important for critical decision making systems such as healthcare. In such a scenario, the model learned by simple algorithms such as FedAvg would have an inconsistent performance over different distributions, which is not acceptable. However, the resulting model from our algorithm will have robust performance over different distributions it has been trained on.
\section{Introduction}
Federated learning (FL) has been a key learning paradigm to train a centralized model from an extremely large number of devices/users without accessing their local data~\cite{konevcny2016federated}. A commonly used approach is to aggregate the individual loss functions usually weighted proportionally to their sample sizes and solve the following optimization problem in a distributed manner: \vspace{-0.0cm}
\begin{equation}
\min_{\boldsymbol{w} \in \mathcal{W}} F (\boldsymbol{w}) :=
\sum_{i=1}^N \frac{n_i}{n}\left\{f_i(\boldsymbol{w}):=\mathbb{E}_{\xi\sim \mathcal{P}_i}[\ell(\bm{w};\xi)]\right\}, \label{weighted loss}
\end{equation}
where $N$ is number of clients each with $n_i$ training samples drawn from some unknown distribution $\mathcal{P}_i$ (possibly different from other clients), $f_i(\bm{w})$ is the local objective at device $i$ for a given loss function $\ell$, $\mathcal{W}$ is a closed convex set, and $n$ is total number of samples.
In a federated setting, in contrast to classical distributed optimization, in solving the optimization problem in Eq.~\ref{weighted loss}, three key challenges need to be tackled including i) communication efficiency, ii) the low participation of devices, and iii) heterogeneity of local data shards. To circumvent the communication bottleneck, an elegant idea is to periodically average locally evolving models as employed in FedAvg algorithm~\cite{mcmahan2017communication}. Specifically, each local device optimizes its own model for $\tau$ local iterations using SGD, and then a subset of devices is selected by the server to communicate their models for averaging. This approach, which can be considered as a variant of local SGD~\cite{stich2018local,haddadpour2019local,haddadpour2019trading} but with partial participation of devices, can significantly reduce the number of communication rounds, as demonstrated both empirically and theoretically in various studies~\cite{li2019convergence,khaled2019better,haddadpour2019convergence,haddadpour2020federated,woodworth2020minibatch}.
While being compelling from the communication standpoint, FedAvg does not necessarily tackle the \textit{data heterogeneity} concern in FL. In fact, it has been shown that the generalization capability of the central model learned by FedAvg, or any model obtained by solving Eq.~\ref{weighted loss} in general, is inevitably plagued by increasing the diversity among local data distributions~\cite{li2019feddane,karimireddy2019scaffold,haddadpour2019convergence}. This is mainly due to the fact the objective in Eq.~\ref{weighted loss} assumes that all local data are sampled from the same distribution, but in a federated setting, local data distributions can significantly vary from the average distribution. Hence, while the global model enjoys a good \textit{average performance}, its performance often degrades significantly on local data when the distributions drift dramatically.
To mitigate the data heterogeneity issue, one solution is to personalize the global model to local distributions. A few notable studies~\cite{deng2020adaptive,mansour2020three} pursued this idea and proposed to learn a mixture of the global and local models. While it is empirically observed that the per-device mixture model can reduce the generalization error on local distributions compared to the global model, however, the learned global model still suffers from the same issues as in FedAvg, which limits its adaptation to newly joined devices. An alternative solution is to learn a model that has uniformly good performance over almost all devices by minimizing the agnostic (distributionally robust) empirical loss:\vspace{-0.1cm}
\begin{equation}
\min_{\boldsymbol{w} \in \mathcal{W}} \max_{\bm{\lambda}\in \Lambda } F (\boldsymbol{w},\boldsymbol{\lambda}) := \sum_{i=1}^N \lambda_i f_i(\boldsymbol{w}), \label{agnostic loss}\vspace{-0.1cm}
\end{equation}
where $\boldsymbol{\lambda} \in \Lambda \doteq \{\boldsymbol{\lambda} \in \mathbb{R}_{+}^N: \sum_{i=1}^N \lambda_i = 1\}$ is the global weight for each local loss function.
The main premise is that by minimizing the robust empirical loss, the learned model is guaranteed to perform well over the worst-case combination of empirical local distributions, i.e., limiting the reliance to only a fixed combination of local objectives\footnote{Beyond robustness, agnostic loss yields a notion of fairness~\cite{mohri2019agnostic}, which is not the focus of present work.}.~\citet{mohri2019agnostic} was among the first to introduce the agnostic loss into federated learning, and provided convergence rates for convex-linear and strongly-convex-strongly-concave functions. However, in their setting, the server has to communicate with local user(s) at each iteration to update the global mixing parameter $\boldsymbol{\lambda}$, which hinders its scalability due to communication cost.
The aforementioned issues, naturally leads to the following question: \emph{\textit{Can we propose a provably communication efficient algorithm that is also distributionally robust?}} The purpose of this paper is to give an affirmative answer to this question by proposing a \textbf{Distributionally Robust Federated Averaging} ({\sffamily{DRFA}}) algorithm that is distributionally robust, while being communication-efficient via periodic averaging,
and partial node participation, as we show both theoretically and empirically. From a high-level algorithmic perspective, we develop an approach to analyze minimax optimization methods where model parameter $\bm{w}$ is trained distributedly at local devices, and mixing parameter $\bm{\lambda}$ is only updated at server periodically. Specifically, each device optimizes its model locally, and a subset of them are adaptively sampled based on $\boldsymbol{\lambda}$ to perform model averaging. We note that since $\bm{\lambda}$ is updated only at synchronization rounds, it will inevitably hurt the convergence rate. Our key technical contribution is the introduction and analysis of a \textit{randomized snapshotting schema} to approximate the accumulation of history of local gradients to update $\bm{\lambda}$ as to entail good convergence.\vspace{-1mm}
\noindent\textbf{Contributions.} We summarize the main contributions of our work as follows:\vspace{-0.0cm}
\begin{itemize}[leftmargin=*]
\item To the best of our knowledge, the proposed {\sffamily{DRFA}} algorithm is the first to solve distributionally robust optimization in a communicationally efficient manner for federated learning, and to give theoretical analysis on heterogeneous (non-IID) data distributions. The proposed idea of decoupling the updating of $\bm{w}$ from $\boldsymbol{\lambda}$ can be integrated as a building block into other federated optimization methods, e.g.~\cite{karimireddy2019scaffold,li2018federated} to yield a distributionally robust solution.\vspace{1mm}
\item We derive the convergence rate of our algorithm when loss function is convex in $\bm{w}$ and linear in $\boldsymbol{\lambda}$, and establish an $O({1}/{T^{3/8}})$ convergence rate with only $O\left(T^{3/4}\right)$ communication rounds. For nonconvex loss, we establish convergence rate of $O({1}/{T^{1/8}})$ with only $O\left(T^{3/4}\right)$ communication rounds. Compared to~\cite{mohri2019agnostic}, we significantly reduce the communication rounds.\vspace{1mm}
\item For the regularized objectives, we propose a variant algorithm, dubbed as {\sffamily{DRFA-Prox}}, and prove that it enjoys the same convergence rate as {\sffamily{DRFA}}. We also analyze an alternative method for optimizing regularized objective and derive the convergence rate in strongly-convex-strongly-concave and non-convex (under PL condition)-strongly-concave settings.\vspace{1mm}
\item We demonstrate the practical efficacy of the proposed algorithm over competitive baselines through experiments on federated datasets.
\end{itemize}
\section{Related Work} \vspace{-1mm}
\textbf{Federated Averaging.} Recently, many federated methods have been considered in the literature. FedAvg, as a variant of local GD/SGD, is firstly proposed in~\cite{mcmahan2017communication} to alleviate the communication bottleneck in FL. The first convergence analysis of local SGD on strongly-convex smooth loss functions has established in~\cite{stich2018local} by showing an $O\left({1}/{T}\right)$ rate with only $O(\sqrt{T})$ communication rounds. The analysis of the convergence of local SGD for nonconvex functions and its adaptive variant is proposed in~\cite{haddadpour2019local}. The extension to heterogeneous data allocation and general convex functions, with a tighter bound, is carried out in~\cite{khaled2020tighter}.~\cite{haddadpour2019convergence} analyzed local GD and SGD on nonconvex loss functions as well as networked setting in a fully decentralized setting. The recent work~\cite{li2019convergence} analyzes the convergence of FedAvg under non-iid data for strongly convex functions. In~\cite{woodworth2020local,woodworth2020minibatch}, Woodworth et al compare the convergence rate of local SGD and mini-batch SGD, under homogeneous and heterogeneous settings respectively.
\textbf{Distributionally Robust Optimization.} There is a rich body of literature on Distributionally Robust Optimization (DRO), and here, we try to list the most closely related work. DRO is an effective approach to deal with the imbalanced or non-iid data~\cite{namkoong2016stochastic,namkoong2017variance,fan2017learning,pmlr-v89-zhu19a,fan2017learning,mohri2019agnostic}, which is usually formulated as a minimax problem. A bandit mirror descent algorithm to solve the DRO minimax problem is proposed in~\cite{namkoong2016stochastic} . Another approach is to minimize top-k losses in the finite sum to achieves the distributional robustness~\cite{fan2017learning}. The first proposal of the DRO in federated learning is~\cite{mohri2019agnostic}, where they advocate minimizing the maximum combination of empirical losses to mitigate data heterogeneity.
\textbf{Smooth Minimax Optimization.} Another related line of work to this paper is the minimax optimization. One popular primal-dual optimization method is \emph{(stochastic) gradient descent ascent} or (S)GDA for short. The first work to prove that (S)GDA can converge efficiently on nonconvex-concave objectives is~\cite{lin2019gradient}. Other classic algorithms for the minimax problem are extra gradient descent (EGD)~\cite{korpelevich1976extragradient} and optimistic gradient descent (OGD), which are widely studied and applied in machine learning (e.g., GAN training~\cite{goodfellow2014generative,daskalakis2017training,liu2019decentralized,liang2018interaction}). The algorithm proposed in~\cite{thekumparampil2019efficient} combines the ideas of mirror descent and Nesterov's accelerated gradient descent (AGD)~\cite{nesterov1983method}, to achieve $\tilde{O}\left({1}/{T^2}\right)$ rate on strongly-convex-concave functions, and $\tilde{O}\left({1}/{T^{1/3}}\right)$ rate on nonconvex-concave functions. A proximally guided stochastic mirror descent and variance reduction gradient method (PGSMD/PGSVRG) for nonconvex-concave optimization is proposed in~\cite{rafique2018non}. Recently, an algorithm using AGD as a building block is designed in~\cite{lin2020near}, showing a linear convergence rate on strongly-convex-strongly-concave objective, which matches with the theoretical lower bound~\cite{zhang2019lower}. The decentralized minimax problem is studied in~\cite{srivastava2011distributed,mateos2015distributed,liu2019decentralized}, however, none of these works study the case where one variable is distributed and trained locally, and the other variable is updated periodically, similar to our proposal.
\section{Distributionally Robust Federated Averaging}\label{sec: DRFA}\vspace{-1mm}
We consider a federated setting where $N$ users aim to learn a global model in a collaborative manner without exchanging their data with each other. However, users can exchange information via a server that is connected to all users. Recall that the distributionally robust optimization problem can be formulated as
$
\min_{\boldsymbol{w} \in \mathcal{W}} \max_{\bm{\lambda}\in \Lambda } F (\boldsymbol{w},\boldsymbol{\lambda}) := \sum_{i=1}^N \lambda_i f_i(\boldsymbol{w})
$, where $f_i(\bm{w})$ is the local objective function corresponding to user $i$, which is often defined as the empirical or true risk over its local data. As mentioned earlier, we address this problem in a federated setting where we assume that $i$th local data shard is sampled from a local distribution $\mathcal{P}_i$-- possibly different from the distribution of other data shards. Our goal is to train a central model $\bm{w}$ with limited communication rounds. We will start with this simple setting where the global objective is linear in the mixing parameter $\bm{\lambda}$, and will show in Section~\ref{sec:regularized} that our algorithm can also provably optimize regularized objectives where a functional constraint is imposed on the mixing parameter, with a slight difference in the scheme to update $\boldsymbol{\lambda}$.
\vspace{-1mm}
\subsection{The proposed algorithm}\vspace{-2mm}
To solve the aforementioned problem, we propose {\sffamily{DRFA}} algorithm as summarized in Algorithm~\ref{alg:1}, which consists of two main modules: local model updating and periodic mixture parameter synchronization. The local model updating is similar to the common local SGD~\cite{stich2018local} or FedAvg~\cite{mcmahan2017communication}, however, there is a subtle difference in selecting the clients as we employ an adaptive sampling schema. To formally present the steps of {\sffamily{DRFA}}, let us define $S$ as the rounds of communication between server and users and $\tau$ as the number of local updates that each user runs between two consecutive rounds of communication. We use $T = S\tau$ to denote the total number of iterations the optimization proceeds.
\begin{algorithm}[t]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\caption{Distributionally Robust Federated Averaging ({\sffamily{DRFA}}) }
\label{alg:1}
\begin{algorithmic}[1]
\REQUIRE $N$ clients , synchronization gap $\tau$, total number of iterations $T$, $S=T/\tau$, learning rates $\eta$, $\gamma$, sampling size $m$, initial model $\bar{\boldsymbol{w}}^{(0)}$ and initial $\boldsymbol{\lambda}^{(0)}$.\\
\ENSURE Final solutions $\hat{\boldsymbol{w}} = \frac{1}{mT} \sum_{t=1}^{T}\sum_{i\in{\mathcal{D}^{(\floor{\frac{t}{\tau}})}}} \boldsymbol{w}^{(t)}_i$, $\hat{\boldsymbol{\lambda}} = \frac{1}{S}\sum_{s=0}^{S-1} \boldsymbol{\lambda}^{(s)}$, or (2) $\bm{w}^T$, $\boldsymbol{\lambda}^S$.
\FOR{ $s = 0$ to $S-1$ }
{\STATE {Server \textbf{samples} $\mathcal{D}^{(s)} \subset [N]$ according to $\boldsymbol{\lambda}^{(s)}$ with size of $m$}\\
\STATE {Server \textbf{samples} $t'$ from $ s\tau+1,\ldots,(s+1)\tau$} uniformly at random\\
\STATE Server \textbf{broadcasts} $\bar{\boldsymbol{w}}^{(s)}$ and $t'$ to all clients $i \in
\mathcal{D}^{(s)}$}\\[5pt]
\algemph{antiquewhite}{1}{
\FOR{clients $i \in \mathcal{D}^{(s)}$ \textbf{parallel}}
\STATE Client \textbf{sets} $\boldsymbol{w}_i^{(s\tau)} = \bar{\bm{w}}^{(s)}$
\FOR{$t = s\tau,\ldots,(s+1)\tau-1 $}
\STATE {$\bm{w}^{(t+1)}_i = \prod_{\mathcal{W}}\left(\bm{w}^{(t)}_i - \eta \nabla f_i(\bm{w}^{(t)}_i;\xi^{(t)}_i)\right) $}\\[-8pt]
\ENDFOR
\ENDFOR
\STATE {Client $i \in \mathcal{D}^{(s)}$ \textbf{sends} $\bm{w}^{((s+1)\tau)}_i$ and $\bm{w}^{(t')}_i$ back to the server}}
{\STATE {Server \textbf{computes} $\bar{\bm{w}}^{(s+1)} = \frac{1}{m} \sum_{i\in \mathcal{D}^{(s)}} \bm{w}^{((s+1)\tau)}_i$ }\\[-5pt]
\STATE Server \textbf{computes} $\bm{w}^{(t')} = \frac{1}{m} \sum_{i\in \mathcal{D}^{(s)}} \bm{w}^{(t')}_i$}\\
\algemph{blizzardblue}{1}{
\STATE {Server uniformly samples a subset $\mathcal{U} \subset [N]$ of clients with size $m$ \hfill\mycommfont{{// Update $\boldsymbol{\lambda}$}}}
\STATE {Server \textbf{broadcasts} $\bm{w}^{(t')}$ to each client $i \in \mathcal{U}$, compute $f_i(\bm{w}^{(t')};\xi_i )$ over a local minibatch}
\STATE {Make $N$-dimensional vector $\bm{v}$: $v_i = \frac{N}{m}f_i(\bm{w}^{(t')};\xi_i $) if $i \in \mathcal{U}$, otherwise ${v}_i = 0$}
\STATE { Server updates $\boldsymbol{\lambda}^{(s+1)} = \prod_{\Lambda}\left( \boldsymbol{\lambda}^{(s)} + \tau \gamma \bm{v} \right) $}
}
\ENDFOR
\end{algorithmic}
\end{algorithm}
\noindent\textbf{Periodic model averaging via adaptive sampling.}~Let $\bar{\bm{w}}^{(s)}$ and $\boldsymbol{\lambda}^{(s)}$ denote the global primal and dual parameters at server after synchronization stage $s-1$, respectively. At the beginning of the $s$th communication stage, server selects $m$ clients $\mathcal{D}^{(s)} \subset [N]$ randomly based on the probability vector $\boldsymbol{\lambda}^{(s)}$ and broadcasts its current model $\bar{\bm{w}}^{(s)}$ to all the clients $i \in \mathcal{D}^{(s)}$. Each client $i$, after receiving the global model, updates it using local SGD on its own data for $\tau$ iterations. To be more specific, let $\bm{w}^{(t+1)}_i$ denote the model at client $i$ at iteration $t$ within stage $s$. At each local iteration $t = s\tau, \ldots, (s+1)\tau$, client $i$ updates its local model according to the following rule
$$\bm{w}^{(t+1)}_i = \prod_{\mathcal{W}}\left(\bm{w}^{(t)}_i - \eta \nabla f_i(\bm{w}^{(t)}_i;\xi^{(t)}_i)\right), $$
where $\prod_{\mathcal{W}}(\cdot)$ is the projection onto $\mathcal{W}$ and the stochastic gradient is computed on a random sample $\xi^{(t)}_i$ picked from the $i$th local dataset. After $\tau$ local steps, each client sends its current model $\bm{w}^{((s+1)\tau)}_i$ to the server to compute the next global average primal model $\bar{\bm{w}}^{(s+1)} = (1/m)\sum_{i\in \mathcal{D}^{(s)}} \bm{w}^{((s+1)\tau)}_i$. This procedure is repeated for $S$ stages. We note that adaptive sampling not only addresses the scalability issue, but also leads to smaller communication load compared to full participation case
\noindent\textbf{Periodic mixture parameter updating.}~The global mixture parameter $\boldsymbol{\lambda}$ controls the mixture of different local losses, and can only be updated by server at synchronization stages. The updating scheme for $\boldsymbol{\lambda}$ will be different when the objective function is equipped with or without the regularization on $\boldsymbol{\lambda}$. In the absence of regularization on $\boldsymbol{\lambda}$, the problem is simply linear in $\boldsymbol{\lambda}$. A key observation is that in linear case, the gradient of $\boldsymbol{\lambda}$ only depends on $\bm{w}$, so we can approximate the sum of history gradients over the previous local period (which does not show up in the real dynamic). Indeed, between two synchronization stages, from iterations $s\tau+1$ to $(s+1)\tau$, in the \textit{fully synchronized} setting~\cite{mohri2019agnostic}, we can update $\boldsymbol{\lambda}$ according to \vspace{-2mm}
\begin{equation*}
\boldsymbol{\lambda}^{(s+1)} =\prod_{\Lambda}\left(\boldsymbol{\lambda}^{(s)} +\gamma \sum_{t = s\tau+1}^{(s+1)\tau} \nabla_{\boldsymbol{\lambda}}F( \bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)})\right)
\end{equation*}
where $\bm{w}^{(t)} = \frac{1}{m}\sum_{i\in \mathcal{D}^{(s)}} \bm{w}_i^{(t)}$ is the average model at iteration $t$.
To approximate this update, we propose a \textit{random snapshotting} schema as follows. At the beginning of the $s$th communication stage, server samples a random iteration $t'$ (snapshot index) from the range of $s\tau+1$ to $(s+1)\tau$ and sends it to sampled devices $\mathcal{D}^{(s)}$ along with the global model. After the local updating stage is over, every selected device sends its local model at index $t'$, i.e., $\bm{w}^{(t')}_i$, back to the server. Then, server computes the average model $\bm{w}^{(t')} = \frac{1}{|\mathcal{D}^{(s)}|} \sum_{i\in\mathcal{D}^{(s)}}\bm{w}^{(t')}_i$, that will be used for updating the mixture parameter $\boldsymbol{\lambda}^{(s)}$ to $\boldsymbol{\lambda}^{(s+1)}$ ($\boldsymbol{\lambda}^{(s+1)}$ will be used at stage $s+1$ for sampling another subset of users $\mathcal{D}^{(s+1)}$). To simulate the update we were supposed to do in the fully synchronized setting, server broadcasts $\bm{w}^{(t')}$ to a set $\mathcal{U}$ of $m$ clients, selected uniformly at random, to stochastically evaluate their local losses $f_i(\cdot), i \in \mathcal{U}$ at $ \bm{w}^{(t')}$ using a random minibatch $\xi_i$ of their local data. After receiving evaluated losses, server will construct the vector $\bm{v}$ as in Algorithm~\ref{alg:1}, where $v_i = \frac{N}{m} f_i(\bm{w}^{(t')};\xi_i), i \in \mathcal{U}$ to compute a stochastic gradient at dual parameter. We claim that this is an \textit{unbiased estimation} by noting the following identity:
\begin{equation}
\mathbb{E}_{t',\mathcal{U},\xi_i}\left[\tau \bm{v}\right] = \mathbb{E}_{t'}\left[\tau \nabla_{\boldsymbol{\lambda}}F\left( \bm{w}^{(t')}, \boldsymbol{\lambda}^{(s)}\right)\right] = \sum_{t = s\tau+1}^{(s+1)\tau} \nabla_{\boldsymbol{\lambda}}F\left(\bm{w}^{(t)}, \boldsymbol{\lambda}^{(s)}\right) \label{eq: approximation gradient}.
\end{equation}
However, the above estimation has a high variance in the order of $O(\tau^2)$, so a crucial question that we need to address is finding the proper choice of $\tau$ to guarantee convergence, while minimizing the overall communication cost. We also highlight that unlike local SGD, the proposed algorithm requires two rounds of communication at each synchronization step for decoupled updating of parameters.
\section{Convergence Analysis}
In this section, we present our theoretical results on the guarantees of the {\sffamily{DRFA}} algorithm for two general class of convex and nonconvex smooth loss functions. All the proofs are deferred to appendix.
\noindent\textbf{Technical challenge.}~Before stating the main results we would like to highlight one of the main theoretical challenges in proving the convergence rate. In particular, a key step in analyzing the local descent methods with periodic averaging is to bound the deviation between local and (virtual) global at each iteration. In minimizing empirical risk (finite sum),~\cite{khaled2019better} gives a tight bound on the deviation of a local model from averaged model which depends on the quantity $\frac{1}{N}\sum_{i=1}^N \|\nabla f_i(\bm{x}^*)\|^2$, where $\bm{x}^*$ is the minimizer of $\frac{1}{N}\sum_{i=1}^N f_i(\bm{x})$. However, their analysis is not generalizable to minimax setting, as the dynamic of primal-dual method will change the minimizer of $F(\cdot, \boldsymbol{\lambda}^{(s)})$ every time $\boldsymbol{\lambda}^{(s)}$ is updated, which makes the analysis more challenging compared to the average loss case. In light of this and in order to subject heterogeneity of local distributions to a more formal treatment in minimax setting, we introduce a quantity to measure dissimilarity among local gradients.
\begin{definition}[Weighted Gradient Dissimilarity]
A set of local objectives $f_i(\cdot), i=1, 2, \ldots, N$ exhibit $
\Gamma$ gradient dissimilarity defined as
$ \Gamma := \sup_{\bm{w} \in \mathcal{W}, \boldsymbol{p}\in \Lambda,i\in [n],} \sum_{j\in [n]}p_j\|\nabla f_i(\bm{w})- \nabla f_{j}(\bm{w})\|^2$.
\end{definition}
The above notion is a generalization of gradient dissimilarity, which is employed in the analysis of local SGD in federated setting~\cite{li2019communication,deng2020adaptive,li2019convergence,woodworth2020minibatch}. This quantity will be zero if and only if all local functions are identical. The obtained bounds will depend on the gradient dissimilarity as local updates only employ samples from local data with possibly different statistical realization.
We now turn to analyzing the convergence of the proposed algorithm. Before, we make the following customary assumptions:
\begin{assumption} [Smoothness/Gradient Lipschitz]\label{assumption: smoothness}
Each component function $f_i(\cdot), i=1,2, \ldots, N$ and global function $F(\cdot, \cdot)$ are $L$-smooth, which implies: $
\| \nabla f_i(\bm{x}_1) - \nabla f_i(\bm{x}_2)\| \leq L \|\bm{x}_1 - \bm{x}_2\|, \forall i \in [N], \forall \bm{x}_1, \bm{x}_2$ and $
\| \nabla F(\bm{x}_1,\bm{y}_1) - \nabla F(\bm{x}_2,\bm{y}_2)\| \leq L \|(\bm{x}_1,\bm{y}_1) - (\bm{x}_2,\bm{y}_2)\|, \forall (\bm{x}_1,\bm{y}_1), (\bm{x}_2,\bm{y}_2)$.
\end{assumption}
\begin{assumption} [Gradient Boundedness] \label{assumption: bounded gradient} The gradient w.r.t $\boldsymbol{w}$ and $\boldsymbol{\lambda}$ are bounded, i.e., $\|\nabla f_i(\boldsymbol{w})\|\leq G_w $ and $\|\nabla_{\boldsymbol{\lambda}}F(\boldsymbol{w},\boldsymbol{\lambda})\|\leq G_{\lambda}$.
\end{assumption}
\begin{assumption} [Bounded Domain] \label{assumption: bounded domain} The diameters of $\mathcal{W}$ and $ \Lambda$ are bounded by $D_{\mathcal{W}}$ and $D_{\Lambda}$.
\end{assumption}
\begin{assumption}[Bounded Variance] \label{assumption: bounded variance} Let $\Tilde{\nabla} F(\bm{w};\boldsymbol{\lambda}) $ be stochastic gradient for $\boldsymbol{\lambda}$, which is the $N$-dimensional vector such that the $i$th entry is $f_i(\bm{w};\xi)$, and the rest are zero. Then we assume $
\|\nabla f_i(\bm{w};\xi) - \nabla f_i(\bm{w})\| \leq \sigma^2_{w}, \forall i \in [N]$ and $
\|\Tilde{\nabla} F(\bm{w};\boldsymbol{\lambda}) - \nabla F(\bm{w};\boldsymbol{\lambda})\| \leq \sigma^2_{\lambda}$.
\end{assumption}
\subsection{Convex losses}
The following theorem establishes the convergence rate of primal-dual gap for convex objectives.
\begin{theorem} \label{theorem1}
Let each local function $f_i$ be convex, and global function $F$ be linear in $\boldsymbol{\lambda}$. Assume the conditions in Assumptions~\ref{assumption: smoothness}-\ref{assumption: bounded variance} hold. If we optimize (\ref{agnostic loss}) using Algorithm~\ref{alg:1} with synchronization gap $\tau = \frac{T^{1/4}}{\sqrt{m}}$, learning rates $\eta = \frac{1}{4L \sqrt{T}}$ and $\gamma = \frac{1}{T^{5/8}}$, for the returned solutions $\hat{\boldsymbol{w}}$ and $\hat{\boldsymbol{\lambda}}$ it holds that
\begin{equation*}
{\small
\begin{aligned}
\max_{\boldsymbol{\lambda}\in \Lambda}\mathbb{E}[F(\hat{\boldsymbol{w}},\boldsymbol{\lambda} )] -\min_{\bm{w}\in\mathcal{W}} \mathbb{E}[F(\boldsymbol{w} ,\hat{\boldsymbol{\lambda}} )] \leq O\Big{(}\frac{D_{\mathcal{W}}^2+G_{w}^2}{\sqrt{T}} +\frac{D_{\Lambda}^2}{T^{3/8}} +\frac{G_{\lambda}^2}{m^{1/2}T^{3/8}} +\frac{\sigma_{\lambda}^2}{m^{3/2}T^{3/8}}+ \frac{\sigma_w^2+\Gamma}{m\sqrt{T} }\Big{)}.
\end{aligned}}
\end{equation*}
\end{theorem}
The proof of Theorem~\ref{theorem1} is deferred to Appendix~\ref{sec: proof DRFA convex}. Since we update $\boldsymbol{\lambda}$ only at the synchronization stages, it will almost inevitably hurt the convergence. The original agnostic federated learning~\cite{mohri2019agnostic} using SGD can achieve an $O(1/\sqrt{T})$ convergence rate, but we achieve a slightly slower rate $O\left(1/ T^{3/8}\right)$ to reduce the communication complexity from $O(T)$ to $O(T^{3/4})$. Indeed, we trade $O(T^{1/8})$ convergence rate for $O(T^{1/4})$ communication rounds. As we will show in the proof, if we choose $\tau$ to be a constant, then we recover the same $O(1/{\sqrt{T}})$ rate as \cite{mohri2019agnostic}. Also, the dependency of the obtained rate does not demonstrate a linear speedup in the number of sampled workers $m$. However, increasing $m$ will also accelerate the rate, but does not affect the dominating term. We leave tightening the obtained rate to achieve a linear speedup in terms of $m$ as an interesting future work.
\subsection{Nonconvex losses}
We now proceed to state the convergence in the case where local objectives $f_i, i \in [N]$ are nonconvex, e.g., neural networks. Since $f_i$ is no longer convex, the primal-dual gap is not a meaningful quantity to measure the convergence. Alternatively, following the standard analysis of nonconvex minimax optimization, one might consider the following functions to facilitate the analysis.
\begin{definition} \label{df1}
We define function $\Phi (\cdot)$ at any primal parameter $\bm{w}$ as:
\begin{eqnarray}
\Phi (\boldsymbol{w}) := F(\boldsymbol{w},\boldsymbol{\lambda}^*(\boldsymbol{w})), \quad \text{where} \; \boldsymbol{\lambda}^*(\boldsymbol{w}) := \arg \max_{\boldsymbol{\lambda}\in \Lambda} F(\boldsymbol{w},\boldsymbol{\lambda}) \label{Phi}.
\end{eqnarray}
\end{definition}
However, as argued in~\cite{lin2019gradient}, on nonconvex-concave(linear) but not strongly-concave objective, directly using $\|\nabla \Phi(\bm{w})\|$ as convergence measure is still difficult for analysis. Hence, Moreau envelope of $\Phi$ can be utilized to analyze the convergence as used in several recent studies~\cite{davis2019stochastic,lin2019gradient,rafique2018non}.
\begin{definition}[Moreau Envelope] A function $\Phi_{p} (\bm{x})$ is the $p$-Moreau envelope of a function $\Phi$ if $ \Phi_{p} (\bm{x}) := \min_{\bm{w}\in \mathcal{W}} \left\{ \Phi (\bm{w}) + \frac{1}{2p}\|\bm{w}-\bm{x}\|^2\right\}$.
\end{definition}
We will use $1/2L$-Moreau envelope of $\Phi$, following the setting in~\cite{lin2019gradient,rafique2018non}, and state the convergence rates in terms of $\|\nabla \Phi_{1/2L} (\bm{w})\|$.
\begin{theorem} \label{thm:nonconvex-linear}
Assume each local function $f_i$ is nonconvex, and global function $F$ is linear in $\boldsymbol{\lambda}$. Also, assume the conditions in Assumptions~\ref{assumption: smoothness}-\ref{assumption: bounded variance} hold. If we optimize (\ref{agnostic loss}) using Algorithm~\ref{alg:1} with synchronization gap $\tau = T^{1/4}$, letting $\bm{w}^t = \frac{1}{m}\sum_{\mathcal{D}^{(\floor{ \frac{t}{\tau}})}} \bm{w}_i^{(t)}$ to denote the virtual average model at $t$th iterate, by choosing $\eta = \frac{1}{4LT^{3/4}} $ and $\gamma = \frac{1}{\sqrt{T}} $, we have:
{\small\begin{align*}{ \frac{1}{T}\sum_{t=1}^T \mathbb{E}\left[\left\| \nabla \Phi_{1/2L}(\bm{w}^{t}) \right\|^2 \right] \leq O\left(\frac{D_{\Lambda}^2}{T^{1/8}}+ \frac{ \sigma_{\lambda}^2}{mT^{1/4}}+ \frac{G_{\lambda}^2}{ T^{1/4}}+ \frac{G_w \sqrt{G_w^2+\sigma_w^2}}{T^{1/8}} + \frac{D_{\mathcal{W}}(\sigma_w+\sqrt{\Gamma})}{T^{1/2} }\right)}.
\end{align*}}
\end{theorem}
The proof of Theorem~\ref{thm:nonconvex-linear} is deferred to Appendix~\ref{sec: proof DRFA nonconvex}. We obtain an $O\left({1}/{T^{1/8}}\right)$ rate here, with $O\left({1}/{T^{3/4}}\right)$ communication rounds. Compared to SOTA algorithms proposed in~\cite{lin2019gradient,rafique2018non} in nonconvex-concave setting which achieves an $O\left({1}/{T^{1/4}}\right)$ rate in a single machine setting, our algorithm is distributed and communication efficient. Indeed, we trade $O\left({1}/{T^{1/8}}\right)$ rate for saving $O\left({1}/{T^{1/4}}\right)$ communications. One thing worth noticing is that in~\cite{lin2019gradient}, it is proposed to use a smaller step size for the primal variable than dual variable, while here we choose a small step size for dual variable too. That is mainly because the approximation of dual gradients in our setting introduces a large variance which necessities to employ smaller rate to compensate for high variance. Also, the number of participated clients will not accelerate the leading term, unlike vanilla local SGD or its variants~\cite{mcmahan2017communication,stich2018local,karimireddy2019scaffold}.
\begin{algorithm}[t]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\caption{Distributionally Robust Federated Averaging: Proximal Method (\sffamily{DRFA-Prox})}
\label{alg:3}
\begin{algorithmic}[1]
\REQUIRE The algorithm is identical to Algorithm~\ref{alg:1} except the updating rule for $\boldsymbol{\lambda}$.
\algemph{blizzardblue}{1} {
\STATE {\quad Server uniformly samples a subset $\mathcal{U} \subset [N]$ of clients with size $m$ \hfill\mycommfont{{// Update $\boldsymbol{\lambda}$}}}
\STATE {\quad Server \textbf{broadcasts} $\bm{w}^{(t')}$ to each client $i \in \mathcal{U}$}
\STATE {\quad Each client $i \in \mathcal{U}$ computes $f_i(\bm{w}^{(t')};\xi_i)$ over a local minibatch $\xi_i$ and sends to server}
\STATE {\quad Server computes $N$-dimensional vector $\bm{v} $: ${v}_i = \frac{N}{m}f_i(\bm{w}^{(t')};\xi_i $) if $i \in \mathcal{U}$, otherwise ${v}_i = 0$ }
\STATE { \quad Server updates $\boldsymbol{\lambda}^{(s+1)} = \arg\max_{\bm{u}\in \Lambda} \left \{\tau g(\bm{u}) - \frac{1}{2\gamma}\|\boldsymbol{\lambda}^{(s)}+\gamma \tau \bm{v} - \bm{u} \|^2 \right \}. $}
}
\end{algorithmic}
\end{algorithm}
\section{DRFA-Prox: Optimizing Regularized Objective}\label{sec:regularized}
As mentioned before, our algorithm can be generalized to impose a regularizer on $\bm{\lambda}$ captured by a regularization function $g(\boldsymbol{\lambda})$ and to solve the following minimax optimization problem:
\begin{equation}
\min_{\boldsymbol{w} \in \mathcal{W}} \max_{\boldsymbol{\lambda}\in \Lambda } F (\boldsymbol{w},\boldsymbol{\lambda}) := \left\{f(\boldsymbol{w},\boldsymbol{\lambda}):=\sum_{i=1}^N \lambda_i f_i(\boldsymbol{w}) \right\} +g (\boldsymbol{\lambda}) . \label{regularized loss}
\end{equation}
The regularizer $g(\boldsymbol{\lambda})$ can be introduced to leverage the domain prior, or to make the $\boldsymbol{\lambda}$ update robust to adversary (e.g., the malicious node may send a very large fake gradient of $\boldsymbol{\lambda}$). The choices of $g$ include KL-divergence, optimal transport~\cite{kantorovich2006translocation,monge1781memoire}, or $\ell_p$ distance.
In regularized setting, by examining the structure of the gradient w.r.t. $\boldsymbol{\lambda}$, i.e.,
$ \nabla_{\boldsymbol{\lambda}} F(\bm{w},\boldsymbol{\lambda}) = \nabla_{\boldsymbol{\lambda}} f(\bm{w},\boldsymbol{\lambda}) + \nabla_{\boldsymbol{\lambda}} g(\boldsymbol{\lambda}) \nonumber.
$, while $\nabla_{\bm{\lambda}} f(\bm{w},\boldsymbol{\lambda})$ is independent of $\boldsymbol{\lambda}$, but $\nabla_{\boldsymbol{\lambda}} g(\boldsymbol{\lambda})$ has dependency on $\boldsymbol{\lambda}$, and consequently our approximation method in Section~\ref{sec: DRFA} is not fully applicable here. Inspired by the proximal gradient methods~\cite{beck2017first,nesterov2013gradient,beck2009fast}, which is widely employed in the problems where the gradient of the regularized term is hard to obtain, we adapt a similar idea, and propose a proximal variant of {\sffamily{DRFA}}, called {\sffamily{DRFA-Prox}}, to tackle regularized objectives. In {\sffamily{DRFA-Prox}}, the only difference is the updating rule of $\boldsymbol{\lambda}$ as detailed in Algorithm~\ref{alg:3}. We still employ the gradient approximation in {\sffamily{DRFA}} to estimate history gradients of $\nabla_{\bm{\lambda}} f$, however we utilize proximity operation to update $\boldsymbol{\lambda}$:
\begin{align}
\boldsymbol{\lambda}^{(s+1)} = \arg\max_{\bm{u}\in \Lambda} \left \{\tau g(\bm{u}) - \frac{1}{2\gamma}\|\boldsymbol{\lambda}^{(s)}+\gamma \tau \bm{v} - \bm{u} \|^2 \right \}.\nonumber
\end{align}
As we will show in the next subsection, {\sffamily{DRFA-Prox}} enjoys the same convergence rate as {\sffamily{DRFA}}, both on convex and nonconvex losses.
\subsection{Convergence of DRFA-Prox}
The following theorems establish the convergence rate of {\sffamily{DRFA-Prox}} for convex and nonconvex objectives in federated setting.
\begin{theorem}[Convex loss] \label{thm: regularized convex-linear}
Let each local function $f_i$ be convex. Assume the conditions in Assumptions~\ref{assumption: smoothness}-\ref{assumption: bounded variance} hold. If we optimize (\ref{regularized loss}) using Algorithm~\ref{alg:3} with synchronization gap $\tau = \frac{T^{1/4}}{\sqrt{m}}$, $\eta = \frac{1}{4L \sqrt{T}}$, and $\gamma = \frac{1}{T^{5/8}}$, for the returned solutions $\hat{\boldsymbol{w}}$ and $\hat{\boldsymbol{\lambda}}$ it holds that:
\begin{equation}
{\small
\begin{aligned}
\min_{\bm{w}\in\mathcal{W}} \max_{\boldsymbol{\lambda}\in \Lambda}\mathbb{E}[F(\hat{\boldsymbol{w}},\boldsymbol{\lambda} ) & - F(\boldsymbol{w} ,\hat{\boldsymbol{\lambda}} )] \leq O\Big{(}\frac{D_{\mathcal{W}}^2+G_{w}^2}{\sqrt{T}} +\frac{D_{\Lambda}^2}{T^{3/8}}+\frac{G_\lambda^2 }{m^{1/2}T^{3/8}} +\frac{\sigma_{\lambda}^2}{m^{3/2}T^{3/8}}+ \frac{\sigma_w^2+\Gamma}{m \sqrt{T} }\Big{)}.\nonumber
\end{aligned}}
\end{equation}
\end{theorem}
The proof of Theorem~\ref{thm: regularized convex-linear} is deferred to Appendix~\ref{sec: proof DRFA-Prox convex}. Clearly, we obtain a convergence rate of $O\left({1}/{T^{3/8}}\right)$, which is same as rate obtained in Theorem~\ref{theorem1} for {\sffamily{DRFA}} in non-regularized case.
\begin{theorem}[Nonconvex loss] \label{thm:regularized nonconvex-linear}
Assume each local function $f_i$ is nonconvex. Also, assume the conditions in Assumptions~\ref{assumption: smoothness}-\ref{assumption: bounded variance} hold. If we optimize (\ref{regularized loss}) using Algorithm~\ref{alg:3} with synchronization gap $\tau = T^{1/4}$, letting $\bm{w}^t = \frac{1}{m}\sum_{\mathcal{D}^{(\floor{ \frac{t}{\tau}})}} \bm{w}_i^{(t)}$ to denote the virtual average model at $t$th iterate, by choosing $\eta = \frac{1}{4LT^{3/4}} $ and $\gamma = \frac{1}{\sqrt{T}} $, we have:
{\small\begin{align*}{ \frac{1}{T}\sum_{t=1}^T \mathbb{E}\left[\left\| \nabla \Phi_{1/2L}(\bm{w}^{t}) \right\|^2 \right] \leq O\left(\frac{D_{\Lambda}^2 }{T^{1/8}}+ \frac{ \sigma_{\lambda}^2}{mT^{1/4}}+ \frac{ G_{\lambda}^2}{T^{1/4}}+ \frac{G_w \sqrt{G_w^2+\sigma_w^2}}{T^{1/8}} + \frac{D_{\mathcal{W}} (\sigma_w+\sqrt{\Gamma})}{T^{1/2} }\right)}.
\end{align*}}
\end{theorem}
The proof of Theorem~\ref{thm:regularized nonconvex-linear} is deferred to Appendix~\ref{sec: proof DRFA-Prox nonconvex}. Note that, we recover the same convergence rate as {\sffamily{DRFA}} on nonconvex losses (Theorem~\ref{theorem2}). However, we should remark that solving the proximal problem will take extra computation time, which is not reflected in the convergence rate.
\subsection{An alternative algorithm for regularized objective}\label{sec:alternative algorithm}
Here we present an alternative method similar to vanilla AFL~\cite{mohri2019agnostic} to optimize regularized objective~(\ref{regularized loss}), where we choose to do the full batch gradient ascent for $\boldsymbol{\lambda}$ every $\tau$ iterations according to $ \boldsymbol{\lambda}^{(s+1)} = \prod_{\Lambda}\left( \boldsymbol{\lambda}^{(s)} + \gamma \nabla_{\lambda} F\left(\bar{\bm{w}}^{(s)}, \boldsymbol{\lambda}^{(s)}\right)\right)
$. We establish convergence rates in terms of $\Phi(\bm{w})$ as in Definition~\ref{df1}, under assumption that $F(\cdot, \bm{\lambda})$ is strongly-convex or satisfies PL-condition~\cite{karimi2016linear} in $\bm{w}$, and strongly-concave in $\boldsymbol{\lambda}$. Due to lack of space, we present a summary of the rates and defer the exact statements to Appendix~\ref{app: alternative algorithm} and the proofs to Appendices~\ref{sec:scsc} and~\ref{sec:ncsc}.
\noindent \textbf{Strongly-convex-strongly-concave case.} In this setting, we obtain an $\tilde{O}\left({\tau}/{T}\right)$ rate. If we choose $\tau = 1$, which is fully synchronized SGDA, then we recover the same rate $\tilde{O}\left({1}/{T}\right)$ as in~\cite{mohri2019agnostic}. If we choose $\tau$ to be $O(\sqrt{T/m})$, we recover the rate $ \Tilde{O}\left( {1}/{\sqrt{mT}} \right)$, which achieves a linear speedup in the number of sampled workers (see Theorem~\ref{theorem2} in Appendix~\ref{app: alternative algorithm}).
\noindent\textbf{Nonconvex (PL condition)-strongly-concave case.} We also provide the convergence analysis when $F$ is nonconvex but satisfying the PL condition~\cite{karimi2016linear} in $\bm{w}$, and strongly concave in $\boldsymbol{\lambda}$. In this setting, we also obtain an $\tilde{O}\left({ \tau}/{T}\right)$ convergence rate which is slightly worse than that of strongly-convex-strongly-concave case. The best known result of non-distributionally robust version of FedAvg on PL condition is $O({1}/{T})$ \cite{haddadpour2019convergence}, with $O(T^{1/3})$ communication rounds. It turns out that we trade some convergence rates to guarantee worst-case performance (see Theorem~\ref{thm3} in Appendix~\ref{app: alternative algorithm}).
\section{Experiments}
In this section, we empirically verify {\sffamily{DRFA}} and compare its performance to other baselines. More experimental results are discussed in the Appendix~\ref{app:add_exp}. We implement our algorithm based on \texttt{Distributed} API of PyTorch~\cite{paszke2019pytorch} using \texttt{MPI} as our main communication interface, and on an \texttt{Intel Xeon E5-2695} \texttt{CPU} with $28$ cores. We use three datasets, namely, Fashion MNIST~\cite{xiao2017/online}, Adult~\cite{adult}, and Shakespeare~\cite{caldas2018leaf} datasets. The code repository used for these experiments can be found at: \url{https://github.com/MLOPTPSU/FedTorch/}
\noindent\textbf{Synchronization gap.}~To show the effects of synchronization gap on {\sffamily{DRFA}} algorithm, we run the first experiment on the Fashion MNIST dataset with logistic regression as the model. We run the experiment with $10$ devices and a server, where each device has access to only one class of data, making it distributionally heterogeneous. We use different synchronization gaps of $\tau\in\{5,10,15\}$, and set $\eta=0.1$ and $\gamma=8\times10^{-3}$. The results are depicted in Figure~\ref{comp_tau_fashion_mnist}, where out of all the test accuracies on each single local distribution, we report the worst one as the worst distribution accuracy.
Based on our optimization scheme, we aim at optimizing the worst distribution accuracy (or loss), thus the measure depicted in Figure~\ref{comp_tau_fashion_mnist} is in accordance with our goal in the optimization.
It can be inferred that the smaller the synchronization gap is, the fewer number of iterations required to achieve $50\%$ accuracy in the worst distribution (Figure~\ref{fig:comp_fashion_mnnist_iteration}). However, the larger synchronization gap needs fewer number of communication and shorter amount of time to achieve $50\%$ accuracy in the worst distribution (Figure~\ref{fig:comp_fashion_mnist_comm_round} and~\ref{fig:comp_fashion_mnist_wall_clock}).
\begin{figure*}[t!]
\centering
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/tau/wors_acc_fashion_mnist_iteration_compare_tau.pdf}
\label{fig:comp_fashion_mnnist_iteration}
}
\hfill
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/tau/wors_acc_fashion_mnist_comm_round_compare_tau.pdf}
\label{fig:comp_fashion_mnist_comm_round}
}
\hfill
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/tau/wors_acc_fashion_mnist_wall_clock_compare_tau.pdf}
\label{fig:comp_fashion_mnist_wall_clock}
}
\caption[]{Comparing the effects of synchronization gap on the {\sffamily{DRFA}} algorithm on the Fashion MNIST dataset with a logistic regression model. The figures are showing the worst distribution accuracy during the training. }
\label{comp_tau_fashion_mnist}
\vspace{-0.4cm}
\end{figure*}
\begin{figure*}[t!]
\centering
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/baseline/worst_acc_comm_compare_fashion_mnist_comm_round_base.pdf}
\label{fig:comp_base_fashion_mnist_comm_round}
}
\hfill
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/baseline/worst_acc_comm_compare_fashion_mnist_iteration_base.pdf}
\label{fig:comp_base_fashion_mnnist_iteration}
}
\hfill
\subfigure[]{
\centering
\includegraphics[width=0.31\textwidth]{Figures/baseline/acc_stdev_comm_compare_fashion_mnist_comm_round_base.pdf}
\label{fig:stdev}
}
\caption[]{Comparing {\sffamily{DRFA}} algorithm with AFL~\cite{mohri2019agnostic}, q-FedAvg~\cite{li2019fair}, and FedAvg on Fashion MNIST dataset with logistic regression. {\sffamily{DRFA}} can achieve the same level of worst distribution accuracy, with fewer number of communication rounds, and hence, lower runtime. It also efficiently decreases the variance among the performance of different nodes with fewer communication rounds.}
\label{comp_base_fashion_mnist1}\vspace{-0.4cm}
\end{figure*}
\noindent\textbf{Comparison with baselines.} From the algorithmic point of view, the AFL algorithm~\cite{mohri2019agnostic} is a special case of our {\sffamily{DRFA}} algorithm, by setting the synchronization gap $\tau=1$. Hence, the first experiment suggests that we can increase the synchronization gap and achieve the same level of worst accuracy among distributions with fewer number of communications. In addition to AFL, q-FedAvg proposed by~\citet{li2019fair} aims at balancing the performance among different clients, and hence, improving the worst distribution accuracy. In this part, we compare {\sffamily{DRFA}} with AFL, q-FedAVG, and FedAvg.
To compare them, we run our algorithm, as well as AFL, q-FedAvg and FedAvg on Fashion MNIST dataset with logistic regression model on $10$ devices, each of which has access to one class of data. We set $\eta=0.1$ for all algorithms, $\gamma=8\times 10^{-3}$ for {\sffamily{DRFA}} and AFL, and $q=0.2$ for q-FedAvg. The batch size is $50$ and synchronization gap is $\tau=10$.
Figure~\ref{fig:comp_base_fashion_mnnist_iteration} shows that AFL can reach to the $50\%$ worst distribution accuracy with fewer number of local iterations, because it updates the primal and dual variables at every iteration.
However, Figure~\ref{fig:comp_base_fashion_mnist_comm_round} shows that {\sffamily{DRFA}} outperforms AFL, q-FedAvg and FedAvg in terms of number of communications, and subsequently, wall-clock time required to achieve the same level of worst distribution accuracy (due to much lower number of communication needed).
\begin{wrapfigure}{r}{0.33\linewidth}
\centering
\vspace{-0.3cm}
\includegraphics[width=0.33\textwidth]{Figures/baseline/acc_comm_compare_fashion_mnist_comm_round_base.pdf}\vspace{-1mm}
\caption[]{Averag global accuracy for each algorithm for $100$ rounds of communication. It shows that {\sffamily{DRFA}} keeps the same level of global accuracy as FedAvg, while it boosts its worst performing distribution accuracy.}\label{comp_base_fashion_mnist}
\vspace{-0.3cm}
\end{wrapfigure}
Note that, q-FedAvg has is very close to AFL in terms of communication rounds, but it is far behind it in terms of local computations.
Also, note that FedAvg has the same computation complexity as {\sffamily{DRFA}} and q-FedAvg at each round but cannot reach the $50\%$ accuracy even after $300$ rounds of communication.
Similar to q-FedAvg, to show how different devices are performing, Figure~\ref{fig:stdev} depicts the standard deviation among the accuracy of different clients, which shows the level of fairness of the learned model among different clients. It can be inferred that {\sffamily{DRFA}} can achieve the same level as AFL and q-FedAvg with fewer number of communication rounds, making it more efficient.
To compare the average performance of these algorithms, Figure~\ref{comp_base_fashion_mnist} shows the global training accuracy of them over $100$ rounds of communication on Fashion MNIST with logistic regression, where {\sffamily{DRFA}} performs as good as FedAvg in this regard. AFL needs more communication rounds to reach to the same level.
\section{Conclusion}
In this paper we propose a communication efficient scheme for distributionally robust federated model training. In addition, we give the first analysis of local SGD in distributed minimax optimization, under general smooth convex-linear, and nonconvex linear, strongly-convex-strongly-concave and nonconvex (PL-condition)-strongly concave settings. The experiments demonstrate the convergence of our method, and the distributional robustness of the learned model. The future work would be improving obtained convergence rates due to gap we observed compared to centralized case. Another interesting question worth exploring will be investigating variance reduction schemes to achieve faster rates, in particular for updating mixing parameter.
\clearpage
\input{impact}
\section*{Acknowledgements}
This work has been done using the Extreme Science and Engineering Discovery Environment (XSEDE) resources, which is supported by National Science Foundation under grant number ASC200045. We are also grateful for the GPU donated by NVIDIA that was used in this research.
|
\section{Introduction}
\label{sec:intro}
\begin{figure}
\centering
\tabcolsep=0.05cm
\begin{tabular}{r|ccc}
{Source } \rotatebox{90}{ Reference}
&\includegraphics[bb=0 0 128 128,width=0.10\textwidth]{figures/figure1/2773005.png}
& \includegraphics[bb=0 0 128 128,width=0.10\textwidth]{figures/figure1/1795111.png}
& \includegraphics[bb=0 0 128 128,width=0.10\textwidth]{figures/figure1/2842126.png}
\\
\hline
\includegraphics[bb=0 0 128 128,width=0.10\textwidth]{figures/figure1/imgHQ11831.png}
&\includegraphics[bb=0 0 128 138,width=0.10\textwidth]{figures/figure1/imgHQ11831_2773005_AniGAN.png}
& \includegraphics[bb=0 0 128 138,width=0.10\textwidth]{figures/figure1/imgHQ11831_1795111_AniGAN.png}
& \includegraphics[bb=0 0 128 138,width=0.10\textwidth]{figures/figure1/imgHQ11831_2842126_AniGAN.png}
\\
\includegraphics[bb=0 0 128 128,width=0.10\textwidth]{figures/figure1/imgHQ11853.png}
&\includegraphics[bb=0 0 128 128,width=0.10\textwidth]{figures/figure1/imgHQ11853_2773005_AniGAN.png}
& \includegraphics[bb=0 0 128 128,width=0.10\textwidth]{figures/figure1/imgHQ11853_1795111_AniGAN.png}
& \includegraphics[bb=0 0 128 128,width=0.10\textwidth]{figures/figure1/imgHQ11853_2842126_AniGAN.png}
\end{tabular}
\caption{\small Illustration of some synthesized results with the proposed AniGAN for style-guided face-to-anime translation. The first row and the first column show reference anime-faces and source photo-faces, respectively. The remaining columns show the high-quality anime-faces with diverse styles generated by AniGAN, given source photo-faces with large pose variations and multiple reference anime-faces with different styles. }
\label{fig:figure1}
\vspace*{-18pt}
\end{figure}
Animations play an important role in our daily life and have been widely used in entertainment, social, and educational applications. Recently, \textit{\textbf{anime}}, aka Japan-animation, has been popular in social media platforms. Many people would like to transfer their profile photos into anime images, whose styles are similar to that of the roles in
their favorite animations such as \textit{Cardcaptor Sakura} and \textit{Sailor Moon}. However, commercial image editing software fails to do this transfer, while manually producing an anime image in specific styles needs professional skills.
In this paper, we aim to automatically translate a photo-face into an anime-face based on the styles of a reference anime-face. We refer to such a task as \textit{Style-Guided Face-to-Anime Translation} ({StyleFAT}). Inspired by the advances of generative adversarial networks (GANs) \cite{GANs}, many GAN-based methods (e.g., \cite{CycleGAN,UGATIT,Pix2Pix,Pix2PixHD}) have been proposed to automatically translate images between two domains.
However, these methods \cite{CycleGAN,UNIT} focused on learning one-to-one mapping between the source and target images, which does not transfer the information of the reference image into a generated image.
Consequently, the styles of their generated anime-faces \cite{CycleGAN,UNIT} are usually dissimilar from that of the reference ones. Recently, a few reference-guided methods \cite{drit_plus,ma2018exemplar} were proposed for multi-modal translation which generates diverse results by additionally taking reference images from the target domain as input.
These methods, however, usually fail to fulfill the StyleFAT task and generate low-quality anime images.
Different from the image translation tasks of reference-guided methods, {StyleFAT} poses new challenges in two aspects. First, an anime-face usually has large eyes, a tiny nose, and a small mouth which are dissimilar from natural ones. The significant variations of shapes/appearances between anime-faces and photo-faces require
translation methods to largely overdraw the local structures (e.g., eyes and mouth) of a photo-face, different from caricature translation \cite{CariGAN} and makeup-face transfer \cite{Li2018BeautyGAN,chen2019beautyglow,CVPR2020PSGAN} which preserve the identity of a photo-face.
Since most reference-guided methods are designed to preserve the local structures/identity of a source image, these methods not only poorly transform the local shapes of facial parts into anime-like ones, but also fail to make the these local shapes style-consistent with the reference anime-face. On the other hand, simultaneously transforming local shapes and transferring anime styles is challenging and has not yet been well explored.
Second, anime-faces involve various appearances and styles (e.g. various hair textures and drawing styles). Such large intra-domain variations poses challenges in devising a generator to translate a photo-face into a specific-style anime-face, as well as in training a discriminator to capture the distributions of anime-faces.
To address the above problems, we propose a novel GAN-based model called \textit{AniGAN} for {StyleFAT}. First,
since it is difficult to collect pairs of photo-faces and anime-faces, we train AniGAN with unpaired data in an unsupervised manner.
Second, we propose a new generator architecture that preserves the global information (e.g., pose) of a source photo-face, while transforming local facial shapes into anime-like ones and transferring colors/textures based on the style of a reference anime-face. The proposed generator does not rely on face landmark detection or face parsing. Our insight is that
the local shapes (e.g., large and round eyes) can be treated as a kind of styles like color/texture. In this way, transforming a face's local shapes can be achieved via style transfer. To transform local facial shapes via style transfer, we explore \textit{where} to inject the style information into the generator. In particular, the multi-layer feature maps extracted by the decoder represent multi-level semantics (i.e., from high-level structural information to low-level textural information). Our generator therefore injects the style information into the multi-level feature maps of the decoder.
Guided the injected style information and different levels of feature maps, our generator adaptively learns to transfer color/texture styles and transform local facial shapes.
Furthermore, two normalization functions are proposed for the generator to further improve local shape transformation and color/texture style transfer.
In addition to the generator, we propose a double-branch discriminator, that explicitly considers large appearance variations between photo-faces and anime-faces as well as variations among anime images.
The double-branch discriminator not only learns domain-specific distributions by two branches of convolutional layers, but also learns the distributions of a common space across domains by shared shallow layers, so as to mitigate artifacts in generated faces. Meanwhile, a domain-aware feature matching loss is proposed to reduce artifacts of generated images by exploiting domain information in the branches.
Our major contributions are summarized as follows:
\begin{enumerate}
\item To the best of our knowledge, this is the first study on the style-guided face-to-anime translation task.
\item
We propose a new generator to simultaneously transfer color/texture styles and transform the local facial shapes of a source photo-face into their anime-like counterparts based on the style of a reference anime-face, while preserving the global structure of the source photo-face.
\item
We devise a novel discriminator to help synthesize high-quality anime-faces via learning domain-specific distributions, while effectively avoiding noticeable distortions in generated faces via learning cross-domain shared distributions between anime-faces and photo-faces.
\item
Our new normalization functions improve the visual qualities of generated anime-faces in terms of transforming local shapes and transferring anime styles.
\end{enumerate}
\section{Related Work}
\label{sec:related}
\textbf{Generative Adversarial Networks.}
Generative Adversarial Networks (GANs) \cite{GANs} have achieved impressive performance for various image generation and translation tasks \cite{cGANs,PGGAN,StyleGAN,SAGAN,BigGAN,yang2019show,song2018contextual,chen2019quality,li2020staged,liu2019swapgan}.
The key to the success of GANs is the adversarial training between the generator and discriminator. In the training stage, networks are trained with an adversarial loss, which constrains the distribution of the generated images to be similar to that of the real images in the training data. To better control the generation process, variants of GANs, such as conditional GANs (cGANs) \cite{cGANs} and multi-stage GANs \cite{PGGAN,StyleGAN}, have been proposed. In our work, we also utilize an adversarial loss to constrain the image generation. Our model uses GANs to learn the transformation from a source domain to a significantly different target domain, given unpaired training data.
\textbf{Image-to-Image Translation.}
With the popularization of GANs, GAN-based image-to-image translation techniques have been widely explored in recent years \cite{Pix2Pix,Pix2PixHD,CycleGAN,UNIT,MUNIT,UGATIT,FUNIT,StarGAN,yang2020one}. For example, trained with paired data, Pix2Pix \cite{Pix2Pix} uses a cGAN framework with an $L1$ loss to learn a mapping function from input to output images. Wang et al. proposed an improved version of Pix2Pix \cite{Pix2PixHD} with a feature matching loss for high-resolution image-to-image translation.
For unpaired data, recent efforts \cite{CycleGAN,UNIT,MUNIT,UGATIT,FUNIT} have greatly improved the quality of generated images. CycleGAN \cite{CycleGAN} proposes a cycle-consistency loss to get rid of the dependency on paired data. UNIT \cite{UNIT} maps source-domain and target-domain images into a shared-latent space to learn the joint distribution between the source and target domains in an unsupervised manner. MUNIT \cite{MUNIT} extends UNIT to multi-modal contexts by incorporating AdaIN \cite{AdaIN} into a content and style decomposition structure. To focus on the most discriminative semantic parts of an image during translation, several works \cite{ContrastGAN,UGATIT} involve attention mechanisms. ContrastGAN \cite{ContrastGAN} uses the object mask annotations from each dataset as extra input data. UGATIT \cite{UGATIT} applies a new attention module and proposes an adaptive layer-instance normalization (AdaLIN) to flexibly control the amount of change in shapes and textures. However, the style controllability of the above methods is limited due to the fact that the instance-level style features are not explicitly encoded. To overcome this, FUNIT \cite{FUNIT} utilizes a few-shot image translation architecture for controlling the categories of output images, but its stability is still limited.
\textbf{Neural Style Transfer.}
StyleFAT can also be regarded as a kind of the neural style transfer (NST) \cite{NST1,NST2,NST3}. In the field of NST, many approaches have been developed to generate paintings with different styles. For example, CartoonGAN \cite{CartoonGAN} devises several losses suitable for general photo cartoonization. ChipGAN \cite{ChipGAN} enforces voids, brush strokes, and ink wash tone constraints to a GAN loss for Chinese ink wash painting style transfer. APDrawingGAN \cite{APDrawingGAN} utilizes a hierarchical GAN to produce high-quality artistic portrait drawings.
CariGANs \cite{CariGAN} and WarpGAN \cite{WarpGan} design special modules for geometric transformation to generate caricatures.
Yaniv et al. \cite{yaniv2019face} proposed a method for geometry-aware style transfer for portraits utilizing facial landmarks.
However, the above methods either are designed for a specific art field which is completely different from animation, or rely on additional annotations (such as facial landmarks).
\section{Proposed Approach}\label{approach}
\textbf{Problem formulation.}
Let $X$ and $Y$ denote the source and target domains, respectively. Given a source image $x \in X$ and a reference image $y \in Y$, our proposed AniGAN learns multimodal mapping functions $G: (x,y) \mapsto \tilde{x}$ that transfer $x$ into domain $Y$.
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\linewidth, keepaspectratio]{figures/generator_v3_1.pdf}
\caption{Architecture of the proposed generator. It consists of a content encoder and a style encoder to translate the source image into the output image reflecting the style of the reference image. The grey trapezoids indicate typical convolution blocks. More detailed notations are described in the contexts of Sec. \ref{approach}. }
\label{framework}
\vspace*{-14pt}
\end{figure*}
To generate high-quality anime-faces for the StyleFAT task, the goal is to generate an anime-face $\tilde{x}$ that well preserves the global information (e.g., face pose) from ${x}$ as well as reflects the styles (e.g., colors and textures) of reference anime-face $y$, while transforming the shapes of facial parts such as eyes and hair into anime-like ones.
To achieve the above goals, a question is posed, \textit{how to
simultaneously transform local shape while transferring color/texture information}?
Existing methods focus on transferring styles while preserving both local and global structures/shapes from the source image, which, however, cannot well address the above problem.
Differently, we here explore \textit{where} to inject style information into the generator, and a novel generator architecture is thereby proposed, as shown in Fig. \ref{framework}.
Different from existing methods which inject style information in the bottleneck of the generator, we introduce two new modules in the decoder of the generator for learning to interpret and translate style information.
Furthermore, we also propose two normalization functions to control the style of generated images while transforming local shapes, inspired by recent work \cite{AdaIN,UGATIT}.
In addition, anime-faces contain significant intra-variations, which poses large challenge for generating high-quality images without artifacts. To further improve the stability of the generated results, a novel double-branch discriminator is devised to better utilize the distribution information of different domains.
Figs. \ref{framework} and \ref{discriminator} illustrate the architectures of our proposed generator and discriminator, respectively. The generator takes a source image and a reference image as inputs and then learns to synthesize an output image. The double-branch discriminator consists of two branches, where one branch discriminates real/fake images for domain $X$, and the other for $Y$.
\subsection{Generator}
The generator of AniGAN consists of a content encoder $E_{c}$, a style encoder $E_{s}$ and a decoder $F$, as shown in Fig. \ref{framework}.
\textbf{Encoder.}
The encoder includes a content encoder $E_c$ and a style encoder $E_s$. Given a source photo-face $\mathbf{x}$ and a reference anime-face $\mathbf{y}$, the content encoder $E_{c}$ is used to encode the content of the source image $\mathbf{x}$, and the style encoder is employed to extract the style information from the reference $\mathbf{y}$.
The functions can be formulated as follows:
\begin{equation}
\alpha=E_{c}(\mathbf{x}),
\end{equation}
\begin{equation}
(\gamma_{s}, \beta_{s}) =E_{s}(\mathbf{y})
\end{equation}
where $\alpha$ is the content code encoded by the content encoder $E_{c}$,
$\gamma_{s}$ and $\beta_{s}$ are the style parameters extracted from the reference anime-face $\mathbf{y}$.
\textbf{Decoder.}
The decoder $F$ constructs an image from a content code and style codes. However, different from typical image translations that transfer styles while preserving both local and global structures of source image, our decoder aims to transform the local shapes of facial parts and preserve the global structure of the source photo-face during style transfer.
\begin{figure}[t]
\centering
\includegraphics[width=0.71\linewidth, keepaspectratio]{figures/discriminator_v2.pdf}
\caption{ Architecture of the proposed double-branch discriminator, where $D_U$ denotes the shared layers between the two branches for domain $X$ and $Y$. The discriminator distinguishes real images from fake ones in the two individual branches. }
\label{discriminator}
\vspace{-14pt}
\end{figure}
Recent image translation methods \cite{FUNIT,AdaIN, MUNIT,UGATIT,ma2018exemplar} transfer the style of the target domain by equipping residual blocks (resblocks) in the bottleneck of the generator with style codes. However, we observe that such decoder architectures cannot well handle the StyleFAT task. Specifically, they are either insufficient to elaborately transfer an anime style, or introduce visually annoying artifacts in generated faces.
To decode high-quality anime-faces for StyleFAT, we propose an adaptive stack convolutional block (denoted as \textit{ASC} block), Fine-grained Style Transfer (FST) block and two normalization functions for the decoder (see Fig. \ref{framework})
\textbf{{ASC block}} Existing methods such as MUNIT~\cite{MUNIT}, FUNIT~\cite{FUNIT}, and EGSC-IT\cite{ma2018exemplar} transfer styles to generate images via injecting the style information of the target domain into the resblocks in the bottleneck of their decoder. However, we observe that resblocks may ignore some anime-style information, which degrades the translation performance of decoder on StyleFAT task. More specifically, although multiple reference images with different anime styles are given, the decoder with resblocks synthesizes similar styles for specific regions, especially on eyes. For example, the decoder with resblocks improperly renders right eyes with the same color in the generated images (see Fig. \ref{fig:ASC}).
We argue that the decoder would "skip" some injected style information due to the residual operation.
To address this issue, we propose an ASC block for the decoder. ASC stacks convolutional layers, activation, and our proposed normalization layers, instead of using resblocks, as shown in Fig. \ref{framework}.
\textbf{{FST block}}
One aim of our decoder is to transform local facial features into anime-like ones,
different from existing methods which preserve local structures from a source photo-face. We here explore how to transform local shapes in the decoder when transferring styles.
One possible solution is to first employ face parsing or facial landmark detection to detect facial features and then transform local facial features via warping like \cite{WarpGan}.
However, since the local structure of anime-faces and photo-faces are significantly dissimilar to each other, warping often leads to artifacts in the generated anime-faces. For example, it is difficult to well warp the mouth of a photo-face into a tiny one of a reference anime-face.
Instead of employing face parsing or facial landmark detection, our insight is that local structures can be treated as a kind of styles like color/texture and can be altered through style transfer.
Therefore, we propose a FST module to simultaneously transform local shapes of the source image and transfer color/texture information from the reference image.
In particular, as revealed in the literature, deep- and shallow-layer feature maps with low and high resolutions in the decoder indicate different levels of semantic information from high-level structures to low-level colors/textures. Motivated by the fact, we argue that the FST block can adaptively learn to transform local shapes and decode color/texture information by injecting the style information into feature maps with different resolutions. In other words, since the feature map contains high-level structural information, FST can learn to transform local shapes into anime-like ones with the specified style information. Thus, as shown in Fig. \ref{framework}, FST consists of a stack of upsampling, convolutional, and normalization layers. Furthermore, style codes are injected after the upsampling layer in our FST block, different from the prior works \cite{MUNIT,UGATIT,FUNIT} that only injects style codes into the bottleneck of the generator. Consequently, FST block enables the decoder to adaptively synthesize color/texture and transform local shapes, according to both the style codes and different levels of feature maps.
\textbf{{Normalization}} Recent image translation methods \cite{AdaIN,MUNIT,ma2018exemplar} normalize the feature statistics of an image via Adaptive
Instance Normalization (AdaIN) to adjust the color/texture styles of the image.
However, we aim to devise normalization functions which not only transfer color/texture styles from the reference, but also transform local shapes of the source image based on the reference.
Recently, it was shown in \cite{UGATIT} that layer normalization (LN) \cite{LN} can transform the structure/shape of an image.
Besides, AdaLIN was proposed in \cite{UGATIT} to control the degree of changes in textures and shapes by adaptively combining AdaIN and LN.
However, AdaLIN is insufficient to simultaneously transfer the color/texture information of a local region and its shape information from a reference image to a generated image.
That is, since AdaLIN combines IN and LN in a per-channel manner, AdaLIN ignores the correlations among channels. For example, the shape styles of eyes and their color/texture styles may respectively dominate in different channels. In such case, the features learned by AdaLIN often ignore shape styles or color/texture styles. In other words, the combination space of AdaLIN tends to be smaller than that of all-channel combinations of IN and LN.
To address the above issue, we propose two novel normalization functions called point-wise layer instance normalization (PoLIN) and adaptive point-wise layer instance normalization (AdaPoLIN) for the generator.
our PoLIN and AdaPoLIN learn to combine all channels of IN and LN, different from AdaLIN \cite{UGATIT}.
To achieve all-channel combination of instance normalization (IN)~\cite{IN} and LN, PoLIN learns to combine IN and LN via a $1 \times 1$ convolutional layer as defined below:
\begin{equation}
\textsc{PoLIN}(z)=Conv\left (\left[\frac{\mathbf{z}-\mu_{I}(\mathbf{z})}{\sigma_{I}(\mathbf{z})},\frac{\mathbf{z}-\mu_{L}(\mathbf{z})}{\sigma_{L}(\mathbf{z})}\right]\right),
\end{equation}
where $Conv(\cdot)$ denotes the $1 \times 1$ convolution operation, $[\cdot, \cdot]$ denotes the channel-wise concatenation, $\mathbf{z}$ is the the feature map of a network layer, $\mu_{I}$, $\mu_{L}$ and $\sigma_{I}$, $\sigma_{L}$ denote the channel-wise and layer-wise means and standard deviations, respectively.
AdaPoLIN adaptively combines IN and LN, while employing the style codes from the reference anime-faces to retain style information:
\begin{equation}
\begin{split}
& \text{AdaPoLIN}(z,\gamma_{s},\beta_{s})\\ & = \gamma_{s} \cdot Conv\left(\left[\frac{\mathbf{z}-\mu_{I}(\mathbf{z})}{\sigma_{I}(\mathbf{z})},\frac{\mathbf{z}-\mu_{L}(\mathbf{z})}{\sigma_{L}(\mathbf{z})}\right]\right) +\beta_{s},
\end{split}
\end{equation}
where $\gamma_{s}$ and $\beta_{s}$ are style codes,
and the bias in $Conv(\cdot)$ is fixed to 0.
Thanks to their all-channel combination of IN and LN, the proposed PoLIN and AdaPoLIN lead to a larger combination space than AdaLIN, thereby making them beneficial to handle color/texture style transfer and local shape transformation for StyleFAT.
\subsection{Discriminator}
It is challenging to design a discriminator which effectively distinguishes real anime-faces from fake ones for StyleFAT. In particular, both the appearances and shapes vary largely among anime-faces, leading to significant intra-variations in the distribution of anime-faces.
Thus, it is difficult for a typical discriminator (e.g. \cite{CycleGAN}) to well learn the distribution of anime-faces.
As a result, the generated anime-faces may contain severely distorted facial parts and noticeable artifacts
To address the above issues, we propose a double-branch discriminator. In particular, we assume that anime-faces and photo-faces partially share common distributions and such cross-domain shared distributions constitute meaningful face information, since these two domains are both about human faces. In other words, by learning and utilizing the cross-domain shared distributions, the discriminator can help reduce distortions and artifacts in translated anime-faces.
Therefore, as shown in Fig. \ref{discriminator}, the proposed double-branch discriminator consists of shared shallow layers followed by two domain-specific output branches: one branch for distinguishing real/fake anime-faces and the other for distinguishing real/fake photo-faces.
With the Siamese-like shallow layers shared by the photo-face and anime-face branches, the additional photo-face branch aims to assist the anime-face branch to learn domain-shared distributions.
As a result, the anime-face branch learns to effectively discriminate those generated anime-faces with distorted facial parts or noticeable face artifacts.
On the other hand, each branch contains additional domain-specific deep layers with an extended receptive field to individually learn the distributions of anime-faces and photo-faces\footnote{Two separate discriminators without shared shallow layers can also individually learn the distributions of anime-faces and photo-faces. However, we observe that such a design not only consumes more computational cost but also performs worse than our double-branch discriminator.}
We formulate the two-branch discriminator in terms of domain $X$ and domain $Y$ for generality. Let $D_{X}$ and $D_{Y}$ denote the discriminator branches corresponding to domain $X$ and $Y$, respectively, and $D_{U}$ denote the shallow layers shared by $D_{X}$ and $D_{Y}$.
An input image $h$ is discriminated either by $D_X$ or by $D_Y$ according to the domain that $h$ belongs to.
The discriminator function is formulated as follows:
\begin{equation}
D(h)=
\begin{cases}
D_{X}(D_U(h)) & \text{if} \ h \in X, \\
D_{Y}(D_U(h)) & \text{if} \ h \in Y.
\end{cases}
\end{equation}
Our discriminator helps significantly improve the quality of generated images and the training stability of the generator, since it not only individually learns domain-specific distributions using separable branches, but also learns domain-shared distributions across domains using shared shallow layers. In addition, our discriminator is scalable and can be easily expended to multiple branches for tasks across multiple domains.
\subsection{Loss Functions}
StyleFAT is different from typical image translation tasks which preserve the identity or the whole structure of the input photo-face. If we directly employ loss functions in existing methods \cite{ma2018exemplar,CariGAN} that preserves both local and global structures or an identity in the source image, the quality of a generated anime-face would be negatively affected.
Instead,
besides adversarial loss like in \cite{Li2018BeautyGAN,ma2018exemplar}, the objective of our model also additionally involves reconstruction loss, feature matching loss, and domain-aware feature matching loss.
\textbf{Adversarial loss.}
Given a photo-face $x \in X$ and a reference anime-face $y \in Y$, the generator aims to synthesize from $x$ an output image $G(x,y)$ with a style transferred from $y$. To this end, we adopt an adversarial loss similar to \cite{cGANs,FUNIT} as follows:
\begin{equation}
\begin{array}{ccc}
& L_{\text{adv}} =\mathbb{E}_{x}[\log D_{X}(x)]+\mathbb{E}_{x,y}[\log(1-D_{X}(G(y,x)))] \\&+
\mathbb{E}_{y}[\log D_{Y}(y)] +\mathbb{E}_{y,x}[\log(1-D_{Y}(G(x,y)))].
\end{array}
\end{equation}
\textbf{Feature matching loss.}
To encourage the model to produce natural statistics at various scales, the feature matching loss \cite{Pix2PixHD,SPADE} is utilized as a supervision for training the generator.
Formally, let $D_{U}^{k}(h)$ denote the feature map extracted from the $k$-th down-sampled version of the shared layers $D_{U}$ of $D_X$ or $D_Y$ for input $h$,
and $\bar D_{U}^k(h)$ denote the global average pooling result of $D_{U}^{k}(h)$, the feature matching loss $L_{fm}$ is formulated below:
\begin{equation}
L_{\text{fm}} = \mathbb{E}_{h}[\sum\limits_{k \in K_1} ||\bar D_{U}^k(h) - \bar D_{U}^k(G(h,h))||_{1}],
\end{equation}
where $K_1$ denotes the set of selected layers in $D_{U}$ used for feature extraction. In our work we set $K_1=\{1,2\}$ according to the architecture of our discriminator.
\textbf{Domain-aware feature matching loss. }
We further utilize the domain-specific information to optimize the generator. In particular, we extract
features by the domain-specific discriminator $D_X$ or $D_Y$, respectively.
Similar to $D_{U}^k(\cdot)$, let $D_{X}^k(\cdot)$ denote the $k$-th down-sampling feature map extracted from the branch $D_X$ in domain $X$, and $\bar D_{X}^k(\cdot)$ denote the average pooling result of $D_{X}^k(\cdot)$ (similar notations are used for domain $Y$).
Then we define a domain-aware feature matching loss $L_{\text{dfm}}$ as follows:
\begin{equation}
\begin{aligned}
& L_{\text{dfm}}=\\
&\begin{cases}
\mathbb{E}_{h}[\sum\limits_{k \in K_2} ||\bar D_{X}^k(D_U(h)) - \bar D_{X}^k(D_U(G(h,h)))||_{1}], \text{ if} \ h \in X \\
\mathbb{E}_{h}[\sum\limits_{k \in K_2} ||\bar D_{Y}^k(D_U(h)) - \bar D_{Y}^k(D_U(G(h,h)))||_{1}],\text{ if} \ h \in Y
\end{cases}
\end{aligned}
\end{equation}
where $K_2$ represents the set of selected layers in $D_{X}$ and $D_{Y}$ used for feature extraction. In our work we set $K_2=\{3\}$ according to the architecture of our discriminator.
With $L_{\text{dfm}}$, the artifacts of generated images can be largely mitigated thanks to the additional domain-specific features.
\begin{figure*}[!htbp]
\centering
\includegraphics[bb=0 0 120 128,width=0.1\textwidth]{figures/figure_dataset/imgHQ11938.png}
\includegraphics[bb=0 0 120 128,width=0.1\textwidth]{figures/figure_dataset/imgHQ11820.png}
\includegraphics[bb=0 0 120 128,width=0.1\textwidth]{figures/figure_dataset/imgHQ11379.png}
\includegraphics[bb=0 0 120 128,width=0.1\textwidth]{figures/figure_dataset/imgHQ12103.png}
\includegraphics[bb=0 0 120 128,width=0.1\textwidth]{figures/figure_dataset/imgHQ11188.png}
\includegraphics[bb=0 0 120 128,width=0.1\textwidth]{figures/figure_dataset/imgHQ11385.png}
\includegraphics[bb=0 0 120 128,width=0.10\textwidth]{figures/figure_dataset/imgHQ11320.png}
\includegraphics[bb=0 0 120 128,width=0.10\textwidth]{figures/figure_dataset/43015.png}
\includegraphics[bb=0 0 120 128,width=0.10\textwidth]{figures/figure_dataset/2585108.png}
\includegraphics[bb=0 0 120 128,width=0.10\textwidth]{figures/figure_dataset/2908067.png}
\includegraphics[bb=0 0 120 128,width=0.10\textwidth]{figures/figure_dataset/367027.png}
\includegraphics[bb=0 0 120 128,width=0.10\textwidth]{figures/figure_dataset/914019.png}
\includegraphics[bb=0 0 120 128,width=0.10\textwidth]{figures/figure_dataset/2848117.png}
\includegraphics[bb=0 0 120 128,width=0.10\textwidth]{figures/figure_dataset/2791115.png}
\caption{ Example photo-faces and anime-faces of our face2anime dataset. From top to bottom: photo-faces and anime-faces. }
\label{dataset}
\vspace*{-17pt}
\end{figure*}
\textbf{Reconstruction loss.}
We aim to preserve the global semantic structure of source photo-face $\mathbf{x}$. Without a well-designed loss on the discrepancy between the generated anime-face and source photo-face, the global structure information of the source photo-face may be ignored or distorted. However, as discussed previously, we cannot directly employ an identity loss to preserve the identity like \cite{Li2018BeautyGAN} or a perceptual loss to preserve the structure of the image like \cite{ma2018exemplar}.
Different from existing methods \cite{Li2018BeautyGAN,ma2018exemplar}, we impose a reconstruction loss to preserve the global information of photo-face. Specifically, given source photo-face $\mathbf{x}$, we also use $\mathbf{x}$ as the reference to generate a face $G(\mathbf{x}, \mathbf{x})$. If the generated face $G(\mathbf{x}, \mathbf{x})$ well reconstructs the source photo-face, we argue that the generator preserves the global structure information from the photo-face. Thus, we define the reconstruction loss as the dissimilarity between $\mathbf{x}$ and $G( \mathbf{x}, \mathbf{x})$ by
\begin{equation}
L_{\text{rec}}=||G( \mathbf{x}, \mathbf{x}) -\mathbf{x}||_{1}.
\end{equation}
This loss encourages the generator to effectively learn global structure information from photo-face, such that the crucial information of x
is preserved in the generated image.
\textbf{Full Objective.} Consequently, we combine all the above loss functions as our full objective as follows:
\begin{equation}
L_{G}=L_{\text{adv}}+\lambda_{\text{rec}} \cdot L_{\text{rec}} + \lambda_{\text{fm}} \cdot (L_{\text{fm}}+L_{\text{dfm}}).
\end{equation}
\begin{equation}
L_{D}=-L_{\text{adv}},
\end{equation}
where $\lambda_{\text{rec}}, \lambda_{\text{fm}}$ are hyper-parameters to balance the losses.
\begin{comment}
\begin{figure*}[t]
\centering
\includegraphics[width=0.67\linewidth, keepaspectratio]{figures/figure3/figure3.png}
\caption{Example natural and anime face images of the face2anime dataset.From top to bottom, photo-faces and anime-faces.}
\label{dataset}
\end{figure*}
\end{comment}
\section{Experimental Results}
We first conduct qualitative and quantitative experiments to evaluate the performance of our approach.
We then conduct user studies to evaluate the subjective visual qualities of generated images. Finally, we evaluate the effectiveness of proposed modules by ablation studies.
\textbf{Baselines}
We compare our method with CycleGAN \cite{CycleGAN}, UGATIT\footnote{We used the full version implementation of UGATIT} \cite{UGATIT}, and MUNIT \cite{MUNIT}, the state-of-the-arts in reference-free image translation. In addition, we compare our method with reference-guided image translation schemes including FUNIT \cite{FUNIT}, EGSC-IT \cite{ma2018exemplar}, and DRIT++ \cite{drit_plus} which are most relevant to our method.
Note that MUNIT also allows to additionally take a reference image as the input at the testing stage. We refer to this baseline as RG-MUNIT. However, instead of using a reference image as part of a training image pair, RG-MUNIT takes a random latent code from the target domain as style information for the training (see more details in its github code\footnote{https://github.com/NVlabs/MUNIT}). Due to such inadequate reference information during training, while performing StyleFAT with reference faces, RG-MUNIT tends to fail for face images with significant inter-variations between domains or data with large intra-variations within a domain. That is, RG-MUNIT performs much poorer than the original MUNIT that takes a random latent code from the target domain at the test stage. We hence only show the visual results with the original MUNIT in the qualitative comparisons, while evaluating quantitative results for both the original MUNIT and RG-MUNIT.
We train all the baselines based on the
open-source implementations provided by the original papers\footnote{As discussed in Section \ref{sec:related}, since neural-style transfer methods focus on specific art-style translation, we do not compare these methods for fair comparison.}.
\subsection{Datasets}
\textbf{Selfie2anime.} We follow the setup in UGATIT \cite{UGATIT} and use the selfie2anime dataset to evaluate our method. For the dataset, only female character images are selected and monochrome anime images are removed manually.
Both selfie and anime images are separated into a training set with 3,400 images and a test set with 100 images.
\textbf{Face2anime.} We build an additional dataset called face2anime, which is larger and contains more diverse anime styles (e.g., face poses, drawing styles, colors, hairstyles, eye shapes, strokes, facial contours) than selfie2anime, as illustrated in Fig. \ref{dataset}. The face2anime dataset contains 17,796 images in total, where the number of both anime-faces and natural photo-faces is 8,898. The anime-faces are collected from the Danbooru2019 \cite{Danbooru} dataset, which contains many anime characters with various anime styles. We employ a pretrained cartoon face detector \cite{AnimeFace} to select images containing anime-faces\footnote{Images containing male characters are discarded, since Danbooru2019 only contains a small number of male characters.}. For natural-faces, we randomly select 8,898 female faces from the CelebA-HQ \cite{CelebA,PGGAN} dataset. All images are aligned with facial landmarks and are cropped to size $128 \times 128$. We separate images from each domain into a training set with 8,000 images and a test set with 898 images.
\begin{figure*}[!htbp]
\centering
\tabcolsep=0.1c
\begin{tabular}{cc:ccccccccc}
{Reference} & {Source} & CycleGAN& UGATIT&MUNIT&FUNIT&DRIT++&EGSC-IT&Ours\\
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c/imgHQ12411_2589084_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11302_2135050_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11257_1282017_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11418_2867071_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ12900_12053_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11248_847113_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11562_1576076_reference.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c/imgHQ12411_2589084_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11302_2135050_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11257_1282017_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11418_2867071_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ12900_12053_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11248_847113_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11562_1576076_source.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c/imgHQ12411_2589084_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11302_2135050_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11257_1282017_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11418_2867071_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ12900_12053_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11248_847113_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11562_1576076_CycleGAN.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c/imgHQ12411_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11302_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11257_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11418_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ12900_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11248_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11562_UGATIT.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c/imgHQ12411_2589084_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11302_2135050_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11257_1282017_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11418_2867071_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ12900_12053_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11248_847113_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11562_1576076_MUNIT.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c/imgHQ12411_2589084_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11302_2135050_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11257_1282017_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11418_2867071_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ12900_12053_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11248_847113_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11562_1576076_FUNIT.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c/imgHQ12411_2589084_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11302_2135050_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11257_1282017_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11418_2867071_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ12900_12053_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11248_847113_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11562_1576076_DRIT.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 256 278,width=1\textwidth]{figures/fig_c/imgHQ12411_2589084_EG.png}
\includegraphics[bb=0 0 256 256,width=1\textwidth]{figures/fig_c/imgHQ11302_2135050_EG.png}
\includegraphics[bb=0 0 256 256,width=1\textwidth]{figures/fig_c/imgHQ11257_1282017_EG.png}
\includegraphics[bb=0 0 256 256,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_EG.png}
\includegraphics[bb=0 0 256 256,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_EG.png}
\includegraphics[bb=0 0 256 256,width=1\textwidth]{figures/fig_c/imgHQ11418_2867071_EG.png}
\includegraphics[bb=0 0 256 256,width=1\textwidth]{figures/fig_c/imgHQ12900_12053_EG.png}
\includegraphics[bb=0 0 256 256,width=1\textwidth]{figures/fig_c/imgHQ11248_847113_EG.png}
\includegraphics[bb=0 0 256 256,width=1\textwidth]{figures/fig_c/imgHQ11562_1576076_EG.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c/imgHQ12411_2589084_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11302_2135050_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11257_1282017_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_aniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11418_2867071_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ12900_12053_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11248_847113_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11562_1576076_AniGAN.png}
\end{minipage}
\end{tabular}
\caption{ Comparison of various image translation methods on the face2anime dataset. From left to right: source photo-face, reference anime-face, the results by CycleGAN \cite{CycleGAN}, MUNIT \cite{MUNIT}, UGATIT \cite{UGATIT}, FUNIT \cite{FUNIT}, DRIT++ \cite{drit_plus}, EGSC-IT \cite{ma2018exemplar} and our AniGAN. }
\label{qualitive_comparison}
\vspace*{-17pt}
\end{figure*}
\subsection{Implementation Details}
We train and evaluate our approach using the face2anime and selfie2anime datasets. We use the network architecture mentioned in Section \ref{approach} as our backbone. We set $\lambda_{fm}=1$ for all experiments, $\lambda_{rec}=1.2$ for the face2anime dataset, and $\lambda_{rec}=2$ for the selfie2anime dataset.
For fast training, the batch size is set to 4 and the model is trained for 100K iterations. The training time is less than 14h on a single Tesla V100 GPU with our implementation in PyTorch \cite{Pytorch}. We use RMSProp optimizer with a learning rate of $0.0001$. To stabilize the training, we use the hinge version of GAN \cite{GeometricGAN,SNGAN,SAGAN,BigGAN} as our GAN loss and also adopt real gradient penalty regularization \cite{WGANGP,RealGP}. The final generator is a historical average version \cite{PGGAN} of the intermediate generators, where the update weight is $0.001$.
\subsection{Qualitative comparison}
Given a source photo-face and a reference anime-face, a good translation result for StyleFAT task should share similar/consistent anime-styles (e.g., color and texture) with the reference without introducing noticeable artifacts, while facial features are anime-like and the global information (e.g., the face pose) from the source is preserved.
\begin{figure*}[!htbp]
\centering
\tabcolsep=0.1c
\begin{tabular}{cc:ccccccccc}
{Reference} & {Source} & CycleGAN& UGATIT&MUNIT&FUNIT&DRIT++&EGSC-IT&Ours\\
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c_self/female_7349_0005_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_7613_0052_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_30499_0048_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_18022_0081_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_24446_0011_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_14587_0076_reference.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c_self/female_7349_0005_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_7613_0052_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_30499_0048_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_18022_0081_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_24446_0011_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_14587_0076_source.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c_self/female_7349_0005_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_7613_0052_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_30499_0048_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_18022_0081_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_24446_0011_CycleGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_14587_0076_CycleGAN.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c_self/female_7349_0005_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_7613_0052_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_30499_0048_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_18022_0081_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_24446_0011_UGATIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_14587_0076_UGATIT.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c_self/female_7349_0005_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_7613_0052_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_30499_0048_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_18022_0081_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_24446_0011_MUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_14587_0076_MUNIT.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c_self/female_7349_0005_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_7613_0052_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_30499_0048_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_18022_0081_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_24446_0011_FUNIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_14587_0076_FUNIT.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c_self/female_7349_0005_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_7613_0052_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_30499_0048_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_18022_0081_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_24446_0011_DRIT.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_14587_0076_DRIT.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c_self/female_7349_0005_EG.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_7613_0052_EG.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_30499_0048_EG.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_18022_0081_EG.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_24446_0011_EG.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_14587_0076_EG.png}
\end{minipage}
&
\begin{minipage}[t]{0.092\textwidth}
\centering
\includegraphics[bb=0 0 128 138,width=1\textwidth]{figures/fig_c_self/female_7349_0005_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_7613_0052_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_30499_0048_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_18022_0081_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_24446_0011_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c_self/female_14587_0076_AniGAN.png}
\end{minipage}
\end{tabular}
\caption{ Comparison with image-to-image translation methods on the selfie2anime dataset. From left to right: source photo-face, reference anime-face, the results by CycleGAN \cite{CycleGAN}, MUNIT \cite{MUNIT}, UGATIT \cite{UGATIT}, FUNIT \cite{FUNIT}, DRIT++ \cite{drit_plus}, EGSC-IT \cite{ma2018exemplar} and our AniGAN. }
\label{fig:qualitive_comparison_self2anime}
\vspace*{-17pt}
\end{figure*}
Fig. \ref{qualitive_comparison} illustrates qualitative comparison results on \textit{face2anime} dataset, where the photo-faces involve various identities, expressions, illuminations, and poses, whereas the reference anime-faces present various anime-styles. Compared with \textit{cat2dog} and \textit{horse2zebra} dataset, face2anime dataset contains larger variations of shape/appearance among anime-faces, whose data distribution is much more complex and is challenging for image translation methods.
The results show that CycleGAN introduces visible artifacts in their generated anime-faces (see
the forehead in the fourth row). MUINT also leads to visible artifacts in some generated anime-faces, as shown in the third and fourth row in Fig.\ref{qualitive_comparison}.
UGATIT better performs than CycleGAN and MUNIT.
However, the anime styles of generated anime-faces by CycleGAN, UGATIT and MUNIT are dissimilar to that of the references.
FUNIT is designed for few-shot reference-guided translation, and hence is not suitable for the StyleFAT task.
Consequently,
the style of generated anime-faces by FUNIT is much less consistent with the references compared with our method.
Although EGSC-IT usually well preserves the poses of photo-faces, it also attempts to preserve the local structures of photo-faces, which often conflicts with the transfer of anime styles, since the local shapes of facial parts like eyes and mouth in an anime-face are dissimilar to the counterparts of the corresponding photo-face. Consequently, EGSC-IT often leads to severe artifacts in the local structures of generated anime-faces (see eyes and hair in the first to third rows).
DRIT++ also introduces artifacts into generated faces, when transferring the styles of the reference anime-faces. For example, DRIT++ generates two mouths in the generated face in the third row and distort eyes in the fifth row.
Outperforming the above state-of-art methods, our method generates the highest-quality anime-faces.
First, compared with reference-guided methods FUNIT, EGSC-IT, and DRIT++, the styles of generated anime-faces with our method are the most consistent with that of reference faces, thanks to our well-designed generator.
Moreover, our method well preserves the poses of photo-faces, although our method does not use perceptual loss like EGSC-IT. Our method also well converts local structures like eyes and mouth into anime-like ones without introducing clearly visible artifacts.
Fig. \ref{fig:qualitive_comparison_self2anime} compares the results of various methods on the selfie2anime dataset. The results show that FUNIT introduces artifacts into some generated faces, as shown in the fourth and sixth rows in Fig.~\ref{fig:qualitive_comparison_self2anime}. Besides, EGSC-IT tends to generate anime-faces with similar styles, although the reference anime-faces are of different styles (see the reference and generated images in the fourth to sixth rows in Fig. \ref{fig:qualitive_comparison_self2anime}). Similarly, DRIT++ tends to synthesize eyes with similar styles in the generated faces. For example, the synthesized eyes by DRIT++ are orange ellipse in the first, second, fourth and fifth rows in Fig. \ref{fig:qualitive_comparison_self2anime}. In contrast, our method generates anime-faces reflecting the various styles of the reference images. In other words, our method achieves the most consistent styles with those of reference anime-faces over the other methods. In addition, our method generates high-quality faces which preserve the poses of source photo-faces, despite a photo-face is
partially occluded by other objects (e.g., the hand in the fifth rows of Fig. \ref{fig:qualitive_comparison_self2anime}).
{\subsection{Quantitative Comparisons}}
In addition to qualitative evaluation, we quantitatively evaluate the performance of our method in two aspects. One is the visual quality of generated images, and the other is the translation diversity.
\textbf{Visual quality.} We evaluate the the quality of our results with Frechet Inception Distance (FID) metric \cite{FID} which has been popularly used to evaluate the quality of synthetic images in image translation works e.g., \cite{starganv2, APDrawingGAN, FUNIT}.
The FID score evaluates the distribution discrepancy between the real faces and synthetic anime-faces. A lower FID score indicates that the distribution of generated images is more similar to that of real anime-faces. That is, those generated images with lower FID scores are more plausible as real anime-faces. Following the steps in \cite{FID}, we compute a feature vector by a pretrained network \cite{InceptionV3} for each real/generated anime-face, and then calculate FID scores for individual compared methods, as shown in Table \ref{fid1}. The FID scores in Table \ref{fid1} demonstrate that our our AniGAN achieves the best scores on both the face2anime and selfie2anime datasets, meaning that the anime-faces generated by our approach have the closest distribution with real anime-faces, thereby making they look similar visually.
\textbf{Translation diversity.}
For the same photo-face, we evaluate whether our method can generate anime-faces with diverse styles, given multiple reference anime-faces with different styles.
We adopt the learned perceptual image patch
similarity (LPIPS) metric, a widely adopted metric for assessing translation methods on multimodal mapping \cite{starganv2, drit_plus} in the perceptual domain, for evaluating the translation diversity.
Following \cite{starganv2}, given each testing photo-face, we randomly sample 10 anime-faces as its reference images and then generate 10 outputs. For these 10 outputs, we evaluate the LPIPS scores between every two outputs\footnote{Following \cite{starganv2}, we uniformly scale all generated images to the size of $256\times 256$.}.
Table \ref{LPLIS} shows the average of pairwise LPIPS over all testing photo-faces. A higher LPIPS score indicates that the translation method generates images with larger diversity.
\begin{table}[t]
\caption{ Comparison of FID scores on the face2anime and selfie2anime datasets: lower is better}
\centering
\begin{tabular}{ccc}
\toprule
Method & Face2anime & Selfie2anime \\
\midrule
CycleGAN & 50.09 & 99.69 \\
UGATIT & 42.84 & 95.63\\
MUNIT & 43.75 & 98.58 \\
EG-MUNIT & 185.23 & 305.33 \\
FUNIT & 56.81 & {117.28} \\
DRIT & 70.59 & 104.49 \\
EGSC-IT & 67.57 & 104.70 \\
AniGAN (Ours) & \textbf{38.45} & \textbf{86.04}\\
\bottomrule
\end{tabular}
\vspace*{-11pt}
\label{fid1}
\end{table}
\begin{table}[t]
\caption{ Comparison of average LPIPS scores on the face2anime and selfie2anime dataset: higher is better}
\centering
\begin{tabular}{ccc}
\toprule
& Face2anime & Selfie2anime \\
\midrule
DRIT++ & 0.184 & 0.201 \\
EGSC-IT & 0.302 & 0.225 \\
AniGAN (Ours) & \textbf{0.414} & \textbf{0.372}\\
\bottomrule
\end{tabular}
\label{LPLIS}
\vspace*{-11pt}
\end{table}
\begin{table}[!htbp]
\centering
\caption{Preference percentages of 20 subjects for four methods in user study. Higher is better.}
\begin{tabular}{cccccc}
\toprule
Method & Face2anime & Selfie2anime\\
\midrule
FUNIT & 17.50\%&40.17\%\\
DRIT++&68.17\%&47.83\%\\
EGSC-IT &27.33\%&30.67\%\\
Ours&\textbf{87.00}\%& \textbf{81.33}\%\\
\bottomrule
\end{tabular}
\label{user_study}
\vspace*{-11pt}
\end{table}
\begin{table*}[t]
\centering
\caption{Quantitative comparison for ablation study using FID score. Lower is better}
\begin{tabular}{p{3.7cm} p{0.4cm}<{\centering} p{0.35cm}<{\centering} p{0.36cm}<{\centering} p{0.67cm}<{\centering} p{1.3cm}<{\centering} p{0.2cm}<{\centering} p{0.3cm}<{\centering} p{0.67cm}<{\centering} p{0.9cm}<{\centering} m{1.45cm}<{\centering} m{1.78cm}<{\centering}}
\toprule
Method & ASC & FST & DB & PoLIN & AdaPoLIN & IN & LIN & AdaIN &AdaLIN & Face2anime & Selfie2anime \\
\midrule
w/o ASC & & $\checkmark$ & $\checkmark$ & $\checkmark$ &$\checkmark$ & & & & & 40.52 & 96.20\\
w/o FST & $\checkmark$ & & $\checkmark$ & & $\checkmark$ & & & & & 44.13 & 99.91\\
w/o DB & $\checkmark$ & $\checkmark$ & & $\checkmark$ & $\checkmark$ & & & & & 40.56 & 92.78\\
w/o PoLIN w IN & $\checkmark$ & $\checkmark$ & $\checkmark$ & & $\checkmark$ & $\checkmark$ & & & & 40.73 & 89.62\\
w/o PoLIN w LIN & $\checkmark$ & $\checkmark$ & $\checkmark$ & & $\checkmark$ & &$\checkmark$ & & & 39.30 & 90.66\\
w/o AdaPoLIN w AdaIN & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$& && & $\checkmark$ & & 40.16 & 90.31\\
w/o AdaPoLIN w AdaLIN & $\checkmark$ & $\checkmark$ &$\checkmark$ & $\checkmark$ & & & & & $\checkmark$ & 39.52 & 91.93\\
AniGAN (Ours) & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & & & & & \textbf{38.45} & \textbf{86.98}\\
\bottomrule
\end{tabular}
\label{fid_ablation}
\vspace*{-6pt}
\end{table*}
\begin{table*}[t]
\centering
\caption{Quantitative comparison for ablation study using LPIPS score. Higher is better}
\begin{tabular}{p{3.7cm} p{0.4cm}<{\centering} p{0.35cm}<{\centering} p{0.36cm}<{\centering} p{0.67cm}<{\centering} p{1.3cm}<{\centering} p{0.2cm}<{\centering} p{0.3cm}<{\centering} p{0.67cm}<{\centering} p{0.9cm}<{\centering} m{1.45cm}<{\centering} m{1.78cm}<{\centering}}
\toprule
Method & ASC & FST & DB & PoLIN & AdaPoLIN & IN & LIN & AdaIN &AdaLIN & Face2anime & Selfie2anime \\
\midrule
w/o ASC & & $\checkmark$ & $\checkmark$ & $\checkmark$ &$\checkmark$ & & & & & 0.375& 0.321\\
w/o FST & $\checkmark$ & & $\checkmark$ & & $\checkmark$ & & & & &0.391 & 0.340 \\
w/o DB & $\checkmark$ & $\checkmark$ & & $\checkmark$ & $\checkmark$ & & & & & 0.395 & 0.342 \\
w/o PoLIN w IN & $\checkmark$ & $\checkmark$ & $\checkmark$ & & $\checkmark$ & $\checkmark$ & & & & 0.409 & 0.362\\
w/o PoLIN w LIN & $\checkmark$ & $\checkmark$ & $\checkmark$ & & $\checkmark$ & &$\checkmark$ & & &0.402 & 0.367\\
w/o AdaPoLIN w AdaIN & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$& && & $\checkmark$ & & 0.400 & 0.356 \\
w/o AdaPoLIN w AdaLIN & $\checkmark$ & $\checkmark$ &$\checkmark$ & $\checkmark$ & & & & & $\checkmark$ & 0.397 & 0.336 \\
AniGAN (Ours) & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & $\checkmark$ & & & & & \textbf{0.414} & \textbf{0.372}\\
\bottomrule
\end{tabular}
\label{LPIPS_ablation}
\vspace*{-12pt}
\end{table*}
Since CycleGAN, UGATIT focus on one-to-one mapping and cannot generate multiple outputs given a source image, we do not include them for comparison of translation diversity.
We also do not compare with MUNIT and FUNIT, since MUNIT does not take a references as the input and FUNIT focuses on few-shot learning instead of translation diversity.
Instead, we compare with DRIT++ and EGSC-IT, which are state-of-the-arts in reference-guided methods.
DRI++ uses a regularization loss which explicitly encourages the generator to generate diverse results. Although our method does not impose such loss, LPIPS scores in Table \ref{LPLIS} shows that our method outperforms DRIT++ and EGSC-IT on translation diversity, thanks to the generator and discriminator of our method.
We compare our method with three reference-guided methods FUNIT, EGSC-IT and DRIT++.
Since translation methods (e.g., CycleGAN and UGATIT) do not transfer the information of a reference image, we do not compare with these methods for fair comparison. The subjective user study is conducted for face2anime and selfie2anime dataset, respectively, where 10 pairs of photo-faces and anime-faces in each dataset are fed into these four translation methods to generate anime-faces.
We receive $1,200$ answers from 20 subjects in total for each dataset, where each method is compared $600$ times.
As shown in Table \ref{user_study}, most subjects are in favor of our method for the results on both face2anime and selfie2anime datasets, demonstrating that the anime-faces translated by our method are usually the most visually appealing to the subjects.
\subsection{User Study}
We conduct a subjective user study to further evaluate our method.
20 subjects are invited to participate in our experiments, whose ages range from 22 to 35.
Following \cite{Pix2PixHD, MUNIT}, we adopt pairwise A/B test.
For each subject, we show a source photo-face, a reference anime-face, and two anime-faces generated by two different translation methods. The generated anime-faces are presented in a random order, such that subjects are unable to infer which anime-faces are generated by which translation methods.
We then ask each subject the following question:
"\textit{Q: Which generated anime-faces has better visual quality by considering the source photo-face and the anime styles of the reference anime-face?} "
We compare our method with three reference-guided methods FUNIT, EGSC-IT and DRIT++.
Since translation methods (e.g., CycleGAN and UGATIT) do not transfer the information of a reference image, we do not compare with these methods for fair comparison. The subjective user study is conducted for face2anime and selfie2anime dataset, respectively, where 10 pairs of photo-faces and anime-faces in each dataset are fed into these four translation methods to generate anime-faces.
We receive $1,200$ answers from 20 subjects in total for each dataset, where each method is compared $600$ times.
As shown in Table \ref{user_study}, most subjects are in favor of our results on both face2anime and selfie2anime datasets, demonstrating that anime-faces translated by our method are usually the most visually appealing to the subjects.
\subsection{Ablation Study} \label{ablation}
We conduct ablation experiments to validate the effectiveness of individual components in our method: (1) ASC block, (2) FST block, (3) PLIN and AdaPLIN, (4) double-branch discriminator.
\begin{figure}[t]
\centering
\begin{tabular}{c:ccc}
\begin{minipage}[t]{0.10\textwidth}
\centering
Reference
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/64_row_0_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/14_row_7_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/71_row_6_reference.png}
\end{minipage}
\begin{minipage}[t]{0.10\textwidth}
\centering
Source
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/64_row_0_content.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/14_row_7_content.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/71_row_6_content.png}
\end{minipage}
&
\begin{minipage}[t]{0.10\textwidth}
\centering
w/o ASC
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/64_row_0_wo_ASC_m.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/14_row_7_wo_ASC_m.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/71_row_6_wo_ASC_m.png}
\end{minipage}
\begin{minipage}[t]{0.10\textwidth}
\centering
Ours
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/64_row_0_aniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/14_row_7_aniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/71_row_6_aniGAN.png}
\end{minipage}
\end{tabular}
\caption{ Visual comparison of the contributions of ASC blocks, where ``w/o ASC" improperly renders ''brown" right eyes in all the generated images.}
\label{fig:ASC}
\vspace*{-12pt}
\end{figure}
\begin{figure*}[t
\centering
\subfigure[]{
\begin{minipage}[t]{0.10\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11973_2773005_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11783_170012_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ12601_426039_reference.png}
\end{minipage}
}
\subfigure[]{
\begin{minipage}[t]{0.1\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11973_2773005_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11783_170012_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ12601_426039_source.png}
\end{minipage}
}
\subfigure[]{
\begin{minipage}[t]{0.1\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11973_2773005_AniGAN_wo_PLIN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11783_170012_AniGAN_wo_PLIN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ12601_426039_AniGAN_wo_PLIN.png}
\end{minipage}
}
\subfigure[]{
\begin{minipage}[t]{0.1\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11973_2773005_AniGAN_wo_PLIN_w_LIN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11783_170012_AniGAN_wo_PLIN_w_LIN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ12601_426039_AniGAN_wo_PLIN_w_LIN.png}
\end{minipage}
}
\subfigure[]{
\begin{minipage}[t]{0.1\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/36_row_0_wo_AdaPLIN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/76_row_7_wo_AdaPLIN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/82_row_6_wo_AdaPLIN.png}
\end{minipage}
}
\subfigure[]{
\begin{minipage}[t]{0.1\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11973_2773005_AniGAN_wo_adaplin_w_adalin.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11783_170012_AniGAN_wo_adaplin_w_adalin.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ12601_426039_AniGAN_wo_adaplin_w_adalin.png}
\end{minipage}
}
\subfigure[]{
\begin{minipage}[t]{0.1\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11973_2773005_aniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ11783_170012_aniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_norm/imgHQ12601_426039_aniGAN.png}
\end{minipage}
}
\caption{ Visual comparison of the contributions of PoLIN and AdaPoLIN. Row from left to right: (a) reference anime-faces, (b) source photo-faces, generated faces by (c) {``w/o PoLIN w/ IN"}, (d) {``w/o PoLIN w/ LIN"}, (e) {``w/o AdaPoLIN w/ AdaIN"}, (f) {``w/o AdaPoLIN w/ AdaLIN"} and (g) our method.}
\label{fig:normalization}
\vspace*{-14pt}
\end{figure*}
\begin{figure}[t
\centering
\begin{tabular}{c:ccc}
\begin{minipage}[t]{0.10\textwidth}
\centering
Reference
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/1122145.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/2815053.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/2606106.png}
\end{minipage}
\begin{minipage}[t]{0.10\textwidth}
\centering
Source
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/imgHQ11430.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/imgHQ11218.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/imgHQ11446.png}
\end{minipage}
&
\begin{minipage}[t]{0.10\textwidth}
\centering
w/o FST
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/imgHQ11430_1122145_AniGAN_wo_FST.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/imgHQ11218_2815053_AniGAN_wo_FST.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/imgHQ11446_2606106_AniGAN_wo_FST.png}
\end{minipage}
\begin{minipage}[t]{0.10\textwidth}
\centering
Ours
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/imgHQ11430_1122145_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/imgHQ11218_2815053_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_fst/imgHQ11446_2606106_AniGAN.png}
\end{minipage}
\end{tabular}
\caption{ Visual comparison of the contributions of FST blocks}
\label{fig_fst}
\vspace*{-16pt}
\end{figure}
\textbf{ASC block.}
We seek to validate whether the ASC block effectively retains the style information of the reference image and helps the generator transfer the style characteristics.
Note that the key differentiating factor in our ASC block is the removal of residual blocks from the bottleneck of our generator, different from start-of-the-art image translation methods (e.g., MUNIT, UGATIT, and FUNIT). We hence implement a baseline called \textbf{``w/o ASC"} which adds residual blocks in the bottleneck of our decoder.
As shown in Fig. \ref{fig:ASC}, \textbf{``w/o ASC"} tends to ignore certain style information due to the additional residual blocks. For example, \textbf{``w/o ASC"} ignores the styles of the right eyes in the references and renders ''brown" right eyes in all the generated images.
In contrast, our method well and consistently renders the style of the left and right eyes, despite no face landmarks or face parsing are used. Clearly, our method outperforms \textbf{``w/o ASC"} in transferring the styles of reference anime-faces, thanks to the ASC block.
\textbf{FST block.}
Different from existing methods (e.g., MUNIT, FUNIT, EGSC-IT) which use typical upsampling blocks in the decoder,
FST is additionally equipped with our normalization functions.
To validate the effectiveness of FST, we build a baseline (called \textbf{``w/o FST"}) which replaces the FST blocks with typical upsampling blocks (like MUNIT, FUNIT, and EGSC-IT) without our normalization functions.
As shown in Fig. \ref{fig_fst}, \textbf{``w/o FST"} performs poorly in converting the shape of local facial features and transferring styles. For example, \textbf{``w/o FST"} poorly converts the face shapes and eyes and introduces artifacts in the generated faces.
In contrast, with FST, our method better converts the local shapes into anime-like ones than \textbf{``w/o FST"}. Moreover, our method also better transfers the styles of the reference anime-faces to the generated faces than \textbf{``w/o FST"} does.
Similarly, the FID of \textbf{``w/o FST"} increases in Table \ref{fid_ablation}, indicating that the anime-faces generated by \textbf{``w/o FST"} are less plausible than those generated by our method.
In addition, the LPIPS scores of \textbf{w/o FST} and our method in Tab. \ref{LPIPS_ablation} shows that FST is helpful for generating diverse anime-faces.
\textbf{PoLIN and AdaPoLIN.}
We build four baselines to evaluate the effectiveness of PoLIN and AdaPoLIN. The first and second baselines are built for evaluating PoLIN, and the third and fourth baseline are for AdaPoLIN. The first baseline, named \textbf{``w/o PoLIN w/ IN"}, is constructed by replacing PoLIN with IN in \cite{IN}. We build \textbf{``w/o PoLIN w/ IN"}, since EGSC-IT, DRIT and FUNIT employ IN in the up-sampling
convolutional layers of their decoder, different from our method with PoLIN. The second baseline, named \textbf{``w/o PoLIN w/ LIN"}, is constructed by replacing PoLIN with layer-Instance Normalization (LIN).
The third baseline, called \textbf{``w/o AdaPoLIN w/ AdaIN"}, replaces AdaPoLIN with AdaIN in \cite{AdaIN}, which was employed by many translation methods e.g., \cite{MUNIT,ma2018exemplar,FUNIT,SketchEccv2020}.
The fourth baseline, called \textbf{``w/o AdaPoLIN w AdaLIN"}, replaces AdaPoLIN with AdaLIN which is used in UGATIT.
The FID scores in Table \ref{fid_ablation} show that our method outperforms the four baselines. Without PoLIN, the performance of transforming local shapes into anime-like ones is degraded, as shown in the results generated by \textbf{``w/o PoLIN w/ IN"} and \textbf{``w/o PoLIN w/ LIN"} in Fig. \ref{fig:normalization}. For example, \textbf{``w/o PoLIN w/ IN"} introduces artifacts in the hair at the right boundary of generated faces in the first row of Figs. \ref{fig:normalization}(c) and (d). Similarly, without AdaPoLIN, both \textbf{``w/o AdaPoLIN w/ AdaIN"} and \textbf{``w/o AdaPoLIN w AdaLIN"} perform worse than our method. Table \ref{fid_ablation} shows that \textbf{``w/o AdaPoLIN w AdaIN"} and \textbf{``w/o AdaPoLIN w AdaLIN"} degrade the performance in terms of translation diversity.
\begin{figure}[t]
\centering
\begin{tabular}{c:ccc}
\centering
\begin{minipage}[t]{0.10\textwidth}
\centering
Reference
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/1480052.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/1997070.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/2528106.png}
\end{minipage}
\begin{minipage}[t]{0.10\textwidth}
\centering
Source
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/imgHQ12469.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/imgHQ12384.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/imgHQ11965.png}
\end{minipage}
&
\begin{minipage}[t]{0.10\textwidth}
\centering
w/o DB
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/imgHQ12469_1480052_AniGAN_oneb.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/imgHQ12384_1997070_AniGAN_oneb.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/imgHQ11965_2528106_AniGAN_oneb.png}
\end{minipage}
\begin{minipage}[t]{0.10\textwidth}
\centering
Ours
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/imgHQ12469_1480052_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/imgHQ12384_1997070_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/figure_discri/imgHQ11965_2528106_AniGAN.png}
\end{minipage}
\end{tabular}
\caption{Visual comparison of the contributions of the double-branch discriminator }
\label{fig:discriminator}
\vspace*{-16pt}
\end{figure}
It is worth noting that all the above baselines, that replace our normalization functions with other normalization functions employed by DRIT, EGSC-IT, UGATIT, etc., still achieve better FID and LPIPS scores than state-of-the-art methods, as shown in Table \ref{fid1}, \ref{LPLIS}, \ref{fid_ablation}, and \ref{LPIPS_ablation}. This indicates that the architectures of our generator and discriminator are more advantageous for StyleFAT task.
\textbf{Double-branch discriminator.}
We implement a baseline \textbf{``w/o DB"} that removes the photo-face branch (i.e., the branch that discriminates real/fake photo-faces) from our discriminator. Table \ref{fid_ablation} shows that it yields poorer FID than our method. Table \ref{LPIPS_ablation} also shows LPIPS scores of \textbf{``w/o DB"} is worse than that of our method.
More specifically, as shown in Fig. \ref{fig:discriminator}, \textbf{``w/o DB"} distorts local facial shapes, leading to low-quality generated faces, especially for challenging source photo-faces.
This is because \textbf{``w/o DB"} would mainly focus on generating plausible anime images, rather than on generating plausible anime human faces.
In contrast, our method generates high-quality anime-faces, thanks to the additional photo-face branch in the discriminator. With the photo-face branch in our discriminator, the photo-face and anime-face branches share the first few shallow layers, which helps the anime-face branch better learn real facial features from photo-faces so as to well discriminate low-quality anime-faces with distorted facial features.
\section{Conclusion}
In this paper, we propose a novel GAN-based method, called AniGAN, for style-guided face-to-anime translation. A new generator architecture and two normalization functions are proposed, which effectively transfer styles from the reference anime-face, preserve global information from the source photo-face and convert local facial shapes into anime-like ones. We also propose a double-branch discriminator to assist the generator to produce high-quality anime-faces. Extensive experiments demonstrate that our method achieves superior performance compared with state-of-the-art methods.
\begin{figure}[!htbp]
\tabcolsep=0.15cm
\centering
\begin{tabular}{cc:cccccc}
\centering
{Reference} & {Source} &StarGAN-v2&Ours\\
\begin{minipage}[t]{0.095\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11606_1916109_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_starv2/imgHQ12556_232096_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11736_635009_reference.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_starv2/imgHQ11973_2773005_reference.png}
\end{minipage}
&
\begin{minipage}[t]{0.095\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11606_1916109_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_starv2/imgHQ12556_232096_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11736_635009_source.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_starv2/imgHQ11973_2773005_source.png}
\end{minipage}
&
\begin{minipage}[t]{0.095\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11606_1916109_StarGANV2.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_StarGANV2.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_StarGANV2.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_starv2/imgHQ12556_232096_StarGANV2.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_starv2/imgHQ11736_635009_StarGANV2.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_starv2/imgHQ11973_2773005_StarGANV2.png}
\end{minipage}
&
\begin{minipage}[t]{0.095\textwidth}
\centering
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11606_1916109_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11677_1486059_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11419_1901004_aniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_starv2/imgHQ12556_232096_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_c/imgHQ11736_635009_AniGAN.png}
\includegraphics[bb=0 0 128 128,width=1\textwidth]{figures/fig_starv2/imgHQ11973_2773005_AniGAN.png}
\end{minipage}
\end{tabular}
\caption{\small Comparison results on the face2anime dataset. From left to right: source photo-face, reference anime-face, the results by StarGAN-v2 \cite{starganv2} and our AniGAN. }
\label{fig:starganv2}
\vspace*{-15pt}
\end{figure}
|
\section{Introduction}
\input{introduction}
\subsection{Contributions and Guide}
This work advances prior art \cite{scobee2019maximum} in inferring state-action constraints:
\begin{itemize}
\item by respecting causality using the principle of maximum \textbf{causal} entropy for likelihood generative models
\item by extending the hypothesis family to include risk-tolerating \textbf{chance} constraints
\item and by streamlining the algorithm into one backwards pass, thereby maintaining the same computational complexity as the non-stochastic version \cite{scobee2019maximum}
\end{itemize}
\section{Background}
\subsection{Maximum Likelihood Infererence}
The inference task is to fit parameters to observed demonstration trajectories.
We follow the maximum likelihood inference framework that models data as samples of a parametrized distribution.
In our case, the distribution can be derived from the task description and agent model. These will be described in subsections \ref{sec:MDP} and \ref{sec:Boltzmann}, respectively.
The parameters to fit will be which candidate constraints $C \in \mathcal{C}$ are effecting the expert and at what risk tolerance $\alpha_C$. These chance constraints will be described in subsection \ref{sec:ChanceConstraints}.
\subsection{Markov Decision Processes}
\label{sec:MDP}
\input{mdp_definition}
\subsection{Chance Constraints}
\label{sec:ChanceConstraints}
\input{transitionChanceConstraints}
\subsection{Boltzmann Distribution on Controller Sequences}
\label{sec:Boltzmann}
\input{softBellmanDistributionOnActionsV2}
\section{Constraint Inference Statistics}
\input{constraintInferenceStats.tex}
\section{Dynamic Programming}
\label{sec:DynDeriv}
\input{dynamicProgrammingMLCIS_v3}
\section{Algorithm}
\input{algorithm}
\section{Results}
\input{results}
\section{Limitations and Future Work}
The algorithms set forth in this paper focused on discretized state and action spaces.
For controlling many systems on practical timescales, the state must be handled as a continuous parameter.
Future work should investigate how gridded state spaces like in Figure \ref{fig:demonstration} could be refined to approximate continuous state spaces.
Reducing the algorithm to a variant Bellman backup, as we did in Theorem \ref{thm:dynprog}, suggests that the continuous variant may just be solving a Hamilton-Jacobi-Bellman equation.
These partial differential equations have a rich literature investigating their solution, including toolsets like \cite{MitchellToolbox}.
Extending constraint inference to stochastic systems raises questions of whether human experts might be better modeled using a prospect-theoretic or risk-sensitive measure as in \cite{mazumdar2017gradient}.
Future work should investigate how human heuristics for statistical prediction might impact the way demonstrations are generated.
The algorithm should be designed to be robust to these biases or even leverage their structure.
\section{Discussion and Conclusion}
\input{conclusion}
\bibliographystyle{plain}
\subsection{Determining Chance Constraint Risk Levels from Demonstrated Transitions}
Corollary \ref{cor:chanceConstraintLevels} states that this set of candidates can be further reduced to only those whose newly added $\psi(x)$ are as exclusive as possible without excluding any of the demonstrations $(\tilde{x}_{0:T},\tilde{a}_{0:T}) \in \mathcal{D}$. That is, when adding state constraints, the newly added exclusion threshold $\psi(x)$ must as low as possible while still being greater than all transition probabilities to $x$ that were chosen by the expert in their demonstrations.
For simplicity, we will bound lowerbound $\psi(x)$ to prevent any precursor states of $x$ from having all its available actions ruled out thereby dooming any trajectory entering that precursor state to necessarily violate the chance constraint on $x$. Therefore this lowerbound $\Psi(x)$ must be defined:
\begin{align}
\Psi(x') =
\max_{ \{
x \mid
\exists \hat{a}
\ni
P(x' |
x,
\hat{a}
)
> 0
\} }
\min_{a}
P(x' | x, a)
\end{align}
This implies that the new $\psi(x)$ should be:
\begin{align}
\psi(x) =
\max \left\{
\max_{
(
\tilde{x}_{0:T},
\tilde{a}_{0:T}
)
\in \mathcal{D}
}
\max_{
t
\in [0:T-1]
}
S(
\tilde{x}(t),
\tilde{a}(t),
x
) ,
\Psi(x)
\right\}
\label{eq:riskLevelSelection}
\end{align}
\begin{algorithm}
\SetAlgoLined
\KwResult{$V_{C^0,t}$ and a column vector $F$ where each entry corresponds to the $F_{C^+,0}$ that adds one state/action constraint on top of $C^+ \in \mathcal{C^+_0}$}
\SetEndCharOfAlgoLine{\relax}
\For{$x \in \mathcal{X}$}{
$Z(T,x) \leftarrow \exp(w(x))$\;
$F(T,x) \leftarrow 1$\;
}
\For{$t \in [T-1,0]$}{
\For{$x \in \mathcal{X}$}{
$Z(t,x) \leftarrow 0$ \;
$F(t,x) \leftarrow 0$ \;
\For{$a \in \mathcal{A}$}{
$Q(t,x,a) \leftarrow r(x,a)$ \;
$D(t,x,a) \leftarrow 0$ \;
\For{$x' \in \mathcal{X}$}{
$Q(t,x,a) += S(x,a,x') \log(Z(t+1,x'))$ \;
$D(t,x,a) += S(x,a,x') \log(F(t+1,x'))$ \;
}
$Z(t,x) += \Phi_{C^0}(x,a) \exp(Q(t,x,a))$ \;
$F(t,x) += \Phi_{C \in \mathcal{C}}(x,a)$
$\exp(Q(t,x,a))$
$\exp(D(t,x,a))$ \;
}
$F(t,x) = F(t,x) / Z(t,x)$
}
}
\caption{\label{alg:Fratio}Modified Bellman Backup with Value Ratio}
\end{algorithm}
The most likely constraint is then whichever one still allows the observed demonstrations and has the smallest normalizing constant from the starting state $F_{C,0}(x_0)$.
Note that Algorithm \ref{alg:Fratio} has computations on the order of $O(|\mathcal{X}|^2 (|\mathcal{X}| + |\mathcal{A}|)$, identical to the computational complexity of prior art in maximum likelihood constraint inference \cite{scobee2019maximum}.
|
\section{Introduction}\label{sec:introduction}}
We consider an optimal steering problem for networks flows over a finite or infinite time horizon. Specifically, we derive discrete counterparts of our {\em cooling} results in \cite{CGPcooling}. The goal is to steer the Markovian evolution to a steady state with desirable properties while minimizing relative entropy, or relative entropy rate, over admissible distributions on paths. This relates to a special Markov Decision Process problem, cf. \cite[Section 6]{CGPAnnualReview} which is referred to as {\em Schr\"{o}dinger's Bridge Problem} (SBP) or {\em regularized optimal mass transport} (OMT). A rich history on this circle of ideas originates from two remarkable papers by Erwin Schr\"odinger in 1931/32 \cite{S1,S2} who was interested in a large deviation problem for a cloud of Brownian particles, and of earlier work by Gaspar Monge in 1781 on the problem of optimally transporting mass between sites \cite{Monge}.
Important contributions on the existence question in SBP, left open by Schr\"odinger, were provided over time by Fortet, Beurling, Jamison and F\"ollmer \cite{For,Beu,Jam2,F2}.
It should be remarked that Fortet's proof in 1940 is algorithmic, establishing convergence of a (rather complex) iterative scheme. This predates by more than twenty years the contribution of Sinkhorn \cite{Sin64} who established convergence in a special case of the discrete problem The latter proof is much simpler than in the continuous (time and space) problem solved by Fortet. Thus, these algorithms should be called Fortet-IPF-Sinkhorn, where IPF stands for the Iterative Proportional Fitting algorithm proposed in 1940 without proof of convergence in \cite{DS1940}. These problems admit a ``static" and a ``dynamic" formulation. While the static, discrete space problem was studied by many, starting from Sinkhorn, see \cite{Cuturi,PC} and references therein, the dynamic, discrete problem was considered in \cite{PT,GP,CGPT1,CGPT2,CGPT3}. We provide in this paper a new fluid-dynamic derivation of the Schr\"odinger system on which the iteration is based.
Convergence of the iterative scheme in a suitable projective metric was also established in \cite{GP} for the discrete case and in \cite{CGP_SIAMAppl.} for the continuous case.
This topic lies nowadays at the crossroads of many fields of science such as probability, statistical physics, optimal mass transport, machine learning, computer graphics, statistics, stochastic control, image processing, etc. Several survey papers have appeared over time emphasizing different aspects of the subject, see \cite{W,leo2,PC,CGP_SIREV,CGPAnnualReview,CGP_CSM}.
The paper is outlined as follows.
In Section \ref{GBProblems}, we present background on the discrete-time and space Schr\"odinger bridge problem while emphasizing its ``fluid dynamic'' formulation (the original proof of a key well-known result is deferred to the Appendix).
In Section \ref{OSSS}, we introduce the corresponding
infinite-horizon steering problem by minimizing the entropy rate with respect to the prior measure. We then discuss existence for the one-step Schr\'odinger system leading to existence for the steering problem. Section \ref{REVERSE} is devoted to an interesting result linking optimality in the steering problem to reversibility of the solution. The final Section \ref{cooling} applies all the previous results to cooling, where the goal is to optimally steer the Markov chain to a steady state corresponding to a lower {\em effective} temperature.
\section{Discrete-time dynamic bridges: background}\label{GBProblems}
We begin by discussing a paradigm of great significance in network flows. It amounts to designing probabilistic transitions between nodes, and thereby probability laws on path spaces, so as to reconcile marginal distributions with priors that reflect on the structure of the network and objectives on transferance of resourses across the network. The basic formulation we discuss amounts to the so called Schr\"odinger bridge problem in discrete-time. This dynamic formulation echoes the fluid dynamic formulation of the classical (continuous time and space) Schr\"odinger bridge problem. Certain proofs that we present are mildly new.
We consider a directed, strongly connected, aperiodic graph ${\bf G}=(\mathcal X,\mathcal E)$ with vertex set $\mathcal X=\{1,2,\ldots,n\}$ and edge set $\mathcal E\subseteq \mathcal X\times\mathcal X$,
and we consider trajectories/paths on this graph over the time set ${\mathcal T}=\{0,1,\ldots,N\}$. The family of feasible paths $x=(x_0,\ldots,x_N)$ of length $N$, namely paths such that $x_ix_{i+1}\in\mathcal E$ for $i=0,1,\ldots,N-1$,
is denoted by ${\mathcal FP}_0^N\subseteq\mathcal X^{N+1}$.
We seek a probability distribution $\mathfrak P$ on the space of paths ${\mathcal {FP}}_0^N$ with prescribed initial and final marginals
$\nu_0(\cdot)$ and $\nu_N(\cdot)$, respectively, and such that the resulting random evolution
is closest to a ``prior'' measure ${\mathfrak M}$ on ${\mathcal {FP}}_0^N$ in a suitable sense.
The prior law for our problem is a Markovian evolution
\begin{equation}\label{FP}
\mu_{t+1}(x_{t+1})=\sum_{x_t\in\mathcal X} \mu_t(x_t) m_{x_{t}x_{t+1}}(t)
\end{equation}
with nonnegative distributions $\mu_t(\cdot)$ over $\mathcal X$, $t\in{\mathcal T}$, and weights $m_{ij}(t)\geq 0$ for all indices $i,j\in{\mathcal X}$ and all times.
In accordance with
the topology of the graph, $m_{ij}(t)=0$ for all $t$ whenever $ij\not\in\mathcal E$. Often, but not always, the matrix
\begin{equation}\label{eq:matrixM}
M(t)=\left[ m_{ij}(t)\right]_{i,j=1}^n
\end{equation}
does not depend on $t$.
Here, $m_{ij}$ plays the same role as that of the heat kernel in continuous-time.
However, herein, the transition matrix $M(t)$ may not represent a {\em probability} transition matrix
in that the rows of $M(t)$ do not necessarily sum up to one. Thus, the ``total transported mass'' is not necessarily preserved,
in that $\sum_{x_t\in\mathcal X} \mu_t(x_t)$ may be afunction of time.
A particular case of interest is when $M$ is the {\em adjacency matrix} of the graph that encodes the topological structure of the network.
The evolution \eqref{FP}, together with measure $\mu_0(\cdot)$, which we assume positive on $\mathcal X$, i.e.,
\begin{equation}\label{eq:mupositive}
\mu_0(x)>0\mbox{ for all }x\in\mathcal X,
\end{equation}
induces
a measure ${\mathfrak M}$ on ${\mathcal {FP}}_0^N$ as follows. It assigns to a path $x=(x_0,x_1,\ldots,x_N)\in{\mathcal {FP}}_0^N$ the value
\begin{equation}\label{prior}{\mathfrak M}(x_0,x_1,\ldots,x_N)=\mu_0(x_0)m_{x_0x_1}(0)\cdots m_{x_{N-1}x_N}(N-1),
\end{equation}
and gives rise to a flow
of {\em one-time marginals}
\[\mu_t(x_t) = \sum_{x_{\ell\neq t}}{\mathfrak M}(x_0,x_1,\ldots,x_N), \quad t\in\mathcal T.\]
\begin{definition} We denote by ${\mathcal P}(\nu_0,\nu_N)$ the family of probability distributions on ${\mathcal {FP}}_0^N$ having the prescribed marginals $\nu_0(\cdot)$ and $\nu_N(\cdot)$.
\end{definition}
We seek a distribution in this set which is closest to the prior ${\mathfrak M}$ in {\em relative entropy} (divergence, Kullback-Leibler index) defined by
\begin{equation*}
{\mathfrak D}(\mathfrak P\|Q):=\left\{\begin{array}{ll} \sum_{x}\mathfrak P(x)\log\frac{\mathfrak P(x)}{\mathfrak M(x)}, & {\rm Supp} (\mathfrak P)\subseteq {\rm Supp} (\mathfrak M),\\
+\infty , & {\rm Supp} (\mathfrak P)\not\subseteq {\rm Supp} (\mathfrak M),\end{array}\right.
\end{equation*}
Here, by definition, $0\cdot\log 0=0$.
The value of ${\mathfrak D}(\mathfrak P\|\mathfrak M)$ may turn out to be negative due to the different total masses in the case when ${\mathfrak M}$ is not a probability measure. The optimization problem, however, poses no challenge as the relative entropy is (jointly) convex over this larger domain and is bounded below.
This brings us to the
so-called
{\em Schr\"odinger Bridge Problem} (SBP):
\begin{problem}\label{prob:optimization}
Determine
\begin{eqnarray}\label{eq:optimization}
\mathfrak P^*[\nu_0,\nu_N]:={\rm argmin}\{ {\mathfrak D}(\mathfrak P\|{\mathfrak M}) \mid \mathfrak P\in {\mathcal P}(\nu_0,\nu_N)
\}.
\end{eqnarray}
\end{problem}
By analyzing the contribution to the entropy functional of transitions along edges, in a sequential manner, we arive at a formulation that parallels the fluid-dynamic formulation of the Schr\"odinger problem in continuous time\footnote{
The fluid dynamic reformulation of Schr\"odinger's bridge problem in continuous space/time drew important connections with Monge-Kantorovich optimal mass transport (OMT) \cite{BB} --the latter seen as a limit when the stochastic excitation reduces to zero \cite{leo2}. In particular, it pointed to
a time-symmetric formulation of Schr\"odinger's bridges \cite{CGPJOTA}.}, considered in \cite[Section 4]{leo2} and \cite{CGPJOTA}.
Suppose we restrict our search in Problem \ref{prob:optimization} to Markovian $\mathfrak P$'s. In analogy to (\ref{prob:optimization}), we then have
\begin{equation}\label{markovianmeasure}
\mathfrak P(x_0,x_1,\ldots,x_N)=\nu_0(x_0)\pi_{x_0x_1}(0)\cdots \pi_{x_{N-1}x_N}.
\end{equation}
Also let $p_t$ be the one-time marginals of $\mathfrak P$. i.e.
\[p_t(x_t) = \sum_{x_{\ell\neq t}}\mathfrak P(x_0,x_1,\ldots,x_N), \quad t\in\mathcal T.\]
We finally have the update mechanism
\begin{equation}\label{update}
p_{t+1}(x_{t+1})=\sum_{x_t\in\mathcal X} p_t(x_t) \pi_{x_{t}x_{t+1}}(t)
\end{equation}
which, in vector form, is
\begin{equation}
\label{eq:fp}
p_{t+1}=\Pi'(t) p_{t}.
\end{equation}
Here $\Pi=\left[ p_{ij}(t)\right]_{i,j=1}^n$ is the transition matrix and prime denotes transposition.
Using (\ref{prior})-(\ref{markovianmeasure}) we obtain
\[{\mathfrak D}(\mathfrak P\|{\mathfrak M})={\mathfrak D}(\nu_0\|\mu_0) + \sum_{t=0}^{N-1}\sum_{x_t}{\mathfrak D}(\pi_{x_tx_{t+1}}(t)\|m_{x_tx_{t+1}}(t))p_t(x_t).
\]
Since ${\mathfrak D}(\nu_0\|\mu_0)$ is invariant over ${\mathcal P}(\nu_0,\nu_N)$, we can now have the following equivalent (fluid dynamic) formulation:
\vspace{.3cm}
\begin{problem}\label{problemfluid-dynamic}
\begin{align
{\rm min}_{(p,\pi)}&\left\{\sum_{t=0}^{N-1}\sum_{x_t}{\mathfrak D}\left(\pi_{x_tx_{t+1}}(t)\|m_{x_tx_{t+1}}(t))p_t(x_t)\right)\right\},\nonumber\\\label{fluid-dynamic2}{\rm \hspace*{-15pt}subject\; to}\; & \nonumber\\
&p_{t+1}(x_{t+1})=\sum_{x_t\in\mathcal X} p_t(x_t) \pi_{x_{t}x_{t+1}}(t),\\
& \sum_{x_{t+1}}\pi_{x_{t}x_{t+1}}(t)=1, \; \forall x_t\in\mathcal X, \mbox{ and}\label{fluid-dynamic3}\\
&p_0(x_0)=\nu_0(x_0), \quad p_N(x_N)=\nu_N(x_N),\label{fluid-dynamic4}\\
&\mbox{for } t\in\{0,1,\ldots,N-1\}.
\end{align}
\end{problem}
\noindent
Here $p=\{p_t; t=0,1,\ldots,N\}$ is a flow of probability distributions on $\mathcal X$ (corresponds to the flow of densities $\rho_t$ in the continuous setting) and $\pi=\{\pi_{x_tx_{t+1}}(t); t=0,1,\ldots,N-1\}$ is a flow of transition probabilities (corresponds to the flow of the drifts).
The relative entropy ${\mathfrak D}\left(\pi_{x_tx_{t+1}}(t)\|m_{x_tx_{t+1}}(t))\right)$ between transition laws, corresponds to the
square of the modulus of the velocity field in the continuous setting \cite{CGPJOTA}, so that ${\mathfrak D}\left(\pi_{x_tx_{t+1}}(t)\|m_{x_tx_{t+1}}(t))p_t(x_t)\right)$ in fact corresponds to kinetic energy.
We do not impose explicitly the non-negativity constraint on the $p_t$ and $\pi_{x_tx_{t+1}}(t)$ as
it will be automatically satisfied (due to the fact that the relative entropy grows unbounded at the boundary of the admissible distributions and transition probabilities). We have the following theorem:
\begin{theorem} \label{maincl}Suppose there exists a pair of nonnegative functions $(\varphi,\hat{\varphi})$ defined on $[0,T]\times{\cal X}$ and satisfying the system
\begin{eqnarray}\label{scsi1}\varphi(t,x_t)=\sum_{x_{t+1}}m_{x_{t}x_{t+1}}(t)\varphi(t+1,x_{t+1}),\\
\hat{\varphi}(t+1,x_{t+1})=\sum_{x_t} m_{x_{t}x_{t+1}}(t)\hat{\varphi}(t,i),\label{scsi2}
\end{eqnarray}
for $0\le t \le(T-1)$, as well as the boundary conditions
\begin{equation}\label{scsi3}
\varphi(0,x_0)\cdot\hat{\varphi}(0,x_0)=\nu_)(x_0),\; \varphi(T,x_T)\cdot\hat{\varphi}(T,x_T)=\nu_T(x_T),
\end{equation}
for $x_t\in\mathcal X$ and $t\in\{0,T\}$, accordingly.
Suppose moreover that $\varphi(t,i)>0,\; \forall 0\le t\le T, \forall i\in{\cal X}$. Then, the Markov distribution $\mathfrak P^*$ in ${\mathcal P}(\nu_0,\nu_N)$ having transition probabilities
\begin{equation}\pi^*_{x_{t}x_{t+1}}(t)=m_{x_{t}x_{t+1}}(t)\frac{\varphi(t+1,x_{t+1})}{\varphi(t,x_{t})}
\end{equation}
solves Problem \ref{problemfluid-dynamic}.
\end{theorem}
\begin{remark}\label{uniqueness}Notice that if $(\varphi,\hat{\varphi})$ satisfy (\ref{scsi1})-(\ref{scsi2})-(\ref{scsi3}), so does the pair $(c\varphi,\frac{1}{c}\hat{\varphi})$ for all $c>0$. Hence, uniqueness for the Schr\"{o}dinger system is always intended as uniqueness of the ray.
\end{remark}
The proof of Theorem \ref{maincl} is deferred to Appendix \ref{A}.
Our derivation there of the Schr\"odinger system appears original and alternative to the two usual approaches. The first one uses Lagrange multipliers for the constraints involving the marginals \cite{GP} very much as in Schr\"odinger's original spirit. The second uses a ``completion of the relative entropy" argument, see \cite[Section IV]{PT}. Let $M(t)=\left(m_{ij}(t)\right)$. Under the assumption that the entries of the matrix product
\begin{equation}\label{posassumption}G:=M(0)M(1) \cdots M(N-2)M(N-1)
\end{equation}
are all positive, there exists a (unique in the projective geometry sense) solution to the system (\ref{scsi1})-(\ref{scsi3}) which can be computed through a Fortet-IPF-Sinkhorn iteration \cite{For,DS1940,Sin64,Cuturi,GP}.
\section{Optimal Steering to a Steady State}\label{OSSS}
Consider a given prior transition rate between the vertices of the graph, represented by a matrix $M$ as in (\ref{eq:matrixM}), which at present we assume as being time-invariant. Suppose $\pi$ is a desired stationary probability distribution that we wish to maintain over the state space $\mathcal X$ for the distribution of resources, goods, packets, vehicles, etc.\ at steady state. We address below the following basic question: {\em How can we modify $M$ so that the new evolution is close to the prior in relative entropy while at the same time it admits $\pi$ as invariant distribution?}
We address the above question next and provide a partial answer
to the complementary question on what conditions ensure that any given $\pi$ can be rendered the stationary distribution with a transition law compatible with the topology of the graph.
\subsection{Minimizing the relative entropy rate}
We consider the following problem inspired by \cite{DL}. Given a (prior) measure ${\mathfrak M}\in {\mathcal {FP}}_0^N$ as in Section \ref{GBProblems}, corresponding to a constant transition matrix $M$, and a probability vector $\pi$ on $\mathcal X$,
we are seeking a
(row)
stochastic matrix $\Pi$ such that
\[
\Pi'\pi=\pi
\]
while the corresponding (time-invariant) measure $\mathfrak P$ on $\mathcal X\times\mathcal X\times\cdots$ is close to $\mathfrak M$ in a suitable sense. Our interest here focuses on the infinite horizon case where $N$ is arbitrarily large.
To this, we let ${\mathcal P}$ be the family of probability distributions on ${\mathcal {FP}}_0^N$, and consider the following problem.
\begin{problem}\label{P1}
\begin{eqnarray}\nonumber
&&\min_{\mathfrak P\in\mathcal P}\lim_{N\rightarrow\infty}\frac{1}{N}{\mathfrak D}\left(\mathfrak P_{[0,N]}\|\mathfrak M_{[0,N]}\right)\\
{\rm subject\, to}\; &&\Pi'\pi=\pi,\nonumber\\\nonumber&&\Pi\mathds{1}=\mathds{1},
\end{eqnarray}
where $\mathds{1}$ is the vector with all entries equal to $1$.
\end{problem}
By formula $(18)$ in \cite{PT}, which applies to general (not necessarily mass preserving) Markovian evolutions, we have the following representation for the relative entropy between Markovian measures
\begin{eqnarray}\nonumber
&&{\mathfrak D}\left(\mathfrak P_{[0,N]}\|\mathfrak M_{[0,N]}\right)\\\nonumber&&={\mathfrak D}(\pi\|m_0)+\sum_{k=0}^{N-1}\sum_{i_k}{\mathfrak D}(\pi_{i_ki_{k+1}}\|m_{i_ki_{k+1}})\pi(i_k).
\end{eqnarray}
Thus, Problem \ref{P1} reduces to the {\em stationary} Schr\"{o}dinger bridge problem:
\begin{problem}\label{P2}
\begin{eqnarray}\nonumber
&& \min_{\left(\pi_{ij}\right)}\sum_{i,j\in\mathcal X}{\mathfrak D}(\pi_{ij}\|m_{ij})\pi(i),\\
{\rm subject\, to}\; &&\sum_{i\in\mathcal X} \pi_{ij}\pi(i)=\pi(j),\quad j\in\mathcal X, \nonumber\\
\nonumber&&\sum_{j\in\mathcal X}\pi_{ij}=1, \quad i\in\mathcal X.
\end{eqnarray}
\end{problem}
Note that in the above we have ignored enforcing the nonnegativity of the $\pi_{ij}$. This problem can readily be shown to be equivalent to a standard {\em one-step} Schr\"odinger bridge problem for the joint distributions $p(i,j)$ and $m(i,j)$ at times $t\in\{0,1\}$ with the two marginals equal to $\pi$. Indeed, a straightforward calculation gives
\begin{eqnarray}\nonumber&&\sum_{ij}\log\frac{\pi(i,j)}{m(i,j)}{\pi(i,j)}=\sum_{ij}\log\frac{\pi_{ij}\pi(i)}{m_{ij}m_0(i)}{\pi_{ij}\pi(i)}\\&&=\sum_{i}{\mathfrak D}(\pi_{ij}\|m_{ij})\pi(i)+ {\mathfrak D}(\pi\|m_0).\label{SBjoint}
\end{eqnarray}
Following Theorem \ref{maincl}, suppose there exist two vectors
\[(\varphi(t,\cdot),\hat{\varphi}(t,\cdot)), \mbox{ for } t\in\{0,1\},
\]
with nonnegative entries, satisfying the Schr\"odinger system
\begin{subequations}\label{eq:DSchroesystem}
\begin{eqnarray}\label{eq:DSchroesystemA}
&&\varphi(0,i)=\sum_j m_{ij}\varphi(1,j),\\
&&\hat{\varphi}(1,j)=\sum_i m_{ij}\hat{\varphi}(0,i)\label{eq:DSchroesystemB},\\&&\varphi(0,i)\cdot\hat{\varphi}(0,i)=\pi(i),\label{eq:DSchroesystemC}\\&&\varphi(1,j)\cdot\hat{\varphi}(1,j)=\pi(j).\label{eq:DSchroesystemD}
\end{eqnarray}
\end{subequations}
Thence, we conclude that
\begin{equation}\label{one-step_opt}\pi^*_{ij}=\frac{\varphi(1,j)}{\varphi(0,i)}m_{ij}
\end{equation}
satisfies both constraints of Problem \ref{P2}. The corresponding measure $\mathfrak P^*\in\mathcal P$ solves Problem \ref{P1}. There is, however, a {\em difficulty}. The assumption on the matrix $G$ in (\ref{posassumption}), which guarantees existence for the Schr\"odinger system, becomes here the fact that $M$ must have all positive elements. The latter property is typically not satisfied
since
$M$ must comply with the topology of the graph. There exists, fortunately, a reasonable condition on $M$ which ensures existence of solutions for (\ref{eq:DSchroesystem}).
\subsection{Existence for the one-step Schr\"odinger system}\label{Existence}
\begin{definition} \cite{FL} A square matrix $A =(a_{ij})$ is called {\em indecomposable} if no permutation matrix $P$ exists such that
\[A=P\left[\begin{array}{cc}A_{11} &0 \\ A_{21} & A_{22}\end{array}\right]P'
\]
where $A_{11}$ and $A_{22}$ are nonvacuous square matrices. $A$ is called {\em fully indecomposable} if there exist no pair of permutation matrices $P$ and $Q$ such that
\begin{equation}\label{structure}A=P\left[\begin{array}{cc}A_{11} &0 \\ A_{21} & A_{22}\end{array}\right]Q
\end{equation}
where $A_{11}$ and $A_{22}$ are nonvacuous square matrices.
\end{definition}
\begin{remark} A square, indecomposable matrix $A$ has a real positive simple eigenvalue equal to its {\em spectral radius} \cite{Varga}. Let $A=(a_{ij})$ be the {\em adjacency matrix} of the graph ${\bf G}=(\mathcal X,\mathcal E)$, namely
\[a_{ij}=\left\{\begin{array}{ll} 1, &(i, j)\in\mathcal E,\\0, & (i,j)\not\in \mathcal E.\end{array}\right.
\]
Then $A$ is indecomposable if and only if ${\bf G}=(\mathcal X,\mathcal E)$ is strongly connected \cite[p.608]{FL}.
\end{remark}
\begin{proposition}\label{PROP} Suppose $M$ is fully indecomposable and $\pi$ has all positive components. Then there exists a solution to (\ref{eq:DSchroesystem}) with $\varphi(0,\cdot)$ and $\varphi(1,\cdot)$ with positive components which is unique in the sense of Remark \ref{uniqueness}.
\end{proposition}
\begin{proof} Let $\varphi_0=\varphi(0,\cdot)$ and $\varphi_1=\varphi(1,\cdot)$. Observe that (\ref{eq:DSchroesystemA})-(\ref{eq:DSchroesystemB})-(\ref{one-step_opt}) admit the matricial form
\begin{subequations}\label{eq:DSchroesystemMatrix}
\begin{eqnarray}\label{eq:DSchroesystemMatrixA}\varphi_0&=&M\varphi_1,\\\label{eq:DSchroesystemMatrixB}\hat{\varphi}_1&=&M'\hat{\varphi}_0,\\
\label{eq:DSchroesystemMatrixC}\Pi^*&=&{\rm Diag}(\varphi_0)^{-1}M{\rm Diag}(\varphi_1).
\end{eqnarray}
\end{subequations}
The proof can now be constructed along the lines of \cite[Theorem 5]{MO}.
\end{proof}
Another interesting question is the following: Given the graph ${\bf G}$, what distributions $\pi$ admit at least one stochastic matrix compatible with the topology of the graph for which they are invariant? Clearly, if all self loops are present ($(i,i)\in\mathcal E,\forall i\in\mathcal X$), any distribution is invariant with respect to the identity matrix which is compatible. Without such a strong assumption, a partial answer is provided by the following result.
\begin{proposition}\label{PROP2}Let $\pi$ be a probability distribution supported on all of $\mathcal X$, i.e. $\pi(i)>0, \forall i\in\mathcal X$. Assume that the adjacency matrix of ${\bf G}=(\mathcal X,\mathcal E)$ is fully indecomposable. Then, there exist stochastic matrices $\Pi$ compatible with the topology of the graph ${\bf G}$ such that
\[\Pi'\pi=\pi.
\]
\end{proposition}
\begin{proof} Take, in Problem \ref{P2}, $M=A=[a_{ij}]$ the adjacency matrix of the graph ${\bf G}$.
By Theorem \ref{maincl} and Proposition \ref{PROP}, Problem \ref{P2} has a solution which admits $\pi$ as invariant distribution.
\end{proof}
One final interesting question is the following: Assume that the adjacency matrix of the graph ${\bf G}$ is fully indecomposable. Let $\mathcal Q_\pi$ be the non empty set of stochastic matrices $Q$ compatible with the topology of ${\bf G}$ and such that $Q'\pi=\pi$. Clearly $\mathcal Q$ is a convex set. Can we characterize the matrix in $\mathcal Q$ which induces a maximum entropy rate measure on the feasible paths? We already know the answer to this question. Indeed, consider the formulation of Problem \ref{P2} with the distribution of the edges of formula (\ref{SBjoint}). Recall that maximizing entropy (rate) is equivalent to minimizing relative entropy (rate) from the uniform. If we take as prior the uniform distribution on the edges, $M$ is just a positive scalar times the adjacency matrix $A$ of the graph. Thus, the solution to Problem \ref{P2} with $A$ as prior transition provides the maximum entropy rate.
\section{Reversibility}\label{REVERSE}
In \cite[Corollary 2]{CGPcooling}, we have shown that, in the case of a reversible prior (Boltzmann density) for a stochastic oscillator, the solution of the continuous countepart of Problem \ref{P1} (minimizing the expected input power) is reversible. We prove next that the same remarkable property holds here in the discrete setting.
\begin{theorem}\label{revers} Let the transition matrix $M$ of the prior measure $\mathfrak M$ be time invariant. Assume that $M$ is reversible with respect to $\mu_0=\mu$, i.e.
\begin{equation}\label{revprior}
{\rm Diag}(\mu)M=M'{\rm Diag}(\mu).
\end{equation}
Then, then the solution $\Pi^*$ of Problem \ref{P2} is also reversible with respect to $\pi$.
\end{theorem}
\begin{proof}
Reversibility of $M$ with respect to $\mu$ is equivalent to the statement that $\Sigma:={\rm Diag}(\mu)M$ is symmetric, as noted. Define the vectors
\[\hat{\psi}_0={\rm Diag}(\mu)^{-1}\hat{\varphi}_0, \quad \psi_0={\rm Diag}(\mu)\varphi_0.
\]
Then, the first two equations in system (\ref{eq:DSchroesystemMatrix}) supplemented with the marginal conditions read
\begin{subequations}\label{eq:DSchroesystemMatrix2}
\begin{eqnarray}\label{eq:DSchroesystemMatrix2A}&&\psi_0=\Sigma\varphi_1,\\\label{eq:DSchroesystemMatrix2B}&&\hat{\varphi}_1=\Sigma'\hat{\psi}_0=\Sigma\hat{\psi}_0,\\
\label{eq:DSchroesystemMatrix2C}&&\psi_0\circ \hat{\psi}_0=\varphi_1\circ \hat{\varphi}_1=\pi,
\end{eqnarray}
\end{subequations}
where the operation $\circ$ in (\ref{eq:DSchroesystemMatrix2C}) represents componentwise between vectors.
Due to the symmetry of $\Sigma$ and the uniqueness in the sense of Remark \ref{uniqueness}, we have that $\hat{\psi}_0=c\varphi_1$ and $\psi_0=\frac{1}{c} \hat{\varphi}_1$ for some positive scalar $c$. It readily follows that
\begin{align*} {\rm Diag}(\pi)\Pi^*
&={\rm Diag}(\pi){\rm Diag}(\varphi_0)^{-1}M{\rm Diag}(\varphi_1)\\
&={\rm Diag}(\hat{\varphi}_0)M{\rm Diag}(\varphi_1)\\
&={\rm Diag}(\hat{\psi}_0){\rm Diag}(\mu)M{\rm Diag}(\varphi_1)\\
&={\rm Diag}(\hat{\psi}_0)M'{\rm Diag}(\mu){\rm Diag}(\varphi_1)\\
&={\rm Diag}(\hat{\psi}_0)M'{\rm Diag}(\varphi_0)^{-1}{\rm Diag}(\psi_0){\rm Diag}(\varphi_1)\\
&={\rm Diag}(\varphi_1)M'{\rm Diag}(\varphi_0)^{-1}{\rm Diag}(\hat{\varphi}_1){\rm Diag}(\varphi_1)\\&=\Pi'{\rm Diag}(\pi).\hspace*{4.2cm}\Box
\end{align*}
\end{proof}
Returning now to the question raised at the end of Subsection \ref{Existence}, in the case when $A$ is symmetric, we can take $\pi_A:=A\mathds{1}$.
\section{Cooling}\label{cooling}
In several advanced applications such as atomic force microscopy and
microresonators,
it is often important to
dampen stochastic excitations (due to vibrations, molecular dynamics, etc.) affecting
experimental apparatuses. This is often accomplished through feedback, steering the system to a desired steady state corresponding to an {\em effective temperature} which is lower than that of the environment (e.g., fluid), see for instance \cite{LMC,DE,BBP,MG,SR,Vin}. Optimal asymptotic and fast cooling of stochastic oscillators was studied in \cite{CGPcooling}. We show next how the results of Section \ref{OSSS} permit to derive corresponding results in the context of discrete space and time considered herein.
\subsection{Optimal fast cooling}\label{FC}
Consider the setting of the previous section where the transition matrix of the prior measure $M$ is time invariant. Let us introduce a {\em Boltzmann distribution} $\pi_T$ on $\mathcal X$
\begin{equation}\label{BOL}
\pi_T(i):=Z(T)^{-1}\exp\left[\frac{-E_i}{kT}\right], \quad Z(T)=\sum_i \exp\left[\frac{-E_i}{kT}\right],
\end{equation}
corresponding to the ``energy function'' $E_x$, for $x\in\mathcal X$,
and let
us assume that $\pi_T$ is invariant for $M=P(T)$, that is,
\[P(T)'\pi_T=\pi_T.
\]
One such class of transition matrices is given by $P(T,Q)=\left(p_{ij}(T,Q)\right)$ with
$$p_{ij}(T,Q)=\left\{\begin{array}{ll} q_{ij}\min\left(\exp\left(\frac{E_i-E_j}{kT}\right),1\right), &i\neq j,\\1-\sum_{l,l\neq i}q_{il}\min\left(\exp\left(\frac{E_i-E_l}{kT}\right),1\right), & i=j.\end{array}\right.,
$$
where $Q=\left(q_{ij}\right)$ is any symmetric transition matrix of an irreducible chain compatible with the topology of ${\bf G}$\footnote{As is well-known, this (Metropolis) chain is actually reversible with respect to the Boltzmann distribution.}. Let $T_{{\rm eff}}<T$ be a desired lower temperature. Given any prior, such as the Ruelle-Bowen measure \cite{CGPT1} or the invariant path space measure $\mathfrak P(T)$ with transition $P(T)$ and marginals $\pi_T$ and a sufficiently long time interval $[0,N]$, we can use a standard Schr\"odinger bridge (see Section \ref{GBProblems}) to steer optimally the network flow from any initial distribution $\nu_0$ to $\nu_N=\pi_{T_{{\rm eff}}}$ at time $N$. At time $N$, however, we need to change the transition mechanism to keep the chain in the steady state $\pi_{T_{{\rm eff}}}$. This is accomplished in the next subsection using the results of Section \ref{OSSS}.
\subsection{Optimal asymptotic cooling}\label{AC}
Consider once again the case where the prior transition $M=P(T)$ remains constant over time and has the Boltzmann distribution (\ref{BOL}) as initial, invariant measure. Consider the equivalent Problems \ref{P1} and \ref{P2}.
\begin{theorem}Assume that $P(T)$ is fully indecomposable. Then, the solution to Problem \ref{P2} is given by
\begin{equation}\label{OPTNEWTRANS}
\Pi^*={\rm Diag}(\varphi_0)^{-1}P(T){\rm Diag}(\varphi_1).
\end{equation}
where
\begin{subequations}\label{eq:DSchroesystemMatrix3}
\begin{eqnarray}\label{eq:DSchroesystemMatrix3A}\varphi_0&=&P(T)\varphi_1,\\\label{eq:DSchroesystemMatrix3B}\hat{\varphi}_1&=&P(T)'\hat{\varphi}_0,\\
\label{eq:DSchroesystemMatrix3C}\varphi_0\circ\hat{\varphi}_0&=&\varphi_1\circ\hat{\varphi}_1=\pi_{T_{{\rm eff}}}.
\end{eqnarray}
\end{subequations}
As earlier, $\circ$ denotes componentwise multiplication of vectors. Moreover, if $P(T)$ is reversible with respect to $\pi_T$, so is $\Pi^*$ with respect to $\pi_{T_{{\rm eff}}}$.
\end{theorem}
\begin{proof}
The result follows from Theorem \ref{maincl}, Proposition \ref{PROP} and Theorem \ref{revers}.
\end{proof}
\section{Conclusion}
The key point of this paper has been to explore analogues, in the setting of discrete space and time, for
Markovian evolutions that match specified marginals and echo well-known results in the continuous time setting
for Sch\"odinger bridges. Specifically, a key result is to show reversibility of the transition rates under the assumption that this is a property of the prior. We consider both, transitioning between marginals over a specified time window as well as the problem to secure a stationary distribution. Both cases are accomplished by suitably adjusting the transition rates from a given prior, while keeping the distance of the new law closest to that corresponding to the prior in the relative entropy sense. Application of the results in the context regulating the flow of resources over a network is contemplated. A notion of temperature when distributions are expressed as Boltzmann distributions with respect to an energy function is brought up, and the problem to transition between distributions corresponding to different temperatures considered, in analogy to the problem of cooling in continuous time and space.
|
\section{Introduction}
At extreme energy densities nuclear matter undergoes a phase transition from a state characterized by confined hadrons to a state in which quarks and gluons become the relevant degrees of freedom. Ultrarelativistic heavy-ion collision experiments at the Relativistic Heavy Ion Collider in New York and the Large Hadron Collider at CERN have now collected a wealth of data concerning the properties of the high energy density phase of nuclear matter, which is called the quark-gluon plasma (QGP) \cite{Averbeck:2015jja}. One of the key observables used in measuring the properties of the QGP is the suppression of heavy quark-antiquark bound states such as bottomonium. The suppression observed by experimentalists gives a measurement of the in-medium breakup rate of bottomonium states and allows one to constrain key plasma observables such as the initial central temperature of the QGP generated in heavy-ion collisions, its shear viscosity to entropy density ratio, and the differential rate at which the QGP expands in non-central collisions (see e.g. \cite{Acharya:2020kls,ATLAS5TeV,Sirunyan:2018nsz,Ye:2017vuw}).
Fundamentally, the computation of the survival probability of a given bottomonium state can be cast into the framework of open quantum systems (OQS) in which there is a probe (bottomonium states) and medium (light quarks and gluons). Within the OQS framework, in order to describe the in-medium evolution of bottomonium states one must trace over the medium degrees of freedom and obtain evolution equations for the reduced density matrix of the system \cite{Akamatsu:2011se,Akamatsu:2012vt,Akamatsu:2014qsa,Katz:2015qja,Brambilla:2016wgg,Kajimoto:2017rel,Brambilla:2017zei,Blaizot:2017ypk,Blaizot:2018oev,Yao:2018nmy,Alund:2020ctu,Islam:2020gdv,Islam:2020bnp,Brambilla:2020qwo}. In the limit that the medium relaxation time scale and the intrinsic time scale of the probe are much smaller than the probe relaxation time scale, the resulting dynamical equation for the reduced density matrix can be cast into a so-called Lindblad form \cite{lindblad,gorini}. A key outcome of such calculations is that the effective heavy-quark potential possesses an imaginary part which can be related to the total in-medium decay width of the states. This imaginary part has been determined using direct quantum field theoretic or effective field theory calculations \cite{Laine:2006ns,Brambilla:2016wgg,Brambilla:2017zei}.
There have been computations of the imaginary part of the heavy-quark potential based on high-temperature quantum chromodynamics (QCD) calculations in the hard thermal loop (HTL) limit \cite{Laine:2006ns,Dumitru:2007hy,Brambilla:2008cx,Burnier:2009yu,Dumitru:2009fy,Dumitru:2009ni,Margotta:2011ta,Guo:2018vwy}, using effective field theory (pNRQCD) \cite{PhysRevD.21.203,Lucha:1991vn,Brambilla:2004jw,Brambilla:2010xn}, finite-temperature lattice QCD \cite{Rothkopf:2009pk,Rothkopf:2011db,Burnier:2012az,Burnier:2013nla,Burnier:2015nsa,Burnier:2015tda,Burnier:2016mxc,Burnier:2016kqm,Bala:2019cqu,Bala:2020tdt}, and real-time classical-statistical solutions of Yang-Mills theory in classical thermal equilibrium
\cite{Laine:2007qy,Lehmann:2020fjt}. In this work we build upon the studies presented in Ref.~\cite{Laine:2007qy} and present findings that are complementary to Ref.~\cite{Lehmann:2020fjt}. In Ref.~\cite{Laine:2007qy} the authors presented first results for the imaginary part of the heavy-quark potential using classical-statistical Yang-Mills simulations on spatially 3D lattices of size $12^3$ and $16^3$. In this paper we extend these results to larger lattices up to $252^3$
and consider SU(2) and SU(3) gauge theories. Additionally, the results of Ref.~\cite{Laine:2007qy} were presented only for a few spatial points $r$ in a table. Due to the use of rather large lattice sizes, we can now compute the imaginary part of the heavy quark potential at larger values of $r/a$ and reconstruct the functional form of the imaginary part of the heavy-quark potential for a much wider range of distances.
This allows us to make more precise comparisons between our lattice-extracted imaginary part and (a) results obtained in the continuum limit using hard thermal loops \cite{Laine:2006ns} and (b) results obtained using the lattice-regularized hard-classical-loop (HCL) theory \cite{Laine:2007qy}.
Herein we will present results for the imaginary part of the heavy-quark potential obtained using classical Yang Mills (CYM) simulations of a thermalized gluonic plasma. The use of CYM simulations is motivated by the fact that in situations where (a) gluonic occupation numbers are large, such as in thermal equilibrium for sufficiently low momenta or in the initial stages of heavy-ion collisions, and (b) the gauge coupling is weak $g^2 \ll 1$, vacuum contributions to observables are suppressed by powers of the gauge coupling. As a consequence, neglecting vacuum contributions in such cases often leads to a good approximation to the full quantum dynamics. \changeflag{However, one should note that in the real-time classical lattice theory rotational invariance is broken and the short and long distance physics do not decouple \cite{Bodeker:1995pp, Arnold:1997yb}. As a consequence, the results that we obtain in our classical simulations may differ from those in thermal equilibrium.}
In order to extract information about thermal systems using CYM, one must prepare the thermalized field configurations. Historically, this is done by preparing 3+1D configurations using Monte-Carlo techniques \cite{Grigoriev:1988bd,Ambjorn:1990pu,Ambjorn:1997jz,Moore:1996qs,Moore:1999fs,Hart:2000ha}. Herein we follow a simpler strategy and initialize the system close to thermal equilibrium using momentum-space initialization and let the fields thermalize in real time before extracting observables.
The advantage of this procedure is that we can run simulations on very large lattices with small lattice spacings $a$ at moderate computational cost. In practice, however, this means that we have to extract the temperature $g^2 T$ numerically in the thermalized system from correlation functions. We find that this only introduces a small uncertainty for $g^2 T$ of the order of a few percent, which justifies this approach.
One additional complication that arises when dealing with CYM treatments is that they do not possess a finite ultraviolet limit due to the Rayleigh-Jeans divergence \changeflag{and the theory is non-renormalizable} \cite{Kajantie:1993ag,Ambjorn:1995xm,Arnold:1996dy,Moore:1999fs,Berges:2013lsa,Epelbaum:2014yja}. For this reason, it is important to identify a suitable manner in which to scale results in order to extract relevant information. We will use the Debye mass $m_D$ computed in the hard-classical-loops framework to demonstrate that, when plotted as a function of $m_D r$, the imaginary part of the heavy-quark potential \changeflag{is only mildly sensitive to the lattice spacing, or more generally to the simulation parameter $\beta \propto 1 / (g^2Ta)$, at small distances $m_D r \lesssim 1$.}
Additionally we find that, when presented in this manner, results obtained using CYM simulations agree well with the semi-analytic hard classical loop result at small quark-antiquark separations.
Our study also suggests that the latter approaches a finite form in the large-$\beta$ limit for all separations.
We present our results for a wide range of temperatures, lattice spacings, and lattice volumes, which map to values of $\beta$ in the range $16 \rightarrow 300$.
This work sets the stage for extracting the out-of-equilibrium imaginary part of the heavy-quark potential in expanding Yang Mills plasmas \cite{Berges:2013eia,Berges:2013fga}.
\changeflag{Classical Yang-Mills simulations are applicable in that case, allowing one to extrapolate to small lattice spacings, because the dynamics of the highly occupied plasmas are governed by hard scales that can be set below the lattice momentum cutoff.}
Such non-equilibrium extractions are necessary since it is currently unknown how to analytically extract the heavy-quark potential in anisotropic plasmas due the presence of a non-Abelian plasma instability called the chromo-Weibel instability \cite{Mrowczynski:1996vh,Romatschke:2003ms,Arnold:2003rq,Burnier:2009yu,Nopoush:2017zbu}.
In particular, recent studies have shown that particularly in highly anisotropic systems nonperturbative effects beyond hard loop calculations are crucial \cite{Boguslavski:2019fsb,Boguslavski:2021buh}.
The structure of our paper is as follows. In Sec.~\ref{sec:theory} we provide the theoretical background necessary for the computation of the imaginary part of the heavy-quark potential. In Sec.~\ref{sec:results} we present the results from lattice simulations and the comparison \changeflag{to perturbative calculations.} We present concluding remarks and outlook for future works in Sec.~\ref{sec:conclusions}.
\section{Theory and numerical setup}
\label{sec:theory}
We consider pure SU($N_c$) gauge theory with the Yang-Mills classical action
\begin{align}
S[A] = -\frac{1}{2}\int d^4 x\, \mrm{Tr} \left(F^{\mu\nu} F_{\mu\nu}\right),
\end{align}
with Einstein sum convention for repeated Lorentz indices $\mu,\nu = 0,\dots,3$.
The field strength tensor is given by $F_{\mu\nu} = \partial_\mu A_\nu - \partial_\nu A_\mu -ig [A_\mu,A_\nu]$ with gauge coupling $g$, gauge field $A_\mu(x)$ and commutator $[,]$. Unless stated otherwise, we will use $N_c=3$.
\subsection{Lattice discretization and equations of motion}
We use a standard real-time lattice discretization approach where fields are discretized on cubic lattices with $N^3$ sites and lattice spacing $a$ (see, e.g., Refs.~\cite{Boguslavski:2018beu,Berges:2013fga} and references therein for more details). In this real-time approach, spatial gauge fields are replaced by gauge links $U_j(t,\mathbf x) \approx \exp\left( i g a A_j(t,\mathbf x) \right)$ at discrete coordinates $x_k = n_k a$ for $n_k = 0, \ldots, N-1$, while temporal gauge with $A_0 = 0$, and thus $U_0 = \mathbb{1}$ is used. The classical equations of motion are written in a gauge-covariant manner as
\begin{align}
\label{eq:classEOM}
U_j(t+a_t/2,\mathbf x) &= \exp\left( i a_t a g E^j(t,\mathbf x) \right)U_j(t-a_t/2,\mathbf x) \, , \nonumber \\
gE^j(t+a_t,\mathbf x) &= gE^j(t,\mathbf x)
- \frac{a_t}{a^3} \sum_{j \neq i} \left[ U_{ij}(t-a_t/2,\mathbf x) + U_{i(-j)}(t-a_t/2,\mathbf x) \right]_{\mathrm{ah}} \, ,
\end{align}
and are solved alternately in a leapfrog scheme~\cite{Krasnitz:1995xi}. The plaquette is defined as $U_{ij}(\mathbf x) = U_i(\mathbf x)U_j(\mathbf x + \mathbf{a}_i)U_i^\dagger(\mathbf x + \mathbf{a}_j)U_i^\dagger(\mathbf x)$, and analogously for $U_{i(-j)}$, with vectors $\mathbf{a}_{i/j}$ of length $a$ in spatial
directions $i,j = 1,2,3$. $[V]_{\mathrm{ah}} \equiv -i(V-V^\dagger - \mrm{Tr}(V-V^\dagger)/N_c)/2$ defines the anti-Hermitian traceless part of the matrix $V$. The time step is taken to be small, typically $a_t/a = 0.01$ in this work, to reduce temporal lattice artifacts. The classical equations \eqref{eq:classEOM} guarantee that the degree of Gauss law violation is conserved at every time step
\begin{align}
D_j E^j(t,\mathbf x) &\equiv \frac{1}{a}\sum_j\left( E^j(t,\mathbf x) - U_j^\dagger(t-a_t/2,\mathbf x - \mathbf{a}_j)E^j(t,\mathbf x - \mathbf{a}_j)U_j(t-a_t/2,\mathbf x - \mathbf{a}_j) \right) \nonumber \\
&= \text{constant} \,.
\end{align}
Moreover, this lattice discretization guarantees gauge invariance of observables like Wilson loops that are supposed to be gauge invariant.
\subsection{Computation of observables}
Classical observables are computed by averaging over different configurations evolved independently
\begin{align}
\langle O \rangle(t) \equiv \frac{1}{n}\sum_{j=1}^n O_j(t)\,,
\end{align}
and their uncertainity corresponds accordingly to the uncertainty of the mean.
\subsection{Imaginary part of the classical potential}
\label{sec:theory_potential}
We are interested in extracting the imaginary part of the classical potential $V_{\mathrm{cl}}(t,r)$ with $r \equiv |\mathbf x|$. Following Refs.~\cite{Laine:2006ns,Laine:2007qy}, it can be calculated using
\begin{align}
i\partial_t C_{\mathrm{cl}}(t,r) = V_{\mathrm{cl}}(t,r) C_{\mathrm{cl}}(t,r) \, ,
\end{align}
as the asymptotic temporal slope of $\log[C_{\mathrm{cl}}(t,r)]$. The classical thermal Wilson loop $C_{\mathrm{cl}}(t,r)$ is defined as
\begin{align}
C_{\mathrm{cl}}(t,r) \equiv \frac{1}{N_c}\,\mrm{Tr} \left\langle W[(t_0,\mathbf x);(t,\mathbf x)]\, W[(t,\mathbf x);(t,\mathbf 0)]\, W[(t,\mathbf 0);(t_0,\mathbf 0)]\, W[(t_0,\mathbf 0);(t_0,\mathbf x)] \right\rangle,
\end{align}
with temporal Wilson lines $W[(t_0,\mathbf x);(t,\mathbf x)] = \mathbb{1}$ and spatial Wilson lines $W[(t,\mathbf 0);(t,\mathbf x)] = U_j(t,\mathbf 0)U_j(t,\mathbf{a}_j)U_j(t,2\,\mathbf{a}_j) \cdots U_j(t,\mathbf x)$ for $\mathbf x = \mathbf{\hat a}_j\,r$ and $\mathbf{\hat a}_j = \mathbf{a}_j/a$ being a spatial unit vector. Since the classical thermal state is homogeneous, the Wilson loop is additionally averaged over all lattice points by averaging over the reference coordinates $\mathbf 0$.
In order to extract the imaginary part of $V_{\mathrm{cl}}(t,r)$ we compute the time-dependence of $C_{\mathrm{cl}}(t,r)$. Due to the imaginary part of the in-medium heavy quark potential, this quantity will decay exponentially at late times with the rate of exponential decay set by the imaginary part of $V_{\mathrm{cl}}(t,r)$. As a result, $V_{\mathrm{cl}}(t,r)$ can be computed using the logarithmic slope of this decay as mentioned above. We note that this is formally applicable only in the large-$t$ limit, but we find that an exponential decay is established rather quickly, particularly when using large lattice sizes. We then define the imaginary part of the static classical potential as the late-time limit
\begin{align}
{\rm Im}[V_{\rm cl}(r)] \equiv \lim_{t \to \infty} {\rm Im}[V_{\rm cl}(t,r)]\,.
\end{align}
In practice, this agrees with the logarithmic slope extracted from $C_{\mathrm{cl}}(t,r)$.
\subsection{Construction of quasi-thermalized and fully-thermalized classical states}
\label{sec:thermalstate}
In order to associate our measurements of $V_{\mathrm{cl}}(t,r)$ with a specific temperature $T$, it is necessary to have a method for self-consistently determining the temperature of the CYM fields. One can construct thermalized CYM configurations using 3+1D Monte-Carlo techniques \cite{Grigoriev:1988bd,Ambjorn:1990pu,Ambjorn:1997jz,Moore:1996qs,Moore:1999fs,Hart:2000ha}. However, as mentioned in the introduction, we use a simpler technique, which amounts to initializing the fields in a quasi-thermal configuration in momentum-space, as will be explained around \eqref{eq:EE_noGauss_init}, and then allowing them to self-thermalize dynamically. On large lattices this method is quite efficient and one finds that the fields self-thermalize quickly. One complication is that the resulting equilibrated temperature $T$ is different than the quasi-thermal initial temperature $T_0$ used to initialize the fields in momentum-space. As a result, one must have a method to extract the time-dependent temperature in order to (a) determine when field thermalization is functionally complete and (b) extract the final thermalized temperature of the CYM fields. Once the system is fully thermalized (within an acceptable uncertainty), one can then proceed with the measurement of $V_{\mathrm{cl}}(t,r)$.
For this purpose, we extract chromo-electric field correlation functions in Fourier space. In practice, the chromo-electric fields are Fourier transformed as
\begin{align}
E^j(t,\mathbf p) = a^3\sum_{\mathbf x} e^{-2\pi i \sum_k n_k m_k/N} E^j(t,\mathbf x) \, ,
\end{align}
and projected onto normalized transverse vectors $v_j^\lambda(\mathbf p)$ with two polarizations $\lambda = 1,2$ transverse to the longitudinal polarization vector $v_j^{3}(\mathbf p) = p_j/p$,
providing $E_\lambda(t,\mathbf p)$. Here we used $x_k = n_k\, a$ with $k=1,2,3$, the lattice momentum definition $p_k\, a = -i(1-
\exp(-2\pi i m_k/N))$, $p^2 = \sum_k |p_k|^2$, and $n_k, m_k = 1,\dots, N$. The correlation functions are then computed as
\begin{align}
\langle EE \rangle_{T}(t, p) &= \frac{1}{2(Na)^3(N_c^2-1)}\sum_{\lambda = 1,2} \left\langle |E_{\lambda}(t,\mathbf p)|^2 \right\rangle \nonumber \\
\langle EE \rangle_{L}(t, p) &= \frac{1}{(Na)^3(N_c^2-1)} \left\langle |E_{\lambda=3}(t,\mathbf p)|^2 \right\rangle\,,
\end{align}
and similarly for $A$ fields. This expression implies averaging over the direction of $\mathbf p$ due to the system's approximate isotropy for not too large momenta and we will consider $\langle EE \rangle_{T/L}$ as functions of $p$.\footnote{
Although the anisotropy of the cubic lattice becomes more prominent close to the lattice cutoff at $|p_k| = 2/a$, we will mostly use these correlators as means to extract parameters by fitting to them over a large range of momenta, such that the anisotropic behavior of the cube does not play a significant role here.
}
Note that these correlation functions are, in general, gauge dependent. We get rid of the residual gauge freedom by gauge-fixing to the Coulomb-type gauge $\partial^j A_j(t,\mathbf x) = 0$ to almost machine precision right before computing the correlators.
As mentioned above, while all measurements will be performed in classical thermal equilibrium, for practical purposes we start our simulations in a state very close to thermal equilibrium and let the system dynamically thermalize. We use two-point correlation functions $\langle EE \rangle$ in Fourier space to check to what extent the system has thermalized. Our procedure is as follows.
We set the fields in Fourier space at initial time $t = 0$ such that they satisfy
\begin{align}
\label{eq:EE_noGauss_init}
g^2\langle EE \rangle_T = g^2T_0\,, \qquad g^2\langle EE \rangle_L = 0\,, \nonumber \\ g^2\langle AA \rangle_T = \frac{g^2T_0}{p^2}\,, \qquad g^2\langle AA \rangle_L = 0\,.
\end{align}
These initial conditions can be implemented efficiently by setting (see, e.g., \cite{Smit:2002yg,Berges:2013fga,Boguslavski:2018beu})
\begin{align}
\label{eq:AE_initial}
gE^j_a(t{=}0,\mathbf p) &= \sqrt{g^2T_0}\,\sum_{\lambda = 1,2} v_j^{\lambda}(\mathbf p)\,\beta_a^{\lambda}(\mathbf p) \nonumber \\
gA_j^a(t{=}0,\mathbf p) &= \sqrt{\frac{g^2 T_0}{p^2}}\,\sum_{\lambda = 1,2} v_j^{\lambda}(\mathbf p)\,\alpha_a^{\lambda}(\mathbf p)\,,
\end{align}
with the adjoint color index $a = 1, \dots, N_c^2-1$, and with Gaussian distributed complex random numbers satisfying $\langle (\alpha_a^{\lambda}(\mathbf p))^* \alpha_b^{\lambda'}(\mathbf p') \rangle = \delta_{ab}\delta_{\lambda \lambda'}\delta_{\mathbf p \mathbf p'} N^3a^3$, $\langle (\beta_a^{\lambda}(\mathbf p))^* \beta_b^{\lambda'}(\mathbf p') \rangle = \delta_{ab}\delta_{\lambda \lambda'}\delta_{\mathbf p \mathbf p'} N^3a^3$, while all other correlations vanish.
Subsequently, we restore the Gauss law $D_j E^j(x) = 0$ with covariant derivative $D_\mu$ to machine precision using the algorithm in Ref.~\cite{Moore:1996qs}. This algorithm changes the chromo-electric field correlators mostly at low momenta
and, for instance, makes $\langle EE \rangle_L$ finite. Then we let the system evolve according to the Yang-Mills classical equations of motion \eqref{eq:classEOM}. Classical thermal equilibrium is reached when the $\langle EE \rangle_{T/L}(t,p)$ correlation functions do not visibly change over an extended amount of time.
\begin{figure}[t!]
\centerline{
\includegraphics[width=0.48\linewidth]{./eet_vart.pdf}
$\;\;\;$
\includegraphics[width=0.48\linewidth]{./T_mD_extract_n0_0375_N_64_a_1_np_224_Jan1.pdf}
}
\caption{
(Left:) Example of thermalization for $g^2 T_0 = 3/8$ and $a=1$. The transverse chromo-electric field correlation is shown at different times of the simulated gauge dynamics.
The inset shows a zoomed region of the plot. The dashed orange line corresponds to the field initialization \eqref{eq:EE_noGauss_init} with $g^2 T_0$ before the Gauss law is restored.
(Right:) Extracting $T$ and $\hat{m}_D$ by fitting \eqref{eq:extractT} to transverse and longitudinal chromo-electric field correlations. The fit functions are given by the black dashed and continuous purple lines. The red dashed line corresponds to the value $g^2 T_{\text{UV}}$ at the hardest momentum, the shaded error band shows the systematic error in the estimation of the temperature $[T-\Delta T, T+\Delta T]$ with $\Delta T = T_{\text{UV}} - T$ .
}
\label{plot:T_md_convergence}
\end{figure}
An example of the results obtained using this thermalization algorithm is shown in the left panel of Fig.~\ref{plot:T_md_convergence} for $\langle EE \rangle_{T}$ and simulation parameters $g^2 T_0 = 0.375$, $N=64$, $a=1$, averaged over $N_{\text{config}} = 224$ configurations. The dashed orange horizontal line shows the initial correlator prior to Gauss law restoration, that equals $g^2 T_0$ for all momenta. After the restoration, the transverse correlator shows some momentum dependence and is given by the blue dashed line at $t/a = 0$. In the subsequent evolution, the system thermalizes rather quickly. Already for times $t/a \gtrsim 30$, the correlation function is almost stationary, for later times $t/a \gtrsim 60$, the deviation between curves at different times lies within the uncertainty of the data, as can be seen in the inset. At that point, the system can be considered approximately thermal.
As shown in Ref.~\cite{KM2012} and recently used in \cite{Matsuda:2020qwx}, the correlation functions $\langle EE \rangle_{T/L}(t,p)$ encode the temperature and the Debye mass of the classical state
\begin{align}
\label{eq:extractT}
g^2\langle EE \rangle_T \approx g^2T\,, \qquad g^2\langle EE \rangle_L \approx g^2T\,\frac{\hat{m}_D^2}{p^2+\hat{m}_D^2}\,.
\end{align}
These relations are not exact but become more reliable at larger momenta $p \gg m_D$. We can use them to extract the temperature $g^2 T$ and an estimate for the Debye (screening) mass $\hat{m}_D$.
A systematic error (for $g^2 T$) is estimated as the deviation of the correlators from these simple functional forms, as will be explained shortly.
The extraction of these parameters
from the transverse and longitudinal chromo-electric field correlations is demonstrated in the right panel of Fig.~\ref{plot:T_md_convergence} for the same parameters as in the left panel at the time $t/a = 140$. One can see that the thermalized correlations are lower than the initial value $g^2 T_0$ for the considered parameters and that they indeed agree well with the fit functions $g^2\langle EE \rangle_T^{\text{fit}} = g^2\hat{T}_{E_tE_t}$ and $g^2\langle EE \rangle_L^{\text{fit}} = g^2\hat{T}_{E_lE_l}/(1+p^2/\hat{m}^2_D)$, as shown by the black dashed and purple continuous lines. Remarkably, these independent fits lead to almost the same value of the temperature and we set
\begin{align}
g^2T = g^2\hat{T}_{E_tE_t}\,.
\end{align}
Since the transverse correlator $\langle EE \rangle_T$ still grows slowly with $p$, we extract the value $g^2 T_{\text{UV}}$ at the largest momentum to estimate the error as $\Delta T = T_{\text{UV}} - T$. This systematic error is shown by the shaded band in the figure. We find that the uncertainty in the extraction of the temperature is generally quite low and becomes smaller with decreasing lattice spacing. For the coarse lattice spacing $a=1$ used in Fig.~\ref{plot:T_md_convergence}, the relative error is only $\Delta T/T < 3\%$.
\begin{figure}[t]
\centerline{
\includegraphics[width=0.48\linewidth]{./TmD_n0.pdf}
$\;\;\;$
\includegraphics[width=0.48\linewidth]{./TmD_n0_a01.pdf}
}
\caption{Extracted values of $g^2\hat T_{E_tE_t}$, $g^2\hat T_{E_lE_l}$ and $\hat{m}_D$ for different values of $g^2 T_0$ for $a = 1$ (left) and $a = 0.1$ (right). All parameters are shown in lattice units.}
\label{plot:Tmd_T0}
\end{figure}
The extracted temperature and mass values, $g^2\hat T_{E_tE_t}$, $g^2\hat T_{E_lE_l}$ and $\hat{m}_D$, are shown in Fig.~\ref{plot:Tmd_T0} as functions of $g^2 T_0$ for the larger lattice spacing $a = 1$ (left) and a fine lattice spacing $a = 0.1$ (right). One observes that both $g^2\hat T_{E_tE_t}$ and $g^2\hat T_{E_lE_l}$, visible as blue and red points, respectively, agree well in both cases. While for the small lattice spacing, $g^2 T_0$ and the extracted $g^2 T$ are almost identical, one observes deviations that grow with temperature for the coarser lattice. Since the lattice simulations depend on the combination $\beta = 2N_c / (g^2Ta)$, as will be explained in Sec.~\ref{sec:coupling_scaling}, we can summarize that deviations increase with decreasing $\beta$, while for growing $\beta$ values, both $\Delta T/T$ and $|T_0-T|/T$ become smaller, enabling a better control of the value of the temperature.
\subsection{The HTL, HCL, and lattice Debye masses}
We will also need to compute the Debye mass in order to compare results obtained using different analytic and numerical approaches. The definition of the Debye mass
depends on the method being used.
In this paper we consider three methods for defining the Debye mass: (1) continuum hard-thermal-loops calculations, (2) lattice-discretized hard classical loops calculations, and (3) direct lattice measurement using chromo-electric field correlators. The definitions in these three cases are:
\begin{enumerate}
\item
The well-known continuum hard thermal loop (HTL) expression is
\begin{align}
\label{eq:HTLmass}
\left(m_{D}^{\text{HTL}}\right)^2 = 4 g^2 N_c \int \frac{d^3 p }{(2\pi)^3}\,\frac{f_{\text{BE}}(p)}{p} = \frac{N_c g^2 T^2}{3}\,,
\end{align}
with the Bose-Einstein distribution $f_{\text{BE}}(p) = (\exp(p/T)-1)^{-1}$.
\item
In classical thermal equilibrium one has $f_{\text{BE}}(p) \rightarrow T/p$, and the discretization on an anisotropic cubic lattice with a finite lattice spacing leads to the following hard classical loop (HCL) expression
\begin{align}
\label{eq:HCLmass}
\left(m_{D}^{\text{HCL}}\right)^2 = \frac{g^2T N_c}{2a}\left( \frac{3\Sigma}{2\pi}-1 \right),
\end{align}
where $\Sigma \approx 3.1759$ is a factor that results from the anisotropy of the lattice \cite{Moore:1999fs}.
Strictly speaking, this expression is the leading-order perturbative result in the infinite volume limit. Since we employ large lattice sizes, the latter effect can be neglected. Moreover, finite volume effects are expected to be less important than sub-leading perturbative effects that typically contribute to \eqref{eq:HTLmass} and \eqref{eq:HCLmass} at low momenta $p \lesssim m_D$. We will use \eqref{eq:HCLmass} in our figures when we plot the CYM lattice results as a function of $m_D r$.
\item
Our final possibility is to measure the Debye mass $\hat{m}_D$ directly from our thermalized CYM configurations using chromo-electric field correlators, which is explained in the previous subsection. Its temperature, and thus $\beta$ dependence is shown in Fig.~\ref{plot:Tmd_T0} for different lattice spacings. One generally finds that it grows with temperature. More precisely, we find an approximately linear connection to $m_{D}^{\text{HCL}}$ with
\begin{align}
\label{eq:mdhat_md}
\hat m_D \approx 1.42\; m_{D}^{\text{HCL}}
\end{align}
for $92 \leq \beta \leq 300$, while at lower $\beta$ deviations from a linear connection become more sizable.
\end{enumerate}
\subsection{Parameters and scaling properties with the coupling}
\label{sec:coupling_scaling}
The coupling $g$ can be scaled out of the dynamical equations of motion \eqref{eq:classEOM} by rescaling all field amplitudes as $gA \mapsto A$ and $gE \mapsto E$. As visible in \eqref{eq:extractT}, the coupling then only enters in the combination $g^2 T$. This dimensionful scale can then be used to make all quantities dimensionless by rescaling them with appropriate powers. In the rescaled version, one then has $\langle EE \rangle_T \mapsto 1$. Note that this rescaling implies that lattice simulations only depend on the lattice size $N^3$ and on the lattice spacings $g^2Ta \equiv 2N_c/\beta$ and $g^2Ta_t$. Since we use $a_t \ll a$, temporal artefacts that may result from a dependence on $g^2Ta_t$ are suppressed, and the simulations mainly depend on
\begin{align}
\beta = \frac{2N_c}{g^2Ta}\,.
\end{align}
If not stated otherwise, we will write our values for $a$, $g^2 T_0$, the extracted temperature $g^2 T$ and all dimensionful variables in lattice units. We will also provide the corresponding $\beta$ values.
\section{Results}
\label{sec:results}
In this section we present the non-perturbative results from classical-statistical simulations of real-time Yang-Mills theory on cubic lattices with $N^3$ sites and the spacing length of $a$. As detailed in Sec.~\ref{sec:theory}, the fields are initialized with initial temperature $T_0$, they thermalize by solving classical equations of motion, and the actual temperature $T$ is extracted with an error estimate from correlation functions of chromo-electric fields.
We present the values of the imaginary part of the classical potential ${\rm Im}[V_{\rm cl}(r)]$ for $SU(3)$ extracted from the evolution of temporal Wilson loops \cite{Laine:2006ns, Laine:2007qy}. For comparison, we start this section by recalling
the corresponding analytic results from {dimensionally regularized HTL} and the expressions from {lattice-regularized HCL} calculations of ${\rm Im}[V_{\rm cl}(r)]$ to the second-order. We compute the HCL potential numerically for a wide range of $\beta$ values and fit the numerical HCL results to the functional form of a suitable analytically available approximation with only two parameters. The fit allows us to extract the $\beta$ dependence of the parameters of the HCL potential, including an estimate for the large-$\beta$ limit with $\beta \to \infty$.
We then discuss our numerical results from classical Yang-Mills simulations for a wide range of $\beta$ values, and compare them with previously published data of ${\rm Im}[V_{\rm cl}(r)]$ obtained from lattice calculations, with simulations in $SU(2)$ theory and with HCL results. Our main simulation results are also summarized in tables in App.~\ref{app:tables}.
\subsection{Hard Thermal Loop result}
\label{sec:HTL_V}
In Ref.~\cite{Laine:2006ns} the authors derived an expression for the imaginary part of the heavy-quark potential to leading-order in the strong coupling constant using the continuum hard thermal loop framework and dimensional regularization. Their final result could be expressed compactly as
\begin{equation}
{\rm Im}[V^{(2)}(r)] = - \frac{C_F g^2 T}{4\pi} \phi\left(m_{D}^{\text{HTL}}\, r\right) \, .
\label{htlimv}
\end{equation}
where $C_F = (N_c^2 -1)/2N_c\,$,
\begin{equation}
\phi\left( x \right) \equiv 2 \int_0^{\infty} {\rm d}z \frac{z}{\left(1+z^2 \right)^2}\left[ 1- \frac{\sin\left(zx\right)}{zx}\right] ,
\label{eq:phidef}
\end{equation}
and $m_{D}^{\text{HTL}}$ given by the continuum hard thermal loop Deybe mass \eqref{eq:HTLmass}. Note that in the large-$r$ limit one has $\lim_{r \rightarrow \infty} \phi(r) = 1$. As a result the asymptotic value of the imaginary part in this case is $-C_F g^2 T/4\pi$.
\begin{figure}[t]
\centerline{
\includegraphics[width=0.48\linewidth]{./HCL_aVar_Separate.pdf}
$\;\;\;$
\includegraphics[width=0.48\linewidth]{./HCL_aVar_DivByAinf.pdf}
}
\caption{(Left) Large-time limit of ${\rm Im}[V^{(2)}_{\rm cl}(r)]/g^2T$ for $g^2T = 0.44$ and $a \in \{1, 0.5, 0.25, 0.2, 0.1 \}$ corresponding to $\beta \in \{13.6, 27.3, 54.6, 68.2, 136.4\}$, calculated using the infinite-volume hard classical loop (HCL) approximation in \eqref{hclimv}. These numerical results are represented by points while fits using the function \eqref{Aphi} are shown as lines.
(Right) Same as in the left panel, but divided by the asymptotic value $A_\infty \equiv \lim_{r \rightarrow \infty} {\rm Im}[V^{(2)}_{\rm cl}(r)]/g^2T$.}
\label{plot:hcl_1}
\end{figure}
\subsection{Hard Classical Loop results}
\label{sec:HCL_results}
The result for the imaginary part of the classical potential in the infinite volume and infinite time limit ($N \rightarrow \infty,\ t \rightarrow \infty$) from second-order perturbation theory regularized on a cubic lattice (HCL: Hard Classical Loop) of size $\left(aN\right)^3$ is calculated in Ref.~\cite{Laine:2007qy} and is given by
\begin{equation}
{\rm Im}[V^{(2)}_{\rm cl}(r)]/g^2T = -\frac{\pi C_F N_c^2}{\beta} \int_0^1 {\rm d}^3{\mathbf{x}}\frac{1-\cos\left( \pi x_3r/a \right)}{ (\tilde{\mathbf{x}}^2 + N_c^2\Sigma/\pi \beta )^2} \int_{-1}^1 {\rm d}^3{\mathbf{y}} \frac{\delta\left( {\tilde{\mathbf{x}} \cdot \mathring{\mathbf{y}} } \right)}{ \left(\tilde{\mathbf{y}}^2\right)^{1/2} }, \label{hclimv}
\end{equation}
where the accented variables are defined as
\begin{equation}
\tilde{x}_i \equiv 2\sin\left( \frac{\pi x_i}{2}\right), \qquad \mathring{x}_i \equiv \sin\left(\pi x_i\right), \qquad x_i \in \left( -1, 1 \right).
\end{equation}
We perform the numerical integration necessary in \eqref{hclimv} using the VEGAS algorithm \cite{vegas}, as implemented in the CUBA library \cite{cuba}.
The results for $g^2T = 0.44$ and $a \in \{1, 0.5, 0.25, 0.2, 0.1 \}$
corresponding to $\beta \in \{13.6, 27.3, 54.6, 68.2, 136.4\}$ are shown as points in the left panel of Fig.~\ref{plot:hcl_1} as functions of $m_D r$ with $m_D$ given by the HCL Debye mass \eqref{eq:HCLmass}.
As can be seen from this panel, the asymptotic large-$r$ value of ${\rm Im}[V_{\rm HCL}]$ depends on $\beta$, however, the small-$r$ behavior is similar for all values of $\beta$ shown. We find that the numerical HCL curves are fit very well by the functional form
\begin{equation}
{\rm Im}[V^{(2)}_{\rm cl}(r)] = g^2T A_{\infty}\phi\left(B\,m_D r\right) .
\label{Aphi}
\end{equation}
with $\phi(x)$ given in Eq.~\eqref{eq:phidef}.
This functional form is motivated by the analytic leading-order perturbative HTL result \eqref{htlimv} for which $A^\text{HTL}_{\infty} = -C_F/4\pi$ and $B^\text{HTL}=1$ with the HTL calculated $m_D$. The fitted curves are shown as lines in Fig.~\ref{plot:hcl_1} and nicely coincide with the numerically integrated results.
\begin{figure}[tp]
\centerline{
\includegraphics[width=0.48\linewidth]{./hcl_A_beta.pdf}
$\;\;\;$
\includegraphics[width=0.48\linewidth]{./hcl_B_beta.pdf}
}
\caption{Fit values $A_\infty^{\text{HCL}}$ (left) and $B^{\text{HCL}}$ (right) of the fit function \eqref{Aphi} to the classical potential ${\rm Im}[V^{(2)}_{\rm cl}(r)]/g^2T$ computed within HCL for a wide parameter range $\beta = 10 \rightarrow 360$ for $a = 0.1$. Fits to the dependence of these parameters on $\beta$ are shown as red curves, with the chosen form of the fit functions indicated in the label. The asymptotic values are $A_\infty^{\text{HCL}}(\beta {\to} \infty)=-0.144$ and $B^{\text{HCL}}(\beta {\to} \infty)=1.55$.}
\label{plot:hcl_AB}
\end{figure}
To make the agreement better visible, we show in the right panel of Fig.~\ref{plot:hcl_1} the same results divided by their corresponding asymptotic values $A_\infty^{\text{HCL}}$. All curves fall on top of each other, demonstrating that ${\rm Im}[V^{(2)}_{\rm cl}(r)]/(g^2TA_\infty^{\text{HCL}})$ is well described by $\phi\left(B^{\text{HCL}}\,m_{D}^{\text{HCL}}\, r\right)$. This also shows that, while fitted values of $A_{\infty}^{\text{HCL}}$ depend on $\beta$, this functional form and, in particular, the parameter $B^{\text{HCL}}$ is barely sensitive to $\beta$ in the considered range of $\beta$ values.
We show the fitted parameters $A_\infty^{\text{HCL}}$ and $B^{\text{HCL}}$ in Fig.~\ref{plot:hcl_AB} for a wide range $10 \leq \beta \leq 360$. Consistent with our earlier observations, $B^{\text{HCL}}$ becomes practically constant at large $\beta$, while $A_\infty^{\text{HCL}}$ is still growing for the values considered here. To better understand the functional form of the parameters and to extrapolate to larger $\beta$, we have performed fits with different functional forms. The functions that seem to describe the results in the best way are
\begin{align}
A_\infty^{\text{HCL}}(\beta) &\approx -0.144 - 0.139\, \beta^{-0.24} \nonumber \\
B^{\text{HCL}}(\beta) &\approx 1.548 + 0.12\,e^{-0.025 \beta}\,,
\end{align}
and are included into Fig.~\ref{plot:hcl_AB} as red continuous lines.%
\footnote{We tried both functional forms for each of the parameters $A_\infty$ and $B$ and found that these forms shown in the plots are preferred by far.}
The fit functions allow us to estimate the asymptotic values
\begin{align}
\label{eq:AB_HCL_contlimit}
A_\infty^{\text{HCL}}(\beta {\to} \infty) \approx -0.144 = -0.108\,C_F\,, \qquad B^{\text{HCL}}(\beta {\to} \infty) \approx 1.548 \,.
\end{align}
Since the large-$\beta$ limit corresponds to the limit $a \to 0$, its existence is very nontrivial for an observable in classical thermal lattice field theory due to the Rayleigh-Jeans divergence. Our analysis suggests that such a limit exists for ${\rm Im}[V^{(2)}_{\rm cl}(r)]$. This was possible in our case because, different from previous studies, we studied the HCL potential as a function of $m_{D}^{\text{HCL}}\, r$ rather than any other combination of the lattice spacing or the temperature. Since $m_{D}^{\text{HCL}}$ diverges for small $a$, our parametrisation hides the divergence and leaves the functional form and the remaining parameters finite. It is interesting to compare the extrapolated parameter values \eqref{eq:AB_HCL_contlimit} to the corresponding HTL values in dimensional regularization of Sec.~\ref{sec:HTL_V}. As mentioned before, these are
\begin{align}
\label{eq:AB_HTL}
A^\text{HTL}_{\infty} \approx -0.08\, C_F\,, \qquad B^\text{HTL}=1\,.
\end{align}
The $A_\infty$ values are quite close, with the estimated HCL value $|A_\infty^{\text{HCL}}(\beta {\to} \infty)|$ being $35\%$ larger than the corresponding HTL value. Note here that our extrapolation is expected to have a large uncertainty for $A_\infty^{\text{HCL}}$ for $\beta \to \infty$ since its value is still quite rapidly changing in the considered $\beta$ interval, as seen in Fig.~\ref{plot:hcl_AB}. While the fit works surprisingly well over a large $\beta$ interval, which gave us confidence to write a limiting value, our analysis is not a proof that $|A_\infty^{\text{HCL}}(\beta {\to} \infty)|$ is indeed finite but rather an observation from this fit function. This is different for the parameter $B$, that agrees well with its estimated large-$\beta$ limit for $\beta \gtrsim 200$. Since $B$ multiplies the divergent Debye mass, the product $B\, m_{D}^{\text{HCL}}$ could be interpreted as a \changeflag{replacement for} the lattice regularized mass.
\subsection{Nonperturbative CYM simulation results for the classical Wilson loop}
We now turn to the results of our CYM lattice simulations. To extract the imaginary part of the classical potential ${\rm Im}[V_{\rm cl}(r)]$ from non-perturbative lattice simulations, we follow the procedure introduced in Ref.~\cite{Laine:2007qy} and revisited in Sec.~\ref{sec:theory_potential}. We consider rectangular Wilson loops of different spatial size $r = n a$ and temporal length $t = n_t a_t$.%
\footnote{Since we work in temporal axial gauge with $U_0 = \mathbb{1}$, the temporal edges of the Wilson loops are trivial and the loops reduce to line-line correlators.}
We use a temporal lattice spacing of $a_t = a/100$ for all results reported in this paper. The imaginary part of the classical potential ${\rm Im}[V_{\rm cl}(r)]$ at separation $r$ for a heavy quarkonium system is calculated from the large-time slope of the logarithm of the Wilson loop $C_{\rm cl}(r,t)$, averaged over lattice sites, orientations and configurations. The time evolution of $C_{\rm cl}(r,t)$ is shown for two example cases in Fig.~\ref{plot::Ccl_example}.
\begin{figure}[t]
\centerline{
\includegraphics[width=0.48\linewidth]{./Ccl_n0_0375_N_12_np_28.pdf}
$\;\;\;$
\includegraphics[width=0.48\linewidth]{./Ccl_a01_N_252.pdf}
}
\caption{Examples of the classical Wilson loop as a function of time for lattice spacing $a = 1$ and spatial separation values $r/a = \{1,\ 2,\ 3,\ 4\}$ (left), and for lattice spacing $a = 0.1$ and spatial separation values $r/a = \{10,\ 20,\ 30,\ 40\}$ (right).
}
\label{plot::Ccl_example}
\end{figure}
In the left panel of Fig.~\ref{plot::Ccl_example} we present typical results obtained on a small lattice which is averaged over $N_\text{config}=28$ classical-statistical configurations. The bands indicate the uncertainty in the mean obtained by dividing the standard deviation of the mean value by $\sqrt{N_\text{config} N_{\text{iso}} N_\text{start.points}}$. As indicated from this last expression, we exploit the isotropy and homogeneity of the lattice in order to obtain increased statistics, where we average over two different Wilson loop orientations
and use $N_\text{start.points} = N^3$ possible spatial starting points from which we can measure the Wilson loop.
As can be seen from this panel, one cannot perform the fits to extract the logarithmic slope at early times. In practice, one must wait a sufficient amount of time beyond the thermalization time ($t_0$). Typically, we use a fit window that
encompasses $(t-t_0)/a > 10$. We find that as long as one uses starting times for the fit window which satisfy this constraint, there is only a very small variation in the extracted logarithmic slopes.
In the right panel of Fig.~\ref{plot::Ccl_example} we present typical results obtained on a large lattice in which we only take into account one configuration for the initial fields. As this panel demonstrates, when using large lattices, the uncertainty is decreased due to the much larger number of starting points that can be used to measure the Wilson loop. As a result, on large lattices one does not need to sample as many initial configurations, which helps to reduce the run time required to extract statistically accurate results. We have explicitly checked that the uncertainty bands obtained with one initial configuration are consistent with results obtained using larger number of initial configurations but a smaller lattice size.
\begin{figure}[t]
\centerline{
\includegraphics[width=0.7\linewidth]{./beta16_n0_044_N12_np224_dT.pdf}
}
\caption{${\rm Im}[V_{\rm cl}(r)]/g^2T$ as a function of $r/a$ for $\beta = 16$. The blue filled disks are the results obtain herein averaged over 28 configurations. The red filled squares are the results reported by Laine et al.\ in Ref.~\cite{Laine:2007qy}.}
\label{plot:beta16}
\end{figure}
\subsubsection{Benchmark results for $\beta = $ 16 and $a=1$}
In Fig.~\ref{plot:beta16} we compare our results to the previously available ones from Ref.~\cite{Laine:2007qy} for $\beta = 16$. We set the parameter $g^2 T_0 = 0.44$ for the initial distribution so that the resulting temperature extracted from $\langle EE \rangle$ correlators is $g^2 T \approx 0.373$, which is close to the value $g^2 T = 0.375$ for the corresponding $\beta = 16$ when setting $a = 1$, $g^2 = 1$.
The error bars indicated in Fig.~\ref{plot:beta16} include uncertainties from averaging the Wilson loops over the starting points, orientations, classical configurations, uncertainties coming from the fitting residuals, and uncertainties from the temperature extraction. As can be seen from this Figure, our results are in agreement within uncertainties with the reported results from Ref.~\cite{Laine:2007qy}.
\begin{figure}[t]
\centerline{
\includegraphics[width=0.48\linewidth]{./Su2vsSu3.pdf}
$\;\;\;$
\includegraphics[width=0.48\linewidth]{./Su2vsSu3_ratio.pdf}
}
\caption{{\em Left:} ${\rm Im}[V_{\rm cl}(r)]/g^2T$ extracted from $SU(2)$ and $SU(3)$ lattice gauge theories simulated with the same values $T_0$. The given values of $T$ correspond to $SU(3)$. The potential is divided by $C_F N_c$ of the respective gauge group. {\em Right:} The ratios of $SU(3)$ to $SU(2)$ results for ${\rm Im}[V_{\rm cl}(r)]/g^2T$ for the same simulations. The orange dashed line corresponds to $8/3$.}
\label{plot:su2_1_2}
\end{figure}
\subsubsection{Comparing results for SU(3) vs.\ SU(2)}
\label{sec:SU3_vs_SU2}
To check for the $N_c$ dependence of the $SU(N_c)$ gauge theories, we perform here similar lattice calculations for $SU(3)$ and $SU(2)$ systems and extract ${\rm Im}[V_{\rm cl}(r)]/g^2T$ for different values of $T_0$. In Fig.~\ref{plot:su2_1_2} we use the same lattice parameters $a=1$ and $N=12$ as we did in Fig.~\ref{plot:beta16} and compare the simulations for the values $g^2 T_0 = \{0.2,0.3,0.4\}$. In the left panel of Fig.~\ref{plot:su2_1_2}, the values of the ratio ${\rm Im}[V_{\rm cl}(r)]/(g^2T\,C_F N_c)$ are shown for the two gauge groups as functions of $r/a$. One finds that they agree well at low $r/a$ while the deviations grow with the distance. However, the agreement improves for smaller temperatures (larger $\beta$ values).
In the right panel, the ratio of the imaginary parts of the potentials of the $SU(3)$ and $SU(2)$ theories, $({\rm Im}[V_{\rm cl}(r)]/(g^2T)[SU(3)]) / ({\rm Im}[V_{\rm cl}(r)]/(g^2T)[SU(2)])$, is plotted as a function of $r/a$. The horizontal dashed line corresponds to $(N_c C_F[SU(3)]) / (N_c C_F[SU(2)])=8/3$, which is the ratio of their $N_c C_F$ values. One observes that the data points are close to this value for all $r$ and $T_0$, which shows
that the imaginary part of the potential in our lattice simulations admits Casimir scaling
\begin{align}
\label{eq:V_propto_CN}
{\rm Im}[V_{\rm cl}(r)] \propto C_F N_c\,.
\end{align}
This is similar to the perturbative HCL expression \eqref{hclimv} where due to $\beta \propto N_c$, one also finds this relation \eqref{eq:V_propto_CN}.
The deviations visible in the right panel of Fig.~\ref{plot:su2_1_2} decrease with increasing $\beta$ (smaller temperature) and with decreasing distance, as we have observed in the left panel.
An important source for these deviations is that we compare here simulations with the same $g^2 T_0$ parameter and not with the same temperature $g^2 T$. As visible in the left panel of Fig.~\ref{plot:Tmd_T0} for $a=1$ and $SU(3)$ theory, the extracted temperatures deviate stronger at larger $g^2 T_0$. The deviations in $SU(2)$ theory are qualitatively similar but can lead to quantitatively different values of $g^2 T$, which implies that we compare the $SU(N_c)$ theories effectively at slightly different temperatures. This effect is reduced when going to smaller temperatures or smaller lattice spacings (larger $\beta$ values).
Moreover, as we will show below, the short-distance behavior is less sensitive to the temperature, which is the reason why the deviations decrease at lower $r/a$.
\subsubsection{SU(3) simulation results for different $\beta$}
We performed CYM lattice simulations with $g^2 T_0 = 0.44$ and $a = \{1,\ 0.5,\ 0.25,\ 0.2\}$ keeping $aN = 12$ fixed and also for $g^2 T_0 = 0.45$ and $a =0.1$ with $N = 252$. Our results for ${\rm Im}[V_{\rm cl}(r)]/g^2T$ are shown in Fig.~\ref{plot:imv_mDr} as points connected by solid colored lines, and the respective $\beta$ values are written in the figure.
The $a = 0.1$ CYM lattice simulation results are compared with the corresponding results (same $a$ and $T$) obtained from HCL perturbation theory \eqref{hclimv}, included as a blue dashed curve. For both our CYM lattice results and the HCL results,
${\rm Im}[V_{\rm cl}(r)]/g^2T$ is plotted as a function of $m_D r$,
where we use the leading-order Debye mass $m_{D}^{\text{HCL}}$ calculated in HCL theory \eqref{eq:HCLmass}.
On the right edge of the figure we indicate the asymptotic $r \to \infty$ values obtained using both the dimensionally-regularized continuum HTL result \eqref{htlimv} and the extrapolated $N,\beta \rightarrow \infty$ HCL result with the parameters \eqref{eq:AB_HCL_contlimit}.
\begin{figure}[t]
\centerline{
\includegraphics[width=0.8\linewidth]{./imv_aVar_with_a1_mike.pdf}
}
\caption{ ${\rm Im}[V_{\rm cl}(r)]/g^2T$ for different lattice spacing from nonperturbative Wilson loops (solid lines) and from HCL perturbation theory for $a=0.1$ (blue dashed line). The temperature values (extracted from $EE$ correlations) are slightly different for different values of $a$, and we write the corresponding $\beta$ values instead. The $r \rightarrow \infty$ limits of HCL for the continuum extrapolation $\beta \to \infty$ \eqref{eq:AB_HCL_contlimit} and of the dimensionally-regularized HTL perturbation theories \eqref{eq:AB_HTL} are indicated by short dotted red lines on the right axis for comparison. The distance in the horizontal axis is scaled using the HCL Debye mass \eqref{eq:HCLmass} for $m_D(a)$.}
\label{plot:imv_mDr}
\end{figure}
As can be seen from this Figure (Fig.~\ref{plot:imv_mDr}), the CYM lattice results seem to \prevchangeflag{approach a finite large-$\beta$ form}
when plotted as a function of $m_D r$, with the CYM results with $\beta=136$ even overlapping with the corresponding HCL curve at $m_D r \lesssim 1$.
At large values of $r$ our CYM simulation results approach the corresponding HCL curve for the respective $\beta$ value. For the largest $\beta = 136$ in the figure, we know from the left panel of Fig.~\ref{plot:hcl_AB} that the amplitude of the potential at large $r$ is still quite far away from the extrapolated value. However, Fig.~\ref{plot:imv_mDr} indicates that \changeflag{with increasing $\beta$, the CYM potential is approaching the perturbative HCL large-$\beta$ limit from below.}
We emphasize that, while previous studies showed the potential for different $\beta$ values as functions of $r/a$, we find that plotting it as a function of $m_D r$ makes the comparison more intuitive. Most importantly, this allows to \changeflag{study the classical potential even at large $\beta$ values while incorporating the dominant UV divergence into the mass.}%
\footnote{\prevchangeflag{By adjusting discretization parameters $a$ and $N$, we ensure that, despite the divergence of $m_{D}^{\text{HCL}} \sim 1/a^{1/2}$, relevant distance scales $r \sim 1/m_D$ lie within our lattice $r_{\text{min}} \ll 1/m_D \ll r_{\text{max}}$. Here the minimal and maximal lattice distances are given by $r_{\rm{min}} = a$ and $r_{\rm{max}} \sim Na$.}}
We have performed such a large-$\beta$ extrapolation for the HCL calculations already in Sec.~\ref{sec:HCL_results} by considering ${\rm Im}[V_{\rm cl}(r)]/g^2T$ as a function of $m_{D}^{\text{HCL}}\, r$ for increasing \changeflag{$\beta$.}
Instead of the leading-order perturbative HCL mass $m_{D}^{\text{HCL}}$, one could instead plot the CYM results as a function of $\hat m_D\, r$, with the values of \changeflag{the screening mass} $\hat{m}_D$ extracted from CYM longitudinal chromo-electric field correlators as discussed in Sec.~\ref{sec:thermalstate}. The latter is an estimate of the Debye mass including higher-order and nonperturbative contributions. We find that, in practice, the HCL Debye mass \eqref{hclimv} is proportional to the correlator-extracted Debye mass $\hat{m}_D$
over a wide range of $\beta$ via Eq.~\eqref{eq:mdhat_md} with proportionality constant $1.42$, and that the resulting rescaling only introduces a constant rescaling of the horizontal axis.
Since there are larger uncertainties inherent in the extraction of $\hat{m}_D$ from the chromo-electric correlators, we have chosen to rescale the classical lattice and HCL results using the same HCL mass definition.
\subsubsection{Small-distance behavior}
\begin{figure}[t]
\centerline{
\includegraphics[width=0.48\linewidth]{./imvr_n0var_N252_mDr_a01.pdf}
$\;\;\;$
\includegraphics[width=0.48\linewidth]{./fit_kappa_plot2fromAB.pdf}
}
\caption{
(Left) The extracted classical potential ${\rm Im}[V_{\rm cl}(r)]/g^2T$ as a function of $m_Dr$ for fixed lattice spacing $a = 0.1$ but for different temperatures $g^2 T = 0.2,\dots, 0.59$,
which correspond to the range $\beta = 300,\dots, 100$.
The continuous curves are fits to our data points using the form of the potential in \eqref{Aphi}.
(Right) The combined data of the left panel, with additional data sets for $g^2 T_0 = 0.25, 0.35, 0.45, 0.55, 0.65$, is compared to the analytical form \eqref{Aphi} with parameter values \eqref{eq:latt_AB_fit} and its low-distance limit \eqref{eq:ImV2_lowX} as black and red curves, respectively.
}
\label{plot:short-dist}
\end{figure}
\changeflag{Let us now compare the fitting function \eqref{Aphi} with parameters $A_\infty$ and $B$ to the imaginary part of the heavy-quark potential.
We saw previously that it} agrees with this form in both the HTL and the HCL formalisms with the parameters
\begin{align}
\label{eq:HTL_AB_fit}
A^\text{HTL}_{\infty} &\approx -0.080\, C_F\,, \qquad \qquad \quad~ B^\text{HTL}=1 \\
A_\infty^{\text{HCL}}(\beta {\to} \infty) &\approx -0.108\,C_F\,, \qquad B^{\text{HCL}}(\beta {\to} \infty) \approx 1.548\,.
\label{eq:HCL_AB_fit}
\end{align}
We \changeflag{provide evidence} that, at small distances, the classical potential extracted from our lattice simulation data also follows the functional form \eqref{Aphi} \changeflag{for different values of $\beta$}. This is shown in the left panel of Fig.~\ref{plot:short-dist}, where ${\rm Im}[V_{\rm cl}(r)]/g^2T$ for fixed lattice spacing $a=0.1$ is plotted
for different temperatures corresponding to $100 \lesssim \beta \leq 300$, as a function of $m_D\, r$. Fits to each data set using \eqref{Aphi} in the considered interval $m_D r \leq 6$ are included as continuous lines. One finds very good agreement between the data and the respective fits. Moreover, although $m_D$ and the extracted potential vary with the temperature $T$ and, equivalently, with $\beta$, all curves are seen to fall on top of each other at low distances $m_D r \lesssim 1$.
\changeflag{We have seen such behavior already in Fig.~\ref{plot:imv_mDr} for even smaller values of $\beta$.}%
\footnote{\changeflag{In Fig.~\ref{plot:imv_mDr}, curves for different $\beta$ values are observed to overlap at short distances $m_D r \lesssim 1$. This works best when the smallest distance on the lattice is much smaller than the inverse Debye mass. For $\beta = 16$, there is only one data point within $m_D r \lesssim 1$, making a comparison more challenging.}}
\changeflag{This indicates} approximately temperature-independent low-distance behavior for \changeflag{a wide} $\beta$ interval.
Averaging over the resulting parameter values \changeflag{from the fits to each data set in Fig.~\ref{plot:short-dist}}, we obtain
\begin{align}
\label{eq:latt_AB_fit}
A_\infty^{\text{latt}} = -0.24 \pm 0.02 = \left( -0.18 \pm 0.015 \right)C_F\,,
\qquad
B^{\text{latt}} = 1.2 \pm 0.05\,.
\end{align}
Here the errors have been estimated by adding the standard error and the fitting error.
In the right panel of Fig.~\ref{plot:short-dist}, we combine all of the curves of the left panel and compare them to the function \eqref{Aphi} with the mean parameter values in \eqref{eq:latt_AB_fit}, shown as a black line. As expected, one observes very good agreement with our data, especially at low distances $m_D r$.
We can also perform a short-distance expansion of the fitting function for the imaginary part of the potential \eqref{Aphi}, neglecting terms $\mathcal{O}((m_D r)^4)$,
\begin{align}
\label{eq:ImV2_lowX}
{\rm Im}[V_{\rm cl}(r)] \simeq -r^2\,\frac{1}{9}\,|A_\infty|B^2\, g^2T\,m_D^2\left( 4-3\gamma - 3\log (B\, m_D\, r) \right).
\end{align}
\changeflag{The resulting curve} is shown as a red line and is observed to agree well with our data points for $m_D\, r \lesssim 0.5$.
\changeflag{Thus, for a wide range of $\beta$ values, the short-distance behavior of our CYM lattice data agrees well with the perturbative functional form \eqref{Aphi} and its leading short-distance expansion, which is parametrically given by $\left|{\rm Im}[V_{\rm cl}(r)]\right| \sim C_F\, g^2T\,\left(m_D\,r\right)^2\,\log (m_D\,r)$.}
\section{Conclusions and outlook}
\label{sec:conclusions}
In this paper we used classical-statistical lattice simulations of pure Yang-Mills fields to extract the imaginary part of the heavy-quark potential as a function of the quark-antiquark separation. In order to carry out our simulations on large lattices, we used a simplified scheme to generate thermalized gauge field configurations which relied on initialization of chromo-electric fields in momentum-space followed by a period of self-thermalization. In order to determine whether the fields were thermalized, we extracted the temperature of the system as a function of time directly from the dynamically generated chromo-electric field correlators.
We went beyond previous classical-statistical lattice calculations of ${\rm Im}[V_{\rm cl}]$ by considering rather large lattice sizes and \changeflag{a wide range of values of $\beta$ (or, equivalently, of the lattice spacing $a$ and temperature)}. We considered both $SU(2)$ and $SU(3)$ gauge groups and found that the potentials obtained obeyed a simple Casimir scaling.
\changeflag{Using our CYM simulations on such large lattices, we argued that ${\rm Im}[V_{\rm cl}]$ should be plotted as a function of $m_D\, r$, with the screening mass $m_D$, when comparing our lattice results for different $\beta$ values or to perturbative calculations. In particular, we showed that ${\rm Im}[V_{\rm cl}]$ approaches the perturbative lattice-regularized HCL results with increasing $\beta$. As a side result, our calculations suggest that a finite large-$\beta$ limit exists for ${\rm Im}[V_{\rm cl}]$ when calculated in the HCL framework and plotted as a function of $m_D\, r$. In the region of small distances $m_D\, r \lesssim 1$, we demonstrated that ${\rm Im}[V_{\rm cl}]$ from our CYM lattice simulations is insensitive to the finite values of $\beta$ over a wide parameter range, and is close to the HCL results.}
We also found that both our lattice simulation and HCL results were very well approximated by a functional form which can be obtained from a leading-order hard-thermal loop calculation. One only needs to take into account a different \changeflag{prefactor $A_\infty$}
and a different scaling of the argument of the functional form, encoded in a parameter $B$. Using fits of this form and then expanding the result at small $m_D r$, we were able to extract \changeflag{small-distance approximations for the imaginary part of the heavy-quark potential}.
Looking to the future, we plan to perform a similar calculation in an anisotropic gluonic plasma that is expanding along the longitudinal direction. This case is of phenomenological interest since the expansion of the quark-gluon plasma in relativistic heavy-ion collisions generates large momentum-space anisotropy. This renders the perturbative analytic calculation of the imaginary part of the heavy-quark potential ill-defined due to the presence of unstable modes in the soft gauge field propagator. \changeflag{In contrast, classical-statistical lattice simulations are applicable in that case, and this work sets the stage to corresponding future studies out of equilibrium}.
\begin{table*}[ht]
{\small
\centering
\begin{tabular}{|c c | c| c| c| c| c| c |}
\cline{3-8}
\multicolumn{2}{c|}{} & \multicolumn{6}{c|}{${\rm Im}[V_{\rm cl}(r, t \rightarrow \infty)]/g^2T \qquad [a = 1]$} \\
\hline
$g^2 T_0$ & $g^2 T$ & $r/a = 1$ & $r/a = 2$ & $r/a = 3$ & $r/a = 4$ & $r/a = 5$ & $r/a = 6$ \\ [0.5ex]
\hline\hline
0.20 & 0.185(2) & -0.030 & -0.081(1) & -0.131(2) & -0.174(2) & -0.210(3) & -0.240(4) \\
\hline
0.25 & 0.227(4) & -0.037(1) & -0.099(2) & -0.158(3) & -0.209(4) & -0.252(5) & -0.288(6) \\
\hline
0.30 & 0.267(5) & -0.045(1) & -0.117(2) & -0.187(4) & -0.246(5) & -0.297(6) & -0.341(8) \\
\hline
0.35 & 0.306(7) & -0.052(1) & -0.136(3) & -0.217(5) & -0.286(7) & -0.345(8) & -0.397(10) \\
\hline
0.40 & 0.343(11) & -0.060(2) & -0.157(5) & -0.249(8) & -0.329(11) & -0.398(13) & -0.461(15) \\
\hline
0.45 & 0.379(11) & -0.069(2) & -0.179(5) & -0.284(8) & -0.376(11) & -0.457(14) & -0.529(16) \\
\hline
0.50 & 0.413(11) & -0.078(2) & -0.204(5) & -0.323(9) & -0.428(12) & -0.515(14) & -0.574(19) \\ [1ex]
\hline
\end{tabular}
}
\caption{Values of ${\rm Im}[V_{\rm cl}(r)]/g^2T$ for different $T_0$ extracted from lattice simulations with $a=1, N = 64$ and 224 configurations.}
\label{tab:table1}
\end{table*}
\begin{table*}[ht]
{\small
\centering
\begin{tabular}{|c c | c| c| c| c| c| c |}
\cline{3-8}
\multicolumn{2}{c|}{} & \multicolumn{6}{c|}{${\rm Im}[V_{\rm cl}(r, t \rightarrow \infty)]/g^2T \qquad [a = 0.1]$} \\
\hline
$g^2 T_0$ & $g^2 T$ & $r/a = 10$ & $r/a = 20$ & $r/a = 30$ & $r/a = 40$ & $r/a = 50$ & $r/a = 60$ \\ [0.5ex]
\hline\hline
0.20 & 0.199(1) & -0.104(1) & -0.165(1) & -0.188(2) & -0.200(2) & -0.207(3) & -0.210(3) \\
\hline
0.25 & 0.248(2) & -0.114(1) & -0.171(1) & -0.191(2) & -0.200(2) & -0.207(2) & -0.212(3) \\
\hline
0.30 & 0.296(2) & -0.128(1) & -0.186(2) & -0.207(2) & -0.224(2) & -0.232(3) & -0.236(4) \\
\hline
0.35 & 0.345(3) & -0.137(1) & -0.197(2) & -0.218(2) & -0.229(2) & -0.243(3) & -0.258(4) \\
\hline
0.40 & 0.394 & -0.146 & -0.206(1) & -0.226(1) & -0.239(2) & -0.255(3) & -0.269(4) \\
\hline
0.45 & 0.442(3) & -0.152(1) & -0.209(2) & -0.225(2) & -0.235(3) & -0.248(3) & -0.255(4) \\
\hline
0.50 & 0.490(8) & -0.161(3) & -0.218(4) & -0.238(4) & -0.242(5) & -0.251(5) & -0.257(10) \\ [1ex]
\hline
\end{tabular}
}
\caption{Values of ${\rm Im}[V_{\rm cl}(r)]/g^2T$ for different $T_0$ extracted from lattice simulations with $a= 0.1, N = 252$ (one configuration).}
\label{tab:table2}
\end{table*}
\acknowledgments{
We would like to thank M.\ Escobedo and J.H.\ Weber for useful discussions.
M.S.\ and B.K.\ were supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics Award No.~DE-SC0013470. We thank the Ohio Supercomputer Center for support under the auspices of Project No.\ PGS0253. The computational results presented have been also achieved in part using the Vienna Scientific Cluster (VSC).
}
|
\section{Introduction}
Recently deep convolutional neural network classifiers have significantly dominated the computer vision field by yielding the state-of-the-art performance in many classification tasks such as visual object, face, scene, and action classification and detection problems. The soft-max loss (more precisely, cross entropy loss on probabilities obtained by applying soft-max function to the output logits of the deep neural network) is the most commonly used classification loss function in the classical convolutional neural networks. However, many studies \cite{R1,R2,R3,R10,R12} demonstrated that soft-max loss works well for closed set recognition problems where the test samples come only from the classes seen during training. However, soft-max typically fails for open set recognition problems, the settings that allow the test samples coming from the unknown classes \cite{R33}. This is partly due to the two important reasons: classical soft-max does not enforce a large-margin between classes and there is no mechanism to control the intra-class variations. In open set recognition problems, novel classes (ones not seen during training) may occur at test time, and classifiers are expected to reject the test samples that do not appear to belong to any of the known training classes. Therefore, compactly clustered and well separated acceptance regions estimated for classes improve accuracies in such open set recognition problems. Here, acceptance region of a class indicates the bounding region where a test sample is assigned to that class label if its feature representation lies in that region and rejected if its feature vector is outside the region.
To return more compact class acceptance regions, \cite{R4,R5} use contrastive loss (CL). Contrastive loss minimizes the Euclidean distance of the positive sample pairs and penalizes the negative pairs that have a distance smaller than a given threshold. In a similar manner, \cite{R6,R7,R8,R9} employ triplet loss function that used a positive sample, a negative sample and an anchor. An anchor is also a positive sample and the goal is simultaneously to minimize the distance between the positive sample and the anchor and to maximize the distance between the anchor and the negative sample. Although methods using both contrastive and triplet loss functions return compact decision boundaries, the number of sample pairs or triplets grows quadratically compared to the total number of samples, which results in slow convergence and instability. To overcome this limitation, a careful sampling/mining of data is required, and the accuracy largely depends on this mining strategy. Recent studies \cite{R8,R9,R40} also reveal that large batch sizes during learning and selecting the same number of samples per class are necessary for good results.
Wen et al. \cite{R1,R2} combined the soft-max loss function with the center loss for face recognition. Center loss reduces the intra-class variations by minimizing the distances between the individual class samples and their corresponding class centers. The resulting method significantly improves the accuracies over the method using soft-max alone in the context of face recognition, which shows the importance of compact class decision boundaries. Despite its good accuracies, this method does not explicitly attempt to maximize the inter-class separability and relies on soft-max for this goal.
Zhang et al. \cite{R10} combined the range loss with soft-max loss to overcome the limitations of the contrastive loss. The range loss consists of two terms that simultaneously minimize the intra-class variations and maximize the inter-class separation.
However, this method is very complex since one has to carefully tune two weights to balance 3 different loss terms. Furthermore, it only penalized one center pair in each batch to maximize the inter-class separation. As pointed out in \cite{R11}, this strategy is not comprehensive since more center pairs may have margins smaller than a designated threshold. Deng et al. \cite{R12} introduced a method combining soft-max loss function with the marginal loss to create compact and well separated classes in Euclidean space. Marginal loss enforces the distances between sample pairs from different classes to be larger than a selected threshold while enforcing the distances between sample pairs coming from the same classes to be smaller than the selected threshold. Wei et al. \cite{R11} combined soft-max loss and center loss functions with the minimum margin loss where the minimum margin loss enforces all class center pairs to have a distance larger than a specified threshold. This method is also complex since there are many parameters that must be carefully set by the user.
Support Vector Data Description (SVDD) method of \cite{R35} aims to find a compact hypersphere that includes the majority of the positive class samples for anomaly detection. This method is extended in \cite{R34} by using deep neural networks, where the proposed classifier trains a deep neural network that minimizes the volume of hypersphere that encloses the deep network representations of the data. However this method has limitations in the sense that the hypersphere center is fixed to a certain vector at the beginning and it is not updated as in center loss method. Also, it requires to use only ReLU activation functions to avoid trivial solutions.
Liu et al. \cite{R14,R15} proposed the SphereFace method which uses the angular soft-max (A-softmax) loss that enables to learn angularly discriminative features. To this end, the proposed method projects the original Euclidean space of features to an angular space, and introduces a multiplicative angular margin to maximize inter-class separation. However, the SphereFace method has limitations in the sense that the decision margins change for each class. As a result, some inter-class features have a larger margin while others have a smaller margin, which reduces the discriminating power \cite{R16}. Zhao et al. \cite{R17} proposed the RegularFace method in which A-softmax term is combined with exclusive regularization term to maximize the inter-class separation. Wan et al. \cite{R16} introduced the CosFace method which imposes an additive angular margin on the learned features. To this end, they normalize both features and learned weight vectors to remove radial variations and then introduce an additive margin term $m$ to maximize the decision margin in the angular space.
A similar method called ArcFace is introduced in \cite{R18}, where an additive angular margin is added to the target angle to maximize the separation in angular space. It should be noted all these aforementioned methods, which maximize the margin in the angular space rather than the Euclidean space, are proposed for face recognition. These methods work well for face recognition since face class samples in specific classes can be approximated by using linear/affine spaces, and the similarities can be measured well by using the angles between sample vectors in such cases. Linear subspace approximation will work as long as the number of the features are much larger than the number of class specific samples which holds for many face recognition problems. However, for many general classification problems, the training set size is much larger compared to the dimensionality of the learned features and therefore these methods cannot be generalized to the classification applications other than face recognition as demonstrated in our experiments.
Regarding open set recognition, in addition to the methods using hand crafted features such as \cite{R36,R37}, recent methods \cite{R38,R39} use deep neural networks for such settings. Neal et al. \cite{R38} introduce a method that uses an encoder-decoder GAN (Generative Adversarial Network) architecture to generate synthetic open set examples. Oza and Patel \cite{R39} propose an open set recognition algorithm using class conditioned auto-encoder with new training and testing methodologies.
\medskip\noindent{\bf Contributions:} In this paper, we propose a general deep neural network classifier that returns compact and well separated acceptance regions for each class. The proposed methodology allows to use a margin-based cost function to maximize the inter-class separation and there is a mechanism that enables to control the compactness of the class regions to minimize the intra-class variations. As opposed to the methods combining soft-max loss function with other loss terms, our deep neural network gracefully integrates intra-class compactness and inter-class separation goals into a single unified framework. Moreover, in contrast to the some of the aforementioned methods above, the proposed method has a nice geometric intuition that helps to interpret the resulting class acceptance regions.
Our proposed method returns a compact polyhedral acceptance region for each class. The samples coming from other classes can be easily rejected based on the distances to the resulting acceptance regions. This makes the proposed method suitable especially for open set recognition and anomaly detection problems. Our proposed method does not need to generate synthetic samples and it significantly outperforms state-of-the-art open set recognition methods such as \cite{R38,R41} relying on synthetically created open set examples.
\section{Method}
This paper introduces a deep neural network classifier that returns well separated compact class decision boundaries. To this end, we propose a new loss function for CNN networks. In the proposed method, inter-class separation is realized by using the hinge loss and we have another loss term to adjust the compactness of the acceptance regions of the classes and class features. We build our method based on polyhedral conic classifiers introduced in \cite{R20,R19}. Therefore, we first briefly overview polyhedral conic functions and then introduce our new deep neural network classifier.
\subsection{Preliminaries}
In polyhderal conic classification methods, each class is approximated with a polyhedral conic region constructed by using polyhedral conic functions (PCFs). PCFs were first introduced in \cite{R20} to separate two arbitrary finite disjoint sets, and its extended version, extended polyhedral conic functions (EPCFs), have been proposed in \cite{R19} for visual object detection, closed and open set recognition. PCFs and EPCFs can be defined as follows:
\begin{align}
\!\! f_{\mathbf{w},\gamma,\mathbf{s},b}(\mathbf{x})&=\mathbf{w}^\top(\mathbf{x}-\mathbf{s})+\gamma\left\|\mathbf{x}-\mathbf{s} \right\|_1 -b, &\textrm{(PCF)}\!\!\label{eq:con_func} \\
\!\! f_{\mathbf{w},\boldsymbol{\gamma},\mathbf{s},b}(\mathbf{x})&=\mathbf{w}^\top(\mathbf{x}-\mathbf{s})+\boldsymbol{\gamma}^\top|\mathbf{x}-\mathbf{s}| -b, &\textrm{(EPCF)}\!\!\label{eq:econ_func}
\end{align}
where $\mathbf{x} \in \R^d$ is a test point, $\mathbf{s}\in\R^d$ is the cone vertex, $\mathbf{w} \in \R^d$ is a weight vector and $b$ is an offset. For PCF, $\left\|\mathbf{u}\right\|_1=\sum_{i=1}^d|u_i|$ denotes the vector L1 norm and $\gamma$ is a corresponding weight, while for EPCF, $\left|\mathbf{u}\right|=\left(|u_1|,\ldots,|u_d|\right)^\top$ denotes the component-wise modulus and $\boldsymbol{\gamma}$ is a corresponding weight vector.
Both \cite{R20} and \cite{R19} introduced polyhedral conic classifiers that use PCFs and EPCFs with decision regions $f(\mathbf{x})\leq 0$ for positives and $f(\mathbf{x})>0$ for negatives (the opposite of the well known SVM - Support Vector Machine - decision rule).
The resulting acceptance region includes an intersection of at most $2^d$ half spaces, and forms “kite-like" polyhedral acceptance regions as seen in Fig. \ref{fig:pcc}. The values of polyhedral conic functions given in (\ref{eq:con_func},\ref{eq:econ_func}) can be regarded as the distances to the acceptance regions in the sense that more negative values indicate that the test samples are closer to the center of the acceptance region and more positive values indicate that they are far from the acceptance region and therefore they must be rejected.
\begin{figure*}[tbh]
\centering
\begin{tabular}[h]{@{}ccc@{}}
\includegraphics[width=45mm]{gfx/a1} &
\includegraphics[width=45mm]{gfx/b1} & \includegraphics[width=45mm]{gfx/f1} \\
(a) & (b) & (c)
\end{tabular}
\caption{Visualization of polyhedral conic classifier (PCC) for 2D synthetic data: The positive acceptance regions are “kite-like” octahedroids containing the points for which a linear hyperplane lies above an $L_1$ cone.(a): 2D positive (yellow) and negative (blue) samples, (b) view of positive-class acceptance region in 3D, (c): Resulting “kite-like” acceptance region in 2D space. (Figure is adopted from \cite{R19})}
\label{fig:pcc}
\end{figure*}
\subsection{Deep Compact Polyhedral Conic Classifier}
Cevikalp and Saglamlar \cite{R19} introduced a deep neural network classifier using extended polyhedral conic function. However, this method possesses two major limitations: As a first limitation, instead of using a different cone vertex for each class, the authors used only a single common cone vertex for all classes in multi-class classification problems. This largely limits the classifier to return more discriminative class decision boundaries as illustrated in Fig. \ref{fig:2}. As seen in the figure, allowing only one common cone vertex for all classes, restrains the classifier to return well localized decision boundaries for the classes whose samples lie far from the selected cone vertex. This is because, the resulting acceptance region is formed by intersecting the cone (whose facets are learned by the classification algorithm) including the vertex $(\mathbf{s},-b)$ with the hyperplane as seen in Fig. \ref{fig:2}. Therefore, it is crucial to use a class-specific cone vertex point $\mathbf{s}$ close to the class samples for each class. This also speeds up the training phase despite it brings a slight additional complexity during testing time. As a second limitation, \cite{R19} did not use any mechanism to return more compact class acceptance regions, which corresponds to minimizing the intra-class variations. Therefore, although the method took the inter-class separation into consideration, it neglected the intra-class variation as in soft-max loss function.
\begin{figure}[tbh]
\centering
\begin{tabular}[h]{@{}cc@{}}
\includegraphics[width=0.5\columnwidth]{gfx/figure1_new.pdf} &
\includegraphics[width=0.5\columnwidth]{gfx/figure2.pdf} \\
\end{tabular}
\caption{Using a single cone vertex for all three classes as seen on the left restrains to return better acceptance regions for classes. Using 3 cones (the separating cone of each class has the same color as the class samples) seen on the left figure still manages to return true acceptance regions determined by the dashed lines. However, this becomes more difficult or impossible for the classes whose samples lie far from the overall mean. In contrast, using a different cone vertex point $\mathbf{s}_c$, $c=1,\ldots,3$, as seen on the right causes to return better acceptance regions easily. It also significantly improves the convergence time during training.}
\label{fig:2}
\end{figure}
In this paper, we propose a new deep neural network classifier that simultaneously maximizes the inter-class margin and minimizes the intra-class variations. To this end, we first extend the method proposed in \cite{R19} such that different cone vertices are used for each class. Secondly, we introduce an elegant mechanism to adjust the compactness of the class acceptance regions, which in turn enables us to minimize the intra-class variations.
It should be noted that the acceptance regions of the classes are bounded and convex only when $b>0$, $\boldsymbol{\gamma}>0$, and $|w_i|<\gamma_i$ for all $i=1,\ldots,d$, i.e., when the hyperplane section has a shallower slope than every facet of the $L_1$ cone (see the proof given at the supplementary material).
Considering that these conditions are satisfied, the hyper-volume of the class acceptance regions depends on two factors: $\gamma_i$ values and $b$. For fixed $\gamma_i$ values, increasing the $b$ value increases the volume since the cone vertex is located at $(\mathbf{s},-b)$. This corresponds to move the $L_1$ cone downward without any change in its facets. Decreasing the $b$ value on the other hand will decrease the volume since this corresponds to move the cone upward closer to the intersecting hyperplane (see the related video given as supplementary material showing the volume change based on $b$ values). For a fixed $b$ value, we can also change the volume by changing the slopes of the rays forming the cone through $\gamma_i$ values. Increasing the $\gamma_i$ values decreases the volume of the acceptance region, whereas the decreasing $\gamma_i$ values increases the volume (see the related video).
Therefore, the overall volume of the class acceptance regions is related to $b/\gamma_i$ since the acceptance region has width $O(b/\gamma_i)$ along axis $i$. Therefore, for fixed $b$ values, if we keep increasing $\gamma_i$ values, the acceptance region of a specific class shrinks and all corresponding class samples gather around a compact region. This effect is similar to the center loss term that enforces the samples of a specific class to cluster around their mean. In fact as seen in the synthetic experiments, the class samples typically collapse onto a single point in the proposed method although we do not explicitly enforce this constraint as in center loss function.
Now, assume we are given labeled training samples in the form $(\mathbf{x}_i,y_i)$, $i=1,\ldots,n$ and $y_i \in \left\{1,\ldots,C\right\}$, where $C$ is the number of classes. The input samples $\mathbf{x}_i$ can be 2D images or some vectoral inputs.
Let $\mathbf{f}_i\in \R^d$ is the deep neural network feature representation of $\mathbf{x}_i$ at the classification layer of the used network.
Consider that $\mathbf{w}_c$ and $\boldsymbol{\gamma}_c$, $c=1,\ldots,C$, are the weights of the extended polyhedral conic classifier used in the last stage of the deep neural network.
By setting $\mathbf{\tilde{w}}_c=-\mathbf{w}_c$, $\tilde{{\boldsymbol{\gamma}}}_c=-\boldsymbol{\gamma}_c$ (this is just for converting decision function into well-known SVM decision form), the loss function of the proposed deep neural network method can be written as
\begin{multline}
\label{eq:mopt}
\Scale[0.80]{\underset{\mathbf{\tilde{W}},\mathbf{b}}{\text{min}} \:\:\:\:\: \frac{\lambda}{2}\sum_{c=1}^C (\mathbf{\tilde{w}}_c^\top \mathbf{\tilde{w}}_c) + \sum_{i=1}^n \sum_{j\in \left\{1,...,C\right\}, j\neq y_i} H_1( (\mathbf{\tilde{w}}_{y_i}^\top (\mathbf{f}_i-\mathbf{s}_{y_i}) }\\
\Scale[0.80]{+\tilde{{\boldsymbol{\gamma}}}_{y_i}^\top \left| \mathbf{f}_i-\mathbf{s}_{y_i} \right| + b_{y_i} ) - ( \mathbf{\tilde{w}}_{j}^\top (\mathbf{f}_i-\mathbf{s}_{j})+\tilde{{\boldsymbol{\gamma}}}_{j}^\top \left| \mathbf{f}_i-\mathbf{s}_{j} \right| + b_j ) )} \\
\Scale[0.80]{+ \eta \sum_{c=1}^C \sum_{m=1}^d \text{max}(0,\kappa-(-\tilde{\gamma}_{cm}-\left|\tilde{w}_{cm}\right|)).}
\end{multline}
Here, $H_\Delta(t) = \text{max}(0,\Delta-t)$ is the classical hinge loss, $\lambda$ is a regularization weight for $\mathbf{\tilde{w}}_c$, $\mathbf{\tilde{W}}=\left[\mathbf{\tilde{w}}_1 \ldots \mathbf{\tilde{w}}_C \right]$, $\mathbf{b}=\left[b_1 \ldots b_C \right]^\top$, and $\eta$ is the weight term for balancing the compactness loss term given as the last term with respect to other loss terms of the objective function. The $\kappa>0$ value given in the compactness term is a user-supplied cost penalty for ensuring $|w_i|<\gamma_i$ (or equivalently $-\tilde{\gamma}_i>|\tilde{w}_i|$) and decreasing values of $\tilde{{\boldsymbol{\gamma}}}$ (increasing values of $\boldsymbol{\gamma}$) to minimize the hyper-volume of the class acceptance regions. The cone vertices, $\mathbf{s}_{c}$, $c=1,\ldots,C$, are set to the mean of the classes. The first term in the optimization function is the regularization term for $\mathbf{\tilde{w}}_c$ weights and it is already implemented in deep neural network classification methods as there are various methods to implement the regularization. The second term guarantees the large-margin separation and ensures that the classes are separated by a margin at least 1. The last term in the optimization controls the compactness of the class distribution, i.e., intra-class variation. To this end we use hinge loss again, and the positive $\kappa$ value acts as a margin to make sure that the positive $\gamma$ value is larger than $|w_i|$ by at least $\kappa$ in all directions, $m=1,\ldots,d$. The classes become more tight and compact as the value of $\kappa$ increases. This is similar to the center loss term in spirit, but we can adjust the compactness of the classes by changing the value of $\kappa$.
\begin{figure*}[t]
\centering
\begin{tabular}[tbh]{@{}cccc@{}}
\includegraphics[width=40mm]{gfx/synt1kapa0.png} &
\includegraphics[width=40mm]{gfx/synt1kapa050.png} & \includegraphics[width=40mm]{gfx/synt1kapa075.png} & \includegraphics[width=40mm]{gfx/synt1kapa1.png} \\
$\kappa=0.0$ & $\kappa=0.35$ & $\kappa=0.70$ & $\kappa=1.0$
\end{tabular}
\caption{The positive class (shown with orange color) acceptance regions for different $\kappa$ values. If we do not enforce any constraint on the learned weights, i.e., $\kappa=0.0$, the acceptance region is unbounded. As we increase the value of $\kappa$, the positive class acceptance regions shrink, and the positive class samples tightly cluster around a compact region as expected. }
\label{fig:3}
\end{figure*}
Our proposed classification loss function allows to control both inter-class separation and intra-class variation, and it has a nice geometry that helps to interpret the resulting acceptance regions, which many recent methods lack.
It should be noted that the Deep Support Vector Data Description (SVDD) method of \cite{R34} approximates a compact bounding hypersphere for a particular class. Compared to the polyhedral acceptance regions found by the proposed method, hyperspheres are looser models and they cannot approximate the class region more accurately compared to more tight bounding polyhedral regions returned by the proposed method.
\subsubsection{Implementation Details}
We used ResNet architecture \cite{R22} unless otherwise stated, and trained it with the SGD algorithm. It should be noted that the regularization term is already implemented in many deep neural network software, and it is commonly called as weight decay.
The common practice is to set weight decay to $0.00001$ for ResNet architectures, thus we tried values closer to this number. The best value is found to be equivalent to $0.0005$. Mini-batches of size $128$ were used in the experiments.
The momentum parameter was set to $0.9$, and we initialized the weights both randomly and by using the weights of a pre-trained network.
The $\kappa$ values control the compactness of the classes and we used validation data to set the best value.
We set the cone vertex of each class to the mean of the corresponding class samples. Since the feature representations change during training iterations, the class means need to be updated in each batch. However, updating the class means in each batch by using entire training set is impractical. Therefore, we update the class centers by using the selected samples in the current batch by using a similar procedure described in \cite{R1} (see supplementary material for details).
\section{Experiments}
We tested the proposed method, Deep Compact Extended Polyhedral Conic Classifier Network (DC-EPCC)\footnote{Source code of the proposed method is provided as supplementary material. The codes and trained models will be made publicly available.}, on both closed and open set recognition problems. We compared the proposed method to the Deep EPCC method of \cite{R19} and other related methods using soft-max loss and center loss functions \cite{R1,R2} as well as SphereFace \cite{R14}, CosFace \cite{R16}, and ArcFace \cite{R18} methods. We used ResNet-50 architecture as backbone architecture for all methods unless otherwise stated, thus they are all directly comparable.
\subsection{Illustrations on Mnist and FaceScrub Datasets}
For the first experiment, we designed a deep neural network where the output of the last hidden layer is set to 2 for visualizing the learned features as in \cite{R1}. This allows to plot the learned feature vectors in 2D space. To visualize the control over the compactness of the classes by changing $\kappa$ values, we selected one class (the digit 2) from the Mnist dataset, and trained a binary deep neural network where the positive class is the selected digit class and the negative class is the remaining classes in the dataset. The learned feature representations of the positive class samples (shown with orange color) and the corresponding acceptance regions for different $\kappa$ values are plotted in Fig. \ref{fig:3}. As seen in the figure, if we do not enforce the compactness constraints on the learned weights (i.e., $\kappa=0.0$), the resulting positive class acceptance region is not necessarily bounded. As we increase the value of the $\kappa$, the area of the acceptance regions decreases, which in turn yields more compact positive class acceptance regions.
Moreover, although we do not explicitly enforce the class samples to collapse onto a single point as in center loss, positive class samples mostly collapse onto a single point, which is the ultimate goal in both center loss \cite{R1} and \textit{metric learning by collapsing classes} method introduced in \cite{R31}.
\begin{figure*}[tbh]
\centering
\begin{tabular}[h]{@{}ccc@{}}
\includegraphics[width=42mm]{gfx/deep_compact_epcc_new.png} &
\includegraphics[width=42mm]{gfx/deep_epcc.png} & \includegraphics[width=42mm]{gfx/softmax_new.png} \\
\includegraphics[width=42mm]{gfx/centerloss_new.png} &
\includegraphics[width=42mm]{gfx/sphereface_new.png} & \includegraphics[width=42mm]{gfx/arcface_new.png}
\end{tabular}
\caption{Learned feature representations of face images for different methods. }
\label{fig:4}
\end{figure*}
In the second experiment, we compared the learned feature representations returned by various deep neural networks including the proposed method. To this end, we used FaceScrub dataset. The FaceScrub dataset \cite{R23} includes face images of 530 celebrities. It has been created by detecting faces based on automated search of public figures on the internet followed by manually checking and cleaning the results. In the dataset, there are 265 male and 265 female celebrities’ face images. In our experiments, we used only 4 classes for better visualization. The learned features are plotted in Fig. \ref{fig:4}. We also plotted the feature representations learned by the deep EPCC method of \cite{R19} that uses a common center for all classes and does not enforce any compactness criterion. As seen in the figure, both the proposed method and the method using center loss return compactly clustered feature representations, whereas deep EPCC and soft-max loss fail to return compact class regions. The SphereFace and ArcFace methods enforce compactness in the angular space, thus the features learned by these two methods lie in the vicinity of lines with different orientations. Therefore, the angular orientations of these lines determine the class memberships during testing. We would like to point out that the proposed method returns more compact feature representations with respect to deep EPCC method of \cite{R19} as expected.
\subsection{Experiments on PASCAL VOC 2007 Dataset}
We conducted experiments on the PASCAL 2007 Visual Object Classification dataset. PASCAL VOC dataset includes 20 object categories. We trained a binary classifier for each category where the object class is taken as positive and the remaining classes are treated as negatives, which is the standard classification setup.
The accuracies are computed in terms of average precision by using the VOC evaluation toolkit.
The results are given in Table \ref{tb:pascalvoc}. SphereFace, CosFace and ArcFace method yielded very poor accuracies, thus we omitted these results. As seen in the table, the proposed DC-EPCC method significantly outperforms CNN networks using both soft-max and center loss functions. The proposed method also significantly outperforms Deep EPCC method of \cite{R19}. Our proposed method achieves the best accuracies for 15 categories of all 20 object categories, and it outperforms the second best performing method by 2\% on the average. The CNN method using the center loss wins for the remaining 5 object categories, and the CNN network using the soft-max is the worst performing method. Deep EPCC method of \cite{R19} also outperforms the CNN method using soft-max loss, but its accuracy is lower than the accuracies of the proposed method and the method using center loss. It should be noted that both deep EPCC methods use positive class center in this setting, thus the only difference between the two methods is that we enforce the class compactness in the proposed methodology. Yet, the accuracy difference is quite big, 3.7\%, which shows the importance of compact acceptance regions.
Obtaining best accuracies by the proposed method and center loss is mostly expected since both methods return compact class acceptance regions for the tested positive object classes and they easily reject the samples coming from the background classes. However, a complete failure of the methods maximizing margin in the angular space was quite unexpected.
\begin{table*} [tb]
\begin{center}
\renewcommand{\arraystretch}{1.2}
\caption{Average Precision scores (\%) on PASCAL VOC 2007 classification dataset.}
\label{tb:pascalvoc}
\def\S#1{\begin{sideways}#1\end{sideways}}
\def\B#1{\textbf{#1}}
\small
\tabcolsep 2.2pt
\resizebox{\textwidth}{!}{
\begin{tabular} {|@{~\strut}l@{~}*{21}{|c}|} \hline
\B{Methods} &
\S{Aeroplane}&\S{Bicycle}&\S{Bird}&\S{Boat}&\S{Bottle}&\S{Bus}&\S{Car}&\S{Cat}&\S{Chair}&\S{Cow}&\S{Dining Table~~}&\S{Dog}&\S{Horse}&\S{Motorbike}&\S{Person}&\S{Potted Plant}&\S{Sheep}&\S{Sofa}&\S{Train}& \S{TV Monitor} & \S{Average}\\ \hline
{DC-EPCC} & 93.9 & \textbf{89.2} & \textbf{90.5} & 90.2 & \textbf{51.9} & \textbf{82.5} & \textbf{91.1} & 87.8 & \textbf{69.5} & \textbf{77.0} & \textbf{75.7} & \textbf{85.0} & \textbf{91.8} & \textbf{86.8} & 94.2 & \textbf{61.4} & \textbf{81.2} & \textbf{73.1} & 94.7 & \textbf{82.4} & \textbf{82.5} \\ \hline
{Deep EPCC \cite{R19}} & 93.2 & 85.6 & 86.3 & 87.4 & 44.1 & 79.9 & 88.0 & 84.6 & 66.9 & 69.5 & 72.0 & 79.8 & 91.6 & 85.0 & 94.1 & 58.8 & 79.2 & 67.5 & 94.4 & 73.1 & 78.8 \\ \hline
{Soft-max} & 91.9 & 85.0 & 84.7 & 86.1 & 40.9 & 73.5 & 88.3 & 80.3 & 65.0 & 60.9 & 64.5 & 75.6 & 88.4 & 81.0 & 91.4 & 49.3 & 66.7 & 63.3 & 90.7 & 72.0 & 75.0 \\ \hline
{Center Loss \cite{R1}} & \textbf{94.3} & 88.0 & 88.0 & \textbf{90.9} & 48.0 & 78.9 & 90.7 & \textbf{88.3} & 68.5 & 71.1 & 74.1 & 84.2 & 90.8 & 85.7 & \textbf{94.5} & 57.0 & 75.3 & 68.4 & \textbf{94.8} & 78.4 & 80.5 \\ \hline
\end{tabular}
}
\end{center}
\end{table*}
\subsection{Experiments on CIFAR-100 and FaceScrub Datasets}
Here we tested the proposed method on CIFAR-100 and FaceScrub datasets. We fine-tuned all tested networks from a network trained on ILSVRC2012 for CIFAR-100 dataset experiment and a network trained on MS-Celeb-1M dataset \cite{R32} is used to fine the CNN networks for FaceScrub dataset experiment. The accuracies of methods tested on CIFAR-100 dataset are given in terms of classical classification rates in Table \ref{cifar100_res}. The proposed DC-EPCC method achieves the best accuracy followed by the CNN method using center loss. Methods maximizing the margin in the angular space, SphereFace, CosFace and ArcFace, perform badly and give the worst accuracies. This shows that general object categories cannot be modeled by subspaces as in face recognition. Deep EPCC method of \cite{R19} also performs poorly mostly due to the fact that using a common single center for all 100 classes restricts to return well localized acceptance regions as illustrated in Fig. \ref{fig:2}.
\begin{table}[t]
\begin{center}
\renewcommand{\arraystretch}{1.3}
\caption{Classification Rates (\%) on the CIFAR-100 and FaceScrub datasets.}
\label{cifar100_res}
\vspace{0.3cm}
\begin{tabular}[tb] {|l|c|c|} \hline
\textbf{Method} & \textbf{CIFAR-100} & \textbf{FaceScrub}\\ \hline
DC-EPCC & \textbf{83.17} & 98.19 \\ \hline
Deep EPCC \cite{R19} & 78.06 & 98.18 \\ \hline
Soft-max & 81.88 & 96.02 \\ \hline
Center Loss \cite{R1} & 82.19 & \textbf{98.99} \\ \hline
SphereFace \cite{R14} & 75.19 & 97.50 \\ \hline
CosFace \cite{R16} & 78.85 & 97.97\\ \hline
ArcFace \cite{R18} & 77.74 & 97.65 \\ \hline
\end{tabular}
\end{center}
\vskip-10pt
\end{table}
The FaceScrub dataset \cite{R23} includes face images of 530 celebrities. In this dataset, there are 265 male and 265 female celebrities' face images.
We manually checked the face images and cleaned non-face images since there were still some annotation mistakes. As a result, we had 67,437 face images of 530 celebrities with an average of 127 images (minimum 39, maximum 201) per person. The face images are mostly high resolution frontal face images and we resized them to 112$\times$112.
We used 60\% of data as training, and the remaining 40\% is used for testing. The classification accuracies of tested methods are given
in Table \ref{cifar100_res}. The best accuracy is obtained by CNN method using center loss and it is closely followed by the proposed method. As opposed to CIFAR-100 results, methods maximizing the angular margin also perform well here. The CNN method using the soft-max is the worst performing method.
\subsection{Experimental Results on Open Set Recognition}
For open set recognition, we follow the settings used in \cite{R38,R39} and evaluate open set recognition performance on SVHN \cite{R42} and CIFAR-10 datasets. In this setting, samples selected from randomly chosen 6 classes are used for training, and test set uses the samples coming from all 10 classes. The goal is to detect the novel samples coming from the unknown 4 classes during testing. This procedure is repeated 5 times and average of the computed accuracies in each run is used for final accuracy. Area under the ROC (Receiver Operating Characteristic) curve (AU-ROC) is used to evaluate the performance. In the proposed method, we first scale classifiers' scores by dividing the maximum of the absolute values for each class such that scores for a particular class will be between -1 and 1 (we also applied sigmoid function to convert the scores to probabilities, the results were similar). Then, the negative of the maximum score is used as the final score value for open-set recognition, i.e., $-\underset {c=1,...,C} {\text{max}} \: s_c(\mathbf{f}_{test})$, where $s_c(\mathbf{f}_{test})$ denotes the classification score of the test sample for the class $c$.
\begin{table}[t]
\begin{center}
\renewcommand{\arraystretch}{1.3}
\caption{AU-ROC Scores (\%) for open-set recognition. Results are averaged over 5 random partitions of known/open set classes. All the values in the table other than the proposed method and Deep EPCC are taken from \cite{R38} and \cite{R39}.}
\label{openset_res}
\vspace{0.3cm}
\begin{tabular}[tb] {|l|c|c|} \hline
\textbf{Method} & \textbf{CIFAR-10} & \textbf{SVHN} \\ \hline
DC-EPCC & 87.7 & \textbf{93.2} \\ \hline
Deep EPCC \cite{R19} & 84.8 & 92.0\\ \hline
Soft-max & 67.7 & 88.6 \\ \hline
OpenMax \cite{R33} & 69.5 & 89.4\\ \hline
G-OpenMax \cite{R41} & 67.5 & 89.6 \\ \hline
OSRCI \cite{R38} & 69.9 & 91.0 \\ \hline
C2AE \cite{R39} & \textbf{89.5} & 92.2 \\ \hline
\end{tabular}
\end{center}
\vskip-10pt
\end{table}
Computed accuracies are given in Table \ref{openset_res}. All the values in the table other than the proposed method and Deep EPCC are taken from \cite{R38} and \cite{R39}. As seen in the table, the proposed method achieves the best accuracy for SVHN dataset and obtains a comparable result on the CIFAR-10 dataset. The results on CIFAR-10 datasets are very promising in the sense that the proposed method and C2AE significantly outperform other open set recognition methods.
It should be noted that, we do not attempt to do any special operation for the open set recognition such as creating synthetic open set examples as in \cite{R38} and \cite{R33}. We train our classifiers as usual as in closed set recognition settings. Yet, we significantly outperform many of the state-of-the-art open set recognition algorithms as seen in the table.
Obtaining good accuracies by the proposed method is expected since the proposed method returns very compact acceptance regions for classes. The samples in the center of these regions have the higher confidence scores, and the scores decrease as the samples move away from the center. Therefore, the samples coming from different classes lie far from these acceptance regions and they have very low confidence scores. As a result, they are easily rejected, which in turn improves the AU-ROC score.
\begin{table}[t!]
\begin{center}
\renewcommand{\arraystretch}{1.2}
\caption{Performance Comparison of various approaches for different views and modalities on the DAD dataset.}
\label{dad_res}
\tabcolsep 1.5pt
\begin{tabular}[tb] {|l|c|c|c|c|} \hline
\multicolumn{1}{|c|}{\multirow{2}{*}{\textbf{Method}}} & \multicolumn{4}{c|}{\textbf{AU-ROC}} \\ \cline{2-5}
& \phantom{i} \textbf{Top-D} \phantom{i} &\phantom{i} \textbf{Top-IR} \phantom{i}& \phantom{i} \textbf{Front-D} \phantom{i} & \textbf{Front-IR} \\ \hline
DC-EPCC & 0.9012 & \textbf{0.8954} & \textbf{0.9087} & \textbf{0.8727} \\ \hline
Deep EPCC \cite{R19} & 0.8859 & 0.8665 & 0.8832 & 0.8715 \\ \hline
CL \cite{R43} & \textbf{0.9200} & 0.8857 & 0.9020 & 0.8666 \\ \hline
Deep SVDD \cite{R34} & 0.5453 & 0.5881 & 0.6509 & 0.6434 \\ \hline
\end{tabular}
\end{center}
\vskip-10pt
\end{table}
\subsection{Experiments on Anomaly Detection Task}
For anomaly detection task, we use recently introduced Driver Anomaly Detection (DAD) dataset \cite{R43}. The DAD dataset consists of 550 minutes of normal driving videos and 100 minutes of anomalous driving containing 8 different distracting actions in its training set. In the test set of the DAD dataset, there are 88 minutes of normal driving videos and 45 minutes of anomalous driving videos containing 24 different distracting actions. The dataset is recorded with 45 fps for two views (front and top) and two modalities (depth and infrared). The task here is to generalize normal driving and detect anomalies although some of the distracting actions has never been introduced to the network at the training.
We have compared our approach with several anomaly detection methods. In order to make fair comparison, all the models use 3D ResNet-18 network that is pretrained on Kinetics-600 dataset. All methods receive 16 frames input downsampled from 32 frames. Comparative results can be seen in Table \ref{dad_res}. The proposed method achieves the best performance for front view with both modalities and top view with infrared modalities. For only top view depth modality, contrastive learning approach used in \cite{R43} achieves the best performance. We must also note that Deep SVDD \cite{R34} method performs significantly inferior compared to other methods. The reason is that this method only uses normal driving videos at the training time and does not leverage anomalous driving samples in the training set (although the classical SVDD method of \cite{R35} allows the negative samples for learning, the deep neural network version does not have this property). This shows the importance of the anomalous samples at the training time to generalize normal driving.
We have also visualized the generalization perfromance of the proposed DC-EPCC method on a continuous video stream from the DAD dataset validation set in Fig. \ref{fig:DAD_scores}. DC-EPCC can successfully produce high scores for normal driving and low scores for anomalous actions. Therefore, utilization of a preset threshold is enough for the detection of anomalies. Although there are some fluctuations at the transition from normal to anomalous (and vice versa), they can be mitigated by applying some post processing similar to \cite{R43}.
\begin{figure}[t]
\centering
\includegraphics[width=85mm]{gfx/dcepcc_scores.png}
\caption{Illustration of recognition for a continuous video stream using front view and dept modality. The frames are classified as anomalous driving if the similarity score is blow the preset threshold. Ground truth annotation is shown on top.}
\label{fig:DAD_scores}
\end{figure}
\section{Conclusion}
This paper introduces a new deep neural network classifier that simultaneously minimizes the intra-class variations and maximizes the inter-class separability. To this end, the proposed method uses a margin based cost function for maximizing the inter-class separation, and a geometrically inspired mechanism that controls the compactness of the class acceptance regions to minimize the intra-class variation. The experimental results on various datasets show that the proposed method outperforms related state-of-the-art deep neural network classifiers on majority of the tested datasets. Especially, the proposed classifier is more suitable for open set recognition tasks and binary classification problems in which one positive object class is separated from a background class composed of many classes other than the object of interest as in PASCAL VOC classification setting.
Moreover, the proposed method achieves the state-of-the-art accuracies on anomaly detection and significantly outperforms state-of-the-art method, Deep SVDD, proposed for anomaly detection.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
Optimal designs are a class of experimental designs that are optimal with respect to some statistical criteria such as minimizing the variance of best linear unbiased estimators in regression problems and maximizing the amount of information obtained from the experiment.
It is desirable to design experiments that provide more information and reduce the uncertainty relating to the goal of the study.
In regression problems, we model the responses of a random experiment, denoted as $Y_1, \ldots, Y_N$, whose inputs are represented by a vector $\bx_i \in \mathcal{X}$ with respect to some known regression functions $\bff(\bx_1),\ldots,\bff(\bx_N)\in \mathbb{R}^{p}$, i.e.,
$$Y(\bx_i)=\bff^{T}(\bx_i)\bbeta+\varepsilon,i=1,2, \ldots,N,$$
where $\bff(\bx)\in \mathbb{R}^{p}$ is the covariates which is independent variable (regressor) associated with $\bx$, the vector $\bbeta \in \mathbb{R}^{p}$ is $p$-dimensional parameter vector, and $\varepsilon$ is the error term with $E[\varepsilon] = 0$ and $Var[\varepsilon]=\sigma^{2} > 0$. For different independent trials, the errors are assumed to be uncorrelated and independent, and the inputs $\bx_{i}$ (a candidate set of design points) are chosen by the experimenter in the design space $\mathcal{X}$. We assume that the model is non-singular in the sense that $\{\bff(\bx):\bx\in \mathcal{X}\}$ spans $\mathbb{R}^{p}$. We wish to pick a small subset of the input vector $\bx_i$ such that querying the corresponding responses will lead to a good estimator of the model. In this paper, we discuss the computation of approximate optimal designs for regression models with uncorrelated errors
\citep[see, for example,][]{Fedorov,Harman2018,Atkinson}.
Assume that the inputs $\tilde{\bx}_{i}$, for $i = 1, 2, \ldots, N$ are chosen within a set of distinct points $\bx_{1},\ldots,\bx_{m}$ with integer $p\leq m \ll N$ (but in some special case $m$ can be equal to $N$, see Setting 5), and let $n_{k}$ denote the number of times the particular points $\bx_{k}$ occurs among $\tilde{\bx}_{1},\ldots,\tilde{\bx}_{N}$, and $\tilde{N}=n_{1}+\cdots+n_{m}$ indicates the number of all candidate experiments. The exact experimental design can be summarized by defining a design $\xi_{\tilde{N}}$ as
\begin{equation}
{\xi}_{\tilde{N}} = \left(\begin{array}{ccc}\bx_{1} & \cdots & \bx_{m} \\ \frac{n_1}{\tilde{N}} & \cdots & \frac{n_{m}}{\tilde{N}}\end{array}\right).
\label{DO1}
\end{equation}
In the design ${\xi}_{\tilde{N}}$ in Eq. (\ref{DO1}), the first row gives the points in the design space $\mathcal{X}$ where the input parameters have to be taken and the second row indicates the proportion of the experimental units assigned or the frequencies of the experiments repeated at these points. Strictly speaking, an exact experimental design $\xi$ of size $\tilde{N}$ can be characterized by a probability distribution on $\mathcal{X}$ in which the probability of $\xi_{\tilde{N}}$ occurs at $\bx_{j}$ is $n_{j}/\tilde{N}$.
If $\xi$ follows a continuous probability distribution or a discrete probability distribution
\begin{equation}
{\xi} = \left(\begin{array}{ccc}\bx_{1} & \cdots & \bx_{m} \\ w_{1} & \cdots & w_{m}\end{array}\right),
\label{DO2}
\end{equation}
where $\Pr(\xi=\bx_{j})=w_{j}, j=1,2,\ldots,m$ and $\sum_{j=1}^{m} w_{j} = 1$, then $\xi$ is a continuous design. The goal of the design of experiment theory is then to pick $m$ out of the given $N$ experiments so as to make the most accurate estimate of the parameter $\bbeta$.
For the review and details related to the determination of optimal experimental designs, the readers can refer to \cite{Dette} and the references therein.
Let $\Xi$ be the set of all exact designs or approximate designs (i.e., probability measures) on the design space $\mathcal{X}$; for a given design $\xi$, we denote the information matrix of $\xi$ for the experimental design by
$$
\bM(\xi)=\sum\limits_{\bx\in\mathcal{X}}\xi(\bx)\bff(\bx)\bff^{T}(\bx) \quad \text{or}\quad \bM(\xi)=\int\limits_{\bx\in\mathcal{X}} \xi(\bx)\bff(\bx)\bff^{T}(\bx)d\xi(\bx)
$$
Based on this formulation, an approximate $D$-optimal design for quadratic polynomial was provided by \cite{Chen} when the design space is a circle and \cite{Duan} provided two efficient computational algorithms for optimal continuous experimental designs for linear models.
Under these model assumptions, the Fisher information matrix corresponding to $\bbeta$ is proportional to the information matrix.
Therefore, to obtain the most accurate estimate of certain parameters, we aim to choose the $\xi$ such that $\bM(\xi)$ is maximized according to some criterion. In the following, we will focus on
a probability measure on $\mathcal{X}$ with support given by the points $\bx_{i}$ and weights $w_{i}$ in Eq. (\ref{DO2}).
In order to obtain the optimal design, a general approach is to consider some generally accepted statistical criteria proposed by \cite{Kiefer1974} namely the $\Phi_q$-criteria.
The $D$-optimality and the $A$-optimality are two of the most commonly used optimality criteria due to their natural statistical interpretations.
It has been shown that the computation of some important prediction-based optimality criteria such as the $I$-optimality criterion \citep{Cook1982,Goos2016} that minimizes the average prediction variance can be converted into the computation of the $A$-optimality \citep[][Section 10.6]{Atkinson}. In particular, $I$-optimal designs on a finite design space can also be computed using the algorithm developed for $A$-optimality. Thus, in this paper, we focus on the $D$- and $A$-optimal designs where the objective functions are in the form of $\Phi_D(M) = \det(\bM)^{-1}$ and $\Phi_A(M) = tr(\bM)^{-1}$, respectively, for any positive definite matrix $\bM$.
The result by \cite{Welch1982} about the NP-hardness of $D$-optimality is only valid for the exact design problem, while in this paper our aim is to develop efficient computational algorithms for searching the solutions $\xi^{*}$ of the optimization problem $\min \log \Phi_D(M(\xi))$ and $\min \log \Phi_A(M(\xi))$ for $D$-optimality and $A$-optimality, respectively.
Optimal design is at the heart of statistical planning and inference using linear models \citep[see, for example,][]{Box1978}. The theory of optimal designs and the development of numerical computational algorithms for obtaining optimal designs have long been studied in the literature under different scenarios. For instance, \cite{Meyer1995} proposed the coordinate exchange algorithm to construct $D$-optimal and linear-optimal experimental designs for exact design. The algorithm uses a variant of the Gauss-Southwell cyclic coordinate-descent algorithm within the $K$-exchange algorithm to achieve substantive reductions in required computing.
\cite{Gao} developed a general class of the multiplicative algorithms for continuous designs, which can be used to obtain optimal allocation for a general regression model subject to the $D$- and $A$-optimal criteria.
For continuous experimental designs, in general, the continuous factors are generated by the vector $\bff(\bx)$ of linearly independent regular functions where the design points $\bx$ filling the design space $\mathcal{X}$. Then, to choose the optimal design points that maximize the information matrix.
There are many analytical methods for obtaining the approximate optimal designs. \cite{Kiefer1959} introduced the equivalence principle and propose in some cases algorithms to solve the optimization problem. Following the early works of \cite{Karlin1966a}, the case of polynomial regression on a compact interval on $\mathbb{R}$ has been widely studied.
The well-known equivalence theorem of \cite{Kiefer1959} led to the development of a practical algorithm called vertex direction methods (VDMs) for the construction of a $D$-optimal design \citep{Fedorov,Wynn}. They also proved the convergence of the sequence to an optimal (in the appropriate sense) design.
\cite{Silvey} proposed a multiplicative algorithm (MUL) for optimal designs on finite design space, of which the analog in the mixture setting with finite, fixed support is an EM algorithm \citep{Dempster1977}. The VDMs and MUL algorithms all are based on the techniques from differentiable optimization. The general idea is to use directional derivatives to find a direction of improvement, and then employ a line search to determine an optimal step length.
\cite{Yu} proposed the cocktail algorithm, which actually is a mixture of multiplicative, vertex-exchange, and VDM algorithms for $D$-optimum design; it includes a nearest-neighbor exchange strategy that helps to apportion weights between adjacent points and has the property that poor support points are quickly removed from the total support points.
\cite{Harman2018} considered an extension and combination of both the VEM algorithm and the $KL$-exchange algorithm that is used to compute exact designs \citep{Atkinson} and developed the randomized exchange method (REX) for the optimal design problem.
Recent progress in this area has been obtained by employing hybrid methods that alternate between steps of the cocktail algorithm, or by using the randomized exchange method. Following the work of \cite{Gao}, \cite{Duan} proposed an efficient computational algorithm for computing continuous optimal experimental designs for linear models.
In this paper, we aim to propose a computational algorithm to obtain approximate $D$-optimal designs and a computational algorithm to obtain approximate $A$-optimal designs on any compact design spaces.
This paper is organized as follows. The statistical inference based on a regression model along with the form of an information matrix and variance-covariance matrix for the model parameters are presented in Section 2.
After a review of the $D$- and $A$-optimal criteria, the proposed algorithms and the theoretical results related to the convergence and monotonicity of the proposed algorithms are also presented in Section 2.
Section 3 presents some numerical illustrations with several linear regression models on different types of design spaces which are more general in practical applications for $D$-optimality and $A$-optimality designs.
The proofs of the main results are presented in the Appendix.
\section{Algorithms for Approximate Optimal Designs}
In this section, we introduce the method for searching for optimal designs when regression analysis is used. We focus on the numerical computation of approximate $D$- and $A$-optimal designs. For notation simplicity, we denote $\bff_{i} = \bff(\bx_{i})$, $w_{i}=w(\bx_{i})$, $y_{i}=y(\bx_{i})$ in the following.
Consider the linear regression model
\begin{equation}
Y(\bx)=\bff^{T}(\bx)\bbeta+\varepsilon,\;\bx\in \mathcal{X},
\label{MD}
\end{equation}
where $\bff(\bx)$ is the covariates, $\bbeta $ is a $p$-dimensional parameter vector, $\mathcal{X}$ is the design space and $\varepsilon$ is the error term with mean $0$ and variance $\sigma^2$. When the observations $(\bx, \by = y(\bx))$ are obtained based on the model in Eq. (\ref{MD}), the ordinary least squares estimator of $\bbeta$ can be expressed as
$$
{\hbbeta}=\left[\sum\limits_{i=1}^{N} w_{i}\bff_i\bff_i^{T}\right]^{-1}\sum\limits_{i=1}^{N} w_{i}y_i\bff_i,
$$
where $w_i\geq 0$ is the mass on the point $\bx_i,\;i=1, 2, \ldots,N$ and $\sum_{i=1}^{N} w_i=1$.
The variance of $\hbbeta$ can be obtained as
$$
Var(\hbbeta)=\left[\sum\limits_{i=1}^{N} w_{i}\bff_i\bff_i^{T}\right]^{-1}\sigma^2.
$$
Most of the existing computational algorithms for obtaining optimal designs discretize the underlying continuous space by considering a finite design space $\mathcal{X}=\{\bx_{1}, \cdots, \bx_{N}\} \subset \mathbb{R}^{q}$. These existing algorithms rely on either complex algorithms or advanced mathematical programming solvers. Here, we proposed algorithms that are simple yet effective in obtaining the optimal design for $D$-optimality and $A$-optimality without relying on other complex algorithms or advanced mathematical programming solvers.
\subsection{Algorithms for $D$-optimal Designs}
In an experiment, researchers often wish to estimate the model parameters with the highest precision.
One of the commonly used optimality criteria in experimental design is the $D$-optimal design which maximizes the determinant of the Fisher information matrix, which results in minimum volume for the Wald-type joint confidence region for the model parameters if the variance is known \citep{Gilmour2012}.
Specifically, the $D$-optimal design maximizes the log-determinant of the information matrix, i.e., it minimizes the log-determinant of the asymptotic variance-covariance matrix $Var(\hat\bbeta)$. In other words, the $D$-optimality criterion results in minimizing the generalized variance of the parameter estimates. The $D$-optimal criterion can be described as follows.
\par
\par
\noindent
\textbf{$D$-optimal criterion:}
\begin{equation}
\min\limits_{w_{1},\cdots,w_{N}}\left\{-\log|\sum\limits_{i=1}^{N}w_{i}\bff_{i}\bff_{i}^{T}|:\;\mbox{subject to}\;w_{i}\geq 0\;\mbox{and}\;\sum\limits_{i=1}^{N}w_{i}=1,\; i=1,2,\ldots,N\right\},
\label{DO}
\end{equation}
\par
For $D$-optimality, we can obtain the following result.
\par
\noindent
{\bf Theorem 1.} $\bw^{*}$ is the $D$-optimal solution for Eq. (\ref{DO}) if and only if
$$
\sum\limits_{i=1}^{N} w_{i}\bff^{T}(\bx_{i})\left[\sum\limits_{j=1}^{N} w^{*}_{j}\bff(\bx_{j})\bff^{T}(\bx_{j})\right]^{-1}\bff(\bx_{i})\leq p
$$
for $w_{i}\geq0$ and $\sum\limits_{i=1}^{N} w_{i} = 1,\;i=1,\cdots,N$.
Theorem 1 is a special case of a part of the general equivalence theorem, and the detailed proof is provided in the Appendix.
For the $D$-optimal criterion in Eq. ($\ref{DO}$), we propose the following algorithm to obtain the optimal choice of $\bw^{*}$ based on Theorem 1.
\let\AND\undefined
\begin{algorithm}[H]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\caption{Algorithm for $D$-optimal design}
\label{alg1}
\begin{algorithmic}[1]
\REQUIRE Regressor $\bff(\bx)\in R^{p}$, design space $\mathcal{X}$, stopping parameter $\gamma$ and tuning parameter $\delta$.\\
\ENSURE Approximate design points $\bx_{s}^{*}$ and corresponding weight $w_{s}^{*}$, $s=1,\cdots,k$.\\
\STATE Generate random design points $\bx_{i} \in \mathcal{X}$ and corresponding starting weights $w^{(0)}_{i}$, $i=1,\cdots,N$.\\
\STATE compute the regressors $\bff_{1},\cdots,\bff_{N}\in R^{p}$.\\
\REPEAT
\STATE $$ w^{(h)}_{i}=\frac{w^{(h-1)}_{i}\bff^{T}_{i}D^{(h-1)}\bff_{i}}{p},i=1,\cdots,N,$$
where
$$ D^{(h-1)}=\left[\sum_{i=1}^{N} w^{(h-1)}_{i}\bff_{i}\bff^{T}_{i}\right]^{-1}.$$\\
\UNTIL $\sum_{i=1}^{N} |w_{i}^{(h)} - w_{i}^{(h-1)} | < \gamma$. \\
\STATE find $w_{i}>\delta$ and corresponding design points $\bx_{i}$, $i=1,\cdots,N$, and write them as $\bx^{*}_{s}$, $s=1,\cdots,k$. \\
\STATE Let $\bx^{*}_{s}$, $s=1,\cdots,k$ as the new design points and repeat Step 2--5.\\
\STATE Output the optimal design points $\bx^{*}_{s}$ and corresponding weights $w^{*}_{s}$, $s=1,\cdots,k$.
\end{algorithmic}
\end{algorithm}
In Algorithm 1, $\bff(\bx)$ is the functional form of the regressors, the stopping parameter $\gamma$ determines the stopping criteria of the algorithm and the tuning parameter $\delta$ is used to choose the design points with weights which are larger than $\delta$. Note that the choice of $\delta$ must such that $k\geq p$, which is crucial for guaranteeing the non-singularity of the information matrix. In general, the experimenter only needs to set the values of $\gamma$ and $\delta$ to be very small. In Section 3, we set $\delta=0.0001$ and $\gamma=0.0005$.
\noindent
{\bf Remark 1.} Algorithm 1 is efficient even when the sample size $N$ is very large because the iteration process relies on the weight of each sample, but it does not rely on the weights of other samples in a particular iteration. Hence, one can use a parallel strategy to speed up the computations required for Algorithm 1.
Most of the existing algorithms do not share this advantage because the design points interact with each other. For example, the VDM algorithm is based on the differentiable optimization techniques in which the basic idea is to move the current design point $\bx$ to the direction of some other design points while decreasing all components of $\bx$. In addition, the algorithms derived from the VDM algorithm are all depending on the sample size $N$. Therefore, these algorithms will suffer from losing efficiency when the sample size $N$ is large.
In the following, we present proof of the convergence of Algorithm 1 for $D$-optimality. To prove the convergence of the proposed algorithm, we first show that the log-determinant of the information matrix is monotonic. To prove the convergence of the proposed algorithm, we need to add the bounded assumption and require the following two lemmas.\\
\noindent
{\bf Lemma 1.} Let $A(\bx)$ be a nonnegative definite matrix function on $\mathcal{X}$, \linebreak $\bw(\bx)=(w(\bx_1),w(\bx_2),\cdots,w(\bx_N))$ and $\tilde{\bw}(\bx)=(\tilde{w}(\bx_1),\tilde{w}(\bx_2),\cdots,\tilde{w}(\bx_N))$ are two probability vectors in $R^{N}$, and
$$\sum\limits_{i=1}^{N} w(\bx_{i}) A(\bx_{i}) \;\;\mbox{and}\;\;\sum\limits_{i=1}^{N} \tilde{w}(\bx_{i}) A(\bx_{i}) $$
are positive definite matrices. Then,
\begin{eqnarray*}
& & \log \left|\sum\limits_{i=1}^{N} w(\bx_{i}) A(\bx_{i})\right|-\log \left|\sum\limits_{i=1}^{N} \tilde{w}(\bx) A(\bx_{i}) \right| \\
& \geq & \sum\limits_{i=1}^{N} \tilde{w}(\bx_{i})\mbox{tr}\left\{A(\bx_{i})\left[\sum\limits_{j=1}^{N} \tilde{w}(\bx_{j}) A(\bx_{j}) \right]^{-1}\right\}\log\frac{w(\bx_{i})}{\tilde{w}(\bx_{i})}.
\end{eqnarray*}
\noindent
\textbf{Proof.} Following the proof of Lemma 1 in \cite{Duan} and Lemma 2 in \cite{Gao}, the results in Lemma 1 can be obtained.\\
\noindent
{\bf Lemma 2.} Suppose $\bw(\bx)=(w(\bx_1),w(\bx_2),\cdots,w(\bx_N))$ and \linebreak $\tilde{\bw}(\bx)=(\tilde{w}(\bx_1),\tilde{w}(\bx_2),\cdots,\tilde{w}(\bx_N))$ satisfy $\sum_{i=1}^{N}w(\bx_i)=\sum_{i=1}^{N}\tilde{w}(\bx_i)=1$ are two probability vectors in $R^{N}$, then
$$
\sum\limits_{i=1}^{N}|w(\bx_i)-\tilde{w}(\bx_i)|\leq\left[2\sum\limits_{i=1}^{N} w(\bx_i)\log\frac{w(\bx_i)}{\tilde{w}(\bx_i)}\right]^{1/2}.
$$
\textbf{Proof.} See \cite{Kullback}. \\
The following theorem shows the convergence of the proposed algorithm for $D$-optimality.\\
\noindent {\bf Theorem 2.} Under the assumption that $\log \left|\sum\limits_{i=1}^{N}\bff(\bx_{i})\bff^{T}(\bx_{i}) \right|$ is bounded, we have
$$
\sum\limits_{i=1}^{N} |w^{(n)}(\bx_{i})-w^{(n-1)}(\bx_{i})|\longrightarrow 0 {\mbox { as }} n \longrightarrow +\infty.
$$
The proof of Theorem 2 is provided in the Appendix.
The algorithm proposed here can be considered as a member of the general class of multiplicative algorithms \citep{Silvey}. Hence, the proposed algorithm shares the simplicity and monotonic convergence property of the class of multiplicative algorithms, and the convergence rate does not depend on $N$ compared to some exact algorithms such as the coordinate-exchange algorithm \citep{Meyer1995}.
Now, we provide a theorem to show that the $D$-optimal approximate design converges to the continuous $D$-optimal design \citep[see, for example,][]{Duan} under certain conditions.
Here, a design is approximate if it is a discrete probability measure and a design is continuous if it is a probability measure with a density with respect to the Lebesgue measure on the observation domain.
Assume that $w^{opt}$ is the $D$-optimal design on the design space $\mathcal{X}$, and it is a continuous probability distribution.\\
\noindent {\bf Theorem 3.} Assume that the random sample points $\bx_i, i=1, 2, \ldots, N$ are generated according to $g(\bx)$ on the design space $\mathcal{X}$ and $g(\bx) > \epsilon > 0$ for all $\bx \in \mathcal{X}$. If $\int_\mathcal{X} \bff^{T}(\bx)\bff(\bx)d\bx$ is bounded on $\mathcal{X}$, then
$$\det\sum\limits_{i=1}^{N} w^*(\bx_i)\bff(\bx_i)\bff^{T}(\bx_i) -\det\int_\mathcal{X} w^{opt}(\bx)\bff(\bx)\bff^{T}(\bx) d\bx\rightarrow 0$$
in probability when $N\to\infty$.\\
\noindent{\bf Proof.} First, we have
\begin{eqnarray*}
\int_\mathcal{X} w^{opt}(\bx) \bff(\bx)\bff^{T}(\bx)d\bx
&=&\int_\mathcal{X} \frac{w^{opt}(\bx)}{g(\bx)} \bff(\bx)\bff^{T}(\bx) g(\bx)d\bx\\
&=&\sum\limits_{i=1}^{N} \frac{w^{opt}(\bx_i)}{Ng(\bx_i)} \bff(\bx_i)\bff^{T}(\bx_i)+\bc_N,
\end{eqnarray*}
where $\bc_N=O_p(N^{-1/2})$. We also have
$0\leq w^{opt}(\bx_i)/\{Ng(\bx_i)\}\leq C/(N\epsilon)<1$
for $N$ sufficiently large.
In addition,
\begin{eqnarray*}
K_N &\equiv& \sum\limits_{i=1}^{N} w^{opt}(\bx_i)/\{Ng(\bx_i)\}\\
&=&\int_\mathcal{X} w^{opt}/g(\bx) g(\bx)d\bx+d_N = 1+d_N,
\end{eqnarray*}
where $d_N=O_p(N^{-1/2})$. Note that $w^{opt}(\bx_i)/\{NK_Ng(\bx_i)\}, i=1, 2, \ldots, N$ are valid weights.
Write $\sum\limits_{i=1}^{N} w^*(\bx_i)\bff(\bx_i)\bff^{T}(\bx_i) = \int_\mathcal{X} w_N(\bx)\bff(\bx)\bff^{T}(\bx) d\mu(\bx)$, where
$w_N(\bx)=w^*(\bx_i)$ if $\bx\in\{\bx_1, \bx_2, \ldots, \bx_N\}$ and $w_N(\bx)=0$
otherwise. Then, because $w^{opt}$ is optimal on $\mathcal{X}$ among all possible bounded probability density functions and probability mass functions on $\mathcal{X}$, we can obtain
\begin{equation}
\det \int_\mathcal{X} w^{opt}(\bx)\bff(\bx)\bff^{T}(\bx) d\bx
\ge \det\sum\limits_{i=1}^{N} w^*(\bx_i)\bff(\bx_i)\bff^{T}(\bx_i).
\label{eq:relation1}
\end{equation}
Given $\bx_i, i = 1, 2, \ldots, N$, $w^*(\bx_i)$ is
optimal, we have
\begin{equation}
\det \sum\limits_{i=1}^{N} \frac{w^{opt}(\bx_i)}{NK_Ng(\bx_i)} \bff(\bx_i)\bff^{T}(\bx_i)
\le \det \sum\limits_{i=1}^{N} w^*(\bx_i)\bff(\bx_i)\bff^{T}(\bx_i).
\label{eq:relation2}
\end{equation}
Combining Eqs. (\ref{eq:relation1}) and (\ref{eq:relation2}),
we have
\begin{eqnarray}
\nonumber
\det \int_\mathcal{X} w^{opt}(\bx)\bff(\bx)\bff^{T}(\bx) d\bx &\geq& \det \sum\limits_{i=1}^{N} w^*(\bx_i)\bff(\bx_i)\bff^{T}(\bx_i)\\
\nonumber
&\geq& \det \frac{1}{K_N}\sum\limits_{i=1}^{N} \frac{w^{opt}(\bx_i)}{Ng(\bx_i)} \bff(\bx_i)\bff^{T}(\bx_i)\\
&=&\det \frac{1}{1+d_N}\left\{ \int_\mathcal{X} w^{opt}(\bx)\bff(\bx)\bff^{T}(\bx) d\bx-\bc_N\right\}.
\label{eq:combined}
\end{eqnarray}
Because $\bc_N\rightarrow\0$ and $d_N\rightarrow0$ in probability when $N\rightarrow\infty$, Eq. (\ref{eq:combined}) leads to
\begin{eqnarray*}
&&\det\int_\mathcal{X} w_N(\bx)\bff(\bx)\bff^{T}(\bx) d\mu(\bx)
-\det\int_\mathcal{X} w^{opt}(\bx)\bff(\bx)\bff^{T}(\bx) d\bx\\
&=&\det\sum\limits_{i=1}^{N} w^*(\bx_i)\bff(\bx_i)\bff^{T}(\bx_i)
-\det\int_\mathcal{X} w^{opt}(\bx)\bff(\bx)\bff^{T}(\bx) d\bx\rightarrow0
\end{eqnarray*}
in probability when $N\rightarrow\infty$. \\
Theorem 3 verifies that the approximate optimal design obtained from the proposed algorithm will eventually converge to the continuous optimal design at the speed of $\sqrt{N}$. Theorem 3 also guarantees the convergence of the proposed algorithm.
\subsection{Algorithms for A-optimal Designs}
Another commonly used optimality criterion is the $A$-optimality criterion that minimizes the trace of the variance-covariance matrix of the maximum likelihood estimates (MLEs). The $A$-optimality criterion provides an overall measure of the variations in the model parameter estimates.
The objective function being minimized in the $A$-optimal design is described as follows. \\
\textbf{$A$-optimal criterion:}
\begin{equation}
\min\limits_{w_1,\cdots,w_{N}} \left\{ {\mbox {tr}} \left( \left[\sum\limits_{i=1}^{N} w_{i}\bff_{i}\bff^{T}_{i} \right]^{-1} \right), \mbox{subject to}\;w_{i}\geq 0\;\mbox{and}\;\sum\limits_{i=1}^{N} w_{i}=1,\; i=1,\cdots,N \right\}.
\label{AOP}
\end{equation}
For $A$-optimality, we can obtain the following theorem. For simplicity, we denote $ I_\mathcal{X}(\bw, \bff) = \sum\limits_{i=1}^{N} w(\bx_{i})\bff_{i}(\bx_{i})\bff_{i}^{T}(\bx_{i})$ and $I_\mathcal{X}(\bw^{*}, \bff)=\sum\limits_{i=1}^{N} w^{*}(\bx_{i})\bff_{i}(\bx_{i})\bff^{T}(\bx_{i})$.
\noindent
{\bf Theorem 4.} $\bw^{*}(\bx)=(w^{*}(\bx_1),\cdots,w^{*}(\bx_N))$ is the $A$-optimal solution for Eq. (\ref{AOP}) if and only if
$$
\frac{ {\mbox {tr}}(\left[I_\mathcal{X}(\bw^{*}, \bff) \right]^{-1} I_\mathcal{X}(\bw, \bff) \left[ I_\mathcal{X}(\bw^{*}, \bff) \right]^{-1})} {{\mbox {tr}}(\left[I_\mathcal{X}(\bw^{*}, \bff)\right]^{-1})} \leq 1,
$$
for $ w(\bx_i)\geq0, \; i=1,2,\ldots,N$ and $\sum\limits_{i=1}^{N} w(\bx_i)=1$.
The proof of Theorem 4 is provided in the Appendix. A similar approach to Theorem 3 can be used to show that the approximate $A$-optimal design converges to the continuous $A$-optimal design. Based on Theorem 4, for the $A$-optimal criteria in Eq. (\ref{AOP}), the following algorithm is proposed to obtain the $A$-optimal design $\bw^{*}(\bx)$.\\
\let\AND\undefined
\begin{algorithm}[H]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\caption{Algorithm for $A$-optimal design}
\label{alg2}
\begin{algorithmic}[1]
\REQUIRE Regressor $\bff(\bx)\in R^{m}$, design space $\mathcal{X}$, stopping parameter $\gamma$ and tuning parameter $\delta$.\\
\ENSURE Approximate design points $\bx_{s}^{*}$ and corresponding weight $w_{s}^{*}$, $s=1,2,\ldots,k$.\\
\STATE Generate a random design points $\bx_{i} \in \mathcal{X}$ and corresponding starting weights $w^{(0)}_{i}$, $i=1,2,\ldots,N$.\\
\STATE compute the regressors $\bff_{1},\bff_{2},\ldots,\bff_{N}\in R^{p}$.\\
\REPEAT
\STATE {$$ w^{(h)}_{i}= w^{(h-1)}_{i} \left[\frac{(p-1)}{p} \frac{tr(D^{(h-1)}\bff_{i}\bff_{i}^{T} D^{(h-1)} )}{tr(D^{(h-1)})} +\frac{1}{p} \right], i=1,2,\ldots,N,$$}
where
$$ D^{(h-1)}=\left[\sum_{i=1}^{N} w^{(h-1)}_{i}\bff_{i}\bff^{T}_{i}\right]^{-1}.$$\\
\UNTIL $\sum_{i=1}^{N} |w_{i}^{(h)} - w_{i}^{(h-1)} | < \gamma$. \\
\STATE find $w_{i}>\delta$ and corresponding design points $\bx_{i}$, $i=1,2,\ldots,N$, and write them as $\bx^{*}_{s}$, $s=1,2,\ldots,k$. \\
\STATE Let $\bx^{*}_{s}$, $s=1,2,\ldots,k$ as the new design points and repeat Step 2--5.\\
\STATE Output the optimal design points $\bx^{*}_{s}$ and corresponding weights $w^{*}_{s}$, $s=1,\cdots,k$.
\end{algorithmic}
\end{algorithm}
In Algorithm 2, $\bff(\bx)$ is the functional form of the regressors, the stopping parameter $\gamma$ determines the stopping criteria of the algorithm and the tuning parameter $\delta$ is to choose the optimal design points with the weight larger than $\delta$. For the same reason as described in Remark 1, a parallel strategy can be used to speed up the computations required for Algorithm 2.
For $A$-optimality, the proposed algorithm provides a convergence solution that is robust to the initial value because the algorithm does not depend on the initial value.
We have attempted to develop the theoretical justification of the convergence of the proposed computational algorithm for $A$-optimality, however, the mathematical justification is not available. Instead, we provide some simulation and numerical results to support the validity and reliability of the proposed algorithm in the subsequent section.
Here, we conjecture the monotonic convergence of the algorithm for $A$-optimality based on the extensive simulation and numerical results.
\section{Numerical Illustrations}
In this section, we consider five different settings \citep{Castro2017} to evaluate the performance of the proposed algorithms for obtaining approximate $D$- and $A$-optimal designs.
We use these numerical examples to illustrate that the proposed algorithms can efficiently identify the optimal design.
For comparative purposes, we also apply the randomized exchange (REX) algorithm proposed by \cite{Harman2018}, the cocktail (CO) algorithm proposed by \cite{Yu}, the vertex direction (VDM) algorithm proposed by \cite{Fedorov} and \cite{Wynn}, and the multiplicative (MUL) algorithm proposed by \cite{Silvey} for computing the $D$- and $A$-optimal designs.
Since the cocktail algorithm is for $D$-optimal only \citep{Yu}, hence, the cocktail algorithm is not applied to obtain the $A$-optimal design.
\subsection{Setting 1: Two-dimensional design space in a square with $p = 6$}
In Setting 1, we consider the model
$$
y(\bx)=\bbeta^{T}\bff(\bx),
$$
where $\bbeta=(\beta_0,\beta_1,\beta_2,\beta_3,\beta_4,\beta_5)^{T}$, $\bff(\bx)=(1,x_1,x_2,x_1^2,x_1x_2,x_2^2)^{T}$ and $\bx\in \mathcal{X}=[-1,1] \times [-1,1]$.
The optimal design points and their corresponding weights obtained from Algorithm \ref{alg1}, Algorithm \ref{alg2} and the REX, CO, VEM and MUL algorithms for $D$- and $A$-optimal designs are presented in Tables \ref{set1_DOP} and \ref{set1_AOP}, respectively. The optimal values of the corresponding objective functions are also presented in Tables \ref{set1_DOP} and \ref{set1_AOP}.
From Tables \ref{set1_DOP} and \ref{set1_AOP}, we observe that the proposed algorithms identify the same optimal design points as the REX, CO, VDM, and MUL algorithms. The weights for optimal points obtained from the REX, CO algorithms and the proposed algorithm are very close, and there is no significant difference between these three algorithms according to the values of the $D$-optimality objective function presented in Table \ref{set1_DOP}. Furthermore, the performance of the VDM and MUL algorithms are not as good as the REX, CO and proposed algorithms. Similar results and conclusions can be observed from Table \ref{set1_AOP} for $A$-optimal designs.
\subsection{Setting 2: Two-dimensional design space in a circle with $p = 6$}
In Setting 2, we consider the model
\begin{eqnarray*}
y(\bx)=\bbeta^{T}\bff(\bx),
\end{eqnarray*}
where $\bbeta=(\beta_0,\beta_1,\beta_2,\beta_3,\beta_4,\beta_5)^{T}$, $\bff(x)=(1,x_1,x_2,x_1^2,x_1x_2,x_2^2)^{T}$ and $\bw\in \mathcal{X}=\{(x_1,x_2)^{T}: \;x_1^2+x_2^2\leq 1\}$.
Figure \ref{set2} presents the $D$-optimal design points in which the weight for the center of the unit circle (0, 0) is 1/6, and the other optimal design points are uniformly distributed on the ring (the vertices of a regular $s$-sided polygon in the circle) with a combined weight $5/6$ in theory. For a more detailed analysis of this setting, one can refer to \cite{Duan}. The optimal design points and their corresponding weights obtained from Algorithm \ref{alg1}, Algorithm \ref{alg2} and the REX, CO, VEM and MUL algorithms for $D$- and $A$-optimal designs are presented in Tables \ref{set2_DOP} and \ref{set2_AOP}. The optimal values of the corresponding objective functions are also presented in Tables \ref{set2_DOP} and \ref{set2_AOP}, respectively.
Note that different algorithms may have different design points because the optimal design points are distributed uniformly on a circle with almost equal weight. Hence, the points may locate at different locations that are symmetrical about the center of the circle.
From Table \ref{set2_DOP}, the REX and CO algorithms, and the proposed algorithm have the same weight as the theoretical value for the center point $(0, 0)$ and the other optimal design points are evenly distributed on the ring with total weight $5/6$, while the VDM and MUL algorithms have distributed the weights for all design points including the points in the center and on the ring.
By comparing the values of the objective function $-\log(|\Sigma(\bx)|)$ of different algorithms, we can see that the proposed algorithm and the REX algorithm give the same value of the objective function, which is a better value compared to the values obtained from other algorithms. Once again, similar results and conclusions can be observed from Table \ref{set2_AOP} for $A$-optimal designs.
\subsection{Setting 3: Two-dimensional Wynn's polygon design space with $p = 6$}
In Setting 3, we consider a two-dimensional irregular design space called Wynn's polygon as
$$
y(\bx)=\bbeta^{T}\bff(\bx),
$$
where $\bbeta=(\beta_0,\beta_1,\beta_2,\beta_3,\beta_4,\beta_5)^{T}$, $\bff(\bx)=(1,x_1,x_2,x_1^2,x_1x_2,x_2^2)^{T}$ and $\bx\in \mathcal{X}=\{(x_1,x_2)^{T}: \;x_1,x_2\geq -\frac{1}{4}\sqrt{2},x_1\leq(x_2+\sqrt{2}),x_2\leq \frac{1}{3}(x_1+\sqrt{2}),x_1^2+x_2^2\leq 1)\}$.\\
Figure \ref{set3} shows the $D$-optimal design points and Table \ref{set3_DOP} shows the design points and corresponding weights by different algorithms. From Table \ref{set3_DOP}, we can see that the proposed algorithm, the REX and CO algorithm can locate the optimal design points with the theoretical weights while the VDM and multiplicative algorithm fail to do so.
For $A$-optimal, the design points and corresponding weights by different algorithms are presented in Table \ref{set3_AOP}. From Table \ref{set3_AOP}, we observe that the proposed algorithm and the REX algorithm provide the same results, and they are superior to other methods.
\subsection{Setting 4: Three-dimensional design space in a cube with $p = 10$}
In Setting 4, we consider the model
$$y(\bx)=\bbeta^{T}\bff(\bx),$$
where $\bbeta=(\beta_0,\beta_1,\beta_2,\ldots, \beta_8,\beta_9)^{T}$, $\bff(\bx)= (1,x_1,x_2,x_3,x_1^2,x_1x_2,x_1x_3,x_2^2,x_2x_3,x_3^2)^{T}$ and $\bx\in \mathcal{X}=[-1,1]\times[-1,1]\times[-1,1]$.
Based on the algorithms considered here, we obtain the same 27 $D$- and $A$-optimal design points with different weights. The weights obtained from different algorithms with the corresponding values of the objective function are presented in Tables \ref{set4_DOP} and \ref{set4_AOP} for the $D$- and $A$-optimality, respectively.
From Tables \ref{set4_DOP} and \ref{set4_AOP}, we observe that the results obtained from the proposed algorithms are very close to the theoretical values \cite{Duan}. Note that although the weights of the proposed optimal design are different from the weights of continuous optimal design provided by \cite{Atkinson}, our optimal design has a smaller $D$-optimal value. Therefore, according to the definition of $D$-optimal criterion, the proposed optimal design should be better.
In this numerical study, we observe that the optimal design points obtained by the existing algorithms are unstable in the sense that the optimal design points may not be unique in multiple runs of the algorithms. This may lead to uncertainty in practical applications. Moreover, we found that the REX, VDM, and MUL algorithms cannot always get all the optimal design points. This may lead to severe problems in some critical experiments. For example, in pharmaceutical or chemical experiments, ignoring some design points may lead to severe consequences.
For illustrative purpose, the $D$-optimal value for the proposed, REX and CO algorithms presented in Tables \ref{set4_DOP} and \ref{set4_AOP} are closest to the theoretical values.
In fact, the REX method may fail when the number of candidate points is small and the number of parameters is large. For example, in Setting 4, if the number of candidate points has 27 design points, the REX method sometimes has a singularity of the $M$ matrix during the calculation process, which causes the failure of obtaining the optimal design. Similarly, in Setting 1, when there are only 7 candidate points, we found that the REX method may also fail sometimes. In contrast, the proposed method is feasible and stable in obtaining the optimal design points even when the number of candidate points is small.
This is a significant advantage of the proposed method in practical application because there are many scenarios that the number of candidate points is small due to high cost or environmental factors.
\subsection{Setting 5: Three-dimensional design space in a sphere with $p = 9$}
In Setting 5, we consider the model
$$y(\bx)=\bbeta^{T}\bff(\bx),$$
where $\bbeta=(\beta_1,\beta_2, \ldots, \beta_8,\beta_9)^{T}$, $\bff(\bx)=(x_1,x_2,x_3,x_1^2,x_1x_2,x_1x_3,x_2^2,x_2x_3,x_3^2)^{T}$ and $\bx\in \mathcal{X}=\{(x_1,x_2,x_3)^{T}: \;x_1^2+x_2^2+x_3^2=1\}$.
For this setting, we use the Fibonacci numbers on the 3-dimensional unit sphere as the initial design points for the algorithms considered here. For more details related to this setting, the reader can refer to \cite{Castro2017}.
Regression problems with a unit sphere design space have many applications in astrophysics, gravity induction, geophysics, climate laws, and global navigation, because there are countless signals on the surface of the earth, and satellite signals also affect our daily lives. Another important application of regression with a unit sphere design space is three-dimensional human faces recognition with sparse spherical representation in authentication and surveillance.
Based on this setting, we find that every design point on the unit sphere can be considered as an optimal design point. By using the proposed algorithms, we obtain all the design points with equal weights.
Figures \ref{set5_1} and \ref{set5_2} display the design points of the $D$-optimal design when the number of supporting points are $500$ and $10$, respectively.
However, when applying the REX, VDM, MUL, and CO algorithms, only a smaller number of design points are identified. For instance, the REX algorithm gives only 128 points as the optimal design points when we use 5000 supporting points in the three-dimensional unit sphere. Moreover, the weights assigned to these 128 design points are not equal based on the REX algorithm.
Figure \ref{set5_3} presents the $D$-optimal design obtained from the REX algorithm when the number of supporting points is 500. We also present the values of the $D$- and $A$-optimality objective functions in Table \ref{set5_DEFF}. From Table \ref{set5_DEFF}, we observe that the $D$-optimal and $A$-optimal values of the proposed algorithm are smaller than the other algorithms considered here. Thus, the proposed method performs well in this case.
To compare the speed of the proposed algorithm for $D$-optimality and $A$-optimality with the REX, CO, VDM and MUL algorithms, we plot the $D$-efficiency and $A$-efficiency (i.e,
$|\sum\limits_{i=1}^{N}w^{*}_{i}\bff_{i}\bff_{i}^{T}|/|\sum\limits_{i=1}^{N}w_{i}\bff_{i}\bff_{i}^{T}|$ and\\ ${\mbox {tr}} \left( \left[\sum\limits_{i=1}^{N} w^{*}_{i}\bff_{i}\bff^{T}_{i} \right]^{-1}\right)/{\mbox {tr}} \left( \left[\sum\limits_{i=1}^{N} w_{i}\bff_{i}\bff^{T}_{i} \right]^{-1}\right)$, where $w^{*}_i$ $i=1,\cdots,N$ is the theoretical optimal design) versus the time (seconds) for Setting 4 with varying sizes from Figures \ref{DOP_N21} -- \ref{AOP_N201}. From Figures \ref{DOP_N21} -- \ref{DOP_N201}, we can see that all of these algorithms will ultimately converge to the theoretical optimal design (given enough time), but the proposed method is superior to the other methods for large size of $N$ of the design space because of the parallel strategy for the computation as mentioned in Remark 1. However, for a small $N$, the REX algorithm tends to perform better than the proposed method. A similar observation can also be drawn for the $A$-efficiency from Figures \ref{AOP_N21} -- \ref{AOP_N201}.
To illustrate the performance of the proposed method in the case that the number of factors is large, we consider the full quadratic regression model
\begin{eqnarray}
y(x_1,x_2,\cdots,x_q)=\beta_0+\sum\limits_{i=1}^q \beta_ix_i+\sum\limits_{j=1}^q\sum\limits_{k=j}^q \beta_{j,k}x_jx_k+\varepsilon,
\label{quadratic_regression}
\end{eqnarray}
where $\beta_{1,1},\beta_{1,2},\cdots,\beta_{q,q}$ correspond to the parameters $\beta_{q+1},\beta_{q+2},\cdots,\beta_{p}$, $p=\frac{(q+1)(q+2)}{2}$.
In Figure \ref{log_eff}, the vertical axis is the value of $-\log_{10} (1-eff)$ for $q=8,10,12,14$ and the horizontal axis denotes the number of iterations, where $eff$ represents the lower bound of the $D$-efficiency \citep{Pukelsheim2006}: $D$-efficiency $\geq \frac{p}{\max_{\bx\in \mathcal{X}} \bff^{'}(\bx) \bM(\xi) \bff(\bx)}$, where $\xi$ is the current design.
In other words, the vertical axis in Figure \ref{log_eff}, the values $1, 2, 3, \cdots$ correspond to $D$-efficiency $0.9, 0.99, 0.999, \cdots$.
In Figure \ref{D_value}, the vertical axis is the $D$-criterion values of designs produced by the proposed method for $q=8,10,12,14$ and the horizontal axis denotes the number of iterations.
From Figures \ref{log_eff} and \ref{D_value}, with the increase of iteration times, $-\log_{10} (1-eff)$ gradually increases and $D$-criterion values converge to the $D$-optimal value when the number of factors is large. Thus, the proposed method is still effective even when the number of factors is large.
Based on the numerical evaluations of the five settings considered in this section, we found that the proposed algorithms for $D$-optimality and $A$-optimality converge in all cases, and the optimal design points, as well as the corresponding weights, are close to the theoretical values. Furthermore, in some cases, the proposed method outperforms some existing algorithms for computing approximate $D$- and $A$-optimal designs. It is noteworthy that the proposed algorithm is simple and it can be implemented without relying on any advanced mathematical programming solvers. Therefore, the proposed algorithms provide a more convenient and effective way to approximate the $D$- and $A$-optimal solutions on the compact design space.
\section{Concluding Remarks}
In this paper, we discuss the approximate optimal design and proposed efficient iterative computational algorithms to obtain the approximate $D$-optimal and $A$-optimal designs for linear models on compact design spaces.
Due to the simplicity and efficiency of the algorithm, the two proposed algorithms are easy to implement. The proposed algorithms are useful tools for many practical applications of optimal design of experiments.
We also provided proof of the monotonic convergence of the proposed algorithm for $D$-optimality and demonstrate that the proposed algorithms provide solution that converges to the optimal design. Furthermore, we prove that the optimal approximate designs converge to the continuous optimal design under certain conditions. A theoretical justification for the convergence of the proposed algorithm for $A$-optimality is not available, but our numerical results strongly support the validity and reliability of the proposed algorithm. These algorithms are implemented in Matlab and the programs are available from the authors upon request. It is worth mentioning that although we focus on $D$-optimal designs and $A$-optimal designs for linear models in this paper, the ideas of the proposed algorithms can be extended to other optimal criteria and other design space in high-dimensional situations.
\section*{Appendix}
\subsection*{Proof of Theorem 1}
Since $\log(|\textbf{A}|)$ is concave in $\textbf{A}$ with $\textbf{A}$ being a positive definite matrix, we have
\begin{eqnarray*}
&&\log \left\{ \left| (1-\lambda)\sum\limits_{i=1}^{N} w^{*}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i})+\lambda\sum\limits_{i=1}^{N} w(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right| \right\}\\
&&\geq (1-\lambda) \log\left\{\left|\sum\limits_{i=1}^{N} w^{*}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right| \right\}+\lambda \log\left\{ \left|\sum\limits_{i=1}^{N} w(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right| \right\}.
\end{eqnarray*}
Then, $w^{*}$ is the optimal solution for the $D$-optimal criterion in (2) if and only if
\begin{eqnarray*}
\frac{\log \left\{ \left| (1-\lambda) I_{\mathcal{X}}(\bw^{*}, \bff) + \lambda I_{\mathcal{X}}(\bw, \bff) \right| \right\} - \log \left\{ \left| I_{\mathcal{X}}(\bw^{*}, \bff) \right| \right\}}{\lambda} \leq 0,
\end{eqnarray*}
where $I_{\mathcal{X}}(\bw, \bff) = \sum\limits_{i=1}^{N} w(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i})$, for all $w(\bx_{i})$ that satisfy $w(\bx_{i})\geq 0, \; i=1,2,\ldots,N$ and $\sum\limits_{i=1}^{N} w(\bx_{i}) =1$, and $\lambda > 0$.
Thus, for $\lambda\downarrow 0$, we have
\begin{eqnarray*}
& &\lim\limits_{\lambda\downarrow 0} \frac{\log(|(1-\lambda) I_{\mathcal{X}}(\bw^{*}, \bff) + \lambda I_{\mathcal{X}}(\bw, \bff)|) - \log(|I_{\mathcal{X}}(\bw^{*},\bff)|)}{\lambda}\\
& = & \left. \frac{\partial \log \{ |(1-\lambda) I_{\mathcal{X}}(\bw^{*}, \bff) +\lambda I_{\mathcal{X}}(\bw, \bff)| \} }{\partial \lambda} \right|_{\lambda=0}\\
& = & \mbox{tr} \left\{ \left[\sum\limits_{i=1}^{N} w^{*}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right]^{-1} \sum\limits_{i=1}^{N} [w(\bx_{i})-w^{*}(\bx_{i})]\bff(\bx_{i})\bff^{T}(\bx_{i}) \right\}\\
& = & \mbox{tr} \left\{ \left[\sum\limits_{i=1}^{N} w^{*}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right]^{-1} \sum\limits_{i=1}^{N} w(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right\} - p \leq 0,
\end{eqnarray*}
which gives the result in Theorem 1.
\subsection*{Proof of Theorem 2}
From Lemma 1, we have
\begin{eqnarray*}
& & \log \left|\sum\limits_{i=1}^{N} w^{(n)}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right|-\log \left|\sum\limits_{i=1}^{N} w^{(n-1)}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right|\\
& \geq & \sum\limits_{i=1}^{N} w^{(n-1)}(\bx_{i})\mbox{trace}\left\{\bff(\bx_{i})\bff^{T}(\bx_{i})D^{(n-1)}\right\}\log \frac{w^{(n)}(\bx_{i})}{w^{(n-1)}(\bx_{i})}\\
& = & \sum\limits_{i=1}^{N} w^{(n-1)}(\bx_{i})\bff^{T}(\bx_{i})D^{(n-1)}\bff(\bx_{i})\log \frac{w^{(n)}(\bx_{i})}{w^{(n-1)}(\bx_{i})}\\
& = & p\sum\limits_{i=1}^{N} w^{(n)}(\bx_{i})\log \frac{w^{(n)}(\bx)}{w^{(n-1)}(\bx)}\geq 0.
\end{eqnarray*}
Thus, we can conclude that
$$\log \left|\sum\limits_{i=1}^{N} w^{(n)}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right|$$
is increasing in $n = 2, 3, \ldots.$ We can get that
\begin{eqnarray*}
\log \left|\sum\limits_{i=1}^{N} w^{(n)}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right| \leq \log \left|\sum\limits_{i=1}^{N} \bff(\bx_{i})\bff^{T}(\bx_{i}) \right|.
\end{eqnarray*}
Therefore, under the bounded assumption, the sequence $\log \left|\sum\limits_{i=1}^{N} w^{(n)}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right|$ is uniformly bounded and increasing, and hence it is convergent.
Using Lemma 1 and Lemma 2, we can obtain
\begin{eqnarray*}
0 & = & \lim_{n \to \infty}\log\left|\sum\limits_{i=1}^{N} w^{(n)}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right| - \log\left|\sum\limits_{i=1}^{N} w^{(n-1)}(\bx_{i})\bff(\bx_{i})\bff^{T}(\bx_{i}) \right|\\
& \geq & p \sum\limits_{i=1}^{N} w^{(n)}(\bx_{i})\log \frac{w^{(n)}(\bx_{i})}{w^{(n-1)}(\bx_{i})}\\
& \geq & \frac{p}{2}[\sum\limits_{i=1}^{N} |w^{(n)}(\bx_{i})- w^{(n-1)}(\bx_{i})|]^{2}\geq 0.
\end{eqnarray*}
Then, we can conclude that
$$
\sum\limits_{i=1}^{N} |w^{(n)}(\bx_{i})-w^{(n-1)}(\bx_{i})|\longrightarrow 0 {\mbox { as }} n \longrightarrow +\infty.
$$
\subsection*{Proof of Theorem 4}
We can check that ${\mbox {tr}}(\textbf{A})^{-1}$ is concave in $\textbf{A}$, where $\textbf{A}$ is positive definite matrix. We have
\begin{eqnarray*}
{\mbox {tr}}[(1-\lambda)I_\mathcal{X}(\bw^{*}, \bff)+\lambda I_\mathcal{X}(\bw, \bff)]^{-1} \leq (1-\lambda){\mbox {tr}}(I_\mathcal{X}^{-1}(\bw^{*}, \bff))+\lambda {\mbox {tr}}(I_\mathcal{X}^{-1}(\bw, \bff)).
\end{eqnarray*}
Then, $w^{*}$ is the optimal solution for the $A$-optimal criterion in Eq. (\ref{AOP}) if and only if for all $w(\bx_{i})(w(\bx_{i})\geq 0\;\mbox{and}\;\sum\limits_{i=1}^{N} w(\bx_{i}) =1)$,
\begin{eqnarray*}
\frac{{\mbox {tr}}[(1-\lambda)I_\mathcal{X}(\bw^{*}, \bff)+\lambda I_\mathcal{X}(\bw, \bff)]^{-1} - {\mbox {tr}}(I_\mathcal{X}^{-1}(\bw^{*}, \bff))}{\lambda} \geq 0
\end{eqnarray*}
for $\lambda>0$.
Thus, for $\lambda\downarrow 0$ we have
\begin{eqnarray*}
&&\lim\limits_{\lambda\downarrow 0} \frac{{\mbox {tr}}[(1-\lambda)I_\mathcal{X}(\bw^{*}, \bff)+\lambda I_\mathcal{X}(\bw, \bff)]^{-1}-{\mbox {tr}}(I_\mathcal{X}^{-1}(\bw^{*}, \bff))}{\lambda}\\
& = & \frac{\partial \{ tr[(1-\lambda)I_\mathcal{X}(\bw^{*}, \bff)+\lambda I_\mathcal{X}(\bw, \bff)]^{-1}\}}{\partial \lambda}|_{\lambda=0}\\
& = & -{\mbox {tr}}[I_\mathcal{X}^{-1}(\bw^{*}, \bff)(I_\mathcal{X}(\bw, \bff)-I_\mathcal{X}(\bw^{*}, \bff))I_\mathcal{X}^{-1}(\bw^{*}, \bff) ]\\
& = & -{\mbox {tr}}( I_\mathcal{X}^{-1}(\bw^{*}, \bff)I_\mathcal{X}(\bw, \bff)I_\mathcal{X}^{-1}(\bw^{*}, \bff)) + {\mbox {tr}}( I_\mathcal{X}^{-1}(\bw^{*}, \bff)) \geq 0,
\end{eqnarray*}
which implies Theorem 4.
|
\subsection{Generic case}
It is expected that graphene has a smaller lattice constant than all
of the realistic WSM materials, as it is made of the lighter carbon
atoms compared to heavier atoms of most realized WSMs. This means
that graphene will have the smaller unit cell (larger BZ in momentum
space). Furthermore, when graphene is added on top of the WSM, the
two lattices will be generally incommensurate due to mismatch in
their lattice constants, and to possible relative rotation between
the two lattices. Another knob to tune for the system is the relative
chemical potential between graphene and the WSM. We require the WSM
to be doped in a way so that the Dirac point in graphene lies in the
bulk states of the WSM which does not need any special tuning.
Focusing on the ${\bf K}$ valley in graphene, we see that for a
continuous range of angles, we can make the ${\bf K}$ points in
graphene coincide with bulk states of the WSM, as shown in
Fig. \ref{fig3}. For states near the ${\bf K}$ valley, they will mix
with the continuum of WSM bulk states, effectively realizing a
resonant level model\cite{Mah00}. Accordingly, the ${\bf K}$ valley
states will get delocalized and extend into the bulk, disappearing
from the surface in the thermodynamic limit. On the other hand,
states near the ${\bf K}'$ valley will lie in the gap of the WSM
energy spectrum and remain on the surface; this is possible due to
the broken time reversal symmetry in the WSM. In this way, our system
works as a valley filter which within the assumptions of our model is
100\% efficient. In principle, this effect may be tested
experimentally by laying a graphene sheet on top of two WSM slabs,
which are rotated by $60^\circ$ relative to one another. This is
expected to block the current completely, in analogy with light
passing through two polarizing filters with perpendicular
polarization axes.\\
\section{Generic hopping between Graphene and the WSM surface}
\label{SM:sec1}
Having treated a highly symmetric model of hopping between the MLG and
the top layer of the WSM in the main text (see Fig. 3 in the main
text), we now treat the case where the $C_3$ symmetry is broken by the
coupling. We also allow the spin to not be conserved in
the tunneling process. We displace the two lattices with respect to
each other to get a configuration where there is no center of rotation
present (See Fig. \ref{fig1}\subref{str}). There are 6 atoms per unit
cell (at the interface) in our model; two of them being the A and the B sublattices in
graphene and the remaining four belong to the WSM which we denote by
numbers from 0 to 3. The coupling will take the general form
\begin{equation}
\begin{pmatrix}
\kappa_{A,0}(\vec{p}) & \kappa_{B,0}(\vec{p})\\
\kappa_{A,1}(\vec{p}) & \kappa_{B,1}(\vec{p})\\
\kappa_{A,2}(\vec{p}) & \kappa_{B,2}(\vec{p})\\
\kappa_{A,3}(\vec{p}) & \kappa_{B,3}(\vec{p})\\
\end{pmatrix}\otimes
\begin{pmatrix}
\nu_{\upuparrows} & \nu_{\uparrow\downarrow}\\
\nu_{\downarrow\uparrow} & \nu_{\downdownarrows}
\end{pmatrix},
\end{equation}
where $\kappa_{A/B,i}(\vec{p})$ denotes the coupling between the A/B
sublattice and the $i$ site of the WSM which is generally a function
of momentum $\vec{p}$. $\nu_{\alpha\beta}$ denotes the coupling
between spins $\alpha$ and $\beta$ in graphene and the WSM,
respectively. Taking the configuration shown in Fig. \ref{fig1}\subref{str} and keeping nearest neighbor terms only, (1.1) will take the form, \begin{equation}
\begin{pmatrix}
\kappa_{A,0} & \kappa_{B,0}\\
0 & 0\\
\kappa_{A,2}e^{i\vec{p}.\vec{b_2}} & \kappa_{B,2}\\
\kappa_{A,3}e^{i\vec{p}.\vec{b_1}+i\vec{p}.\vec{b_2}} & \kappa_{B,3}\\
\end{pmatrix}\otimes
\begin{pmatrix}
\nu_{\upuparrows} & \nu_{\uparrow\downarrow}\\
\nu_{\downarrow\uparrow} & \nu_{\downdownarrows},
\end{pmatrix}
\end{equation}
where $\vec{b_{1,2}}$ are the lattice bases vectors. The behavior of states in the ${\bf K}'$ valley depends
on the details of the coupling, as shown in
Fig. \ref{fig1}\subref{c3break1},\subref{c3break2}.
\begin{figure}[h]
\subfloat[\label{str}]{%
\includegraphics[width=0.3\linewidth]{suppfig0}}
\hspace{1em}
\subfloat[\label{c3break1}]{%
\includegraphics[width=0.3\linewidth]{suppfig1}}
\hspace{1em}
\subfloat[\label{c3break2}]{%
\includegraphics[width=0.3\linewidth]{suppfig2}}
\hspace{1em}
\subfloat[\label{c3break3}]{%
\includegraphics[width=0.3\linewidth]{suppfig3}}
\hspace{1em}
\subfloat[\label{c3break4}]{%
\includegraphics[width=0.3\linewidth]{suppfig4}}
\\
\caption{\label{fig1} General interface between graphene and the WSM. (a) Displacement of graphene and the WSM lattices lead to a structure with no $C_3$ symmetry present. (b),(c) A zoom-in on the $\bf{K}'$ valley band structure with the tunneling conserving spin. (d),(e) A zoom-in on the $\bf{K}'$ valley band structure with spin non-conserving processes present. The colors code the average of the spin operator's $z$ component with red being $\uparrow$ and blue being $\downarrow$. The dilution of the colors in (d),(e) implies less spin polarization. Values of the coupling in (b),(d) are $\kappa_{A,0}=\kappa_{B,2}=\kappa_{B,3}=0.2$, $\kappa_{A,2}=\kappa_{A,3}=0.15$ and $\kappa_{B,0}=0.1$. Values of the coupling in (c),(e) are $\kappa_{A,0}=\kappa_{B,2}=\kappa_{B,3}=0.3$, $\kappa_{A,2}=\kappa_{A,3}=0.25$ and $\kappa_{B,0}=0.2$.}
\end{figure}
No qualitative difference from the highly symmetric model of the main
text is found in the spectrum near the ${\bf K}'$ valley. We conclude
that the spectrum presented in Fig. 4(c) of the main text is robust to
making the tunneling matrix elements between the MLG and the WSM
generic.
\section{Dissolution of Graphene States into Weyl Semimetal}
\label{SM:sec2}
It is well-known that unless there are symmetry restrictions, or
couplings are fine-tuned \cite{BIC-review}, a localized state which is
tunnel-coupled to a continuum of delocalized states will generically
hyrbridize with them and dissolve into them \cite{Mah00}. Here we
present numerical evidence that there is no hidden
symmetry/fine-tuning in our model that prevents this dissolution.
To find the effect on a graphene state after the coupling is turned
on, for a given value of ${\vec k}$ near the ${\bf K}$ valley, we look
for states localized near the top surface, in a range of energies near
that of the graphene unperturbed states. We use the inverse
participation ratio ($IPR$) as a measure for localization which is
defined for a wave function $\psi$ as, \begin{equation}
IPR=\frac{1}{\sum\limits_{i}\abs{\psi_i}^4}.
\end{equation}
where $i$ is a joint index representing the layer, spin, and for
graphene sites, sublattice as well. A completely localized state will
give an $IPR$ of order 1 whereas a completely delocalized state will
give an $IPR$ of order $N$, with $N$ being the dimension of the
Hilbert space. Using this definition, we conduct a search for
localized states at a momentum near the ${\bf K}$ valley and in the
vicinity of the unperturbed energy. We then observe how the probability of the electron to be on the MLG
behaves as the
thickness of the WSM slab is increased and the bulk states become denser (See Fig. \ref{amp}). For
different momenta near the ${\bf K}$ valley, one sees clearly that the
states become delocalized as the the system size increases. One clearly sees the trend towards complete delocalization in the thermodynamic limit.
\begin{figure}[h]
\subfloat[\label{amp}]{%
\includegraphics[width=0.3\linewidth]{amp}}
\hspace{1em}
\subfloat[\label{amp2}]{%
\includegraphics[width=0.3\linewidth]{amp2}}
\hspace{1em}
\subfloat[\label{amp3}]{%
\includegraphics[width=0.3\linewidth]{amp3}}\\
\caption{\label{fig2}Delocalization of the $\bf{K}$ valley. (a)-(c) The graphene wave function amplitudes as a function of the system size for momenta $\vec{p}=(2\pi/3,0)$, $\vec{p}=(2\pi/3+0.025,0.014)$ and $\vec{p}=(2\pi/3-0.02,0.012)$ respectively. Calculations were done in the general configuration with the coupling in equation (1.2). $\kappa=0.1$ represents the set of parameters $\kappa_{A,0}=\kappa_{B,3}=0.1$ and $\kappa_{A,2}=\kappa_{A,3}=\kappa_{B,0}=\kappa_{B,2}=0.08$. $\kappa=0.2$ represents the set of parameters $\kappa_{A,0}=\kappa_{B,3}=0.2$ and $\kappa_{A,2}=\kappa_{A,3}=\kappa_{B,0}=\kappa_{B,2}=0.18$. $\kappa=0.3$ represents the set of parameters $\kappa_{A,0}=\kappa_{B,3}=0.3$ and $\kappa_{A,2}=\kappa_{A,3}=\kappa_{B,0}=\kappa_{B,2}=0.28$}
\end{figure}
\section{Conductance}
\label{SM:sec3}
We next discuss how one calculates the conductance in a graphene-WSM device \cite{beenakker}, for which the result is given in Fig. 4(d) in the main text. We work within the effective model for the $\bf{K}'$ valley introduced in equation (6) in the main text. In this simple model, there are four spin polarized bands and the spin carried by the bands alternates between up and down so it emulates the band inversion seen in the bands obtained from the tight binding calculation. In this model, however, the two middle bands touch at zero energy where there is a tiny gap in the bands obtained numerically. So as long as the chemical potential is away from this point (which is the regime we are interested in), the model should give qualitatively the same result as the numerical bands.
Looking at the schematic picture in Fig. 1 in the main
text, we see that the strip is divided into three regions with
graphene leads on the left ($x<0$) and right ($x>L$) at potential 0
while the middle region ($0<x<L$) at potential $V_g$. We consider a
scattering problem where we assume the potential to take the
form \begin{equation} V(x) = \begin{cases} 0 & x\leq 0 \\ eV_g & 0\leq
x\leq L \\ 0 & L\leq x
\end{cases}\; .
\end{equation}
For the ease of notation we denote the x-momentum $k$ and the transverse momentum $q$. We also set $\hbar$ and graphene Fermi velocity $v$ to 1. We show the conductance resulting from three different values for the chemical potential in Fig. \ref{fig3} to show that the result obtained is not too sensitive to the chemical potential. We present the detailed calculation for the chemical potential $\mu_1$ in Fig. \ref{fig3}; the others can be obtained in a similar fashion.\\ To start, we need to find the dispersion of electrons in the leads and in the middle region. We have in the leads \begin{equation}
\varepsilon=\sqrt{k^2+q^2}
\end{equation}
For $0<x<L$, we have (using the continuum model of Eq. (6) in the main text)\begin{equation}
\varepsilon=eV_g-\lambda_2+\sqrt{\lambda_1^2+\tilde{k}^2+q^2},
\end{equation}
for the spin up band. For the spin down band, we have \begin{equation}
\varepsilon=eV_g+\lambda_2-\sqrt{\lambda_1^2+\tilde{k}^2+q^2}
\end{equation}
Note that $k$ in the leads is always real while $\tilde{k}$ can be imaginary.
Consider a scattering state coming from the left. We have for $x<0$,
\begin{align}
\Phi_L &= \begin{pmatrix}
-1 \\
z_k
\end{pmatrix}e^{ikx+iqy}+r\begin{pmatrix}
-1 \\
z_{-k}
\end{pmatrix}e^{-ikx+iqy}
\end{align}
For $0<x<L$, we have
\begin{align}
\tilde{\Phi} = \alpha\begin{pmatrix}
-1 \\
w_{\tilde{k}}
\end{pmatrix}e^{i\tilde{k}x+iqy}+\beta\begin{pmatrix}
-1 \\
w_{-\tilde{k}}
\end{pmatrix}e^{-i\tilde{k}x+iqy}
\end{align}
For $x>L$, we have
\begin{align}
\Phi_R &= t\begin{pmatrix}
-1 \\
z_k
\end{pmatrix}e^{ik(x-L)+iqy}
\end{align}
$z_k=\frac{k-iq}{\sqrt{k^2+q^2}}$. $w_{\tilde{k}}=\frac{\tilde{k}-iq}{\lambda_1+\sqrt{\lambda_1^2+\tilde{k}^2+q^2}}$
for the spin up states, and
$w_{\tilde{k}}=\frac{\tilde{k}-iq}{\lambda_1-\sqrt{\lambda_1^2+\tilde{k}^2+q^2}}$
for the spin down states. $r$ and $t$ define the reflection and the
transmission probabilities, respectively.\\ Using the continuity of
the wave functions at $x=0$ and $x=L$, we can solve for the
transmission amplitude $t$, resulting in \begin{equation}
t=\frac{(1+z_k^2)(w_{\tilde{k}}-w_{-\tilde{k}})}{e^{i\tilde{k}L}[(1+z_{k}w_{-\tilde{k}})(w_{\tilde{k}}-z_{k})]+e^{-i\tilde{k}L}[(1+z_{k}w_{\tilde{k}})(z_{k}-w_{-\tilde{k}})]}
\end{equation}
The transmission probability is given by $T=|t|^2$. Plotting this
equation as a function of the conserved transverse momentum ($k_y$)
gives Fig. (4)(d) in the main text.
\begin{figure}[h]
\subfloat[\label{bands}]{%
\includegraphics[width=0.4\linewidth]{model}}
\hspace{1em}
\subfloat[\label{con}]{%
\includegraphics[width=0.4\linewidth]{cond}}\\
\caption{\label{fig3}Conductance at different chemical potentials. (a) The bands obtained from the continuum model of equation (6) in the main text. The three horizontal black lines represent three values for the chemical potential. (b) The conductance measured in units of ($e^2/h$) for the three respective chemical potential. Solid, $\mu_1=0.02$, Dashed, $\mu_2=0.04$ and Dotted, $\mu_3=0.06$. In (a),(b) red denotes spin $\uparrow$ and blue denotes spin $\downarrow$.}
\end{figure}
\section{Graphene-Fermi arc hybridization}
\label{SM:sec4}
\begin{figure}[h]
\centering
\includegraphics[width=0.3\linewidth]{arcs}
\caption{an incommensurate structure of graphene and the WSM shown in momentum space. The 3 equivalent graphene $\bf{K}$ vectors coincide with the Fermi arcs.}
\label{fig:bz2}
\end{figure}
Another possibility that this system offers is the hybridization between graphene's Dirac cones and the Fermi arc states localized on the surface of the WSM. This again will occur in only one of graphene's valleys due to the breaking of time reversal symmetry. This construction, however, requires some fine-tuning as we need the chemical potential to be at the Dirac points in graphene and at the Weyl nodes in the WSM. Furthermore, we need the $K$ point in graphene to coincide with the Fermi arcs as shown in Fig. \ref{fig:bz2} which happens at certain orientations of the graphene-WSM system.\\
In a fashion similar to our treatment in the main text, we consider a commensurate structure of graphene and the WSM with the Dirac $\bf{K}$ point overlapping a Fermi arc in momentum space. We then perform our tight binding calculations to study the electronic properties of this system. A plot for the band structure near the $\bf{K}$ valley shows the effect of the WSM Fermi arc on the Dirac cone is shown in Fig. \ref{fig5}\subref{kbands}. \begin{figure}[h]
\subfloat[\label{kbands}]{%
\includegraphics[width=0.3\linewidth]{kvalley}}
\subfloat[\label{recon}]{%
\includegraphics[width=0.3\linewidth]{fermiarc}}\\
\subfloat[\label{linear}]{%
\includegraphics[width=0.3\linewidth]{lin}}
\subfloat[\label{quadratic}]{%
\includegraphics[width=0.3\linewidth]{quad}}\\
\caption{\label{fig5} Fermi Arcs hybridization with the $\bf{K}$ valley. (a) The Dirac cone near the $\bf{K}$ valley along the $k_x$ axis in momentum hybridizing with Fermi arc states present on the top surface of the WSM. (b) A plot of the reconstructed Fermi arcs near the $\bf{K}$ point which shows the detachement of the arcs into three closed loops. (c) The gap in the spin $\uparrow$ Dirac cone scales linearly with the coupling as the data points fits nicely to a straight line (in red). (d) The gap in the spin $\downarrow$ Dirac cone scales quadratically with the coupling as the data is fitted nicely to a parabola (in red).}
\end{figure}
We can understand the resulting band structure based on the fact that the Fermi arc states are spin polarized in our simple model with the spins being up for the top WSM surface. Consequently, the spin up Dirac cone acquires a gap which is shown to scale linearly with the coupling as shown in Fig. \ref{fig5}\subref{linear}. The spin down cone, however, remains relatively intact (See Fig. \ref{fig5}\subref{kbands}) but upon careful examination, we can see it acquires a small gap which scales quadratically with the coupling as shown in Fig. \ref{fig5}\subref{quadratic}.\\
Finally, looking at the zero energy contour, we can see a reconstruction of the Fermi arcs as shown in Fig. \ref{fig5}\subref{recon}.
The existence of a Fermi loop at the surface can lead to interesting resonance signatures in the optical conductivity of the system in a magnetic field perpendicular to the surface.
|
\subsection{Problem Formulation and Notation}
We consider again the example from Fig.~\ref{fig:fig1}, where we want to classify a task, $T_t$, and an attribute, $A_a$. An image containing a woman is denoted by the presence of a binary indicator variable, $A_0=1$, or for notational convenience, simply $A_0$. On this dataset women ($A_0$) are correlated with painting ($T_0$), and men ($A_1$) with not painting ($T_1$). The set of attributes, $a\in\mathcal{A}=\{0, 1\}$ in this example, corresponds to $A_0$ and $A_1$, each of which is a binary indicator variable. When we introduce our proposed metric in Sec.~\ref{sec:our_metrix_exp}, this notation expresses our explicit and deliberate allowance for situations in which a person may identify with zero, one, or more than one attributes. The set of tasks, $t\in\mathcal{T}=\{0, 1\}$ in this example, follows a similar formulation to allow for multi-label settings. We imagine $N$ images each of $(A_0,T_0)$ and $(A_1,T_1)$ but only $N/2$ images of $(A_0,T_1)$ and $(A_1,T_0)$. A classifier trained to recognize painting on this data is likely to learn these associations and over-predict painting on images of women and under-predict painting on images of men; however, algorithmic interventions may counteract this and in fact result in the opposite behavior.
\subsection{Overview of Existing Fairness Metrics}
\label{sec:error_types}
We begin with a review of existing fairness metrics in a concrete classification setting.
Four fairness metrics we consider are: {\it False Positive Rate (FPR) and True Positive Rate (TPR) difference:} the difference in false positive (true positive) rate of predicting the label $T_0$ on images of $A_1$ versus $A_0$~\citep{chouldechova2016recidivism, hardt2016equalopp}, {\it accuracy difference:} difference between the task prediction accuracy on images of $A_1$ versus on images of $A_0$~\citep{berk2017risk}, and {\it mean accuracy across subgroups:} mean task prediction accuracy across the four image subgroups ($(A_0, T_0), (A_1, T_0), (A_0, T_1), \text{and } (A_1, T_1)$)~\citep{buolamwini2018gendershades}.
Fig.~\ref{fig:metric_comp} shows the behavior of fairness metrics under varying amounts of learned amplification.
However, these four metrics are not designed to account for the training correlations, and unable to distinguish between cases of increased or decreased learned correlations.
\cite{zhao2017menshop} introduced an alternative to these that explicitly captures the notion of bias amplification. Concretely, they consider $P(A_a=1|T_t=1)$ of the training data as the fraction of times a protected attribute $A_a \in \mathcal{A}$ appears on images corresponding to task $T_t \in \mathcal{T}$. They then compare this with the test-time predictions made by the model, $P(\hat{A}_a=1|\hat{T}_t=1)$, or the number of times attribute $A_a$ is predicted on images where the task is predicted as $T_t$, which allows them to measure bias amplification in the absence of any additional annotations on the hold-out set. Note that in this formulation they are assuming the model is making predictions for both the task and the attribute. The full bias amplification metric (reformulated in our terms), is
\begin{multline}
\text{\biasamptheirs}=
\frac{1}{|\mathcal{T}|}
\sum_{t=1}^{|\mathcal{T}|} \sum_{a=1}^{|\mathcal{A}|} \underbrace{\mathbbm{1}\left( P(A_a=1|T_t=1) > \frac{1}{|\mathcal{A}|}\right)}_{y(t, a)}
\\ \underbrace{\left ( P(\hat{A}_a=1|\hat{T}_t=1) - P(A_a=1|T_t=1) \right )}_{\Delta(t, a)}
\label{eqn:menshop}
\end{multline}
Fig.~\ref{fig:metric_comp} empirically demonstrates that this metric is able to capture the level of increasing bias amplification. (For consistency in comparison with prior metrics, we assume the model always correctly predicts the protected attribute $A$.) However, as we discuss in the next section, there are some properties of bias amplification that this metric is not able to capture: for example, it does not distinguish between errors in predicting the protected attribute versus errors in predicting the task. Thus we introduce a new metric (last graph of Fig.~\ref{fig:metric_comp}) which maintains the desirable properties from \cite{zhao2017menshop} while including a number of innovations.
\begin{figure}[t]
\centering
\begin{subfigure}{}
\includegraphics[width=1.\linewidth]{Images/metric_comp.png}
\end{subfigure}
\caption{\label{fig:metric_comp}
Different fairness metrics vary in how they respond to model errors. In our image dataset (Fig.~\ref{fig:fig1}) of predicting someone who is a woman or man to be painting or not, we consider a classifier that always predicts the task correctly for men, but varies for women. The x-axes of the graphs correspond to the fraction of women predicted to be painting, where the ground-truth is $2/3$, and the model does not predict false positives before this point. The first two metrics, FPR and TPR difference, only capture one of under- or overclassification. The next two metrics are symmetric around $2/3$, being unable to differentiate under- or overclassification. Both bias amplification metrics are able to distinguish between under- and overclassification.
}
\end{figure}
\subsection{Shortcomings of BiasAmp$_{\textnormal{MALS}}$}
\label{sec:biasamptheirs}
Despite the advantages of \biasamptheirs{} (Eqn.~\ref{eqn:menshop}) and its ability to distinguish under- and overclassifications of training correlations, this metric also suffers from a number of shortcomings. To ground our discussion, we will work directly with the model outputs released by~\citet{zhao2017menshop} from their Conditional Random Field (CRF) model on COCO~\citep{lin14coco}, which has predictions for gender and objects detected for each image.
\subsubsection{Non-binary attributes}
\label{sec:nonbinary}
The first shortcoming is the metric assumes the protected attributes are binary, limiting its use: the indicator variable $y(t,a)$ only counts an increase in positive correlations, but not a decrease in negative correlations, in multi-attribute scenarios.
Consider a task $T_0$ such that $A_0$ is associated with it, but none of the other $A_i$, where $i\in\mathcal{A}, i\neq 0$. In this scenario, $\Delta(T_0, A_i)$ is only measured when there is one other $A_i$ such that $A_i=\lnot A_0$, since $\Delta(t, a)=-\Delta(t, \lnot a)$.
A simple addition of $-\Delta$
when $y$ is 0 would count all bias amplification when there are more than two groups.
\subsubsection{Base rates}
\label{sec:baserates}
The second shortcoming is that the metric does not take into account the base rates of each attribute. Concretely, when determining in $y(t,a)$ of Eqn.~\ref{eqn:menshop} whether the attribute $A_a$ is correlated with the task $T_t$, $P(A_a=1|T_t=1)$ is compared to $\frac{1}{|\mathcal{A}|}$. However, this assumes that all $A_a$'s within $\mathcal{A}$ are evenly distributed, which may not be the case. For example, in COCO there are about 2.5x as many men as women, so it would appear that most objects positively correlate with men simply by nature of there being an overrepresentation of men. Consider the object \texttt{oven}; \biasamptheirs{} calculates $P(A_{\texttt{man}}=1|T_{\texttt{oven}}=1)=0.56>\frac{1}{2}$ and thus considers this object to be correlated with men rather than women. However, computing $P(A_{\texttt{man}}=1,T_{\texttt{oven}}=1) = 0.0103 < 0.0129 = P(A_{\texttt{man}}=1)P(T_{\texttt{oven}}=1)$
reveals that men are in fact \emph{not} correlated with oven, and the seeming overrepresentation comes from the fact that men are overrepresented in the dataset more generally. Not surprisingly, the model trained on this data learns to associate women with ovens and underpredicts men with ovens at test time, i.e., $P(\hat{A}_{\texttt{man}}=1 | \hat{T}_{\texttt{oven}}=1) - P(A_{\texttt{man}}=1 | T_{\texttt{oven}}=1) = -0.10$. \biasamptheirs{} erroneously counts this as \emph{negative} bias amplification.
\subsubsection{Entangling directions}
\label{sec:sub_disentangle}
Another shortcoming is the inability to distinguish between different \emph{types} of bias amplification. \citet{zhao2017menshop} discovers that ``Technology oriented categories
initially biased toward men such as \texttt{keyboard}... have each increased their bias toward
males by over 0.100." Concretely, from Eqn.~\ref{eqn:menshop}, $P(A_{\texttt{man}}=1|T_{\texttt{keyboard}}=1) = .70$ and $P(\hat{A}_{\texttt{man}}=1|\hat{T}_{\texttt{keyboard}}=1) = .83$, demonstrating an amplification of bias. However, the \emph{direction} or cause of bias amplification remains unclear: is the presence of man in the image increasing the probability of predicting a keyboard, or vice versa? Looking more closely at the model's disentangled predictions, we see that when conditioning on the attribute, $P(\hat{T}_{\texttt{keyboard}}=1|A_{\texttt{man}}=1) = 0.0020 < 0.0032 = P(T_{\texttt{keyboard}}=1|A_{\texttt{man}}=1)$, and when conditioning on the task,
$P(\hat{A}_{\texttt{man}}=1|T_{\texttt{keyboard}}=1) = 0.78 &> 0.70 = P(A_{\texttt{man}}=1|T_{\texttt{keyboard}}=1)$,
indicating that keyboards are under-predicted on images with men yet men are over-predicted on images with keyboards. Thus the root cause of this amplification appears to be in the gender predictor rather than the object detector. Such disentangement allows us to properly focus algorithmic intervention efforts.
This also highlights the need to consider the ground truth labels in addition to the predictions on the hold-out set, since when considering only the predictions, it is impossible to decouple the different sources of bias.
\subsection{Threshold}
We fully replicate the original experiment from~\citet{zhao2017menshop} using \biasamptheirs{} on their model predictions and measure $.040$. However, we observe that ``man" is being predicted at a higher rate (75.6\%) than is actually present (71.2\%).
With this insight, we tune the decision threshold on the validation set such that the gender predictor is well-calibrated to be predicting the same percentage of images to have men as
the dataset actually has. When we calculate \biasamptheirs{} on these newly thresholded predictions for the test set, we see bias amplification drop from $0.040$ to $0.001$ just as a result of this threshold change, outperforming even the solution proposed in \citet{zhao2017menshop} of corpus-level constraints, which achieved a drop to only $0.021$. Fairness can be quite sensitive to the threshold chosen~\citep{chen2020threshold}, so careful threshold selection should be done, rather than using a default of $.5$.
When a threshold is needed in our experiments, we pick it to be well-calibrated on the validation set. Although we do not take this approach, because at deployment time it is often the case that discrete predictions are required, one could also imagine integrating bias amplification across proportion or threshold to have a threshold-agnostic measure of bias amplification, similar to what is proposed by \citet{chen2020threshold}.
\subsection{Directions for Mitigation}
\begin{table}
\includegraphics[width=.95\linewidth]{Images/coco_masks.png}
\caption{\biasampours{} changes as a result of three conditions of images: full, noisy person mask, full person mask. As less of the person is visible, A$\rightarrow$T decreases from less human attribute visual cues to bias the task prediction. T$\rightarrow$A increases because the model must rely on visual cues from the task to predict the attribute.}
\label{fig:coco_mask}
\end{table}
We introduce a scenario for validating the decoupling aspect of our metric, that simultaneously serves as inspiration for an intervention approach to mitigating bias amplification. We use a baseline amplification removal idea of applying segmentation masks (noisy or full) over the people in an image to mitigate bias stemming from human attributes~\citep{wang2019balanced}. We train on the COCO classification task, with the same 66 objects from~\citet{zhao2017menshop}, a VGG16~\citep{simonyan2014vgg} model pretrained on ImageNet~\citep{imagenet} to predict objects and gender, with a Binary Cross Entropy Loss over all outputs, and measure \biasampoursTA{} and \biasampoursAT{}; we report 95\% confidence intervals for 5 runs of each scenario. In Tbl.~\ref{fig:coco_mask} we see, as expected, that as less of the person is visible, A$\rightarrow$T decreases because there are less human attribute visual cues to bias the task prediction. On the other hand, T$\rightarrow$A increases because the model must lean into task biases to predict the person's attribute. However, we can also see from the overlapping confidence intervals that the difference between noisy and full masks does not
appear to be particularly robust; we continue a discussion of this phenomenon in Sec.~\ref{sec:consistency}.
Further mitigation techniques are outside of our scope, but we look to works like~\citet{singh2020context, wang2019balanced, agarwal2020spurious}.
\subsection{Considerations of T $\rightarrow$ A Bias Amplification}
\label{sec:ta_bias}
\begin{figure}[t]
\centering
\includegraphics[width=1.\linewidth]{Images/new_captions.png}
\caption{\label{fig:caption}Illustrative captions from the Equalizer model~\citep{hendricks2018snowboard}, which in these captions decreases $T\rightarrow A$ bias amplification from the Baseline, but inadvertently increases $A\rightarrow T$. Green underlined words are correct, and red italicized words are incorrect. In the images, the Equalizer improves on the Baseline for the gendered word, but introduces biased errors in the captions.}
\end{figure}
If we think more deeply about these bias amplifications, we might come to a normative conclusion that $T\rightarrow A$, which measures sensitive attribute predictions conditioned on the tasks, should not exist in the first place. There are very few situations in which predicting sensitive attributes makes sense~\citep{scheuerman2020identity, larson2017gender},
so we should carefully consider if this is strictly necessary for target applications. For the image domains discussed, by simply removing the notion of predicting gender, we trivially remove all $T\rightarrow A$ bias amplification. In a similar vein, there has been great work done on reducing gender bias in image captions~\citep{hendricks2018snowboard, tang2020caption}, but it is often focused on targeting $T\rightarrow A$ rather than $A\rightarrow T$ amplification. When disentangling the directions of bias, we find that the Equalizer model~\citep{hendricks2018snowboard}, which was trained with the intention of increasing the quality of gender-specific words in captions, inadvertently increases $A\rightarrow T$ bias amplification for certain tasks. We treat gender as the attribute and the objects as different tasks. In Fig.~\ref{fig:caption} we see examples where the content of the Equalizer's caption exhibits bias coming from the person's attribute.
Even though the Equalizer model reduces $T\rightarrow A$ bias amplification in these images, it inadvertently increases $A\rightarrow T$. It is important to disentangle the two directions of bias and notice that while one direction is becoming more fair, another is actually becoming more biased. Although this may not always be the case, depending on the downstream application~\cite{bennett2021accessibility}, perhaps we could consider simply replacing all instances of gendered words like ``man" and ``woman" in the captions with ``person" to trivially eliminate $T\rightarrow A$, and focus on $A\rightarrow T$ bias amplification.
Specifically when gender is the sensitive attribute, ~\citet{keyes2018agr} thoroughly explains how we should carefully think about why we might implement Automatic Gender Recognition (AGR), and avoid doing so.
On the other hand, sensitive attribute labels, ideally from self-disclosure, can be very useful. For example, these labels are necessary to measure $A\rightarrow T$ amplification, which is important to discover, as we do not want our prediction task to be biased for or against people with certain attributes.
\subsection{Variance in Estimator Bias}
\label{sec:consistency}
Evaluation metrics, ours included, are specific to each model on each dataset. Under common loss functions such as cross entropy loss, some evaluation metrics like average precision are not very sensitive to random seed. However, bias amplification, along with other fairness metrics like FPR difference, often fluctuates greatly across runs. Because the loss functions that machine learning practitioners tend to default to using are proxies for accuracy, it makes sense that various local minima, while equal in accuracy, are not necessarily equal for other measurements. The phenomena of differences between equally predictive models has been termed the Rashomon Effect~\citep{breiman2001cultures}, or predictive multiplicity~\citep{marx2020multiplicity}.
Thus, like previous work~\citep{fisher2019wrong}, we urge transparency, and advocate for the inclusion of confidence intervals. To illustrate the need for this, we look at the facial image domain of CelebA~\citep{liu2015faceattributes}, defining two different scenarios of the classification of \texttt{big nose} or \texttt{young} as our task, and treating the gender labels as our attribute.
Note that we do not classify gender, for reasons raised in Sec.~\ref{sec:ta_bias}, so we only measure $A \rightarrow T$ amplification. For these tasks, women are correlated with no big nose and being young, and men with big nose and not being young. We examine two different scenarios, one where our independent variable is model architecture, and another where it is the ratio between number of images of the majority groups (e.g., young women and not young men) and minority groups (e.g., not young women and young men). By looking at the confidence intervals, we can determine which condition allows us to draw reliable conclusions about the impact of that variable on bias amplification.
For model architecture, we train 3 models pretrained on ImageNet~\citep{imagenet} across 5 runs: ResNet18~\citep{he16resnet}, AlexNet~\citep{alexnet}, and VGG16~\citep{simonyan2014vgg}. Training details
are in
Appendix A.2.
In Fig.~\ref{fig:models} we see from the confidence intervals that while model architecture does not result in differing enough of bias amplification to conclude anything about the relative fairness of these models, across-ratio differences are significant enough to draw conclusions about the impact of this ratio on bias amplification. We encourage researchers to include confidence intervals so that findings are more robust to random fluctuations. Concurrent work covers this multiplicity phenomenon in detail~\citep{damour2020underspecification}, and calls for more application-specific specifications that would constrain the model space. However, that may not always be feasible, so for now our proposal of error bars is more general and immediately implementable. In a survey of recently published fairness papers from prominent machine learning conferences, we found that 25 of 48 (52\%) reported results of a fairness metric without error bars (details in
Appendix A.2.
Even if the model itself is deterministic, error bars could be generated through bootstrapping~\citep{efron1992bootstrap} to account for the fact that the test set itself is but a sample of the population, or varying the train-test splits~\citep{friedler2019comparative}.
\begin{figure*}[t]
\includegraphics[width=0.94\linewidth]{Images/multiplicity.png}
\caption{\label{fig:models}We investigate the consistency of various metrics by looking at 95\% confidence intervals as we manipulate two independent variables: model architecture (left three graphs), and majority to minority groups ratio (right graph). The top row (blue) is for the attribute of \texttt{big nose}, and bottom row (orange) is for \texttt{young}. For model architecture, across 5 runs, the accuracy measure of average precision retains a consistent ranking across models, but two different fairness measures (FPR difference and $A\rightarrow T$ bias amplification) have overlapping intervals. This does not allow us to draw conclusions about the differing fairness of these models. However, across-ratio differences in bias amplification are significant enough to allow us to draw conclusions about the differing levels of fairness.}
\end{figure*}
\subsection{Limitations of Bias Amplification}
\label{sec:usecase}
An implicit assumption that motivates bias amplification metrics, including ours, is that the ground truth exists and is known. Further, a perfectly accurate model can be considered perfectly fair, despite the presence of task-attribute correlations in the training data. This allows us to treat the disparity between the correlations in the input vs correlations in the output as a fairness measure.
It follows that bias amplification would {\em not} be a good way to measure fairness when the ground truth is either unknown or does not correspond to desired classification. In this section, we discuss two types of applications where bias amplification should not necessarily be used out-of-the-box as a fairness metric.
\smallsec{Sentence completion: no ground truth}
\label{sec:nlp}
Consider the fill-in-the-blank NLP task, where there is no ground truth for how to fill in a sentence. Given ``The [blank] went on a walk", a variety of words could be suitable. Therefore, to measure bias amplification in this setting, we need to subjectively set the base correlations, i.e., $P(T_t=1|A_a=1), P(A_a=1|T_t=1)$.
To see the effect of adjusting base correlations, we test the bias amplification between occupations and gender pronouns, conditioning on the pronoun and filling in the occupation and vice versa. In Tbl.~\ref{tbl:nlp}, we report our measured bias amplification results on the FitBERT (Fill in the blanks BERT)~\citep{havens2019fitbert, devlin2019bert} model using various sources as our base correlation of bias from which amplification is measured. The same outputs from the model are used for each set of pronouns, and the independent variable we manipulate is the source of base correlations: 1) equality amongst the pronouns, using two pronouns (he/she) 2) equality amongst the pronouns, using three pronouns (he/she/they) 3) co-occurrence counts from English Wikipedia (one of the datasets BERT was trained on), and 4) WinoBias~\citep{zhao2018winobias} with additional information supplemented from the 2016 U.S. Labor Force Statistics data.
Additional details are in
Appendix A.3.
We find that relative to U.S. Labor Force data on these particular occupations, FitBERT actually exhibits no bias amplification. Yet it would be simplistic to conclude that FitBERT presents no fairness concerns with respect to gender and occupation. For one, it is evident from Fig.~\ref{fig:occu_bert} that there is an overall bias towards ``he" (this translates to a bias amplification for some occupations and a bias reduction for others; the effects roughly cancel out in our bias amplification metric when aggregated). More importantly, whether U.S. labor statistics are the right source of base correlations depends on the specific application of the model and the cultural context in which it is deployed. This is clear when noticing that the measured \biasampoursTA{} is much stronger when the gender distribution is expected to be uniform, instead of gender-biased Labor statistics.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{Images/occu_scatter.png}
\caption{\label{fig:occu_bert}Each point represents an occupation's probability at being associated with the pronoun for a man. FitBERT perpetuates gender-occupation biases seen in the U.S. Labor Force, and additionally over-favors the pronoun for men.}
\end{figure}
\begin{table}
\begin{center}
\begin{tabular}{ |>{\centering\arraybackslash}p{2.94cm}|>{\centering\arraybackslash}p{2.1cm}|>{\centering\arraybackslash}p{2.1cm}| }
\hline
Base Correlation Source (\# pronouns) & \biasampoursTA{} & \biasampoursAT{} \\
\hline
Uniform (2) & .1368 $\pm$ .0226 & .0084 $\pm$ .0054 \\
\hline
Uniform (3) & .0914 $\pm$ .0151 & .0056 $\pm$ .0036 \\
\hline
Wikipedia (2) & .0372 $\pm$ .0307 & -.0002 $\pm$ .0043 \\
\hline
2016 U.S. Labor Force (WinoBias) (2) & -.1254 $\pm$ .0026 & -.0089 $\pm$ .0054 \\
\hline
\end{tabular}
\end{center}
\caption{\label{tbl:nlp}\biasampours{} for different base correlation sources. The value-laden choice of base correlation source depends on the downstream application.}
\end{table}
\smallsec{Risk prediction: future outcomes unknown}
Next, we examine the criminal risk prediction setting. A common statistical task in this setting is predicting the likelihood a defendant will commit a crime if released pending trial. This setting has two important differences compared to computer vision detection tasks: 1) The training labels typically come from arrest records and suffer from problems like historical and selection bias~\citep{suresh2019framework, olteanu2019social, green2020recidivism}, and 2) the task is to predict future events and thus the outcome is not knowable at prediction time. Further, the risk of recidivism is not a static, immutable trait of a person. Given the input features that are used to represent individuals, one could imagine an individual with a set of features who does recidivate, and one who does not. In contrast, for a task like image classification, two instances with the same pixel values will always have the same labels (if the ground truth labels are accurate).
As a result of these setting differences, risk prediction tools may be considered unfair even if they exhibit no bias amplification. Indeed, one might argue that a model that shows no bias amplification is necessarily unfair as it perpetuates past biases reflected in the training data. Further, modeling risk as immutable misses the opportunity for intervention to change the risk~\cite{barabas2018intervention}. Thus, matching the training correlations should not be the intended goal~\citep{wick2019tradeoff, hebertjohnson2018multicalibration}.
To make this more concrete, in Fig.~\ref{fig:compas} we show the metrics
of \biasampoursAT{} and False Positive Rate (FPR) disparity measured on COMPAS predictions~\citep{angwin2016compas}, only looking at two racial groups, for various values of the risk threshold. A false positive occurs when a defendant classified as high risk but does not recidivate; FPR disparity has been interpreted as measuring how unequally different groups suffer the costs of the model's errors \cite{hardt2016equalopp}. The figure shows that bias amplification is close to 0 for almost all thresholds. This is no surprise since the model was designed to be calibrated by group~\cite{flores2016compas}. However, for all realistic values of the threshold, there is a large FPR disparity. Thus, risk prediction is a setting where a lack of bias amplification should not be used to conclude that a model is fair.
Like any fairness metric, ours captures only one perspective, which is that of not amplifying already present biases. It does not require a correction for these biases. Settings that bias amplification are more suited for include those with a known truth in the labels, where matching them would desired. For example, applicable contexts include certain social media bot detection tasks where the sensitive attribute is the region of origin, as bot detection methods may be biased against names from certain areas. More broadly, it is crucial that we pick fairness metrics thoughtfully when deciding how to evaluate a model.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{Images/recidivism_thresholds.png}
\caption{\label{fig:compas}COMPAS risk predictions exhibit FPR disparities, but little bias amplification. Bias amplification measures only whether the model matches the (biased) training data, not the bias of the overall system.}
\end{figure}
\subsection{Notation}
We describe the existing metric~\cite{zhao2017menshop} and highlight shortcomings that we address in Sec.~\ref{sec:our_metrix_exp}.
\subsection{Notation}
\label{sec:biasamptheirs-notation}
Let $\mathcal{A}$ be the set of protected demographic groups: for example, $\mathcal{A} = \{$woman, man$\}$ in Fig.~\ref{fig:fig1}. $A_a$ for $a\in \mathcal{A}$ is the binary random variable corresponding to the presence of the group $a$; thus $P(A_{\texttt{woman}}= 1)$ can be empirically estimated as the fraction of images in the dataset containing women. Note that this formulation is generic enough to allow for multiple protected attributes and intersecting protected groups. Let $T_t$ with $t\in\mathcal{T}$ similarly correspond to binary target tasks, e.g., $\mathcal{T}= \{$painting$\}$ in Fig.~\ref{fig:fig1}.
\subsection{Formulation and shortcomings}
\label{sec:biasamptheirs-shortcomings}
Using this notation, \citet{zhao2017menshop}'s metric is:
\begin{align}
\label{eqn:menshop}
\text{\biasamptheirs}&=
\frac{1}{|\mathcal{T}|}\sum_{a\in \mathcal{A}, t\in\mathcal{T}}y_{at}\Delta_{at}\\
\text{with } y_{at} &= \mathbbm{1}\left[ P(A_a=1|T_t=1) > \frac{1}{|\mathcal{A}|} \right ]
\notag \\
\Delta_{at} &= P(\hat{A}_a=1|\hat{T}_t=1) - P(A_a=1|T_t=1) \notag
\end{align}
where $\hat{A}_a$ and $\hat{T}_t$ denote model predictions for the protected group $a$ and the target task $t$ respectively.
One attractive property of this metric is that it doesn't require any ground truth test labels: assuming the training and test distributions are the same, $P(A_a=1|T_t=1)$ can be estimated on the training set, and $P(\hat{A}_a=1|\hat{T}_t=1)$ relies only on the \emph{predicted} test labels. However, it also has a number of shortcomings.\
\smallsec{Shortcoming 1: The metric focuses only on \emph{positive} correlations} This may lead to numerical inconsistencies, especially in cases with multiple protected groups.
To illustrate, consider a scenario with 3 protected groups $A_1$, $A_2$, and $A_3$ (disjoint; every person belongs to one), one binary task $T$, and the following dataset\footnote{For the rest of this subsection, for simplicity since we have only one task, we drop the subscript $t$ so that $T_t$, $y_{at}$ and $\Delta_{at}$ become $T$, $y_a$ and $\Delta_a$ respectively. Further, assume the training and test datasets have the same number of examples (exs.).}
:
\begin{itemize}[noitemsep,topsep=0pt]
\item When $A_1=1$: {\bf 10} exs. of $T=0$ and {\bf 40} exs. of $T=1$
\item When $A_2=1$: {\bf 40} exs. of $T=0$ and {\bf 10} exs. of $T=1$
\item When $A_3=1$: {\bf 10} exs. of $T=0$ and {\bf 20} exs. of $T=1$
\end{itemize}
From Eq.~\ref{eqn:menshop}, we see $y_1=1$, $y_2=0$, $y_3=0$. Now consider a model that always makes correct predictions of the protected attribute $\hat{A}_a$, always correctly predicts the target task when $A_1=1$, but predicts $\hat{T}=0$ whenever $A_2=1$ and $\hat{T}=1$ whenever $A_3=1$. Intuitively, this would correspond to a case of overall learned bias amplification. However, Eqn.~\ref{eqn:menshop} would measure bias amplification as 0 since the strongest positive correlation (in the $A_1=1$ group) is not amplified.
Note that this issue doesn't arise as prominently when there are only 2 disjoint protected groups (binary attributes), which was the case implicitly considered in \citet{zhao2017menshop}. However, even with two groups there are miscalibration concerns. For example, consider the dataset above but only with the $A_1=1$ and $A_2=1$ examples. A model that correctly predicts the protected attribute $\hat{A}_a$, correctly predicts the task on $A_1=1$, yet predicts $\hat{T}=0$ whenever $A_2=1$ would have bias amplification value of $\Delta_1 = \frac{40}{40}-\frac{40}{50} = 0.2$. However, a similar model that now correctly predicts the task on $A_2=1$ but always predicts $\hat{T}=1$ on $A_1=1$ would have a much smaller bias amplification value of $\Delta_1 = \frac{50}{60}-\frac{40}{50} = 0.033$, although intuitively the amount of bias amplification is the same.
\smallsec{Shortcoming 2: The chosen protected group may not be \emph{correct} due to imbalance between groups} To illustrate, consider a scenario with 2 disjoint protected groups:
\begin{itemize}[noitemsep,topsep=0pt]
\item When $A_1=1$: {\bf 60} exs. of $T=0$ and {\bf 30} exs. of $T=1$
\item When $A_2=1$: {\bf 10} exs. of $T=0$ and {\bf 20} exs. of $T=1$
\end{itemize}
We calculate $y_1 = \mathbbm{1}\left[\frac{30}{50} > \frac{1}{2} \right ] = 1$ and $y_2 = 0$, even though the correlation is actually the reverse. Now a model, which always predicts $\hat{A}_a$ correctly, but intuitively amplifies bias by predicting $\hat{T}=0$ whenever $A_1=1$ and predicting $\hat{T}=1$ whenever $A_2=1$ would actually get a \emph{negative} bias amplification score of $\frac{0}{30}-\frac{30}{50} = -0.6$.
\biasamptheirs{} erroneously focuses on the protected group with the most examples ($A_1=1$) rather than on the protected group that is actually correlated with $T=1$ ($A_2=1$). This situation manifests when $\min\left(\frac{1}{|\mathcal{A}|}, P(A_a=1)\right)<P(A_a=1|T_t=1)<\max\left(\frac{1}{|\mathcal{A}|}, P(A_a=1)\right)$, which is more likely to arise as the the distribution of attribute $A_a=1$ becomes more skewed.
\smallsec{Shortcoming 3: The metric entangles \emph{directions} of bias amplification} By considering only the predictions rather than the ground truth labels at test time, we are unable to distinguish between errors stemming from $\hat{A}_a$ and those from $\hat{T}$. For example, looking at just the test predictions we may know that the prediction pair $\hat{T}=1,\hat{A_1}=1$ is overrepresented, but do not know whether this is due to overpredicting $\hat{T}=1$ on images with $A_1=1$ or vice versa.
\subsection{Experimental analysis}
\label{sec:biasamptheirs-COCO}
To verify that the above shortcomings manifest in practical settings, we revisit the analysis of \citet{zhao2017menshop} on the COCO~\cite{lin14coco} image dataset with two disjoint protected groups $A_{\texttt{woman}}$ and $A_{\texttt{man}}$, and 66 binary target tasks, $T_t$, corresponding to the presence of 66 objects in the images. We directly use the released model predictions of $\hat{A}_a$ and $\hat{T}_t$ from \citet{zhao2017menshop}.
First, we observe that in COCO there are about 2.5x as many men as women, leading to shortcoming 2 above. Consider the object \texttt{oven}; \biasamptheirs{} calculates
$ P(A_{\texttt{man}}=1|T_{\texttt{oven}}=1)=0.56>\frac{1}{2}$
and thus considers this to be correlated with men rather than women. However, computing $P(A_{\texttt{man}}=1,T_{\texttt{oven}}=1) = 0.0103 < 0.0129 = P(A_{\texttt{man}}=1)P(T_{\texttt{oven}}=1)$
reveals that men are in fact \emph{not} correlated with oven, and this result stems from the fact that men are overrepresented in the dataset generally. Not surprisingly, the model trained on this data associates women with ovens and underpredicts men with ovens at test time, i.e., $P(\hat{A}_{\texttt{man}}=1 | \hat{T}_{\texttt{oven}}=1) - P(A_{\texttt{man}}=1 | T_{\texttt{oven}}=1) = -0.10$, erroneously measuring \emph{negative} bias amplification.
In terms of directions of bias amplification, we recall that \citet{zhao2017menshop} discovers that ``Technology oriented categories
initially biased toward men such as \texttt{keyboard}... have each increased their bias toward
males by over 0.100." Concretely, from Eqn.~\ref{eqn:menshop}, $P(A_{\texttt{man}}=1|T_{\texttt{keyboard}}=1) = .70$ and $P(\hat{A}_{\texttt{man}}=1|\hat{T}_{\texttt{keyboard}}=1) = .83$, demonstrating an amplification of bias. However, the \emph{direction} or cause of bias amplification remains unclear: is the presence of man in the image increasing the probability of predicting a keyboard, or vice versa? Looking more closely at the model's disentangled predictions, we see that when conditioning on the attribute, $P(\hat{T}_{\texttt{keyboard}}=1|A_{\texttt{man}}=1) = 0.0020 < 0.0032 = P(T_{\texttt{keyboard}}=1|A_{\texttt{man}}=1)$, and when conditioning on the task,
$P(\hat{A}_{\texttt{man}}=1|T_{\texttt{keyboard}}=1) = 0.78 > 0.70 = P(A_{\texttt{man}}=1|T_{\texttt{keyboard}}=1)$,
indicating that while keyboards are under-predicted on images with men, men are over-predicted on images with keyboards. Thus the root cause of this amplification appears to be in the gender predictor rather than the object detector. Such disentangement allows us to properly focus algorithmic intervention efforts.
Finally, we make one last observation regarding the results of \citet{zhao2017menshop}. The overall bias amplification is measured to be $.040$. However, we observe that ``man" is being predicted at a higher rate (75.6\%) than is actually present (71.2\%). With this insight, we tune the decision threshold on the validation set such that the gender predictor is well-calibrated to be predicting the same percentage of images to have men as the dataset actually has. When we calculate \biasamptheirs{} on these newly thresholded predictions for the test set, we see bias amplification drop from $0.040$ to $0.001$ just as a result of this threshold change, outperforming even the solution proposed in \citet{zhao2017menshop} of corpus-level constraints, which achieved a drop to only $0.021$. Fairness can be quite sensitive to the threshold chosen~\citep{chen2020threshold}, so careful threshold selection should be done, rather than using a default of $0.5$.
When a threshold is needed in our experiments, we pick it to be well-calibrated on the validation set. In other words, we estimate the expected proportion $p$ of positive labels from the training set and choose a threshold such that on $N$ validation examples, the $Np$ highest-scoring are predicted positive. Although we do not take this approach, because at deployment time it is often the case that discrete predictions are required, one could also imagine integrating bias amplification across threshold to have a threshold-agnostic measure of bias amplification, similar to what is proposed by \citet{chen2020threshold}.
\iffalse
To illustrate, first consider a scenario with 2 (disjoint) protected groups with the following dataset and labels:
\begin{tabular}{p{0.4\linewidth}p{0.4\linewidth}}
$A_1=1,T=0$: 10 exs. & $A_1=1,T=1$: 40 exs.\\
$A_2=1,T=0$: 40 exs. & $A_2=1,T=1$: 10 exs.\\
\end{tabular}
We compute $P(A_1=1|T=1) = \frac{40}{50} = 0.8$, so $y_1 = 1$ while $y_0 = 0$. Eq.~\ref{eqn:menshop} thus reduces in this case to $\text{\biasamptheirs} = P(\hat{A}_1=1|\hat{T}=1) - 0.8$.
Now consider two models, which both make mostly correct predictions of $\hat{A}_a$ and $\hat{T}$, with two exceptions: the first model always predicts the task as $\hat{T}=1$ on protected group $A_1=1$, and the second model always predicts the task as $\hat{T}=0$ on protected group $A_2=1$. Intuitively, these models amplify bias in symmetric ways and by the same amount. However, the metric is \emph{not} symmetric, so the first model has $\text{\biasamptheirs}$ of $\frac{50}{60} - 0.8 = 0.167$ while the second model's is higher at
$\frac{40}{40}-0.8 = 0.2$.
For a more troubling example, consider further a scenario with 3 (still disjoint) protected groups, so in addition to the examples above we also have
\begin{tabular}{p{0.4\linewidth}p{0.4\linewidth}}
$A_3=1,T=0$: 10 exs. & $A_3=1,T=1$: 20 exs.\\
\end{tabular}
. Now consider a model that always makes correct predictions of the protected attribute $\hat{A}_a$, always correctly predicts the target task on group 1, but always predicts $\hat{T}=0$ on group 2 and $\hat{T}=1$ on group 3. Intuitively, this would correspond to a case of overall learned bias amplification. However, Eqn.~\ref{eqn:menshop} would measure bias amplification at 0 since the strong positive correlation of group 1 is not amplified.
Note that this issue doesn't arise as prominently when there are only 2 protected groups (with every example corresponding to exactly one group), which was the case implicitly considered in \cite{zhao2017menshop}. However, even with two groups there are miscalibration concerns. For example, consider a dataset consisting only of Groups 1 and 2 above. A model that correctly predicts the protected attribute $\hat{A}_a$, correctly predicts the task on group 1, and always predicts $\hat{T}=0$ on group 2 would have bias amplification value of $\Delta_1 = \frac{40}{40}-\frac{40}{50} = 0.2$. However, a similar model that now correctly predicts the task on group 2 but always predicts $\hat{T}=1$ on group 1 would have a bias amplification value of $\Delta_1 = \frac{40}{40}-\frac{50}{60} = 0.17$, despite the fact that intuitively the amount of bias amplification is the same.
To illustrate, consider a scenario with 3 (disjoint) protected groups with the following dataset\footnote{Assume for simplicity the training and test datasets have the same number of examples (exs.).}:
\begin{itemize}[noitemsep,topsep=0pt]
\item Group 1: 10 exs. with $T=0$ and 40 exs. with $T=1$
\item Group 2: 40 exs. with $T=0$ and 10 exs. with $T=1$
\item Group 3: 10 exs. with $T=0$ and 20 exs. with $T=1$
\end{itemize}
From Eq.~\ref{eqn:menshop}, we see $y_a=1$ only for group $a=1$, and $0$ otherwise. Now consider a model that always makes correct predictions of the protected attribute $\hat{A}_a$, always correctly predicts the target task on group 1, but always predicts $\hat{T}=0$ on group 2 and $\hat{T}=1$ on group 3. Intuitively, this would correspond to a case of overall learned bias amplification. However, Eqn.~\ref{eqn:menshop} would measure bias amplification at 0 since the strong positive correlation of group 1 is not amplified.
Note that this issue doesn't arise as prominently when there are only 2 protected groups (with every example corresponding to exactly one group), which was the case implicitly considered in \cite{zhao2017menshop}. However, even with two groups there are miscalibration concerns. For example, consider a dataset consisting only of Groups 1 and 2 above. A model that correctly predicts the protected attribute $\hat{A}_a$, correctly predicts the task on group 1, and always predicts $\hat{T}=0$ on group 2 would have bias amplification value of $\Delta_1 = \frac{40}{40}-\frac{40}{50} = 0.2$. However, a similar model that now correctly predicts the task on group 2 but always predicts $\hat{T}=1$ on group 1 would have a bias amplification value of $\Delta_1 = \frac{40}{40}-\frac{50}{60} = 0.17$, despite the fact that intuitively the amount of bias amplification is the same.
\subsection{Overview of Existing Fairness Metrics}
\label{sec:error_types}
We begin with a summary of existing fairness metrics and then demonstrate their behavior (and shortcomings) with an example. Four fairness metrics we consider are: {\it False Positive Rate (FPR) and True Positive Rate (TPR) difference:} the difference in false positive (true positive) rate of predicting a single label $T$ on images of $A_0$ versus $A_1$ (for two protected groups)~\citep{chouldechova2016recidivism, hardt2016equalopp}, {\it accuracy difference:} difference between the task prediction accuracy on images of $A_0$ versus on images of $A_1$~\citep{berk2017risk}, and {\it mean accuracy across subgroups:} mean task prediction accuracy across the image subgroups in $\mathcal{A} \times \mathcal{T}$
the four image subgroups ($(A_0, T_0), (A_1, T_0), (A_0, T_1), \text{and } (A_1, T_1)$)~\citep{buolamwini2018gendershades}.
Fig.~\ref{fig:metric_comp} shows the behavior of fairness metrics under varying amounts of learned amplification.
However, these four metrics are not designed to account for the training correlations, and unable to distinguish between cases of increased or decreased learned correlations.
\cite{zhao2017menshop} introduced an alternative to these that explicitly captures the notion of bias amplification. Concretely, they consider $P(A_a=1|T_t=1)$ of the training data as the fraction of times a protected attribute $A_a \in \mathcal{A}$ appears on images corresponding to task $T_t \in \mathcal{T}$. They then compare this with the test-time predictions made by the model, $P(\hat{A}_a=1|\hat{T}_t=1)$, or the number of times attribute $A_a$ is predicted on images where the task is predicted as $T_t$, which allows them to measure bias amplification in the absence of any additional annotations on the hold-out set. Note that in this formulation they are assuming the model is making predictions for both the task and the attribute. The full bias amplification metric (reformulated in our terms), is
\begin{multline}
\text{\biasamptheirs}=
\frac{1}{|\mathcal{T}|}
\sum_{t=1}^{|\mathcal{T}|} \sum_{a=1}^{|\mathcal{A}|} \underbrace{\mathbbm{1}\left( P(A_a=1|T_t=1) > \frac{1}{|\mathcal{A}|}\right)}_{y(t, a)}
\\ \underbrace{\left ( P(\hat{A}_a=1|\hat{T}_t=1) - P(A_a=1|T_t=1) \right )}_{\Delta(t, a)}
\label{eqn:menshop}
\end{multline}
We consider again the example from Fig.~\ref{fig:fig1} where on this dataset women ($A_0=1$, or $A_0$ for convenience) are correlated with painting ($T_0$), and men ($A_1$) with not painting ($T_1$). The set of attributes, $\mathcal{A}=\{0, 1\}$ in this example, corresponds to $A_0$ and $A_1$, each of which is a binary indicator variable. Later on in our proposed metric, this expresses our explicit and deliberate allowance for situations in which a person may identify with zero, one, or more than one attributes. The set of tasks, $\mathcal{T}=\{0, 1\}$ in this example, follows a similar formulation of binary indicators to allow for multi-label settings. We imagine $N$ images each of $(A_0,T_0)$ and $(A_1,T_1)$ but only $N/2$ images of $(A_0,T_1)$ and $(A_1,T_0)$. A classifier trained to recognize painting on this data is likely to learn these associations and over-predict painting on images of women and under-predict painting on images of men; however, algorithmic interventions may counteract this and in fact result in the opposite behavior.
Fig.~\ref{fig:metric_comp} empirically demonstrates that this metric is able to capture the level of increasing bias amplification. (For consistency in comparison with prior metrics, we assume the model always correctly predicts the protected attribute $A$.) However, as we discuss in the next section, there are some properties of bias amplification that this metric is not able to capture: for example, it does not distinguish between errors in predicting the protected attribute versus errors in predicting the task. Thus we introduce a new metric (last graph of Fig.~\ref{fig:metric_comp}) which maintains the desirable properties from \cite{zhao2017menshop} while including a number of innovations.
\begin{figure}[t]
\centering
\begin{subfigure}{}
\includegraphics[width=.9\linewidth]{Images/metric_comp.png}
\end{subfigure}
\caption{\label{fig:metric_comp}
Different fairness metrics vary in how they respond to model errors. In our image dataset (Fig.~\ref{fig:fig1}) of predicting someone who is a woman or man to be painting or not, we consider a classifier that always predicts the task correctly for men, but varies for women. The x-axes of the graphs correspond to the fraction of women predicted to be painting, where the ground-truth is $2/3$, and the model does not predict false positives before this point. The first two metrics, FPR and TPR difference, only capture one of under- or overclassification. The next two metrics are symmetric around $2/3$, being unable to differentiate under- or overclassification. Both bias amplification metrics are able to distinguish between under- and overclassification.
}
\end{figure}
\subsection{Shortcomings of BiasAmp$_{\textnormal{MALS}}$}
\label{sec:biasamptheirs}
Despite the advantages of \biasamptheirs{} (Eqn.~\ref{eqn:menshop}) and its ability to distinguish under- and overclassifications of training correlations, this metric also suffers from a number of shortcomings. To ground our discussion, we will work directly with the model outputs released by~\citet{zhao2017menshop} from their Conditional Random Field (CRF) model on COCO~\citep{lin14coco}, which has predictions for gender and objects detected for each image.
\subsubsection{Non-binary attributes}
\label{sec:nonbinary}
The first shortcoming is the metric assumes the protected attributes are binary, limiting its use: the indicator variable $y(t,a)$ only counts an increase in positive correlations, and not a decrease in negative correlations, in multi-attribute scenarios.
Consider a task $T_0$ such that $A_0$ is associated with it, but none of the other $A_i$, where $i\in\mathcal{A}, i\neq 0$. In this scenario, $\Delta(T_0, A_i)$ is only measured when there is one other $A_i$ such that $A_i=\lnot A_0$, since $\Delta(t, a)=-\Delta(t, \lnot a)$.
A simple addition of $-\Delta$ for all $A_i$'s when $y$ is 0 ensures that when there are more than two groups, their bias amplification is also counted.
\subsubsection{Base rates}
\label{sec:baserates}
The second shortcoming is that the metric does not take into account the base rates of each attribute.
\fi
\subsection{Directions for Mitigation}
\begin{table}
\includegraphics[width=.95\linewidth]{Images/coco_masks.png}
\caption{Bias amplification, as measured on the test set, changes across three image conditions: original, noisy person mask, full person mask. \biasamptheirs{} misleadingly makes it appear as if bias amplification \emph{increases} as the gender cues are removed. In reality, A$\rightarrow$T decreases with less visual attribute cues to bias the task prediction, while it is T$\rightarrow$A that increases as the model relies on visual cues from the task to predict the attribute.}
\label{fig:coco_mask}
\end{table}
\subsection{Experimental analysis}
We verify that our metric successfully resolves the empirical inconsistencies of Sec.~\ref{sec:biasamptheirs-shortcomings}. As expected, \biasampoursAT{} is positive at .1778 in shortcoming 1 and .3333 in 2; \biasampoursTA{} is 0 in both.
We further introduce a scenario for empirically validating the decoupling aspect of our metric. We use a baseline amplification removal idea of applying segmentation masks (noisy or full) over the people in an image to mitigate bias stemming from human attributes~\citep{wang2019balanced}. We train on the COCO classification task, with the same 66 objects from~\citet{zhao2017menshop}, a VGG16~\citep{simonyan2014vgg} model pretrained on ImageNet~\citep{imagenet} to predict objects and gender, with a Binary Cross Entropy Loss over all outputs, and measure \biasampoursTA{} and \biasampoursAT{}; we report 95\% confidence intervals for 5 runs of each scenario. In Tbl.~\ref{fig:coco_mask} we see that, misleadingly, \biasamptheirs{} reports increased amplification as gender cues are removed. However what is actually happening is, as expected, that as less of the person is visible, A$\rightarrow$T decreases because there are less human attribute visual cues to bias the task prediction. It is T$\rightarrow$A that increases because the model must lean into task biases to predict the person's attribute. However, we can also see from the overlapping confidence intervals that the difference between noisy and full masks does not
appear to be particularly robust; we continue a discussion of this phenomenon in Sec.~\ref{sec:consistency}.\footnote{This simultaneously serves as inspiration for an intervention approach to mitigating bias amplification. In
Appendix A.4
we provide a more granular analysis of this experiment, and how it can help to inform mitigation. Further mitigation techniques are outside of our scope, but we look to works like~\citet{singh2020context, wang2019balanced, agarwal2020spurious}. }
\subsection{Additional Metric Details}
\label{app:metricdetails}
We provide additional details here about \biasampours{}, as defined in Sec. 4.
In practice the indicator variable, $y_{at}$, is computed over the statistics of the training set, whereas everything else is computed over the test set. The reason behind this is that the direction of bias is determined by the existing biases in the training set.
Comparisons of the values outputted by \biasampours{} should only be done relatively. In particular, within one of the directions at a time, either $A\rightarrow T$ or $T\rightarrow A$, on one dataset. Comparing $A\rightarrow T$ to $T\rightarrow A$ directly is not a signal as to which direction of amplification is stronger.
\subsection{Details and Experiment from Variance in Estimator Bias}
\label{app:multiplicity_details}
For the models we trained in Sec. 5.2, we performed hyperparameter tuning on the validation set, and ended up using the following: ResNet18 had a learning rate of .0001, AlexNet of .0003, and VGG16 of .00014. All models were trained with stochastic gradient descent, a batch size of 64, and 10 epochs. We use the given train-validation-test split from the CelebA dataset.
Our method for surveying prominent fairness papers is as follows: on Google Scholar we performed a search for papers containing the keywords of ``fair", ``fairness", or ``bias" from the year 2015 onwards, sorted by relevance. We did this for the three conferences of 1) Conference on Neural Information Processing Systems (NeurIPS), 2) International Conference on Machine Learning (ICML), and 3) ACM Conference on Fairness, Accountability, and Transparency (FAccT). We picked these conferences because of their high reputability as machine learning conferences, and thus would serve as a good upper bound for reporting error bars on fairness evaluation metrics. We also looked at the International Conference on Learning Representations (ICLR), but the Google Scholar search turned up very few papers on fairness. From the three conferences we ended up querying, we took the first 25 papers from each, pruning those that were either: 1) not related to fairness, or 2) not containing fairness metrics for which it error bars could be relevant (e.g., theoretical or philosophical papers). Among the 48 papers that were left of the 75, if there was at least one graph or table containing a fairness metric that did not appear to be fully deterministic, and no error bars were included (even if the number reported was a mean across multiple runs), this was marked to be a ``non-error-bar" paper, of which 25 of the 48 papers looked into met this criteria.
\subsection{Details on Measuring Bias Amplification in FitBERT}
\label{app:nlp}
Here we provide additional details behind the numbers presented in Tbl. 2 in Sec. 5.3.
As noted, and done, by~\cite{liang2020big}, a large and diverse corpus of sentences is needed to sample from the large variety of contexts. However, that is out of scope for this work, where we run FitBERT on 20 sentence templates of the form ``[1) he/she/(they)] [2) is/was] a(n) [3) adjective] [4) occupation]". By varying 2) and using the top 10 most frequent adjectives from a list of adjectives~\citep{english_adj} that appear in the English Wikipedia dataset (one of the datasets BERT was trained on) that would be applicable as a descriptor for an occupation (pruning adjectives like e.g., ``which", ``left") for 3), we end up with 20 template sentences. We then alternate conditioning on 1) (to calculate A$\rightarrow$T) and 4) (to calculate T$\rightarrow$A). The 10 adjectives we ended up wtih are: new, known, single, large, small, major, French, old, short, good. We use the output probabilities rather than discrete predictions in calculating $P(\hat{A}_a=1|T_t=1)$ and $P(\hat{T}_t=1|A_a=1)$ because there is no ``right" answer in sentence completion, in contrast to object prediction, and so we want the output distribution.
When calculating the amount of bias amplification when the base rates are equal, we picked the direction of bias based on that provided by the WinoBias dataset. In practice, this can be thought of as setting the base correlation, $P(A_a=1|T_t=1)$ for a men-biased job like ``cook" to be $.5+\epsilon$ for ``he" and $.5-\epsilon$ for ``she" when there are two pronouns, and $.33+\epsilon$ for ``he" and $.33-\epsilon$ for ``she" and ``they", where in practice we used $\epsilon=1\mathrm{e-}7$. This ensures that the indicator variable, $y_{at}$ from Eq. 2, is set in the direction fo the gender bias, but the magnitudes of $\Delta_{at}$ are not affected to a significant degree.
To generate a rough approximation of what training correlation rates could look like in this domain, we look to one of the datasets that BERT was trained on, the Wikipedia dataset. We do so by simply counting the cooccurrences of all the occupations along with gendered words such as ``man", ``he", ``him", etc. There are flaws with this approach because in a sentence like ``She went to see the doctor.", the pronoun is in fact not referring to the gender of the person with the occupation. However, we leave a more accurate measurement of this to future work, as our aim for showing these results was more for demonstrative purposes illustrating the manipulation of the correlation rate, rather than in rigorously measuring the training correlation rate.
We use 32 rather than 40 occupations in WinoBias~\citep{zhao2018winobias}, because when we went to the 2016 U.S. Labor Force Statistics data~\citep{labor2016} to collect the actual numbers of each gender and occupation in order to be able to calculate $P(T_t=1|A_a=1)$, since WinoBias only had $P(A_a=1|T_t=1)$, we found 8 occupations to be too ambiguous to be able to determine the actual numbers. For example, for ``attendant", there were many different attendant jobs listed, such as ``flight attendants" and ``parking lot attendant", so we opted rather to drop these jobs from the list of 40. The 8 from the original WinoBias dataset that we ignored are: supervisor, manager, mechanician, CEO, teacher, assistant, clerk, and attendant. The first four are biased towards men, and the latter four towards women, so that we did not skew the distribution of jobs biased towards each gender.
\subsection{COCO Masking Experiment Broken Down by Object}
\label{app:brokendown_mask}
In Table~\ref{fig:coco_mask} of Sec.~\ref{sec:our_metrix_exp} we perform an experiment whereby we measure the bias amplification on COCO object detection based on the amount of masking we apply to the people in the images. We find that \biasampoursAT decreases when we apply masking, but \biasampoursTA increases when we do so. To better inform mitigation techniques, it is oftentimes helpful to take a more granular look at which objects are actually amplifying the bias. In Table~\ref{tbl:breakdown} we provide such a granular breakdown. If our goal is to target \biasampoursAT, we might note that objects like \texttt{tv} show decreasing bias amplification when the person is masked, while \texttt{dining table} stays relatively stagnant.
\begin{table}[t]
\caption{A breakdown of \biasampoursAT and \biasampoursTA by object for the masking experiment done on COCO in Table~\ref{fig:coco_mask}.}
\label{tbl:breakdown}
\centering
\begin{tabular}{p{.65in}|>{\centering\arraybackslash}p{.78in}|>{\centering\arraybackslash}p{.78in}|>{\centering\arraybackslash}p{.78in}|>{\centering\arraybackslash}p{.78in}|>{\centering\arraybackslash}p{.78in}|>{\centering\arraybackslash}p{.78in}}
\toprule
\multirow{2}{*}{Object} & \multicolumn{2}{c|}{Original} & \multicolumn{2}{c|}{Noisy Person Mask}& \multicolumn{2}{c}{Full Person Mask} \\
\cline{2-7}
& $A\rightarrow T$ & $T\rightarrow A$& $A\rightarrow T$ & $T\rightarrow A$& $A\rightarrow T$ & $T\rightarrow A$ \\
\hline
teddy bear & $-0.13 \pm 0.04$ & $1.23 \pm 0.32$ & $0.13 \pm 0.04$ & $2.27 \pm 0.34$ & $-0.09 \pm 0.04$ & $1.93 \pm 0.43$ \\
handbag & $0.44 \pm 0.14$ & $7.88 \pm 0.67$ & $0.4 \pm 0.17$ & $5.62 \pm 2.13$ & $0.24 \pm 0.13$ & $2.96 \pm 3.03$ \\
fork & $0.62 \pm 0.22$ & $7.67 \pm 1.31$ & $0.63 \pm 0.18$ & $7.45 \pm 1.65$ & $0.76 \pm 0.24$ & $6.22 \pm 0.62$ \\
cake & $-0.29 \pm 0.06$ & $5.7 \pm 0.59$ & $-0.09 \pm 0.04$ & $5.2 \pm 0.59$ & $-0.16 \pm 0.06$ & $3.3 \pm 1.7$ \\
bed & $0.01 \pm 0.04$ & $1.33 \pm 1.43$ & $0.06 \pm 0.08$ & $5.33 \pm 1.43$ & $-0.09 \pm 0.03$ & $6.67 \pm 0.0$ \\
umbrella & $-0.05 \pm 0.07$ & $9.52 \pm 3.92$ & $0.08 \pm 0.06$ & $13.33 \pm 2.64$ & $-0.04 \pm 0.11$ & $11.24 \pm 2.76$ \\
spoon & $0.06 \pm 0.06$ & $-2.91 \pm 2.78$ & $0.03 \pm 0.06$ & $-10.91 \pm 4.83$ & $0.04 \pm 0.03$ & $-15.27 \pm 4.11$ \\
giraffe & $0.21 \pm 0.13$ & $5.32 \pm 1.51$ & $0.2 \pm 0.05$ & $4.05 \pm 2.46$ & $0.02 \pm 0.04$ & $4.95 \pm 1.34$ \\
bowl & $0.28 \pm 0.04$ & $2.2 \pm 1.02$ & $0.06 \pm 0.07$ & $4.8 \pm 2.56$ & $0.18 \pm 0.12$ & $7.4 \pm 2.26$ \\
knife & $-0.22 \pm 0.12$ & $-11.74 \pm 2.39$ & $-0.35 \pm 0.12$ & $-10.43 \pm 3.15$ & $-0.31 \pm 0.05$ & $-8.84 \pm 2.18$ \\
wine glass & $-0.41 \pm 0.14$ & $-5.24 \pm 0.83$ & $-0.62 \pm 0.09$ & $-7.14 \pm 3.49$ & $-0.69 \pm 0.07$ & $-10.95 \pm 3.64$ \\
dining table & $-0.75 \pm 0.14$ & $4.53 \pm 1.02$ & $-0.76 \pm 0.14$ & $3.18 \pm 2.12$ & $-0.74 \pm 0.17$ & $4.58 \pm 1.38$ \\
cat & $0.07 \pm 0.04$ & $2.41 \pm 2.05$ & $0.19 \pm 0.05$ & $10.0 \pm 2.42$ & $0.17 \pm 0.02$ & $20.0 \pm 1.81$ \\
sink & $0.18 \pm 0.15$ & $-4.29 \pm 2.4$ & $0.11 \pm 0.07$ & $-5.03 \pm 2.6$ & $-0.12 \pm 0.1$ & $-3.19 \pm 1.5$ \\
cup & $-0.3 \pm 0.09$ & $-12.36 \pm 3.42$ & $-0.15 \pm 0.08$ & $-12.0 \pm 1.32$ & $-0.12 \pm 0.03$ & $-14.42 \pm 2.94$ \\
potted plant & $0.21 \pm 0.18$ & $6.36 \pm 5.1$ & $0.34 \pm 0.07$ & $0.0 \pm 2.82$ & $0.32 \pm 0.08$ & $-4.55 \pm 2.52$ \\
refrigerator & $-0.06 \pm 0.03$ & $9.6 \pm 1.36$ & $-0.07 \pm 0.06$ & $7.47 \pm 1.75$ & $0.01 \pm 0.03$ & $12.0 \pm 4.25$ \\
microwave & $-0.01 \pm 0.02$ & $-3.5 \pm 5.3$ & $-0.01 \pm 0.03$ & $13.0 \pm 13.32$ & $-0.03 \pm 0.03$ & $6.0 \pm 9.05$ \\
couch & $-1.35 \pm 0.16$ & $-0.25 \pm 1.24$ & $-0.94 \pm 0.23$ & $1.62 \pm 1.06$ & $-1.21 \pm 0.16$ & $0.15 \pm 0.8$ \\
oven & $0.07 \pm 0.09$ & $7.67 \pm 1.73$ & $-0.33 \pm 0.12$ & $10.67 \pm 3.0$ & $0.03 \pm 0.13$ & $13.12 \pm 1.49$ \\
sandwich & $-0.98 \pm 0.15$ & $-8.93 \pm 0.92$ & $-2.6 \pm 0.18$ & $-15.23 \pm 2.78$ & $-2.46 \pm 0.4$ & $-15.67 \pm 1.74$ \\
book & $-0.43 \pm 0.08$ & $-3.07 \pm 0.57$ & $-0.48 \pm 0.13$ & $-3.34 \pm 1.24$ & $-0.85 \pm 0.11$ & $-3.18 \pm 2.0$ \\
bottle & $0.05 \pm 0.11$ & $-7.73 \pm 1.54$ & $-0.13 \pm 0.14$ & $-8.33 \pm 3.87$ & $0.06 \pm 0.06$ & $-11.52 \pm 1.65$ \\
cell phone & $-0.09 \pm 0.13$ & $3.6 \pm 1.92$ & $0.05 \pm 0.15$ & $13.72 \pm 0.89$ & $-0.1 \pm 0.12$ & $18.72 \pm 2.36$ \\
pizza & $-0.19 \pm 0.1$ & $6.85 \pm 1.81$ & $-0.09 \pm 0.03$ & $15.17 \pm 2.41$ & $-0.38 \pm 0.12$ & $9.3 \pm 1.56$ \\
banana & $0.35 \pm 0.08$ & $4.42 \pm 0.66$ & $0.56 \pm 0.09$ & $6.1 \pm 1.06$ & $0.1 \pm 0.19$ & $5.19 \pm 0.72$ \\
toothbrush & $-0.47 \pm 0.11$ & $-2.42 \pm 2.63$ & $-0.55 \pm 0.13$ & $-4.32 \pm 5.14$ & $-0.5 \pm 0.12$ & $-11.85 \pm 4.43$ \\
tennis racket & $-0.09 \pm 0.08$ & $9.22 \pm 3.05$ & $0.0 \pm 0.09$ & $14.75 \pm 2.74$ & $0.09 \pm 0.12$ & $14.75 \pm 1.38$ \\
chair & $-0.31 \pm 0.11$ & $1.87 \pm 0.22$ & $-0.31 \pm 0.06$ & $3.85 \pm 0.37$ & $-0.31 \pm 0.05$ & $4.69 \pm 0.0$ \\
dog & $0.14 \pm 0.04$ & $0.43 \pm 0.19$ & $0.3 \pm 0.07$ & $1.6 \pm 0.29$ & $0.3 \pm 0.04$ & $1.7 \pm 0.35$ \\
donut & $-0.3 \pm 0.08$ & $-1.4 \pm 0.29$ & $-0.39 \pm 0.15$ & $-0.0 \pm 0.59$ & $-0.41 \pm 0.15$ & $0.09 \pm 0.37$ \\
suitcase & $-0.43 \pm 0.08$ & $1.96 \pm 1.12$ & $-0.09 \pm 0.01$ & $7.3 \pm 0.65$ & $-0.11 \pm 0.14$ & $8.43 \pm 1.4$ \\
laptop & $0.27 \pm 0.07$ & $4.58 \pm 3.57$ & $0.06 \pm 0.04$ & $7.67 \pm 6.33$ & $0.1 \pm 0.06$ & $17.39 \pm 5.52$ \\
hot dog & $1.48 \pm 0.12$ & $7.63 \pm 2.72$ & $1.51 \pm 0.07$ & $9.37 \pm 1.75$ & $1.86 \pm 0.14$ & $9.16 \pm 3.03$ \\
remote & $0.33 \pm 0.02$ & $9.14 \pm 2.73$ & $0.15 \pm 0.09$ & $11.03 \pm 2.59$ & $0.15 \pm 0.07$ & $18.62 \pm 3.12$ \\
clock & $0.77 \pm 0.16$ & $4.48 \pm 3.05$ & $0.62 \pm 0.08$ & $9.61 \pm 3.02$ & $0.88 \pm 0.11$ & $11.44 \pm 3.78$ \\
bench & $-0.02 \pm 0.05$ & $11.49 \pm 3.61$ & $-0.06 \pm 0.06$ & $14.68 \pm 4.73$ & $-0.04 \pm 0.06$ & $16.6 \pm 2.61$ \\
tv & $0.35 \pm 0.09$ & $5.16 \pm 4.59$ & $0.27 \pm 0.12$ & $14.19 \pm 2.58$ & $0.21 \pm 0.09$ & $18.06 \pm 3.03$ \\
mouse & $-0.22 \pm 0.06$ & $0.95 \pm 5.76$ & $-0.26 \pm 0.05$ & $4.57 \pm 4.99$ & $-0.18 \pm 0.06$ & $5.14 \pm 4.11$ \\
horse & $0.07 \pm 0.03$ & $8.13 \pm 5.17$ & $0.11 \pm 0.07$ & $13.63 \pm 1.96$ & $0.16 \pm 0.04$ & $16.59 \pm 4.92$ \\
fire hydrant & $-0.21 \pm 0.07$ & $4.71 \pm 2.98$ & $-0.15 \pm 0.07$ & $1.76 \pm 4.64$ & $-0.18 \pm 0.06$ & $5.0 \pm 6.19$ \\
keyboard & $0.01 \pm 0.08$ & $1.64 \pm 3.4$ & $-0.08 \pm 0.08$ & $17.38 \pm 4.51$ & $0.02 \pm 0.08$ & $31.15 \pm 6.74$ \\
bus & $0.02 \pm 0.04$ & $-11.33 \pm 2.15$ & $-0.17 \pm 0.07$ & $-9.0 \pm 3.53$ & $-0.1 \pm 0.04$ & $3.0 \pm 6.62$ \\
toilet & $0.26 \pm 0.06$ & $7.65 \pm 4.07$ & $0.2 \pm 0.09$ & $12.17 \pm 4.15$ & $0.19 \pm 0.06$ & $17.22 \pm 5.67$ \\
person & $-0.04 \pm 0.1$ & $-1.47 \pm 4.54$ & $0.1 \pm 0.07$ & $-4.12 \pm 5.25$ & $0.03 \pm 0.08$ & $-3.53 \pm 4.65$ \\
traffic light & $-0.2 \pm 0.03$ & $4.44 \pm 2.48$ & $-0.27 \pm 0.06$ & $6.06 \pm 1.86$ & $-0.27 \pm 0.11$ & $11.52 \pm 2.28$ \\
sports ball & $-1.44 \pm 0.2$ & $3.41 \pm 1.56$ & $-0.95 \pm 0.27$ & $3.86 \pm 1.83$ & $-1.27 \pm 0.32$ & $4.95 \pm 1.97$ \\
bicycle & $-0.23 \pm 0.07$ & $13.58 \pm 2.02$ & $0.01 \pm 0.15$ & $13.09 \pm 3.98$ & $-0.21 \pm 0.08$ & $12.72 \pm 1.83$ \\
car & $0.2 \pm 0.2$ & $3.7 \pm 1.49$ & $0.57 \pm 0.1$ & $-0.74 \pm 2.2$ & $0.32 \pm 0.1$ & $0.37 \pm 2.66$ \\
backpack & $0.01 \pm 0.03$ & $7.57 \pm 1.63$ & $0.0 \pm 0.06$ & $17.84 \pm 3.09$ & $-0.05 \pm 0.04$ & $18.38 \pm 2.53$
\end{tabular}
\end{table}
\setcounter{table}{2}
\begin{table}[t]
\caption{\textit{Continued}}
\label{tbl:breakdown}
\centering
\begin{tabular}{p{.65in}|>{\centering\arraybackslash}p{.78in}|>{\centering\arraybackslash}p{.78in}|>{\centering\arraybackslash}p{.78in}|>{\centering\arraybackslash}p{.78in}|>{\centering\arraybackslash}p{.78in}|>{\centering\arraybackslash}p{.78in}}
\toprule
\multirow{2}{*}{Object} & \multicolumn{2}{c|}{Original} & \multicolumn{2}{c|}{Noisy Person Mask}& \multicolumn{2}{c}{Full Person Mask} \\
\cline{2-7}
& $A\rightarrow T$ & $T\rightarrow A$& $A\rightarrow T$ & $T\rightarrow A$& $A\rightarrow T$ & $T\rightarrow A$ \\
\hline
train & $0.15 \pm 0.07$ & $8.47 \pm 3.33$ & $0.27 \pm 0.08$ & $16.82 \pm 1.86$ & $0.14 \pm 0.16$ & $19.28 \pm 3.97$ \\
kite & $-0.09 \pm 0.04$ & $-2.67 \pm 3.35$ & $-0.16 \pm 0.07$ & $-7.56 \pm 3.4$ & $-0.22 \pm 0.05$ & $-4.89 \pm 3.78$ \\
cow & $-0.17 \pm 0.08$ & $1.6 \pm 1.95$ & $0.11 \pm 0.08$ & $0.86 \pm 3.0$ & $0.15 \pm 0.07$ & $-3.58 \pm 2.55$ \\
skis & $0.12 \pm 0.09$ & $0.24 \pm 2.13$ & $0.25 \pm 0.03$ & $-1.31 \pm 3.01$ & $0.23 \pm 0.1$ & $-7.5 \pm 1.76$ \\
truck & $-0.27 \pm 0.04$ & $-13.33 \pm 1.68$ & $-0.25 \pm 0.12$ & $-10.26 \pm 1.42$ & $-0.16 \pm 0.03$ & $-17.95 \pm 2.46$ \\
elephant & $-0.58 \pm 0.15$ & $9.09 \pm 1.25$ & $-0.24 \pm 0.08$ & $4.29 \pm 2.94$ & $-0.63 \pm 0.06$ & $1.69 \pm 3.13$ \\
boat & $0.03 \pm 0.05$ & $2.8 \pm 1.79$ & $-0.02 \pm 0.07$ & $-0.0 \pm 2.48$ & $0.05 \pm 0.05$ & $3.2 \pm 3.06$ \\
frisbee & $0.09 \pm 0.17$ & $-1.36 \pm 1.18$ & $-0.14 \pm 0.22$ & $0.7 \pm 2.45$ & $-0.79 \pm 0.17$ & $-6.88 \pm 2.28$ \\
airplane & $0.14 \pm 0.07$ & $4.23 \pm 3.27$ & $0.15 \pm 0.07$ & $5.77 \pm 4.4$ & $0.16 \pm 0.05$ & $5.77 \pm 3.69$ \\
motorcycle & $-0.06 \pm 0.04$ & $-6.35 \pm 3.94$ & $0.01 \pm 0.06$ & $-9.04 \pm 2.42$ & $0.12 \pm 0.04$ & $-1.73 \pm 6.64$ \\
surfboard & $-0.02 \pm 0.05$ & $3.83 \pm 4.79$ & $-0.06 \pm 0.07$ & $1.74 \pm 1.36$ & $-0.08 \pm 0.06$ & $2.43 \pm 4.83$ \\
tie & $0.16 \pm 0.06$ & $3.29 \pm 2.1$ & $0.08 \pm 0.07$ & $3.53 \pm 3.45$ & $0.23 \pm 0.04$ & $4.71 \pm 4.47$ \\
snowboard & $0.4 \pm 0.13$ & $10.05 \pm 2.1$ & $0.53 \pm 0.16$ & $9.64 \pm 1.0$ & $0.41 \pm 0.22$ & $7.9 \pm 1.7$ \\
baseball bat & $0.46 \pm 0.04$ & $-3.72 \pm 2.47$ & $0.45 \pm 0.1$ & $-3.36 \pm 1.8$ & $0.23 \pm 0.12$ & $1.24 \pm 2.17$ \\
baseball glove & $-0.03 \pm 0.05$ & $27.06 \pm 4.12$ & $-0.16 \pm 0.04$ & $34.9 \pm 1.29$ & $-0.1 \pm 0.03$ & $36.47 \pm 0.84$ \\
skateboard & $-0.28 \pm 0.03$ & $11.37 \pm 3.51$ & $-0.34 \pm 0.08$ & $10.98 \pm 3.86$ & $-0.35 \pm 0.11$ & $5.88 \pm 3.77$ \\
\bottomrule
\end{tabular}
\end{table}
\section{Introduction}
\input{Input_Files/Intro}
\section{Related Work}
\input{Input_Files/Related_Work}
\section{Existing Bias Amplification Metric}
\input{Input_Files/Existing_Metrics}
\section{BiasAmp$_{\rightarrow}$ (Directional Bias Amplification)}
\input{Input_Files/Our_Metric}
\section{Analysis and Discussion}
\input{Input_Files/Normative}
\section{Conclusion}
\input{Input_Files/Conclusion}
\section*{Acknowledgements}
This material is based upon work supported by the National Science Foundation under Grant No. 1763642. We thank Sunnie S. Y. Kim, Karthik Narasimhan, Vikram Ramaswamy, Brandon Stewart, and Felix Yu for feedback. We especially thank Arvind Narayanan for significant comments and advice.
We also thank the authors of Men Also Like Shopping~\citep{zhao2017menshop} and Women Also Snowboard~\citep{hendricks2018snowboard} for uploading their model outputs and code online in a way that made it easily reproducible, and for being prompt and helpful in response to clarifications.
|
\section{Introduction}
Combinatorial Optimization Problems (COPs) arise in many real-life applications such as scheduling \cite{Mak2021,Velez2013}, planning \cite{Akartunali2015,Kruijff2018}, resource allocation \cite{Lalbakhsh2018,Mak2017}, routing \cite{Mak2007,Seixas2013}, and time-tabling \cite{Ghoniem2016,Zhou2020}. See, for example, \cite{chen2010,Gorman2020,Williams2013}, for more examples of mathematical programming applied in real-life business COPs where millions or even billions of dollars were saved.
There are several commonly-employed solution approaches for COPs. The two main branches are exact methods and heuristic methods. Mathematical Programming is an exact method that can provide proven optimal solutions, and even when it fails to produce an optimal solution within a predetermined time and memory limit, it can still provide a proven optimality gap. Heuristic approaches (such as trial-and-error, simulation, learning, meta-heuristic or custom-made problem-specific heuristic) on the other hand, do not have a solution guarantee. With meta-heuristics or learning methods, with parameters properly tuned, they may be able to provide reasonably good quality solutions within a much shorter time. Exact algorithms will always be preferred in applications where a proven optimal solution matter. For instance, in Kidney Exchange Optimization, an increment of one unit in the objective function means one more kidney transplant can be carried out, and undoubtably will have a significant impact on the health outcome of the patient with a kidney failure.
When solving a COP, mathematical Programming-based exact algorithms essentially implement an exhaustive tree search with smart pruning strategies. In the case of Integer Programming (IP)-family of methods, the theoretical basis is algebra, whereas in the case of Constraint Programming (CP), the theoretical basis is logical inferences. CP and IP each has their strengths and weaknesses. IP-family of methods include Pure Integer Programming where all decision variables are integers, Binary Integer Programming where all decision variables are binary, and Mixed-integer Linear Programming (MILP) where some decision variables are continuous, and the rest are binary or general integers. MILP can also model some nonlinear terms (e.g., quadratic, bilinear, and piecewise linear terms), and therefore MILP is a very practical technique in modelling and solving real-life COPs. This is evidenced by the fact that in the history of Franz Edelman Awards, 20\% of the finalists applied IP-family of methods \cite{Gorman2020}, and that the top two algorithms in the 2$^{nd}$ Nurse Rostering Competition are MILP-based methods \cite{Ceschia2019}.
The formal mathematical specification of a MILP problem is given as follows:
$\{\min, \ \max\}$
$\{ \mathbf{c} \cdot \mathbf{x} + \mathbf{d} \cdot \mathbf{y} \ : \ A\mathbf{x} + B\mathbf{y} \leq \mathbf{f}, \ \mathbf{x} \in \mathbb{Z}^n_+, \ \mathbf{y} \in \mathbb{Q}^p_+\}$,
with $\mathbf{x}=(x_1,\ldots,x_n)$ and $\mathbf{y}=(y_1,\ldots,y_p)$ the decision variables; $\mathbf{c}$ and $\mathbf{d}$ the cost coefficients, for $\mathbf{c}$ a $n$-vector of $\mathbb{Q}$, $\mathbf{d}$ a $p$-vector of $\mathbb{Q}$; $A \in \mathbb{Z}_{m \times n}$ and $B \in \mathbb{Q}_{m \times p}$ the constraint coefficient matrices; and $\mathbf{f}$ a $m$-vector of $\mathbb{Q}$. For a thorough exposition of MILP, see, for example, \cite{Nemhauser1998,Pochet2006}.
An MILP comprises four main components: a set of decision variables, an objective function that is a linear combination of the decision variables, a set of constraints (each containing a linear combination of decision variables), and the index sets that enumerate the decision variables and constraints. We proposed an MILP ontology in \cite{Ofoghi2020}, see Figure \ref{fig:ontology} below.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=12cm]{Ontology.png}
\caption{The proposed ontology of mixed integer linear programming models for Combinatorial Optimization problems. Note: The visualization is courtesy of WebVOWL~\cite{VOWLpaper}.
}
\label{fig:ontology}
\end{center}
\end{figure*}
\section{Constraint types}
Here, we ask a fundamental question: is there a finite number of MILP constraint types, and if the answer is yes, how many are there?
Classic families of COPs such as routing, scheduling, planning have dozens or hundreds of variations from real-life applications. Every COP is different. Numerous MILPs for real-life COPs found and solved, and many more yet to be discovered and formulated. We are not able to examine every single MILP ever developed in history, however we performed two simple studies for obtaining insights. 1) We examined all constraints used in the production planning problems listed in H. Paul Williams' textbook ``Model building in mathematical programming''. 2) We examined constraints used in a number of publications.
We considered the production planning examples in H. Paul Williams' book ``Model building in mathematical programming'', and observed that constraints that represent limits (bounds), those that blending from raw materials to products, those that balance two quantities, those that governs logic conditions, and the classic binary integer programming constraints such as set partition, set packing, and set covering and their weighted variations cover all the production planning problems presented therein. In Figure \ref{fig:Williams}, we present a table where we listed the meaning of the constraints used and the section number in the textbook where the examples were discussed. These constraints are reasonably easy to interpret in the sense that the mathematical specification of the constraints is either very close to or can be directly translated from a natural language (NL) description of a n\"{a}ive end-user, (a domain-expert end-user who is not trained to develop MILP models)--we call these explicit constraints.
\begin{figure*}[ht]
\begin{center}
\includegraphics[max size={\textwidth}]{ConstraintTypes.png}
\caption{A table of constraints and the sections in which they are used in the H Paul Williams book \cite{Williams2013} for production planning problems.
}
\label{fig:Williams}
\end{center}
\end{figure*}
There are many ways to classify explicit constraints into different types. For instance, if we classify MILP constraint by their mathematical form, they can only be in one of the following forms: $\mathbf{a} \cdot \mathbf{x} \leq b$, $\mathbf{a} \cdot \mathbf{x} = b$, and $\mathbf{a} \cdot \mathbf{x} \geq b$ (here,
$b\geq 0$). For simplicity, we will write $\mathbf{a} \cdot \mathbf{x}$ as $ax$ for the rest of the paper.
{\bf Type I: Bound constraints (Demand and Supply)}
Resource limit (supply) constraints $ax \leq b$ and demand constraints $ax \geq d$ are very commonly used in MILPs, particularly in production planning-type of problems. For resource limit (supply) constraints, an upper bound on the supply can be fixed (e.g., a Knapsack constraint) or depends on the value of a decision variable. Similar for the lower bound on a demand that needs to be satisfied.
{\bf Type II: Balancing constraints}
Equality constraints $ax = b$ has many variations in its usage: to balance (equate) input and output quantity; to balance the flow or quantities over two consecutive time periods, to set initial conditions, to assign values, and so on.
{\bf Set packing/partitioning/covering constraints}
The set packing/partitioning/covering constraints are subtypes of Type I and Type II constraints, typically used for assignment or allocation. They allow us to model the choice of at most/exactly/at least one out of many. The weighted version of the set packing/partitioning/covering constraints allow us to model the choice of $n>1$ out of many.
{\bf Logic constraints}
The three main subtypes of {\bf logic constraints} are the {\em Big-M}, {\em If-then}, and {\em Either-or} constraints, each has a number of varieties. (Some of these varieties were discussed in \cite{Ofoghi2020}).
So, the next question is, what about real-life COPs other than the ones in the \cite{Williams2013} and how do we represent the knowledge in order to enable automatic mapping from business requirements to the mathematical specification (or that of a general purpose modelling language such as OPL or Minizinc)?
\section{The optimization modelling tree}
We designed an Optimization Modelling Tree (OMT) and examined a number of COPs in published journal articles to ascertain whether the OMT is adequate in the sense that by traversing through the tree all elements for the MILPs can be found. The focus of this exercise is to evaluate whether the constraint types and subtypes in the OMT are enough to represent these example COPs.
\begin{figure*}[ht]
\begin{center}
\includegraphics[width=16.5cm]{OM-Tree.png}
\caption{The Optimization Modelling Tree (OMT)
}
\label{fig:OMT}
\end{center}
\end{figure*}
{\bf A chemical production scheduling example}
A chemical production scheduling MILP model is presented in \cite{Velez2013}. The problem considers a given planning horizon, partitioned into a number of time slots. The decisions to be made are whether a unit (a machine or equipment) should start processing a task at a particular time slot, what the batch size should be, and the inventory level of each material at each time slot. A basic MILP is presented with 4 constraints. Constraint Set (1) is a set packing constraint ensuring that each unit will be starting at most one task at a time (Number 11 on the OMT). Constraint Set (2) is a combined logic (Big-M) and upper/lower bound constraint--if a unit starts processing a task at a given time, then the capacity of the batch size must be observed, otherwise, the task will not be processed on this machine at this time (Numbers 3 and 9 on the OMT). Constraint Set (3) presented in the paper should have been two constraints. One is to equate the inventory (storage) of a material at a time slot to the inventory at the previous time slot plus the new production and minus the consumption (Number 14 on the OMT). The second part of Constraint Set (3) is an upper bound on the storage limit (Number 7 on the OMT). These constraints are reasonably straight forward to describe by an end-user, and all requirements can be found in the constraint types on the OMT.
{\bf A supply chains production planning example}
An MILP model for mid-term production planning for high-tech low-volume supply chains is presented in \cite{Kruijff2018}. The decision variables are mostly general integer variables, and the six constraint sets are as follows. Constraint Set (1) are to balance two quantities, in specific, quantities between two consecutive time slots (Number 12 on the OMT). Constraint Sets (2) and (5) are equality constraints for assigning quantities (Number 13 on the OMT). Constraint Sets (3) and (6) are variable upper bounding and lower bounding (Numbers 2 and 8 on the OMT) whereas Constraint set (4) has the logic condition that the upper and lower bounds on decision variables for quantities apply only when the associated binary decision variables is non-zero (Numbers 3 and 9 on the OMT).
{\bf A university course timetabling problem example}
A university course timetabling problem was modelled as an MILP in \cite{Ghoniem2016}. The decision variables are binary. One set of the decision variables represent yes/no answers to whether a particular section of a course should be assigned to a particular professor in a particular time slot. Translating them from NL to formal specifications should be reasonably straight forward. Constraint sets (2) and (3) are Set Partitioning constraints (for choice of exactly one out of many, Number 17 on the OMT), Constraint Sets (4) to (9) are Set Packing constraints (for choice of at most one out of many, Number 11 on the OMT)). Constraint sets (11)--(13), and (15) are general if-then constraints regulating if $X$ occurs, then both of $Y$ and $Z$ must occur. The constraints are in the form of $2X \leq Y + Z$ (Number 24 on the OMT), although $X \leq Y$ and $X \leq Z$ are better constraints to use. This brings an important aspect for knowledge engineering MILPs: multiple feasible MILP constraints exist for the same requirement, some are strong for computational use than others. The OMT in its current state has some limitations, as we can see from the next example.
{\bf A multitrip vehicle routing problem with time windows example}. The COP described in \cite{Seixas2013} is a routing-type problem. An end-user not trained with MILP knowledge does not normally describe that a yes/no decision is associated with each pair of locations (e.g., $i$ and $j$ with a yes answer indicating Location $j$ must be visiting immediate after Location $i$). However, commonly-used MILPs for routing problems typically use a binary variable for each of these decisions. Once the hurdle in decision variable definition is overcome, the rest of the constraints can be found in the constraint types or subtypes described in the OMT. Constraint Sets (1), (3), (4) are all Set Partitioning Constraints, i.e., to choose exactly one out of many (Number 17 on the OMT). Constraint Set (2) is to set to zero variables that represent impossible decisions (Number 19 on the OMT). Constraint Sets (6) to (8) are to regulate the time of arrival of a vehicle route to visit a customer, and the constraints are {\em if-then} subtype be found in the OMT. The last constraint set (10) ensures is a straight forward upper bounding constraint on total time used, and the bound itself is a variables (Number 2 on the OMT).
Constraint Set (9) is a special type of demand - capacity constraint commonly used in routing-type of problems. The requirement is not trivial to describe in NL by an end-user but the mathematical constraint itself can be found in the OMT (it is in fact a Set Packing Constraint, Number 11 on the OMT). Constraint Set (5) is a flow-balance constraint (which is covered by the OMT), the mathematical meaning is that if a customer is visited, then there must be a customer that was visited before him/her and one after him/her. An end-user would not describe the requirement like this. We call these implicit constraints.
\section{Summary remarks and future research}
What the OMT contains, is not just the mathematical specification of the MILP constarints. Mathematically, $ax \leq b$, $ax = b$, and $ax \geq b$ are enough to cover all MILP constraints. However, the OMT we designed branches by usage, (or, the meaning of the constraints in application). A beginner MILP modeller, for example, can traverse through the tree to elicit business requirements from a non-expert end-user. We have tested some COP instances and the constraints in the OMT do in fact cover all the ``explicit'' (or, straight-forward) constraints. Even for constraints in our test cases that are not straight forward, i.e., the ``implicit'' constraints, they too are covered by the OMT, though the mapping mechanism is not represented on the OMT.
We have the same results for the ACs and the SECs of an ATSP, they can appear in the form of Set Partitioning and Set Covering constraints respectively, but the mapping is not explicitly represented in the tree.
In summary, we hypothesize that the OMT is scalable, however, we are unable to prove it in this paper.
\subsection{Future research}
Now, what is required appears to be the compilation of a list of mappings for commonly-used implicit constraints. For example, the knowledge that ``visiting each city exactly once and return to the home city'' is equivalent to ``each entity has one that precedes it and one that succeeds it'' is needed to be represented on this OMT, and consequently the Set Partitioning and Set Covering constraints be identified as the right ones to use. This will be the subject of our next research project: to perform a survey of literature for commonly-used implicit constraints and the usage they map to, and to represent such a mapping on the OMT in an efficient way.
\input{main.bbl}
\end{document}
|
\section{Introduction}
Noisy Intermediate-Scale Quantum (NISQ) technology is being developed rapidly and poses a great challenge to come up with efficient quantum algorithms \cite{nisq}, which will operate on the NISQ computers and perform better than classical algorithms. Many real-world use cases are associated with machine learning and optimization, for which variational quantum circuits offer an appropriate framework.
The typical optimization tasks can be formulated as a search for the ground state of a
Hamiltonian $H$, which may encode an exact combinatorial problem \cite{QAOA,Lucas_2014}.
The variational quantum algorithms (VQA) consist of two elements \cite{Peruzzo_2014}. The first part is quantum, where one constructs a parameterized quantum circuit composed of $L$ unitary layers on the product state of $n$ qubits, $|0\rangle^{\otimes n}$.
The layer unitaries and quantum gates therein depend on continuous parameters, each initialized with the uniform measure on $[0, 2\pi)$. Denoting all the circuit parameters collectively by $\theta$,
the variational state is written as
\begin{equation}
|\psi_c(\theta)\rangle = U(\theta)|0\rangle^{\otimes n} \ .
\end{equation}
The second part of the variational quantum algorithm is classical, where we estimate the Hamiltonian expectation value with the variational circuit state, i.e.,
\begin{equation}
E(\theta) = \langle \psi_c(\theta)| H |\psi_c(\theta) \rangle \ ,
\label{E}
\end{equation}
and minimize it in the $nL$-dimensional parameter space using the gradient descent method.
\begin{figure}[t]
\centering
\subfloat[]{
\centering
\label{fig:circuit_diagram_left}
\includegraphics[height=1.85cm]{./fig1a.pdf}}
\hfill
\subfloat[]{
\centering
\label{fig:circuit_diagram_right}
\includegraphics[height=1.85cm]{./fig1b.pdf}}
\caption{The circuit architecture used in Sections~\ref{sec:randomcircuit}~and~\ref{sec:opt}. (a) The horizontal axis can be interpreted as the discrete time $L$. We call the commuting set of simultaneous 2-qubit gates as the circuit layer. (b) Each gate consists of the single-qubit Pauli-$y$ rotations \eqref{eq:paulirot} followed by the CZ operation \eqref{eq:cz}.}
\label{fig:circuit_diagram}
\end{figure}
Entanglement encodes information in the qubit correlations, which are generated by the successive application of the circuit layers.
Given two complementary systems $A/B$, the Renyi-$k$ entropy of the reduced density matrix
\begin{equation}
{\cal R}^{k}_A = \frac{1}{1-k} \log \text{Tr} \left(\rho^k_A\right)
\label{Re}
\end{equation}
measures their entanglement, so does the von Neumann entropy that corresponds to \eqref{Re} in the special limit $k\rightarrow 1$:
\begin{equation}
S_{EE} = - \text{Tr}\, \rho_A \log \rho_A \ .
\label{SEE}
\end{equation}
The reduced density matrix $\rho_{A}$
is obtained from the full circuit density matrix $\rho_c(\theta) = |\psi_c(\theta)\rangle\langle \psi_c(\theta)|$ by taking a partial trace over the subsystem $B$.
The performance of the variational quantum algorithm depends largely on whether the quantum circuit can prepare an initial variational state $|\psi_c(\theta) \rangle$ that is close to the target ground state $|\psi_g\rangle$ of the Hamiltonian.
In this paper we argue that the average entanglement entropy \eqref{Re} or \eqref{SEE} of random circuit states provides
a distance measure that can quantify a successful minimization of the energy function. Note that, for their computation,
we specifically use the equal partition $n_A=n_B=n/2$ and the binary logarithm.
\begin{figure}[t]
\centering
\includegraphics[width=6cm]{./fig2.pdf}
\caption{
A schematic plot for the growth of mean entanglement entropies and geometric distance of circuit-generated states. We often distinguish the low, intermediate, high-depth circuits in Sections~\ref{sec:randomcircuit}~and~\ref{sec:opt} and denote their corresponding depth ranges by $A$, $B$, and $C$, respectively.
}
\label{fig:random_growth}
\end{figure}
The evolution of the entanglement entropies as a function of the circuit depth $L$ is schematically drawn in Figure~\ref{fig:random_growth}. It is convenient to divide the range of $L$ into three regions $A$, $B$, and $C$.\footnote{Such usage of $A$, $B$, $C$ should be distinguishable from the other usage of $A$, $B$ that denotes a subset of $n$ qubits.} $A$ is where the entanglement entropy continues to grow, while $C$ is where it has saturated to a constant value. As for their scaling behavior in $n$, the random circuit states in $A$/$C$ obey the area/volume law scaling of the entanglement entropies, respectively. Since the ground states of gapped local Hamiltonians are expected to have an area law entanglement entropy, we expect that an initial variational state $|\psi_c(\theta) \rangle$ in $A$ would lead to efficient VQA optimization in contrast to those circuits in $C$.
We also identify $B$ as a transition region between $A$ and $C$, where the entanglement entropy has already reached saturation yet the initial random parameter can determine the success/failure of the VQA optimization.
The technical reason why the circuit optimization fails in region $C$ is the vanishing gradient problem. When the circuit distribution is approximately 2-design, such that the first and second moments are indistinguishable from those of the Haar distribution, the energy gradient at initial random values cannot deviate from zero, except for an exponentially decaying probability in $n$ \cite{Mc,cost-dep-bp,entanglement-bp, holmes2021connecting}. It happens for the circuit ensemble in $B/C$, where the Renyi-$2$ entropy as a diagnostic of the quantum 2-design is closest to $n_A=n/2$ that corresponds to the Haar ensemble.
Until now, we assumed that the entanglement entropy of the target state follows the area law scaling, as in gapped local one-dimensional systems \cite{Hastings_2007}. However, it does not always hold, and the variational circuit in $A$ cannot minimize the circuit energy \eqref{E} to the ground level. For the Sachdev-Ye-Kitaev model \cite{SYK,2015Kitaev,Maldacena_2016}, whose ground state exhibits a volume law entanglement \cite{Huang_2019}, the optimization does fail no matter to which of $A/B/C$ the variational circuit belongs. Incidentally, we argue that higher-dimensional parameter space can assist the circuit optimization even at high level of circuit state entanglement, so that over-parameterized circuits can offer a high precision approximation of volume law entangled target states including the SYK ground state \cite{highdepth}.
The rest of this paper is organized as follows: Section~\ref{sec:ent} motivates the entanglement diagnostics as the initialization condition to arrange variational states close to the target. Section~\ref{sec:randomcircuit} studies the average entanglement growth of circuit states as a function of the circuit depth. Section~\ref{sec:opt} examines the importance of the entanglement diagnostics in the local gradient search of optimal circuit parameters. Section~\ref{sec:other} checks the validity of the entanglement diagnostics by testing them against different circuit architectures and also discuss the impact of shrinking the circuit parameter dimension. The paper concludes with discussion and outlook in Section~\ref{sec:dis}.
Additional details are given in the appendices.
\section{Entanglement Diagnostics}
\label{sec:ent}
Using the density matrix of the quantum circuit $\rho_c(\theta)$, the expectation value of the Hamiltonian (\ref{E}) reads:
\begin{equation}
E(\theta) = \text{Tr} \left(\rho_c(\theta) H\right) . \label{E1}
\end{equation}
Our optimization task is to get as close as possible to the ground state of the Hamiltonian by minimizing \eqref{E1}.
It can be achieved by multiple iterations of evaluating the density matrix $\rho_{c}(\theta)$ and updating the parameters via the gradient descent \eqref{eq:gd} that will finally stop at $\theta = \theta_f$.
We would like to reach the final parameter $\theta_{f}$ such that
\begin{equation}
\label{eq:error}
\Delta E \equiv \text{Tr} \left((\rho_{c}(\theta_{f}) - \rho_{g})H \right) \simeq 0
\end{equation}
where $\rho_{g}$ is the exact ground state of the Hamiltonian.
A simple upper bound of this approximation error $\Delta E$ follows from the Cauchy-Schwarz inequality,
\begin{equation}
\text{Tr} \left((\rho_{c}(\theta_{f}) - \rho_{g})H \right) \leq
{\norm{\rho_{c}(\theta_{f}) - \rho_{g}}}_1 \cdot {\norm{H}}_1,
\label{BH}
\end{equation}
where the trace norm $\norm{O}_1$ is the sum of singular values of an operator $O$, i.e., eigenvalues of $(O^{\dagger}O)^{1/2}$.
A natural condition for efficient reduction of $\Delta E$ is arranging an initial circuit state $\rho_{c}(\theta_{in})$ to
be in the proximity of the ground state with a small enough trace distance ${\norm{\rho_{c}(\theta_{in}) - \rho_{g}}}_1$. However, we will confront two issues.
First, we generally do not know the ground state, thus being unable to estimate the trace distance ${\norm{\rho_{c}(\theta_{in}) - \rho_{g}}}_1$.
Second, the trace distance can be very sensitive to tiny changes of quantum states. So the above condition is often over-restrictive, discarding most reasonable initial states.
Instead, we want to relax the condition by using the entanglement entropy of an initial circuit state as a distance proxy between $\rho_c(\theta_{in})$ and $\rho_{g}$, from which one can expect the success/failure of circuit optimization. It can be motivated as follows: The inequalities on the von Neumann and Renyi-$k$ entropy differences \cite{Fannes1973, raggioa1995properties,Audenaert_2007,Chen2016}:
\begin{align}
|S_{EE}(\rho_A) - S_{EE}(\sigma_A)|
\leq &
\ n\norm{\rho_A - \sigma_A}_1 + (e\ln{2})^{-1},\\
|\mathcal{R}^k(\rho_A) - \mathcal{R}^k(\sigma_A)|
\leq &
\ k\cdot 2^{n_A(k-1)}\norm{\rho_A - \sigma_A}_1,
\end{align}
show that, for given two quantum states $\rho_A$ and $\sigma_A$, being close in their quantum entropies is necessary for being close in their trace distance $\norm{\rho_A - \sigma_A}_1$. Suppose now that $\rho_A$ and $\sigma_A$ are the density matrices
of a subsystem $A$, which can be combined with a complementary part $B$ to constitute the entire $n$ qubit system, i.e., $\rho_A \equiv \text{Tr}_B \left(\rho\right)$ and $\sigma_A \equiv \text{Tr}_B \left(\sigma\right)$. Monotonicity of the trace distance under the partial trace,
\begin{equation}
{\norm{\rho_{A} - \sigma_{A}}}_1 \leq {\norm{\rho - \sigma}}_1 \ ,
\label{B1}
\end{equation}
implies in turn that, for the trace distance between two quantum states $\rho$ and $\sigma$ to be small, the difference in their entanglement entropies should necessarily be small. Hence, the entanglement diagnostics of initial circuit states can be considered as a weaker version of the proximity measure.
We usually cannot estimate the trace distance from the exact ground state $\rho_g$ due to our ignorance of $\rho_g$. However, we expect that the ground states of gapped local Hamiltonians are far from typical quantum states $\sigma$, whose $\sigma_A$ are approximately maximally mixed \cite{Goldstein_2006}. Thus, we require the trace distance between the equiprobable state and the reduced density matrix of the circuit state $\rho_A$ to be large. This requires non-maximal entanglement entropies of the circuit states, i.e. they should not scale with the subsystem size $n_A$ \cite{Page_1993}. This is encoded in the following:
\begin{theorem}
\noindent
(i) The trace distance between the reduced density matrix $\rho_A$
and the maximally mixed state ${2^{-n_A}}\,I_{n_A}$ satisfies the following inequality:
\begin{align}
\mathcal{L}(n_A, \rho_A) \leq \frac{1}{2}\left\Vert\rho_A - \frac{I_A}{2^{n_A}}\right\Vert_1 \leq
\mathcal{U}(n_A, \rho_A)
\end{align}
with
\begin{align}
\mathcal{U}(n_A, \rho_A) &= \left(\frac{n_A - S_{EE}(\rho_A)}{2}\right)^{1/2}\label{B4}\\
\mathcal{L}(n_A, \rho_A) &= \frac{1}{2k} \left( 2^{ (1-k)\mathcal{R}^k_A(\rho_A)} - 2^{(1-k)n_A}\right)
\label{eq:lower_bound}
\end{align}
where $S_{EE}(\rho_A)$ and $\mathcal{R}^q_A(\rho_A)$ are the von Neumann and Renyi-$k$ entropies of the reduced state $\rho_A$, respectively.
\\[0.2cm]\noindent
(ii) In the large size limit $n_A \gg 1$ of the subsystem $A$, the following lower bound holds asymptotically:
\begin{equation}
1 - \frac{S_{EE}\left(\rho_{A}\right)}{n_A}
\lesssim \frac{1}{2}\left\Vert{\rho_A - \frac{I_A}{2^{n_A}}}\right\Vert_1 \ .
\label{eq:lower_bound2}
\end{equation}
\end{theorem}
\begin{proof}
\noindent
(i)
We start from the Pinsker's inequality:
\begin{equation}
\tfrac{1}{2} {\norm{\rho - \sigma}}^2_1 \leq S(\rho ||\sigma)
\label{B0}
\end{equation}
on the trace distance between two states $\rho$ and $\sigma$ and their relative entropy. Plugging $\rho=\rho_A$ and $\sigma = 2^{-n_A}\,{I}_{A}$,
\begin{align}
S(\rho_A||\sigma_A) = n_A + \text{Tr}(\rho_A \log \rho_A) = n_A - S_{EE}(\rho_A)
\end{align}
such that (\ref{B0}) becomes (\ref{B4}).
The continuity bound of the Tsallis-$k$ entropy implies \cite{raggioa1995properties}:
\begin{equation}
|\text{Tr} \left(\rho_A^k\right) - \text{Tr} \left(\sigma_A^k\right)|\leq k \,\norm{\rho_A - \sigma_A}_1 \ ,
\label{ineqn}
\end{equation}
which can turn into \eqref{eq:lower_bound} by inserting $\sigma_A = 2^{-n_A}\,{I}_{A}$ and $\text{Tr} \left(\rho_A^k\right) = 2^{(1-k)\mathcal{R}^k_A(\rho_A)}$.
\\[0.2cm]\noindent
(ii)
Recall the Fannes–Audenaert inequality \cite{Audenaert_2007}:
\begin{align}
|S_{EE}(\rho_A) - S_{EE}(\sigma_A)|
\leq &
\tfrac{1}{2}\norm{\rho_A - \sigma_A}_1 \log (2^{n_A}-1) \nonumber\\
&+ H(\tfrac{1}{2}\norm{\rho_A - \sigma_A}_1) \label{eq:fannes}
\end{align}
where $H(t) \equiv -t \log{t} - (1-t)\log(1-t)$. Substituting $\sigma_A = 2^{-n_A}\,{I}_{A}$ and taking the large system size limit $n_A\gg 1$, the LHS of \eqref{eq:fannes} becomes $(n_A - S_{EE}(\rho_A))$, which leads to the asymptotic inequality \eqref{eq:lower_bound2}.
\end{proof}
We stress that the entanglement diagnostic for circuit states is only a necessary condition to keep the initial and target states close. Remarkably, as we will see in Section~\ref{sec:opt}, the gradient-based optimization indeed works efficiently for those variational circuits whose average entanglement entropy scales slower than the volume law.
Concerning the circuit depth, this suggests to avoid intermediate-depth and high-depth circuits, respectively corresponding to $B/C$ of Figure~\ref{fig:random_growth}, and favor the circuits with fewer layers that belong to $A$.
We will estimate the critical depth $L_s$ that divides $A$ and $B/C$ in the following Section~\ref{sec:randomcircuit}.
\section{Random Quantum Circuit}
\label{sec:randomcircuit}
In this section, we study the growth of entanglement entropy for the circuit states generated by a random circuit evolution of the initial product state $|0\rangle^{\otimes n}$.
Figure~\ref{fig:circuit_diagram} is the quantum circuit architecture used in this paper. It defines a $(1+1)$-dimensional discrete quantum system, where the $n$ qubits along the vertical axis represent the space, and the $L$ layers along the horizontal axis span the time. The qubits are arranged identically with period $n$, i.e., $i\simeq i+n$, imposing a periodic boundary condition along the spatial direction. At each time step, the wavefunction evolves by a chain of the 2-qubit unitary gates, acting alternatingly on all neighboring odd-even/even-odd qubit pairs. The 2-qubit gate is made of independent Pauli-$y$ rotations acting on single qubits,
\begin{equation}
R(\varphi) = \exp(i \sigma_y \varphi) = \begin{pmatrix}
\cos\varphi & \sin\varphi \\
-\sin\varphi & \cos\varphi
\end{pmatrix} \ ,
\label{eq:paulirot}
\end{equation}
followed by the controlled-Z operation
\begin{align}
CZ = \text{diag}\left(1, 1, 1, -1\right)
\label{eq:cz}
\end{align}
that generically creates a pairwise entanglement. We will collectively denote all rotation angles by $\theta$ while using $\theta_{l,i}$ to indicate a specific angle that rotates the $i$'th qubit at the $l$'th layer, where $1\leq i \leq n$ and $1\leq \ell\leq L$. These variables are randomly chosen from $\mathcal{U}(0,2\pi)$, the uniform distribution between $0$ to $2\pi$.
\begin{figure}[t]
\subfloat[Von Neumann Entropy]{
\includegraphics[height=4.1cm]{./fig3a.pdf}\label{fig:vn}}
\subfloat[Renyi-2 Entropy]{
\includegraphics[height=4.1cm]{./fig3b.pdf}\label{fig:r2}}
\caption{Von Neumann and Renyi-2 entropies for $8 \leq n \leq 20$ averaged over 50 independent circuit states, as a function of the circuit depth $L$. }
\label{fig:entropy-growth-generic}
\end{figure}
\subsection{Linearity of the
Initial Entanglement Growth}
Let us consider the evolution
of the $n$-qubit state under the random circuit unitaries of Figure~\ref{fig:circuit_diagram}, as a function of the number of layers $L$.
We measure the average growth of the bipartite entanglement of random circuit states by decomposing the $n$ qubits into two equal-size subsystems, $n_A = n_B= n/2$, and calculating the sample statistics of various Renyi entropies for different $n$ and $L$.
Figures~\ref{fig:vn}~and~\ref{fig:r2} show the von Neumann and Renyi-2 entropies averaged over $50$ random circuit states
with different numbers of qubits $n$. Figure~\ref{fig:entropies-generic} compares the Renyi entropies of different orders averaged over $50$ random circuit states with $n=12$ and $20$ qubits.
They all exhibit the linear growth of the entanglement entropies at initial times. The curves then slow down in growth and eventually reach the plateaus.
See Figures~\ref{fig:pmax}--\ref{fig:geometry-generic} in the appendix for the growth curves of several other entanglement quantities with different system sizes $n$.
The early linear growth of the entanglement entropy,
\begin{equation}
{\cal R}^k_{A} (\rho_A) = v_{k} L \ ,
\label{eq:area-law}
\end{equation}
is a characteristic feature of the global quench dynamics \cite{Calabrese_2005}, which in our case is driven by the successive application of the layer unitaries $U(\theta_L)$ to the $n$-qubit product state $|0\rangle^{\otimes n}$. The coefficient $v_k$ is known as the entanglement velocity and generally depends on $k$.
We determine $v_k$ by the linear regression of the early-time entropies on the range of $0 \leq L \leq n/2$. The estimated values of $v_k$, computed at different $n$'s and $k$'s, are summarized in the third columns of Tables~\ref{tbl:random}~and~\ref{tbl:random-extra}.
We find that
$v_k$ is independent of $n$ except for minor fluctuations, identifying \eqref{eq:area-law} with the area-law entanglement of the early-time circuit states.
Furthermore, $v_k$ decreases when the order $k$ of the Renyi entropy increases, i.e., $v_{k_1} > v_{k_2}$ for $k_1 < k_2$.
On the other end, at a late time, the Renyi-$k$ entropy saturates to a constant $r_{n,k}$ for any $n$ and $k$. We compute the saturated value of $\mathcal{R}^{k}_A$ by averaging it over the time frame $200 \leq L \leq 250$ and record that in the fifth columns of Tables~\ref{tbl:random}~and~\ref{tbl:random-extra}. The resulting constants $r_{n,k}$ manifest the following simple dependency on $n_A = n/2$:
\begin{equation}
{\cal R}^k_{A} (\rho_A) = r_{n,k} \simeq n/2 - c_k = \text{Vol}(A) - c_k \
\label{eq:volume-law}
\end{equation}
implying the volume-law entanglement of the late-time circuit states \cite{Liu_2018}. We also find that, as the entropy order $k$ increases, the saturated value $r_{n,k}$ declines monotonically, so the shift constant $c_k>0$ can be only larger.
\begin{figure}[t]
\subfloat[$n=12$]{
\includegraphics[height=4.1cm]{./fig4a.pdf}\label{fig:n12}}
\subfloat[$n=20$]{
\includegraphics[height=4.1cm]{./fig4b.pdf}\label{fig:n20}}
\caption{Various Renyi-$k$ entropies for $n=12$ and $n=20$ averaged over 50 independent circuit states, as a function of the circuit depth $L$.}
\label{fig:entropies-generic}
\end{figure}
Combined with the discussion in Section~\ref{sec:ent}, the average entanglement curves suggest to refrain from using a variational circuit in the region of the plateau, i.e., $L \geq L_s$, in order to prepare an initial circuit state in proximity to the target ground state that follows the area-law entanglement. We now turn to examine the scaling behavior of the early-time and late-time scales, i.e., $L_l$ and $L_s$.
\begin{table}[b]
\centering
\begin{tabular}{@{}ccccccccccccc@{}}
\toprule
Type & $n$ & $v_k$ & $L_l$ & $r_{n,k}$ & $L_s$ & Type & $n$ & $v_k$ & $L_l$ & $r_{n,k}$ & $L_s$\\ \midrule
\multirow{7}{*}{$S_{EE}$}
& $8$ & 0.3669 & 7 & 3.2526 & 29 & \multirow{7}{*}{$\mathcal{R}^2_A$} & $8$ & 0.2771 & 9 & 2.9722 & 29\\
& $10$ & 0.3480 & 6 & 4.2639 & 33 && $10$ & 0.2581 & 12 & 3.9821 & 36\\
& $12$ & 0.3533 & 13 & 5.2703 & 43 && $12$ & 0.2645 & 17 & 4.9896 & 43\\
& $14$ & 0.3551 & 9 & 6.2743 & 49 && $14$ & 0.2711 & 15 & 5.9944 & 49\\
& $16$ & 0.3522 & 17 & 7.2766 & 58 && $16$ & 0.2689 & 23 & 6.9973 & 60\\
& $18$ & 0.3459 & 12 & 8.2776 & 65 && $18$ & 0.2682 & 24 & 7.9986 & 65\\
& $20$ & 0.3411 & 23 & 9.2781 & 71 && $20$ & 0.2619 & 33 & 8.9993 & 71\\\bottomrule
\end{tabular}\qquad
\caption{
Phenomenological estimation of the entanglement velocity, saturation value, early-time scale, and late-time scale.}
\label{tbl:random}
\end{table}
\subsection{Timescale for the Entanglement Growth}
Let us study the early-time $L_l$ and late-time scales $L_s$, respectively, as the depth scales where the linear growth \eqref{eq:area-law} ends and where the saturation \eqref{eq:volume-law} begins. We measure $L_l$ and $L_s$ using the following operational definitions:
\begin{align}
\textstyle L_l &= \max\left\{L : |\mathcal{R}^k_A(L) - v_k L| \leq 2\, \text{RMS}_{k}({0, n/2})\right\}
\label{pheno}\\
\textstyle L_s &= \min\left\{L : |\mathcal{R}^k_A(L) - r_{n,k}| \leq 2\,\text{RMS}_{k}({200, 250})\right\},\nonumber
\end{align}
$L_l$ is the maximum depth $L$ where the gap $|\mathcal{R}^k_A(L) - v_k L|$ between the Renyi entropy and its linear approximation maintains smaller than two times the RMS deviation,
\begin{align}
\textstyle\text{RMS}_{k}({a, b}) = \sqrt{\frac{1}{b-a+1} \sum_{\ell=a}^{b}\ (\mathcal{R}^k_A(\ell) - v_k \ell)^2},
\label{eq:RMS}
\end{align}
for $0 \leq L \leq n/2$. Similarly, $L_s$ is the minimum depth $L$ whose difference $|\mathcal{R}^k_A(L) - r_{n,k}|$ between the Renyi entropy and its saturated value remains to be smaller than two times the RMS deviation \eqref{eq:RMS} for $200 \leq L \leq 250$.
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth]{./fig5.pdf}
\caption{
Measurements averaged over 50 independent circuits, before/after the VQA optimization with the nearest-neighbor Ising Hamiltonian \eqref{Ising1} at $g=2$, as a function of the number of circuit layers $L$.}
\label{fig:IsingQ12Optimb}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth]{./fig6.pdf}
\caption{Measurements averaged over 50 independent circuits, before/after the VQA optimization with the nearest-neighbor Ising Hamiltonian \eqref{Ising1} at $g=1$, as a function of the number of circuit layers $L$.}
\label{fig:IsingQ12G1Optim}
\end{figure*}
The estimated values of $L_l$ and $L_s$ for different values of $n$ and $k$ are summarized in the fourth and sixth columns of Tables~\ref{tbl:random}~and~\ref{tbl:random-extra}. We make three observations: First, both timescales $L_l$ and $L_s$ increase as the entropy order $k$ goes higher. Second, the saturation time $L_s$ scales linearly in the system size $n$, i.e., $L_s \sim \mathcal{O}(n)$, because
\begin{equation}
L_s \gtrsim \frac{\text{Vol}(A) - c_k}{v_k \cdot \text{Area}(\partial A)} \sim \frac{n_A}{v_k} \sim \mathcal{O}(n).
\end{equation}
This is consistent with \cite{Liu_2018} that a unitary design that maximizes all Renyi entropies can be reached within a linear complexity in the system size $n$. Third, there exists a transient gap between $L_l$ and $L_s$, at least for finite-sized systems, in which the entanglement growth is slower. Details of the entanglement curves in this crossover region are largely model-dependent. See \cite{Liu_2014} for an example.
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth]{./fig7.pdf}
\caption{Measurements averaged over 50 independent circuits, before/after the VQA optimization with the non-local Ising Hamiltonian \eqref{Ising2} at $\alpha=g=1$, as a function of the number of circuit layers $L$.}
\label{fig:IsingQ12Optim}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=0.85\textwidth]{./fig8.pdf}
\caption{Measurements averaged over 30 independent circuits, before/after the VQA optimization with a particular instance of the SYK${}_{4}$ Hamiltonian \eqref{eq:syk_ham}, as a function of the number of circuit layers $L$.}
\label{fig:SYKQ12Optim}
\end{figure*}
\section{Circuit Optimization}
\label{sec:opt}
Our focus in this section is on the classical component of the hybrid quantum/classical algorithm. The objective is to find circuit parameters $\theta^*$ that closely approximate the ground state energy, $E(\theta^*) \simeq E_g$, by taking iterative steps proportional to the negative gradient of the energy function \eqref{E1} at each point $\theta_\tau$, i.e.,
\footnote{Estimating the gradient requires the readout of the circuit state $\rho_{c}$ at shifted gate parameters \cite{parameter_shift1} conducted by repeated measurements of Pauli strings. We will not consider the effect of the readout noise in this paper.}
\begin{align}
\theta_{\tau + 1} = \theta_{\tau} - \eta \, \nabla E(\theta_{\tau}).
\label{eq:gd}
\end{align}
The learning rate $\eta$
scales the step size
of each update.
A too-large $\eta$
can cause overshooting near the minimum $\theta^*$,
while a too-small $\eta$
can make the optimization trajectory
stuck at local minima.
We will use
$\eta=0.005$ for most experiments.
When the parameters update is small, each step of the gradient descent
can reduce the energy by
\begin{align}
\Delta E(\theta_{\tau}) \equiv E(\theta_{\tau+1}) - E(\theta_{\tau}) = -\eta\, \norm{\nabla E(\theta_\tau)}_2^2 \ .
\label{eq:energy-dec}
\end{align}
Due to the constant decrease of the energy \eqref{eq:energy-dec}, we expect to reach $E_g$ eventually if there are no other obstacles. We will terminate the iteration after updating the circuit parameter $10^4$ times in all our numerical experiments.
\subsection{Results}
\label{sec:res}
Let us discuss the eigensolver optimization results that aim to solve the ground state of many-body systems. We specifically consider the 1d transverse-field Ising models with nearest-neighbor and long-range interactions and the Sachdev-Ye-Kitaev (SYK) models. See Appendix~\ref{sec:ham} for a brief review of their Hamiltonians and ground-state entanglement properties.
\subsubsection{The Transverse-Field Ising Models}
\label{sec:isings}
We search the ground states of interacting 1d spin-chain systems. To break the degeneracy of ground states, we turn on the magnetic coupling $g$ to all the spin variables, choosing it to be $g=1$ or $2$.
As we are interested in finding a general correlation between the entanglement diagnostics and the success of optimization, not relying on specific characteristics of Hamiltonians, we study the optimization for the following three Ising models:
\begin{enumerate}[label=(\roman*),itemsep=-1ex,partopsep=1ex,parsep=1ex]
\item the nearest-neighbor spin coupling \eqref{Ising1} with $g=2$
\item the nearest-neighbor spin coupling \eqref{Ising1} with $g=1$
\item the long-range spin coupling \eqref{Ising2} with $\alpha=g=1$.
\end{enumerate}
We repeatedly perform the circuit optimization 50 times, to remove fluctuation made by random parameter initialization, and record the circuit outputs in Figures~\ref{fig:IsingQ12Optimb}-\ref{fig:IsingQ12Optim} as a function of the circuit depth $L$.
Each figure consists of three panels. The left ones represent the energy difference $E(\theta) - E_g$ between the circuit state \eqref{E1} and the exact ground state \eqref{eq:ising_eg}. The middle ones show the trace distance between the reduced circuit state $\rho_{c,A}$ and the reduced ground state $\rho_{g,A}$. The right ones display the Renyi-2 entropy of the reduced circuit state. All the orange/blue curves therein represent a corresponding quantity before/after the optimization.
\begin{figure*}[t]
\centering
\includegraphics[width=0.90\textwidth]{./fig9.pdf}
\caption{The evolution of the Renyi-$2$ entropy by the gradient-based VQA optimization with the Ising and SYK${}_4$ Hamiltonians, i.e., \eqref{Ising1}, \eqref{Ising2}, \eqref{eq:syk_ham}, for $n=12$ qubits. The horizontal direction denotes the number of parameter updates $\tau$. The dashed line shows the Renyi-$2$ entropy of the exact ground states for each Hamiltonian.}
\label{fig:training-curve}
\end{figure*}
Figure~\ref{fig:IsingQ12Optimb} is for the nearest-neighbor Ising model \eqref{Ising1} with $g=2$. It reveals the relation between the average entanglement entropy of initial circuit states and the success of gradient-based optimization: The optimization works well for the circuits with $L \lesssim 36$. However, in the intermediate range of $36 \lesssim L \lesssim 52$, the success rate gradually lowers as the circuit becomes deeper. Beyond that, i.e., $L \gtrsim 52$, it always fails to close the gap between the exact ground state and the circuit state as to their energy and entanglement entropies. Such relation shows an advantage of using the circuits with $L < L_s$, whose entanglement curve has not reached the plateau. The above pattern
also persists in Figures~\ref{fig:IsingQ12G1Optim}~and~\ref{fig:IsingQ12Optim}, which correspond to the nearest-neighbor and long-range Ising models with $g=1$.
A notable difference of Figure~\ref{fig:IsingQ12G1Optim} from Figures~\ref{fig:IsingQ12Optimb} or~\ref{fig:IsingQ12Optim} appears in the trace distance curve, where the optimization fails to narrow the distance even when the circuit energy is close to the exact ground state energy. It is related to the fact that the ground state entanglement entropy in the $g=1$ nearest-neighbor Ising model is higher than those in other Ising models, as shown in Figure~\ref{re}. When the entanglement entropy of target ground states is higher, the local search of approximating circuit parameters becomes increasingly difficult. Such difficulty leads to deviations between the post-optimization circuit state and the exact ground state, to which the trace distance reflects much more sensitively than the energy and entanglement entropy differences.
\subsubsection{The Sachdev-Ye-Kitaev Model}
\label{sec:sykoptim}
We will now discuss the circuit optimization in a situation where the Hamiltonian ground state exhibits
a volume law scaling of the entanglement entropy.
The SYK${}_4$ Hamiltonian \eqref{eq:syk_ham} defined with an instance of random coupling constants has a ground state that follows the volume law scaling of entanglement \cite{Huang_2019}, as reviewed in Appendix~\ref{sec:ham} and specifically in Figure~\ref{re}. We optimize the variational circuit to approximate the SYK${}_4$ ground state and summarize the output in Figure~\ref{fig:SYKQ12Optim} as a function of the circuit depth $L$.
Since the approximation target state itself behaves in terms of entanglement like a generic quantum state, the optimization task is now much more challenging. Unlike the optimization towards the Ising ground state, even choosing a less entangled circuit within the range $L \lesssim 36$ does not lead to success. Figure~\ref{fig:SYKQ12Optim} manifests this failure, not only in the trace distance between the circuit and exact ground states but also their differences of energy and entanglement entropy.
\subsubsection{Optimization Speed}
As another indicator of how difficult the circuit optimization is, we draw in Figure~\ref{fig:training-curve} the evolution of Renyi-2 entanglement entropy $\mathcal{R}^2_A(\rho_{c,A})$ as a function of the number of parameter updates $\tau$. The three curves therein are for the circuits with $L=12$, $40$, $68$, which represent the characteristics of low-, intermediate-, high-depth circuits. The entanglement entropy of the ground states is marked by the dashed lines.
Towards the Ising ground states, the gradient descent works efficiently for the $L=12$ circuit, rapidly reducing the entanglement entropy within $10^3$ steps of the update. However, the same gradient descent takes a much longer time for the $L=40$ circuit and even fails to reach closely the target state for the $L=68$ circuit. As the average entanglement entropy of initial circuit states increases, the optimization difficulty becomes more evident not only in the trace distance, as in Figure~\ref{fig:IsingQ12G1Optim}, but also through the entanglement diagnostics.
The optimization task towards the SYK${}_4$ ground state is inherently more challenging such that all three curves leave a large entanglement gap from the target state. Interestingly, the gradient descent constantly reduces the entanglement entropy of the $L=12$ circuit state, enlarging the gap over the optimization steps $\tau$.
In general, over-parameterization can assist the circuit optimization that starts from/ends at a highly-entangled typical quantum state. An exponentially high-dimensional parameter space
was needed for the SYK${}_4$ model to approximate its ground-level energy with very high precision \cite{highdepth}.
\subsection{Entanglement Diagnostics and Optimization}
Our results shown in Section~\ref{sec:res} exemplify the difficulty in finding a successful optimization trajectory that starts from or ends at a typical quantum state that takes up the vast majority of the Hilbert space. This has been best described through the evolution of the entanglement entropies, \eqref{Re} and \eqref{SEE}, over the optimization steps, rather than a more commonly-used sensitive metric such as the trace distance between the circuit and target states.
Suppose we can divide the Hilbert space into two subregions distinguished by their entanglement entropies, say $A$ and $B/C$, in accordance with Figure~\ref{fig:random_growth}. Generic random states belong to the region $B/C$ whose entanglement entropies are approximately maximal.
For many interesting applications, the target state is a non-generic state that resides in the region $A$, i.e., following the area-law scaling of the entanglement \cite{Lucas_2014}. Along an optimization path inside the region $A$, the circuit state entanglement tends to decrease regularly. However, when an initial state $\rho_{c}(\theta_\text{in})$ belongs to the region $B/C$, the local parameter update \eqref{eq:gd} is unable to cross over to the region $A$, thus failing to reach the ground state energy. We make these observations from the optimization result in Section~\ref{sec:isings} that discusses the Ising Hamiltonians.
\begin{figure}[tb]
\centering
\subfloat[Von Neumann Entropy]{
\includegraphics[height=4.1cm]{./fig10a.pdf}}
\subfloat[Renyi-2 Entropy]{
\includegraphics[height=4.1cm]{./fig10b.pdf}}
\caption{Von Neumann and Renyi-2 entropies for $8 \leq n \leq 20$ averaged over 50 independent $p=1/2$ circuit states, as a function of the circuit depth $L$.}
\label{fig:entropy-growth-stochastic}
\end{figure}
Even when the target state is maximally entangled and lies in the region $B/C$, the entanglement entropy of the circuit state $\rho_{c}(\theta_{\tau})$ still tends to decrease on average. It means that the entanglement gap between the circuit and target state can become larger, if an initial circuit state has a smaller entanglement entropy than the target state, i.e., $\mathcal{R}^{k}_A(\rho_{c, A}) \lesssim \mathcal{R}^{k}_A(\rho_{g, A})$.
When $\mathcal{R}^{k}_A(\rho_{c, A}) > \mathcal{R}^{k}_A(\rho_{g, A})$, the optimization moves towards narrowing the gap, but often failing to match a desired level of the entanglement. These observations are based on the optimization results in Section~\ref{sec:sykoptim}, obtained for the 1d SYK Hamiltonian.
The numerical results suggest that the Hilbert space can be partitioned into multiple layers, distinguished by the supported amount of the bipartite entanglement entropy. It is a very demanding task to move across distant layers via the gradient descent \eqref{eq:gd}, which is doable only for the over-parameterized circuits that involve exponentially large parameter space \cite{highdepth}.
\begin{figure}[tb]
\centering
\subfloat[$n=12$]{
\includegraphics[height=4.1cm]{./fig11a.pdf}}
\subfloat[$n=20$]{
\includegraphics[height=4.1cm]{./fig11b.pdf}}
\caption{Various Renyi-$k$ entropies for $n=12$ and $n=20$ averaged over 50 independent $p=1/2$ circuit states, as a function of the circuit depth $L$.}
\label{fig:entropies-generic1}
\end{figure}
\section{Other Circuit Architectures}
\label{sec:other}
We discussed the importance of choosing the circuit to avoid the saturation of its average initial entanglement entropy, for a generic optimization task that finds a target state with the area law entanglement.
This section examines if the entanglement diagnostic still serves as an indicator of efficient circuit optimization with different circuit architectures. We also consider the effect of reducing the number of circuit parameters while maintaining a similar degree of entanglement.
\begin{figure*}[!t]
\centering
\includegraphics[width=0.85\textwidth]{./fig12.pdf}
\caption{Measurements averaged over 50 independent $p=1/2$ circuits, before/after the VQA optimization with the nearest-neighbor Ising Hamiltonian \eqref{Ising1} at $g=1$, as a function of the number of circuit layers $L$.}
\label{fig:IsingQ12Optim-stochastic}
\end{figure*}
\begin{figure*}[t!]
\centering
\includegraphics[width=0.85\textwidth]{./fig13.pdf}
\caption{Measurements averaged over 50 independent restricted circuits \eqref{sub}, before/after the VQA optimization with the nearest-neighbor Ising Hamiltonian \eqref{Ising1} at $g=1$, as a function of the number of circuit layers $L$.}
\label{fig:IsingQ12Optim-layerequal}
\end{figure*}
\subsection{Random Graph Architecture}
\label{sec:random}
Let us study a simple stochastic variation of the circuit architecture that omits the $CZ$ entangler \eqref{eq:cz} inside the 2-qubit gate of Figure~\ref{fig:circuit_diagram_right} with a fixed probability $p=1/2$.
\subsubsection{Entanglement Growth}
Since the average number of the entangler is cut in half, we expect that the entanglement growth rate would be halved. Accordingly, the circuit depth $L$ to reach the saturation of the entanglement entropy would be doubled.
Figures~\ref{fig:entropy-growth-stochastic}~and~\ref{fig:entropies-generic1} show the evolution of entanglement diagnostics as a function of the circuit depth $L$, estimated by the sample averages of 50 random states.
The overall shape of the curves remains the same, but the growth rate has significantly decreased. Reaching a certain level of the entanglement diagnostics requires twice the circuit depth compared to the non-stochastic architecture, i.e., $p=1$, as expected.
See Figure~\ref{fig:geometry-prob} for the curve of the geometric measure whose growth rate has been halved.
\subsubsection{Optimization}
Given the optimization task that reaches the nearest-neighbor Ising ground energy \eqref{eq:ising_eg} with the background field coupling $g=1$, the outputs of the $p=1/2$ stochastic circuit are all collected in Figure~\ref{fig:IsingQ12Optim-stochastic} as a function of $L$.
The depth range of the $p=1/2$ circuits where the gradient descent remains successful has increased to $L \lesssim 96$. Beyond that, the optimization success rate continues to drop until it reaches $0\%$ at $L \sim 144$ and above. This is consistent with the entanglement growth curves, which maintain the same overall shape as in Section~\ref{sec:randomcircuit} but only with a lower growth rate. We remark that the low and intermediate ranges, in which the optimization may succeed with a non-zero probability, has been extended to $L \lesssim 136$, more than mere doubling. It is the impact of the expanded parameter space whose dimension has been doubled, as required for the $p=1/2$ circuit to hold the same level of entanglement.
Over the entire range of $L$, unlike the trace distance, the entropy diagnostic holds a robust correlation with the successful minimization of the circuit energy \eqref{E1}, showing its usefulness regardless of circuit-specific details.
\subsection{Restricted Circuit Parametrization}
\label{sec:param}
Recall that an additional circuit layer increases both the average entanglement entropy of initial circuit states and the number of classical parameters. To isolate the effect of the classical parameter space, we study the consequence of imposing the following restriction:
\begin{equation}
\theta_{\ell,1} = \theta_{\ell, 2} = \cdots = \theta_{\ell, n} \quad \text{ for all } 1\leq \ell \leq L \ ,
\label{sub}
\end{equation}
which equates all the parameters in each layer,
yet maintains the same growth rate of entanglement diagnostics.
\begin{figure}[!t]
\centering
\subfloat[Von Neumann Entropy]{
\includegraphics[height=4.1cm]{./fig14a.pdf}}
\subfloat[Renyi-2 Entropy]{
\includegraphics[height=4.1cm]{./fig14b.pdf}}
\caption{Von Neumann and Renyi-2 entropies for $8 \leq n \leq 20$ averaged over 50 independent restricted circuit states \eqref{sub}, as a function of the circuit depth $L$.}
\label{fig:entropy-growth-layerequal}
\end{figure}
The basic 2-qubit gate ${\cal O}_{i,j}$
in Figure~\ref{fig:circuit_diagram_right} reads:
\begin{equation}
{\cal O}_{i,j} = CZ_{i,j} \cdot R(\theta_{l,i})\otimes R(\theta_{l,j}) \ ,
\end{equation}
where
$CZ=\text{diag}(1,1,1-1)$ and $R(\theta)$ is the Pauli rotation \eqref{eq:paulirot} around the $y$-axis.
It is curious to note that the constraint
(\ref{sub}) is equivalent
to imposing $[{\cal O}_{i,j},Q_{i,j}] = 0$
on the Hilbert space of $(i,j)$ qubits, where:
\begin{equation}
Q_{i,j} =
\begin{pmatrix}
q_1+q_2 & 0 & 0 & 0\\
0 & q_1 & q_2 & 0 \\
0 & q_2 & q_1 & 0 \\
0 & 0 & 0 & q_1+q_2 \\
\end{pmatrix}
\end{equation}
in the computational basis of $(i,j)$ qubits.
Still, there is no globally conserved charge written as a tensor product sum of $Q_{i,i+1}$,
because it does not generically commute with ${\cal O}_{i-1,i}$ and ${\cal O}_{i,i+1}$ on the next layer.
\subsubsection{Entanglement Growth}
The entanglement entropies averaged over 50 random circuit states under the parameter space restriction \eqref{sub} are illustrated in Figures~\ref{fig:entropy-growth-layerequal}~and~\ref{fig:entropies-layerequal} as a function of the number of circuit layers $L$. Except small extra wiggles, the overall growth shape and speed of the entanglement diagnostics are similar to those of the unconstrained circuit.
Such correspondence of the entanglement growth curves renders the restricted circuit an appropriate setup to study separately the effect of the parameter space dimension on the circuit optimization.
As a side remark, we have seen that the evolution curve of the geometric measure, illustrated in Figure~\ref{fig:geometry-rest}, for the restricted circuit is far more fluctuating than as for the unconstrained circuit, while their saturation depth scales remain largely the same.
\begin{figure}[t]
\centering
\subfloat[$n=12$]{
\includegraphics[height=4.1cm]{./fig15a.pdf}}
\subfloat[$n=20$]{
\includegraphics[height=4.1cm]{./fig15b.pdf}}
\caption{Various Renyi-$k$ entropies for $n=12$ and $n=20$ averaged over 50 independent restricted circuit states \eqref{sub}, as a function of the circuit depth $L$.}
\label{fig:entropies-layerequal}
\end{figure}
\subsubsection{Optimization}
We optimize the restricted circuit to approximate the ground state of the nearest-neighbor Ising model at $g=1$ using the gradient descent. The results are summarized in Figure~\ref{fig:IsingQ12Optim-layerequal} as a function of $L$. It is notable that even the circuit optimization with $L \lesssim 20$ often stops at $\Delta E \gtrsim 1$, not giving a reliable approximation of the ground energy. Furthermore, starting from $L\gtrsim 24$, an increasingly large proportion of the randomly initialized circuits fails to reach the ground level energy $E_g$ and leave a large gap, i.e., $\Delta E \gtrsim 8$. Such transitional result emerges at a much lower depth than $L=44$ of the unconstrained circuit.
It emphasizes the importance of having enough parameters in applying the gradient descent to optimization tasks, even for those circuits that remain within a suitable range of the entanglement diagnostics.
\section{Discussion and Outlook}
\label{sec:dis}
In this paper we considered the variational circuit model of quantum computation, arguing for the effectiveness of entanglement diagnostics in finding the circuit architecture for efficient parameter optimization that minimizes the Hamiltonian expectation value. Introduced as a distance measure between the circuit and target states, the entanglement diagnostic has shown its usefulness by illustrating that quantum circuits states within a suitable range of entanglement entropies can successfully reach the ground level energy of local Hamiltonians. It also says that, while entanglement is a valuable non-local resource for quantum computation, circuit states being highly entangled do not necessarily have an advantage but it can be rather the opposite.\footnote{See also the discussion in \cite{GM}.}
One way to control the average entanglement entropy of circuit states is to adjust the number of circuit layers. The mean entropy grows linearly with the circuit depth, then gradually slows down, and finally converges to a constant near the theoretical maximum. Denoting by $L_s$ the saturation depth beyond which the average entanglement entropy has converged, we divided the depth range into two intervals, $L < L_s$ and $L \geq L_s$, and called them respectively $A$ and $B/C$. $A$ is typically the optimal region that leads to efficient VQA computation, e.g., when we search an area-law entangled target state, while $B/C$ can suffer from the barren plateau problem,
One can further differentiate $C$ from $B$ based on whether the optimization success rate has become $0\%$ or not yet.
Although the assumption of an area-law entangled target state covers most of the interesting VQA applications \cite{Lucas_2014}, the ground states of some important Hamiltonians exhibit volume-law entanglement scaling. Matching the entanglement diagnostic alone is not sufficient to approximate such states due to the overwhelming population of highly-entangled quantum states. We need deep variational circuits whose depth $L$ lies in $B/C$ and that are equipped with a large parameter space that can assist high-resolution specification and approximation of the desired target state \cite{highdepth}. Having more circuit parameters can generally help to approximate the ground state better, as exemplified by the decreased accuracy for a reduced number of independent variables\footnote{It is not a conflict with \cite{Funcke_2021} which reduces parameter space redundancy by identifying the principal directions carefully for maximal expressibility. Our reduction is rather arbitrary and without a guarantee that the circuit with remaining independent parameters can be maximally expressible.} in Section~\ref{sec:param}, as well as the increased success rate for circuits with extra single-qubit rotation parameters \cite{highdepth, circuit_hessian}.
There are many follow-up directions for further investigation: First, for having additional substantial evidence to the validity of the entanglement diagnostics, it would be crucial to consider 2d gapped local Hamiltonians whose ground states follow the area law entanglement scaling but are difficult to approximate. Second, we would like to explore various circuit architectures, e.g., using other rotation and entangling gates \cite{haug2021capacity}, built on different graph structures \cite{Harrow:2019lyw,Cervera_Lierta_2019}, or conserving diffusive charges \cite{sub,diffusive}.
Especially, symmetry-preserving circuits can work efficiently for the VQA optimization if the target state is known to respect the imposed built-in symmetries \cite{ising_circuit,ising_circuit2}. Third, the layered circuit defines a discrete dynamical system. We would like to investigate the appearance of quantum chaos in the circuit wavefunction, such as the emergence of random matrix ensemble for the reduced density matrix \cite{Chen_2018} and the operator spreading \cite{Mezei:2016wfz}, relating them to the efficient VQA optimization \cite{KO2}. Finally, it is important to study different types of noise and analyse how they
affect the VQA optimization performance \cite{barron2020measurement,fontana2020optimizing}.
\section*{Acknowledgements}
We would like to thank Khen Cohen, Tom Levy, Eun Gook Moon, Muli Safra and Lior Wolf
for valuable discussions.
We are grateful to Jaedeok Kim and Dario Rosa for the collaboration at the early stage of this project. The work of J.K. is supported by the NSF grant PHY-1911298 and the Sivian fund.
The work of Y.O. is supported in part by Israel Science Foundation Center
of Excellence,
the IBM Einstein Fellowship and John and Maureen Hendricks Charitable Foundation
at the Institute for Advanced Study in Princeton.
Our Python code for the numerical experiments is written in TensorFlow Quantum \cite{TFQ}. The experimental data are managed by using Weights \& Biases \cite{wandb}.
\begin{figure*}[p]
\includegraphics[height=5.4cm]{./fig16.pdf}
\caption{The ground energy of the following Hamiltonian systems of different sizes $n$: (Left) Nearest-neighbor Ising Hamiltonian with different $g$. (Middle) Long-range Ising Hamiltonian at $g=1$ with different $\alpha$. (Right) SYK${}_4$ Hamiltonian with 100 different instances of Gaussian random couplings.
}
\label{gse}
\end{figure*}
\subsection{Complete Graph Conserved Charges}
Consider first the case $n=2$ qubits with $L$ layers.
The network operator reads:
\begin{equation}
O_{net} = \prod_{i=1}^L Z R(\theta_{2i})\otimes R(\theta_{1i}) \ ,
\end{equation}
where
$Z=I\oplus P$ with $P$ the reflection
\begin{equation}
P = \begin{pmatrix}
1 & 0 \\
\\
0 & -1
\end{pmatrix} \ ,
\end{equation}
and $R(\theta)$ is a rotation around the $y$-axis
\begin{equation}
R(\theta) = \begin{pmatrix}
Cos[\frac{\theta}{2}] & -Sin[\frac{\theta}{2}] \\
\\
Sin[\frac{\theta}{2}] & Cos[\frac{\theta}{2}]
\end{pmatrix} \ .
\end{equation}
We look for integrable directions in the $2L$-dimensional
parameter space $(\theta_{11},\theta_{21},...,\theta_{1L},\theta_{2L})$ by searching for conserved charges.
For a classical system with $k$ degrees of freedom to be integrable one needs $k$ conserved charges in involution.
The charges should satisfy:
\begin{equation}
\left[Q,O_{net}\right] = 0 \ .
\label{commute}
\end{equation}
One can solve (\ref{commute}) on the $n$-dimensional submanifold
$\theta_{2i}=0, i=1,...,L$ where:
\begin{equation}
R(\theta_{1i})\otimes R(\theta_{2i}) = R(\theta_{1i})\otimes I
,~~~~ i=1,...,L \ .
\end{equation}
and find a one parameter family of conserved charges:
\begin{equation}
Q = \hat{q}\oplus\hat{q},~~~~
\hat{q} =
\begin{pmatrix}
1 & 0 \\
0 & -1 \\
\end{pmatrix}
\ .
\end{equation}
\begin{itemize}
\item This calculation raises a question. If we started by
not acting on the second qubit with the rotation
gate around the $y$-axis, will we get a better convergense ?
It seems that this subspace is not good and there is no entanglement.
\end{itemize}
Another solution is on the subspace
$\theta_{1i} = \theta_{2i} = c_i, i=1,...,L$.
On this subspace one finds a two-parameter family of symmetry charges:
\begin{equation}
Q =
\begin{pmatrix}
q_1+q_2 & 0 & 0 & 0\\
0 & q_1 & q_2 & 0 \\
0 & q_2 & q_1 & 0 \\
0 & 0 & 0 & q_1+q_2 \\
\end{pmatrix}
\ .
\end{equation}
\begin{itemize}
\item Is this the most general solution ?
\end{itemize}
\begin{itemize}
\item How do we know that the subspace $U(\theta)|0>$
is close to the ground state of the Hamiltonian ?
\item In order to characterize the spreading of an operator
we can expand it in a basis of local operators
\begin{equation}
U^{\dagger}H U = \sum_{\alpha}C_{\alpha} O_{\alpha} \ ,
\end{equation}
and calculate the entropy
\begin{equation}
S = -\sum_{\alpha}|C_{\alpha}|^2 Log |C_{\alpha}|^2 \ .
\end{equation}
In fact we can ask more generally about the scaling with time
of the Renyi entropy associated with the operator spread.
\item One can also inquire what is the relation between these
entropies and the entanlgement entropies that have been generated
by the quantum gates.
We can try to check how much entanglement we generate in the circuit.
Suppose we have $n$ qubits and let's divide to two groups
$n_A,n_B, n_A \leq n_B, n_A+n_B=n$.
Let is calculate the reduced density matrix of our state
as a functiin of time:
\begin{equation}
\rho_A(t) = Tr_B \left(\rho(t)\right) \ .
\end{equation}
Question: how does the purity
\begin{equation}
I(t) = Tr_A \left(\rho_A^2\right) \ ,
\end{equation}
decay with time.
If we average over the uniform distribution
it should probably decay exponentially.
What happens along the trajectory that we are following
trying to get to the minimum.
\item Do conserved charges limit the growth of entanglement ?
\end{itemize}
\begin{itemize}
\item We can change the architecture of the circuit - viewed as a graph we should not have a complete
graph but a much more sparse one, like an Expander graph.
This has two advantages: first, from theoretical point of view the more local ${U}^{\dagger}(\theta) H U$
the longer it takes for it to spread.
Second, a complete graph cannot be designed as a physical quantum circuit.
Can we identify better circuit architectures based on graph properties ?
If we are concerned with operator mixing then it seems that the right think to do is
to study random walk on a graph.
\item A possible direction to follow is to view the spread of the operator as a random motion and estimate the mean and the variance
of the spread. This depends on the circuit architecture.
We can analyze the functional relationship between the
variance of the spread and the Hessian of the gradient motion in parameter space.
\item We can change the distribution from which we draw the parameters. Since the random quantum circuit is a graph, we
can think about a (biased) random motion on a graph.
\end{itemize}
We expect that the entanglement property of initial states should be sufficient to predict whether the optimization outcome will be good or bad. Let's pick some specific states with distinct optimization outcomes, and calculate their Renyi-2 entropies with different choices of subregions.
\begin{itemize}
\item Does the optimization respect the "foliation" structure by entanglement?
When $p=\frac{1}{2}$ we should have zero $v_B$ and zero spread. Can we check that ?
Can we see what does it does it do for optimization ?
The other extreme limit is $v_B=1$, where there is a maximal spread.
\item It is possible that the complexity of the mixing caused by the diffusion is actually secondary to the more complex mixing of entanglement.
In particular, the scaling of the scrambling time due to diffusion is different
than the one due to entanglement. That means that these two diagnostics of chaos
are very different on different graphs. More later.
\item In general, it would be interesting to quantify the difference in the operator
spreading between chaotic and integrable systems (there is already some work in this).
\item It seems that the only exponent in the problem is the
velocity $v_b$. Is it true the scaling of the entanglement entropy
is also determined by $v_b$?
This seems to be not the case and these are independent quantities.
\item The importance of the Hamiltonian: intuitively it is clear that the more local the Hamiltonian the less spread is expected.
We need to charaterize it.
\end{itemize}
For the entanglement entropy:
\begin{equation}
|S_A(\rho_{g}) - S_A(\rho_{c})|
\leq \Delta \log (2^{n_A}-1)- \Delta\log \Delta - (1-\Delta) \log (1-\Delta) \simeq
n_A \Delta \ ,
\end{equation}
and the Renyi entropies:
\begin{equation}
|{\cal R}_A^k(\rho_{g}) - {\cal R}^k_A(\rho_{c})|
\leq \frac{2^{n_A(k-1)}}{k-1}\left(1 - (1-\Delta)^k - (2^{n_A}-1)^{k-1} \Delta^k \right)
\simeq \frac{k \Delta 2^{n_A(k-1)}}{k-1} \
\end{equation}
where the approximation is for $\Delta\ll 1$.
\section{Opreator Spreading}
Operator spreading is another diagnostic of the chaotic dynamics and information scrambling. A local operator $\mathcal{O}$ that acts on a small number of qubits at time
step $t=0$ evolves at time step $t$ to
\begin{equation}
\mathcal{O}(t) = U(t)^{\dagger} \mathcal{O}(0) U(t)
\label{evo}
\end{equation}
that acts on a large number of qubits.
This spread can be quantified and its growth is ballistic with a characteristic velocity
called the butterfly velocity.
In our case, the value of the butterfly velocity depends on the circuits architecures
and the type of its quantum gates.
In this section we will study this diagnostic and discuss the relationship and differences
from the entanglement diagnostic.
\subsection{Butterfly Velocity}
Any Hermitian operator $\mathcal{O}(t)$ acting on $n$ qubit systems can be written in the Pauli string basis:
\begin{align}
\mathcal{O}(t) = \frac{1}{2^{n/2}}\sum_{j_1, \cdots, j_n} h_{j_1,\cdots,j_n}(t)\, \sigma_{j_1}^{(1)}\otimes \cdots \otimes \sigma_{j_n}^{(n)} \ ,
\end{align}
where
\begin{align}
h_{j_1,\cdots,j_n}(t) \equiv\frac{1}{2^{n/2}} \text{Tr}(\sigma_{j_1}^{(1)}\otimes \cdots \otimes \sigma_{j_n}^{(n)} \cdot \mathcal{O}(t)) \ .
\end{align}
Under the unitary time evolution (\ref{evo})
\begin{align}
\text{Tr}(\mathcal{O}(t)^\dagger \mathcal{O}(t)) = \text{Tr}(\mathcal{O}(0)^\dagger \mathcal{O}(0)) = \frac{1}{2^n}\sum \vert h_{j_1,\cdots,j_n}(t)\vert^2 = \text{constant} \ .
\end{align}
The size of the operator $\mathcal{O}(t)$ is measured by the size of the region where $\mathcal{O}(t)$ does not commute with
a typical local operator $\sigma_a^{(x)}$ at position $x$, which can be quantified as
\begin{align}
\mathcal{C}(x, t) &= \frac{1}{2}\text{Tr} (\rho_\infty [\mathcal{O}(t), \sigma_a^{(x)}]^\dagger [\mathcal{O}(t), \sigma_a^{(x)}]) = \frac{1}{2}\text{Tr} ( [\mathcal{O}(t), \sigma_a^{(x)}]^\dagger [\mathcal{O}(t), \sigma_a^{(x)}]) \nonumber \\
&= \sum_{\substack{j_1,\cdots,j_n \\j_x \neq 0, a}} 2 \vert h_{j_1,\cdots,j_n}(t)\vert^2
\label{c}
\end{align}
We would like to measure (\ref{c}) where the operator ${\cal O}(0)$ is the
Hamiltonian whose expectation value we are minimizing.
It has been argued that the operator spreading in the random quantum circuit
follows
a one-dimensional biased random walk with probability $p$ of shrinking
and $1-p$ for expanding \cite{Nahum:2017yvy}.
The mean of the spread reads:
\begin{equation}
\mu = (1-2p)t = v_B t \ ,
\label{\mu}
\end{equation}
where $t$ is the time (number of a gates), $v_B$ is the butterfly velocity $0 \leq v_B \leq 1$
and the variance
\begin{equation}
\sigma = \left(2p(1-p)t\right)^{\frac{1}{2}} =
\left((1-v_B^2)t\right)^{\frac{1}{2}} \ .
\end{equation}
Using $p=\frac{1}{5}$ for generic four by four unitary matrices (the two-qubit gates), one evaluates the mean and variance
which are consistent with Figure *. We see
that the operator spread growth follows the same pattern as the entropy diagnostics but reaches
the plateau faster. This can be different for other graph architectures.
The value of $p$ depends on the circuit architecture and we can change it by a biased selection of matrices, hence changing the butterfly velocity.
\subsection{Numerical Simulation}
\section{Conserved Charges}
Conserved charges can slow down the process of information scrambling
and entanglement.
They can slow the ballistic growth of operator spreading \cite{sub}.
We will see that while the use of conserved charges allows for an efficient optimization, one has the risk that the ground state of the Hamiltonian
is not and eigenstate of the conserved charge in which case we are limited
by how close we can get to it.
\subsection{Random Quantum Circuit Conserved Charges}
The basic 2-qubit gate ${\cal O}_2$ that we use in the circuit is depicted in figure 1b. The operator reads:
\begin{equation}
{\cal O}_2 = Z R(\theta_{l,i})\otimes R(\theta_{l,j}) \ ,
\end{equation}
where
$Z=diag[1,1,1-1]$
and $R(\theta)$ is a rotation around the $y$-axis.
It is straightforward to see that $[{\cal O}_2,Q]$ on the subspace
$\theta_{l,i} = \theta_{l,j}$ with $Q$ being
a two-parameter family:
\begin{equation}
Q =
\begin{pmatrix}
q_1+q_2 & 0 & 0 & 0\\
0 & q_1 & q_2 & 0 \\
0 & q_2 & q_1 & 0 \\
0 & 0 & 0 & q_1+q_2 \\
\end{pmatrix}
\ .
\end{equation}
In general, $Q$ are symmetry charges on the subspace:
\begin{equation}
\theta_{1i} = \theta_{2i} = c_i, i=1,...,L \ .
\label{sub}
\end{equation}
We would like to see the effect of constraining to the subspace
(\ref{sub}) on the growth of the entanglement entropy and
on the efficiency and success of the optimization.
Note, that the symmetry charges have a similar effect also on the operator spreading growth.
\subsection{Numerical Simulation}
{\color{red} We need the growth of the entanglement entropy and the form of the optimization}.
\section{Spectral Form Factor}
The spectral for factor of the eigenvalue spectrum of the reduced density
$\rho_A$ can be used as a diagnostic of quantum chaos. In this section we will use
it as yet another characteristic of region $A$.
{\color{red} Does the Ising ground state exhibits a "ramp" ?
Can this be a distinguishing feature from the SYK}
Define:
\begin{equation}
Z(\tau) = Tr e^{-i \tau \rho_A} \ ,
\end{equation}
where $\rho_A$ is the reduced density matrix of the circuit. The dimension
of the Hilbert space is $dim H = Z(0) = N$ and $\rho_A$ is an $N\times N$ matrix.
Define:
\begin{equation}
g(\tau) = \langle Z(\tau \bar{Z(\tau)} \rangle \ .
\end{equation}
$g(\tau)$ exhibits a ramp-like behaviour and it serves as a diagnostic
of quantum chaos.
We will be interested in $\frac{|Z(\tau)|^2}{Z(0)^2}$.
Consider first the case that $\rho$ is a reduced density matrix
of a product state and hence a pure state.
In this case:
\begin{equation}
Z(\tau) = \frac{(N-1) + e^{-i \tau }}{N},~~~|Z(\tau)|^2 = N -2 + \frac{2}{N} + \frac{2(N-1)}{N^2} Cos(\tau) \ .
\end{equation}
Consider next the completely mixed state $\rho = \frac{I}{N}$.
\begin{equation}
Z(\tau) = N e^{-i \frac{\tau}{N} },~~~|Z(\tau)|^2 = N^2 \ .
\end{equation}
Consider the thermal distribution:
$\rho= diag[\lambda_1,...,\lambda_N]$
where $\lambda_k= \frac{e^{-\beta E_k}}{\sum_i e^{-\beta E_i}}$ \ .
In this case:
\begin{equation}
|Z(\tau)|^2 = N + \sum_{i\neq j}^N Cos(\frac{(\lambda_i-\lambda_j)\tau}{N}) \leq N^2 \ ,
\end{equation}
where equality is when $\beta=0$.
Thus, in the range between the non-mixed and fully mixed density matrices
we have at large $N$:
\begin{equation}
\frac{1}{N} \leq \frac{|Z(\tau)|^2}{Z(0)^2} \leq 1 \ .
\end{equation}
{\color{red} Can we plot spectral form factor for the circuit ?}
\begin{figure}
\centering
\includegraphics[height=6cm]{figs/SFFQ12.pdf}
\includegraphics[height=6cm]{figs/SFFQ14.pdf}
\caption{(a) The averaged spectral form factor ${|Z(\tau)|^2}/{Z(0)^2}$ at $N=12$. (b) The averaged spectral form factor ${|Z(\tau)|^2}/{Z(0)^2}$ at $N=14$}
\label{fig:SFF}
\end{figure}
\subsection{Complete Graph}
The circuit is depicted in figure *.
The plot of entanglement as a function of the number of layers is
given in figure *.
We see that in this case the fact that all the qubits are linked to each other
already in the first layer implies that the entanglement reaches saturation
by a very shallow circuit, i.e. effective one layer {\color{red} True ?}
\bibitem{bound}
Chen, Z., Ma, Z., Nikoufar, I. et al. Sharp continuity bounds for entropy and conditional entropy. Sci. China Phys. Mech. Astron. {\bf 60}, 020321 (2017).
https://doi.org/10.1007/s11433-016-0367-x
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{figs/optim-Q8-SYK2.pdf}
\caption{The optimization to reach the ground state of the SYK model for different numbers of circuit layers with $n=8$ qubits.
Left: The difference between the circuit state and the ground state energy as a function of the number of optimization steps. Middle: The trace distance between circuit state and the ground state as a function of the number of optimization steps. Right: The difference between the circuit state and the ground state Renyi entropies
as a function of the number of optimization steps. All the three diagnsotics show that optimization does not work.}
\end{figure}
|
\section{Introduction}
The COVID-19 is an infectious disease caused by the coronavirus SARS-CoV-2 \cite{chen_TL,rothan_JA,guo_MMR,tu_IJMS,wang_IJAA}. It has reached the status of a pandemic in March of 2020. Up to January of 2021, it has already infected more than 100 million people, leading to the death of more than 2 million ones \cite{oms_url}. Since the earlier stages of this pandemic, a worldwide effort has been devoted to producing vaccines and antiviral drugs to combat this virus. Some successful investigations yielded vaccines that have started to be applied very recently \cite{sharma_FiI,voysey_TheLancet,corbett_NEJM,baden_NEJM,anderson_NEJM,zhang_medrxiv,jones_TheLancet,burki_TheLancet,jones_TheLacent}. Despite the beginning of vaccination, no consensus about an efficient treatment for already infected patients has been reached so far.
Sars-CoV-2 has a crown-like (spherical) form, and its surface protein (Spike) is directly involved in the infectious process \cite{wu_BBaI,chhikara_CBL,liang_FiI}. The receptor of this virus in human cells is the angiotensin-converting enzyme 2 (ACE2) \cite{zhang2020angiotensin,li2003angiotensin,kuba2005crucial}. Sars-CoV-2 surface protein has two subdivisions named S1 and S2, being S1 the receptor-binding domain (RBD) \cite{xiu2020inhibitors,tai2020characterization,spinello2020rigidity,han2020computational}. The RBD plays a major role in the attachment mechanism of Spike protein to ACE2 \cite{singh2020serine}. After the attachment between them, the virus enters the cell and starts the replication process \cite{xiu2020inhibitors}. In this sense, the strategy of virtual screening for possible inhibitors for the RBD/ACE2 attachment \cite{basu2020molecular} may pave the way for novel therapeutic approaches for the treatment of COVID-19.
Drug repurposing is a feasible way to combat diseases with some similarities \cite{zhou2020network,gordon2020sars,pandey2020potential}. In this scenario, the use of phytochemicals is always an important option to be considered \cite{swain2020phytochemicals}. Among their sub-classes, the flavonoids --- a class of small molecules found in fruits, vegetables, flowers, honey, teas, and wines --- stand out \cite{panche2016flavonoids,harborne2013flavonoids,hertog1992content}. Their pharmacological properties include antimicrobial, antioxidant, anti-inflammatory, and antiviral functions \cite{kumar2013chemistry,sharma2006pharmacological,kim2004anti}.
Flavonoids have been employed as inhibitors for the infection mechanism of several diseases \cite{cushnie2005antimicrobial}. Among them, one can mention malaria, leishmaniasis, Chagas, and dengue \cite{lehane2008common,muzitano2009oral,marin2011vitro,sanchez2000antiviral,kiat2006inhibitory,de2015flavonoids}. They have also been considered in studies aimed at developing therapeutic approaches for cancer treatment \cite{abotaleb2019flavonoids,wang2000therapeutic,eunjung2012}. Very recently, it was reported that Luteolin (a flavonoid found in leaves and shells) is efficient as an anti-inflammatory that can interact with the Sars-CoV-2 surface \cite{yu2020computational} and its main protease \cite{muchtaridi2020natural}. More specifically, it is adsorbed in the Spike protein, inhibiting the Sars-CoV-2 attachment to the ACE2, thus preventing infection. Ngwa and colleagues used computer simulations to address the feasibility of Caflanone, Hesperetin, and Myricetin flavonoids in acting as inhibitors for the ACE2 active site attachment \cite{ngwa_M}. Their results pointed to the ability of Caflanone in inhibiting the transmission of the Sars-CoV-2 virus from mother to fetus in pregnancy. Pandey \textit{et. al.} conducted molecular docking and dynamics simulations considering ten flavonoid and non-flavonoid compounds (by using phytochemicals and hydroxychloroquine, respectively) to verify their performance in inhibiting the RBD/ACE2 interaction \cite{pandey_JBSD}. Their findings indicate that Fisetin, Quercetin, and Kamferol molecules couple to RBD/ACE2 complex with good binding affinities. In this sense, they can be explored as possible anti-Sars-CoV-2 agents. Despite the success of these molecules inhibiting the RBD/ACE2, other flavonoids should be tested to broaden the list of possible inhibitors and to confirm their potential in developing new therapeutic approaches for the treatment of COVID-19.
Herein, in silico molecular docking analysis was carried out to propose potential flavonoid candidates in preventing the RBD/ACE2 attachment. These candidates are sixteen different flavonoids present in the Peppermint (Mentha piperita) leaf \cite{DOLZHENKO201067,McKayAreviewof,bodalska2020analysis,areias2001phenolic,riachi2015peppermint,mahendran2020ethnomedicinal,peterson1998flavonoids}. Peppermint is a perennial herb and medicinal plant native to Europe widely used for treating stomach pains, headaches, and inflammation of muscles \cite{McKayAreviewof,mahendran2020ethnomedicinal,peterson1998flavonoids}. Well-known for their flavoring and fragrance traits, peppermint leaves and the essential oil extracted from them are used in food, cosmetic and pharmaceutical products \cite{McKayAreviewof,bodalska2020analysis,DOLZHENKO201067,areias2001phenolic}. Our results revealed that Luteolin 7-O-neohesperidoside is the peppermint flavonoid with a higher binding affinity regarding the RBD/ACE2 complex (about -9.18 Kcal/mol). On the other hand, Sakuranetin was the one with the lowest affinity (about -6.38 Kcal/mol). Binding affinities of the other peppermint flavonoids ranged from -6.44 Kcal/mol up to -9.05 Kcal/mol. These binding affinities are equivalent to other ones reported in the literature for the interaction between flavonoids and the RBD/ACE2 complex \cite{DeOliveira2020,Muhseen2020,Hu2020,muchtaridi2020natural,Basu2020,Istifli2020,Peterson2020,yu2020computational,dourav_JBSD,RUSSO2020109211}. Moreover, the binding site surface analysis showed pocket-like regions on the RBD/ACE2 complex that yield several interactions (mostly hydrogen bonds) between the flavonoid and the amino acid residues of the proteins. Definitively, experimental studies and clinical trials should be further performed to evaluate the efficacy of these compounds in the inhibition of the RBD/ACE2 attachment.
\section{Materials and Methods}
\begin{figure*}[htb]
\centering
\includegraphics[width=1.0\linewidth]{figures/fig01}
\caption{Schematic representation of the (a) main proteins involved RBD/ACE2 interaction. These proteins were obtained from Protein Data Bank, ID 6M0J \cite{lan2020structure}. (b) The binding site surface has the following color scheme: gray, red, blue, and white for carbon, oxygen, nitrogen, and hydrogen atoms, respectively. Only the three regions (R1, R2, and R3) were considered in the docking processes since they define the whole RBD/ACE2 interface. The TYR4, GLN24, PHE486, and ASN487 are the residues present in the region R1; GLU35, LYS31, and GLN493 are the residues present in the region R2; ASP355, THR500, GLY502, GLN498, LYS353, and ASP38 are the residues present in the region R3.}
\label{fig:main_protein}
\end{figure*}
Since Sars-CoV-2 infects human cells through the RBD/ACE2 coupling, the idea of checking for small molecules that may inhibit this interaction is recurring and can be useful to propose a combatant drug \cite{Bojadzic2020.10.22.351056}. Here, we used molecular docking to study the interaction between the peppermint flavonoids with the RBD/ACE2 complex. Below, we present the proteins, inhibitors (flavonoids), and the computational protocol involved in our study.
\subsection{Protein Preparation}
Figure \ref{fig:main_protein} presents the main proteins involved RBD/ACE2 interaction that were obtained from Protein Data Bank, ID 6M0J \cite{lan2020structure}. In the left panel of this figure, the ACE2 protein is in blue, while the RBD Sars-CoV-2 one is in red. Three essential regions of inhibition between these proteins were highlighted with the black squares R1, R2, and R3. In the right side of Figure \ref{fig:main_protein} we show the binding site surface colored as gray, red, blue, and white for carbon, oxygen, nitrogen, and hydrogen atoms, respectively. The yellow rectangle highlights the total surface for inhibition with a clear cavity within region R2. The protein resolution is 2.45 \r{A}, and no pKa prediction was carried out. The modeled structure has 41 residues less than the deposited one, but all the important residues in the RDB/ACE2 interface were considered in our study. Just metal ions were considered in the docking study, water molecules were not included.
\subsection{Ligand Preparation}
\begin{figure*}[!htb]
\centering
\includegraphics[width=0.8\linewidth]{figures/fig02}
\caption{Chemical structure of peppermint leaf flavonoids: (a )Acacetin, (b) Apigenin, (c) Apigenin 7-O-neohesperidoside (Apigenin*), (d) Chryseoriol, (e) Hesperidin, (f) Hesperitin, (g) Ladanein, (h) Luteolin, (i) Luteolin 7-O-glucoside (Luteolin*), (j) Luteolin 7-O-glucuronide (Luteolin**), (k) Luteolin 7-O-neohesperidoside (Luteolin***), (l) Narigenin, (m) Pebrellin, (n) Sakuranetin, (o) Thymusin, and (p) Xanthomicrol.}
\label{fig:flavonoids}
\end{figure*}
The peppermint leaf contains sixteen flavonoids \cite{DOLZHENKO201067,areias2001phenolic}, classified into three subcategories: Flavones (Flavonols), Flavorings, and Flavanones \cite{DOLZHENKO201067,areias2001phenolic}. The flavonoids studied here are Acacetin, Apigenin, Apigenin 7-O-neohesperidoside (Apigenin*), Chryseoriol, Hesperidin, Hesperitin, Ladanein, Luteolin, Luteolin 7-O-glucoside (Luteolin*), Luteolin 7-O-glucuronide (Luteolin**), Luteolin 7-O-neohesperidoside (Luteolin***), Narigenin, Pebrellin, Sakuranetin, Thymusin, and Xanthomicrol. Their 3D structures were extracted from PubChem \cite{pub_url}. The chemical structures of these flavonoids can be seen in figure \ref{fig:flavonoids}, while relevant information such as PubChem ID, molecular weight, molecular formula, and subcategory of the flavonoid is presented in table \ref{tab:lig-id}.
\begin{table*}[htb]
\centering
\begin{tabular}{|c|c|c|c|c|}
\hline
\textbf{Compound} & \textbf{PubChem CID} & \textbf{Mol. Weight (g/mol)} & \textbf{Mol. Formula} & \textbf{Type} \\ \hline
Acacetin & 5280442 & 284.26 & \formula{16}{12}{5} & Flavones and Flavonols \\ \hline
Apigenin & 5280443 & 270.24 & \formula{15}{10}{5} & Flavones and Flavonols \\ \hline
Apigenin* & 5282150 & 578.5 & \formula{27}{30}{14} & Flavones and Flavonols \\ \hline
Chryseoriol & 5280666 & 300.26 & \formula{16}{12}{6} & Flavones and Flavonols \\ \hline
Hesperidin & 10621 & 610.6 & \formula{28}{34}{15} & Flavorings \\ \hline
Hesperitin & 72281 & 302.28 & \formula{16}{14}{6} & Flavanones \\ \hline
Ladanein & 3084066 & 314.29 & \formula{17}{14}{6} & Flavones and Flavonols \\ \hline
Luteolin & 5280445 & 286.24 & \formula{15}{10}{6} & Flavones and Flavonols \\ \hline
Luteolin* & 5280637 & 448.4 & \formula{21}{20}{11} & Flavones and Flavonols \\ \hline
Luteolin** & 5280601 & 462.4 & \formula{21}{18}{12} & Flavones and Flavonols \\ \hline
Luteolin*** & 5282152 & 594.5 & \formula{27}{30}{15} & Flavones and Flavonols \\ \hline
Naringenin & 932 & 272.25 & \formula{15}{12}{5} & Flavorings \\ \hline
Pebrellin & 632255 & 374.3 & \formula{19}{18}{8} & Flavones and Flavonols \\ \hline
Sakuranetin & 73571 & 286.28 & \formula{16}{14}{5} & Flavanones \\ \hline
Thymusin & 628895 & 330.29 & \formula{17}{14}{7} & Flavones and Flavonols \\ \hline
Xanthomicrol & 73207 & 344.3 & \formula{18}{16}{7} & Flavones and Flavonols \\ \hline
\end{tabular}
\caption{Potential inhibitors (peppermint leaf flavonoids) of RBD/ACE2 complex and their compound information.}
\label{tab:lig-id}
\end{table*}
\subsection{Molecular Docking Simulation}
Molecular docking consists of computationally analyze the non-covalent binding between macromolecules (receptor) and small molecules (ligand). Here, the macromolecule is the RBD/ACE2 protein complex (Figure \ref{fig:main_protein}), while the ligands are the sixteen flavonoids present in the peppermint leaf (Figure \ref{fig:flavonoids}). SWISSDOCK server was used for the docking simulations \cite{grosdidier2011fast,grosdidier2011swissdock}. In SWISSDOCK, the docking energies are obtained through the CHARMM (Chemistry at HARvard Macromolecular Mechanics) force field \cite{grosdidier2011swissdock,grosdidier2011fast} using a blind docking strategy that spans over 100 trial configurations for each target/ligand input \cite{cheng2020discovery}. The target/ligand configuration with higher binding affinity is selected using the UCFS CHIMERA software \cite{pettersen2004ucsf}, a visualization tool capable of directly import data from the SWISSDOCK server. Finally, the Protein-Ligand Interaction Profiler (PLIP) server \cite{salentin2015plip} is used to characterize the target/ligand interaction for the configuration with a higher binding affinity for each flavonoid regarding the RBD/ACE2 complex. It is worth mentioning that the screening for the ligand position was limited just to the ACE2/RDB interface (regions R1, R2, and R3 in the left panel of Figure \ref{fig:main_protein}). This interface is the crucial region to be considered for blocking the coronavirus entry and replication cycle. The simulation (docking) box used in the screening for the ligand position was limited just to the ACE2/RDB interface (regions R1, R2, and R3 in the left panel of Figure \ref{fig:main_protein}). The docking box has 27.5 \r{A} $\times$ 9.0 \r{A} $\times$ 8.5 \r{A} of dimension and it was centered at (31.5,-36.0,1.5) \r{A}. These parameters cover the three regions depicted in Figure \ref{fig:main_protein}. The accuracy in estimating the ligand positions and related binding affinities are $\pm$2 \r{A} and $\pm$0.01 Kcal/mol, respectively.
\section{Results}
\begin{table}[htb]
\centering
\begin{tabular}{|c|c|}
\hline
\textbf{Compound} & \textbf{$ \Delta G $ {[}Kcal/mol{]}} \\ \hline
Acacetin & -6.70 \\ \hline
Apigenin & -6.87 \\ \hline
Apigenin 7-O-neohesperidoside & -8.08 \\ \hline
Chryseoriol & -6.78 \\ \hline
Hesperidin & -8.67 \\ \hline
Hesperitin & -6.80 \\ \hline
Ladanein & -6.56 \\ \hline
Luteolin & -7.24 \\ \hline
Luteolin 7-O-glucoside & -8.01 \\ \hline
Luteolin 7-O-glucuronide & -7.74 \\ \hline
Luteolin 7-O-neohesperidoside & -9.18 \\ \hline
Naringenin & -6.44 \\ \hline
Pebrellin & -7.07 \\ \hline
Sakuranetin & -6.38 \\ \hline
Thymusin & -6.94 \\ \hline
Xanthomicrol & -6.83 \\ \hline
\end{tabular}
\caption{Peppermint leaf-based flavonoid candidates undergoing docking experiment with their most favorable conformation (lowest binding affinity $\Delta G$ in Kcal/mol).}
\label{tab:deltaG}
\end{table}
\begin{figure*}[!htb]
\centering
\includegraphics[width=0.8\linewidth]{figures/fig03}
\caption{binding site surface (BSS) for the putative best docking target/ligand configurations of (a) Acacetin, (b) Apigenin, (c) Apigenin$^{*}$, (d) Chryseoriol, (e) Hesperidin, (f) Hesperetin, (d) Ladanein, and (d) Luteolin.}
\label{fig:protein_adsorption-01}
\end{figure*}
\begin{figure*}[!htb]
\centering
\includegraphics[width=0.8\linewidth]{figures/fig04}
\caption{binding site surface (BSS) for the putative best docking target/ligand configurations of (a) Luteloin$^{*}$, (b) Luteloin$^{**}$, (c) Luteloin$^{***}$, (d) Naringenin, (e) Pebrellin, (f) Sakuranetin, (d) Thymusin, and (d) Xanthomicrol.}
\label{fig:protein_adsorption-02}
\end{figure*}
\begin{figure*}[!htb]
\centering
\includegraphics[width=0.8\linewidth]{figures/fig05}
\caption{PLIP docked poses for the RBD/ACE2 interaction with (a) Acacetin, (b) Apigenin, (c) Apigenin$^{*}$, (d) Chryseoriol, (e) Hesperidin, (f) Hesperetin, (d) Ladanein, and (d) Luteolin. The hydrogen, hydrophobic, and $\pi$-staking bonds are denoted by the blue, dashed gray, and dashed yellow lines, respectively. The yellow sphere represents the charge center.}
\label{fig:protein-ligand-01}
\end{figure*}
\begin{figure*}[!htb]
\centering
\includegraphics[width=0.8\linewidth]{figures/fig06}
\caption{PLIP docked poses for the RBD/ACE2 interaction with (a) Luteloin$^{*}$, (b) Luteloin$^{**}$, (c) Luteloin$^{***}$, (d) Naringenin, (e) Pebrellin, (f) Sakuranetin, (d) Thymusin, and (d) Xanthomicrol. The hydrogen, hydrophobic, and $\pi$-staking bonds are denoted by the blue, dashed gray, and dashed yellow lines, respectively. The yellow sphere represents the charge center. ACE2 and RBD moieties are shown above and below the horizontal line, respectively.}
\label{fig:protein-ligand-02}
\end{figure*}
After successful docking of the peppermint flavonoids to the RBD/ACE2 complex, several modes of ligand/target interactions were generated with a particular docking score (binding affinity). The binding mode with the lowest binding affinity is regarded as the best one, once it tends to be the most stable. The binding affinity results ($\Delta G$) obtained here are summarized in Table \ref{tab:deltaG}. SWISSDOCK simulations for all the ligands in Figure \ref{fig:flavonoids} revealed significant binding affinities with the target RBD/ACE2 proteins. Luteolin 7-O-neohesperidoside is the peppermint flavonoid with a higher binding affinity regarding the RBD/ACE2 complex (about -9.18 Kcal/mol). On the other hand, Sakuranetin was the one with the lowest affinity (approximately -6.38 Kcal/mol). Binding affinities of the other peppermint flavonoids ranged from -6.44 Kcal/mol up to -9.05 Kcal/mol. As one can note in Tables \ref{tab:lig-id} and \ref{tab:deltaG}, the best docked flavonoids have greater molecular weight. All the binding affinities are close to the ones reported for the RBD/ACE2 interaction with other species of flavonoids \cite{DeOliveira2020,Muhseen2020,Hu2020,muchtaridi2020natural,Basu2020,Istifli2020,Peterson2020,yu2020computational,dourav_JBSD,RUSSO2020109211}. Moreover, they can outperform the binding affinities reported by docking studies using other types of compounds targeting RBD/ACE2 \cite{Basu2020,luan2020targeting,alexpandi2020quinolines,choudhary2020identification,trezza2020integrated,wei2020silico,panda2020structure,utomo2020revealing}, such as Chloroquine and Hydroxychloroquine, which are lower than -8.0 Kcal/mol \cite{Basu2020}. This fact can be attributed to the abundant phenolic hydroxyl group in flavonoids. The hydroxyl group in the sugar group of flavonoids tends to bind more easily with the heteroatoms of amino acids from RBD/ACE2, as will be shown later. In this sense, peppermint flavonoids can compose the list of potential phytochemical inhibitors for the RBD/ACE2 interaction.
Figures \ref{fig:protein_adsorption-01} and \ref{fig:protein_adsorption-02} illustrate the binding site surface (BSS) for the putative best docking target/ligand configurations. For the sake of clarity, these figures show the BSS only for the RBD/ACE2 region highlighted by the yellow rectangle in Figure \ref{fig:main_protein}(b). The following color scheme is adopted for the BSSs: gray, red, blue, and white for carbon, oxygen, nitrogen, and hydrogen atoms, respectively. In the ball-stick representation for the flavonoids, the carbon, oxygen, and hydrogen atoms are shown in the colors cyan, red, and white, respectively. As a general trend, one can note that the flavonoids fit inside the core pocket region (cavity) of the RBD/ACE2 complex. This cavity is displayed as region 2 in Figure \ref{fig:main_protein}(a). Acacetin, Luteolin$^{*}$, Luteolin$^{**}$, Thymusin, and Xanthomicrol were adsorbed on region 1 (see Figure \ref{fig:main_protein}(a)) of the RBD/ACE2 complex. The ligands tend to interact with the oxygen atoms (red spots in the BSS) in regions 1 and 2. These regions establish pocket-like media that yield interactions (mostly hydrogen bonds) between flavonoids and amino acid residues of proteins.
Figures \ref{fig:protein-ligand-01} and \ref{fig:protein-ligand-02} provide a clear picture of the interaction between the amino acid residues of the proteins and peppermint flavonoids. The docked poses (obtained using PLIP \cite{salentin2015plip} show the residues names and the bond types. In the stick representation of flavonoids, the carbon and oxygen atoms are in the orange and red colors, respectively. The hydrogen, hydrophobic, and $\pi$-staking bonds are denoted by the blue, dashed gray, and dashed yellow lines, respectively. The yellow sphere represents the charge center. In Figure \ref{fig:protein-ligand-01} one can note that Acacetin, Apigenin, Apigenin$^{*}$, Chryseoriol, Hesperidin, Hesperetin, Ladenein, and Luteolin interact with RBD/ACE2 mainly through 4, 5, 5, 6, 12, 5, 4, and 8 hydrogen bonds with distinct amino acid residues in both RBD and ACE2 proteins. Similarly, Figure \ref{fig:protein-ligand-02} shows the interaction mechanism between Luteloin$^{*}$, Luteloin$^{**}$, Luteloin$^{***}$, Naringenin, Pebrellin, Sakuranetin, Thymusin, and Xanthomicrol with RBD/ACE2 is mediated by 7, 5, 9, 8, 5, 5, 4, and 4 hydrogen bonds with distinct amino acid residues in both RBD and ACE2 proteins, respectively. In total, 12 hydrophobic bonds were found. The flavonoids and amino acid residues of the proteins involved in this kind of interaction are highlighted below. Some $\pi$-stacking bonds are also present in the RBD/ACE2 interactions with flavonoids expecting for the Hesperidin (Figure \ref{fig:protein-ligand-01}(e)), Luteolin$^{*}$ (Figure \ref{fig:protein-ligand-02}(a)), and Xanthomicrol (Figure \ref{fig:protein-ligand-02}(h)) cases.
Generally speaking, we identified 31 distinct amino acid residues of the RBD/ACE2 interacting with the peppermint flavonoids. The RBD amino acid residues (and their occurrence) are TYR738 (4), LYS682 (5), GLU761 (6), GLN674 (6), TYR770 (6), ARG688 (8), ASP670 (2), GLY761 (4), GLY741 (2), GLN39 (1), ALA740 (1), LYS723 (3), ARG673 (1), and SER759 (1). The ACE2 amino acid residues (and their occurrence) are GLU5 (3), SER1 (5), ASP12 (7), PHE372 (4), ARG375 (9), ASN15 (8), GLU19 (9), PRO371 (1), ANS15 (1), THR71 (1), ALA369 (4), ARG37 (1), ALA368 (1), LYS335 (2), ASP20 (1), TYR760 (1), and LYS8 (1). This result suggests that the target RBD/ACE2 amino acid residues for this class of phytochemicals are ARG375, ASN15, and GLU19 from ACE2, and ARG668 from RBD, based on their higher occurrence. The flavonoids that present hydrophobic bonds with the RBD/ACE2 amino acids, highlighted in the following as (flavonoid/residue), are Ladanein/GLU19, Luteolin/LYS682, Hesperetin/ASN15, Hesperetin/GLU19, Pebrellin/TYR760, Sakuranetin/GLU19, Thymusin/LY58, Acacetin/GLU5, Apigenin/ASN15, Apigenin/PRO371, Apigenin/TYR770, and Chryseoriol/LYS682.
\section{Conclusions}
In summary, a set of phytochemicals (peppermint flavonoids) were screened against the SARS-CoV-2 Spike receptor-binding domain interacting with the human ACE2 receptor. The approach is based on computationally fitting small molecules for the target RBD/ACE2 complex proteins using the 3D structure of the active site with SWISSDOCK \cite{grosdidier2011fast,grosdidier2011swissdock}, subsequently the ranking of the docked compounds with Quimera \cite{pettersen2004ucsf} and interaction analysis with PLIP \cite{salentin2015plip}. Results revealed that Luteolin 7-O-neohesperidoside has a binding affinity of about -9.18 Kcal/mol, the higher one among the flavonoids studied here. On the other hand, Sakuranetin was the one with the lowest affinity (about -6.38 Kcal/mol). Binding affinities of the other peppermint flavonoids ranged from -6.44 Kcal/mol up to -9.05 Kcal/mol. These values outperform the binding affinities reported by docking studies using other types of compounds in which the RBD/ACE2 complex was also the target \cite{cherrak2020potential,bhowmik2021evaluation}.
The binding site surface analysis showed pocket-like regions on the RBD/ACE2 complex that yield several interactions (mostly hydrogen bonds) between the flavonoid and the amino acid residues of the proteins. The interaction mechanism between the flavonoids and amino acid residues of the proteins is mediated by hydrogen bonds, essentially. The presence of some hydrophobic and $\pi-stacking$ bonds was also observed. In total, we identified 31 distinct amino acid residues of the RBD/ACE2 interacting with the peppermint flavonoids. The target RBD/ACE2 amino acid residues for this class of phytochemicals are ARG375, ASN15, and GLU19 from ACE2, and ARG668 from RBD, based on their higher occurrence.
Some \textit{in vitro} studies investigated the antiviral activity of flavonoids in combating SARS-CoV \cite{badshah2021antiviral,RUSSO2020109211} and SARS-CoV2 \cite{abian2020structural,liskova2021flavonoids,seadawy2021vitro,solnier2020flavonoids} infection. Hesperetin, Luteolin, and Apigenin have been demonstrated as potent inhibitors of SARS-CoV-2 3CLpro \textit{in vitro} and can be considered proper candidates for further optimization and development of therapeutic interventions, particularly those related to inflammation processes and immunity \cite{solnier2020flavonoids}. A Luteolin derivative and Apigenin showed the best docking scores in our study.
\begin{acknowledgement}
The authors gratefully acknowledge the financial support from Brazilian Research Councils CNPq, CAPES, and FAPDF and CENAPAD-SP for providing the computational facilities. M.L.P.J. gratefully acknowledge the financial support from CAPES grant 88882.383674/2019-01. L.A.R.J. gratefully acknowledges respectively, the financial support from FAP-DF grant 00193.0000248/2019-32, CNPq grant 302236/2018-0, and UnB/DPI/DEX Edital 01/2020 grant 23106.057604/2020-05. The molecular graphics and analyses were performed with UCSF Chimera, developed by the Resource for Biocomputing, Visualization, and Informatics at the University of California, San Francisco, with support from NIH P41-GM103311.
\end{acknowledgement}
\bibliographystyle{spphys}
|
\section{Introduction}
Spie\ss {\cite{Spiess}} gives some identities including the types of $\sum_{\ell=1}^n\ell^k H_\ell$, $\sum_{\ell=1}^n\ell^k H_{n-\ell}$ and $\sum_{\ell=1}^n\ell^k H_\ell H_{n-\ell}$. In particular, explicit forms for $r=0,1,2,3$ are given.
In this paper, several identities including $\sum_{\ell=1}^n\ell^k H_\ell^{(r)}$ and $\sum_{\ell=1}^n\ell^k H_{n-\ell}^{(r)}$ are shown as special cases of more general results, where $H_\ell^{(r)}$ denotes hyperharmonic numbers defined in (\ref{def:hyperharmonic}). When $r=1$, $H_n=H_n^{(1)}$ is the original harmonic number defined by $H_n=\sum_{j=1}^n 1/j$.
This paper is also motivated from the summation $\sum_{\ell=1}^n\ell^k$, which is related to Bernoulli numbers. In \cite{Adam}, Stirling numbers are represented via harmonic numbers and hypergeometric functions related to Euler sums. In this paper, the sums involving harmonic numbers and their $q$-generalizations are expressed by using Stirling numbers and their $q$-generalizations.
There are many generalizations of harmonic numbers. Furthermore, some $q$-generalizations of hyperharmonic numbers have been proposed. In this paper, based upon a certain type of $q$-harmonic numbers $H_n^{(r)}(q)$ defined in (\ref{def:qhyperharmonic}), several formulas of $q$-hyperharmonic numbers are also derived as $q$-generalizations. These results are also motivated from the $q$-analogues of the sums of consecutive integers (\cite{GZ,Schlosser,Warnaar}).
In order to consider the weighted summations, we are motivated by the fact that the sum of powers of consecutive integers $1^k+2^k+\cdots+n^k$ can be explicitly expressed in terms of Bernoulli numbers or Bernoulli polynomials. After seeing the sums of powers for small $k$:
$$
\sum_{\ell=1}^n\ell=\frac{n(n+1)}{2},~ \sum_{\ell=1}^n\ell^2=\frac{n(n+1)(2 n+1)}{6},~ \sum_{\ell=1}^n\ell^3=\left(\frac{n(n+1)}{2}\right)^2,~ \dots\,,
$$
the formula can be written as
\begin{align}
\sum_{\ell=1}^{n}\ell^{k}&=\frac{1}{k+1}\sum_{j=0}^k\binom{k+1}{j}B_j n^{k+1-j}\label{ber}\\
&=\frac{1}{k+1}(B_{k+1}(n+1)-B_{k+1}(1))\quad\hbox{\cite{CFZ}}\,,
\label{ber1}
\end{align}
where Bernoulli numbers $B_n$ are determined by the recurrence formula
$$
\sum_{j=0}^k\binom{k+1}{j}B_j=k+1\quad (k\ge 0)
$$
or by the generating function
$$
\frac{t}{1-e^{-t}}=\sum_{n=0}^\infty B_n\frac{t^n}{n!}\,,
$$
and Bernoulli polynomials $B_n(x)$ are defined by the following generating function
$$
\frac{te^{xt}}{e^{t}-1}=\sum_{n=0}^\infty B_n(x)\frac{t^n}{n!}\,.
$$
If Bernoulli numbers $\mathfrak B_n$ are defined by
$$
\frac{t}{e^t-1}=\sum_{n=0}^\infty\mathfrak B_n\frac{t^n}{n!}\,,
$$
we can see that $B_n=(-1)^n\mathfrak B_n$. Then
$$
\sum_{\ell=1}^{n}\ell^k=\frac{1}{k+1}\sum_{j=0}^k\binom{k+1}{j}(-1)^{j}\mathfrak B_j n^{k+1-j}\,.
$$
We recall the well-known Abel's identity, which is frequently used in the present paper.
\begin{Lem}{(Abel's identity)}
For any positive integer $n$,
$$
\sum_{\ell=1}^{n} a_{\ell}b_{\ell}=s_{n}b_{n}+\sum_{\ell=1}^{n-1} s_{\ell}(b_{\ell}-b_{\ell+1})\,.
\label{abel}
$$
where
$$
s_{n}=\sum_{\ell=1}^{n}a_{\ell}.
$$
\end{Lem}
In the weight of harmonic numbers $H_n$, we have the following formulas.
\begin{Prop}
For $n,k\ge 1$,
$$
\sum_{\ell=1}^n\ell^k H_\ell=\frac{H_n}{k+1}\sum_{j=0}^k\binom{k+1}{j}B_j n^{k+1-j}
-\sum_{\ell=1}^{n-1}(H_n-H_\ell)\ell^k\,.
$$
\end{Prop}
\begin{proof}
Set $a_{\ell}={\ell^k}$ and $b_{\ell}={H_\ell}$ in Lemma \ref{abel}. With
\begin{align*}
\sum_{\ell=1}^{n-1} s_{\ell}(H_{\ell}-H_{\ell+1})&=s_{1}(H_{1}-H_{2})+\cdots +s_{n-1}(H_{n-1}-H_{n})\\
&=1^kH_{1}+\cdots +{(n-1)}^{k}H_{n-1}-s_{n-1}H_{n}\\
&=-\sum_{\ell=1}^{n-1}(H_n-H_{\ell})\ell^k\,,
\end{align*}
formula (\ref{ber}) gives the result.
\end{proof}
\begin{Prop}
For $n,k\ge 1$,
$$
\sum_{\ell=1}^n\ell^k H_\ell=\frac{H_n}{k+1}(B_{k+1}(n+1)-B_{k+1}(1))
-\sum_{\ell=1}^{n-1}\frac{B_{k+1}(\ell+1)-B_{k+1}(1)}{(k+1)(\ell+1)}\,.
$$
\end{Prop}
\begin{proof}
Set $a_{\ell}={\ell^k}$ and $b_{\ell}={H_\ell}$ in
Lemma \ref{abel}. Formula (\ref{ber1}) gives the result.
\end{proof}
\section{Weighted summations of $q$-hyperharmonic numbers}
Many types of $q$-generalizations have been studied for harmonic numbers (e.g.,\cite{KL,WG}). In this paper, a $q$-hyperharmonic number $H_n^{(r)}(q)$ (see \cite{MS}) is defined by
\begin{equation}
H_n^{(r)}(q)=\sum_{j=1}^n q^j H_j^{(r-1)}(q)\quad(r,n\ge 1)
\label{def:qhyperharmonic}
\end{equation}
with
$$
H_n^{(0)}(q)=\frac{1}{q [n]_q}
$$
and
$$
[n]_q=\frac{1-q^n}{1-q}\,.
$$
Note that
$$
\lim_{q\to 1}[n]_q=n\,.
$$
In this $q$-generalization,
$$
H_n(q)=H_n^{(1)}(q)=\sum_{j=1}^n\frac{q^{j-1}}{[j]_q}
$$
is a $q$-harmonic number. When $q\to 1$, $H_n=\lim_{q\to 1}H_n(q)$ is the original harmonic number and $H_n^{(r)}=\lim_{q\to 1}H_n^{(r)}(q)$ is the $r$-th order hyperharmonic number, defined by
\begin{equation}
H_n^{(r)}=\sum_{\ell=1}^n H_\ell^{(r-1)}\quad\hbox{with}\quad H_n^{(1)}=H_n\,.
\label{def:hyperharmonic}
\end{equation}
Mansour and Shattuck \cite[Identity 4.1, Proposition 3.1]{MS} give the following identities
\begin{align}
H_n^{(r)}(q)&=\binom{n+r-1}{r-1}_q\bigl(H_{n+r-1}(q)-H_{r-1}(q)\bigr)
\label{ph01}\\
&=\sum_{j=1}^n\binom{n+r-j-1}{r-1}_q\frac{q^{r j-1}}{[j]_q},
\label{ph02}
\end{align}
where
$$
\binom{n}{k}_q=\frac{[n]_q!}{[k]_q![n-k]_q!}
$$
is a $q$-binomial coefficient with $q$-factorials $[n]_q!=[n]_q[n-1]_q\cdots[1]_q$.
Note that the identities (\ref{ph01}) and (\ref{ph02}) are $q$-generalization of the identities (\ref{h01}) and (\ref{h02}), respectively.
\begin{align}
H_n^{(r)}&=\binom{n+r-1}{r-1}(H_{n+r-1}-H_{r-1})\quad\hbox{\cite{Conway}}
\label{h01}\\
&=\sum_{j=1}^n\binom{n+r-j-1}{r-1}\frac{1}{j}\quad\hbox{\cite{BGG}}\,.
\label{h02}
\end{align}
So, we can see the recurrence relation
for $r\ge 1$
$$
H_n^{(r+1)}=\frac{n+r}{r}H_n^{(r)}-\frac{1}{r}\binom{n+r-1}{r}\,.
$$
The generating function of this type of $q$-hyperharmonic numbers is given by
\begin{equation}
\sum_{n=1}^\infty H_n^{(r)}(q)z^n=\frac{-\log_q(1-q^r z)}{q(z;q)_r}\quad(r\ge 0)
\label{gen:qhyperharmo}
\end{equation}
(\cite[Theorem 3.2]{MS}), where
$$
-\log_q(1-t)=\sum_{m=1}^\infty\frac{t^m}{[m]_q}
$$
is the $q$-logarithm function and
$$
(z;q)_k:=\prod_{j=0}^{k-1}(1-z q^j)
$$
is the $q$-Pochhammer symbol. When $q\to 1$, (\ref{gen:qhyperharmo}) is reduced to the generating function of hyperharmonic numbers:
$$
\sum_{n=1}^\infty H_n^{(r)}z^n=\frac{-\log(1-z)}{(1-z)^r}\quad(r\ge 0)\,.
$$
In fact, the same form is given by Knuth \cite{Knuth} as
$$
\sum_{n=r-1}^\infty\binom{n}{r-1}(H_n-H_{r-1})z^{n-r+1}=\frac{-\log(1-z)}{(1-z)^r}\quad(r\ge 0)\,.
$$
By (\ref{ph01}), we have
$$
H_n^{(r+1)}(q)-\frac{[n+r]_q}{[r]_q}H_n^{(r)}(q)
=-\frac{q^{r-1}}{[r]_q}\binom{n+r-1}{r}_q\,.
$$
Hence,
\begin{equation}
H_n^{(r+1)}(q)=\frac{[n+r]_q}{[r]_q}H_n^{(r)}(q)-\frac{q^{r-1}}{[r]_q}\binom{n+r-1}{r}_q\,.
\label{qh11}
\end{equation}
By replacing $n$ by $n+1$ and $r$ by $r-1$ in (\ref{qh11}), together with the definition in (\ref{def:qhyperharmonic}), we have
\begin{align}
&[n+r]_q H_n^{(r)}(q)\notag\\
&=[n+1]_q H_{n+1}^{(r)}(q)-q^{n+r-1}\binom{n+r-1}{r-1}_q\,.
\label{qh13}
\end{align}
Mansour and Shattuck {\cite[Theorem 3.3]{MS}} also give the following formula,
\begin{equation}
H_n^{(r)}(q)=\sum_{j=1}^n q^{j(r-m)}\binom{n+r-m-j-1}{r-m-1}H_j^{(m)}(q).
\label{ph022}
\end{equation}
When $q\to 1$, (\ref{ph022}) is reduced to
$$
H_n^{(r)}=\sum_{j=1}^n\binom{n+r-m-j-1}{r-m-1}H_j^{(m)}
$$
(see also \cite{BGG},\cite[2.4.Theorem]{BS}). When $m=0$, (\ref{ph022}) is reduced to (\ref{ph02}).
We prove a more general result of (\ref{ph01}).
\begin{theorem}
For nonnegative integers $n$ and $k$ and a positive integer $r$, we have
$$
\binom{k+r-1}{k}_q H_n^{(k+r)}(q)=\binom{n+k}{n}_q H_{n+k}^{(r)}(q)-\binom{n+k+r-1}{n}_q H_k^{(r)}(q)\,.
$$
\label{gph01}
\end{theorem}
\noindent
{\it Remark.}
If $r=1$ and $k$ is replaced by $r-1$ in Theorem \ref{gph01}, we have the identity (\ref{ph01}).
If $q\to 1$ in Theorem \ref{gph01}, we have the version of the original hyperharmonic numbers in \cite[Theorem 1]{MD}.
\begin{proof}[Proof of Theorem \ref{gph01}.]
The proof is done by induction on $k$. When $k=0$, the identity is clear since both sides are equal to $H_n^{(r)}(q)$. Assume, then, that the identity has been proved for $0,1,\cdots,k$. We give some explanations for the following calculation. Firstly, by replacing $r$ by $k+r$ in (\ref{qh11}), we get the first identity. Secondly, by using the inductive assumption, we get the second identity. Thirdly, by replacing $n$ by $n+k$ and $n$ by $k$ respectively in (\ref{qh13}), we get the third identity. Then, we have
\begin{align*}
&\binom{k+r}{k+1}_q H_n^{(k+r+1)}(q)\\
&=\binom{k+r}{k+1}_q\frac{[n+k+r]_q}{[k+r]_q}H_n^{(k+r)}(q)-\binom{k+r}{k+1}_q\frac{q^{k+r-1}}{[k+r]_q}\binom{n+k+r-1}{k+r}_q\\
&=\frac{[n+k+r]_q}{[k+1]_q}\binom{n+k}{n}_q H_{n+k}^{(r)}(q)-\frac{[n+k+r]_q}{[k+1]_q}\binom{n+k+r-1}{n}_q H_k^{(r)}(q)\\
&\quad -\frac{q^{k+r-1}}{[k+r]_q}\binom{k+r}{k+1}_q\binom{n+k+r-1}{k+r}_q\\
&=\frac{[n+k+1]_q}{[k+1]_q}\binom{n+k}{n}_q H_{n+k+1}^{(r)}(q)-\frac{q^{n+k+r-1}}{[k+1]_q}\binom{n+k}{n}_q\binom{n+k+r-1}{r-1}_q\\
&\quad -\frac{[n+k+r]_q}{[k+1]_q}\binom{n+k+r-1}{n}_q\frac{[k+1]_q}{[k+r]_q}H_{k+1}^{(r)}(q)\\
&\quad +\frac{[n+k+r]_q}{[k+1]_q}\binom{n+k+r-1}{n}_q\frac{q^{k+r-1}}{[k+r]_q}\binom{k+r-1}{r-1}_q\\
&\quad -\frac{q^{k+r-1}}{[k+r]_q}\binom{k+r}{k+1}_q\binom{n+k+r-1}{k+r}_q\\
&=\binom{n+k+1}{n}_q H_{n+k+1}^{(r)}(q)-\binom{n+k+r}{n}_q H_{k+1}^{(r)}(q)\,.
\end{align*}
We used the relation $[n+k+r]_q-q^n[k+r]_q=[n]_q$ in the final part.
\end{proof}
Cereceda \cite{Cereceda} gives the following formula,
$$
\lim_{n\to\infty}\frac{H_{n+1}^{(n+1)}}{H_n^{(n)}}=4\,.
$$
However, the ratio of $q$-hyperharmonic numbers of type $H_n^{(n)}(q)$ has a different phenomenon.
\begin{Prop}
For $|q|<1$, we have
$$
\lim_{n\to\infty}\frac{H_{n+1}^{(n+1)}(q)}{H_n^{(n)}(q)}=q\,.
$$
\label{ratio:ph01}
\end{Prop}
\begin{proof}
Since
$$
\frac{(1-q^{2 n+1})(1-q^{2 n})}{(1-q^{n+1})(1-q^{n})}\to 1\quad(|q|<1,~n\to\infty)
$$
and
\begin{align*}
\frac{H_{2 n+1}(q)-H_n(q)}{H_{2 n-1}(q)-H_{n-1}(q)}
&=\frac{q\left(\dfrac{1}{[n+1]_q}+\dfrac{q}{[n+2]_q}+\cdots+\dfrac{q^n}{[2 n+1]_q}\right)}{\dfrac{1}{[n]_q}+\dfrac{q}{[n+1]_q}+\cdots+\dfrac{q^{n-1}}{[2 n-1]_q}}\\
&\to q\quad(n\to\infty)\,,
\end{align*}
from (\ref{ph01}),
\begin{align*}
\frac{H_{n+1}^{(n+1)}(q)}{H_n^{(n)}(q)}&=\frac{\binom{2 n+1}{n}_q\bigl(H_{2 n+1}(q)-H_n(q)\bigr)}{\binom{2 n-1}{n-1}_q\bigl(H_{2 n-1}(q)-H_{n-1}(q)\bigr)}\\
&\to 1\cdot q=q\,.
\end{align*}
\end{proof}
\begin{theorem}
For positive integers $n$ and $r$,
\begin{align}
\sum_{\ell=1}^n q^{\ell-1}[\ell]_q H_\ell^{(r)}(q)&=\frac{[n]_q[n+r]_q}{[r+1]_q}H_n^{(r)}(q)-\frac{q^r[n-1]_q[n]_q}{([r+1]_q)^2}\binom{n+r-1}{r-1}_q\notag\\
&=\frac{[n]_q[r]_q}{[r+1]_q}H_n^{(r+1)}(q)+\frac{q^{r-1}}{[r+1]_q}\binom{n+r}{r+1}_q\,.
\label{eq:hq1}
\end{align}
\label{th:q110}
\end{theorem}
\begin{proof}
Set $a_{\ell}=q^{\ell-1}\binom{\ell+r-1}{r}_q$ and $b_{\ell}=H_{\ell+r-1}(q)$. By using Lemma \ref{abel}, we have
\begin{align}
&\sum_{\ell=1}^{n}q^{\ell-1}\binom{\ell+r-1}{r}_q H_{\ell+r-1}(q)\notag\\
&=\sum_{\ell=1}^{n}q^{\ell-1}\binom{\ell+r-1}{r}_q H_{n+r-1}(q)-\sum_{\ell=1}^{n-1}\frac{q^{\ell+r-1}}{[\ell+r]_q}\binom{\ell+r}{r+1}_q\notag\\
&=\binom{n+r}{r+1}_q H_{n+r-1}(q)-\frac{q^r}{[r+1]_q}\binom{n+r-1}{r+1}_q\,.
\label{ph04}
\end{align}
Hence,
\begin{align}
&\sum_{\ell=1}^{n}q^{\ell-1}{[\ell]_q} H_{\ell}^{(r)}(q)\notag\\
&=\sum_{\ell=1}^{n}q^{\ell-1}{[\ell]_q} \binom{\ell+r-1}{r-1}_q (H_{\ell+r-1}(q)-H_{r-1}(q))\notag\\
&=[r]_q \sum_{\ell=1}^{n} q^{\ell-1} \binom{\ell+r-1}{r}_q (H_{\ell+r-1}(q)-H_{r-1}(q))\notag\\
&=[r]_q \sum_{\ell=1}^{n} \binom{\ell+r-1}{r}_q H_{\ell+r-1}(q)-[r]_q H_{r-1}(q)\binom{n+r}{r+1}_q\,.
\label{ph05}
\end{align}
With the help of (\ref{ph01}), (\ref{ph04}) and (\ref{ph05}), we get the desired result.
\end{proof}
When $q\to 1$, Theorem \ref{th:q110} is reduced to the following.
\begin{Cor}
For $n,r\ge 1$,
\begin{align*}
\sum_{\ell=1}^{n} \ell H_{\ell}^{(r)}&=\frac{n(n+r)}{r+1}H_n^{(r)}-\frac{(n-1)^{(r+1)}}{(r-1)!(r+1)^2}\\
&=\frac{n r}{r+1}H_n^{(r+1)}+\frac{1}{r+1}\binom{n+r}{r+1}\,,
\end{align*}
where $(x)^{(n)}=x(x+1)\cdots(x+n-1)$ ($n\ge 1$) denotes the rising factorial with $(x)^{(0)}=1$.
\label{th:110}
\end{Cor}
In order to establish similarly structured theorems of $q$-hyperharmonic numbers, we recall the $q$-Stirling numbers of the second kind, denoted by $S_{q}(n,m)$, defined by Carlitz (see e.g. \cite{Carlitz}) as
\begin{align}
([x]_{q})^{n}=\sum_{m=0}^{n} q^{\binom{m}{2}} S_{q}(n,m) ([x]_q)_{(m)}, \quad (n \in \mathbb N),
\label{qstirling}
\end{align}
where $([x]_q)_{(m)}=[x]_q [x-1]_q \cdots [x-m+1]_q$ denotes the $q$-falling factorial with $([x]_q)_{0}=1$. The $q$-Stirling numbers of the second kind $S_q(n, m)$ satisfy the recurrence relation
$$
S_q(n+1,m)=S_q(n,m-1)+ [m]_q \cdot S_q(n,m)
$$
with boundary values
$$
S_q(n,0)=S_q(0,n)=\delta_{n 0}, \quad (n \geq 0)
$$
(\cite{Gould}).
We need a $q$-version of the relation by Spie\ss {\cite{Spiess}}, which is essential in the proof of the following structured theorem of $q$-hyperharmonic numbers of type
$\sum_{\ell=0}^{n} q^{\ell-1} {([\ell]_q)}^{p} H_{\ell}^{(r)}(q)$.
\begin{Lem}
Given summation formulas $\sum_{\ell=0}^{n} q^{\ell-1} \binom{\ell}{j}_q[c_{\ell}]_q=F_q(n,j)$ for $n,j\in\mathbb N$, one has
$$
\sum_{\ell=0}^{n} q^{\ell-1} {([\ell]_q)}^{p} [c_{\ell}]_q=\sum_{\ell=0}^{p} q^{\binom{\ell}{2}} S_q(p,\ell) \cdot {[\ell]_q}! \cdot F_q(n,\ell)\,.
\label{qspiesslemma}
$$
where $S_q(p,\ell)$ denote the $q$-Stirling numbers of the second kind.
\end{Lem}
\begin{proof}
Using (\ref{qstirling}), we have
\begin{align*}
\sum_{\ell=0}^{n} q^{\ell-1} {([\ell]_q)}^{p} [c_{\ell}]_q
&=\sum_{\ell=0}^{n} q^{\ell-1} [c_{\ell}]_q \sum_{j=0}^{p} q^{\binom{j}{2}} S_{q}(p,j) \cdot ([\ell]_q)_{(j)}\\
&=\sum_{j=0}^{p} q^{\binom{j}{2}} S_{q}(p,j) {[j]_q}! \sum_{\ell=0}^{n} q^{\ell-1} \binom{\ell}{j}_q [c_{\ell}]_q \\
&=\sum_{j=0}^{p} q^{\binom{j}{2}} S_{q}(p,j) \cdot {[j]_q}! \cdot F_q(n,j)\,.
\end{align*}
\end{proof}
We introduce some notations.
For $n,r, p \in \mathbb N$, set
\begin{align}
\sum_{\ell=0}^{n} q^{\ell-1} {[\ell]_q}^{p} H_{\ell}^{(r)}(q)=A_q(p,r,n)H_{n}^{(r)}(q)-B_q(p,r,n)\,.\notag
\end{align}
From (\ref{qh11}), for $p=0$, $A_q(0,r,n)=\frac{[n+r]_q}{[r]_q}, B_q(0,r,n)=\frac{q^{r-1}}{[r]_q}\binom{n+r-1}{r}_q$.
From Theorem \ref{th:q110}, for $p=1$, we know that
\begin{align*}
&A_q(1,r,n)=\frac{[n]_q[n+r]_q}{[r+1]_q},\\
&B_q(1,r,n)=\frac{q^r[n-1]_q[n]_q}{([r+1]_q)^2}\binom{n+r-1}{r-1}_q\,.
\end{align*}
\begin{theorem}\label{qhypersturc1}
For $n,r,p\ge 1$,
$$
\sum_{\ell=0}^{n} q^{\ell-1} {[\ell]_q}^{p} H_{\ell}^{(r)}(q)=A_q(p,r,n)H_{n}^{(r)}(q)-B_q(p,r,n)\,,
$$
where
\begin{align*}
&A_q(p,r,n)\\
&=\sum_{\ell=0}^{p} q^{\binom{\ell}{2}+p-1} S_q(p,\ell){[\ell]_q!} \binom{n+r-1}{r-1}_q^{-1} \binom{r+\ell-1}{\ell}_q \binom{r+n}{r+\ell}_q,\\
&B_q(p,r,n)=\sum_{\ell=0}^{p} \frac{q^{\binom{\ell}{2}+r+2p-2}}{[r+\ell]_q}S_q(p,\ell){[\ell]_q!}\binom{r+\ell-1}{\ell}_q\binom{r+n-1}{r+\ell}_q\,.
\end{align*}
\end{theorem}
\begin{proof}
Set $[c_{\ell}]_q=H_{\ell}^{(r)}(q)$ in Lemma \ref{qspiesslemma}. Then by using Lemma \ref{abel}, we have
\begin{align}
F_q(n,p)&=\sum_{\ell=0}^{n} q^{\ell-1} \binom{\ell}{p}_{q} H_{\ell}^{(r)}(q) \notag\\
&=\sum_{\ell=1}^{n} q^{\ell-1} \binom{\ell}{p}_{q} \binom{\ell+r-1}{r-1}_{q} (H_{\ell+r-1}(q)-H_{r-1}(q))\notag\\
&=\sum_{\ell=1}^{n} q^{\ell-1} \binom{r+p-1}{p}_{q} \binom{\ell+r-1}{r+p-1}_{q} (H_{\ell+r-1}(q)-H_{r-1}(q))\notag\\
&=q^{p-1} \binom{r+p-1}{p}_q \binom{r+n}{r+p}_q H_{n+r-1}(q)\notag\\
&\quad -q^{p+r-1} \binom{r+p-1}{p}_q \sum_{\ell=1}^{n-1} \frac{q^{\ell-1}}{[\ell+r]_q}\binom{r+\ell}{r+p}_q\notag\\
&\quad \quad-\binom{r+p-1}{p} \binom{r+n}{r+p} H_{r-1}\notag\\
&=q^{p-1} \binom{r+p-1}{p}_q \binom{r+n}{r+p}_q (H_{n+r-1}(q)-H_{r-1}(q))\notag\\
&\quad \quad -\frac{q^{r+2p-2}}{[r+p]_q}\binom{r+p-1}{p}_q \binom{r+n-1}{r+p}_q. \label{ph06}
\end{align}
With the help of (\ref{ph01}) and (\ref{ph06}), Lemma \ref{qspiesslemma} gives the result.
\end{proof}
When $q\to 1$, Theorem \ref{qhypersturc1} is reduced to the following.
\begin{Cor}\label{hypersturc1}
For $n,r,p\ge 1$,
$$
\sum_{\ell=0}^{n} {\ell}^{p} H_{\ell}^{(r)}=A(p,r,n)H_{n}^{(r)}-B(p,r,n)\,,
$$
where \begin{align*}
&A(p,r,n)=\sum_{\ell=0}^{p} S(p,\ell){\ell!}{\binom{n+r-1}{r-1}^{-1}\binom{r+\ell-1}{\ell}\binom{r+n}{r+\ell}},\\
&B(p,r,n)=\sum_{\ell=0}^{p} \frac{1}{r+\ell}S(p,\ell){\ell!}\binom{r+\ell-1}{\ell}\binom{r+n-1}{r+\ell}\,.
\end{align*}
\end{Cor}
\begin{Example}
$p=2$ gives
\begin{align}
&\sum_{\ell=1}^n q^{\ell-1}([\ell]_q)^2 H_\ell^{(r)}(q)\notag\\
&=\frac{[n]_q[n+r]_q(1+q[r+1]_q[n]_q)}{[r+1]_q[r+2]_q}H_n^{(r)}(q)\notag\\
&\quad-q^{r}[n-1]_q[n]_q\binom{n+r-1}{r-1}_q\frac{q[r+1]_q^2[n]_q-q^3[r]_q^2+[2]_q}{[r+1]_q^2[r+2]_q^2}\,.
\label{ph08}
\end{align}
\end{Example}
Note that $[\ell+1]_q=1+q \cdot [\ell]_q$ and $[\ell+2]_q=[2]_q+q^2 \cdot [\ell]_q$. With the help of Theorem \ref{qhypersturc1} and identities (\ref{eq:hq1}) and (\ref{ph08}), we have the following identities.
For positive integers $n$ and $r$,
\begin{align*}
&\sum_{\ell=1}^n q^{\ell-1}[\ell]_q[\ell+1]_q H_\ell^{(r)}(q)\notag\\
&=\frac{[n]_q[n+r]_q([2]_q[n+2]_q+q^3[r-1]_q[n+1]_q)}{[r+1]_q[r+2]_q}H_n^{(r)}(q)\notag\\
&\quad-q^r[n-1]_q[n]_q\binom{n+r-1}{r-1}_q\frac{[2]_q[r+2]_q^2+q^4[r+1]_q^2[n-2]_q}{[r+1]_q^2[r+2]_q^2}\,.
\end{align*}\
{\scriptsize
\begin{align*}
&\sum_{\ell=1}^nq^{\ell-1}[\ell]_q[\ell+1]_q[\ell+2]_q H_\ell^{(r)}(q)\\
&=\frac{[n]_q[n+r]_q\bigl((r+1)(r+2)n^2+3(r+1)(r+4)n+2(r^2+6 r+11)\bigr)}{[r+1]_q[r+2]_q[r+3]_q}H_n^{(r)}(q)\\
&\quad -q^r[n-1]_q[n]_q\binom{n+r-1}{r-1}_q\frac{(r+1)^2(r+2)^2 n^2+(r+1)^2(r^2+16 r+34)n+12(3 r^2+12 r+11)}{([r+1]_q)^2([r+2]_q)^2([r+3]_q)^2}\,.
\end{align*}
}
To give a more general result, we need the $q$-unsigned Stirling numbers of the first kind $s_{uq}(n,k)$ defined by
$$
[\ell]_q^{(n)}= [\ell]_q [\ell+1]_q\cdots[\ell+n-1]_q=\sum_{k=0}^{n} s_{uq}(n,k) ([\ell]_q)^{k}, \quad (n \in \mathbb N).
$$
The $q$-unsigned Stirling numbers of the first kind $s_{uq}(n,k)$ are well defined since $[\ell+m]_q=[m]_q+q^{m} \cdot [\ell]_q$.
\begin{theorem}\label{qharmonicconsec}
For positive integers $n, p$ and $r$,
\begin{align*}
\sum_{\ell=1}^n q^{\ell-1} [\ell]_q^{(p)} H_\ell^{(r)}(q)=A_{1q}(p,r,n)H_{n}^{(r)}-B_{1q}(p,r,n)\,,
\end{align*}
where
\begin{align*}
&A_{1q}(p,r,n)=\sum_{m=0}^p s_{uq}(p.m) A_{q}(m,r,n),\\
&B_{1q}(p,r,n)=\sum_{m=0}^p s_{uq}(p.m) B_{q}(m,r,n).\\
\end{align*}
\end{theorem}
\begin{proof}
\begin{align*}
&\sum_{\ell=1}^n q^{\ell-1} [\ell]_q^{(p)} H_\ell^{(r)}(q)\\
&=\sum_{\ell=1}^n q^{\ell-1} \sum_{m=0}^{p} s_{uq}(p,m) {[\ell]_q}^{m}H_\ell^{(r)}(q)\\
&=\sum_{m=0}^{p} s_{uq}(p,m) \sum_{\ell=1}^n q^{\ell-1} {[\ell]_q}^{m}H_\ell^{(r)}(q)\\
&=\sum_{m=0}^{p} s_{uq}(p,m) (A_{q}(m,r,n)H_{n}^{(r)}(q)-B_{q}(m,r,n))\\
&=\left( \sum_{m=0}^p s_{uq}(p,m) A_{q}(m,r,n) \right) H_{n}^{(r)}(q)
-\left( \sum_{m=0}^p s_{uq}(p,m) B_{q}(m,r,n) \right).\\
\end{align*}
\end{proof}
When $q\to 1$, Theorem \ref{qharmonicconsec} is reduced to the following.
\begin{Cor}\label{harmonicconsec}
For positive integers $n, p$ and $r$,
\begin{align*}
\sum_{\ell=1}^n (\ell)^{(p)}H_\ell^{(r)}=A_{1}(p,r,n)H_{n}^{(r)}-B_{1}(p,r,n)\,,
\end{align*}
where
\begin{align*}
A_{1}(p,r,n)=\sum_{m=0}^p (-1)^{p+m} s(p.m) A(m,r,n),\\
B_{1}(p,r,n)=\sum_{m=0}^p (-1)^{p+m} s(p.m) B(m,r,n).\\
\end{align*}
\end{Cor}
\subsection{Backward summations}
Now we consider backward summations of $q$-hyperharmonic numbers.
\begin{theorem}
For positive integers $n$ and $r$,
\begin{multline*}
\sum_{\ell=1}^n q^{2n-2\ell} [\ell]_q H_{n-\ell}^{(r)}(q)\\
=\frac{[n]_q[n+r]_q}{[r]_q[r+1]_q}H_n^{(r)}(q)
-\binom{n+r}{r+1}_q \left(\frac{q^{r-1}}{[r]_q}+\frac{q^{r}}{[r+1]_q}-\frac{q^{n+r-1}}{[n+r]_q}\right)\,.
\end{multline*}
\label{qbhh1}
\end{theorem}
\begin{proof}
Set $a_{\ell}=q^{n-\ell}{H_{n-\ell}^{(r)}(q)}$, and $b_{\ell}=q^{n-\ell}{[\ell]_q}$. By using Lemma \ref{abel} and $[\ell+1]_q-q[\ell]_q=1$, we have
\begin{align*}
&\sum_{\ell=1}^n q^{2n-2\ell} [\ell]_q H_{n-\ell}^{(r)}(q)\\
&=[n]_q \cdot H_{n-1}^{(r+1)}(q)+\sum_{\ell=1}^{n-1}(H_{n-1}^{(r+1)}(q)-H_{n-\ell-1}^{(r+1)}(q))(q^{n-\ell}{[\ell]_q}-q^{n-\ell-1}{[\ell+1]_q})\notag\\
&=[n]_q \cdot H_{n-1}^{(r+1)}(q)+\sum_{\ell=1}^{n-1}H_{n-1}^{(r+1)}(q)(q^{n-\ell}{[\ell]_q}-q^{n-\ell-1}{[\ell+1]_q})\\
&\quad +\sum_{\ell=1}^{n-1}H_{n-\ell-1}^{(r+1)}(q)(-q^{n-\ell}{[\ell]_q}+q^{n-\ell-1}{[\ell+1]_q})\\
&=q^{n-1}H_{n-1}^{(r+1)}(q)+\sum_{\ell=1}^{n-1}q^{n-\ell-1}H_{n-\ell-1}^{(r+1)}(q)\\
&=H_{n-1}^{(r+2)}(q)\,.
\end{align*}
With the help of (\ref{ph01}), we get the desired result.
\end{proof}
When $q\to 1$, Theorem \ref{qbhh1} is reduced to the following.
\begin{Cor}
For positive integers $n$ and $r$,
$$
\sum_{\ell=1}^n\ell H_{n-\ell}^{(r)}=\frac{n(n+r)}{r(r+1)}H_n^{(r)}-\frac{(n)^{(r)}\bigl((2 r+1)n+r^2\bigr)}{(r-1)!r^2(r+1)^2}\,.
$$
\label{bhh1}
\end{Cor}
It is more complicated to get a summation formula for the backward summations of higher power. In the case where $q\to 1$, we have more relations, including the following.
\begin{theorem}
For positive integers $n, p$ and $r$,
\begin{align*}
\sum_{\ell=0}^{n} {\ell}^{p} H_{n-\ell}^{(r)}=A_2(p,r,n)H_{n}^{(r)}-B_2(p,r,n)\,.
\end{align*}
\label{bhh2}
where $A_2(p,r,n)$ and $B_2(p,r,n)$ satisfy the following relations:
\begin{align*}
&A_2(p,r,n)=A_2(0,r,n)\left(1+\sum_{j=0}^{p-1}\binom{p}{j}A_2(j,r+1,n-1)\right),\\
&B_2(p,r,n)\\
&=B_2(0,r,n)\left(1+\sum_{j=0}^{p-1}\binom{p}{j}A_2(j,r+1,n-1)\right)+\sum_{j=0}^{p-1}\binom{p}{j}B_2(j,r+1,n-1)\,,
\end{align*}
with the initial values $A_2(0,r,n)=\frac{n}{r}$ and $B_2(0,r,n)=\frac{1}{r}\binom{n+r-1}{r}$.
\end{theorem}
Nevertheless, we can have a different backward summation formula without weights.
\begin{theorem}
For positive integers $n, p$ and $r$,
$$
\sum_{\ell=1}^n q^{p(n-\ell)} H_{n-\ell}^{(r)}(q)=C_q(p,r,n)H_n^{(r)}(q)-D_q(p,r,n)\,,
$$
where $C_q(p,r,n)$ and $D_q(p,r,n)$ satisfy the following recurrence relation.
\begin{align*}
&C_q(p,r,n)=\frac{[n]_q}{[r]_q}\left(q^{(p-1)(n-1)}+(1-q^{p-1})C_q(p-1,r+1,n-1)\right)\\
&D_q(p,r,n)\\
&=\frac{q^{r-1}[n]_q}{([r]_q)^2}\binom{n+r-1}{r}_q \left(q^{(p-1)(n-1)}+(1-q^{p-1})C_q(p-1,r+1,n-1)\right)\\
&\quad +(1-q^{p-1})D_q(p-1,r+1,n-1).
\end{align*}
\end{theorem}
\begin{proof}
Set $a_{\ell}=q^{n-\ell}{H_{n-\ell}^{(r)}(q)}$ and $b_{\ell}=q^{(p-1)(n-\ell)}$. By using Lemma \ref{abel} and $[\ell+1]_q-q[\ell]_q=1$, we have
\begin{align*}
&\sum_{\ell=1}^n q^{p(n-\ell)} H_{n-\ell}^{(r)}(q)\\
&=H_{n-1}^{(r+1)}(q)+\sum_{\ell=1}^{n-1}(H_{n-1}^{(r+1)}(q)-H_{n-\ell-1}^{(r+1)}(q))(q^{(p-1)(n-\ell)}-q^{(p-1)(n-\ell-1)})\notag\\
&=H_{n-1}^{(r+1)}(q)+\sum_{\ell=1}^{n-1}H_{n-1}^{(r+1)}(q)(q^{(p-1)(n-\ell)}-q^{(p-1)(n-\ell-1)})\\
&\quad +\sum_{\ell=1}^{n-1}H_{n-\ell-1}^{(r+1)}(q)(-q^{(p-1)(n-\ell)}+q^{(p-1)(n-\ell-1)})\\
&=q^{(p-1)(n-1)}H_{n-1}^{(r+1)}(q)+(1-q^{p-1})\sum_{\ell=1}^{n-1}q^{(p-1)(n-\ell-1)}H_{n-\ell-1}^{(r+1)}(q)\\
&=q^{(p-1)(n-1)}H_{n-1}^{(r+1)}(q)\\
&\quad +(1-q^{p-1})\left(C_q(p-1,r+1,n-1)H_{n-1}^{(r+1)}(q)-D_q(p-1,r+1,n-1)\right)\,.
\end{align*}
With the help of (\ref{ph01}), we get the desired result.
\end{proof}
\section*{Acknowledgement}
Authors are grateful to the anonymous referee for helpful comments.
|
\section{Introduction}\label{Background}
The Kohn-Sham (KS) density functional theory (DFT) treats an interacting electronic system by mapping it into a system of non-interacting electrons moving in an effective single-particle KS potential~\cite{Giuliani08}. The validity of the KS scheme relies on the assumption of the form of the universal Hohenberg-Kohn energy functional, or the exchange-correlation (XC) functional, the search for which has motivated extensive studies. Despite the success of DFT, for example at the level of local density approximation (LDA), its application to strongly interacting systems has been more challenging and requires more sophisticated treatments~\cite{Seidl99b,Gori-Giorgi09,Malet12,Yu2016}.
The system of interest to us is the fractional quantum Hall effect (FQHE)~\cite{Tsui82}, which occurs when electrons in two dimensions are subjected to a strong perpendicular magnetic field which quenches their kinetic energy and as a result enhances the effects of Coulomb interaction.
Very few papers~\cite{Ferconi95,Heinonen95,Zhang14c,Zhang15b,Zhao17,Hu19} have been written applying DFT to the FQHE since its discovery over the past four decades.
The difficulty of applying DFT to FQHE traces back to the construction of KS DFT by mapping into non-interacting electrons, because the non-interacting system possesses a large degeneracy. That can be seen by considering the canonical Hamiltonian that describes the bulk of the FQHE sample:
\begin{equation}
\mathcal{H}_{\rm LLL}={V}_{\rm ee}=\sum_{i<j}^{N}\frac{e^2}{\epsilon |\vec{r}_i-\vec{r}_j|}\;. \label{bareHam}
\end{equation}
Here $N$ electrons have been taken to be confined in the lowest Landau level (LLL), as appropriate in the limit of very large magnetic fields. The quantity $\epsilon$ is the dielectric constant of the host material. In Eq.~\eqref{bareHam}, the external potential due to a neutralizing background has been suppressed, and the constant kinetic energy of the LLL has been dropped. In the absence of interaction, the ground state for free electrons has a large degeneracy that counts all the possible ways of occupying $N$ of the $N_{\phi}$ Landau level (LL) orbitals, where $N_{\phi}$ is the single-particle degeneracy of the LLL. The switching-on of the interaction leads to correlated ground states at certain special values of the filling factor $\nu=N/N_{\phi}$. This correlated ground states are certain very complicated linear superpositions of the large number of basis functions, such that the average occupation of each single particle orbital is fractional (equal to $\nu=N/N_{\phi}$).
In the KS DFT treatment of the FQHE, mapping the problem into free electrons in a KS potential is equivalent to replacing ${V}_{\rm ee}$ by ${V}_{\rm KS}$ in the Hamiltonian of Eq.~\eqref{bareHam}. The main point is that the KS solution picks out a single Slater determinant, which is insufficient to describe the FQHE state. In fact, the KS solution can only be a nonuniform state that locally describes an integer quantum Hall effect, rather than a state in which each LL orbital has a fractional occupation. This physics has been illustrated in Fig.~\ref{FQHEmap}. In the DFT literature, two different directions have been developed to treat strong correlation effects: the improvement of XC and the replacement of the Slater determinant in the Kohn-Sham system by a multiconfiguration function~\cite{Yu2016}. In the FQHE, one can imagine addressing this issue with the help of an exchange correlation potential that has cusps at certain densities~\cite{Ferconi95,Heinonen95}; Ref.~\onlinecite{Ferconi95} implements an ensemble average over successive iterations to produce, on average, fractional occupation of LL orbitals.
We circumvent this problem by mapping interacting electrons into an auxiliary system of non-interacting composite fermions~\cite{Hu19}, referred to as the KS* system below whenever it is necessary to differentiate it from the standard KS system of non-interacting electrons. Composite fermions (CFs)~\cite{Jain89,Jain07,Halperin20}, often thought of as the bound state of electrons and an even number (2p) of flux quanta [see Fig.~\ref{FQHEmap}(c)], are the emergent weakly-interacting particles of the FQHE. In particular, the FQHE of electrons is an integer quantum Hall effect (IQHE) of CFs. The advantage of this method is that the integrally occupied orbitals of composite fermions represent fractionally occupied levels of electrons.
The KS* equation for composite fermions has been derived in a way analogous to the standard formulation in the KS scheme of DFT.
It is important in the CF DFT to incorporate properly the non-local gauge interaction between CFs, which arises due to the attached fluxes. The long-range nature of the gauge interaction is crucial for capturing the topological properties of the FQHE. An advantage of CF DFT is that it simplifies the modeling of the XC energy. Like in any DFT method, the exact form of the XC enengy of CFs is not known and must be approximated. However, since CFs are weakly interacting, it is reasonable to assume that their XC energy is a smooth function of density within an LDA. (This corresponds to an XC energy for electrons that has cusps at the Jain fillings~\cite{Ferconi95,Heinonen95,Price96}.) Applications of the CF DFT scheme have obtained not only the ground state density and energy, but also the topological properties of the excitations~\cite{Hu19}, including their fractional charge and fractional braiding statistics, which are robust against the specific choice of the XC potential~\cite{Hu21}.
The goal of the present work is to investigate the behavior of the solution as a function of the strength of the XC potential. The XC potential of composite fermions should be dependent on the quantum well width and LL mixing, but the dependence is likely to be complex, and we have not studied that here. Instead, we vary the strength of the XC potential through an {\it ad hoc} multiplicative factor. To allow for most general solutions we implement the CF DFT in a manner that does not impose any symmetry on the solutions. (In Ref.~\onlinecite{Hu19}, we had applied CF DFT to a rotationally symmetric system, while also assuming, for the convenience of numerical calculation, that the solution also has a rotational symmetry.) Our primary finding is that the liquid state of composite fermions, which occurs for weak XC potentials, yields to a crystal phase when the magnitude of exchange-correlation potential is raised relative to the CF cyclotron energy.
One may ask to what extent our study applies to realistic systems. It is possible that a given choice of the XC potential may correspond to some interaction between electrons, but we have not made any attempt to identify the latter. Our results in this work are only to be viewed at a qualitative level.
Previous studies have considered, in a variational calculation, transitions between liquid and crystal states of composite fermions as a function of the filling factor or Landau level mixing~\cite{Archer13,Zhao18}. It is plausible that increasing the strength of the XC potential enhances mixing between the Landau levels of composite fermions, called $\Lambda$ levels ($\Lambda$L), and thus causes a crystal in the same fashion as LL mixing has been shown to do~\cite{Zhao18}.
There is an important difference, however. Those studies compared energies of liquid and crystal states of composite fermions carrying different numbers of vortices. In our current study, we assume that the state is always described in terms of a given species of composite fermions (which will be assumed to be composite fermions carrying two quantized vortices below). In other words, the liquid as well as the crystal states in our study are states of the same species of composite fermions. It is therefore unclear how our results relate to previous studies or to experiments.
\begin{figure}[t]
\includegraphics[width=\columnwidth]{FQHEtoCFv1.pdf}
\caption{Schematic illustration of our KS approach for the FQHE. (a) The real system under consideration, taken to be at $1/3$-filling in an external confining potential $V_{\rm ext}$. In the bulk, each orbital is occupied, on average, by 1/3 of an electron. The fractional occupation is illustrated by the partial coloring of the otherwise empty circles that represent unoccupied orbitals.
(b) The KS spectrum of an auxiliary problem of non-interacting electrons in a KS potential $V_{\rm KS}$, with the renormalized LLs illustrated by purple lines. The orbitals are either fully occupied or empty. (c) The occupation configuration of the auxiliary system of emergent composite fermions (shown as electrons plus two flux quanta) in the composite-fermion Landau level (called $\Lambda$ level) spectrum. The integer quantum Hall effect of CFs in a KS$^*$ potential $V_{\rm KS}^*$ can reproduce the uniform density of the real system in (a). The DFT treatment in this paper maps the real system of FQHE to an auxiliary system in (c). }\label{FQHEmap}
\end{figure}
The plan of the manuscript is as follows. In Sec.~\ref{HowToCalculate} we review the Kohn-Sham equations. In Sec.~\ref{SquareLattice} we explain in detail how to numerically solve the Kohn-Sham equations using a finite difference method on a square lattice. As an application of our method, in Sec.~\ref{results}, we study the ground state density as a function of the strength of the exchange-correlation energy. The effects of temperature, weak disorder, the form of XC energy, and system size are also considered. We summarize our findings in Sec.~\ref{summary}.
\section{KS$^*$ equations for composite fermions}\label{HowToCalculate}
We consider the following Hamiltonian of a FQHE system in a 2D $x$-$y$ plane with an external potential $V_{\rm ext}$:
\begin{equation}
\mathcal{H}={T}_{\rm ee}+{V}_{\rm ee}+{V}_{\rm ext}\;,\label{fullHam}
\end{equation}
where ${T}_{\rm ee}=\sum_{i=1}^{N}\frac{1}{2m_b}(\vec{p}+\frac{e}{c}\vec{A})^2 $ is the kinetic energy operator, $m_b$ is the band mass of electrons and $\vec{A}$ is the vector potential due to a uniform external magnetic field $\vec{B}=B \vec{e}_{z}=\nabla\times \vec{A}(\vec{r})$ along the $z$ direction.
Following the magnetic-field DFT (BDFT)~\cite{Grayce94,kohn04,Tellgren12}, the total energy functional $E_{\rm tot}$ of the FQHE system can be expressed as a functional of the ground state density $\rho(\vec{r})$ as:
\begin{equation}
E_{\mathrm{tot}}[\rho]=E_{\rm K}[\rho]+E_{\rm xc}[\rho]+E_{\rm H}[\rho]+\int d^2\vec{r} V_{\rm ext}(\vec{r})\rho(\vec{r})\;.\label{Interpretation}
\end{equation}
Here the Hartree energy $E_{\rm H}[\rho]$ takes the standard form:
\begin{equation}\label{EH}
E_{\rm H}(\rho)=\frac{1}{2}\int d^2\vec{r} d^2\vec{r}'\frac{\rho(\vec{r})\rho(\vec{r}')}{\epsilon|\vec{r}-\vec{r}'|}\;.
\end{equation}
In order to study the FQHE in the LLL, throughout this paper we have defined the non-interacting kinetic energy functional to be $E_{\rm K}[\rho]\equiv \frac{1}{2}\hbar\omega_B \int d^2 \vec{r}\rho(\vec{r})=\frac{N}{2}\hbar\omega_B$, where $\omega_B=\frac{eB}{m_bc}$ is the cyclotron frequency. The electron XC energy functional $E_{\rm xc}$ is defined by Eq.~\eqref{Interpretation}. It can be equivalently defined through a constrained search formalism~\cite{Levy79,Lieb83}
\begin{equation}
E_{\rm xc}[\rho]\equiv\min_{\Psi\rightarrow \rho(\vec{r})}\langle \Psi|{T}_{\rm ee}+{V}_{\rm ee}|\Psi\rangle-E_{\rm K}[\rho]-E_{\rm H}[\rho]\;,\label{ExcDef}
\end{equation}
which is further simplified by using $\langle \Psi_{\rm LLL}|{T}_{\rm ee}|\Psi_{\rm LLL}\rangle=\frac{N}{2}\hbar\omega_B$ in the LLL:
\begin{equation}
E_{\rm xc}[\rho]=\min_{\Psi_{\rm LLL}\rightarrow \rho(\vec{r})}\langle \Psi_{\rm LLL}|{V}_{\rm ee}|\Psi_{\rm LLL}\rangle-E_{\rm H}[\rho]\;,\label{ExcDefLLL}
\end{equation}
where the many-body wave function $\Psi_{\rm LLL}$ searches for an energy minimum of ${V}_{\rm ee}$ within the LLL Hilbert space. We adopt Eq.~\eqref{ExcDefLLL} in the following.
The above formulation is the standard version of the Hohenberg-Kohm (HK) theorem.
Here $E_{\rm xc}$ depends on the external magnetic field $\vec{B}$ in the BDFT formalism, but is otherwise a universal functional of density that does not depend on the external potential.
We next construct an auxiliary KS* system of CFs (rather than electrons), which is not typical of the standard KS scheme, but is within the formulation of the generalized KS scheme~\cite{Seidl96} and the concepts in the standard KS scheme apply as usual.
We imagine that there exists a reference system that consists of non-interacting CFs, whose ground state density is the same as the ground state density of the FQHE system and is expressed as the sum of the contribution from the occupied KS orbitals as
\begin{equation}
\rho(\vec{r})=\sum_{\alpha} c_{\alpha}|\psi_{\alpha}(\vec{r})|^2\;,\label{rhodef}
\end{equation}
where $c_{\alpha}$ is the occupation number of the KS orbital labeled by $\alpha$. We then define the ``non-interacting" kinetic energy $T_{\rm s}^*[\rho]$ of CFs as
\begin{equation}\label{KSstartotalEKrequireHKform}
T_{\rm s}^*[\rho]=\sum_{\alpha}\langle\psi_\alpha | T^* | \psi_\alpha\rangle,
\end{equation}
where the kinetic energy operator $T^*$ of CFs is
\begin{equation}
T^*= \frac{1}{2m^*}\left(\vec{p}+\frac{e}{c}\vec{A}^*(\vec{r};[\rho])\right)^2\;.
\end{equation}
The important physics of CFs is incorporated through the density-dependent effective vector potential $\vec{A}^*(\vec{r})$, or effective magnetic field $\vec{B}^*$, through
\begin{equation}
\nabla \times \vec{A}^*(\vec{r})=B^*(\vec{r})\vec{e}_{z}=\left[ B- 2\rho(\vec{r})\phi_0\right]\vec{e}_{z}\;.\label{BstarDef1}
\end{equation}
As in the standard KS scheme, a further connection between the FQHE system and the KS system of CFs is made by rewriting $E_{\rm xc}$ as
\begin{equation}
E_{\rm xc}[\rho]=T_{\rm s}^*[\rho]+E^*_{\rm xc}[\rho]\;,\label{ExcPartition}
\end{equation}
which also defines $E^*_{\rm xc}[\rho]$ as the XC energy of CFs. To clarify, despite the absence of the kinetic energy of electrons, the kinetic energy of CFs arises from Coulomb interaction and is included as part of $E_{\rm xc}[\rho]$. In Ref.~\onlinecite{Hu19}, the CF XC energy was approximated in LDA as:
\begin{equation}
E_{\rm xc}^*(\rho)=\varsigma\int d^2\vec{r}\left[a\nu^{1/2}+(b-f/2)\nu+g\right]\rho(\vec{r})\;,\label{EXCStarintroduction0602}
\end{equation}
with parameters
$a=-0.78$, $b=0.28$, $f=0.33$, $g=-0.050$ in units of $\frac{e^2}{\epsilon l_B}$, and $\nu(\vec{r})=2\pi l^2_B\rho(\vec{r})$ is the local filling factor, where $l_B=\sqrt{\frac{\hbar c}{eB}}$ is the magnetic length. The parameter $\varsigma$ will be used to control the strength of the XC potential; $\varsigma=1$ corresponds to the choice in Ref.~\onlinecite{Hu19}. The first term $a\nu^{1/2}$ in $E_{\rm xc}^*$ is chosen to match with the known classical value of energy of the Wigner crystal in the limit $\nu\rightarrow 0$~\cite{Bonsall77}, and the coefficients for higher orders of $\nu$ are chosen to fit the electronic XC energies that are obtained using trial wave functions at the Jain fillings $\nu=n/(2n+1)$. This XC form is suited for the filling factor range $1/3<\nu<1/2$, but we will use it uncritically for arbitrary filling factors below. The value of $g$ gives a constant energy offset and does not affect the KS orbitals or the ground state densities. (We note the a similar multiplicative factor to tune the strength of the XC potential has been used in other contexts, for example for a model hydrogen molecule~\cite{Holst2019}.)
Minimization of $E_{\rm tot}$ is achieved by variations with respect to the KS orbitals, which leads to the KS* equation:
\begin{widetext}
\begin{equation}
H^*\psi_{\alpha}(\vec{r})=\left[T^*
+V_{\rm H}(\vec{r})+V_{\rm ext}(\vec{r})+V_{\rm xc}^{*}(\vec{r})+V^*_{\rm T}(\vec{r}) \right] \psi_{\alpha}(\vec{r}) = \epsilon_{\alpha} \psi_{\alpha}(\vec{r})\;,
\label{singleCFKSintroduction}
\end{equation}
\end{widetext}
where the Hartree potential takes the standard form
\begin{equation}
V_{\rm H}(\vec{r})=\frac{e^2}{\epsilon}\int d^2\vec{r}'\,\frac{\rho(\vec{r}')}{|\vec{r}-\vec{r}'|}\;.\label{VH}
\end{equation}
The CF XC potential is obtained through $V_{\rm xc}^*(\vec{r})\equiv\delta E^*_{\rm xc}/\delta \rho(\vec{r})$ as:
\begin{equation}
V_{\rm xc}^*(\vec{r})=\varsigma\left[\frac{3}{2}a\nu^{1/2}(\vec{r})+(2b-f)\nu(\vec{r})+g\right].\label{VXCStarintroduction}
\end{equation}
In the KS potential experienced by the CFs, there is a non-standard term $V^*_{\rm T}$ that is defined as:
\begin{equation}
V^*_{\rm T}(\vec{r})=\sum_{\alpha}c_{\alpha}\langle\psi_{\alpha}|\frac{\delta {T}^*}{\delta \rho(\vec{r})}|\psi_{\alpha}\rangle\;,\label{DeltaTStarDefintroduction}
\end{equation}
which comes from the density-dependence of the vector potential $\vec{A}^*$ inside the kinetic energy operator ${T}^*$ of CFs. This term is typically much smaller than the other terms in the KS potential. In particular, the effect of $V^*_{\rm T}$ is irrelevant for the topological properties~\cite{Hu19,Hu21}. It is worth emphasizing that while the kinetic energy operator of electrons, ${T}_{\rm ee}$, is absent in the KS* equation, the kinetic energy of CFs enters the KS* equation, and incorporates the non-perturbative effect of the Coulomb interaction.
At finite temperatures, $\{c_{\alpha}\}$ and the chemical potential $\mu$ are determined by
\begin{eqnarray}
c_{\alpha} &=& \frac{1}{1+\exp[(\epsilon_{\alpha}-\mu)/k_\mathrm{B}\tau]}\;, \label{FDdistribution1}\\
N &=& \sum_{\alpha} c_{\alpha}\;.\label{FDdistribution2}
\end{eqnarray}
The occupation number $c_{\alpha}$ reduces to either $0$ or $1$ in the limit of zero temperature.
\section{Numerical procedure for solving the Kohn-Sham equations}\label{SquareLattice}
In this section, we outline the numerical procedure adopted to find the KS solutions. We show how the finite-difference method is implemented on a discretized lattice. We also discuss the algorithm applied in successive iterations to achieve convergence.
\subsection{Choice for the magnetic vector potential $\vec{A}^* $}
We consider a rectangular 2D system of sides $L_x$ and $L_y$.
We discretize the system into a lattice and label each lattice point as $ \vec{r}=(x,y)$, where $x=i a_x$, $y=j a_y$, with $i=1,2,\cdots,N_x$ and $j=1,2,\cdots,N_y$ and the lattice constants are $a_x=L_x/N_x$ and $a_y=L_y/N_y$, respectively. This allows a discretization of the physical quantities.
The effective magnetic field for CFs is given by
\begin{equation}
\vec{B}^*(\vec{r})=B^*(\vec{r})\vec{e}_{z}=\left[1-2\nu(\vec{r}) \right]B\vec{e}_{z}\;, \label{BstarDef2}
\end{equation}
which is equivalent to Eq.~\eqref{BstarDef1}.
In order to write down the KS Hamiltonian explicitly, we pick the symmetric gauge for the bound flux of CFs. The vector potential $\vec{A}^*(\vec{r})$ reads
\begin{eqnarray}\label{AstarComponentSum}
\vec{A}^*(\vec{r}) & = & \int d^2\vec{r}' \, \frac{ B^*(\vec{r}')} {2 \pi | \vec{r}-\vec{r}' |^2 } \, {\vec{e}_z \times (\vec{r}-\vec{r}')} \nonumber\\
& = & \sum_{\vec{r}^\prime {\ne \vec{r}}}\frac{B^*(\vec{r}^\prime)a_xa_y}{2\pi |\vec{r}-\vec{r}^\prime|^2} \, (y^\prime-y,x-x^\prime).
\end{eqnarray}
The sum over $\vec{r}^\prime$ extends over all space. This choice satisfies the Coulomb gauge condition $\nabla\cdot \vec{A}^*=0$, which can be checked explicitly and implies the commutation relation $[\vec{p}, \vec{A}^*]=0$.
\subsection{The discretized Hamiltonian}
The discretized form of the KS Hamiltonian is straightforward. We show here the form explicitly for the $V_{\mathrm{T}}$ term, which is non-standard and also the most complex.
We proceed as follows
\begin{eqnarray}
&&V_{\rm T}(\vec{r})=\sum_{\alpha}c_{\alpha}\langle\psi_{\alpha}|\frac{1}{2m^*}\frac{\delta \left(\vec{p}'+\frac{e}{c}\vec{A}^*(\vec{r}')\right)^2}{\delta \rho(\vec{r})}|\psi_{\alpha}\rangle\label{GeneralDeltaTStarDef1}\\
&&=\sum_{\alpha}c_{\alpha}\langle\psi_{\alpha}|\frac{e}{m^* c}\frac{\delta \vec{A}^*(\vec{r}')}{\delta \rho(\vec{r})}\cdot\left(\vec{p}'+\frac{e}{c}\vec{A}^*(\vec{r}')\right)|\psi_{\alpha}\rangle\label{GeneralDeltaTStarDef4}\\
&&=\frac{\hbar eB}{2m^*c}\sum_{\alpha}c_{\alpha}\int d^2 \bar{\vec{r}}'\,\psi^*_{\alpha}(\bar{\vec{r}}')\nonumber \\
&&\left[\frac{\bar{y}-\bar{y}'}{|\bar{\vec{r}}'-\bar{\vec{r}}|^2}\left(-i\frac{\partial}{\partial\bar{x}'}+\bar{A}_x^*(\bar{\vec{r}}')\right)\right. \nonumber \\
&&\left.
+\frac{\bar{x}'-\bar{x}}{|\bar{\vec{r}}'-\bar{\vec{r}}|^2}\left(-i\frac{\partial}{\partial\bar{y}'}+\bar{A}_y^*(\bar{\vec{r}}') \right)\right]\psi_{\alpha}(\bar{\vec{r}}')\;,
\end{eqnarray}
where $\langle\psi_{\alpha}|{O}(\vec{r}')|\psi_{\alpha}\rangle \equiv \int d^2 \vec{r}'\,\psi^*_{\alpha}(\vec{r}') {O}(\vec{r}')\psi_{\alpha}(\vec{r}')$, $\bar{\vec{r}}=\vec{r}/l_B$, and $\bar{\vec{A}}^* = e \vec{A}^* l_B/c$.
In Eq.~\eqref{GeneralDeltaTStarDef4}, $\left(\vec{p}'\cdot\frac{\delta \vec{A}^*(\vec{r}')}{\delta \rho(\vec{r})}\right)=0$ is used, which can be checked explicitly by noticing that
\begin{equation}
\frac{\partial}{\partial x'}\frac{\delta A_x^*(\vec{r}')}{\delta \rho(\vec{r})}=-\frac{\partial}{\partial y'}\frac{\delta A_y^*(\vec{r}')}{\delta \rho(\vec{r})}=\frac{2\phi_0}{\pi }\frac{(y-y')(x'-x)}{|\vec{r}'-\vec{r}|^4}\;,\label{Coulomb1}
\end{equation}
where we have used
\begin{equation}
\delta [B^*(\vec{r}')]/\delta \rho(\vec{r})= -2\phi_0\delta(\vec{r}-\vec{r}')\;,
\end{equation}
and
\begin{equation}
\frac{\delta \vec{A}^*(\vec{r}')}{\delta \rho(\vec{r})}= \frac{\phi_0}{\pi |\vec{r}'-\vec{r}|^2}(y'-y,x-x')\;.
\end{equation}
The conversion $\frac{\hbar eB}{2m^*c}=0.0010\frac{e^2}{\alpha_{m^*}\epsilon l_B}$ is used in our numerical calculation, where $\alpha_{m^*}$ relates the CF mass to electron mass $m_e$ by $m^*=\alpha_{m^*}\sqrt{B[T]}m_e$. We take $\alpha_{m^*}=0.08$, which is a good approximation for theoretical transport gaps~\cite{Jain07}.
The Hartree potential is calculated using a discretized form of Eq.~\eqref{VH}. To avoid the singularity in the self energy at $\vec{r}=\vec{r}'$, we replace point charge by a uniformly distributed charge on a square region of size $a_x \times a_y$ centered around $\vec{r}$.
In this paper, we consider an external potential $V_{\mathrm{ext}}$ generated by a uniform positive background charge inside a circular region around the origin. For a system of $N$ electrons, the background charge density is chosen as $\rho_{\mathrm{b}}=\frac{\nu_{\mathrm{b}}}{2\pi l^2_B}$ with a radius $R_{\mathrm{b}}=\sqrt{\frac{2N}{\nu_{\mathrm{b}}}}l_B$, where $\nu_{\mathrm{b}}$ is the average ion filling factor $\nu_{\mathrm{b}}=2\pi l^2_B \rho_{\mathrm{b}}$. We make sure that the rectangle $L_x\times L_y$ is chosen to be large enough so as to comfortably enclose the electron system.
\subsection{Numerical procedure for iterations}\label{iteration}
We obtain the self-consistent solution of Eq.~\eqref{singleCFKSintroduction} using the following iterative procedure. (i) We start with an input density $\rho_{\rm in}$. (ii) We obtain $T^*$ and $V^*_{\rm KS}(\vec{r})$ on the left-hand side of Eq.~\eqref{singleCFKSintroduction}, diagonalize the Hamiltonian to obtain the KS$^*$ orbitals, and determine the output density $\rho_{\rm out}=\sum_{\alpha} c_{\alpha}|\psi_{\alpha}|^2$ according to Eq.~\eqref{rhodef}, \eqref{FDdistribution1} and \eqref{FDdistribution2}. Note that we work with a fixed particle number, and therefore need to adjust the chemical potential suitably in each iteration. (iii) The relative difference $\frac{\Delta N}{N}$ between the input and output $\rho$, where $\Delta N=\int |\rho_{\rm in}-\rho_{\rm out}|d^2 \vec{r}$, is called the absolute difference. We accept $\rho_{\rm out}$ as converged if the relative differences between any of the two output densities for $2000$ successive iterations satisfy $\frac{\Delta N}{N}<0.001$. This ensures that the solution is stable and not altered by further iterations. We find that the energy also converges when the above criterion is satisfied.
(iv) If $\rho$ has not converged, we prepare new input density $\rho_{\rm in}$ by mixing some output density into the previous input: $\rho_{\rm in}\rightarrow \eta\rho_{\rm in}+(1-\eta)\rho_{\rm out}$, where the mixing coefficient is $\eta\geq 0.9$. The choice of $\eta$ close to one helps avoid the so-called {\it occupation sloshing}, which can occur due to the large degeneracy in our system~\cite{Woods2019}. We iterate the process until convergence is reached. It is worth mentioning that the calculation of $V_{\rm T}$ requires the information of the KS orbitals. $V_{\rm T}$ is set to zero in the initial input, but in later iterations, it is necessary to also mix the input and output $V_{\rm T}$ in the same way as the mixing of density in each iteration in order to ensure convergence. We always start at a sufficiently high temperature, where convergence is straightforward, and slowly go to lower temperatures, while using the converged density of the previous temperature as the input.
The KS Hamiltonian needs to be updated in each iteration. We notice that a direct calculation of the $\vec{A}^*$, $V_{\rm T}$ and $V_{\rm H}$ terms on each lattice site using for-loops can be time consuming. To increase efficiency, we have utilized the convolution algorithm that is available in the Intel\textsuperscript{\textregistered} Math Kernel Library (MKL) to calculate these terms. For the diagonalization of the Hamiltonian, we use the Feast algorithm~\cite{Polizzi09}, which is also available in the MKL and can take a guess of eigenstates as input to increase efficiency. This is suitable for our purpose because the KS orbitals from the previous iteration serve as a good guess of eigenvectors for the new diagonalization. In this paper, the typical system we consider has $N=40$, $L_x=L_y=35$, $N_x=N_y=210$. The corresponding Hamiltonian is a sparse matrix with a dimension of 44100. With $\eta= 0.95$, the convergence takes several thousand to tens of thousands of iterations depending on whether the converged density is liquid-like or crystal-like. Liquid-like solutions are largely uniform in the bulk and converge quickly. In contrast, crystal-like solutions require significantly larger number of iterations for convergence, in order to adjust the position and shape of the crystalline sites. The corresponding computation time can range from half a day for liquid to one week for crystal solutions respectively for the above typical system size in a single cluster node with 10 cores.
\begin{figure*}[t]
\includegraphics[width=6.5in]{densityVsXc.pdf}
\caption{The ground state density ($\rho$) and the density of states (DOS) of KS* orbitals for a system of $N=40$ electrons with average filling $\nu=1/3$ at temperature $k_{\rm B}\tau=0.01\frac{e^2}{\epsilon l_B}$. The system size is $L_x=L_y=35\; l_B$, and $N_x=N_y=210$. Panels (a-d) depict how the density varies as a function of the exchange-correlation (XC) potential, whose strength is tuned by the prefactor $\varsigma$; these panels correspond to $\varsigma=0, 1, 1.5, 2$. The density is quoted in units of $(2\pi l_B^2)^{-1}$. The FQHE liquid evolves into a crystal with increasing XC energy. Panels (e-f) show the corresponding density of states, $\rho_E$. The cumulative state count $\mathcal{N}(E)$ gives the number of Kohn-Sham orbitals below energy $E$. Only the lowest 85 orbitals are shown. The dashed line marks the location of the Fermi energy. All the energies are in units of $\frac{e^2}{\epsilon l_B}$. }\label{XCcompare}
\end{figure*}
\section{Results}\label{results}
The validity of our DFT results depends on the accuracy of the choice of the XC energy for CFs. We restrict the approximation of $E_{\rm xc}^*$ to the level of LDA and the form of XC energy in Eq.~\eqref{EXCStarintroduction0602} is obtained by fitting to the ground state energies of the uniform systems in the filling factor range $1/3<\nu<1/2$, which is the range where composite fermions carrying two flux quanta are relevant. Since the CFs are weakly interacting, it is reasonable to use a smooth fitting curve for the XC energy, although there has been no study of the exact constraints~\cite{Pittalis11,Dufty11} on the proper choice of the CF XC energy. Eq.~\eqref{EXCStarintroduction0602} is not unique and a different form has been applied in Ref.~\onlinecite{Zhao17}. These slightly different fitting forms do not influence the ground state energy as well as the topological properties when the system is in the filling factor range $1/3<\nu<1/2$. However, we will use this form for the XC energy uncritically for all filling factors, and all of our results are subject to this approximation.
An important point for our discussion below is that the form of the exact $E_{\rm xc}^*$ also depends on various physical factors~\cite{Price96}. For example, one possible factor is LL mixing, which is absent in the theoretical limit of very strong magnetic fields but is relevant for typical magnetic fields and can be quite significant. With LL mixing, one can reformulate the problem in terms of electrons still residing in the LLL but with an effective interaction, which is less repulsive than the Coulomb interaction at short distances. A similar correction arises due to finite width of the quantum well. In principle, one then needs to evaluate the CF cyclotron energy and CF XC energy for the effective interaction, which is likely to change the relative importance of the two terms. We have not made a realistic determination of these effects.
We will tune the strength of the XC energy $E_{\rm xc}^*$ in Eq.~\eqref{EXCStarintroduction0602} by varying $\varsigma$. Notice that $E_{\rm xc}^*$ is negative, so the XC potential increases in magnitude when $\varsigma>1$. We refer to $\varsigma\rightarrow \infty$ and $\varsigma\rightarrow 0$ as the strong and weak XC energy limits, respectively.
\subsection{Appearance of a crystal phase}
We consider a problem with rotational symmetry by choosing the external potential to be generated by a uniform positive charge in a circular region around the origin, as explained in Sec.~\ref{iteration}. The ground state densities for a system of $N=40$ are shown in Fig.~\ref{XCcompare}(a-d) for certain choices of $\varsigma$. The corresponding density of states (DOS) are shown in Fig.~\ref{XCcompare}(e-h). Here we assume a small temperature of $k_B\tau=0.01\frac{e^2}{\epsilon l_B}$, which is less than $10\%$ of the cyclotron gap of $\Lambda$Ls and is useful for finding converged solutions. (Temperature dependence of KS solutions is discussed later.) In the weak XC energy limit ($\varsigma=0$), the ground state density of electrons almost perfectly screens the background density, due to the dominance of the Hartree term and the external potential. The spectrum of the KS solutions shows the formation of $\Lambda$Ls, where the positions of the lowest two $\Lambda$Ls can be seen from the two peaks in the DOS plot in Fig.~\ref{XCcompare}(e).
In order to obtain a smooth curve for DOS, we have replaced the $\delta(E-\epsilon_i)$ in the standard definition of DOS, $\rho_{E}=\sum_i \delta(E-\epsilon_i)$ (where $\epsilon_i$ is the eigenvalue of KS orbitals sorted by $\epsilon_1<\epsilon_2<\ldots$, increasing with $i$), by a Gaussian to define:
\begin{equation}
\rho_{E}=\sum_i\exp[-(E-\epsilon_i)^2/\sigma^2]/\sqrt{2\pi}\sigma\;,
\end{equation}
where $\sigma=0.008\frac{e^2}{\epsilon l_B}$ throughout this paper. The discrete points of $\{\epsilon_i\}$ can be seen from the plot of the cumulative state count
\begin{equation}
\mathcal{N}(E)=\int_{-\infty}^{E}\sum_i \delta(E'-\epsilon_i)dE'\;.
\end{equation}
When $\varsigma=1$ in Eq.~\eqref{EXCStarintroduction0602}, the total density shows stronger oscillations near the edge but still respects a rotational symmetry. In particular, the density profile near the edge of the system first shoots up before it comes down to zero, which is also seen in results from exact diagonalization (ED) in a rotationally symmetric system~\cite{Tsiper01}. For the stronger XC potential of $\varsigma=1.5$, the bulk becomes a crystal; the absence of a crystal at the boundary is a finite temperature effect, as discussed below. The system fully crystalizes when $\varsigma=2$. The formation of crystalline structures breaks the rotational symmetry of the system, which is allowed in our numerical method where no symmetry is assumed. (In contrast, the calculations in Ref.~\onlinecite{Hu19} choose the angular momentum as a good quantum number and reduce the 2D system to effectively a 1D system along the radial direction; the results therein are rotationally symmetric by construction regardless of the strength of the XC potential or the choice of the CF mass. Rotation symmetry is also imposed in the ED calculation in the disk geometry~\cite{Tsiper01}.)
The results are stable and driven by the XC energy. We discuss in Appendix~\ref{error} that our results, in particular the appearance of a crystal phase, are not a numerical artifact of discretization and lattice configuration.
\subsection{Nature of the crystal phase}
The competition between the correlated Wigner crystal state and the liquid state has been studied theoretically in many articles~\cite{Lam84,Levesque84,Price93,Platzman93,Filinov01,Jeon04a,Jeon04b,Zhao18}. In particular, Ref.~\onlinecite{Zhao18} studies the role of LL mixing and finds that the FQHE liquid yields to a crystal when the LL mixing is large. However, there is an important difference between the crystal found in that study and that in our study. In Ref.~\onlinecite{Zhao18}, the $n/(2n+1)$ FQHE liquid of composite fermions carrying two vortices freezes into an {\it electron} crystal with increasing LL mixing. In our study, on the other hand, we obtain a CF crystal of composite fermions carrying two vortices~\cite{Jeon04a,Jeon04b}.
The Wigner crystallization of a system of 2D electrons in a circularly symmetric confining potential has been studied in Ref.~\onlinecite{Filinov01}. They find that crystallization occurs in two stages: first in the radial ordering, and then in the angular ordering. The situation is similar to our findings, though we have CF crystals rather than electron crystals. We calculate the evolution of variances in bulk density along the radial and the azimuthal directions respectively as we increase $\varsigma$ (results not shown). For a small $\varsigma$ ($\varsigma<1$), both variances remain negligible in the liquid phase. When $\varsigma$ is larger than a threshold value, the radial variance first increases significantly; beyond a greater threshold, the azimuthal variance also increases abruptly. This indicates a two-stage crystallization in our results.
It is interesting to ask if our crystal is an example of the so-called Hall crystal~\cite{Kivelson86,Halperin86,Fradkin99}, which is the quantum Hall effect counterpart of the putative supersolid phase of $^4$He atoms. Because our crystal is a correlated crystal of composite fermions, it may appear to be a promising candidate for the Hall crystal phase. One feature that may distinguish the Hall crystal from the Wigner crystal is that, in the former, the number of particles per unit cell is not necessarily an integer~\cite{Tevsanovic89}. We find, for all cases we have studied, that the number of crystal sites in our KS solution at the smallest temperature is equal to the number of composite fermions (which also justifies the term crystal rather than a charge density wave). Another character of the Hall crystal is that, similarly to the Hall liquid state, it hosts chiral edge states. We find that for our crystal, there are no gapless edge states; this is indicated by the presence of a gap at the chemical potential. We thus conclude that our crystal phase is generically not a Hall crystal, but we do not rule out the possibility that the Hall crystal state could be stabilized for some forms of the XC energy.
\begin{figure*}[t]
\includegraphics[width=6.5in]{densityVskBT.pdf}
\caption{The temperature dependence of the ground state density ($\rho$) and the density of states (DOS) of KS* orbitals for a system with $N=40$ particles at average filling $\nu=1/3$. We choose the XC potential with $\varsigma=1.5$. The density is quoted in units of $(2\pi l_B^2)^{-1}$. The panels (a-d) correspond to temperatures $k_{\rm B}\tau=0.1, 0.05, 0.02, 0.001$ in units of $\frac{e^2}{\epsilon l_B}$. The panels (e-f) show the corresponding density of states and the cumulative state count for Kohn-Sham solutions in (a-d). Only the lowest 100 orbitals are shown. Other parameters are the same as those in Fig.~\ref{XCcompare}.}\label{kBTcompare}
\end{figure*}
\subsection{Effects of temperature, disorder, form of XC interaction and system size}
We ask how temperature influences the density in our calculation. Fig.~\ref{kBTcompare} depicts the evolution of a system with $\varsigma=1.5$ as a function of temperature. At the lowest temperature of $k_B\tau=0.001\frac{e^2}{\epsilon l_B}$, the system is crystalline (panel d). As the temperature is raised, the system melts from the edge into the bulk and becomes a liquid-like state when $k_B{\tau}=0.1\frac{e^2}{\epsilon l_B}$, which is approximately the value of the $\Lambda$L gap ($\approx 0.11\frac{e^2}{\epsilon l_B}$), as can be seen from the DOS plot in Fig.~\ref{kBTcompare}(e). It has been proposed that the re-entrant of a solid state can occur in certain parameter regimes when the temperature of a liquid state is raised~\cite{Platzman93}; we have not explored that physics in our calculations.
Next, we test the stability of the results against a weak disorder. We consider onsite disorder by adding to the external potential a term $\sum_i\delta V_{\rm ext}(\vec{r}_i)$, where $V_{\rm ext}(\vec{r}_i)$ is randomly chosen according to a uniform distribution in the range $[-W,W]$, where $W$ is the strength of disorder. In the $\nu=1/3$ state, we find that both the liquid and the crystal states largely remain unaffected for a disorder strength of up to $W = 0.01 \frac{e^2}{\epsilon l_B}$, although we expect that the phase boundary will be slightly modified by disorder~\cite{Price93}.
One may ask how the detailed form of the XC energy/potential influences the results. For that we consider another form of the XC energy $E_{\rm xc}^{*\prime}=\varsigma\int d^2\vec{r}[-0.61\nu^{0.39}-0.165\nu]\rho(\vec{r})$, which gives an XC potential $V_{\rm xc}^{*\prime}=\varsigma(-0.85\nu^{0.39}-0.33\nu)$. (These forms are different approximations for the exact energies in the range $1/3<\nu<1/2$, but have significant differences outside this range.) The qualitative behavior, namely a liquid for small $\varsigma$ and a crystal at large $\varsigma$ is also seen for the new XC potential. However, the phase boundaries are different; for example, the low-temperature KS solution is a crystal for $V_{\rm xc}^{*\prime}$ with $\varsigma=1$.
(We note that both forms of XC energy produce a uniform liquid state in the bulk for $\varsigma=1$ when the system is constrained to be rotationally symmetric~\cite{Hu19}.)
We have also studied the effects of the CF mass $m^*$, and the density rings and crystal sites emerge for a large $m^*$ that decreases the $\Lambda$L gap. This suggests that the formation of $\Lambda$Ls is also important for the stability of the liquid phase.
We have also investigated the behavior as a function of the system size. Away from the transition region, we find that the nature of the ground state is not sensitive to system size. This is illustrated in Fig.~\ref{Ncompare}. Here, the qualitative features of the solution in Fig.~\ref{XCcompare}(c) are retained in smaller systems. For the systems with the same number of particles in Fig.~\ref{Ncompare}, the solutions remain liquid-like for $\varsigma=1$ and crystal-like for $\varsigma=2$ (results not shown).
\begin{figure}[t]
\includegraphics[width=\columnwidth]{densityVsN.pdf}
\caption{The ground state density $\rho$ for different particle numbers $N=5, 10, 15, 20, 25, 30$ with $\varsigma=1.5$ and a uniform background charge of $\nu_{\mathrm b}=1/3$. The qualitative features of a crystalline structure in the bulk and a liquid-like ring along the edge are consistent with those in the large system of $N=40$ shown in Fig.~\ref{XCcompare}(c). We have used $L_x=L_y=30\; l_B$, $N_x=N_y=180$, $k_B\tau=0.01\frac{e^2}{\epsilon l_B}$. Other parameters are the same as those in Fig.~\ref{XCcompare}.}\label{Ncompare}
\end{figure}
\begin{figure*}[t]
\includegraphics[width=6.5in]{SingleParticle.pdf}
\caption{
This figure shows the wave functions of single-particle Kohn-Sham (KS) orbitals for $\varsigma=1, 1.5, 2$. For each value of $\varsigma$, we show eight orbitals of increasing energy, with state indices $i=1 ,9 \ldots, 57$.
The numbers in the parenthesis represent the pair $(\varsigma, i)$. The height and color in the plot represent the magnitude and phase the KS orbitals, respectively. The broken rotational symmetry in the magnitude of a wave function indicates that angular momentum $L_z$ is no longer a good quantum number. In the liquid-like phase in (a-h), the expectation value of $L_z$ (which can be estimated from the number of phase windings over the azimuthal angle) is positively correlated with the average radius of an orbital. This correlation is absent in the crystal-like phases, where KS orbitals are delocalized over a few crystalline peaks. We choose the background charge at $\nu_b=1/3$, $N=40$ and $k_B\tau=0.01 \frac{e^2}{\epsilon l_B}$. Other parameters are the same as those in Fig.~\ref{XCcompare}. }\label{single}
\end{figure*}
\section{Conclusions}\label{summary}
We have studied how the strength of the XC potential between composite fermions dictates their state. For this purpose we develop a numerical procedure to solve the KS equations of CFs in a fashion that allows for crystalline solutions. Our primary finding is that the state evolves from a liquid-like state to a crystal-like state as the strength of XC energy increases.
We mention again that our study is not to be taken as a quantitative treatment of the physics of crystals in the FQHE regime. A notable limitation is that we only consider states of composite fermions carrying two vortices, and do not consider the possibility of a crystal or a liquid of electrons, or of composite fermions with greater number of attached vortices (as might be relevant in regions of small densities).
An obvious direction for future study will be to build better XC potentials that apply to a larger range of filling factors and also include the effects of finite thickness and Landau level mixing. It is possible that the strength of the XC energy of composite fermions relative to their cyclotron energy may also depend on the filling factor, which may be relevant to the formation of a crystal at low fillings. It would be interesting to apply the DFT method to study the edge structure, the effect of disorder and/or anisotropy, spin physics, screening, and of fractional quantum Hall effect in mesoscopic devices.
\section*{Acknowledgements}
Y. H. thanks Junyi Zhang and Jiabin Yu for helpful discussions. J.K.J. thanks Steve Kivelson for an insightful discussion. The work at Penn State was made possible by financial support from the US Department of Energy under Grant No.~DE-SC0005042. Y. H. acknowledges partial financial support from China Scholarship Council. Computations for this research were performed on the Pennsylvania State University’s Institute for Computational and Data Sciences’ Roar supercomputer.
|
\section{Setting and Notation}
We consider learning in infinite-horizon undiscounted ergodic MDPs $(\mathcal{X}, \mathcal{A}, r, P)$, where $\mathcal{X}$ is the state space, $\mathcal{A}$ is a finite action space, $r:\mathcal{X}\times \mathcal{A}\to[0, 1]$ is an unknown reward function, and $P:\mathcal{X}\times \mathcal{A} \to \Delta_{\mathcal{X}}$ is the unknown probability transition function. A policy $\pi:\mathcal{X}\to\Delta_{\mathcal{A}}$ is a mapping from a state to a distribution over actions. Let $\{(x_t^{\pi}, a_t^{\pi})\}_{t=1}^{\infty}$
denote the state-action sequence obtained by following
policy $\pi$. The expected average reward of policy $\pi$ is defined as
\begin{equation}\label{eq:def_jpi}
J_{\pi}:=\lim_{T\to\infty}\mathbb E\left[\frac{1}{T}\sum_{t=1}^T r(x_t^{\pi}, a_t^{\pi})\right].
\end{equation}
Let $\mu_\pi$ denote the stationary state distribution of a policy $\pi$, satisfying $\mu_\pi = \mathbb{E}_{x\sim \mu_{\pi}, a \sim \pi}[P(\cdot|x, a)]$. We will sometimes write $\mu_\pi$ as a vector, and use $\nu_\pi = \mu_\pi \otimes \pi$ to denote the stationary state-action distribution. In ergodic MDPs, $J_\pi$ and $\mu_\pi$ are well-defined and independent of the initial state.
The optimal policy $\pi_*$ is a policy that maximizes the expected average reward.
We will denote by $J_*$ and $\mu_*$ the expected average reward and stationary state distribution of $\pi_*$, respectively.
The value function of a policy $\pi$ is defined as:
\begin{equation}\label{eq:def_v}
V_{\pi}(x) := \mathbb E\left[\sum_{t=1}^{\infty} (r(x_t^{\pi}, a_t^{\pi}) - J_{\pi})|x_1^{\pi} = x\right].
\end{equation}
The state-action value function $Q_{\pi}(x,a)$ is defined as
\begin{equation}\label{eq:def_q}
Q_{\pi}(x,a) := r(x,a) - J_\pi + \sum_{x'}P(x'|x, a) V_\pi(x').
\end{equation}
Notice that
\begin{equation}\label{eq:v_and_q}
V_{\pi}(x) = \sum_{a}\pi(a|x)Q_{\pi}(x, a).
\end{equation}
Equations~\eqref{eq:def_q} and~\eqref{eq:v_and_q} are known as the Bellman equation. If we do not use the definition of $V_{\pi}(x)$ in Eq.~\eqref{eq:def_v} and instead solve for $V_{\pi}(x)$ and $Q_{\pi}(x, a)$ using the Bellman equation, we can see that the solutions are unique up to an additive constant. Therefore, in the following, we may use $V_{\pi}(x)$ and $Q_{\pi}(x, a)$ to denote the same function up to a constant. We will use the shorthand notation $Q_\pi(x, \pi') = \mathbb{E}_{a \sim \pi'(\cdot|x)}[Q_{\pi}(x, a)]$; note that $Q_\pi(x, \pi) = V_\pi(x)$.
The agent interacts with the environment as follows: at each round $t$, the agent observes a state $x_t\in\mathcal{X}$, chooses an action $a_t \sim \pi_t(\cdot | x_t)$, and receives
a reward $r_t:= r(x_t,a_t)$. The environment then transitions to the next
state $x_{t+1}$ with probability $ P(x_{t+1}|x_t,a_t)$. Recall that $\pi_*$ is the optimal policy and $J_*$ is its expected average reward. The regret of an algorithm with respect to this fixed policy is
defined as
\begin{equation}\label{def:regret}
R_T := \sum_{t=1}^T \Big(J_{*} - r(x_t, a_t)\Big).
\end{equation}
The learning goal is to find an algorithm that minimizes the long-term regret $R_T$.
Our analysis will require the following assumption on the mixing rate of policies.
\begin{assumption}[Uniform mixing]
\label{ass:mixing}
Let $H_\pi$ be the state-action transition matrix of a policy $\pi$. Let $\gamma(H_\pi)$ be the corresponding \emph{ergodicity coefficient}~\citep{seneta1979coefficients}, defined as
\[
\gamma(H_{\pi}) := \max_{z: z^\top {\bf 1}=0, \norm{z}_1=1} \norm{z^\top H_{\pi}}_1.
\]
We assume that there exists a scalar $\gamma < 1$ such that for any policy $\pi$, $\gamma(H_\pi) \leq \gamma < 1$.
\end{assumption}
Assumption~\ref{ass:mixing} implies that for any pair of distributions $\nu, \nu'$, $\norm{(\nu - \nu')^\top H_{\pi}}_1 \leq \gamma \norm{\nu - \nu'}_1 $; see Lemma~\ref{lem:contraction} in Appendix~\ref{app:prelim} for a proof.
\section{Algorithm}\label{sec:api}
Our algorithm is similar to the {\textsc{Politex}} schema~\citep{politex}. In each phase $k$, {\textsc{Politex}} obtains an estimate $\widehat Q_{\pi_k}$ of the action-value function $Q_{\pi_k}$ of the current policy $\pi_k$, and then sets the next policy using the mirror descent update rule:
\begin{align}
\pi_{k+1}(\cdot|x)
& = \argmax_{\pi} \widehat Q_{\pi_k}(x, \pi) - \eta^{-1} D_{KL}(\pi \parallel \pi_{k}(\cdot|x)) \notag \\
& = \argmax_{\pi} \sum_{i=1}^k \widehat Q_{\pi_i}(x, \pi) + \eta^{-1} \mathcal{H}(\pi) \notag \\
& \propto \exp \bigg( \eta \sum_{i=1}^k \widehat Q_{\pi_i}(x, \cdot) \bigg),
\label{eq:kl_politex}
\end{align}
where $\mathcal{H}(\cdot)$ is the entropy function.
When the functions $\{ \widehat Q_{\pi_i}\}_{i=1}^k$ are tabular or linear, the above update can be implemented efficiently by simply summing all the table entries or weight vectors. However, with neural network function approximation, we need to keep all networks in memory and evaluate them at each step, which quickly becomes inefficient in terms of storage and computation. Some of the efficient implementations proposed in literature include keeping a subset of the action-value functions \citep{politex}, and using a parametric policy and optimizing the KL-regularized objective w.r.t. the parameters over the available data \citep{tomar2020mirror}.
Our proposed method, presented in Algorithm~\ref{alg:api}, attempts to directly approximate the average of all previous action-value functions
\begin{align*}
Q_k(x, a) := \frac{1}{k} \sum_{i=1}^k Q_{\pi_i}(x, a).
\end{align*}
To do so, we only use a single network, and continually train it to approximate $Q_k$. At each iteration $k$, we obtain a dataset of tuples $\mathcal{D}_k =\{(x_t, a_t, R_t)\}$, where $R_t$ is the empirical return from the state-action pair $(x_t, a_t)$.
We initialize $\widehat Q_k$ to $\widehat Q_{k-1}$ and update it by minimizing the squared error over the union of $\mathcal{D}_k$ and the replay buffer $\cR$. We then update the replay buffer with all or a subset of data in $\mathcal{D}_k$.
In the sequel, in Section~\ref{sec:regret}, we first show that by focusing on estimating the average Q-function, we can improve the regret bound of {\textsc{Politex}} under nearly identical assumption; in Section~\ref{sec:linear}, we instantiate this bound for linear value functions, where we can estimate the average Q-function simply by weight averaging; in Section~\ref{sec:implementation}, we focus on practical implementations, in particular, we discuss the limitations of weight averaging, and provide details on how to leverage replay data when using non-linear function approximation; in Section~\ref{sec:experiments} we present our experimental results; and in Section~\ref{sec:discussion} we make final remarks.
\begin{algorithm}[t!]
\caption{Schema for policy iteration with replay}
\begin{algorithmic}[1]\label{alg:api}
\STATE \textbf{Input:} phase length $\tau$, num. phases $K$, parameter $\eta$
\STATE \textbf{Initialize:} $\pi_1(a|x) = 1/|\mathcal{A}|$, empty replay buffer $\mathcal{R}$
\FOR{$k=1,\ldots, K$}
\STATE
Execute $\pi_k$ for $\tau$ time steps and collect data $\mathcal{D}_k$
\STATE Compute $\widehat Q_k$, an estimate of $Q_{k} = \frac{1}{k}\sum_{i=1}^k Q_{\pi_i}$, from data $\mathcal{D}_k$ and replay $\cR$
\STATE Set $\pi_{k+1}(a|x) \propto \exp\big(\eta k \widehat Q_k(x,a)\big)$
\STATE Update replay $\mathcal{R}$ with $\mathcal{D}_k$
\ENDFOR
\STATE \textbf{Output:} $\pi_{K+1}$
\end{algorithmic}
\end{algorithm}
\section{Discussion}\label{sec:discussion}
The main contributions of our work are an improved analysis and practical implementation of \textsc{Politex}.
On the theoretical side, we show that {\textsc{Politex}} obtains an $O(\sqrt{T})$ high-probability regret bound in uniformly mixing average-reward MDPs with linear function approximation, which is the first such bound for a computationally efficient algorithm. The main limitation of these result is that, similarly to previous works, they hold under somewhat strong assumptions which circumvent the need for exploration. An interesting future work direction would be to relax these assumptions and incorporate explicit exploration instead.
On the practical side, we propose an efficient implementation with neural networks that relies on experience replay, a standard tool in modern deep RL.
Our work shows that experience replay and KL regularization can both be viewed as approximately implementing mirror descent policy updates within a policy iteration scheme. This provides an online-learning justification for using replay buffers in policy iteration, which is different than the standard explanation for their success.
Our work also suggests a new objective for storing and prioritizing replay samples, with the goal of approximating the average of value functions well. This goal has some similarities with continual learning, where experience replay has also lead to empirical successes. One interesting direction for future work would be to explore other continual learning techniques in the approximate implementation of mirror descent policy updates.
\section{Preliminaries}
\label{app:prelim}
We state some useful definitions and lemmas in this section.
\begin{lemma}
\label{lem:contraction}
Let $X_1$ and $X_2$ be a pair of distribution vectors. Let $H$ be the transition matrix of an ergodic Markov chain with a stationary distribution $\nu$, and ergodicity coefficient (defined in Assumption~\ref{ass:mixing}) upper-bounded by $\gamma < 1$. Then
\begin{align*}
\norm{ (H^m)^\top (X_1 - X_2 )}_1 \leq \gamma^m \norm{X_1 - X_2}_1 \,.
\end{align*}
\end{lemma}
\begin{proof}
Let $\{v_1, ..., v_{n}\}$ be the normalized left eigenvectors of $H$ corresponding to ordered eigenvalues $\{\lambda_1, ..., \lambda_n\}$. Then $v_1 = \nu$, $\lambda_1 = 1$, and for all $i \geq 2$, we have that $\lambda_i < 1$ (since the chain is ergodic) and $v_i^\top {\bf 1} = 0$. Write $X_1$ in terms of the eigenvector basis as:
\begin{align*}
X_1 = \alpha_1 \nu + \sum_{i=2}^n \alpha_i v_i \quad \text{and} \quad
X_2 = \beta_1 \nu + \sum_{i=2}^n \beta_i v_i \,.
\end{align*}
Since $X_1^\top {\bf 1} = 1$ and $X_2^\top {\bf 1} = 1$, it is easy to see that $\alpha_1 = \beta_1 = 1$. Thus we have
\begin{align*}
\norm{H^\top (X_1 - X_2)}_1
= \norm{ H^\top \sum_{i=2}^n (\alpha_i - \beta_i) v_i }_1
\leq \gamma \norm{ \sum_{i=2}^n (\alpha_i - \beta_i) v_i}_1 = \gamma \norm{X_1 - X_2}_1
\end{align*}
where the inequality follows from the definition of the ergodicity coefficient and the fact that ${\bf 1}^\top v_i =0$ for all $i \geq 2$. Since
\begin{align*}
{\bf 1}^\top H^\top \sum_{i=2}^n (\alpha_i - \beta_i) v_i = {\bf 1}^\top \sum_{i=2}^n \lambda_i (\alpha_i - \beta_i) v_i = 0,
\end{align*}
the inequality also holds for powers of $H$.
\end{proof}
\begin{lemma}[Doob martingale]
\label{lem:doob}
Let Assumption~\ref{ass:mixing} hold,
and
let $\{(x_t,a_t)\}_{t=1}^T$
be the state-action sequence obtained when following policies $\pi_1, ..., \pi_k$ for $\tau$ steps each from an initial distribution $\nu_0$. For $t\in [T]$,
let $X_t$ be a binary indicator vector with a non-zero element at the linear index of the state-action pair $(x_t, a_t)$.
Define for $i \in [T]$,
\begin{align*}
B_i & = \mathbb{E} \left[ \sum_{t=1}^T X_t | X_1, ..., X_i \right], \quad \text{ and }\,\,
B_0 = \mathbb{E} \left[ \sum_{t=1}^T X_t\right].
\end{align*}
Then, $\{B_i\}_{i=0}^T$ is a vector-valued martingale: $\mathbb{E}[B_i-B_{i-1}|B_0,\dots,B_{i-1}]=0$ for $i=1,\dots,T$,
and $\norm{B_i - B_{i-1}}_1 \le 2 (1 - \gamma)^{-1}$ holds for $i\in [T]$.
\end{lemma}
The constructed martingale is known as the Doob martingale underlying the sum $\sum_{t=1}^T X_t$.
\begin{proof}
That $\{B_i\}_{i=0}^T$ is a martingale follows from the definition. We now bound its difference sequence. Let $H_t$ be the state-action transition matrix at time $t$, and let $H_{i:t} =\prod_{j=i}^{t-1} H_j$, and define $H_{i:i} = I$.
Then, for $t=0, \dots, T-1$, $\mathbb{E}[X_{t+1} | X_t] = H_{t}^\top X_t$ and by the Markov property, for any $i\in [T]$,
\begin{align*}
B_i &
= \sum_{t=1}^i X_t + \sum_{t=i+1}^{T} \mathbb{E}[X_t | X_i]
= \sum_{t=1}^i X_t + \sum_{t=i+1}^{T} H_{i:t}^{\top} X_i, \quad \text{and} \quad
B_0 = \sum_{t=1}^T H_{0:t}^\top X_0.
\end{align*}
For any $i\in[T]$,
\begin{align}
B_i - B_{i-1} &= \sum_{t=1}^i X_t - \sum_{t=1}^{i-1} X_t + \sum_{t=i+1}^{T} H_{i:t}^{\top} X_i - \sum_{t=i}^{T} H_{i-1:t}^\top X_{i-1} \notag \\
& = \sum_{t=i}^{T} H_{i:t}^{\top} (X_i - H_{i-1}^\top X_{i-1}). \label{eq:diffseq}
\end{align}
Since $X_i$ and $H_{i-1}^\top X_{i-1}$ are distribution vectors, under Assumption \ref{ass:mixing} and using Lemma~\ref{lem:contraction},
\begin{align*}
\norm{B_i - B_{i-1}}_1
\leq \sum_{t=i}^T \norm{ H_{i:t}^{\top} (X_i - H_{i-1}^\top X_{i-1})}_1
\leq 2 \sum_{j=0}^{T-i} \gamma^j \leq 2 (1 - \gamma)^{-1} \,.
\end{align*}
\end{proof}
Let $(\cF_k)_k$ be a filtration and define $\mathbb{E}_k[\cdot] := \mathbb{E}[\cdot|\cF_k]$. We will make use of the following concentration results for the sum of random matrices and vectors.
\begin{theorem}[Matrix Azuma, \citet{tropp2012user} Thm 7.1]
\label{thm:azuma}
Consider a finite $(\cF)_k$-adapted sequence $\{X_k\}$ of Hermitian matrices of dimension $m$, and a fixed sequence $\{A_k\}$ of Hermitian matrices that satisfy $\mathbb{E}_{k-1} X_k = 0$ and $X_k^2 \preceq A_k^2$ almost surely. Let $v = \norm{\sum_k A_k^2}$.
Then with probability at least $1-\delta$, $ \norm{ \sum_k X_k }_2 \leq 2\sqrt{2v \ln (m / \delta) }$.
\end{theorem}
A version of Theorem~\ref{thm:azuma} for non-Hermitian matrices of dimension $m_1 \times m_2$ can be obtained by applying the theorem to a Hermitian dilation of $X$,
$\mathcal{D}(X) = \big[\begin{smallmatrix}0 & X \\ X^* & 0 \end{smallmatrix} \big]$, which satisfies $\lambda_{\max}(\mathcal{D}(X)) = \norm{X}$ and $\mathcal{D}(X)^2 = \big[\begin{smallmatrix}XX^* & 0 \\ 0 & X^*X \end{smallmatrix} \big]$. In this case, we have that $v = \max \left(\norm{\sum_k X_k X_k^*}, \norm{\sum_k X_k^* X_k} \right)$.
\begin{lemma} [Hoeffding-type inequality for norm-subGaussian random vectors, \citet{jin2019short}]
\label{lemma:nsg}
Consider random vectors $X_1, \ldots, X_n \in \mathbb{R}^d$ and corresponding filtrations $\cF_i = \sigma(X_1, \ldots, X_i)$ $i \in [n]$, such that $X_i |\cF_{i-1}$ is zero-mean norm-subGaussian with $\sigma_i \in \cF_{i-1}$. That is:
\begin{align*}
\mathbb{E}[X_i |\cF_i] = 0, \quad P(\norm{X_i} \geq t | \cF_{i-1}) \leq 2 \exp(-t^2 / 2\sigma_i^2) \quad \forall t \in \mathbb{R}, \forall i \in [n].
\end{align*}
If the condition is satisfied for fixed $\{\sigma_i\}$, there exists a constant $c$ such that for any $\delta > 0$, with probability at least $1-\delta$,
\begin{align*}
\norm{\sum_{i=1}^n X_i} \leq c \sqrt{\sum_{i=1}^n \sigma_i^2 \log(2d/\delta)} \,.
\end{align*}
\end{lemma}
\section{Proof of Lemma~\ref{lem:linear_est}}
\label{app:linear}
\begin{proof}
Recall that we split each phase into $2m$ blocks of size $b$ and let $\mathcal{H}_i$ and $\mathcal{T}_i$ denote the starting indices of odd and even blocks, respectively.
We let $R_t$ denote the empirical $b$-step returns from the state action pair $(x_t, a_t)$ in phase $i$:
\begin{align*}
R_t = \sum_{i=t}^{t+b} (r_i - \widehat J_{\pi_i}), \quad \widehat{J}_{\pi_i} = \frac{1}{|\mathcal{T}_i|} \sum_{t \in \mathcal{T}_i} r_t.
\end{align*}
We start by bounding the error in $R_t$. Let $X$ be a binary indicator vector for a state-action pair $(x, a)$. Let $H_{\pi}$ be the state-action transition kernel for policy $\pi$, and let $\nu_{\pi}$ be the corresponding stationary state-action distribution. We can write the action-value function at $(x, a)$ as
\begin{align*}
Q_{\pi}(x, a) & = r(x, a) - J_{\pi} + X^\top H_\pi Q_{\pi} \\
&= (X - \nu_\pi)^\top r + X^\top H_{\pi}( r - J_\pi {\bf 1} + H_{\pi} Q_\pi) \\
& = \sum_{i=0}^\infty (X - \nu_\pi)^\top H_{\pi}^i r \,.
\end{align*}
Let $Q_{\pi}^b(x, a) = \sum_{i=0}^b (X - \nu_\pi)^\top H_{\pi}^i r$ be a version of $Q_{\pi}$ truncated to $b$ steps. Under uniform mixing, the difference to the true $Q_\pi$ is bounded as
\begin{align}
|Q_{\pi}(x, a) - Q_{\pi}^b(x, a)| \leq \sum_{i=1}^\infty \left|(X - \nu_\pi)^\top H_{\pi}^{i+b} r \right| \leq \frac{2 \gamma^{b+1}}{1-\gamma} \,.
\end{align}
Let $b_t = Q_{\pi_i}^b(x_t, a_t) - Q_{\pi_i}(x_t, a_t)$ denote the truncation bias at time $t$, and let $z_t = \sum_{i=t}^{t+b}r_i - X_t^\top H_{\pi_i}^{(i-t)}r$ denote the reward noise. We will write
\begin{align*}
R_t = Q_{\pi_i}(x_t, a_t) + b(J_{\pi_i} - \widehat J_{\pi_i}) + z_t + b_t.
\end{align*}
Note that $m= |\mathcal{H}_i|$ and let
\[\widehat M_i = \frac{1}{m} \sum_{t \in \mathcal{H}_i} \phi_t \phi_t^\top + \frac{\alpha}{m} I \,. \]
We estimate the value function of each policy $\pi_i$ using data from phase $i$ as
\begin{align*}
\widehat w_{\pi_i}
&= \widehat M_i^{-1} m^{-1} \sum_{t\in \mathcal{H}_i} \phi_t R_t \\
&= \widehat M_i^{-1} m^{-1} \sum_{t\in \mathcal{H}_i} \phi_t (\phi_t^\top w_{\pi_i} + b_t + z_t + b (J_{\pi_i} - \widehat J_{\pi_i})) + \widehat M_i^{-1} \frac{\alpha}{m} (w_{\pi_i} - w_{\pi_i}) \\
&= w_{\pi_i} + \widehat M_i^{-1} m^{-1} \sum_{t\in \mathcal{H}_i} \phi_t (z_t + b_t + b (J_{\pi_i} - \widehat J_{\pi_i})) - \widehat M_i^{-1} m^{-1} \alpha w_{\pi_i}
\end{align*}
Our estimate $\widehat w_k$ of $w_k = \frac{1}{k} \sum_{i=1}^k w_{\pi_i}$ can thus be written as follows:
\begin{align*}
\widehat w_k - w_k &= \frac{1}{km} \sum_{i=1}^k \sum_{t \in \mathcal{H}_i} \widehat M_i^{-1}\phi_t (z_t + b_t + b (J_{\pi_i} - \widehat J_{\pi_i}))
- \frac{\alpha}{km} \sum_{i=1}^k \widehat M_i^{-1} w_{\pi_i}.
\end{align*}
We proceed to upper-bound the norm of the RHS.
Set $\alpha = \sqrt{m/k}$. Let $C_w$ be an upper-bound on the norm of the true value-function weights $\norm{w_{\pi_i}}_2$ for $i=1,..., K$. In Appendix~\ref{app:m_bound}, we show that with probability at least $1-\delta$, for $m \geq 72 C_{\Phi}^4 \sigma^{-2} (1 -\gamma)^{-2} \log (d/\delta)$, $\norm{\widehat{M}_i^{-1}}_2 \leq 2\sigma^{-2}$. Thus with probability at least $1-\delta$, the last error term is upper-bounded as
\begin{align}
\label{eq:bias_alpha}
\frac{\alpha}{km} \left \| \sum_{k=1}^k \widehat{M}_i^{-1} w_{\pi_i}\right\|_2 \leq 2 \sigma^{-2}C_w (km)^{-1/2}.
\end{align}
Similarly, for
\begin{equation}\label{eq:b_condi}
b \geq \frac{\log ((1-\gamma)^{-1}\sqrt{km})}{ \log (1/\gamma) },
\end{equation}
the norm of the truncation bias term is upper-bounded as
\begin{align}
\label{eq:bias_truncation}
\frac{1}{km} \sum_{i=1}^k \sum_{t \in \mathcal{H}_i} \norm{\widehat M_i^{-1}\phi_t b_t }_2
&\leq
\frac{2 \gamma^b}{km (1 - \gamma)} \sum_{i=1}^k \sum_{t \in \mathcal{H}_i} \norm{\widehat M_i^{-1}\phi_t}_2 \leq 2 \sigma^{-2} C_\Phi (km)^{-1/2}.
\end{align}
To bound the error terms corresponding to reward noise $z_t$ and average-error noise $J_{\pi_i} - \widehat J_{\pi_i}$, we rely on the independent blocks techniques of \citet{Yu94}. We show in Sections \ref{app:zt} and \ref{app:jbias} that with probability $1-2\delta$, for constants $c_1$ and $c_2$, each of these terms can be bounded as:
\begin{align*}
\frac{1}{km} \left \| \sum_{i=1}^k \sum_{t \in \mathcal{H}_i} \widehat M_i^{-1}\phi_t z_t \right \|_2 & \leq 2 c_1 C_\Phi \sigma^{-2} \sqrt{\frac{b \log(2d/\delta)}{km}} \\
\frac{b}{km} \left \| \sum_{i=1}^k (J_{\pi_i} - \widehat J_{\pi_i}) \sum_{t \in \mathcal{H}_i} \widehat{M}_i^{-1} \phi_t \right \|_2 &\leq 2 c_2 C_\Phi \sigma^{-2} b \sqrt{\frac{ \log(2d/\delta)}{km}}.
\end{align*}
Thus, putting terms together, we have for an absolute constant $c$, with probability at least $1-\delta$,
\begin{align*}
\norm{\widehat w_k - w_k}_{2} \leq c \sigma^{-2}(C_w + C_{\Phi})b \sqrt{\frac{ \log(2d/\delta)}{km}}.
\end{align*}
Note that this result holds for every $k\in[K]$ and thus also holds for $k=K$.
\end{proof}
\subsection{Bounding $\|\sum_{i=1}^k \widehat M_i^{-1} \sum_{t\in\mathcal{H}_i} \phi_t z_t\|_2$}
\label{app:zt}
Let $\norm{\cdot}_{\mathrm{tv}}$ denote the total variation norm.
\begin{definition}[$\beta$-mixing]
\label{def:betamix}
Let $\{Z_t\}_{t=1,2,\ldots}$ be a stochastic process.
Denote by $Z_{1:t}$ the collection $(Z_1,\ldots,Z_t)$, where we allow $t=\infty$.
Let $\sigma(Z_{i:j})$ denote the sigma-algebra generated by $Z_{i:j}$ ($i\le j$).
The $k^{\rm th}$ $\beta$-mixing coefficient of $\{Z_t\}$, $\beta_k$, is defined by
\begin{align*}
\beta_k
& = \sup_{t\ge 1} \EE{ \sup_{B\in\sigma(Z_{t+k:\infty})} |P(B|Z_{1:t})-P(B)| } \\
& = \sup_{t\ge 1} \EE{ \norm{P_{Z_{t+k:\infty}|Z_{1:t}}(\cdot|Z_{1:t})-P_{Z_{t+k:\infty}}(\cdot)}_{\mathrm{tv}} }\,.
\end{align*}
$\{Z_t\}$ is said to be $\beta$-mixing if $\beta_k \rightarrow 0$ as $k\rightarrow\infty$.
In particular, we say that
a $\beta$-mixing process mixes at an {\em exponential} rate with parameters $\overline{\beta}, \alpha , \gamma>0$
if $\beta_k \le \overline{\beta} \exp(-\alpha k^\gamma)$ holds for all $k\ge 0$.
\end{definition}
Let $X_{t}$ be the indicator vector for the state-action pair $(x_t, a_t)$ as in Lemma~\ref{lem:doob}. Note that the distribution of $(x_{t+1}, a_{t+1})$ given $(x_t, a_t)$ can be written as $\mathbb{E}[X_{t+1} | X_t]$. Let $H_t$ be the state-action transition matrix at time $t$, let $H_{i:t} =\prod_{j=i}^{t-1} H_j$, and define $H_{i:i} = I$. Then we have that $\mathbb{E}[X_{t+k} | X_{1:t}] = H_{t:t+k}^\top X_t$ and $\mathbb{E}[X_{t+k}] = H_{1:t+k}^\top \nu_0$, where $\nu_0$ is the initial state distribution. Thus, under the uniform mixing Assumption ~\ref{ass:mixing}, the $k^{th}$ $\beta$-mixing coefficient is bounded as:
\begin{align*}
\beta_k & \le \sup_{t\ge 1} \mathbb{E} \sum_{j=k}^\infty \norm{H_{t:t+j}^\top X_t - H_{1:t+j}^\top \nu_0}_1
\leq \sup_{t\ge 1} \mathbb{E} \sum_{j=k}^\infty \gamma^j \norm{X_t - H_{1:t}^\top \nu_0}_1 \leq \frac{ 2 \gamma^k}{1 - \gamma} \,.
\end{align*}
We bound the noise terms using the independent blocks technique of \citet{Yu94}.
Recall that we partition each phase into $2m$ blocks of size $b$. Thus, after $k$ phases we have a total of $2km$ blocks. Let $\mathbb{P}$ denote the joint distribution of state-action pairs in \emph{odd} blocks.
Let $\cI_i$ denote the set of indices in the $i^{th}$ block, and let $x_{\cI_i}, a_{\cI_i}$ denote the corresponding states and actions. We factorize the joint distribution according to blocks:
\begin{align*}
\mathbb{P}(x_{\cI_1},a_{\cI_1}, x_{\cI_3},a_{\cI_3},\ldots, x_{\cI_{2km-1}}, a_{\cI_{2km-1}})
=& \;\; \mathbb{P}_1(x_{\cI_1},a_{\cI_1}) \times \mathbb{P}_3(x_{\cI_3},a_{\cI_3} | x_{\cI_1},a_{\cI_1}) \times
\cdots \\
& \times
\mathbb{P}_{2km - 1} (x_{\cI_{2km-1}},a_{\cI_{2km-1}} | x_{\cI_{2km-3}},a_{\cI_{2km-3}}).
\end{align*}
Let $\tilde{\mathbb{P}}_i$ be the marginal distribution over the variables in block $i$, and let $\tilde \mathbb{P}$ be the product of marginals of odd blocks.
Corollary~2.7 of \citet{Yu94} implies that for any Borel-measurable set $E$,
\begin{equation}
\label{eq:indep-blocks}
|\mathbb{P}(E) - \widetilde\mathbb{P}(E)| \leq (km-1) \beta_b,
\end{equation}
where $\beta_b$ is the $b^{th}$ $\beta$-mixing coefficient of the process.
The result follows since the size of the ``gap'' between successive blocks is $b$; see Appendix~\ref{app:ib} for more details.
Recall that our estimates $\widehat w_{\pi_i}$ are based only on data in odd blocks in each phase. Let $\widetilde \mathbb{E}$ denote the expectation w.r.t. the product-of-marginals distribution $\tilde \mathbb{P}$. Then $\widetilde\mathbb{E}[\widehat M_i^{-1} \sum_{t\in \mathcal{H}_i} \phi_t z_t] =0$ because for $t\in \mathcal{H}_i$ and under $\widetilde \mathbb{P}$, $z_t$ is zero-mean given $\phi_t$ and is independent of other feature vectors outside of the block. Furthermore, by Hoeffding's inequality $\tilde{\mathbb{P}}(|z_t|/b \geq a) \leq 2 \exp(-2ba^2)$.
Since $\norm{\phi_t}_2 \leq C_\Phi$ and $\norm{\widehat M_i^{-1}}_2 \leq 2 \sigma^{-2}$ for large enough $m$, we have that
\[
\tilde{\mathbb{P}}(\norm{\widehat M_i^{-1}\phi_t z_t}_2 \geq 2b \sigma^{-2} C_{\Phi} a) \leq 2\exp(-2 b a^2).
\]
Since $\widehat{M}_i^{-1}\phi_t z_t$ are norm-subGaussian vectors, using Lemma~\ref{lemma:nsg}, there exists a constant $c_1$ such that for any $\delta \geq 0$
\begin{align*}
\tilde{\mathbb{P}}\left( \left\| \sum_{i=1}^k \widehat M_i^{-1} \sum_{t\in \mathcal{H}_i} \phi_t z_t \right\|_2 \geq 2 c_1 C_\Phi \sigma^{-2} \sqrt{b km \log(2d/\delta)} \right) \leq \delta \,.
\end{align*}
Thus, using \eqref{eq:indep-blocks},
\[
\mathbb{P} \left(\left\| \sum_{i=1}^k \widehat M_i^{-1} \sum_{t\in \mathcal{H}_i} \phi_t z_t \right\|_2 \ge 2 c_1 C_\Phi \sigma^{-2} \sqrt{b km \log(2d/\delta)} \right) \le \delta + (km-1) \beta_b \;.
\]
Under Assumption~\ref{ass:mixing}, we have that $\beta_b \leq 2 \gamma^b (1 - \gamma)^{-1}$. Setting $\delta =2 km \gamma^b(1 - \gamma)^{-1}$ and solving for $b$ we get
\begin{align}\label{eq:b_condi_2}
b = \frac{\log ( 2km \delta^{-1}(1-\gamma)^{-1})}{\log (1/\gamma)}.
\end{align}
Notice that when $b$ is chosen as in Eq.~\eqref{eq:b_condi_2}, the condition~\eqref{eq:b_condi} is also satisfied.
Plugging this into the previous display gives that with probability at least $1-2\delta$,
\[
\left\| \sum_{i=1}^k \widehat M_i^{-1} \sum_{t\in \mathcal{H}_i} \phi_t z_t \right\|_2 \le 2 c_1 C_\Phi \sigma^{-2} \sqrt{b km \log(2d/\delta)}.
\]
\subsection{Bounding $\| \sum_{i=1}^k \widehat M_i^{-1} \sum_{t\in \mathcal{H}_i} \phi_t ( J_{\pi_i} - \widehat J_{\pi_i}) \|_2$}
\label{app:jbias}
Recall that the average-reward estimates $\widehat J_{\pi_i}$ are computed using time indices corresponding to the starts of even blocks, $\mathcal{T}_i$. Thus this error term is only a function of the indices corresponding to block starts.
Now let $\mathbb{P}$ denote the distribution over state-action pairs $(x_t, a_t)$ for indices $t$ corresponding to block starts, i.e. $t \in \{1, b+1, 2b+1, ..., (2km -1)b+1\}$. We again factorize the distribution over blocks as $\mathbb{P} = \mathbb{P}_1 \otimes \mathbb{P}_2 \otimes \cdots \otimes \mathbb{P}_{2km} $.
Let $\tilde{\mathbb{P}} = \tilde{\mathbb{P}}_1 \otimes \tilde{\mathbb{P}}_2 \otimes \cdots \otimes \tilde{\mathbb{P}}_{2km} $ be a product-of-marginals distribution defined as follows. For odd $j$, let $\tilde \mathbb{P}_j$ be the marginal of $\mathbb{P}$ over $(x_{jb+1}, a_{jb+1})$. For even $j$ in phase $i$, let $\tilde \mathbb{P}_j = \nu_{\pi_i}$ correspond to the stationary distribution of the corresponding policy $\pi_i$. Using arguments similar to independent blocks, we show in Appendix~\ref{app:ib} that
\[\norm{\mathbb{P} - \tilde{\mathbb{P}}}_1 \leq 2(2km - 1)\gamma^{b-1}.
\]
Let $\widetilde \mathbb{E}$ denote expectation w.r.t. the product-of-marginals distribution $\tilde \mathbb{P}$.
Then $\widetilde\mathbb{E} [\widehat M_i^{-1} \sum_{t\in \mathcal{H}^i} \phi_t (J_{\pi_i} - \widehat J_{\pi_i})] =0$, since under $\tilde \mathbb{P}$, $\widehat J_{\pi_i}$ is the sum of rewards for state-action pairs distributed according to $\nu_{\pi_i}$, and these state-action pairs are independent of other data.
Using a similar argument as in the previous section, for $b=1+\frac{\log(4km/\delta)}{\log(1/\gamma)}$, there exists a constant $c_2$ such that with probability at least $1-2\delta$,
\[
\left\| \sum_{i=1}^k \widehat M_i^{-1} \sum_{t\in \mathcal{H}_i} \phi_t (J_{\pi_i} - \widehat J_{\pi_i}) \right\|_2 \le 2 c_2 C_\Phi \sigma^{-2} \sqrt{km \log(2d/\delta)} \;.
\]
\subsection{Bounding $\|\widehat M_i^{-1}\|_2$}
\label{app:m_bound}
In this subsection, we show that with probability at least $1-\delta$, for $m \geq 72 C_{\Phi}^4 \sigma^{-2} (1 -\gamma)^{-2} \log (d/\delta))$, $\norm{M_i^{-1}}_2 \leq 2\sigma^{-2}$.
Let $\Phi$ be a $|\mathcal{X}||\mathcal{A}| \times d$ matrix of all features. Let $D_i = {\rm diag}(\nu_{\pi_i})$, and let $\widehat D_i = {\rm diag}(\sum_{t \in \mathcal{H}_i} X_t )$, where $X_t$ is a state-action indicator as in Lemma~\ref{lem:doob}. Let $M_i = \Phi^\top D_i \Phi + \alpha m^{-1} I$.
We can write $\widehat M_i^{-1}$ as
\begin{align*}
\widehat M_i^{-1} &= ( \Phi^\top \widehat D_i \Phi + {\alpha}{\tau}^{-1}I + \Phi^\top (D_i - D_i) \Phi)^{-1} \\
&= (M_i + \Phi^\top (D_i - D_i) \Phi)^{-1} \\
&= (I + M_i^{-1}\Phi^\top (D_i - D_i) \Phi)^{-1} M_i^{-1}
\end{align*}
By Assumption~\ref{ass:excite} and \ref{ass:linear}, $\norm{M_i^{-1}}_2 \leq \sigma^{-2}$. In Appendix~\ref{app:matrix_azuma}, we show that w.p. at least $1-\delta$,
\begin{align*}
\norm{\Phi^\top (\widehat D_i - D_i) \Phi}_2
\leq 6 m^{-1/2} C_{\Phi}^2 (1 - \gamma)^{-1} \sqrt{2\log (d /\delta)}
\end{align*}
Thus
\begin{align*}
\norm{\widehat M_i^{-1}}_2 \leq
\sigma^{-2} (1 - \sigma^{-2} 6 m^{-1/2} C_{\Phi}^2 (1 - \gamma)^{-1} \sqrt{2\log (d /\delta)})^{-1}
\end{align*}
For $m \geq 72 C_{\Phi}^4 \sigma^{-2} (1 -\gamma)^{-2} \log (d/\delta))$, the above norm is upper-bounded by $\norm{\widehat M_i^{-1}}_2 \leq 2 \sigma^{-2}$.
\subsection{Bounding $\|\Phi^\top (\widehat D_i - D_i) \Phi^\top\|_2$}
\label{app:matrix_azuma}
For any matrix $A$,
\begin{align}
\label{eq:bounded_matrix}
\norm{\Phi^\top A \Phi}_2 = \bigg\| \sum_{ij} A_{ij} \phi_i \phi_j^\top \bigg\|_2 \leq \sum_{i, j} |A_{ij}| \norm{\phi_i \phi_j^\top}_2 \leq C_\Phi^2 \sum_{i, j} |A_{ij}| = C_\Phi^2 \norm{A}_{1, 1} \,.
\end{align}
where $\norm{A}_{1, 1}$ denotes the sum of absolute entries of $A$. Using the same notation for $X_t$ as in Lemma~\ref{lem:doob},
\begin{align*}
\norm{\Phi^\top (\widehat D_i - D_i) \Phi}_2
& = \frac{1}{m} \sum_{t \in \mathcal{H}_i} \Phi^\top {\rm diag}(X_t - \nu_t + \nu_t -\nu_{\pi_i}) \Phi \\
&\leq \frac{1}{m} \bigg\| \sum_{t\in \mathcal{H}_i} \Phi^\top {\rm diag}(X_t - \nu_t) \Phi\bigg\|_2 + \frac{C_\Psi^2}{m}\sum_{t\in \mathcal{H}_i} \norm{\nu_t - \nu_{\pi_i}}_1 \,.
\end{align*}
Under the fast-mixing assumption~\ref{ass:mixing}, the second term is bounded by $2 C_{\Psi}^2 m^{-1} (1 - \gamma)^{-1}$.
For the first term, we can define a martingale $(B_i)_{i=0}^{m}$ similar to the Doob martingale in Lemma~\ref{lem:doob}, but defined only on the $m$ indices $\mathcal{H}_i$.
Note that $ \sum_{t \in \mathcal{H}_i} \Phi^\top{\rm diag}(X_t - \nu_t) \Phi = \Phi^\top {\rm diag} (B_{m} - B_{0})\Phi $. Thus we can use matrix-Azuma to bound the difference sequence. Given that
\begin{align*}
\norm{ (\Phi^\top (B_i - B_{i-1}) \Phi)^2}_2 \leq 4 C_{\Phi}^4 (1 - \gamma)^{-2},
\end{align*}
combining the two terms, we have that with probability at least $1 - \delta$,
\begin{align*}
\norm{\Phi^\top (\widehat D_i - D_i) \Phi}_2 &\leq 4 m^{-1/2} C_{\Phi}^2 (1 - \gamma)^{-1} \sqrt{2\log (d /\delta)} + 2 m^{-1} C_{\Phi}^2 (1 - \gamma)^{-1} \\
& \leq 6 m^{-1/2} C_{\Phi}^2 (1 - \gamma)^{-1} \sqrt{2\log (d /\delta)} \,.
\end{align*}
\section{Bounding $\mathbb{E}_{x \sim \mu*}[\widehat V_K(x) - V_K(x)]$}
\label{app:linear_vpi}
We write the value function error as follows:
\begin{align*}
\mathbb{E}_{x \sim \mu*}[\widehat V_K(x) - V_K(x)]
&= \sum_x \mu_*(x) \sum_a \phi(x, a)^\top \frac{1}{K} \sum_{i=1}^K \pi_i(a|x) (\widehat w_{\pi_i} - w_{\pi_i}) \\
&\leq \frac{1}{K} \sum_x \mu_*(x) \sum_a \norm{ \phi(x, a)}_2 \left \| \sum_{i=1}^K \pi_i(a|x) (\widehat w_{\pi_i} - w_{\pi_i}) \right\|_2
\end{align*}
Note that for any set of scalars $\{p_i\}_{i=1}^{K}$ with $p_i \in [0, 1]$, the term $\left \| \sum_{i=1}^K p_i (\widehat w_{\pi_i} - w_{\pi_i}) \right\|_2$ has the same upper bound as $\norm{\sum_{i=1}^K (\widehat w_{\pi_i} - w_{\pi_i})}_2$. The reason is as follows. One part of the error includes bias terms \eqref{eq:bias_alpha} and \eqref{eq:bias_truncation}, whose upper bounds are only smaller when reweighted by scalars in $[0, 1]$. Thus we can simply upper-bound the bias by setting all $\{p_i\}_{i=1}^{K}$ to 1. Another part of the error, analyzed in Appendices \ref{app:zt} and \ref{app:jbias} involves sums of norm-subGaussian vectors. In this case, applying the weights only results in these vectors potentially having smaller norm bounds. We keep the same bounds for simplicity, again corresponding to all $\{p_i\}_{i=1}^{K}$ equal to 1. Thus, reusing the results of the previous section, we have
\begin{align*}
\mathbb{E}_{x \sim \mu*}[\widehat V_K(x) - V_K(x)]
&\leq C_{\Phi} |\mathcal{A}| c \sigma^{-2}(C_w + C_{\Phi})b \sqrt{\frac{ \log(2d/\delta)}{Km}}.
\end{align*}
\section{Independent Blocks}
\label{app:ib}
{\bf Blocks.} Recall that we partition each phase into $2m$ blocks of size $b$. Thus, after $k$ phases we have a total of $2km$ blocks. Let $\mathbb{P}$ denote the joint distribution of state-action pairs in odd blocks.
Let $\cI_i$ denote the set of indices in the $i^{th}$ block, and let $x_{\cI_i}, a_{\cI_i}$ denote the corresponding states and actions. We factorize the joint distribution according to blocks:
\begin{align*}
\mathbb{P}(x_{\cI_1},a_{\cI_1}, x_{\cI_3},a_{\cI_3},\ldots, x_{\cI_{2km-1}}, a_{\cI_{2km-1}})
=&
\;\; \mathbb{P}_1(x_{\cI_1},a_{\cI_1}) \times \mathbb{P}_3(x_{\cI_3},a_{\cI_3} | x_{\cI_1},a_{\cI_1}) \times
\cdots \\
& \times
\mathbb{P}_{2km - 1} (x_{\cI_{2km-1}},a_{\cI_{2km-1}} | x_{\cI_{2km-3}},a_{\cI_{2km-3}}).
\end{align*}
Let $\tilde{\mathbb{P}}_i$ be the marginal distribution over the variables in block $i$, and let $\tilde \mathbb{P}$ be the product of marginals. Then the difference between the distributions $\tilde{\mathbb{P}}$ and $\mathbb{P}$ can be written as
\begin{align*}
\mathbb{P} - \tilde \mathbb{P} =&
\;\; {\mathbb{P}_1} \otimes \mathbb{P}_3 \otimes \cdots \otimes \mathbb{P}_{2km-1} - {\mathbb{P}_1} \otimes \tilde{\mathbb{P}}_3 \cdots \otimes \tilde{\mathbb{P}}_{2km-1} \\
=&
\;\; {\mathbb{P}_1} \otimes (\mathbb{P}_3 - \tilde{\mathbb{P}}_3) \otimes \mathbb{P}_5 \otimes \cdots \otimes\mathbb{P}_{2km-1}\\
& + \mathbb{P}_1 \otimes \tilde \mathbb{P}_3 \otimes(\mathbb{P}_5 - \tilde{\mathbb{P}}_5) \otimes\mathbb{P}_7 \otimes\ldots \otimes\mathbb{P}_{2km-1} \\
& + \cdots \\
& +{\mathbb{P}}_1 \otimes \tilde{\mathbb{P}}_3 \otimes \tilde{\mathbb{P}}_5 \otimes \cdots \otimes \tilde{\mathbb{P}}_{2km-3} \otimes (\mathbb{P}_{2km-1} - \tilde{\mathbb{P}}_{2km-1}).
\end{align*}
Under $\beta$-mixing, since the gap between the blocks is of size $b$, we have that
\begin{align*}
\norm{\mathbb{P}_i(x_{\cI_i},a_{\cI_i} | x_{\cI_{i-2}},a_{\cI_{i-2}}) - \tilde{\mathbb{P}}_i(x_{\cI_i},a_{\cI_i})}_1 \leq \beta_b = \frac{2 \gamma^b}{1-\gamma}.
\end{align*}
Thus the difference between the joint distribution and the product of marginals is bounded as
\begin{align*}
\norm{\mathbb{P} -\tilde{\mathbb{P}}}_1 \leq (km - 1) \beta_b.
\end{align*}
{\bf Block starts.} Now let $\mathbb{P}$ denote the distribution over state-action pairs $(x_t, a_t)$ for indices $t$ corresponding to block starts, i.e. $t \in \{1, b+1, 2b+1, ..., (2km -1)b+1\}$. We again factorize the distribution over blocks:
\begin{align*}
\mathbb{P}(x_1,a_1, x_{b+1}, a_{b+1}, \ldots, x_{(2km-1)b+1}, a_{(2km-1)b+1}) = \mathbb{P}_1(x_1, a_1) \prod_{j=2}^{2km} \mathbb{P}_i(x_{jb + 1}, a_{jb+1} | x_{(j-1)b+1}, a_{(j-1)b+1}).
\end{align*}
Define a product-of-marginals distribution $\tilde \mathbb{P} = \tilde{\mathbb{P}}_1\otimes \tilde{\mathbb{P}}_2 \otimes \cdots \otimes \tilde{\mathbb{P}}_{2km}$ over the block-start variables as follows. For odd $j$, let $\tilde \mathbb{P}_j$
be the marginal of $\mathbb{P}$ over $(x_{jb+1}, a_{jb+1})$. For even $j$ in phase $i$, let $\tilde \mathbb{P}_j = \nu_{\pi_i}$ correspond to the stationary distribution of the policy $\pi_i$.
Using the same notation as in Appendix~\ref{app:prelim},
let $X_t$ be the indicator vector for $(x_t, a_t)$
and let $H_{i:j}$ be the product of state-action transition matrices at times $i+1,..., j$. For odd blocks $j$, we have
\begin{align*}
\norm{ \mathbb{P}_j(\cdot| x_{(j-1)b+1}, a_{(j-1)b+1}) - \tilde \mathbb{P}_j(\cdot)}_1 = \norm{ H_{(j-1)b+1:jb}^\top (X_{(j-1)b+1} - \tilde{\mathbb{P}}_{j-1})}_1 \leq 2 \gamma^{b-1} \,.
\end{align*}
Slightly abusing notation, let $H_{\pi_i}$ be the state-action transition matrix under policy $\pi_i$. For even blocks $j$ in phase $i$, since they always follow an odd block in the same phase,
\begin{align*}
\norm{ \mathbb{P}_j(\cdot| x_{(j-1)b+1}, a_{(j-1)b+1}) - \tilde \mathbb{P}_j(\cdot)}_1 = \norm{ (H_{\pi_i}^{b-1})^\top( X_{(j-b)+1} - \nu_{\pi_i})}_1 \leq 2 \gamma^{b-1} \,.
\end{align*}
Thus, using a similar distribution decomposition as before, we have that $\norm{\mathbb{P} - \tilde{\mathbb{P}}}_1 \leq 2(2km-1) \gamma^{b-1}$.
\section{Related Work}
\textbf{Regret bounds for average-reward MDPs.}
Most no-regret algorithms for infinite-horizon undiscounted MDPs are only applicable to tabular representations and model-based
\citep{bartlett09regal,jaksch2010near,ouyang2017learning
fruit2018efficient,jian2019exploration,talebi2018variance}.
For weakly-communicating MDPs with diameter $D$, these algorithms nearly achieve the minimax lower bound $\Omega(\sqrt{D|\mathcal{X}||\mathcal{A}|T})$ \citep{jaksch2010near} with high probability.
\citet{wei2020model} provide model-free algorithms with regret bounds in the tabular setting.
In the model-free setting with function approximation,
the {\textsc{Politex}} algorithm \citep{politex} achieve $O(d^{1/2}T^{3/4})$ regret in uniformly ergodic MDPs, where $d$ is the size of the compressed state-action space. \citet{hao2020provably} improve these results to $O(T^{2/3})$.
More recently, \citet{wei2020learning} present three algorithms for average-reward MDPs with linear function approximation. Among these, FOPO achieves $O(\sqrt{T})$ regret but is computationally inefficient, and OLSVI.FH is efficient but obtains $O(T^{3/4})$ regret. The MDP-EXP2 algorithm is computationally efficient, and under similar assumptions as in \citet{politex} it obtains a $O(\sqrt{T})$ bound on \emph{expected regret} (a weaker guarantee than the high-probability bounds in other works). Our analysis shows a high-probability $O(\sqrt{T})$ regret bound under the same assumptions.
{\bf KL-regularized approximate policy iteration.} Our work is also related to approximate policy iteration algorithms which constrain each policy to be close to the previous policy in the sense of KL divergence. This approach was popularized by TRPO \citep{schulman2015trust}, where it was motivated as an approximate implementation of conservative policy iteration \citep{kakade2002approximately}.
Some of the related subsequent works include PPO \citep{schulman2017proximal}, MPO \citep{abdolmaleki2018maximum}, V-MPO \citep{vmpo}, and CPO \citep{achiam2017constrained}. While these algorithms place a constraint on consecutive policies and are mostly heuristic, another line of research shows that using KL divergence as a regularizer has a theoretical justification in terms of either regret guarantees \citep{politex,hao2020provably,wei2020model,wei2020learning} or error propagation \citep{vieillard2020leverage,vieillard2020momentum}.
{\bf Experience replay.} Experience replay \citep{lin1992self} is one of the central tools for achieving good performance in deep reinforcement learning. While it is mostly used in off-policy methods such as deep Q-learning \citep{mnih2013playing,mnih2015human}, it has also shown benefits in value-gradient methods \citep{heess2015learning,lillicrap2016continuous}, and has been used in some variants of KL-regularized policy iteration \citep{abdolmaleki2018maximum,tomar2020mirror}. Its success has been attributed to removing some temporal correlations from data fed to standard gradient-based optimization algorithms. \citet{schaul2015prioritized} have shown that non-uniform replay sampling based on the Bellman error can improve performance. Unlike these works, we motivate experience replay from the perspective of online learning in MDPs \citep{even2009online} with the goal of approximating the average of past value functions well.
{\bf Continual learning.} Continual learning (CL) is the paradigm of learning a classifier or regressor that performs well on a set of tasks, where each task corresponds to a different data distribution. The tasks are observed sequentially, and the learning goal is to avoid forgetting past tasks without storing all the data in memory. This is quite similar to our goal of approximating the average of sequentially-observed Q-functions, where the data for approximating each Q-function has different distribution.
In general, approaches to CL can be categorized as regularization-based \citep{kirkpatrick2017overcoming,zenke2017continual,farajtabar2020orthogonal,yin2020sola}, expansion-based \citep{rusu2016progressive}, and replay-based \citep{lopez2017gradient,chaudhry2018efficient,borsos2020coresets}, with the approaches based on experience replay typically having superior performance over other methods.
\section{Introduction}
Model-free reinforcement learning (RL) algorithms combined with powerful function approximation have achieved impressive performance in a variety of application domains over the last decade. Unfortunately, the theoretical understanding of such methods is still quite limited.
In this work, we study single-trajectory learning in infinite-horizon undiscounted Markov decision processes (MDPs), also known as average-reward MDPs, which capture tasks such as routing and the control of physical systems.
One line of works with performance guarantees for the average-reward setting follows the ``online MDP'' approach proposed by \citet{even2009online}, where the agent selects policies by running an online learning algorithm in each state, typically mirror descent. The resulting algorithm is a version of approximate policy iteration (API), which alternates between (1) estimating the state-action value function (or Q-function) of the current policy and (2) setting the next policy to be optimal w.r.t. the sum of all previous Q-functions plus a regularizer. Note that, by contrast, standard API sets the next policy only based on the most recent Q-function.
The policy update can also be written as maximizing the most recent Q-function minus KL-divergence to the previous policy, which is somewhat similar to recently popular versions of API \citep{schulman2015trust,schulman2017proximal,achiam2017constrained,abdolmaleki2018maximum,vmpo}.
The original work of \citet{even2009online} studied this scheme with known dynamics, tabular representation, and adversarial reward functions. More recent works \citep{politex,hao2020provably,wei2020learning} have adapted the approach to the case of unknown dynamics, stochastic rewards, and value function approximation. With linear value functions, the {\textsc{Politex}} algorithm of \citet{politex} achieves $O(T^{3/4})$ high-probability regret in ergodic MDPs, and the results only scale in the number of features rather than states. \citet{wei2020learning} later show an $O(\sqrt{T})$ bound on \emph{expected} regret for a similar algorithm named MDP-EXP2. In this work, we revisit the analysis of {\textsc{Politex}} and show that it can be sharpened to $O(\sqrt{T})$ under nearly identical assumptions, resulting in the first $O(\sqrt{T})$ \emph{high-probability} regret bound for a computationally efficient algorithm in this setting.
In addition to improved analysis, our work also addresses practical implementation of {\textsc{Politex}} with neural networks. The policies produced by {\textsc{Politex}} in each iteration require access to the sum of all previous Q-function estimates. With neural network function approximation, exact implementation requires us to keep all past networks in memory and evaluate them at each step, which is inefficient in terms of memory and computation. Some practical implementation choices include subsampling Q-functions and/or optimizing a KL-divergence regularized objective w.r.t. a parametric policy at each iteration. We propose an alternative approach, where we approximate the average of all past Q-functions by training a single network on a replay buffer with past data. We demonstrate that this choice often outperforms other approximate implementations, especially in terms of run-time. When available memory is constrained, we propose to subsample transitions using the notion of coresets \citep{bachem2017practical}.
Our work also provides a novel perspective on the benefits of experience replay.
Experience replay is a standard tool for stabilizing learning in modern deep RL, and typically used
in \emph{off-policy} methods like Deep Q-Networks \citep{mnih2013playing}, as well as ``value gradient'' methods such as DDPG \citep{lillicrap2016continuous} and SVG \citep{heess2015learning}.
A different line of \emph{on-policy} methods typically does not rely on experience replay; instead, learning is stabilized by constraining consecutive policies to be close in terms of KL divergence \citep{schulman2015trust,vmpo,degrave2019quinoa}. We observe that both experience replay and KL-divergence regularization can be viewed as approximate implementations of \textsc{Politex}.
Thus, we provide a theoretical justification for using experience replay in API, as an approximate implementation of online learning in each state. Note that this online-learning view differs from the commonly used justifications for experience replay, namely that it ``breaks temporal correlations'' \citep{schaul2015prioritized, mnih2013playing}. Our analysis also suggests a new objective for subsampling or priority-sampling transitions in the replay buffer, which differs priority-sampling objectives of previous work \citep{schaul2015prioritized}.
In summary, our main contributions are (1) an improved analysis of \textsc{Politex}, showing an $O(\sqrt{T})$ regret bound under the same assumptions as the original work, and (2) an efficient implementation that also offers a new perspective on the benefits of experience replay.
\section{Regret Analysis of {\textsc{Politex}} }\label{sec:regret}
In this section, we revisit the regret analysis of {\textsc{Politex}} \citep{politex} in ergodic average-reward MDPs, and show that it can be improved from $O(T^{3/4})$ to $O(\sqrt{T})$ under similar assumptions. Our analysis relies in part on a simple modification of the regret decomposition. Namely, instead of including the estimation error of each value-function $Q_{\pi_k}$ in the regret, we consider the error in the running average $Q_k$. When this error scales as $O(1/\sqrt{k})$ in a particular weighted norm, the regret of {\textsc{Politex}} is $O(\sqrt{T})$. As we show in the following section, this bound can be instantiated for linear value functions under the same assumptions as \citet{politex}.
\begin{assumption}[Boundedness]
\label{ass:bounded_diff}
Let $\widehat Q_{\pi_k} := k \widehat Q_k - (k-1) \widehat Q_{k-1}$. We assume that there exists a constant $Q_{\max}$ such that for all $k=1, ..., K$ and for all $x \in \mathcal{X}$,
\[
\max_{a} \widehat Q_{\pi_k}(x, a) - \min_{a} \widehat Q_{\pi_k}(x, a) \leq Q_{\max}.
\]
\end{assumption}
For the purpose of our algorithm, functions $\widehat Q_{\pi_k}$ are unique up to a constant. Thus we can equivalently assume that $\norm{\widehat Q_{\pi_k}(x, \cdot)}_\infty \leq Q_{\max}$ for all $x$.
Define $\widehat V_{\pi_k}(x) := \widehat Q_{\pi_k}(x, \pi_k)$. Let $V_k := \frac{1}{k} \sum_{i=1}^k V_{\pi_i}(x)$ and $\widehat V_k(x) := \sum_{i=1}^k \widehat V_{\pi_i}(x)$ be the average of the state-value functions and its estimate. We will require the estimation error of the running average to scale as in the following assumption.
\begin{assumption}[Estimation error]
\label{ass:est_err}
Let $\mu_*$ be the stationary state distribution of the optimal policy $\pi_*$.
With probability at least $1-\delta$, for a problem-dependent constant $C$, the errors in $\widehat Q_K$ and $\widehat V_K$ are bounded as
\begin{align*}
\mathbb{E}_{x \sim \mu_*} [ \widehat V_K(x) - V_K(x)]
&\leq C \sqrt{\log(1/\delta) / K} \\
\mathbb{E}_{x \sim \mu_* }[ Q_K(x, \pi_*) - \widehat Q_K(x, \pi_*)]
&\leq C \sqrt{\log(1/\delta) / K} \,.
\end{align*}
\end{assumption}
Define $S_\delta(|\mathcal{A}|, \mu_*)$ as in \citet{politex}:
\begin{align*}
S_\delta(|\mathcal{A}|, \mu_*) := \sqrt{\frac{\log|\mathcal{A}|}{2}} + \big\langle \mu_*, \sqrt{\frac{1}{2}\log\frac{1}{\delta\mu_*}}\big\rangle .
\end{align*}
We bound the regret of {\textsc{Politex}} in the following theorem. Here, recall that $\tau$ is the length of each phase, and $\gamma$ and $\eta$ are defined in Assumption~\ref{ass:mixing} and Eq.~\eqref{eq:kl_politex}, respectively.
\begin{theorem}[Regret of \textsc{Politex}]
\label{thm:regret}
Let Assumptions \ref{ass:mixing}, \ref{ass:est_err}, and \ref{ass:bounded_diff} hold.
For $\tau \geq \frac{\log T}{2 \log(1/\gamma)}$ and $\eta = \frac{\sqrt{8\log|\mathcal{A}|}}{Q_{\max}\sqrt{K}}$, for a constant $C_1$, with probability at least $1-4\delta$, the regret of {\textsc{Politex}} in ergodic average-reward MDPs is bounded as
\begin{align*}
R_T \leq \frac{C_1 (1+ Q_{\max}) S_\delta(|\mathcal{A}|, \mu_*)\sqrt{\tau}}{(1-\gamma)^{2}} \sqrt{T}\,.
\end{align*}
\end{theorem}
\begin{proof}
We start by decomposing the cumulative regret, following similar steps as \citet{politex}:
\begin{equation}\label{eqn:regret_dec}
R_T = \sum_{k=1}^K \sum_{t=(k-1)\tau+1}^{k\tau} (J_* - J_{\pi_k}) + (J_{\pi_k} - r_t).
\end{equation}
The second term $V_T = \sum_{k=1}^K \sum_{t=(k-1)\tau+1}^{k\tau}(J_{\pi_k} -r_t)$ captures the sum of differences between observed rewards and their long term averages. In previous work, this term was shown to scale as $O(K\sqrt{\tau})$. We show that the analysis can in fact be tightened to $O(\sqrt{T})$ using improved concentration bounds and the slow-changing nature of the policies. See Lemma~\ref{lem:vtbound} in Appendix~\ref{sec:vtbound} for precise details.
The first term, which is also called \emph{pseudo-regret} in literature, measures the difference between the expected reward of the reference policy and the policies produced by the algorithm.
Applying the performance difference lemma~\citep{cao1999single}, we can write each pseudo-regret term as
\begin{align*}
J_{*} -J_{\pi_k} &= \mathbb{E}_{x \sim \mu_{*}} \left[ Q_{\pi_{k}}(x, \pi_{*})-Q_{\pi_{k}}(x, \pi_k) \right].
\end{align*}
Now, notice that {\textsc{Politex}} is running exact mirror descent for loss functions $\widehat{Q}_{\pi_k}$. We bridge the pseudo-regret by the $\widehat{Q}_{\pi_k}$ terms:
\begin{align}
R_{T1a}&= \tau\sum_{k=1}^K \mathbb{E}_{x \sim \mu_{*}} \left[ Q_{\pi_{k}}(x, \pi_{*})- \widehat{Q}_{\pi_k}(x, \pi_*) \right] \label{eq:rt1a} \\
R_{T1b} &= \tau\sum_{k=1}^K \mathbb{E}_{x \sim \mu_{*}} \left[ \widehat{Q}_{\pi_k}(x, \pi_k)-Q_{\pi_{k}}(x, \pi_k) \right] \label{eq:rt1b} \\
R_{T2}&=
\tau\sum_{k=1}^K \mathbb{E}_{x \sim \mu_{*}} \left[ \widehat{Q}_{\pi_k}(x, \pi_*)-\widehat{Q}_{\pi_k}(x, \pi_k)\right]. \label{eq:rt2}
\end{align}
$R_{T2}$ can be bounded using the regret of mirror descent as in previous work \citep{politex}. Setting $\eta = \frac{\sqrt{ \log |\mathcal{A}| }}{Q_{\max}\tau\sqrt{2K}}$, and using a union bound over all states, with probability at least $1-\delta$, $R_{T2}$ is bounded as
\begin{align*}
R_{T2} \leq \tau Q_{\max} S_{\delta}(|\mathcal{A}|, \mu_*) \sqrt{K} \,.
\end{align*}
{\bf Bounding regret due to estimation error.} We now focus on bounding $R_{T1} = R_{T1a} + R_{T1b}$ under Assumption~\ref{ass:est_err}. We have the following:
\begin{align*}
R_{T1a} &= \tau \sum_{k=1}^K \mathbb{E}_{x \sim \mu_*} \big[ Q_{\pi_k}(x, \pi_*) - \widehat{Q}_{\pi_k}(x, \pi_*) \big] \\
&= (\tau K) \mathbb{E}_{x \sim \mu_*, a \sim \pi_*}\left[Q_K(x, a) - \widehat Q_K(x, a)\right] \\
& \leq C \tau \sqrt{K\log(1/\delta) }.\\
R_{T1b} &= \tau \sum_{k=1}^K \mathbb{E}_{x \sim \mu_*} \big[\widehat V_{\pi_k}(x) - V_{\pi_k}(x) \big] \\
&= \tau K \mathbb{E}_{x \sim \mu_*} \big[ \widehat V_K(x) - V_K(x) \big] \\
& \leq C \tau \sqrt{K\log(1/\delta)}.
\end{align*}
We can then obtain the final result by combining the bounds on $R_{T1a}$, $R_{T1b}$, $R_{T2}$, $V_T$ from Appendix~\ref{sec:vtbound}, and using union bound as well as the fact that $K = T / \tau$.
\end{proof}
\section{Linear Value Functions}\label{sec:linear}
In this section, we show that the estimation error condition in Assumption~\ref{ass:est_err} (and thus $O(\sqrt{T})$ regret) can be achieved under similar assumptions as in \citet{politex} and \citet{wei2020learning}, which we state next.
\begin{assumption}[Linear value functions]
\label{ass:linear}
The action-value function $Q_\pi$ of any policy $\pi$ is linear:
$Q_\pi(x, a) = w_\pi^\top \phi(x, a)$,
where $\phi: \mathcal{S} \times \mathcal{A} \rightarrow \mathbb{R}^d$ is a known feature function such that $\max_{x, a} \norm{\phi(x, a)} \leq C_\Phi$.
\end{assumption}
\begin{assumption}[Feature excitation]
\label{ass:excite}
There exists a constant $\sigma^2$ such that for any policy $\pi$,
\[
\lambda_{\min} \left( \mathbb{E}_{(x, a) \sim \mu_\pi \otimes \pi} [\phi(x, a) \phi(x, a)^\top] \right) \geq \sigma^2 > 0.
\]
\end{assumption}
We now describe a simple procedure for estimating the average action-value functions $Q_k(x, a) = \phi(x, a)^\top w_k$, where $w_k = \frac{1}{k} \sum_{i=1}^k w_{\pi_i}$, such that the conditions of Assumption~\ref{ass:est_err} are satisfied. Essentially, we estimate each $Q_{\pi_i}$ using least-squares Monte Carlo and then average the weights. We will use the shorthand notation $\phi_t = \phi(x_t, a_t)$. Let $\mathcal{H}_i$ and $\mathcal{T}_i$ be subsets of time indices in phase $i$ (defined later).
We estimate $Q_{\pi_i}$ as follows:
\begin{align}
\widehat w_{\pi_i}
&= \bigg(\sum_{t \in \mathcal{H}_i} \phi_t \phi_t^\top + \alpha I\bigg)^{-1} \sum_{t \in \mathcal{H}_i} \phi_t R_t
\end{align}
where $R_t$ are the empirical $b$-step returns ($b$ is specified later), computed as
\begin{align}\label{eq:linear_reward}
R_t = \sum_{j=t}^{t+b} (r_t - \widehat J_{\pi_i}), \quad
\widehat J_{\pi_i} = \frac{1}{|\mathcal{T}_i|} \sum_{t \in \mathcal{T}_i} r_t.
\end{align}
We then estimate $w_k$ as $\widehat w_k = \frac{1}{k} \sum_{i=1}^k \widehat w_{\pi_i}$. Note that for this special case of linear value functions, we do not need to use the replay buffer in Algorithm~\ref{alg:api}. For analysis purposes, we divide each phase of length $\tau$ into $2m$ blocks of size $b$ and let $\mathcal{H}_i$ ($\mathcal{T}_i$) be the starting indices of odd (even) blocks in phase $i$. Due to the gaps between indices and fast mixing, this makes the data almost independent (we make this precise in Appendix~\ref{app:linear}) and the error easier to analyze. In practice, one may simply want to use all data.
For a distribution $\mu$, let $\norm{x}_\mu$ denote the distribution-weighted norm such that $\norm{x}_\mu^2 = \sum_i \mu_i x_i^2$. Using Jensen's inequality, we have that $(\mathbb{E}_{x\sim \mu}[q(x)])^2 \leq \mathbb{E}_{x \sim \mu}[q(x)^2]$. Thus, it suffices to bound the $Q$-function error in the distribution-weighted norm, $\norm{Q_K - \widehat Q_K}_{\mu_* \otimes \pi_*}$. Furthermore, given bounded features,
\begin{align*}
\norm{\widehat Q_K - Q_K}_{\mu_* \otimes \pi_*}
&\leq C_{\Phi}\norm{\widehat w_K - w_K}_2 \,,
\end{align*}
so it suffices to bound the error in the weights.
We bound this error in the following Lemma, proven in Appendix~\ref{app:linear}.
\begin{lemma}[Estimation error for linear functions]
\label{lem:linear_est}
Suppose that Assumptions \ref{ass:mixing}, \ref{ass:linear} and \ref{ass:excite} hold and that true action-value weights are bounded as $\norm{w_{\pi_i}}_2 \leq C_w$ for all $i =1,\ldots K$.
Then for any policy $\pi$, for $\alpha = \sqrt{\tau/K}$, $m \geq 72 C_{\Phi}^4 \sigma^{-2} (1 -\gamma)^{-2} \log (d/\delta)$, and $b \geq \frac{\log (T\delta^{-1}(1-\gamma)^{-1} ) }{ \log(1/\gamma)}$, there exists an absolute constant $c$ such that with probability at least $1-\delta$,
\begin{align*}
\norm{\widehat w_K - w_K}_2 \leq c \sigma^{-2}(C_w + C_{\Phi})b \sqrt{\frac{ \log(2d/\delta)}{Km}}.
\end{align*}
\end{lemma}
Furthermore, in Appendix~\ref{app:linear_vpi}, we show that the error in the average state-value function satisfies the following:
\begin{align*}
\mathbb{E}_{\mu_*} [\widehat V_K(x) - V_K(x)] \leq
cC_{\Phi}|\mathcal{A}|(C_w + C_{\Phi})\frac{b}{\sigma^2} \sqrt{\frac{ \log(2d/\delta)}{Km}}.
\end{align*}
We have demonstrated that Assumption~\ref{ass:est_err} can be satisfied with linear value functions.
For Assumption~\ref{ass:bounded_diff}, it suffices for the weight estimates $\{\widehat w_{\pi_i}\}$ to be bounded. This will be true, since we assume that the true weights are bounded and we can bound the error in the weight space. Thus {\textsc{Politex}} has an $O(\sqrt{T})$ regret in this setting (though note that we incur an extra $|\mathcal{A}|$ factor coming from the $\widehat V_K$ error bound).
\section{Bounding the Difference Between Empirical and Average Rewards}
\label{sec:vtbound}
In this section, we bound the second term in Equation~\ref{eqn:regret_dec}, corresponding to the difference between empirical and average rewards.
\begin{lemma}
\label{lem:vtbound}
Let Assumption~\ref{ass:mixing} hold, and assume that $\tau \geq \frac{\log T}{2 \log (1/\gamma)}$ and that $r(x, a) \in [0, 1]$ for all $x, a$. Then, by choosing $\eta = \frac{\sqrt{8\log |\mathcal{A}|}}{Q_{\max} \sqrt{K}}$, we have with probability at least $1-\delta$,
\begin{align*}
\sum_{k=1}^K \sum_{t=(k-1)\tau+1}^{k\tau} (r_t - J_{\pi_k}) \leq 2(1 - \gamma)^{-1} \sqrt{2T \log(2/\delta)}
+ 2\sqrt{T} + (1 - \gamma)^{-2} \sqrt{8K\log|\mathcal{A}|} \,.
\end{align*}
\end{lemma}
\begin{proof}
Let $r$ denote the vector of rewards, and recall that $J_{\pi} = \nu_\pi^\top r$. Let $X_t$ be the indicator vector for the state-action pair at time $t$, as in Lemma~\ref{lem:doob}, and let $\nu_t = \mathbb{E}[X_t]$. We have the following:
\begin{align*}
V_T &:= \sum_{k=1}^K \sum_{t=(k-1)\tau+1}^{k\tau} (r_t - J_{\pi_k})
= \sum_{k=1}^K \sum_{t=(k-1)\tau+1}^{k\tau} r^\top (X_t - \nu_t + \nu_t - \nu_{\pi_k})
\end{align*}
We slightly abuse the notation above by letting $\nu_t$ denote the state-action distribution at time $t$, and $\nu_\pi$ the stationary distribution of policy $\pi$.
Let $\{B_i\}_{i=0}^T$ be the Doob martingale in Lemma~\ref{lem:doob}. Then $B_0 =\sum_{t=1}^T \nu_t$ and $B_T = \sum_{t=1}^T X_t$, and the first term can be expressed as
\begin{align*}
V_{T1} := \sum_{t=1}^T r^\top (X_t - \nu_t) = r^\top (B_T - B_0).
\end{align*}
By Lemma~\ref{lem:doob},
$ |\langle B_i - B_{i-1}, r\rangle| \leq \norm{B_i - B_{i-1}}_1\norm{r}_\infty \leq 2 (1 -\gamma)^{-1} $.
Hence by Azuma's inequality, with probability at least $1-\delta$,
\begin{equation}\label{eq:vt1}
V_{T1} \leq 2(1 - \gamma)^{-1} \sqrt{2T \log(2/{\delta})}.
\end{equation}
For the second term we have
\begin{align*}
V_{T2} & := \sum_{k=1}^K \sum_{t=(k-1)\tau+1}^{k\tau} r^\top (\nu_t - \nu_{\pi_k}) \\
& = \sum_{k=1}^K r^\top \bigg(\sum_{i=1}^\tau (H_{\pi_k}^i)^\top \nu_{(k-1)\tau} - \nu_{\pi_k}\bigg) \\
& \leq \sum_{k=1}^K \norm{r}_\infty \sum_{i=1}^\tau \left\| (H_{\pi_k}^i)^\top (\nu_{(k-1)\tau} - \nu_{\pi_{k-1}} + \nu_{\pi_{k-1}}) - \nu_{\pi_k} \right\|_1 \\
& \leq \sum_{k=1}^K \sum_{i=1}^\tau \norm{ \nu_{(k-1)\tau} - \nu_{\pi_{k-1}} }_1 + \norm{(H_{\pi_k}^i)^\top \nu_{\pi_{k-1}} - \nu_{\pi_k}}_1 \\
& \leq \sum_{k=1}^K \sum_{i=1}^\tau \norm{ (H_{\pi_{(k-1)}}^\tau)^\top \nu_{(k-2)\tau} - \nu_{\pi_{k-1}} }_1 + \gamma^{i} \norm{\nu_{\pi_{k-1}} - \nu_{\pi_k}}_1 \\
& \leq 2T \gamma^{\tau} + \frac{1}{1-\gamma} \sum_{k=1}^K \norm{\nu_{\pi_k} -\nu_{\pi_{k-1}}}_1 \,.
\end{align*}
For $\tau \geq \frac{\log T}{2 \log (1/\gamma)}$, the first term is upper-bounded by $2\sqrt{T}$.
Using results on perturbations of Markov chains \citep{seneta1988perturbation,cho2001comparison}, we have that
\begin{align*}
\norm{\nu_{\pi_k} - \nu_{\pi_{k-1}}}_1
\leq \frac{1}{1 - \gamma} \norm{H_{\pi_k} - H_{\pi_{k-1}}}_\infty \leq \frac{1}{1 - \gamma} \max_x \norm{\pi_{k}(\cdot|x) - \pi_{k-1}(\cdot|x)}_1.
\end{align*}
Note that the policies $\pi_k(\cdot|x)$ are generated by running mirror descent on reward functions $\widehat{Q}_{\pi_k}(x, \cdot)$.
A well-known property of mirror descent updates with entropy regularization (or equivalently, the exponentially-weighted-average algorithm) is that the difference between consecutive policies is bounded as
\begin{align*}
\norm{\pi_{k+1}(\cdot|x) - \pi_k(\cdot|x)}_1 \leq \eta \norm{\widehat{Q}_{\pi_k}(x,\cdot)}_\infty \,.
\end{align*}
See e.g. \citet{NeGySzA13} Section V.A for a proof, which involves applying Pinsker's inequality and Hoeffding's lemma (\citet{cesa2006prediction}
Section A.2 and Lemma A.6).
Since we assume that $\norm{\widehat{Q}_{\pi_k}}_\infty \leq Q_{\max}$, we can obtain
\[
V_{T2} \leq 2\sqrt{T} + (1 - \gamma)^{-2} K \eta Q_{\max}.
\]
By choosing $\eta = \frac{\sqrt{8\log |\mathcal{A}|}}{Q_{\max} \sqrt{K}}$, we can bound the second term as
\begin{equation}\label{eq:vt2}
V_{T2} \leq 2\sqrt{T} + (1 - \gamma)^{-2} \sqrt{8K\log|\mathcal{A}| }.
\end{equation}
Putting Eq.~\eqref{eq:vt1} and~\eqref{eq:vt2} together, we obtain that with probability at least $1-\delta$, \[
V_T \leq 2(1 - \gamma)^{-1} \sqrt{2T \log(2/\delta)}
+ 2\sqrt{T} + (1 - \gamma)^{-2} \sqrt{8K\log|\mathcal{A}|} \,.
\]
\end{proof}
\section{Practical Implementation}\label{sec:implementation}
As mentioned earlier, the key idea in our policy update is to obtain an estimate $\widehat{Q}_k$ of the average of all the Q-functions in previous phases. We have seen that when we use linear functions to approximate the Q-functions, we can simply average the weights in order to get an estimate of the average Q-function. However, in practice, we often need to use non-linear functions, especially neural networks, to approximate complex Q-functions. In this section, we discuss how to efficiently implement our algorithm with non-linear function approximation.
\subsection{Weight Averaging}\label{sec:weight_ave}
The simplest idea may be averaging the weights of neural networks. However, a crucial difference from the linear setting is that averaging the weights of the neural networks is not equivalent to averaging the functions that they represent: the function that a neural network represent is invariant to the permutation of the hidden units, and thus two networks with very different weights can represent similar functions. Therefore, this implementation may only succeed when all the Q-function approximations are around the same local region in the weight space. Thus, when we learn the new Q-function $\widehat Q_{\pi_k}$ in phase $k$, we should initialize it with $\widehat Q_{k-1}$, run SGD with new data, and then average with $\widehat Q_{k-1}$.
\subsection{Experience Replay}\label{sec:replay}
Another natural idea is to leverage the data from replay buffer to obtain an estimate of the average Q-function. We elaborate the details below. We use simple $b$-step Monte Carlo estimate for the $Q$ value of each state-action pair. For any $(i-1)\tau + 1 \le t \le i\tau - b $, we can estimate the state-action value of $(x_t, a_t)$ by $b$-step cumulative reward\footnote{This is a practical implementation of Eq.~\eqref{eq:linear_reward}, i.e., we do not split the data in each phase into blocks.}
\[
R_t = \sum_{j = t}^{t+b} (r_j - \widehat J_{\pi_i}), \quad \widehat J_{\pi_i} = \frac{1}{\tau} \sum_{j=(i-1)\tau+1}^{i\tau} r_j.
\]
In the following, we denote by $\tau':=\tau-b$ the maximum number of data that we can collect from every phase. At the end of each phase, we store all or a subset of the $(x_t, a_t, R_t)$ tuples in our replay buffer $\mathcal{R}$.
We extract feature $\phi(x, a)\in\mathbb{R}^d$ for the state-action pair $(x, a)$ and let $\mathcal{F}\subseteq\{f:\mathbb{R}^d\mapsto\mathbb{R} \}$ be a class of functions that we use to approximate the Q-functions. For phase $i$, we propose the following method to estimate $Q_{\pi_i}$: $\widehat Q_{\pi_i}(x, a)=\widehat f(\phi(x, a))$, where $\widehat f \in \arg\min_{f\in\mathcal{F}} \ell_i(f)$ and
\begin{equation}\label{eq:squared}
\ell_i(f):=\frac{1}{\tau'}\sum_{t=(i-1)\tau + 1}^{i\tau - b}(f(\phi(x_t, a_t)) - R_t)^2.
\end{equation}
Suppose that we store all the data from the previous phases in the replay buffer, then in order to estimate the average of the Q-functions of the first $k$ phases, i.e., $\widehat Q_k$, we propose to use the heuristic that minimizes the average of the $k$ squared loss functions defined in Eq.~\eqref{eq:squared}, i.e., $\frac{1}{k}\sum_{i=1}^k \ell_i(f)$.
{\bf Subsampling and coreset.}
In practice, due to the high memory cost, it may be hard to store all the data from the previous phases in the replay buffer. We found that a simple strategy to resolve this issue is to begin with storing all the data from every phase, and add a limit on size of the replay buffer. When the buffer size exceeds the limit, we eliminate a subset of the data uniformly at random.
Another approach is to sample a subset of size $s$ from the data collected in each phase. Denote this subset by $\mathcal{R}_i$ for phase $i$. Thus in the $k$-th phase, we have $\tau'$ data $\mathcal{D}_k$ from the current phase as well as $s(k-1)$ data from the replay buffer $\mathcal{R}$. First, suppose that the $s$ data points are sampled uniformly at random. We can then minimize the following objective:
\begin{equation}\label{eq:uniform_downsample}
\min_{f\in\mathcal{F}}\frac{1}{k}\Big(\ell_i(f) + \sum_{i=1}^{k-1}\widehat{\ell}_i(f)\Big),
\end{equation}
where $\widehat{\ell}_i(f) := \frac{1}{s}\sum_{(x_t, a_t, R_t)\in\mathcal{R}_i} (f(\phi(x_t, a_t)) - R_t)^2$ is an unbiased estimate of $\ell_i(f)$. Further, uniform sampling is not the only way to construct an unbiased estimate. In fact, for any discrete distribution, with PMF $q=\{q_t\}$, over the $\tau'$ data in $\mathcal{D}_i$, we can sample $s$ data points according to $q$ and construct
\begin{equation}\label{eq:hatell_q}
\widehat{\ell}_i(f) := \frac{1}{\tau'}\sum_{(x_t, a_t, R_t)\in\mathcal{R}_i} \frac{1}{q_t} (f(\phi(x_t, a_t)) - R_t)^2,
\end{equation}
in order to obtain an unbiased estimate of $\ell_i(f)$. As shown by~\citet{bachem2017practical}, by choosing $q_t \propto (f(\phi(x_t, a_t)) - R_t)^2$, we can minimize the variance of $\widehat{\ell}_i(f)$ for any fixed $f$. In the following, we call the subset of data sampled according to this distribution a \emph{coreset} of the data. In the experiments in Section~\ref{sec:experiments}, we show that thanks to the variance reduction effect, sampling a coreset often produces better performance than sampling a subset uniformly at random, especially when the rewards are sparse.
{\bf Comparison to \citet{politex}.}
Our algorithm can be considered as an efficient implementation of {\textsc{Politex}} ~\citep{politex} via experience replay. In the original {\textsc{Politex}} algorithm, the Q-functions are estimated using Eq.~\eqref{eq:squared} for each phase, and all the functions are stored in memory. When an agent interacts with the environment, it needs to evaluate all the Q-functions in order to obtain the probability of each action. This implies that the time complexity of computing action probabilities increases with the number of phases, and as a result the algorithm is hard to scale to a large number of phases. Although we can choose to evaluate a random subset of the Q-functions to estimate the action probabilities in \textsc{Politex}, our implementation via experience replay can still be faster since we only need to evaluate a single function, trained with replay data, to take actions.
\section{Experiments}\label{sec:experiments}
In this section, we evaluate our implementations empirically. We make comparisons with several baselines in two control environments.
{\bf Environments.}
We use two control environments with the simulators described in~\citet{tassa2018deepmind}. Both environments are episodic with episode length $1000$.
The environments we evaluate are:
\begin{itemize}
\item \emph{Cart-pole}~\citep{barto1983neuronlike}: The goal of this environment is to balance an unactuated pole by applying forces to a cart at its base. We discretize the continuous force to $5$ values: $\{-2, -1, 0, 1, 2\}$.
The reward at each time step is a real number in $[0, 1]$. We end episodes early when the pole falls (we use rewards less than $0.5$ an indicator for falling), and assume zero reward for the remaining steps when reporting results.
\item \emph{Ball-in-cup}: Here, an actuated planar receptacle can translate in the vertical plane in order to swing and catch a ball attached to its bottom. This task has a sparse reward: $1$ when the ball is in the cup, and $0$ otherwise. We discretize the two-dimensional continuous action space to a $3\times 3$ grid, i.e., $9$ actions in total.
\end{itemize}
\begin{figure*}[ht]
\centering
\includegraphics[width=0.32\textwidth]{arxiv/cartpole_reward_phase.pdf}
{\includegraphics[width=0.32\textwidth]{arxiv/cartpole_runtime.pdf}}
{\includegraphics[width=0.32\textwidth]{arxiv/cartpole_coreset.pdf}}
\subfigure[Iteration complexity] {\includegraphics[width=0.32\textwidth]{arxiv/ball_in_cup_reward_phase.pdf}}
\subfigure[Training speed]
{\includegraphics[width=0.32\textwidth]{arxiv/ball_in_cup_runtime.pdf}}
\subfigure[Replay subsampling strategies]{\includegraphics[width=0.32\textwidth]{arxiv/ball_in_cup_coreset.pdf}}
\caption{Experiments on the Cart-pole environment (top) and Ball-in-cup (bottom).}
\label{fig:experiments}
\end{figure*}
{\bf Algorithms.}
We compare the following algorithms: our proposed implementation of {\textsc{Politex}} using experience replay, {\textsc{Politex}} using weight averaging, the original {\textsc{Politex}} algorithm, the variant of {\textsc{Politex}} that averages $10$ randomly selected Q-functions, the mirror descent policy optimization (MDPO) algorithm~\citep{tomar2020mirror}, the constrained policy optimization (CPO)~\citep{achiam2017constrained}, and the V-MPO algorithm~\citep{song2019v}.
For all the algorithms, we extract Fourier basis features~\citep{konidaris2011value} from the raw observations for both environments: for Cart-pole, we use $4$ bases and for Ball-in-cup we use $2$ bases. For variants of {\textsc{Politex}} we construct the state-action features $\phi(x, a)$ by block one-hot encoding, i.e., we partition the $\phi(x, a)$ vector into $|\mathcal{A}|$ blocks, and set the $a$-th block to be the Fourier features of $x$ and other blocks to be zero.
We approximate $Q$-functions using neural networks with one hidden layer and ReLU activation: the width of the hidden layer is $50$ for Cart-pole and $250$ for Ball-in-cup.
For MDPO, CPO and V-MPO, the algorithms use a policy network whose input and output are the state feature and action probability, respectively. We use one hidden layer networks with width $50$ for Cart-pole and $250$ for Ball-in-cup. These three algorithms also need a value network, which takes the state feature as input and outputs its value for the current policy. For both environments, we use one hidden layer network with width $50$.
For Cart-pole, we choose phase length $\tau=10^4$ and for Ball-in-cup, we choose $\tau=2\times 10^4$. Since the environments are episodic, we have multiple episodes in each phase. For Cart-pole, since the training usually does not need a large number of phases, we do not set limit on the size of the replay buffer, whereas for Ball-in-cup, we set the limit as $2\times 10^6$. For our algorithm and the variants of \textsc{Politex}, we choose parameter $\eta$ in $\{5, 10, 20, 40, 80, 160\}$ and report the result of each algorithm using the value of $\eta$ that produces the best average reward. For MDPO, CPO and V-MPO, we note that according to Eq.~\eqref{eq:kl_politex}, the KL regularization coefficient between policies is the reciprocal of the $\eta$ parameter in our algorithm, we also run these baseline algorithms in the same range of $\eta$ and report the best result. In addition, in CPO, we set the limit in KL divergence between adjacent policies $0.001\eta$, which, in our experiments, leads to good performance. We treat the length of Monte Carlo estimate $b$ as a hyper parameter, so we choose it in $\{100, 300\}$ and report the best performance.
{\bf Results.} We run each algorithm in each environment $20$ times and report the average results as well as the standard deviation of the $20$ runs (shaded area). We report the average reward in each phase before training on the data collected during the phase. The results are presented in Figure~\ref{fig:experiments}. Every run is conducted on a single P100 GPU.
From Fig.~\ref{fig:experiments}(a), we can see that the iteration complexity (average reward vs number of phases) and final performance of the experience replay implementation using all the data from past phases (blue curve) are similar to many state-of-the-art algorithms, such as {\textsc{Politex}} and V-MPO. Meanwhile, according to Fig.~\ref{fig:experiments}(b), the experience replay implementation achieves strong performance in training speed, i.e., best performance at $30$ minute in Cart-pole and second best performance at $200$ minute in Ball-in-cup.
We note here that the training time includes both the time for data collection, i.e., interacting with the environment simulator and the training of a new policy. The observation that the experience replay based implementation is faster than the original {\textsc{Politex}} and the implementation that uses $10$ Q-functions can be explained by the fact that the replay-based algorithm only uses a single Q-function and thus is faster at data collection, as discussed in Section~\ref{sec:replay}. In addition, that the replay-based algorithm runs faster than MDPO, CPO, and V-MPO can be explained by its simplicity: in variants of \textsc{Politex}, we only need to approximate the Q-function, whereas in MDPO, CPO, and V-MPO, we need to train both the policy and value networks.
The {\textsc{Politex}} weight averaging scheme achieves
the best performance on Ball-in-cup in both iteration complexity and training speed, but does not converge to a solution that is sufficiently close to the optimum on Cart-pole. Notice that for weight averaging, we used the initialization technique mentioned in Section~\ref{sec:weight_ave}. Considering the consistent strong performance of experience replay in both environments, we still recommend applying experience replay when using non-linear function approximation.
In Fig.~\ref{fig:experiments}(c), we can see that when we only sample a small subset ($1\%$ or $4\%$) of the data to store in the replay buffer, the coreset technique described in Section~\ref{sec:replay} achieves better performance due to the variance reduction effect. This improvement is quite significant in the Ball-in-cup environment, which has sparse rewards. Notice that sampling coreset only adds negligible computational overhead during training, and thus we recommend the coreset technique when there exists a strict memory constraint.
\section*{Acknowledgements}
We would like to thank Mehrdad Farajtabar, Dilan G\"or\"ur, Nir Levine, and Ang Li for helpful discussions.
\bibliographystyle{abbrvnat}
|
\section{Introduction}
Capacity-approaching forward error correction (FEC) based on low-density parity-check (LDPC) codes\cite{Smith-ite, Kudekar-CC, Koike-ite, Liva-PEXIT, Chang-PEXIT, Wei-NB, EXIT, SEXIT, Thorpe-proto03, Thorpe-proto04, Fossorier-girth, Tanner-girth, Kim-girth, Wang-girth} have made a great contribution to increasing data rates of wireless and optical communication systems.
However, the pursuit of high FEC performance has led to a significant increase in power consumption and circuit size.
Hence, attaining a good trade-off between performance and computational complexity is of great importance.
In addition, recent high-performance LDPC codes usually require very large codeword lengths,
whereas shorter FEC codes are preferred\cite{Liva-short} for latency-constrained systems, such as Internet-of-Things (IoT) applications.
Polar codes\cite{Arikan-polar, Tal-list, Shin-polar, Tal-design, Arikan-sys, Sarkis-sys, Li-list, Seidl-polar, Mori-DE, Trifonov-GA, He-beta, Elkelesh-GA, Ebada-DL, Alamdar-SSC, Doan, Elkelesh-BPL, Presman-mixed, Mori-NB, Gabry-kernel, relaxedPolar, Koike-JLT18, Koike-ICC18, Zhang-CRC, Wang-RS, Li-RM, Bourassa} have drawn much attention as alternative capacity-approaching codes in place of LDPC codes for short block lengths, in particular for the fifth-generation (5G) networks.
Besides encoder design methods\cite{Tal-design, Mori-DE, Trifonov-GA, He-beta, Elkelesh-GA, Ebada-DL},
a number of decoder algorithms were developed\cite{Alamdar-SSC, Doan, Elkelesh-BPL}.
With successive cancellation list (SCL) decoding\cite{Tal-list}, polar codes can be highly competitive with state-of-the-art LDPC codes.
To date, various extended versions based on polar coding have also been proposed in literature; e.g.,
nonbinary\cite{Mori-NB}, mixed-kernel\cite{Presman-mixed, Gabry-kernel}, irregular\cite{relaxedPolar, Koike-JLT18},
concatenated\cite{Zhang-CRC, Wang-RS, Li-RM}, convolutional\cite{Bourassa}, and turbo product coding\cite{Koike-ICC18}.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{fig/polar16.pdf}
\caption{Four-stage polarization: QC polar codes $(2^4,2^3,2^8)$ having circulant shift values for $32$ proto-polarization units.}
\label{fig:polar16}
\end{figure}
In this paper, we introduce a novel family of protograph-based polar codes, which we call quasi-cyclic (QC) polar codes having circulant permutation at proto-polarization units as illustrated in Fig.~\ref{fig:polar16}.
With a proper circulant shift value, we show that the QC polar codes can eliminate short cycles in the code graph, which achieves a remarkable breakthrough toward resolving the long-standing issue that the belief-propagation (BP) decoding does not perform well for polar codes.
In the QC polar codes, highly parallel short polar codes are coupled to achieve performance comparable to longer polar codes while maintaining the computational complexity as low as that of short polar codes.
The contributions of this paper are summarized as follows:
\begin{itemize}
\item {\bf{Protograph-based polar codes}}:
We propose a new family of protograph polar codes.
To the best of the authors' knowledge, the concept of protograph codes has never been applied to
such non-parity-check codes.
\item {\bf{QC polar codes}}:
We introduce highly parallelizable QC polar codes, a special case of protograph, with circulant permutations at the proto-polarization units.
\item {\bf{Complexity analysis}}:
We show that the computational complexity of the proposed QC polar codes can be significantly decreased with a protograph lifting factor.
\item {\bf{Girth analysis}}:
We analyze short cycles of the protograph polar codes, and develop a design method to increase the girth.
Eliminating short cycles enables BP decoding to properly work for the QC polar codes.
\item {\bf{State-of-the-art performance}}:
We demonstrate that the QC polar codes with shallow polarization can achieve competitive performance of deep polarization codes.
\item {\bf{Irregular pruning}}:
Further complexity reduction and performance improvement are shown with irregular pruning of polarization to cut loops in the protograph.
\end{itemize}
\section{Basics of Polar Codes}
\subsection{Polar Encoding}
An $n$-stage polar code with $K$ information bits and $N=2^n$ encoded bits uses an $N\times N$ generator matrix $\boldsymbol{G}^{ \otimes n}$ for encoding, where
$[\cdot]^{\otimes n}$ denotes the $n$-fold Kronecker power and $\boldsymbol{G}$ is a binary kernel matrix defined as
\begin{align}
\boldsymbol{G} &=
\begin{bmatrix}
1 & 1 \\
0 & 1 \\
\end{bmatrix}
.
\end{align}
Let ${\boldsymbol{u}} = [{u_1},{u_2},\ldots,{u_N}]^\mathrm{T}$ and
${\boldsymbol{x}}= [{x_1},{x_2},\ldots,{x_N}]^\mathrm{T}$ respectively denote the column vectors of input bits and encoded bits.
The codeword (for non-systematic codes) is given
by $\boldsymbol{x} = \boldsymbol{G}^{\otimes n} \boldsymbol{B} \boldsymbol{u}$, where
the matrix multiplications are carried out over the binary field (i.e., modulo-$2$ arithmetic),
and $\boldsymbol{B}$ denotes an $N \times N$ bit-reversal permutation matrix\cite{Arikan-polar}.
Due to the nature of the Kronecker product, polar encoding and decoding
can be performed at a complexity on the order of $\mathcal{O}(N \log_2 N)$.
The multi-stage operation of the Kronecker products gives rise to the so-called
polarization phenomenon to approach capacity in arbitrary channels\cite{Arikan-polar}.
The polar coding maps the information bits to the $K$ most reliable locations in $\boldsymbol{u}$.
The remaining $N-K$ input bits are frozen bits, known to both encoder and decoder.
We use $\mathbb{K}$ and $\bar{\mathbb{K}}$ to denote the subsets of $\{1, 2, \ldots, N\}$ that correspond to the information bit and frozen bit locations, respectively.
The lowest reliability can be selected to be in $\bar{\mathbb{K}}$ for frozen bits, e.g., by Bhattacharyya parameter\cite{Arikan-polar}, density evolution\cite{Mori-DE, Tal-design},
Gaussian approximation\cite{Trifonov-GA}, beta expansion\cite{He-beta}, genetic algorithm\cite{Elkelesh-GA}, and deep learning\cite{Ebada-DL}.
\subsection{Polar Decoding}
The original SC decoder\cite{Arikan-polar} proceeds sequentially over the bits, from $u_1$ to $u_N$.
For each index $i \in \{1,2, \ldots, N\}$, an estimate $\hat{u}_i$ for bit $u_i$ is made as follows.
If $i \notin \mathbb{K}$, then $\hat{u}_i$ is set to the known value of $u_i$, otherwise, when $i \in \mathbb{K}$, $\hat{u}_i$ is set to the most likely value for $u_i$ given the channel outputs,
assuming that the previous estimates $[\hat{u}_1, \hat{u}_2, \ldots, \hat{u}_{i-1}]$ are correct.
The SC decoding was improved by the SCL decoder\cite{Tal-list},
which proceeds similarly to the SC decoder, except that for each data bit index $i \in \mathbb{K}$, the decoder retains both possible estimates, $\hat{u}_i = 0$ and $\hat{u}_i = 1$, in subsequent decoding paths.
The list-decoding approach limits the number of paths to a fixed-size list $L$ of the most likely partial paths.
The combination of SCL decoding with an embedded cyclic redundancy check (CRC) to reject invalid paths yields significantly improved performance\cite{Tal-list, Zhang-CRC}.
Various other decoding algorithms were proposed in the literature, e.g., simplified SC decoding\cite{Alamdar-SSC}, neural SC decoding\cite{Doan}, and BP list decoding\cite{Elkelesh-BPL}.
\subsection{Computational Complexity}
\label{sec:comp}
It is known that short LDPC codes do not perform well as shown in~\cite{Liva-short}, where
nonbinary (NB) coding can improve LDPC codes in the short-length regimes.
However, the computational complexity of NB-LDPC decoding is generally higher than binary counterparts, in particular for large Galois field sizes.
It is thus of great importance to realize low computational complexity in addition to high coding gain.
We evaluate the computational complexity of polar decoding and show that it is competitive with LDPC decoding.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{fig/complexity.pdf}
\caption{Computational complexity per coded bit as a function of block length $N$ for standard polar SCL decoding (per list), QC polar BP decoding and LDPC BP decoding (per iteration).}
\label{fig:comp}
\end{figure}
The polar SCL decoding has a log-linear complexity; specifically, $\mathcal{O}[L N \log_2(N)/2]$) for a list size of $L$.
This nonlinearity is a major drawback in comparison to the linear complexity of LDPC BP decoding, i.e.,
$\mathcal{O}[2I d_\mathrm{v} N]$ where $d_\mathrm{v}$ denotes the average degree of variable nodes (VNs).
Note that the factor of $2I d_\mathrm{v}$ comes from the bidirectional message passing, whereas SCL decoding uses unidirectional message passing over $N\log_2(N)/2$ VNs.
Due to the nonlinear complexity, polar codes can eventually be less effective than LDPC codes as we increase the block lengths $N$.
However, it turns out to be an advantage when we aim to reduce the block sizes in order to decrease decoding latency.
This is illustrated in Fig.~\ref{fig:comp}, where complexity per coded bit (i.e., divided by $N$) is plotted as a function of block length $N$ for polar and LDPC decoding.
Because per-bit complexity is constant depending on average degree $d_\mathrm{v}$ for LDPC codes, there is no
motivation to decrease the block length.
In contrast, polar decoding becomes simpler when we reduce block sizes.
Remarkably, polar decoding will be more efficient than typical LDPC decoding at short block lengths of $N < 10^4$.
This promotes polar codes as a strong candidate for latency-critical systems.
Although the actual computational complexity may vary depending on hardware implementation,
most prototyping studies\cite{Sarkis-sys} have revealed that polar codes can compete favorably with LDPC codes in terms of complexity.
Note that the LDPC decoding is more complicated for higher rates because
the average check-node (CN) degree is larger, whereas polar codes have at most three degrees at CNs.
Nevertheless, polar SCL decoding is not amenable to parallel implementation.
In this paper, we propose a highly parallelizable polar code family whose complexity is $\mathcal{O}[LN\log_2(N/Q)/2]$ for a parallelism factor of $Q$.
From Fig.~\ref{fig:comp}, we can observe the significant advantage in its decoding complexity.
The details of our proposed QC polar codes will be described in the next section.
\section{Protograpah-Based QC Polar Codes}
\subsection{Protograph Codes}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{fig/lifting.pdf}
\caption{Lifting operation for proto-polarization units:
(a) regular polar units,
(b) replication of $Q$-parallel encoders,
(c) permutation for interleaving intermediate encoding bits.}
\label{fig:lifting}
\end{figure}
Thorpe\cite{Thorpe-proto03} introduced the concept of protograph codes, a class of LDPC codes constructed from a protograph in such a way that the $1$'s in the parity-check matrix are replaced by $(Q\times Q)$-permutation matrices and the $0$'s by $(Q\times Q)$-zero matrices.
The permutation size $Q$ is also called a lifting size.
If the permutation matrices are circulant, the protograph code reduces to a well-known QC LDPC code\cite{Thorpe-proto04}.
To the best of authors' knowledge, no studies have been reported for designing the protographs for polar codes.
Analogously in lifting operations of the parity-check matrix for LDPC codes, we replace the generator matrix of polar codes.
For example, the following generator matrix for $2$-stage polar codes is replaced with permutation matrices $\boldsymbol{P}_{i,j}$:
\begin{align}
\boldsymbol{G}^{\otimes 2} &=
\begin{bmatrix}
1 & 1 & 1 & 1\\
0 & 1 & 0 & 1\\
0 & 0 & 1 & 1\\
0 & 0 & 0 & 1\\
\end{bmatrix}
\mathop{\Longrightarrow}_{\mathrm{Lifting}}
\begin{bmatrix}
\boldsymbol{P}_{1,1} & \boldsymbol{P}_{1,2} & \boldsymbol{P}_{1,3} & \boldsymbol{P}_{1,4} \\
\boldsymbol{0} & \boldsymbol{P}_{2,2} & \boldsymbol{0} & \boldsymbol{P}_{2,4}\\
\boldsymbol{0} & \boldsymbol{0} & \boldsymbol{P}_{3,3} & \boldsymbol{P}_{3,4}\\
\boldsymbol{0} & \boldsymbol{0} & \boldsymbol{0} & \boldsymbol{P}_{4,4}\\
\end{bmatrix}
,\notag
\end{align}
where $\boldsymbol{0}$ is an all-zero matrix of size $Q\times Q$.
The simplest choice of permutation matrices is a weight-$1$ circulant matrix:
$
\boldsymbol{P}_{i,j} =
\boldsymbol{I}(s'_{i,j})
$,
where $\boldsymbol{I}(s)$ denotes the $s$th circulant permutation matrix obtained by cyclically right-shifting a $Q\times Q$ identity matrix by $s$ positions, and $s'_{i,j}$ is a shift value to design.
For this special case, we may call the protograph polar codes as QC polar codes.
It can be regarded as a generalized low-density generator matrix (LDGM) based on polar codes.
We consider a hardware-friendly lifting operation at each polarization stage with identity diagonal matrices $\boldsymbol{P}_{i,i} = \boldsymbol{I}(0)$.
Our lifting operation is illustrated in Fig.~\ref{fig:lifting}, where we replicate $Q$-parallel polar encoders and permute exclusive-or (XOR) incident bits among the parallel encoders.
Fig.~\ref{fig:polar16} shows an example of QC polar codes having a shift base matrix of size $n\times 2^{n-1}$ as follows:
\begin{align}
\boldsymbol{S} &=
\begin{bmatrix}
139 & 252 & 234 & 156 & 157 & 142 & 50 & 68 \\
134 & 25 & 178 & 20 & 254 & 101 & 146 & 212 \\
79 & 192 & 144 & 129 & 204 & 71 & 237 & 252 \\
37 & 235 & 140 & 72 & 255 & 137 & 203 & 133 \\
\end{bmatrix}
,
\label{eq:shift16}
\end{align}
whose $(i,j)$th shift value is assigned for the $j$th proto-polarization unit at the $i$th stage.
Note that the QC polar codes still hold most benefits of original polar codes such as structured encoding and decoding.
As we will discuss below, the QC polar codes have a number of remarkable features.
\subsection{High-Girth Design}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{fig/cycle.pdf}
\caption{Short cycle examples in QC polar codes:
(a) cycle-$4$ message passing loop,
(b) cycle-$6$ message passing loop.}
\label{fig:cycle}
\end{figure}
In order to achieve good performance, we shall design the shift values of QC polar codes.
One obviously poor choice is the case when we use all zeros for shifting, leading to mutually independent $Q$-parallel short polar codes without any coupling gain.
The protograph codes are often designed to achieve a high girth --- the ``girth'' of a code is the length of the shortest cycle in the code graph.
It is known\cite{Fossorier-girth} that the girth of any conventional QC LDPC code is upper bounded by $12$.
Tanner\cite{Tanner-girth} proposed a systematic way to optimize shift values to achieve girth-12.
It was further shown in~\cite{Kim-girth} that an irregular QC LDPC code can achieve a girth larger than 12.
For QC polar codes $(2^n, K, Q)$ of code length $N=2^nQ$, there are $n 2^{n-1}$ shift values to design as in (\ref{eq:shift16}).
Unfortunately, the factor graph of polar codes are inherently loopy and there exist a large number of short cycles as illustrated in Fig.~\ref{fig:cycle}.
Nonetheless, by optimizing shift values, we can increase the girth for QC polar codes with $Q>1$.
For example, the cycle-$4$ loop in Fig.~\ref{fig:cycle}(a) can be eliminated if the shift values satisfy the condition\cite{Wang-girth}:
\begin{align}
{}-s_{1,1} - s_{2,2} + s_{1,2} + s_{2,1} \neq 0 \pmod Q,
\label{eq:girth4}
\end{align}
where we accumulate shift values of all proto-CNs along the loop.
Note that the shift values are negated if the path goes downward.
This explains the long-lasting problem that the BP decoding performs very poorly for the conventional polar codes ($Q=1$), i.e., the accumulated shifts will be always zero, resulting in a small girth of $4$.
Our QC polar codes resolve this issue by maximizing the girth in the protograph.
Similarly, the cycle-$6$ loop in Fig.~\ref{fig:cycle}(b) can be removed if we can satisfy
\begin{align}
{}-s_{1,3} -s_{3,4} + s_{2,4} + s_{1,2} + s_{3,1} - s_{2,1} \neq 0 \pmod Q.
\notag
\end{align}
Note that irregular polar coding\cite{Koike-JLT18} is an alternative that could remove some, but not all, short-cycle loops.
We extended the greedy design method used for irregular polar coding to jointly optimize frozen bit locations and circulant shift values by means of protograph extrinsic information transfer (P-EXIT)\cite{Liva-PEXIT, Chang-PEXIT} and a hill-climbing girth maximization\cite{Wang-girth}.
\subsection{Error-Rate Performance}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{fig/resQCm2.pdf}
\caption{BER performance of $32$-iteration BP decoding for $2$-stage half-rate QC polar codes $(2^2,2^1,2^q)$.
Frozen bit indications are $[1,1,0,0]$.}
\label{fig:m2}
\end{figure}
Fig.~\ref{fig:m2} shows bit-error rate (BER) performance as a function of signal-to-noise ratio (SNR) for short polar codes $(2^n,2^{n-1},2^q)$ with $n=2$ polarization stages in additive white Gaussian noise channels.
We here use $32$-iteration BP decoding with two-way round-robin scheduling from the first to the last stages and its reversed direction alternatingly (parallel flooding updates per stage).
The first two bits $[u_1, u_2]$ are frozen.
The following shift base matrices are considered:
\begin{align}
\begin{bmatrix}
0 & 0 \\
0 & 0 \\
\end{bmatrix}
,\quad
\begin{bmatrix}
0 & 1 \\
1 & 0 \\
\end{bmatrix}
,\quad
\begin{bmatrix}
0 & 0 \\
0 & 1 \\
\end{bmatrix}
,\quad
\begin{bmatrix}
0 & 0 \\
-1 & 1 \\
\end{bmatrix}
\end{align}
for standard polar codes ($Q=2^q=1$), un-optimized QC polar codes,
optimized QC polar codes, and irregular QC polar codes, respectively.
We denote a pruned polarization by a negative shift value.
As the first two cases do not satisfy the condition in (\ref{eq:girth4}) to eliminate girth-$4$,
the BER performance is worse than the other two cases.
By eliminating the cycle-$4$ loop, the QC polar codes can achieve a gain of $0.9$~dB without sacrificing any computational complexity.
We found that frozen bit locations are also important; specifically, no gain was achieved with $[u_1, u_3]$ being frozen.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{fig/resQCm4.pdf}
\caption{BER performance of $32$-iteration BP decoding for $4$-stage half-rate QC polar codes $(2^4,2^3,2^q)$.
Frozen bit indications are $[1,1,1,1,1,1,1,0,1,0,0,0,0,0,0,0]$.}
\label{fig:m4}
\end{figure}
The performance improvement can be more significant for a deeper stage $n$ and larger lifting size $Q$.
We plot the BER performance for half-rate $4$-stage QC polar codes $(2^4,2^3,2^q)$.
We can see that the increase of the lifting size $Q=2^q$ can significantly improve performance by up to $2.4$~dB gain over the standard polar codes.
It should be noted that the per-bit complexity is identical for all of these QC polar codes regardless of the lifting size $Q$;
specifically, $Q$ parallel decoding of short polar codes requires a total complexity on the order of $Q\times \mathcal{O}[n2^{n-1}]$ for a total codeword length of $Q\times 2^n$ bits.
For $Q=256$, our girth design method could remove all short cycles up to $6$.
The designed shift base matrix is written in (\ref{eq:shift16}), and also depicted in Fig.~\ref{fig:polar16}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{fig/resQCn16.pdf}
\caption{BER performance for long half-rate QC polar codes $(2^n, 2^{n-1}, 2^q)$ of block length $N=2^{n+q}=2^{16}$.}
\label{fig:n16}
\end{figure}
We next demonstrate that our QC polar codes using shallow polarization stages can compete
against long standard polar codes with deeper polarization stages.
Fig.~\ref{fig:n16} shows the BER performance of $16$-stage standard polar codes and $6$-/$8$-stage QC polar codes for a total block length of $N=2^{16}$ bits.
We also present the shallow $6$-/$8$-stage polar codes without protograph lifting.
Noticeably, shallow $6$-stage QC polar codes with $Q=1024$ parallel BP decoding can outperform SC decoding of $16$-stage polar codes.
Furthermore, our $8$-stage QC polar codes with $Q=256$ can achieve performance competitive with state-of-the-art SCL decoding (with a list size of $L=32$) for the long $16$-stage polar codes.
In addition, it was verified that the QC polar codes can resolve the issues of BP decoding to offer comparable performance to SCL decoding with large list sizes.
These results are practically impactful because the encoding, decoding, and code design of shallower polar codes are much simpler and more efficient.
\subsection{Irregular Pruning}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{fig/irreg_m6.pdf}
\caption{Impact of pruning proto-polarizations for irregular QC polar codes $(2^n,2^{n-1},2^q)$.}
\label{fig:irreg}
\end{figure}
We finally investigate the irregular QC polar codes which deactivate polarization units.
As discussed, pruning proto-polarization units may also assist removing short cycles.
It was shown in\cite{Koike-JLT18} that the conventional irregular polar codes are often capable of reducing the encoding/decoding complexity, decoding latency, and even BER (due to improved Hamming weight distributions).
Fig.~\ref{fig:irreg} shows performance of QC polar codes $(2^n, 2^{n-1}, 2^q)$ when proto-polarization units are gradually pruned by a greedy algorithm\cite{Koike-JLT18}.
For $4$-stage polar codes, the performance degrades as the number of inactive polarizations increases.
Nevertheless, the QC polar codes with $Q=256$ are still better than the standard polar codes ($Q=1$) until $64$\% of the polar units are removed.
For $6$-stage polar codes with $Q=1024$ lifting, it was observed that pruning up to $7$\% of the proto-polarization units improves the BER performance over the regular counterpart.
In consequence, the irregular QC polar codes can further reduce the decoding complexity with potential performance improvement.
\subsection{Discussion}
Some major advantages of the proposed protograph polar codes are listed below:
\begin{itemize}
\item The girth of polar codes can be increased significantly.
\item The BP decoding can compete with SCL decoding.
\item Multiple short polar encoders and decoders are implemented in a fully parallel fashion with no additional complexity besides circulant message exchanges.
\item It realizes a low computational complexity equivalent to $Q$-fold shorter polar codes.
\item Shallow polarization offers comparable performance to deeper polarization.
\item Code design is simpler using shallower polarization.
\item There is a higher flexibility in codeword lengths of non powers-of-two, i.e., $N=2^n Q$.
\item Irregular polarization is straightforward to apply with the shift value matrix design.
\item Well-established techniques such as girth design and P-EXIT from LDPC codes are applicable.
\end{itemize}
We however found that the recent BP list decoding\cite{Elkelesh-BPL} was not compatible with the QC polar codes as it is.
As we focused on the proof-of-principle study in this paper, there remain many research directions, including extensions to BP list decoding, systematic encoding, nonbinary codes, systematic circulant shift design, BP scheduling optimization, and multi-weight permutation.
In particular, it is interesting to consider inhomogeneous polar codes, e.g., non-identical frozen bit locations across $Q$ polar codes, and circulant permutations among different polarization stages.
We also note that our QC polar codes are similar to polar product codes\cite{Koike-ICC18} in the sense that parallel short polar codes are coupled, but the fundamental difference lies in its mechanism of coupling (QC polar codes do not need additional row/column polar codes but computation-free circulant).
Also, mixture-kernel and nonbinary polar codes are analogous to the QC polar codes in the sense that a single polarization unit processes multiple bits at once in parallel.
However, nonbinary polar codes require additional complexity and there is a limited flexibility to choose a Galois field size $Q$.
More importantly, $Q$-ary polar codes have only $\log_2(Q)$ bits parallelism, whereas fully $Q$-parallel encoding is possible in the QC polar codes.
We believe that the protograph design for generalized LDGM (including QC polar codes) will stimulate the research community.
\section{Conclusions}
We proposed a novel class of polar codes called QC polar codes,
which replicate parallel short polar encoders and decoders with circulant permutations to exchange intermediate messages among them.
We developed a protograph-based design method to optimize the girth to achieve high coding gain.
By removing short cycles, the proposed QC polar codes can outperform the standard polar codes.
We believe that the QC polar coding offers a breakthrough to resolve the long-standing issue that BP decoding performs poorly for conventional polar codes due to the inherent girth-4.
It was demonstrated that the QC polar codes with shallow polarization can achieve the state-of-the-art performance of deeper polar codes at a considerably reduced complexity.
The QC polar codes are hardware friendly as highly parallel encoding/decoding is feasible with a reduced polarization stage.
We also evaluated the impact of irregular QC polar codes, which can further decrease the complexity and BER for some cases.
For the proposed protograph codes, we addressed a number of interesting extensions for future research directions.
\section*{Acknowledgment}
We would like to thank
Drs. David S. Millar, Keisuke Kojima, and Kieran Parsons at MERL for their useful discussion.
|
\section{INTRODUCTION}
While end-to-end (E2E) \cite{graves2014towards} deep learning (DL) models brought great
improvements to the field of automatic speech recognition (ASR)
in recent years and reduced word error rates (WER) on benchmark
datasets significantly \cite{li2020comparison}, they also come with a set of problems.
Even though they are designed to work well with unknown words,
since they usually concatenate subword-units and can thus essentially
produce any textual transcription, they tend to perform even worse on
rare words than classical DL ASR systems \cite{spell2019}. This is due to them
having an internal latent language model, which is biased and cannot
be tweaked. For example, an ASR trained on the LibriSpeech corpus
is very likely to transcribe the name \textit{Hilde} as \textit{Hilda} or the name \textit{Josef}
as \textit{Joseph}.
Even though those words are rare, they tend to be of great importance
for many tasks. A good example for this is automatic meeting
transcription. People frequently address each other by name in a
meeting. And meeting participants can have very unique names with
unique pronunciations. So for the transcriptions, the prior knowledge
which names to expect could be very helpful. Incorporating it
however is a non-trivial task, because E2E approaches cannot be fine-tuned
easily, since they lack a lexicon or a pronunciation dictionary.
Proper nouns have been identified as a challenging problem in
ASR for a while now \cite{proper2014}. Recently some approaches have arisen to
tackle this challenge with E2E ASR using a specialised architecture
and losses \cite{peyser2020improving} or using specific data and training procedures to better
represent contextual information \cite{alon2019contextual}. Our approach is meant for rare
words in general, however in this work we choose rare proper nouns
as exemplary data and use few-shot learning to improve performance
on them.
In meta learning \cite{schmidhuber1987evolutionary} using the metric space approach (MSML)
\cite{Goldberger2004NeighbourhoodCA}, an embedding function is trained to transform datapoints into a
metric space, where comparisons purely based on positions are possible.
Since structural knowledge about a task is within the embedding
function, new information can be considered by making comparisons
to reference samples of data we want to adapt to. Essentially clustering
is performed in a latent space. MSML could provide solutions
for our goal of improving recognition of rare words in highly specific
contexts, since it can operate with as little as one reference datapoint
of the data we want to adapt to, it doesn’t need any retraining or time
consuming tweaking mechanisms for new reference samples, and the
decisions made on the basis of the comparisons in the metric space
are simple and interpretable and can thus be integrated into an E2E
workflow with few complications.
We propose a two part ASR system that integrates metric space
representations of expected difficult keywords into the E2E pipeline.
An ASR system is built using the Transformer architecture \cite{vaswani2017attention} which
has been shown to outperform any prior architecture for most speech
related tasks\cite{karita2019comparative, li2019neural, vila2018end, dong2018speech} by heavily relying on attention mechanisms
\cite{bahdanau2015neural}. The intermediate representation that the Transformer-encoder
produces is used as prior embedding and then embedded
further into a metric space on a frame-shifted window basis inspired
by previous work on keyword spotting \cite{chen2014small}. Prior approaches of embedding
audio include using a skip-gram model for speech \cite{chung2018speech2vec} and
using convolutional neural nets (CNNs) for general purpose audio
detection \cite{audiomax}. By using the Transformer as embedding function, we
hope to get more phonetically rich embeddings. The metric space
embeddings are then compared to reference samples using various
renown MSML approaches such as Siamese Networks \cite{bromley1994signature}, Relation
Networks \cite{sung2018learning}, Prototypical Networks \cite{snell2017prototypical} and Matching Networks
\cite{vinyals2016matching}. The results of these comparisons are taken into account when
decoding the orthographic representation of the utterance.
Our main contributions are as follows: we propose a new method
of generating meaningful embeddings of spoken language, we redesign
MSML approaches to spot keywords in a continuous signal
and we showcase an approach to integrate the results into an E2E
ASR system. We verify our proposed methods on two severely different
datasets and observe the embeddings working well, the keyword
spotting achieving high F$_1$ scores and the keyword spotting ASR
interface improving rare-word WER by 1.6\% on average.
\section{PROPOSED APPROACH}
An overview of the entire system we built can be seen in figure \ref{sys}.
The ASR encoder feeds into the ASR decoder, and also into each of
the four MSML agents in a windowed fashion. The ASR decoder
produces hypotheses and the MSML agents (usually only the one you
favor) produce a list of recognized keywords. The selection of the
best hypotheses that the decoder’s beamsearch produces takes the list
of recognized keywords into consideration.
\begin{figure}[htb]
\centering
\centerline{\includegraphics[width=.48\textwidth]{sys.pdf}}
\caption{Overview of how the components are combined}
\label{sys}
\end{figure}
\subsection{Semantics of Audio}
We need to abstract away from speaker, microphone, recording environment
and so on, but retain the phonetic information. To bypass
this challenge, we use the output of the encoder of a Transformer
ASR as the input to the metric space embedding function. The Transformer
encoder has been trained on thousands of hours of speech
beforehand and thus abstracts away from surface level features and is
robust against data variations already.
To use the ASR system as embedding function, a feature matrix
derived from an arbitrary audio signal is fed to the encoder of the
Transformer, yielding an embedding matrix with 512 dimensions
per timestep in our case. The Transformer applies attention with
respect to the entire input sequence itself, so in order to get stable
embeddings we have to limit the attention span to the size of our
window. This is done by windowing over the signal first and encoding
one appropriately sized chunk at a time using the ASR.
\subsection{Meta Learning for Spotting}
MSML approaches are usually used for classification on clean data,
mostly images. We however want a network that can be applied
efficiently to a sequence of windows of a signal and rather than
classifying the correct class, spot whether something is present or
not. This requires some major changes to the architectures of the four
meta learning approaches used.
All of the approaches are redesigned to look at one class at a
time. Each network assigns a similarity score to a pair of query and
a member of the support set of a class. The query is a window of
ASR-encoded signal. For each class, the highest similarity score that
any of its supports achieves with any of the windows is saved. In
the end, the score of each class is compared against a manually set
threshold and if it peaks above, it is added to the list of recognized
keywords that is passed on to the ASR decoder.
For the Siamese Network, the support and the query are embedded
into the metric space, and the cosine similarity (CosSim,
normalized dot product, measure how well vectors align) of the resulting
vectors is computed. The scoring mechanism is displayed in
equation \ref{eq:s}. An encoded matrix M is denoted as enc(M), the query as
$q$, the support as $s$
\begin{equation}
\text{score}_\text{class}^{q} = \max_{s \in \text{class}} \cfrac{\text{enc}(q)\cdot\text{enc}(s)}{||\text{enc}(q)||\cdot||\text{enc}(s)||}
\label{eq:s}
\end{equation}
In our Relation Network implementation, the query and a support are again encoded using the metric space encoder. Here however the similarity function is learnt, not static. The two vectors are concatenated and passed though a small multi layer perceptron to get the similarity score. The relation scoring mechanism can be seen in \ref{eq:r}, $relat$ denotes the relation function, $concat$ denotes vector-concatenation.
\begin{equation}
\text{score}_\text{class}^{q} = \max_{s \in \text{class}} \text{relat}(\text{concat}(\text{enc}(q),\text{enc}(s)))
\label{eq:r}
\end{equation}
Our Prototypical implementation takes in multiple supports at a time, which are then averaged after they have been encoded to build a prototype of the corresponding class, which is then compared to the query using CosSim. For the case of k=1, it is equivalent to the Siamese Network. The scoring mechanism is displayed in equation \ref{eq:pn}, the calculation of the prototype is displayed in equation \ref{eq:p}.
\begin{align}
\text{score}_\text{class}^{q} &= \cfrac{\text{enc}(q)\cdot\text{enc}(p_{\text{class}})}{||\text{enc}(q)||\cdot||\text{enc}(p_{\text{class}})||}
\label{eq:pn}
\\
p_{class} &= \sum_{s \in \text{class}} \cfrac{s}{|\text{class}|}
\label{eq:p}
\end{align}
The Matching Network is usually designed to benefit from fully conditional encoding functions into a metric space, which rely on attention. For this task however, this setup did not converge at all. Instead we use again a simple encoder like the other networks, however with the output being sequences. We apply attention from the query to the support sequence. The conceptual matching lies within the properties of the resulting attention matrix, namely how well the two sequences align. And to get a similarity score from the attention matrix, we pass it through a small long short-term memory layer \cite{hochreiter1997long}. The matching scoring mechanism is displayed in equation \ref{eq:m}, $att$ is an attention mechanism and $reduce$ is any method to reduce over a sequence axis.
\begin{equation}
\text{score}_\text{class}^{q} = \max_{s \in \text{class}} \text{reduce}(\text{att}(\text{enc}(q),\text{enc}(s)))
\label{eq:m}
\end{equation}
\subsection{Integration with Speech Recognition}
The decoder of the ASR produces a set of hypotheses using a beamsearch.
One of them is chosen as the final ASR output. This choice
usually just depends on the ranking that the beamsearch itself assigns
to the hypotheses. In our system, this choice is influenced by a list of
recognized keywords that the MSML agents produce. If a keyword is
spotted, hypotheses that contain the word are moved to the front of
the ranking. The spotting at this level is binary.
To illustrate how this works, assume we have a sentence that
contains the name Noirtier. Shown below are the beginnings of the
hypotheses that our ASR yields for the sentence.
\begin{itemize}
\item $<$eos$>$ nautier was near the bed ...
\item $<$eos$>$ natier was near the bed ...
\item $<$eos$>$ nartier was near the bed ...
\item $<$eos$>$ noirtier was near the bed ...
\end{itemize}
The correct name appears in the hypotheses, however only as the
fourth option. Now if our keyword spotting system detects that the
sentence contains the name, we pick the most probable hypothesis
that contains the desired transcription, which is the fourth here.
\section{DATASETS}
\subsection{Speech Commands Dataset}
\label{data} We train the meta learning agents on the Speech Commands dataset \cite{warden2018speech}. It contains a total of 35 different words with 1000 to 3000 samples each. To further enrich the variety of the dataset, we add 160 more names as keywords with 4 to 6 samples each. Those samples come from three different speech synthesis systems and random data augmentations are applied to them. The additional keywords are carefully selected in order to avoid having names in there that are orthographically different, yet pronounced the same.
\subsection{LibriSpeech Names}
The LibriSpeech corpus \cite{panayotov2015librispeech} is the training set for the Transformer ASR and the development set of the MSML agents. It contains over 1000 hours of read speech. For tuning the meta learning agents, a subset is created from the dev-clean portion. A total of 40 proper nouns are selected, which occur between 4 and 10 times in the whole corpus. For each of them, 4 sentences containing them are selected. From each of the utterances, the name that occurs in them is manually cut out and labelled to create a collection of high quality supports for each name. While we choose proper nouns as exemplary data, what we really want to explore is the performance of rare words, thus the upper limit for their occurrence counts.
\subsection{Singapore English National Speech Names}
The Singapore English National Speech Corpus \cite{koh2019building} contains over 3000 hours of speech. The speakers have varying native languages and are recorded with a variety of microphones in different recording environments. We thus believe that the small sub-corpus of sentences that contain proper nouns which we derived from it is well suited to reflect the performance of each of the components in a challenging close-to-real-world scenario. Our derived corpus contains 120 utterances which contain 30 proper nouns and 4 utterances per name. Again, the samples we use as supports are selected and cut manually.
\section{EXPERIMENTS}
\subsection{Setup}
For the audio embeddings, we train a Transformer ASR on the LibriSpeech corpus using the ESPNet toolkit \cite{watanabe2018espnet, hayashi2020espnet}. The recipe uses 80 Mel-frequency buckets as the input and the training is capped to 40 epochs, similar to the recipe of the IMS-Speech system \cite{denisov2019ims}. The training took 96 hours on 4 Titan X GPUs.
The two experiments use the exact same setup, but look at different aspects. The metric space encoder that is used for all of the MSML agents is built as follows: Two 1D convolutional layers with a kernel size of 3 and 20 filters using hyperbolic tangent as activation function are stacked on top of each other, separated by a maximum pooling layer with a window size of 2. A global max pooling reduces over the sequence axis. The contrastive loss function \cite{lecun2006tutorial, hadsell2006dimensionality} is found to work best. RMSProp is used as the optimizer with a learning rate of 0.001 and a discounting factor of 0.9.
To train the networks, we sample pairs of audio and assign the label 1 if they are instances of the same word or 0 otherwise. We always alternate between positive and negative examples, even though we expect mostly negative labels during inference. We do so for 800,000 steps, even though convergence shows much earlier, since the MSML networks contain only around 40,000 parameters each.
For the actual experiment we perform a random sampling, since sampling every possible combination of queries and supports is not feasible. This means for any N and k, a random query and k random supports for N classes are selected. Then all four MSML agents are applied to the same randomized setup, as well as the unmodified system as baseline in the second experiment. We get one result for every combination of N, k and agent per run. The random sampling is performed a total of 10 times and the results are averaged.
\subsection{Continuous Signal Spotting}
The first experiment aims to explore the impact that k and N have on the performance of the various architectures for spotting keywords in full utterances. The performance is measured in the F$_1$ score, which is the harmonic mean of precision and recall.
\begin{figure}[htb]
\centering
\centerline{\includegraphics[width=.5\textwidth]{fscore_k1_mini.pdf}}
\caption{Spotting F$_1$ scores across different N and agents for k=1}
\label{k1f}
\end{figure}
\begin{figure}[htb]
\centering
\centerline{\includegraphics[width=.5\textwidth]{fscore_k4_mini.pdf}}
\caption{Spotting F$_1$ scores across different N and agents for k=4}
\label{k4f}
\end{figure}
The results on the test set (as seen in figures \ref{k1f} and \ref{k4f}) show Matching and Relation Networks achieving the highest scores. Siamese Networks only take the top spot when given enough supports for each task. The drop in precision and thus in F$_1$ score is severe for an increase in N, however considering that the random baseline decreases from 0.5 down to 0.032, the results are still well above chance. Prototypical Networks do not seem to perform too well overall for this kind of task on the test set, which we suspect is due to some supports disturbing the integrity of the prototype.
\subsection{Speech Recognition Integration}
\begin{table}[thb]
\centering
\begin{tabular}{l c c c c c c c}
\toprule
\hspace{\fill}N = & 1 & 5 & 10 & 15 & 20 & 25 & 30\\
\midrule
\multicolumn{7}{l}{\textit{k = 1}} \\
Siamese & \textbf{12.2} & \textbf{9.1} & \textbf{10.2} & \textbf{12.6} & \textbf{9.3} & \textbf{20.1} & \textbf{13.3} \\
Relation & \textbf{12.2} & \textbf{9.1} & \textbf{10.2} & \textbf{12.6} & \textbf{9.3} & \textbf{20.1} & \textbf{13.3} \\
Proto & \textbf{12.2} & \textbf{9.1} & \textbf{10.2} & \textbf{12.6} & \textbf{9.3} & \textbf{20.1} & \textbf{13.3} \\
Matching & \textbf{12.2} & \textbf{9.1} & \textbf{10.2} & \textbf{12.6} & \textbf{9.3} & 21.0 & 14.3 \\
Vanilla & 16.4 & 10.9 & 12.4 & \textbf{12.6} & \textbf{9.3} & 21.0 & \textbf{13.3} \\
\midrule
\multicolumn{7}{l}{\textit{k = 4}} \\
Siamese & \textbf{11.7} & \textbf{8.8} & \textbf{10.1} & 14.2 & \textbf{6.9} & \textbf{13.2} & \textbf{6.1} \\
Relation & \textbf{11.7} & \textbf{8.8} & \textbf{10.1} & 14.2 & \textbf{6.9} & \textbf{13.2} & \textbf{6.1} \\
Proto & \textbf{11.7} & \textbf{8.8} & \textbf{10.1} & 14.2 & \textbf{6.9} & \textbf{13.2} & \textbf{6.1} \\
Matching & \textbf{11.7} & 9.7 & \textbf{10.1} & \textbf{13.4} & 10.0 & \textbf{13.2} & 8.6\\
Vanilla & 15.2 & 12.4 & 15.1 & 14.2 & \textbf{6.9} & 15.3 & \textbf{6.1} \\
\bottomrule
\end{tabular}
\caption{Word Error Rates in \% across N, k and approaches}
\label{tab:wer}
\end{table}
The second experiment focuses on the ASR performance and the
impact that the meta learning agents’ results have on it. This is
measured using the WER of the ASR with and without the meta
learning agents as baseline (called Vanilla here). Inconsistencies in
the Vanilla WERs originate from the randomized setup to ensure
comparability of the approaches within each N and k. The WER
in this evaluation is only counting mistakes that originate from the
keywords. All other mistakes are ignored, as the meta learning agents
have no effect on them. The reranking in this setup considers all of
the hypotheses that the beam search of the decoder (beamsize = 4)
produces
As can be seen in table \ref{tab:wer}, the meta learning agents help the
ASR slightly in most cases for any N and k. Interestingly there is
very little difference between the meta learning agents. One more
remarkable thing is that the Matching Networks seem to be the only
cases where the WER actually increases over the vanilla system with
no augmentations. This is probably due to them having low precision,
but very high recall in the first experiment, so a lot of false positives.
\section{DISCUSSION}
\noindent \textbf{Impact of Transformer based Embeddings} To begin we want to address insights into the usefulness of the Transformer encoder as embedding function. The results on the development set and the results on the test set for both experiments are remarkably close, even though the data is severely different. This good performance across corpora hints at the Transformer ASR embeddings working as intended. In comparison, when the same MSML nets are trained on Mel frequency buckets directly, the performance across corpora shows a massive drop from development to test set.
\noindent \textbf{The Spotting Process} To provide some insight into the spotting process over a whole signal, figure \ref{sot} shows the Siamese similarity score given a support of the word Prometheus over time. At the point in the sentence where the keyword actually appears, the score peaks over 0.95. In most longer signals however, there are random peaks aside from this main peak. Looking at the transcriptions, unfortunately no specific reason becomes apparent. The margin between an actual peak and a random peak across signals is however big enough for a consistent threshold to be set (we choose 0.8).
\begin{figure}[h]
\centering
\centerline{\includegraphics[width=.48\textwidth]{sot.pdf}}
\caption{Siamese similarity score over frames of an utterance given one support for the keyword Prometheus}
\label{sot}
\end{figure}
\noindent \textbf{Impact of N, k and Support Quality on MSML Agents} Next we
want to give some insights into cases of MSML agent performance
which we find interesting. The MSML agents outperform the random
baseline by a huge margin for any N and k, with the performance decreasing
for large N. The Relation Network manages to perform well
even with few supports and surprisingly even in cases where the only
support it is given is pronounced slightly wrong. The Prototypical and
Siamese Networks greatly benefit from more supports, however the
prototype is sometimes disrupted by heavily mispronounced names
in the supports. This is not the case in the development set, where the
Siamese and the Prototypical approach actually yield the best results,
together with the Relation approach. So robustness against imperfect
supports appears to be important for close-to-real-world scenarios.
\noindent \textbf{Impact of imperfect Speech on ASR} One interesting insight about the keyword spotting ASR interface are the cases where the regular ASR tends to make mistakes. Looking into the data and the hypotheses of the ASR, we find that usually the most probable ASR hypothesis matches exactly what the speaker of the utterance articulated. This is however not always the desired transcription, since speakers tend to be sloppy, especially with names, or mispronounce them entirely.
\noindent \textbf{Impact of the Keyword Spotting Integration} Another interesting observation can be made with the cases where the hypothesis was not changed, despite the MSML agents correctly spotting a keyword. This happens quite frequently for names that follow phonotactics that are very different from English, since they do not appear in the hypotheses at all and could thus also not be reranked. At the same time this is also a good thing, since especially the Matching Network often spotted keywords which were not present. The hypothesis does not change then, because the word in question simply does not appear in the selection. So the ASR hypotheses act as a filter to the MSML agents and the MSML agents act as a filter to the ASR hypotheses.
\section{Conclusions}
In this paper we introduce the idea of using a Transformer ASR encoder as an embedding function, which enables stable good performance across severely different corpora. We also discuss a setup that enables four commonly used MSML approaches to perform keyword recognition on arbitrary reference samples in continuous signals. Finally we introduce a system that combines ASR with keyword spotting to enable modular and simple fine-tuning of neural E2E ASR to expected vocabulary. Experiments show a keyword recognition performance much higher than chance across different setups and an improvement in WER, indicating that the fine-tuning works despite the mechanism being a rather simple proof of concept.
\bibliographystyle{IEEEbib}
|
\section{Introduction}
Randomness in quantum systems has been driving recent progress of quantum information science~\cite{AE2007,D2005,DW2004,GPW2005,ADHW2009,DBWR2010,SDTR2013,HOW2005,HOW2007, AS2004,HLSW2004,S2005,BH2013,KRT2014,KL15,KZD2016,OAGKAL2016,B2018, G2019,BFNV2019,OSH2020,EAZ2005,KLRetc2008,MGE2011,MGE2012,B2018, PhysRevLett.112.240504, PhysRevA.93.012301, garion2020experimental, PhysRevLett.122.200502, PhysRevA.87.030301, PhysRevLett.109.240504, PhysRevLett.109.080505,OWE2019,HROWE2020} as well as fundamental physics~\cite{PSW2006,dRARDV2011,dRHRW2016, HP2007,SS2008,S2011,LSHOH2013,HQRY2016,RY2017,NWK2020,LFSLYYM2019,M2021}.
Theoretically, quantum randomness is often formulated by a unitary drawn uniformly at random, also known as a \emph{Haar} random unitary. However, the Haar randomness is physically unfeasible in large quantum systems.
From the viewpoint of applications, the unitaries that have similar properties of a Haar random unitary are of great importance since they can be used instead of the Haar one. When a random unitary has the same $t$-th order statistics as a Haar random unitary on average, it is called a \emph{unitary $t$-design}. For instance, when a protocol exploits the $t$-th power of the measurement probability after applying a Haar random unitary on any state, the protocol also works even if the Haar random unitary is replaced with a unitary $t$-design.
A unitary $t$-design can be regarded as a quantum generalization of $t$-wise independence~\cite{AE2007}, and have many applications, ranging from communication~\cite{D2005,DW2004,GPW2005,ADHW2009,DBWR2010,SDTR2013,HOW2005,HOW2007}, cryptography~\cite{AS2004,HLSW2004}, algorithms~\cite{S2005,BH2013}, sensing~\cite{KRT2014,KL15,KZD2016,OAGKAL2016}, to potentially quantum supremacy~\cite{B2018, G2019,BFNV2019}.
A unitary $t$-design is also related to another important concept in quantum information science, \emph{epsilon-net}~\cite{OSH2020}, implying more applications yet-to-be-discovered.
Furthermore, the concept of unitary designs has opened a novel research field over quantum information science, quantum thermodynamics, strongly correlated physics, and quantum gravity~\cite{PSW2006,dRARDV2011,dRHRW2016, HP2007,SS2008,S2011,LSHOH2013,HQRY2016,RY2017,NWK2020}.
Experimentally, unitary designs and related methods have been exploited for benchmarking noisy quantum devices~\cite{EAZ2005,KLRetc2008,MGE2011,MGE2012,B2018, PhysRevLett.112.240504, PhysRevA.93.012301, garion2020experimental, PhysRevLett.122.200502, PhysRevA.87.030301, PhysRevLett.109.240504, PhysRevLett.109.080505,OWE2019,HROWE2020}, realizing quantum supremacy~\cite{G2019}, demonstrating quantum chaos and quantum holography~\cite{LFSLYYM2019,M2021}.
It is also worthwhile to mention that unitary designs have been studied in combinatorial mathematics~\cite{DGS1975,DGS1977,RS2009,R2010,RS2011,BNRT2020}. Hence, developing the theory of unitary designs is of substantial interest in a wide range of science, both theoretically and experimentally.
An important question about unitary $t$-designs is how to implement them by quantum circuits.
Many implementations of unitary $2$-designs, both approximate and exact ones, were proposed~\cite{DLT2002,BWV2008a,WBV2008,GAE2007,TGJ2007,DCEL2009,HL2009,DJ2011,BWV2008a,WBV2008,CLLW2015,NHMW2015-1}. In contrast, only \emph{approximate} implementations of unitary $t$-designs for general $t$ were known~\cite{HL2009TPE,BHH2016,NHKW2017,HM2018,HMHEGR2020}. Explicit constructions of \emph{exact} unitary designs were left open except special cases~\cite{RS2009,BNRT2020,BNZZ2019}.
Approximate ones typically suffice in applications, but exact designs are more preferable in certain protocols especially when they are used multiple times in a single run of the protocol. If this is the case, the error from each approximate implementation accumulates and eventually spoils the protocol.
One of such protocols is a randomized benchmarking (RB) protocol~\cite{EAZ2005,KLRetc2008,MGE2011}, a standard method for experimentally estimating quantum noise, where unitary $2$-designs are used multiple times. Although the RB is experimentally-friendly and is widely used in various experimental systems, it reveals only the average gate fidelity. To obtain more information about the noise, a number of variants were proposed and experimentally implemented (see, e.g., Ref.\,\cite{HHFFW2019} and the references therein), which are all based on $2$-designs. It is highly expected that, by using higher-designs, much more information about the noise in quantum systems can be extracted. To this end, explicit constructions of exact unitary $t$-designs are important.
Constructing exact designs is, however, by far non-trivial. The difficulty is illustrated by a spherical $t$-design, a random real unit vector analogous to a unitary $t$-design.
The existence of exact spherical $t$-designs was proven in a non-constructive manner more than three decades ago~\cite{SZ1984}. Since then, more concise proofs and explicit constructions have been under intense investigation in combinatorial mathematics (see e.g., Refs.\,\cite{RB1991,WV1991,BB2009,BRV2013,CXX2019} and the references therein). In particular, it was only recently that constructions in general cases~\cite{X2020} and explicit constructions, in the sense that all the algorithms are given in a computable form~\cite{BNOZ2020}, were proposed.
Finding explicit constructions of exact unitary designs, since they are more complicated than spherical designs, is a rather non-trivial problem.
In this paper, we provide for the first time an explicit quantum circuit that generates an exact unitary $t$-design for any $t$ on the arbitrary number $N$ of qubits.
More specifically, we show that an exact unitary $t$-design on $d$-dimensional Hilbert space, i.e., \emph{qudit}, can be generated from those on smaller spaces, which is obtained based on the recent mathematical results by some of the authors~\cite{BNOZ2020}.
Using this result, we provide an inductive construction of quantum circuits for exact unitary $t$-designs on $N$ qubits: we first construct a unitary $t$-design on a single qubit and then extend it to $N$ qubits. Unfortunately, the circuit fails to be efficient, but is still of practical use when the size of the system is small.
As an application of exact unitary designs, we introduce the \emph{$t$-th order RB}, or the \emph{$t$-RB} for short, that harnesses the power of exact unitary $2t$-designs. The standard RB corresponds to the $1$-RB. The $t$-RB enables us to experimentally characterize the higher order properties of quantum noises in the manner free from state-preparation and measurement (SPAM) errors.
We especially investigate the $2$-RB in detail and show that it reveals \emph{self-adjointness} of the noise in the system. The self-adjointness is a new metric of the noise related to the feasibility of quantum error correction (QEC): small self-adjointness implies that the noise cannot be approximated by any stochastic Pauli noise.
The noise on the system being stochastic Pauli is desirable both in theory and in practice. Stochastic Pauli noises are the commonly-used noise models in theoretical studies of QEC, since they are easy to numerically handle, and the properties of QEC, such as error thresholds and logical error rates, for stochastic Pauli noises are well-understood. Also, there is a practical advantage if the noise on the system is stochastic Pauli since they can be corrected simply by applying Pauli operators, making the error correcting scheme easier in general.
After numerically demonstrating the feasibility of the $2$-RB, we perform the $2$-RB in a superconducting system and estimate the self-adjointness of background noise, showing that the $2$-RB experiments are feasible. From the experiment, we find that the interactions with adjacent qubits especially decrease the self-adjointness, which may lead to degradation of the performance of QEC with standard decoders. Hence, either improving the system or extending the noise model in theoretical studies of QEC, or both, is important for further experimental developments of quantum information technology.\\
This paper is organized as follows. In Sec.\,\ref{S:design}, we provide a general introduction of unitary $t$-designs.
Our main results are summarized in Sec.\,\ref{S:main1} for the quantum-circuit construction of exact unitary $t$-designs, and in Sec.\,\ref{S:main2} for the $t$-RB protocols.
A summary of the experiment of the $2$-RB is provided in Sec.\,\ref{S:Main3}.
After we explain the structure of the remaining paper in Sec.\,\ref{S:structure}, we provide a proof of the explicit construction in Sec.\,\ref{Sec:ConstDesign} and the theory of the $t$-RB in Sec.\,\ref{S:higher-orderRB}. The details of the experiment are provided in Sec.\,\ref{S:NumExpDetail}.
We conclude our paper with summary and discussions in Sec.\,\ref{S:SD}. Technical statements are provided in Appendices.
\section{Unitary $t$-designs} \label{S:design}
Let ${\sf U}(d)$ be the unitary group of degree $d < \infty$. The Haar measure ${\sf H}$ on ${\sf U}(d)$ is the unique unitarily invariant measure on the unitary group, i.e., it satisfies
\begin{multline}
\forall \mathcal{W} \subset {\sf U}(d), \forall V \in {\sf U}(d), \\
{\sf H}(V \mathcal{W}) = {\sf H} (\mathcal{W}V) = {\sf H}(\mathcal{W}).
\end{multline}
When it is needed to clarify the degree of the unitary group, we denote the Haar measure by ${\sf H}(d)$.
A unitary $t$-design ${\sf U}_t(d)$ is defined by a finite set of unitaries that mimics the $t$-th order statistical moment of the Haar measure ${\sf H}$.
Amongst several equivalent definitions~\cite{L2010}, we here adopt the following definition.
\begin{Definition}[Unitary $t$-design] \label{Def:des}
For $t \in \mathbb{Z}^+$, a finite set ${\sf U}_{t}(d) \subset {\sf U}(d)$ of unitaries is a unitary $t$-design if
\begin{equation}
\mathbb{E}_{U \sim {\sf U}_{t}(d)} [U^{\otimes t} \otimes U^{\dagger \otimes t}]
=
\mathbb{E}_{U \sim {\sf H}(d)} [U^{\otimes t} \otimes U^{\dagger \otimes t}], \label{Eq:tDesign}
\end{equation}
where $\mathbb{E}_{U \sim {\sf U}_{t}(d)}$ is a uniform average over ${\sf U}_t(d)$, and $\mathbb{E}_{U \sim {\sf H}(d)}$ is the average over the Haar measure.
\end{Definition}
From an operational viewpoint, this definition implies that a unitary $t$-design cannot be distinguished from a Haar random unitary on average even when $t$ copies of the unitary are given. To clasify this, let us define a quantum operation, i.e., a completely-positive and trace-preserving (CPTP) map, ${\cal G}^{\mu}_t$ by
\begin{equation}
{\cal G}_{\mu}^{(t)}(\varrho) := \mathbb{E}_{U \sim \mu} \bigl[ U^{\otimes t} \varrho U^{\otimes t \dagger} \bigr], \label{Eq:gg}
\end{equation}
for any quantum state $\varrho$ on $t$ qudits, where $\mu$ is either the Haar measure ${\sf H}(d)$ on a qudit or a uniform measure over a unitary $t$-design ${\sf U}_t(d)$. Then, we can show that Definition~\ref{Def:des} is equivalent to that (see e.g.\,\cite{L2010})
\begin{equation}
{\cal G}^{(t)}_{\sf H} = {\cal G}^{(t)}_{{\sf U}_t}. \label{Eq:g}
\end{equation}
This implies that, in \emph{any} experiments that use $t$ copies of a random unitary, no difference will be observed on average when a $t$-design is used instead of the Haar one.
For instance, let us consider the probability distribution $\{ p_i(U) := \tr[ P_i U \rho U^{\dagger}] \}$ when a one-qudit state is measured by a given POVM $\{P_i \}_i$ after the application of a unitary $U$. By setting the $t$-qudit state $\varrho$ in Eq.\,\eqref{Eq:gg} to $\rho^{\otimes t}$ and using Eq.\,\eqref{Eq:g}, it follows that, for any $s = 1, \dots, t$,
\begin{equation}
\mathbb{E}_{U \sim {\sf H}} \biggl[ \prod_{r=1}^s p_{i_r} (U) \biggr] = \mathbb{E}_{U \sim {\sf U}_t} \biggl[ \prod_{r=1}^s p_{i_r} (U) \biggr].
\end{equation}
Thus, the distribution of the measurement outcomes for a Haar random unitary and that for an unitary $t$-design exactly coincide up to the $t$-th order on average. Note that this is merely an example, and Eq.\,\eqref{Eq:g} implies much more: a Haar random unitary cannot be differentiated from a unitary $t$-design even by more complicated experiments over $t$ qudits.
The existence of an exact unitary $t$-design on ${\sf U}(d)$ for any $t$ and $d$ follows from the Carath\'{e}odoty's theorem and the fact that the dimension of the space, on which $U^{\otimes t} \otimes U^{\dagger \otimes t}$ is defined, is finite. Note however that the proof indicates only the existence of an exact unitary $t$-design. How to explicitly construct an exact unitary $t$-design has been a highly non-trivial problem.
In our construction, it is convenient to introduce a \emph{strong} unitary $t$-design.
\begin{Definition}[Strong unitary $t$-design]
For $t \in \mathbb{Z}^+$, a finite set ${\sf U}_{\leq t}(d)$ of unitaries on ${\sf U}(d)$ is called a strong unitary $t$-design if
\begin{equation}
\mathbb{E}_{U \sim {\sf U}_{\leq t}(d)} [U^{\otimes r} \otimes U^{\dagger \otimes s}]
=
\mathbb{E}_{U \sim {\sf H}} [U^{\otimes r} \otimes U^{\dagger \otimes s}],
\end{equation}
for $0 \leq \forall r \leq t$ and $0 \leq \forall s \leq t$.
\end{Definition}
Clearly, a strong unitary $t$-design is a unitary $t$-design.
Unlike standard unitary designs, strong unitary designs do not have a clear operational interpretation in quantum information processing, but we use it in the intermediate step of our construction.
\section{Main result 1 -- quantum circuits for exact unitary designs --} \label{S:main1}
In this section, we provide explicit constructions of strong unitary $t$-designs for any $t$. In particular, a quantum circuit for a strong unitary $t$-design on $N$ qubits is provided.
We start with preliminaries in Subsec.\,\ref{SS:Pre}, and provide the construction in Subsec.\,\ref{SS:InductiveConst}. We then comment on the circuit complexity of the construction in Subsec.\,\ref{SS:Efficiency}.
\subsection{Preliminaries} \label{SS:Pre}
Unitary designs have been studied in terms of representation theory~\cite{RS2009,RS2011} since the operator $U^{\otimes t} \otimes U^{\dagger \otimes t}$ in the definition can be regarded as a representation of the unitary group. Our construction is based on representation theory, where irreducible decomposition of the operator plays an important role. A brief introduction of irreducible representations (irreps) will be provided in Section~\ref{SS:unitaryrep}. Here, we mention a couple of well-known facts that are necessary to state our main result.
Any irrep of the unitary group can be indexed by a non-increasing integer sequence $\lambda :=(\lambda_1, \lambda_2, \dots, \lambda_d)$ of length $d$, i.e., $\lambda_i \in \mathbb{Z}$ for $i=1,\dots, d$, and $\lambda_1 \geq \lambda_2 \geq \dots \geq \lambda_d$~\cite{B2004}. In particular, \emph{spherical representations} of ${\sf U}(d)$ with respect to ${\sf K}:= {\sf U}(d_1) \times {\sf U}(d - d_1)$ are of great importance in the construction. Let $\Lambda_{\rm sph}(d_1,d,t)$ be a set of all non-increasing integer sequences $\lambda$ in the form of
\begin{equation}
\lambda = (\lambda_1, \dots, \lambda_{d_1}, 0, \dots, 0, -\lambda_{d_1}, \dots, -\lambda_{1}),
\end{equation}
where $d_1 \leq d/2$ and $t \geq \lambda_1 \geq \dots \geq \lambda_{d_1} \geq 0$. The spherical representation is the irrep indexed by $\lambda \in \Lambda_{\rm sph}(d_1,d,t)$~\cite{GW2009}. For a spherical representation $\lambda \in \Lambda_{\rm sph}(d_1,d,t)$, a \emph{zonal spherical function} $Z^{(d_1)}_{\lambda}(z_1, \dots, z_{d_1})$ is defined by the unique bi-${\sf K}$-invariant function~\cite{JC1974,R2010,BNOZ2020}. The zonal spherical functions are a certain type of symmetric polynomials, and can be explicitly written down (see, e.g., Appendix A of Ref.\,\cite{BNOZ2020}).
\subsection{Inductive constructions} \label{SS:InductiveConst}
Our main technical result is to construct a strong unitary $t$-design on ${\sf U}(d)$ from those on ${\sf U}(d_1)$ and on ${\sf U}(d-d_1)$.
\begin{Theorem} \label{Thm:General}
Let $d_1$ be a positive integer such that $d_1 \leq d/2$. Define a set of unitaries ${\sf W}_{d_1 \oplus d-d_1}$ in ${\sf U}(d)$ by
\begin{equation}
{\sf W}_{d_1 \oplus d-d_1}:= \{ U \oplus V | U \in {\sf U}_{\leq t}(d_1), V \in {\sf U}_{\leq t}(d-d_1)\},
\end{equation}
where ${\sf U}_{\leq t}(d_1)$ and ${\sf U}_{\leq t}(d-d_1)$ are strong unitary $t$-designs on ${\sf U}(d_1)$ and ${\sf U}(d-d_1)$, respectively.
Let $\boldsymbol{\theta}_{\lambda} := (\theta_{\lambda}^{(0)}, \dots, \theta_{\lambda}^{(d_1-1)})$ $(\theta_{\lambda}^{(i)} \in [0, \pi/2])$ be such that
\begin{equation}
Z_{\lambda}^{(d_1)}\bigl( \cos^2 \theta_{\lambda}^{(0)}, \dots, \cos^2 \theta_{\lambda}^{(d_1-1)} \bigr) = 0,
\end{equation}
where $Z_{\lambda}^{(d_1)}$ is the zonal spherical function for $\lambda \in \Lambda_{\rm sph}(d_1,d,t)$.
Let $R_{\lambda}$ be a unitary defined by
\begin{equation}
R_{\lambda} =
\begin{pmatrix}
C(\boldsymbol{\theta}_{\lambda}) & i S(\boldsymbol{\theta}_{\lambda}) & 0 \\
i S(\boldsymbol{\theta}_{\lambda}) & C(\boldsymbol{\theta}_{\lambda}) & 0 \\
0 & 0 & I_{d-2d_1}
\end{pmatrix},
\end{equation}
where $C(\boldsymbol{\theta}_{\lambda}) = {\rm diag}(\cos\theta_{\lambda}^{(0)}, \dots, \cos\theta_{\lambda}^{(d_1-1)})$ and $S(\boldsymbol{\theta}_{\lambda}) = {\rm diag}(\sin\theta_{\lambda}^{(0)}, \dots, \sin\theta_{\lambda}^{(d_1-1)})$, and $I_{d-2d_1}$ is the identity matrix of size $d- 2d_1$.
Then,
\begin{equation}
{\sf W}_d := {\sf W}_{d_1 \oplus d-d_1} \prod_{\lambda \in \Lambda_{\rm sph}(d_1,d,t)} (R_{\lambda} {\sf W}_{d_1 \oplus d-d_1}) \label{Eq:W}
\end{equation}
is a strong unitary $t$-design on ${\sf U}(d)$.
\end{Theorem}
Theorem~\ref{Thm:General} follows from a more general result~\cite{BNOZ2020} shown by some of the authors, which works not only for the unitary group but also for a broader class of compact groups. For the sake of completeness, we provide a direct proof of Theorem~\ref{Thm:General} in Sec.\,\ref{Sec:ConstDesign}.
We then claim that
\begin{equation}
{\sf W}_1=\{ 1, \omega, \omega^2, \dots, \omega^t \}, \label{Eq:W_1}
\end{equation}
where $\omega=\exp[\frac{2\pi}{t+1}]$ is the $(t+1)$-th root of unity, is a strong unitary $t$-design on ${\sf U}(1)$ for any $t$. This is easily checked by direct calculations:
\begin{align}
&\mathbb{E}_{U \sim {\sf W}_1}[ U^{\otimes r} \otimes U^{\dagger \otimes s} ]
=\frac{1}{t+1} \sum_{z \in {\sf W}_1} z^{r} \bar{z}^s = \delta_{rs},\\
&\mathbb{E}_{U \sim {\sf H}(1)}[ U^{\otimes r} \otimes U^{\dagger \otimes s} ]
=\int_{U(1)} z^r \bar{z}^s dz=\delta_{rs},
\end{align}
where $\delta_{rs}$ is the Kronecker delta.
Hence, we have $\mathbb{E}_{U \sim {\sf W}_1}[ U^{\otimes r} \otimes U^{\dagger \otimes s} ] = \mathbb{E}_{U \sim {\sf H}(1)}[ U^{\otimes r} \otimes U^{\dagger \otimes s} ]$ for $0 \leq \forall s, r \leq t$, implying that ${\sf W}_1$ is a strong unitary $t$-design.
From Theorem~\ref{Thm:General} and ${\sf W}_1$, a strong unitary $t$-design on a qudit can be inductively constructed.
\begin{Corollary}\label{Cor:tDesign_qudit}
For $d \geq 1$, let ${\sf W}_{1 \oplus d-1}$ be a set of unitaries given by
\begin{equation}
{\sf W}_{1 \oplus d-1}= \{ z \oplus V | z \in {\sf W}_1, V \in {\sf U}_{\leq t}(d-1)\},
\end{equation}
where ${\sf W}_1 = \{ 1, \omega, \dots, \omega^t \}$ with $\omega$ being the $(t+1)$-th root of unity, and $\theta_{\lambda} \in [0, \pi/2]$ be such that
\begin{equation}
Z^{(1)}_{\lambda}(\cos^2 \theta_{\lambda}) =0.
\end{equation}
Using a unitary $R_{\lambda} = e^{i \theta_{\lambda} X} \oplus I_{d-2}$, where $X$ is the Pauli-$X$ operator, we obtain that
\begin{equation}
{\sf W}_{1 \oplus d-1} \prod_{\lambda \in \Lambda_{\rm sph}(1,d,t)} (R_{\lambda} {\sf W}_{1 \oplus d-1})
\end{equation}
is a strong unitary $t$-design on a qudit.
\end{Corollary}
In this construction, it is important to obtain zeros for the zonal spherical functions $Z_{\lambda}^{(1)}$. This is computationally feasible since they are polynomials of one variable and are explicitly given (see Appendix A of Ref.\,\cite{BNOZ2020}). Furthermore, $\Lambda_{\rm sph}(1,d,t)$ contains only $t$ elements. Hence, we need to solve $t$ polynomials with one variable, which is tractable as far as $t$ is not too large.\\
We now consider a strong unitary $t$-design on $N$ qubits.
Again using Theorem~\ref{Thm:General}, we obtain the quantum circuit on $(N+1)$ qubits based on that on $N$ qubits. See also Fig.\,\ref{Fig:QCforD}.
\begin{Corollary}\label{Cor:tDesign_qubit}
Let ${\sf Q}_{N}$ be a strong unitary $t$-design on $N$ qubits, and ${\rm Ctrl \mathchar`-}{\sf Q}_N$ be a set of controlled-unitaries on $N+1$ qubits, defined by
\begin{equation}
{\rm Ctrl \mathchar`-}{\sf Q}_N := \{\ketbra{0}{0} \otimes U_0 + \ketbra{1}{1} \otimes U_1 : U_0, U_1 \in {\sf Q}_N \}.
\end{equation}
For $\lambda \in \Lambda_{\rm sph}(D,2D,t)$, where $D = 2^N$, let $\boldsymbol{\theta}_{\lambda}:=(\theta_{\lambda}^{(0)}, \dots, \theta_{\lambda}^{(D-1)})$ be such that
\begin{equation}
Z^{(D)}_{\lambda}( \cos^2 \theta_{\lambda}^{(0)}, \dots, \cos^2 \theta_{\lambda}^{(D-1)}) = 0. \label{Eq:ZPF_qubits}
\end{equation}
Representing $\{0, \dots, D-1\}$ in binary form such as $\{\boldsymbol{j} \}_{\boldsymbol{j} \in \{ 0,1\}^N}$, we write $\theta_{\lambda}^{(j)}$ as $\theta_{\lambda}^{(\boldsymbol{j})}$. Let $R_X(\boldsymbol{\theta}_{\lambda})$ be a single-qubit $X$-rotation controlled by $N$ qubits, defined by
\begin{equation}
R_X(\boldsymbol{\theta}_{\lambda}) = \sum_{\boldsymbol{j} \in \{0, 1\}^N} e^{i \theta_{\lambda}^{(\boldsymbol{j})} X} \otimes \ketbra{\boldsymbol{j}}{\boldsymbol{j}}.
\end{equation}
Then,
\begin{equation}
{\rm Ctrl \mathchar`-}{\sf Q}_N \prod_{\lambda \in \Lambda_{\rm sph}(D,2D,t)} \bigr( R_X(\boldsymbol{\theta}_{\lambda}) {\rm Ctrl \mathchar`-}{\sf Q}_N \bigl)
\end{equation}
is a strong unitary $t$-design on $N+1$ qubits.
\end{Corollary}
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{QCforD.pdf}
\caption{The quantum circuit that generates an exact unitary $t$-design on $N+1$ qubits from those on $N$ qubits. The unitary ${\sf Q}_N$ is a quantum circuit for an exact unitary $t$-design on $N$ qubits. The gate $X(\boldsymbol{\theta}_{\lambda})$ is the single-qubit $X$-rotation controlled by the other $N$ qubits, which corresponds to $R_X(\boldsymbol{\theta}_{\lambda})$ in the main text. Note that this gate can be decomposed into a sequence of two-qubit gates using a classical oracle that provides $\theta_{\lambda}^{(\boldsymbol{j})}$ from $\boldsymbol{j}$.
The rotation angles $\boldsymbol{\theta}_{\lambda}$ are obtained by solving $Z_{\lambda}^{(D)}=0$, where $Z_{\lambda}^{(D)}$ is the zonal spherical function for the spherical representation $\lambda \in \Lambda_{\rm sph}(D, 2D, t)$ with $D = 2^N$. The number of the controlled-$X$ rotations is $|\Lambda_{\rm sph}(D, 2D, t)|$.
By the inductive use of this quantum circuit in terms of $N$, we can decompose the circuit to that consisting only of two-qubit gates.
}
\label{Fig:QCforD}
\end{figure}
Corollary~\ref{Cor:tDesign_qubit} implies that a quantum circuit for an exact unitary $t$-design can be inductively constructed from a strong unitary $t$-design on one qubit, i.e., that on ${\sf U}(2)$.
Furthermore, a strong unitary $t$-design on ${\sf U}(2)$ can be constructed using Corollary~\ref{Cor:tDesign_qudit}. Thus, combining Corollaries~\ref{Cor:tDesign_qudit} and~\ref{Cor:tDesign_qubit}, we obtain a quantum circuit for an exact unitary $t$-design for any $t$ and on an arbitrary number of qubits.
Note that the circuit, constructed in this way, can be explicitly decomposed into two-qubit gates. The controlled unitary ${\rm Ctrl \mathchar`-}{\sf Q}_N$ part contains up to three-qubit gates, if the circuit ${\sf Q}_N$ on $N$ qubit is already decomposed into two-qubit gates. The three-qubit gates can be easily rewritten as a series of two-qubit gates.
Also, the $X$-rotation controlled by $N$ qubits, $R_X(\boldsymbol{\theta}_{\lambda})$, can be decomposed into a sequence of two-qubit gates of polynomial length using sufficiently many number of ancillary qubits, which is based on a classical oracle that computes the angle $\theta_{\lambda}^{(\boldsymbol{j})}$ from $\boldsymbol{j}$ (see Appendix~\ref{App:RX}).
In special cases, we can find a much more concise construction based on a similar technique.
\begin{Proposition}~\label{Prop:Clliford4}
Let ${\sf C}(4)$ be the Clifford group on $2$ qubits. There exists a fixed two-qubit unitary $U_c$, such that ${\sf C}(4) U_c {\sf C}(4)$ is an exact unitary $4$-design on $2$ qubits.
\end{Proposition}
Analytically, we can prove that there exist unitaries $V_1$ and $V_2$ such that ${\sf C}(4) V_1 {\sf C}(4) V_2 {\sf C}(4)$ is an exact unitary $4$-design on $2$ qubits\,\cite{BNOZ2020}. Also, an algorithm for computing the unitaries $V_1$ and $V_2$ is given. It, however, turns out from numerics that it is not necessary to apply two extra unitaries if we choose a proper unitary $U_c$, which leads to Proposition\,\ref{Prop:Clliford4}. An explicit form of the unitary $U_c$ is numerically obtained and is provided in Appendix~\ref{App:4DesClifford}. Note that the existence of $U_c$ is confirmed numerically, so the statement holds up to the numerical precision.
This construction is only for a $4$-design on $2$ qubits, but the number of unitaries in the $4$-design is much smaller than that of Corollary~\ref{Cor:tDesign_qubit}.
It is an open problem whether a similar construction works for higher-designs on a larger number of qubits.
\subsection{Efficiency and comparison with a Haar unitary} \label{SS:Efficiency}
To quantitatively evaluate the complexity of the quantum circuit for an exact unitary $t$-design obtained in Corollary~\ref{Cor:tDesign_qubit}, we provide an order estimate of the number $G(N)$ of two-qubit gates in the circuit. Assuming $2^N \gg t$ and using the fact that $|\Lambda_{\rm sph}(D, 2D, t)|= O(e^{\pi \sqrt{2t/3}})$ due to the Hardy and Ramanujan formula for the asymptotics of the number of partitions, we obtain
\begin{equation}
G(N) \approx \exp \biggl[\pi \sqrt{\frac{2t}{3}} (N-1) \biggr], \label{Eq:G}
\end{equation}
to the leading order of $N$. Hence, it is necessary to use exponentially many two-qubit gates as the number of qubits increases. This inefficiency of the quantum circuit may be intrinsic since the construction is inductive.
There is another source of inefficiency. In Corollary~\ref{Cor:tDesign_qubit}, it is necessary to find zeros of zonal spherical functions (see Eq.\,\eqref{Eq:ZPF_qubits}) for all $\lambda \in \Lambda_{\rm sph}(D, 2D, t)$. The zonal spherical function is given in terms of the summation of the (normalized) Schur polynomials (see Appendix A in Ref.\,\cite{BNOZ2020}).
It is unlikely that the Schur polynomials have polynomial size algebraic formulas in general~\cite{SchurPoly2019}. Moreover, the number of variables for each $Z_{\lambda}^{(D)}$ is $D=2^N$. Hence, finding zeros of zonal spherical functions is computationally intractable.
In total, the construction for an exact unitary $t$-design on a large number of qubits is inefficient from both quantum-circuit and classical-computation viewpoints. We, however, think that our construction and our proof technique will form a solid basis of searching more efficient constructions of exact, as well as approximate, unitary designs.
We also emphasize that, despite its inefficiency, our construction is of practical use on a few-qubit system, as we seek in the following sections.
Before we conclude this section, we comment on advantages of our construction of an exact unitary $t$-design over a direct implementation of a Haar random unitary.
A naive way of implementing a Haar random unitary by a quantum circuit consists of three steps. First, we sample a Haar random unitary as a matrix by a classical computer. A classical algorithm for this is known~\cite{M2007}, but it is trivially inefficient since the size of the matrix is exponentially large. We then classically compute a decomposition of the unitary matrix into a sequence of two-qubit unitaries, providing a classical description of a quantum circuit for the unitary. This step is also inefficient, and the resulting quantum circuit is almost surely composed of the exponentially many number of two-qubit gates. Finally, we implement the circuit in practice.
This quantum circuit for a Haar random unitary is inefficient in terms of the number of qubits, and so, cannot be of practical use in a large system. Even in a small system, this naive implementation has a crucial difficulty that, every time a unitary is sampled, the above protocol outputs a quantum circuit with a rather different sequence of various two-qubit gates. This implies that, in each sampling, one needs to significantly modify the quantum circuit. This is in a sharp contrast to our quantum circuit for a unitary $t$-design based on Corollary~\ref{Cor:tDesign_qubit} since it has a \emph{fixed} structure. In each sampling, only what one needs to do is to randomly choose single-qubit gates, or more precisely elements of ${\sf U}(1)$ from ${\sf W}_1$ (see Eq.~\eqref{Eq:W_1}), and to plug them into the quantum circuit with a fixed structure. This will help practical implementations of the circuit in small systems.
It should be also noted that the single-qubit gates in our construction can be sampled from a discrete set, though sampling from a continuous gate set is necessary in the direct implementation of a Haar random unitary. This is another advantage of our construction.
\section{Main result 2 -- Higher-Order Randomized Benchmarking --} \label{S:main2}
We here introduce a higher-order generalization of the standard RB that uses exact unitary $2t$-designs. We call it the \emph{$t$-th order} RB, or simply $t$-RB. The standard RB corresponds to $1$-RB. From the higher-order RB, more information about the noise can be extracted. In particular, we show that a new characterization of the noise, which we call \emph{self-adjointness}, can be estimated from the $2$-RB.
Before we proceed, we emphasize that exact unitary designs, not approximate ones, are of crucial importance in the RB-type protocols. This is because the protocol uses unitary designs multiple times. Hence, if each unitary has an error due to the approximation, it accumulates in the whole process and results in a large error at the end.
Since the goal of the RB-type protocol is typically very high, such as benchmarking the fidelity $> 95\%$, the error originated from the approximate designs would spoil the protocol. Hence, the use of exact unitary designs is of key importance. This point is more elaborated on in Subsec.\,\ref{SS:ExactImportant}.\\
In Subsec.\,\ref{SS:defdef}, we overview a couple of metrics of the noise, i.e., the average fidelity and unitarity, and introduce the self-adjointness.
The importance of the self-adjointness in QEC is argued in Subsec.\,\ref{SS:impSA}. We then introduce the $t$-RB in Subsec.\,\ref{SS:tRB}.
We argue the importance of exact designs in more detail in Subsec.\,\ref{SS:ExactImportant}.
We focus on the $2$-RB in Subsec.\,\ref{SS:2RB} and show that the self-adjointness and the unitarity of the noise can be estimated from the $2$-RB at the same time.
We briefly comment on the scalability of the $t$-RB in Subsec.\,\ref{SS:Scalability}.
\subsection{Characterizing noises} \label{SS:defdef}
A noise $\mathcal{E}$ acting on a $q$-qubit system is formulated by a completely-positive and trace-preserving (CPTP) map. Let $d$ be defined as $d:=2^q$.
The average fidelity and the unitarity are defined by
\begin{align}
&F(\mathcal{E}) := \int d \varphi \bra{\varphi} \mathcal{E} \bigl( \ketbra{\varphi}{\varphi} \bigr) \ket{\varphi},\\
&u(\mathcal{E}) := \frac{d}{d-1} \int d \varphi \bigl|\! \bigr| \mathcal{E}' ( \ketbra{\varphi}{\varphi} ) \bigl|\! \bigr|_2^2,
\end{align}
respectively, where $\mathcal{E}'(\rho) := \mathcal{E}(\rho - I/d)$, and $|\!| A |\!|_2 = (\tr[A^{\dagger} A])^{1/2}$ is the Schatten 2-norm.
The average fidelity satisfies $1/(d+1) \leq F(\mathcal{E}) \leq 1$, and $F(\mathcal{E})=1$ if and only if the system is noiseless, i.e., $\mathcal{E}$ is the identity channel, while the unitarity satisfies $0 < u(\mathcal{E}) \leq 1$, and $u(\mathcal{E}) = 1$ if and only if the noise is coherent, i.e., $\mathcal{E}$ is a unitary channel.
The unitarity is an important metric in the context of QEC since coherent noise is known to be hard to correct in general~\cite{KLDF2016,SWS2015,SFK2017}.
In the RB-type protocols, it is more natural to use a \emph{fidelity parameter} $f(\mathcal{E})$ rather than the average fidelity itself. It is defined by
\begin{equation}
f(\mathcal{E}) = \frac{d F(\mathcal{E}) - 1}{d-1},
\end{equation}
and satisfies $-1/(d^2-1) \leq f(\mathcal{E}) \leq 1$.
We next introduce a self-adjointness of the noise. For any linear map $\mathcal{E}$, an \emph{adjoint} map $\mathcal{E}^{\dagger}$ is defined by $\tr[ A \mathcal{E}(B)] =\tr[ \mathcal{E}^{\dagger}(A) B]$. A noise $\mathcal{E}$ is called self-adjoint if $\mathcal{E} = \mathcal{E}^{\dagger}$, which is equivalent to that all the Kraus operators of $\mathcal{E}$ are self-adjoint.
The \emph{self-adjointness} $H(\mathcal{E})$ of the noise $\mathcal{E}$ is defined by
\begin{equation}
H(\mathcal{E}) := 1- \frac{d+1}{2d} \int d \varphi \bigl|\! \bigr| \mathcal{E}( \ketbra{\varphi}{\varphi}) - \mathcal{E}^{\dagger} ( \ketbra{\varphi}{\varphi}) \bigl|\! \bigr|_2^2.
\end{equation}
The normalization constant is chosen such that $0 \leq H(\mathcal{E}) \leq 1$. Obviously, $H(\mathcal{E}) = 1 $ if and only if $\mathcal{E}$ is self-adjoint, i.e. $\mathcal{E}=\mathcal{E}^{\dagger}$.
Note that the self-adjointness has two contributions from the noisy map $\mathcal{E}$, one is from the unital part and the other from the non-unital part. The non-unital part of the noise makes the self-adjointness less than one since, if $\mathcal{E}$ is not unital, then $\mathcal{E}^{\dagger}$ is not trace-preserving, which implies that $\mathcal{E} \neq \mathcal{E}^{\dagger}$.
To clearly separate the two contributions, we introduce a \emph{self-adjointness parameter} $h(\mathcal{E})$. Using $\mathcal{E}'(\rho) = \mathcal{E}(\rho- I/d)$, we defined it by
\begin{equation}
h(\mathcal{E}) := \frac{d}{d-1} \int d \varphi \tr[\mathcal{E}'(\varphi) \mathcal{E}'^{\dagger}(\varphi)].
\end{equation}
The self-adjointness parameter $h(\mathcal{E})$ is related to the self-adjointness $H(\mathcal{E})$ and the unitarity $u(\mathcal{E})$ by
\begin{align}
&H(\mathcal{E}) = 1 - \frac{d^2-1}{d^2}\bigl( u(\mathcal{E}) - h(\mathcal{E}) \bigr) - \frac{d+1}{2d^2} | \alpha_{\mathcal{E}} |^2, \label{Eq:66}
\end{align}
where $| \alpha_{\mathcal{E}} |$ is a measure of the non-unital part of the noise (see Subsec.\,\ref{SS:L} for the definition). We can clearly observe that $H(\mathcal{E})$ consists of two factors, the unital part $h(\mathcal{E})$ and the non-unital part $|\alpha_{\mathcal{E}}|$.
The three metrics of noises, namely, fidelity, unitarity, and self-adjointness, all capture different properties of the noises. The fidelity reveals the first order property of the noises, while the unitarity and the self-adjointness, which are independent to each other, reveal the second-order.
In order to improve noisy quantum devices, it is of crucial importance to obtain the information of noise as much as possible. Hence, it is certainly of practical use to introduce the self-adjointness as a new metric of noise. In addition, we argue in the next subsection that the self-adjointness has important implications for QEC.
\subsection{Importance of self-adjointness in QEC} \label{SS:impSA}
The most important family of self-adjoint noises is stochastic Pauli noises, whose Kraus operators are all proportional to Pauli matrices. In QEC, Pauli noises are the standard yet most important class of noises both in theory and in practice. From a theoretical perspective, Pauli noises are easy to numerically handle. Hence, most numerical calculations have been carried out by assuming Pauli noises, and it has been confirmed that QEC has preferable features, such as exponential decreases and threshold behaviors of logical error rates, if the noise is Pauli.
The noise being Pauli is also practically preferable in experimental realizations of QEC since it typically simplifies the decoding tasks.
This is especially the case for stabilizer codes, such as surface and color codes, whose standard decoders are to estimate what types of Pauli operators should be applied on which physical qubits during recovery operations.
For stochastic Pauli noises, if the estimation goes well, the state is fully retrieved with high probability by applying Pauli operators to the suitable physical qubits. In contrast, it is not possible to fully correct non-Pauli noises by applying Pauli operators since they generate undesired coherence between different code spaces. Thus, QEC of non-Pauli noises generally suffers from degradation of logical error rates when the standard decoders are used~\cite{SFK2017,BEKP2018} or requires more complicated algorithms for retrieving the performance of QEC. Neither of them is preferable in practice since it induces additional experimental difficulties.
For these reasons, it is desirable to check that the noise on an experimental system is stochastic Pauli. To this end, the self-adjointness provides useful information since, if $H(\mathcal{E}) \ll 1$, then the noise is far from self-adjoint and cannot be approximated by Pauli noises.
This implies that the practical situation differs from the standard assumption in theoretical studies of QEC and incurs additional difficulties on decoding procedure. Thus, the self-adjointness provides practical information about the feasibility of QEC using Pauli-based decoders.
Note that the difficulty of QEC for non-Pauli noises, captured by the self-adjointness, highly depends on the assumptions in quantum error correction schemes. When \emph{any} decoding procedure is available, it would not be so important whether the noise is Pauli or non-Pauli. When this is the case, the unitarity will be a more suitable metric of noise relevant to the feasibility of QEC~\cite{KLDF2016,SWS2015,SFK2017}. Note also that non-Pauli noises can be always transformed to a Pauli noise by Pauli-twirling. However, Pauli-twirling induces additional noise onto the system and, as a result, the performance of QEC will degrade. Thus, it is practically desirable to manufacture the system so that the noise is stochastic Pauli.
We also provide a pedagogical example of noise, where performance of QEC can be directly captured by the self-adjointness but not by fidelity nor unitarity.
Consider a $\theta$-rotation error around the $X$-axis on one qubit, i.e., $\exp[i \theta X/2]$, where $X$ is the Pauli-$X$ operator. The average fidelity $F_{\theta}$ and the self-adjointness $H_{\theta}$ can be obtained as
\begin{align}
1/3 = F_{\pi} &< F_{\pi/2}=2/3,\\
0 = H_{\pi/2} &< H_{\pi}=1.
\end{align}
The unitarity is $1$ for any $\theta$.
One may expect that the $\pi/2$-rotation error is easier to correct than the $\pi$-rotation since the former has higher fidelity than the latter.
However, this is not the case since $\pi$-rotation is simply a perfect bit-flip that can be trivially corrected, while the $\pi/2$-rotation error is known to be particularly hard to correct~\cite{DP2017}.
Thus, neither the average fidelity nor the unitarity, which is $1$ for both errors, is a good metric of the error correctability. In contrast, the self-adjointness clearly captures whether the error can be corrected, at least in this case, since $H_{\pi/2}$ and $H_{\pi}$ are the minimum and the maximum values of the self-adjointness, respectively.
\subsection{General description of the $t$-th order RB} \label{SS:tRB}
We now introduce the $t$-RB using an exact unitary $2t$-design ${\sf U}_{2t}:= \{ U_i \}_i$. As is the case for the standard RB, we assume that the noise is gate- and time-independent, so that the noisy implementation of ${\sf U}_{2t}$ is given by $\{ {\cal G}_i := \mathcal{E} \circ \mathcal{U}_i\}_i$, where $\mathcal{E}$ is the CPTP map that represents the noise, and we used the notation that $\mathcal{U}(\rho) := U \rho U^{\dagger}$.
Let $O_{\rm ini}$ and $O_{\rm meas}$ be the initial and measurement operators, respectively, which we assume to be Hermitian.
We first apply a sequence of unitaries $U_{\boldsymbol{i}} = U_{i_m} \dots U_{i_1}$ onto the initial operator $O_{\rm ini}$. Each $U_{i_n}$ is chosen uniformly at random from ${\sf U}_{2t}$, which we denote by $U_{\boldsymbol{i}} \sim {\sf U}_{2t}^{\times m}$. We then apply its inverse $U_{i_{m+1}} := U_{\boldsymbol{i}}^{-1}$, and measure $O_{\rm meas}$.
If the system is noiseless, $\mathcal{E} = {\rm id}$, this protocol results in a trivial expectation value that
\begin{equation}
\tr \bigl[ O_{\rm meas} \mathcal{U}_{i_{m+1}} \circ \mathcal{U}_{i_m} \circ \dots \circ \mathcal{U}_{i_1}(O_{\rm ini}) \bigr] = \tr[O_{\rm meas} O_{\rm ini}]
\end{equation}
due to the inverse unitary $U_{i_{m+1}}$.
However, when the system is noisy, the expectation value becomes
\begin{equation}
\langle O_{\rm meas} \rangle_{O_{\rm ini}, \boldsymbol{i}}
:=
\tr \bigl[ O_{\rm meas} {\cal G}_{i_{m+1}} \circ {\cal G}_{i_m} \circ \dots \circ {\cal G}_{i_1}(O_{\rm ini}) \bigr],
\end{equation}
which in general differs from $\tr[O_{\rm meas} O_{\rm ini}]$. The basic idea of the RB-type protocol is to extract some information about the noise $\mathcal{E}$ from the difference.
In the $t$-RB, we especially focus on the average of the $t$-th power of the expectation value over all choices of the unitary sequence. That is,
\begin{equation}
V^{(t)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas}) :=
\mathbb{E}_{U_{\boldsymbol{i}} \sim {\sf U}_{2t}^{\times m}} \bigl[
\bigl( \langle O_{\rm meas} \rangle_{O_{\rm ini}, \boldsymbol{i}} \bigr)^t
\bigr].
\end{equation}
Using the representation-theoretic technique, it can be shown that $V^{(t)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas})$ is generally given in the following form:
\begin{equation}
V^{(t)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas})
= \sum_{\lambda} \tr \bigl[ \hat{A}^{(t)}_{\lambda} \bigl( \hat{C}^{(t)}_{\lambda}(\mathcal{E}) \bigr)^m \bigr], \label{Eq:general}
\end{equation}
where $\lambda$ labels the irreps of the unitary group, $\hat{A}^{(t)}_{\lambda}$ and $\hat{C}^{(t)}_{\lambda}(\mathcal{E})$ are $m_{\lambda} \times m_{\lambda}$ matrices with $m_{\lambda}$ being the multiplicity of the irrep $\lambda$.
This is well-known in the literature of RB-type protocols, but we provide a proof in Sec.~\ref{SS:generaltRB} for completeness.
Despite its abstract expression, Eq.\,\eqref{Eq:general} has an important implication that the matrix $\hat{C}^{(t)}_{\lambda}(\mathcal{E})^m$ depends only on $\mathcal{E}$ and $m$, but not on $O_{\rm ini}$ and $O_{\rm meas}$. Hence, from the experimental data of $V^{(t)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas})$ for various $m$, it is in principle possible to estimate the matrix $\hat{C}^{(t)}_{\lambda}(\mathcal{E})$, which contains certain information of the noise $\mathcal{E}$, in the way independent of $O_{\rm ini}$ and $O_{\rm meas}$.
In practice, the most important situation is when the representation is multiplicity-free, i.e., $m_{\lambda} = 1$ for any $\lambda$. In this case, $V^{(t)}$ reduces to a much simpler form:
\begin{equation}
V^{(t)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas}) = \sum_{\lambda} A^{(t)}_{\lambda} \bigl(C^{(t)}_{\lambda}(\mathcal{E})\bigr)^m, \label{Eq:tRBExpDecay}
\end{equation}
where $A^{(t)}, C^{(t)}_{\lambda}(\mathcal{E}) \in \mathbb{R}$. Note that $|C^{(t)}_{\lambda}(\mathcal{E})| \leq 1$ since $V^{(t)}$ is a bounded function. Hence, in this case, $V^{(t)}$ becomes a sum of some exponentially decreasing functions with respect to $m$.
To be more concrete, let us recall the standard RB, corresponding to the $1$-RB. As shown in Ref.\,\cite{EAZ2005}, $V^{(1)}$ is given by
\begin{equation}
V^{(1)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas}) = A_0^{(1)} + A_1^{(1)} f(\mathcal{E})^m, \label{Eq:korekore}
\end{equation}
where $A_0^{(1)}$ and $A_1^{(1)}$ depend only on $O_{\rm ini}$ and $\mathcal{E}(O_{\rm meas})$, and $f(\mathcal{E})$ is the fidelity parameter of the noise $\mathcal{E}$.
Thus, by fitting experimentally obtained data of $V^{(1)}$ for different $m$ with the fitting function $F(m)=A + B \alpha^m$, we can estimate the fidelity parameter $f(\mathcal{E})$.
\subsection{Importance of exact designs in RB} \label{SS:ExactImportant}
In the RB protocol, it is important to use \emph{exact} unitary designs because designs are used many times, sometimes a few hundreds to a thousand, in a single run of the protocol.
To illustrate this, let us consider the $1$-RB when the unitary $2$-design in the protocol is $\epsilon$-approximate.
Let $m$ be the length of the unitary sequence as above. It is straightforward to show that,
\begin{multline}
V^{(1)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas}) \\
\approx
A_0' + A_1 f^m + \epsilon (m-2) (E_2 f^2 + E_1 f + E_0) + O(m^2 \epsilon^2), \label{Eq:korekoredame}
\end{multline}
where $E_i$'s are some constants that depend on $O_{\rm ini}$, $\mathcal{E}(O_{\rm meas})$, $f$, and how the design differs from the exact one.
See Subsec.\,\ref{SS:1RBdetail} for the derivation.
Compared to the $1$-RB with exact ones, i.e., Eq.\,\eqref{Eq:korekore}, fitting this function with respect to $m$ is much harder since it is not a simple exponential decay.
The fitting may go well if $\epsilon \ll f^m/m$. This requires a very high precision of the design since $m$ can be a few hundreds in actual experiments. For instance, when $f=0.95$, the degree $\epsilon$ of approximation of the unitary design should be order $10^{-5}$ or so. Although it is possible to achieve this degree of approximation by a sufficiently long quantum circuit~\cite{BHH2016,NHKW2017,HMHEGR2020}, the RB becomes unpractical if we use such a long circuit at every use of a unitary design in the protocol and repeat it a few hundreds times.
There might be a possibility to improve Eq.\,\eqref{Eq:korekoredame} by using different constructions of approximate unitary designs at every step, by which the differences from the exact design may become random so that they cancel out in total. This will be an interesting question, but at this point, it is not clear if such a technique works. Also, even if it works, we need to assume additional structures of approximate constructions.
The higher-order RB with approximate designs will incur more difficulty in practice. Since it uses higher moment of the outcomes, the fitting function becomes more complicated than Eq.\,\eqref{Eq:korekoredame} when one uses approximate designs. Similarly to the $1$-RB with approximate $2$-designs, much better degree of approximation, that is, longer quantum circuits, will be needed, which is not practical.
Thus, we conclude that exact unitary designs are of crucial importance in a practical implementation of the $t$-RB.
\subsection{Second-order RB} \label{SS:2RB}
We next focus on the $2$-RB using exact unitary $4$-designs, and show that the $2$-RB reveals the self-adjointness of the noise.
To this end, we set the initial operator $O_{\rm ini}$ to a traceless one, i.e., $\tr[O_{\rm ini}] = 0$. This setting, together with the fact that the noise is trace-preserving, makes the representation multiplicity-free (see Appendix~\ref{App:Irreps}). Hence, the expectation value $V^{(2)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas})$ for the $2$-RB is given by a sum of exponentially decaying functions as shown in Eq.\,\eqref{Eq:tRBExpDecay}.
Note that the expectation value for a traceless initial operator can be obtained by performing the same experiment for two different quantum states $\rho$ and $\rho'$, and by taking the difference of the expectation values before they are squared. That is,
\begin{multline}
V^{(2)}(m, \mathcal{E}| \Delta, O_{\rm meas})\\
=
\mathbb{E}_{U_{\boldsymbol{i}}} \bigl[
\bigl( \langle O_{\rm meas} \rangle_{\rho, \boldsymbol{i}}
-
\langle O_{\rm meas} \rangle_{\rho', \boldsymbol{i}}
\bigr)^2
\bigr]
\end{multline}
where $\Delta = \rho - \rho'$ is a traceless operator.
Our second main result in this paper is about $V^{(2)}(m, \mathcal{E}| \Delta, O_{\rm meas})$ as summarized in Theorem~\ref{Thm:2RB}.
\begin{Theorem} \label{Thm:2RB}
In the above setting, $V^{(2)}(m, \mathcal{E}| \Delta, O_{\rm meas})$ is given as follows.
For single-qubit systems,
\begin{multline}
V^{(2)}(m, \mathcal{E}| \Delta, O_{\rm meas}) \\= A_0 u(\mathcal{E})^m + A_1 \biggl( \frac{9}{10}f(\mathcal{E})^2 - \frac{1}{5} u(\mathcal{E}) + \frac{3}{10} h(\mathcal{E}) \biggr)^m, \label{Eq:15}
\end{multline}
where $f(\mathcal{E}), u(\mathcal{E})$, and $h(\mathcal{E})$ are the fidelity parameter, the unitarity, and the self-adjointness parameter of the noise $\mathcal{E}$, respectively.
For multi-qubit systems,
\begin{equation}
V^{(2)}(m, \mathcal{E}| \Delta, O_{\rm meas}) = A_0 u(\mathcal{E})^m + \sum_{\lambda={\rm I}, {\rm II}, {\rm III}} A_{\lambda} C_{\lambda}(\mathcal{E})^m, \label{Eq:16}
\end{equation}
where $0 \leq C_{\lambda}(\mathcal{E}) \leq 1$ depend only on the noise $\mathcal{E}$.
Moreover, they satisfy
\begin{multline}
\sum_{\lambda={\rm I}, {\rm II}, {\rm III}} D_{\lambda} C_{\lambda}(\mathcal{E}) \\
=
\frac{(d^2-1)^2}{2} f(\mathcal{E})^2 - u(\mathcal{E}) + \frac{d^2-1}{2} h(\mathcal{E}), \label{Eq:19p}
\end{multline}
where
\begin{align}
&D_{\rm I} = \frac{d^2(d-1)(d+3)}{4},\\
&D_{\rm II} = \frac{d^2(d+1)(d-3)}{4},\\
&D_{\rm III} = d^2-1.
\end{align}
\end{Theorem}
See Subsec.\,\ref{SS:2RBdetail} for the proof.\\
In the single-qubit case, $V^{(2)}$ is a sum of two exponentially decaying functions with respect to $m$. Hence, from the double-exponential fitting of the experimental data of $V^{(2)}$, we can simultaneously estimate $u(\mathcal{E})$ and $\frac{9}{10}f(\mathcal{E})^2 - \frac{1}{5} u(\mathcal{E}) + \frac{3}{10} h(\mathcal{E})$.
Since it can be shown that the former is not less than the latter, we can estimate which of the two decaying rates corresponds to which quantity without any ambiguity.
It is also possible to estimate the fidelity parameter $f(\mathcal{E})$ from the same data set by computing $V^{(1)}(m, \mathcal{E}| \Delta, O_{\rm meas})$ because a unitary $2$-design is also a unitary $1$-design. Thus, from the experiment of the $2$-RB on a single qubit, all of $f(\mathcal{E}),$ $u(\mathcal{E})$, and $h(\mathcal{E})$ can be estimated simultaneously.
In multi-qubit systems, $V^{(2)}$ has a little more complicated form and consists of four exponentially decaying functions. Also, the decaying rates do not directly correspond to neither the unitarity nor the self-adjointness parameter. We observe from Eq.\,\eqref{Eq:19p} that $h(\mathcal{E})$ can be obtained from a linear combination of the decaying rates $C_{\lambda}(\mathcal{E})$, the value of $u(\mathcal{E})$, and $f(\mathcal{E})$.
One may think that, in the case of multiple qubits, it is practically intractable to accurately fit four exponentially decaying functions from experimental data because each data point has an error.
This difficulty can be circumvented by choosing appropriate initial and measurement operators. By doing so, we can set some of $A_{\lambda}$ zero in the ideal situation (see Tab.\,\ref{Tab:coeff}). This allows us to estimate the decaying rates one by one.
Note that the initial and measurement operators in Tab.\,\ref{Tab:coeff} are all diagonal in the computational basis. Hence, it suffices to perform the experiments for the four initial operators $\ket{00}, \ket{01}, \ket{10}$, and $\ket{11}$, with the measurement in the computational basis. From the data of these experiments, it is possible to reproduce all cases listed in Tab.\,\ref{Tab:coeff} by post-processing.
In the multi-qubit case, the ambiguity remains to decide which of the decaying rates corresponds to which quantity. This is the case even when we use the above step-by-step estimation of the rates since, for instance, it is not clear if the unitarity $u(\mathcal{E})$ is larger or smaller than $C_{\rm I}$. In this case, we need to additionally perform the unitarity benchmarking~\cite{WGHF2015,HHFFW2019} to separately estimate $u(\mathcal{E})$.
If we have an estimated value of $u(\mathcal{E})$, the step-by-step estimation allows us to decide all decaying rates without any ambiguity.
See Sec.\,\ref{SS:NumericalEval} and Sec.\,\ref{SS:NumericalEvalDetail} for the performance of 2-RB in concrete cases.
\begin{table}[t!]
\centering
\begin{tabular}{c||c|c|c}
$(\Delta, O_{\rm meas})$ & $\bigl(ZZ, \ketbra{00}{00}\bigr)$ & $(ZZ, ZZ)$ & $\bigl(\rho_-, \rho_- \bigr) $\\ \hline \hline
$A_0$ & $1/5$ & 16/15 & 4/15 \\
$A_{\rm I}$ & $4/5$ & 48/5 & 41/15\\
$A_{\rm II}$ & $0$ & 16/3 & 1/3 \\
$A_{\rm III}$ & $0$& 0 & 2/3 \\
\end{tabular}
\caption{A table of coefficients $A_0, A_{\rm I}, A_{\rm II}$, and $A_{\rm III}$ appeared in Eq.\,\eqref{Eq:16}, for the $2$-qubit case. The first row provides a pair of the initial and measurement operators $(\Delta, O_{\rm meas})$. We have assumed that the average fidelity of the noise $\mathcal{E}$ is close to $1$, so that the inverse unitary $U_{i_{m+1}}$ can be applied nearly noiseless (see Subsec.\,\ref{SS:2RBdetail} for the detail).
The operator $ZZ$ is $Z \otimes Z$, and $\rho_- := \ketbra{00}{00}-\ketbra{11}{11}$. By choosing proper operators, we can set some coefficients zero, so that experimental estimations of $C_{\lambda}(\mathcal{E})$ become easy.}\label{Tab:coeff}
\end{table}
\begin{figure*}[tb!]
\centering
\includegraphics[width=\textwidth]{fig_result_1q_PQshift.pdf}
\caption{The estimated values of $F(\mathcal{E}_1), u(\mathcal{E}_1),$ and $H(\mathcal{E}_1)$ obtained by 2-RB on one-qubit system for various parameters $p$ and $q$, where we have taken $5000$ samplings both for measurement and for unitary sequences. The dots show the fitting results, and the dashed lines represent theoretical values.}
\label{Fig:numerical_overview_1q}
\end{figure*}
\subsection{Scalability} \label{SS:Scalability}
The $t$-RB for $t \geq 2$ inherits most of the desired properties of the RB-type protocols. For instance, it is experimentally-friendly since, apart from using higher-designs, the difference of the $t$-RB from the standard RB ($1$-RB) is only taking the $t$-th power of the expectation value before the average. It is also true that the $t$-RB is free from SPAM errors (see Eqs.\,\eqref{Eq:general} and~\eqref{Eq:tRBExpDecay}).
The property that the standard RB does have and the $t$-RB does not in general is the scarability. This is for two reasons.
First, no efficient construction of exact unitary $2t$-designs is known for $t \geq 2$ so far.
Second, in the $t$-RB protocol, it is necessary to apply the inverse unitary at the end of the unitary sequence. Hence, we need to beforehand compute the inverse of each sequence. When the system is large, the task is intractable in general. This difficulty is avoided in the standard RB by using the Clifford group, which is an exact unitary $2$-design. Since the inverse is contained in the group, we can find the inverse relatively easily.
One may expect that the difficulty of finding the inverse could be also avoided in the $t$-RB by using the $2t$-design that is also a group, which is called a unitary $2t$-group~\cite{BNRT2020}. However, it is known that unitary $2t$-groups do not exist for $t \geq 2$ if the number of qubits $ \geq 3$. Thus, in the $t$-RB for $t \geq 2$, the hardness of finding the inverse in a large system is inevitable.
Nonetheless, we emphasize that, in the current experimental situations, the RB-type protocols for more than three qubits are practically intractable due to the limitation of the coherent time. Thus, the experimental use of the RB-type protocols is currently aiming to characterize the noise on one- or two-qubit systems in a concise manner. Considering this fact, even if the $t$-RB is not scalable, it is practically useful and beneficial: it is as concise as the standard RB and provides more information about the noise, such as self-adjointness.
\section{Main result 3 -- $2$-RB in a superconducting system --} \label{S:Main3}
We finally implement the $2$-RB in a superconducting system and estimate the self-adjointness of background noise.
Unlike the analytical studies, the expectation values and the average over a unitary $4$-design cannot be taken with arbitrary precision in experiments since the number of repetitions of experiment is practically limited. To check that this limitation does not cause any problem in the evaluation of the self-adjointness, we start with numerically investigating the feasibility of the $2$-RB in Subsec.\,\ref{SS:NumericalEval}. We then provide a summary of experimental results in Subsec.\,\ref{SS:Experiments}.
In recent years, a number of experiments have been performed to characterize various noises on superconducting quantum systems in detail\,\cite{Wilen2021, HAN202110, McEwen2021, mcewen2021resolving}. From our experiments, we show that
the interactions with the adjacent qubits particularly decrease the self-adjoinenss and may cause problems toward realizations of QEC. In particular, our result implies that there exists a gap between the superconducting system and the common noise model used in theoretical studies of QEC, and also that the standard decoders of stabilizer codes may suffer from degradation of logical errors. Hence, toward the realization of QEC, it is desired to further improve the system or to develop the theory of QEC.
\subsection{Numerical evaluation} \label{SS:NumericalEval}
When the $2$-RB is practically implemented, there are two additional concerns. One is originated from the fact that the expectation value $\langle O_{\rm meas} \rangle_{O_{\rm ini}, \boldsymbol{i}}$ is obtained from a limited number of measurements in the basis of $O_{\rm meas}$, resulting in an error due to a finite number of measurements.
The other originates from the evaluation of the average $\mathbb{E}_{U_{\boldsymbol{i}} \sim {\sf U}_{4}^{\times m}}$ over the sequence of unitaries in the $4$-design. Ideally, all sequences in ${\sf U}_{4}^{\times m}$ should be taken, but practically, the average is often evaluated from a small subset in ${\sf U}_{4}^{\times m}$ of randomly chosen sequences, leading to an additional error of estimation.
Taking sufficiently many measurements and samplings of unitary sequences will reproduce the analytical results with high accuracy. However, it is complicated to analytically derive the numbers sufficient for achieving a desired accuracy. We hence perform numerical experiments and show that experimentally-tractable numbers of samplings are sufficient for a reliable $2$-RB.
\subsubsection{One-qubit cases}
In the case of single-qubit systems, we consider a specific noisy map given by
\begin{equation}
\mathcal{E}_1(\rho) = q e^{i\theta X} \rho e^{- i\theta X} + (1-q)((1-p) \rho + p X \rho X), \label{Eq:Noise1}
\end{equation}
which is characterized by three parameters $p, q, \theta$.
The first term of the right-hand side represents a unitary part and the second term represents a stochastic part of the noise. A parameter $q$ determines a ratio between them. Hence, we can consider $q$ as a coherent parameter of noise, e.g., noise is unitary when $q=1$ and is a probabilistic Pauli noise when $q=0$.
The parameters $\theta$ and $p$ represent the rotation angle of the unitary part and the error probability of the stochastic part, respectively. For simplicity, we choose $\theta$ such that the fidelity parameters of unitary and stochastic parts are equal, that is, $p = \sin^2 \theta$. Then, the fidelity parameter $f(\mathcal{E}_1)$ becomes independent of the coherent parameter $q$.
To perform the $2$-RB for this noise, we may use the exact $4$-design constructed in Corollary~\ref{Cor:tDesign_qubit}. However, it is known that the icosahedral group, which we denote by ${\sf I}$, forms an exact $4$-design on one qubit~\cite{RS2009}. Since the icosahedral group has less cardinality than our inductive construction, we use it in the following analysis.
The numerical results for the $2$-RB on a single qubit are shown in Fig.\,\ref{Fig:numerical_overview_1q}. For each sequence length $m$, we have taken $5000$ random unitary sequences from ${\sf I}^{\times m}$ and have had $5000$ measurements to obtain a single data point of $V^{(1)}$. A detailed fitting procedure is provided in Subsec.\,\ref{SS:NumericalEvalDetail}.
To check the accuracy of the $2$-RB, we consider the relative errors $|y-\tilde{y}|/(1-y)$, where $y$ and $\tilde{y}$ are the theoretical value and the fitting value, respectively. Note that $1-y \sim 0$ for all the fitting values when a fidelity close to unity is achieved.
For almost all data points of $F(\mathcal{E}), u(\mathcal{E}),$ and $H(\mathcal{E})$, we find that the relative errors are less than $5.0\%$, except the case when $p$ is large, or equivalently, when the fidelity is small. The relative error becomes moderately large, such as $35\%$, when $p=0.4$ and $q\geq 0.1$, corresponding to $F(\mathcal{E}_1)=0.7$.
This is because the decaying rate of the second term in Eq.\,\eqref{Eq:15} is rather small, making the fitting difficult. However, such a case is not practically relevant since the fidelity is typically $>90\%$. Thus, we conclude that the 2-RB on 1-qubit systems works well in practice.
To analyze the dependence of the accuracy of the $2$-RB on the number of measurements and samplings of random unitary sequences, we additionally perform the 2-RB on one qubit with the various numbers of measurements and samplings. The results are summarized in Tab.\,\ref{Tab:SR_shift_for_1q}, where we set the noise parameters to $p=0.02$ and $q=0.02$. From these results, it appears that setting the numbers of measurements and samplings of random sequences to a few hundreds is sufficient for a good estimate. These results further indicate that increasing the number of random sequences rather than the number of measurements is preferable to improve the accuracy. See Subsec.\,\ref{SS:NumericalEvalDetail} for the details.
\begin{table}[t!]
\centering
\caption{Numerically estimated values of the fidelity, unitarity, and self-adjointness from the single-qubit $2$-RB with the finite numbers of measurements and samplings of unitary sequences. We set $p=q=0.02$. The theoretical values are shown at the bottom of the table.}
\begin{tabular}{c||c|r|r|r}
\# of meas. & \# of sequences & $F$ & $u$ & $H$ \\ \hline \hline
10 & 100 & $0.986(6)$ & $0.9985(5)$ & $0.92(6)$\\
10 & 500 & 0.986(1) & 0.9984(1) & 0.93(1)\\
10 & 1000 & 0.986(6) & 0.9984(3) & 0.92(3) \\
100 & 100 & 0.986(3) & 0.9980(8) & 0.92(2) \\
100 & 500 & 0.986(5)& 0.9980(4)& 0.92(3)\\
100 & 1000 & 0.986(6) & 0.9979(4)& 0.92(1) \\
1000 & 100 & 0.986(3) & 0.9980(1)& 0.92(3)\\
1000 & 500 & 0.986(4) & 0.9979(5)& 0.92(3)\\
1000 & 1000 & 0.986(6) & 0.9978(9)& 0.92(2) \\ \hline
& & 0.9866 & 0.99793 & 0.9247
\end{tabular}
\label{Tab:SR_shift_for_1q}
\end{table}
\subsubsection{Two-qubit cases}
\begin{figure*}[tbh!]
\centering
\includegraphics[width=\textwidth]{fig_result_2q.pdf}
\caption{Four coefficients for exponentially decaying functions in the 2-RB are plotted according to coherence parameter $q$, where we had $10^4$ measurement and $10^4$ samplings of unitary sequences. Dashed lines are theoretical values. Each color corresponds to each value of the parameter $p$.}
\label{Fig:numerical_overview_2q}
\end{figure*}
For two-qubit systems, we consider the noise given by
\begin{align}
\mathcal{E}_2(\rho) &= q e^{i \theta (X\otimes X)} \rho e^{- i\theta (X \otimes X)} \nonumber \\
&+ (1-q)((1-p) \rho + p (X \otimes X) \rho (X\otimes X) ),
\end{align}
which is similar to the one-qubit case. We choose $\theta$ as $p = \sin^2 \theta$, so that $f(\mathcal{E}_2)$ is independent of the coherent parameter $q$. In this case, we use the construction of exact unitary $4$-designs given in Proposition~\ref{Prop:Clliford4}.
In the two-qubit case, it is needed to fit the experimental data by a sum of four exponentially decaying functions, which is in general not easy especially when each data point has errors caused by the finite number of measurements and samplings of unitary sequences. To avoid this difficulty, we use the method explained in Subsec.\,\ref{SS:2RB}, and determine the four decaying rates, i.e., $u(\mathcal{E}_2), C_{\rm I}(\mathcal{E}_2), C_{\rm II}(\mathcal{E}_2)$, and $C_{\rm III}(\mathcal{E}_2)$ in Eq.\,\eqref{Eq:16}, one by one.
The results are shown in Fig.\,\ref{Fig:numerical_overview_2q}.
We have taken $10^4$ random unitary sequences for each sequence length $m$ and the parameters $p,q$. A detailed process of fittings are explained step by step in Sec.\,\ref{SS:NumericalEvalDetail}. In the figure, fitted values are shown as data points. Dashed lines are drawn with theoretically calculated values.
Similarly to the case of the single-qubit $2$-RB, we have checked the relative errors of the fitting results to the theoretical values. The errors are all below $3\%$ for all the points except the case when the theoretical value is exactly zero.
As in the case of single-qubit 2-RB, when we calculate $F(\mathcal{E}), u(\mathcal{E}),$ and $H(\mathcal{E})$ from the fitting values, the relative values of almost all the data points are less than $4.0\%$. While the relative errors become large when $p=0.4$, such a case is not a problem in typical calibration scenario. Thus, the $2$-RB works in actual situations also in the case of two-qubit systems.
\subsection{Experimental implementations of the 2-RB} \label{SS:Experiments}
We demonstrate the $2$-RB in a superconducting-qubit system.
We first explain the setup of our experiments, and then verify the feasibility of the $2$-RB experiment by comparing the unitarity obtained from the $2$-RB with that from the unitarity benchmarking (UB)~\cite{WGHF2015,HHFFW2019}.
We finally characterize background noise of the system. As the background noise is gate- and time-independent, it satisfies the assumptions of the $2$-RB (see Subsec.\,\ref{SS:ExperimentDetail} for the detail).
\subsubsection{Experimental setup}
We use two superconducting qubits ($Q_1$ and $Q_2$) coupled with each other via an electric dipole interaction, which are a part of our $16$-qubit device~\cite{tamate2021scalable}. In all the experiments below, we use the qubit $Q_1$ as a target qubit of the single-qubit $2$-RB and, in some experiments, $Q_2$ as an environmental qubit that induces additional error onto $Q_1$.
The simplified system Hamiltonian $H$ is formulated as follows,
\begin{align}
\frac{H}{\hbar} = \frac{\omega_1}{2} Z\otimes I + \frac{\omega_2}{2} I\otimes Z + \frac{\chi_{ge}}{2} Z\otimes Z,
\label{Eq:ExpHam}
\end{align}
where $\omega_i/2\pi$ is the eigenfrequency of the $i$-th qubit and $\chi_{ge}/2\pi=-0.760~{\rm MHz}$ is an effective interaction strength between the qubits~\cite{gambetta2006qubit}.
It can be interpreted that the eigenfrequency of $Q_1$ switches depending on the quantum state of $Q_2$.
When $Q_2$ is in the $\ket{0}$ ($\ket{1}$) state, $Q_1$ has the eigenfrequency $(\omega_1 + \chi_{ge})/2\pi$ ($(\omega_1 - \chi_{ge})/2\pi$).
In the Bloch sphere representation, the state vector of the qubit rotates around the $Z$-axis with its eigenfrequency as the angular velocity.
We use a local oscillator synchronized with the eigenfrequency of the qubit for observation.
The state vector is stationary in a rotating frame of the local oscillator since the $Z$-axis rotation speed of the Bloch vector matches with that of the measurement basis.
The rotation frame picture also holds when the qubit $Q_1$ couples to the adjacent qubit $Q_2$ when the qubit $Q_2$ is in the $\ket{0}$ or $\ket{1}$ state.
For instance, when the qubit $Q_2$ is always in the $\ket{0}$ state, the eigenfrequency of $Q_1$ is $(\omega_1 + \chi_{ge})/2\pi$. We can detune the frequency of the local oscillator from the qubit frequency $\omega_1$ by $\chi_{ge}$ to make the state vector of $Q_1$ stationary.
It is, however, impossible to keep track of the eigenfrequency of the qubit
when the state of the adjacent qubit varies. This results in an inevitable $Z$-rotation occurring in the quantum state.
In an actual experiment involving multiple qubits, the frequency of the local oscillator is usually set to $\omega_1/2\pi$ to minimize the average $Z$-rotation angle.
See Subsec.\,\ref{SS:ExperimentDetail} for the detail.
\subsubsection{Comparison with the UB}
\begin{table}[t!]
\centering
\caption{The estimated values of the fidelity, unitarity, and self-adjointness from the experiment of the $2$-RB and that of the UB in the superconducting qubit system.}
\begin{tabular}{c||c|c|c|c|c}
Experiment & Group & $F$ & $u$ & $H$ \\ \hline \hline
$2$-RB & Icosahedral & 0.926(6) & 0.970(1) & 0.6(1) \\
UB & Clifford & -- & 0.977(1) & -- \\ \hline
Theoretical & -- & 0.936 & 1 & 0.655
\end{tabular}
\label{Tab:Experiment_Sanity_Check}
\end{table}
In the experiment aiming to compare the $2$-RB and the UB on a single qubit, we use only $Q_1$ and add an artificial noise after applying each gate.
The isolation of the qubit $Q_1$ from the qubit $Q_2$ can be done by keeping the qubit $Q_2$ in the state $\ket{0}$ and by setting the frequency of the local oscillator to $(\omega_1 + \chi_{ge})/2\pi$, which effectively cancel the coupling between $Q_1$ and $Q_2$. About the noise, we especially choose a single-qubit $Z$-rotation by angle $0.2 \pi$, denoted by $R_Z(0.2\pi)$.
Both in the case of the $2$-RB and the UB, we use the icosahedral group ${\sf I}$ and the Clifford group on a single qubits, respectively. Note that the former is an exact $4$-design on a single qubit, and the latter is an exact $2$-design.
We have taken $100$ and $1000$ random sequences for the $2$-RB and the UB, respectively. This is because the UB with the Clifford group converges slower than the $2$-RB with the icosahedral group, which is likely due to the fact that the former and the latter are based on unitary $2$- and $4$-designs, respectively. A higher-design typically leads to a quick convergence since it is more concentrating around the average~\cite{L2009LDB}.
A faster convergence of the UB with $4$-design is expected, which highlights the potential use of a higher-design also for the UB.
We have taken $10^4$ measurements for each random sequence to obtain a data point of $V^{(2)}$.
The results are summarized in Tab.\,\ref{Tab:Experiment_Sanity_Check}.
From the results, we observe that the unitarity characterized by the $2$-RB matches with that by the UB. This indicates that the $2$-RB on our single-qubit system works to characterize the gate performance.
Note that the difference between the unitarity from the $2$-RB and that from the UB is slightly beyond the standard deviation. This is likely because the noise property varies in the UB experiment. As mentioned, we have taken $1000$ random sequences in the UB to ensure the convergence of the statistical average, which has taken more than 10 hours in total. Since the noise in the experimental system drifts in such a long timescale, the situation of the experiment deviates from the ideal situation, where time-independence of the noise is assumed. Indeed, unlike the theoretical prediction of the UB, the data is slightly different from a single-exponential decay. This deviation is expected to be the origin a less precise value of the unitarity estimated from the UB.
\begin{figure*}[tbh!]
\centering
\includegraphics[width=\textwidth]{210308_fig_param_1q_exp_sweep_delay.pdf}
\caption{(a) Experimental results for the single-qubit $2$-RB on an isolated qubit when the inverleaved delay time $t$ are swept.
We have taken $100$ random unitary sequences from ${\sf I}^{\times m}$ for each sequence length $m$ and have had $10^4$ measurements for each sequence to obtain a data point of $V^{(2)}$.
The stars represent the values of $V^{(2)}$ and the dashed lines are fitting results.
The error bars represent the standard deviation of $V^{(2)}$.
Since the vertical axis of the figure is logarithmic notation, the error bars at the bottom of the figure are displayed larger.
(b) Estimated values of $F(\mathcal{E}_{t}), u(\mathcal{E}_{t}),$ and $H(\mathcal{E}_{t})$ obtained by the $2$-RB.
The stars show the fitting results, and the dashed lines are the predicted values from the phenomenological model.}
\label{Fig:experimental_1q_sweep_delay}
\end{figure*}
\subsubsection{Characterizing background noise}
We next perform the single-qubit $2$-RB, aiming to characterize background noise of the qubit $Q_1$ in the experimental system.
We intentionally insert a delay time $t$ after each application of a gate to extract the information of the background noise.
In the following experiments, we have taken $100$ random unitary sequences from ${\sf I}^{\times m}$, where ${\sf I}$ is the icosahedral group, for each sequence length $m$ and have had $10^4$ measurements for each random sequence to get a data point of $V^{(2)}$.\\
\begin{figure*}[tbh!]
\centering
\includegraphics[width=\textwidth]{210308_fig_param_1q_exp_sweep_delay_zz.pdf}
\caption{(a) Experimental results for the single-qubit $2$-RB on the qubit coupled to another qubit, where we sweep the interleaved delay time $t$.
We have taken $100$ random unitary sequences from ${\sf I}^{\times m}$ for each sequence length $m$ and have had $10^4$ measurements for each random sequence to take a data point of $V^{(2)}$.
The stars correspond to the values of $V^{(2)}$ and the dashed lines are fitting results.
(b) Estimated values of $F(\mathcal{E}_{t}), u(\mathcal{E}_{t}),$ and $H(\mathcal{E}_{t})$ obtained by the $2$-RB.
The stars show the fitting results, and the dashed lines are the predicted values from the phenomenological model.}
\label{Fig:experimental_1q_sweep_delay_zz}
\end{figure*}
In the first experiment, we set the frequency of the local oscillator to $(\omega_1 + \chi_{ge})/2\pi$ and treat the qubit $Q_1$ as a target qubit isolated from the qubit $Q_2$.
The background noise of the isolated qubit is often phenomenologically modeled by the Lindblad Master equation given by
\begin{align}
\frac{d\rho}{dt} &=
\sum_{k\in [1,2]}
L_k^\dagger \rho L_k - \frac{1}{2}\left\{L_k^\dagger L_k, \rho\right\}, \label{Eq:NoiseTheoretical1}
\end{align}
where
$L_1 = \hat{a}/\sqrt{T_1}$ represents the energy dissipation with the relaxation time $T_1$, $\hat{a}=(X+iY)/2$ is an annihilation operator of the qubit, and $L_2 = Z/\sqrt{2T_{\phi}}$ represents the phase dissipation with the relaxation time $T_\phi=1/(1/T_2 - 1/2T_1)$.
By solving the Eq.\,\eqref{Eq:NoiseTheoretical1}, we can obtain phenomenological predictions about the background noise $\mathcal{E}_{t}$ corresponding to the delay time $t$.
We sweep the delay time $t$ from $100$ to $500~{\rm ns}$.
The value $V^{(2)}$ obtained from the experiments is shown in Fig.\,\ref{Fig:experimental_1q_sweep_delay}~(a). We estimate the unitarity and the self-adjointness from $V^{(2)}$ through a fitting based on a sum of \emph{two} exponentially-decaying curves given in Eq.\,\eqref{Eq:15}.
However, we observe single-exponential decays from the results.
This indicates two possibilities. One is that the two decaying rates are nearly the same. The other is that one of the two decaying rates is much smaller than the other, so that one exponentially-decaying curve becomes quickly negligible as $m$ increases.
In our experiment, the former is the case because the average fidelity is high, which is confirmed from the $1$-RB.
We can analytically show that the two decaying rates typically coincide when the fidelity is sufficiently high.
More specifically, we have (see Eq.\,\eqref{Eq:97779} in Subsec.\ref{SS:PropSA})
\begin{equation}
1-4 \epsilon \lesssim \frac{h(\mathcal{E}) + u(\mathcal{E})}{2}, \label{Eq:eerrwwww}
\end{equation}
to the first order of $\epsilon$, where $\epsilon$ is the infidelity $1-F(\mathcal{E})$.
This implies that the two decaying rates in Theorem~\ref{Thm:2RB} are approximately greater than $1-4 \epsilon$ and $1- 6 \epsilon$, respectively.
Thus, if $\epsilon \ll 1$, which is indeed the case in our system, the two decaying rates are hard to distinguish, making the curve of $V^{(2)}$ a single-exponential decay.
We, hence, estimate the single-exponential decay rate from the experimental data of the $2$-RB and derive $u(\mathcal{E}_{t})$ and $h(\mathcal{E}_{t})$ from
\begin{equation}
u(\mathcal{E}_t) = \frac{9}{10}f(\mathcal{E}_t)^2 - \frac{1}{5}u(\mathcal{E}_t) + \frac{3}{10}h(\mathcal{E}_t).
\end{equation}
Here, $u(\mathcal{E}_t)$ is obtained from the estimated decaying rate, and $f(\mathcal{E}_t)$ from the $1$-RB (See Eq.\,\eqref{Eq:15}).
The obtained fidelity, unitarity, and self-adjointness are summarized in Fig.\,\ref{Fig:experimental_1q_sweep_delay}~(b).
In calculating self-adjointness, we solved Eq.\,\eqref{Eq:66}, where we substituted $\alpha_{\mathcal{E}}$ of the phenomenological prediction.
They reveal that the background noise of the isolated qubit has the unitarity $u(\mathcal{E}_{t})$ that slowly decreases as the delay increases, while its self-adjoingness $H(\mathcal{E}_{t})$ is nearly independent of the delay.
As we have explained in Subsec.\,\ref{SS:defdef}, a problem may occur when the unitarity is high and the self-adjointness is low, which is not observed in this experiment. Hence, we conclude that the background noise in this case would not cause any problem toward the realization of QEC. \\
In the second experiment, we set the frequency of the local oscillator to $\omega_1/2\pi$ and treated $Q_1$ as a target qubit exposed to the noise induced by the adjacent qubit $Q_2$. In this experiment, no control pulses are applied to $Q_2$, so that $Q_2$ is expected to remain in the $\ket{0}$ state. This leads to a continuous rotation of the state vector of $Q_1$ by the interaction Hamiltonian term of $\chi_{ge} Z/2$.
In this case, the background noise with the interaction Hamiltonian is modeled by the Lindblad Master equation written as follows,
\begin{align}
\frac{d\rho}{dt} =\left[\frac{\chi_{ge}}{2}Z,\rho\right] +
\sum_{k\in [1,2]}
L_k^\dagger \rho L_k - \frac{1}{2}\left\{L_k^\dagger L_k, \rho\right\},
\label{Eq:NoiseTheoretical2}
\end{align}
providing a phenomenological model.
Similarly to the first experiment, we sweep the delay time $t$ from $60~{\rm ns}$ to $180~{\rm ns}$.
The delay time is set to a shorter time than the first experiment because the fidelity deteriorates due to the Z-rotation error.
Since the Z-rotation error does not affect the unitarity, we conclude that the decay rate, which is less sensitive to the delay time than the other, corresponds to the unitarity.
The results of the experiment are shown in Fig.\,\ref{Fig:experimental_1q_sweep_delay_zz}~(a).
As seen from the results, the curves $V^{(2)}$ obey double-exponential decay. From the two decaying rates, we obtain the unitarity $u(\mathcal{E}_{t})$ and the self-adjointness $H(\mathcal{E}_{t})$ as a function of the delay time as depicted in Fig.\,\ref{Fig:experimental_1q_sweep_delay_zz}~(b).
Note that, although the unitarity may seem different from the former experiment, it is merely due to the different time scale of the horizontal axis. The unitarities in the two experiments indeed coincide within the standard deviation (see, e.g, the delay time $100$ (ns)).
The experimental results qualitatively coincide with the phenomenological predictions obtained from Eq.\,\eqref{Eq:NoiseTheoretical2} (see Fig.\,\ref{Fig:experimental_1q_sweep_delay_zz}~(b)). However, the experimental values tend to be smaller. This indicates that there exist noise sources not included in the phenomenological model.
The candidates of the additional noise sources are calibration errors in the $R_X(\pi/2)$ gates, the initial thermal excitation rate of $Q_2$ ($7.2~\%$), and the interaction of $Q_1$ with adjacent qubits other than $Q_2$.
Note that the initial thermal excitation of $Q_2$ makes the noise time-dependent due to the relaxation, and hence, makes the result different from the theoretical prediction of the $2$-RB.
Compared to the first experiment (Fig.\,\ref{Fig:experimental_1q_sweep_delay}), we observe from Fig.\,\ref{Fig:experimental_1q_sweep_delay_zz} that the fidelity $F(\mathcal{E}_t)$ and the self-adjointness $H(\mathcal{E}_t)$ quickly decrease as the delay time $t$ increases. The latter decreases especially quickly: $H(\mathcal{E}_t) \approx 0.48$ at $t=140$ (ns).
This implies that, even if the fidelity is moderately high ($F(\mathcal{E}_t) \approx 0.91$ at $t=140$ (ns)), the extra $Z$-rotation induced by the interaction with another qubit radically changes the property of the noise and makes the noise far from self-adjoint. Consequently, as the delay increases, the noise quickly becomes the one that cannot be approximated by any stochastic Pauli noise.
This result has an important implication toward a realization of QEC.
As mentioned in Subsec.\,\ref{SS:defdef}, theoretical studies of QEC commonly assume stochastic Pauli noises to numerically compute error thresholds and error rates.
Our result implies that, when the interaction with another qubit is non-negligible, we cannot directly apply the theoretical predictions based on Pauli noises.
This problem will be more prominent when the system size grows since, in a large system, a qubit interacts with more qubits in an uncontrolled manner, making the noise much less self-adjoint and much far from Pauli noises.
To circumvent this, effective cancellation of the dipole interaction is of great importance in the further improvement since the dominant interaction between qubits should be originated from the electric dipole interaction.
This feature of the noise, i.e., interactions with other qubits induce small self-adjointness and the difficulty of approximating the noise by a Pauli noise, is expected to be common in any experimental systems. The $2$-RB experiment and the self-adjointness offer a useful method and measure, respectively, to experimentally evaluate the noise in the system from this perspective.
\section{Structure of the remaining paper} \label{S:structure}
The remaining of this paper is organized as follows.
In Sec.\,\ref{Sec:ConstDesign}, a proof of Theorem~\ref{Thm:General} is provided. A brief introduction of representations of the unitary group is also provided before the proof. We then explain the higher-order RB in Sec.\,\ref{S:higher-orderRB}, including the proof of Theorem~\ref{Thm:2RB}. The methods used in the numerical analysis, and the experimental demonstrations are provided in Sec.\,\ref{S:NumExpDetail}.
After we summarize the paper in Sec.\,\ref{S:SD}, we prove technical statements in Appendices.
\section{Constructions of exact designs} \label{Sec:ConstDesign}
In this section, we provide a proof of Theorem~\ref{Thm:General}. We start with a brief introduction of representations of the unitary group in Subsec.\,\ref{SS:unitaryrep}, and prove Theorem~\ref{Thm:General} in Subsec.\,\ref{SS:ProofThm1}.
\subsection{Unitary $t$-designs and representation theory} \label{SS:unitaryrep}
Unitary $t$-designs are closely related to representations of the unitary group since the operator $U^{\otimes t} \otimes U^{\dagger \otimes t}$ in the definition can be regarded as a representation $\rho$ of $U \in {\sf U}(d)$ on ${\mathcal H}_d^{\otimes 2t}$ with ${\mathcal H}_d$ being the Hilbert space with dimension $d$, i.e., $\rho(U) = U^{\otimes t} \otimes U^{\dagger \otimes t}$.
It is natural to consider irreps of the unitary group.
A well-known fact is that each irrep can be indexed by a non-increasing integer sequence $\lambda :=(\lambda_1, \lambda_2, \dots, \lambda_d)$, i.e. $\lambda_1 \geq \lambda_2 \geq \dots \geq \lambda_d$, of length $d$.
In particular, each irrep in $U^{\otimes t} \otimes U^{\dagger \otimes t}$ can be indexed by an element of a set $\Lambda(d,t)$ defined by
\begin{multline}
\Lambda(d, t) := \\
\{\lambda= (\lambda_1, \lambda_2, \dots, \lambda_d)| \lambda_1 \geq \dots \geq \lambda_d, \lambda^+ = \lambda^- \leq t\},
\end{multline}
where $\lambda^+$ and $\lambda^-$ are the absolute value of sum of positive and negative $\lambda_i$'s, respectively.
Using this notation, the representation space ${\mathcal H}_d^{\otimes 2t}$ is irreducibly decomposed into
\begin{equation}
{\mathcal H}_d^{\otimes 2t} = \bigoplus_{\lambda \in \Lambda(d, t)} V_{\lambda}^{\oplus m_{\lambda}}, \label{Eq:IrrepsVectorSpace}
\end{equation}
where $m_{\lambda}$ is the multiplicity of the irrep $\lambda$.
Accordingly, the map $\rho$ is also decomposed into the irreducible ones $\rho_{\lambda}$.
Based on the irrep $(\rho_{\lambda}, V_{\lambda})$ of the unitary group, a unitary $t$-design ${\sf U}_t(d)$ can be characterized in a representation-theoretic manner: for any $\lambda \in \Lambda(d,t)$,
\begin{equation}
\mathbb{E}_{U \sim {\sf U}_t(d)} [\rho_{\lambda}(U)]
=
\mathbb{E}_{U \sim {\sf H}(d)} [\rho_{\lambda}(U)]. \label{Eq:RepDesign}
\end{equation}
The strong unitary $t$-designs are similarly characterized in terms of irreps~\cite{RS2009}. To this end, let $\Lambda_{\leq}(d,t)$ be
\begin{equation}
\Lambda_{\leq}(d, t) := \{\lambda= (\lambda_1, \lambda_2, \dots, \lambda_d)| \lambda_1 \geq \dots \geq \lambda_d, \lambda^{\pm} \leq t\},
\end{equation}
where $\lambda^+$ is not necessarily equal to $\lambda^-$. Then, a strong unitary $t$-design ${\sf U}_{\leq t}(d)$ satisfies
\begin{equation}
\mathbb{E}_{U \sim {\sf U}_{\leq t}} [\rho_{\lambda}(U)]
=
\mathbb{E}_{U \sim {\sf H}} [\rho_{\lambda}(U)], \label{Eq:RepHaarAv}
\end{equation}
for any $\lambda \in \Lambda_{\leq}(d,t)$.
One of the merits in this characterization is that the right-hand-sides of Eqs.\,\eqref{Eq:RepDesign} and~\eqref{Eq:RepHaarAv} are zero for all non-trivial irreps due to the Schur's orthogonality relation, which states that, for any unitarily inequivalent irreps $\lambda$ and $\lambda'$,
\begin{equation}
\mathbb{E}_{U \sim {\sf H}} \bigl[(\rho_{\lambda}(U))_{ij} (\rho_{\lambda'}(U))_{i'j'} \bigr] = 0,
\end{equation}
for any $i,j, i', j'$, where $(\rho_{\lambda}(U))_{ij}$ is the $(i,j)$ element of the matrix.
By setting the irrep $\rho_{\lambda'}$ to a trivial irrep, i.e., $\rho_{\lambda'}(U) = 1$ for any $U \in {\sf U}(d)$, we have
\begin{equation}
\mathbb{E}_{U \sim {\sf H}} \bigl[\rho_{\lambda}(U) \bigr] = 0,
\end{equation}
for any non-trivial irrep $\lambda$.
On the other hand, for any trivial irrep $\lambda$, it is trivial that
\begin{equation}
\mathbb{E}_{U \sim {\sf H}} \bigl[\rho_{\lambda}(U) \bigr] = 1.
\end{equation}
From these facts, (strong) unitary $t$-designs can be defined in terms of representation as follows:
\begin{Definition}[Unitary designs in representation theory] \label{Def:zero}
An ensemble ${\sf U}_{t}(d)$ of unitaries is an exact unitary $t$-design if it holds for any irrep $\rho_{\lambda}$ with $\lambda \in \Lambda(d,t)$ that
\begin{equation}
\mathbb{E}_{U \sim {\sf U}_t(d)} [\rho_{\lambda}(U)] =
\begin{cases}
1 & \text{if the irrep is trivial,}\\
0 & \text{otherwise.}
\end{cases} \label{Eq:aaartcct}
\end{equation}
An ensemble ${\sf U}_{\leq t}(d)$ is a strong unitary $t$-design if Eq.\,\eqref{Eq:aaartcct} holds for any irrep $\rho_{\lambda}$ with $\lambda \in \Lambda_{\leq}(d,t)$.
\end{Definition}
\subsection{Proof of Theorem~\ref{Thm:General}} \label{SS:ProofThm1}
We now prove Theorem~\ref{Thm:General}, which states that ${\sf W}_d$ defined by
\begin{equation}
{\sf W}_d := {\sf W}_{d_1 \oplus d-d_1} \prod_{\lambda \in \Lambda_{\rm sph}(d_1,d,t)} (R_{\lambda} {\sf W}_{d_1 \oplus d-d_1}),
\end{equation}
is a strong unitary $t$-design on ${\sf U}(d)$. Here,
\begin{align}
&{\sf W}_{d_1 \oplus d-d_1}= \{ U \oplus V | U \in {\sf U}_{\leq t}(d_1), V \in {\sf U}_{\leq t}(d-d_1)\},
\end{align}
where ${\sf U}_{\leq t}(d)$ and ${\sf U}_{\leq t}(d-d_1)$ are strong unitary $t$-designs on ${\sf U}(d)$ and ${\sf U}(d-d_1)$, respectively, and $R_{\lambda}$ is constructed by solving the zonal spherical function $Z_{\lambda}$.
It suffices to show
\begin{equation}
\mathbb{E}_{U \sim {\sf W}_d} [\rho_{\lambda}(U)] = 0, \label{Eq:Goal}
\end{equation}
for all non-trivial irreps indexed by $\lambda \in \Lambda_{\leq}(d, t)$.
Note that the average over $U \sim {\sf W}_d$ consists of the independent averages over all ${\sf W}_{d_1 \oplus d-d_1}$, further consisting of those over the strong unitary $t$-designs ${\sf U}_{\leq t}(d_1)$ and ${\sf U}_{\leq t}(d-d_1)$.
Let us first fix a non-trivial irrep $\lambda \in \Lambda_{\leq}(d,t)$ and consider $W_{\lambda}$ defined by
\begin{equation}
W_{\lambda} := \mathbb{E}_{U \sim {\sf W}_{d_1 \oplus d-d_1}} [\rho_{\lambda}(U)].
\end{equation}
Since we consider only irreps $\lambda \in \Lambda_{\leq}(d, t)$, this average can be replaced with the averages over the product ${\sf H}(d_1) \times {\sf H}(d-d_1)$ of the Haar measures on ${\sf K}:= {\sf U}(d_1) \times {\sf U}(d - d_1)$. That is,
\begin{align}
W_{\lambda} &=\mathbb{E}_{U \sim {\sf H}(d_1) \times {\sf H}(d-d_1)} [\rho_{\lambda}(U)].
\end{align}
To investigate $W_{\lambda}$, we consider the irreps of ${\sf K}$.
Since ${\sf K}$ is a subgroup of ${\sf U}(d)$, each irreducible space $V_{\lambda}$ of ${\sf U}(d)$ is decomposed into a direct sum of those of irreps of ${\sf K}$.
For the same reason as in Definition~\ref{Def:zero}, every non-trivial irrep of ${\sf K}$ becomes zero by taking the average over ${\sf H}(d_1) \times {\sf H}(d-d_1)$.
Hence, if the non-trivial irreducible representation space $V_{\lambda}$ of ${\sf U}(d)$ does not contain trivial irreps of ${\sf K}$, $W_{\lambda} = 0$.
In contrast, if a non-trivial irrep $\lambda$ of ${\sf U}(d)$ contains trivial irreps of ${\sf K}$, then the matrix elements of $W_{\lambda}$ corresponding to the trivial irreps of ${\sf K}$ are one, and the others are zero.
Trivial irreps of ${\sf K}$ in a non-trivial irrep $\lambda$ of ${\sf U}(d)$ were studied in a great detail since $({\sf K}, {\sf U}(d))$ is an example of a Gelfand pair~\cite{T1994,W2007}. It is known that the irreps of ${\sf U}(d)$ indexed by $\lambda \in \Lambda_{\rm sph}(d_1,d,t)$ contains only one trivial irreps of ${\sf K}$, and that other irreps of ${\sf U}(d)$ contain no trivial irrep of ${\sf K}$~\cite{GW2009}.
Since trivial irreps are one-dimensional, we denote by $\ket{w_{\lambda}} \in V_{\lambda}$ a unit vector that spans the trivial irrep of ${\sf K}$ in the spherical representation $\lambda \in \Lambda_{\rm sph}(d_1,d,t)$ of ${\sf U}(d)$.
Then, we have
\begin{equation}
W_{\lambda} =
\begin{cases}
0 & \text{\ if\ } \lambda \notin \Lambda_{\rm sph}(d_1,d,t),\\
\ketbra{w_{\lambda}}{w_{\lambda}} & \text{\ if\ } \lambda \in \Lambda_{\rm sph}(d_1,d,t).
\end{cases} \label{Eq:AvW}
\end{equation}
If $\lambda \notin \Lambda_{\rm sph}(d_1,d,t)$, we immediately obtain $\mathbb{E}_{U \sim {\sf W}_{d_1 \oplus d-d_1}} [\rho_{\lambda}(U)]=0$ from the definition of $W_{\lambda}$, which implies Eq.\,\eqref{Eq:Goal}.
If $\lambda \in \Lambda_{\rm sph}(d_1,d,t)$, we define a matrix $M_{\lambda}(U)$ on $V_{\lambda}$ ($U \in {\sf U}(d)$) by
\begin{align}
M_{\lambda}(U) &:= \mathbb{E}_{V_1, V_2 \sim {\sf H}(d_1) \times {\sf H}(d-d_1)} [\rho_{\lambda}(V_1 U V_2)].
\end{align}
Importantly, for any $\lambda \in \Lambda_{\rm sph}(d_1,d,t)$, there exists at least one $R_{\lambda} \in {\sf U}(d)$ such that $\bra{w_{\lambda}} M_{\lambda}(R_{\lambda}) \ket{w_{\lambda}}=0$. This follows from the fact that
\begin{align}
\mathbb{E}_{U \sim {\sf H}(d)}[ \bra{w_{\lambda}} M_{\lambda}(U) \ket{w_{\lambda}}]
&= \bra{w_{\lambda}} \mathbb{E}_{U \sim {\sf H}(d)}[ \rho_{\lambda}(U) ] \ket{w_{\lambda}}\\
&=0,
\end{align}
where we have used the unitary invariance of ${\sf H}(d)$ and that the irrep $\rho_{\lambda}$ is non-trivial, so that $\mathbb{E}_{U \sim {\sf H}(d)}[ \rho_{\lambda}(U) ]=0$.
Due to the intermediate value theorem, there always exists at least one unitary $R_{\lambda} \in {\sf U}(d)$ such that $\bra{w_{\lambda}} M_{\lambda}(R_{\lambda}) \ket{w_{\lambda}}=0$.
Using such $R_{\lambda} \in {\sf U}(d)$ and Eq.\,\eqref{Eq:AvW}, it is straightforward to observe that $W_{\lambda} M_{\lambda} (R_{\lambda}) W_{\lambda} = 0$.
Furthermore, it follows that
\begin{align}
W_{\lambda} M_{\lambda} (R_{\lambda}) W_{\lambda}
&=
\mathbb{E}_{U, U', V_1, V_2 \sim {\sf W}_{d_1 \oplus d-d_1}} [\rho_{\lambda}(U V_1 R_{\lambda} V_2 U')]\\
&=
\mathbb{E}_{W_1, W_2 \sim {\sf W}_{d_1 \oplus d-d_1}} [\rho_{\lambda }(W_1 R_{\lambda} W_2)]\\
&=
\mathbb{E}_{U \sim {\sf W}_{d_1 \oplus d-d_1} R_{\lambda} {\sf W}_{d_1 \oplus d-d_1}} [\rho_{\lambda}(U)].
\end{align}
We, hence, obtain
\begin{equation}
\mathbb{E}_{U \sim {\sf W}_{d_1 \oplus d-d_1} R_{\lambda} {\sf W}_{d_1 \oplus d-d_1}} [\rho_{\lambda}(U)] = 0. \label{Eq:eerrcr}
\end{equation}
Thus, the finite set of unitaries ${\sf W}_{d_1 \oplus d-d_1} R_{\lambda} {\sf W}_{d_1 \oplus d-d_1}$ satisfies the condition for the design, i.e., Eq.\,\eqref{Eq:Goal} for any non-trivial irrep $\lambda \in \Lambda_{\leq}(d,t)$, which leads to the statement that the set of unitaries defined by
\begin{equation}
{\sf W}_d = {\sf W}_{d_1 \oplus d-d_1} \prod_{\lambda \in \Lambda_{\rm sph}(d_1,d,t)} (R_{\lambda} {\sf W}_{d_1 \oplus d-d_1}),
\end{equation}
is a strong $t$-design on ${\sf U}(d)$.
Finally, let us clarify the relation between $R_{\lambda}$ and the zero of the zonal spherical function. To this end, we first observe that the matrix element $\bra{w_{\lambda}} M_{\lambda}(U) \ket{w_{\lambda}}$ of $M_{\lambda}$ is the zonal spherical function $Z_{\lambda}(U)$.
This can be checked by a simple calculation: for any $W_1, W_2 \in K$ and $U \in {\sf U}(d)$, we have
\begin{align}
Z_{\lambda}(W_1 U W_2)&= \bra{w_{\lambda}} \mathbb{E}[\rho_{\lambda} (V_1 W_1 U W_2V_2)] \ket{w_{\lambda}}\\
&= \bra{w_{\lambda}} \mathbb{E}[\rho_{\lambda}(V_1 U V_2)] \ket{w_{\lambda}} \label{Eq:ZSF}\\
&=Z_{\lambda}(U),
\end{align}
where the averages are all taken over $V_1, V_2 \sim {\sf H}(d_1) \times {\sf H}(d-d_1)$.
Thus, $Z_{\lambda}(U)$ is bi-${\sf K}$-invariant, and so, is the zonal spherical function. This implies that $R_{\lambda}$ is indeed a zero of the zonal spherical function.
Based on this fact, we can provide a matrix form of $R_{\lambda}$ in the fixed basis in which a unitary in ${\sf W}_{d_1 \oplus d-d_1}$ is represented as $U \oplus V$.
To this end, it is important to notice that the bi-{\sf K}-invariance of the zonal spherical function implies that it is characterized by the cosets of ${\sf K} = {\sf U}(d_1) \times {\sf U}(d-d_1)$ in ${\sf U}(d)$. The cosets can be further identified with $d_1$-dimensional subspaces corresponding to the support on which ${\sf U}(d_1)$ acts. For instance, the identity element in the coset of ${\sf K}$ corresponds to the subspace $V_0$ spanned by the first $d_1$ vectors of the fixed basis.
The matrix form of $R_{\lambda}$ is obtained by specifying the relation between $V_0$ and the subspace corresponding to another representative of the coset.
To characterize the relation between two subspaces, we use the \emph{principal} angles. For two subspaces $X$ and $Y$, let us refer to $\theta = \min {\rm argcos} |\braket{x}{y}|$, where the minimum is taken over all unit vectors $\ket{x} \in X, \ket{y} \in Y$, as the minimum angle between $X$ and $Y$.
The principal angles $(\theta_0, \dots, \theta_{m-1})$ between two $m$-dimensional subspaces $X$ and $Y$ are then defined as follows: $\theta_0$ is the minimum angle between $X$ and $Y$, and $\theta_{i+1}$ is the minimum angle between $X \cap {\rm span}\{\ket{x_0}, \dots, \ket{x_{i}}\}$ and $Y \cap {\rm span}\{\ket{y_0}, \dots, \ket{y_{i}}\}$, where $(\ket{x_j}, \ket{y_j})$ is a pair of the unit vectors that leads to $\theta_j$.
The cosine of the principal angles between $V_0$ and the subspace corresponding to another representative in the coset determines the value of the zonal spherical function $Z_{\lambda}$, and so, $Z_{\lambda}$ can be written as $Z_{\lambda}(\cos^2 \theta_0, \dots, \cos^2 \theta_{d_1-1})$~\cite{R2010,BNOZ2020}. See, e.g., Refs.\,\cite{JC1974,R2010,BNOZ2020} for the explicit form of $Z_{\lambda}$ as a polynomial of $(\cos^2 \theta_1, \dots, \cos^2 \theta_{d_1})$.
By solving the polynomial, we obtain the principal angles $(\theta^{(0)}_{\lambda}, \dots, \theta^{(d_1-1)}_{\lambda})$ between $V_0$ and the subspace corresponding to the zero of $Z_{\lambda}$.
Recalling the definition of the principal angles and using the left- and right-invariance of the coset by any unitary in ${\sf K}$, we can take a matrix form of $R_{\lambda}$ as follows:
\begin{equation}
R_{\lambda} =
\begin{pmatrix}
C(\boldsymbol{\theta}_{\lambda}) & i S(\boldsymbol{\theta}_{\lambda}) & 0 \\
i S(\boldsymbol{\theta}_{\lambda}) & C(\boldsymbol{\theta}_{\lambda}) & 0 \\
0 & 0 & I_{d-2d_1}
\end{pmatrix},
\end{equation}
where $C(\boldsymbol{\theta}_{\lambda}) = {\rm diag}(\cos\theta_{\lambda}^{(0)}, \dots, \cos\theta_{\lambda}^{(d_1-1)})$ and $S(\boldsymbol{\theta}_{\lambda}) = {\rm diag}(\sin\theta_{\lambda}^{(0)}, \dots, \sin\theta_{\lambda}^{(d_1-1)})$, and $I_{d-2d_1}$ is the identity matrix of size $d- 2d_1$. Note that $R_{\lambda}$ is not necessarily in this form since the coset is invariant under the action of ${\sf K}$.
$\hfill \blacksquare$
\section{Higher-order RB} \label{S:higher-orderRB}
In this section, we investigate the higher-order RB in detail.
We begin with a preliminary in Subsec.\,\ref{SS:L} and explain several basic properties of the self-adjointness in Subsec.\,\ref{SS:PropSA}. We consider the $t$-RB for general $t$ and the $2$-RB in Subsecs.\,\ref{SS:generaltRB} and~\ref{SS:2RBdetail}, respectively.
\subsection{Liouville representation} \label{SS:L}
Let $\sigma_0 = I/\sqrt{2}$, $\sigma_1 = X/\sqrt{2}$, $\sigma_2 = Y/\sqrt{2}$, and $\sigma_3 = Z/\sqrt{2}$ be normalized Pauli operators on one qubit, where normalization is in terms of the Hilbert-Schmidt inner product.
For $q$ qubits, we introduce a vector $\vec{n} = (n_1, n_2, \dots, n_{q})$ ($n_i \in \{0,1,2,3\}$) and use the notation that
\begin{equation}
\sigma_{\vec{n}}:= \sigma_{n_1} \otimes \dots \otimes \sigma_{n_{q}}.
\end{equation}
We also denote $2^q$ by $d$ in this section.
The Liouville representation is a matrix representation of quantum channels, also known as the Pauli transfer matrix. See, e.g., Refs.\,\cite{WGHF2015,KLDF2016,DHW2019}.
Let $\kett{\cdot}$ be a linear map from a set of all linear operators on a $d$-dimensional Hilbert space to a $d^2$-dimensional vector space that specifically maps $\sigma_{\vec{n}}$ to a canonical orthonormal basis vector $e_{\vec{n}}$.
Since the map is linear, we have
\begin{equation}
\kett{A} := \sum_{\vec{n}} \tr[\sigma_{\vec{n}} A] \kett{\sigma_{\vec{n}}},
\end{equation}
for any linear operator $A$. Note that $\braakett{A}{B} = \tr[A^{\dagger} B]$.
Based on this vector representation of linear operators, a linear supermap $\mathcal{E}$ can be represented by a matrix.
The Liouville representation of a linear supermap $\mathcal{E}$ is defined by
\begin{equation}
L_{\mathcal{E}} := \sum_{\vec{n}} \kettbraa{\mathcal{E}(\sigma_{\vec{n}})}{\sigma_{\vec{n}}},
\end{equation}
which is a regular matrix of size $d^2$.
The matrix element in the canonical basis of $\{\kett{\sigma_{\vec{n}}}\}_{\vec{n}}$ is given by
\begin{equation}
\bigl( L_{\mathcal{E}} \bigr)_{\vec{n} \vec{m}} = \braakett{\sigma_{\vec{n}}}{\mathcal{E}(\sigma_{\vec{m}})} = \tr[\sigma_{\vec{n}} \mathcal{E}(\sigma_{\vec{m}})].
\end{equation}
The vector and Liouville representations satisfy the following properties:
\begin{enumerate}
\item $L_{\mathcal{E}} \kett{\rho} = \kett{\mathcal{E}(\rho)}$,
\item $L_{\mathcal{E}_2 \circ \mathcal{E}_1} = L_{\mathcal{E}_2}L_{\mathcal{E}_1}$,
\item $L_{\alpha \mathcal{E}_1 + \beta \mathcal{E}_2} = \alpha L_{\mathcal{E}_1} + \beta L_{\mathcal{E}_2}$ ($\alpha, \beta \in \mathbb{C}$),
\item $L_{\mathcal{E}_1 \otimes \mathcal{E}_2} = L_{\mathcal{E}_1} \otimes L_{\mathcal{E}_2}$,
\item $L_{\mathcal{E}^{\dagger}} = L_{\mathcal{E}}^{\dagger}$.
\end{enumerate}
Properties of a linear supermap $\mathcal{E}$ can be also expressed in terms of the Liouville representation. For instance, the linear map $\mathcal{E}$ is TP if and only if $(L_{\mathcal{E}})_{\vec{0}\vec{0}}=1$ and $(L_{\mathcal{E}})_{\vec{0}\vec{n}} = 0$ for any $\vec{n} \neq \vec{0}$.
Since we are interested in the CPTP map $\mathcal{E}$ that represents a noise, its Liouville representation is always in the form of
\begin{equation}
L_{\mathcal{E}} =
\begin{pmatrix}
1 & 0 \\
\alpha_{\mathcal{E}} & \tilde{L}_{\mathcal{E}}
\end{pmatrix}, \label{Eq:Li}
\end{equation}
where $0$ is a row vector of length $d^2-1$ with all elements being zero, $\alpha_{\mathcal{E}}$ is a column vector of length $d^2-1$, called a \emph{non-unital part} of the noise, and $\tilde{L}_{\mathcal{E}}$ is a $(d^2-1) \times (d^2-1)$ matrix.
The non-unital part $\alpha_{\mathcal{E}}$ of the noise is the zero vector if and only if the map $\mathcal{E}$ is unital, i.e., $\mathcal{E}(I)=I$ with $I$ being the identity operator.
In the Liouville representation, the fidelity parameter $f(\mathcal{E})$ and the unitarity $u(\mathcal{E})$ of a noisy CPTP map $\mathcal{E}$ are given by
\begin{align}
f(\mathcal{E}) &= \frac{1}{d^2-1} \sum_{\vec{n} \neq \vec{0}} \braa{\sigma_{\vec{n}}} L_{\mathcal{E}} \kett{\sigma_{\vec{n}}}\\
&= \frac{1}{d^2-1} \tr[\tilde{L}_{\mathcal{E}}],\\
u(\mathcal{E}) &= \frac{1}{d^2-1} \sum_{\vec{n} \neq \vec{0}} \braa{\sigma_{\vec{n}}} L_{\mathcal{E}}^{\dagger} L_{\mathcal{E}} \kett{\sigma_{\vec{n}}},\\
&= \frac{1}{d^2-1} \tr[\tilde{L}_{\mathcal{E}}^{\dagger} \tilde{L}_{\mathcal{E}}],
\end{align}
respectively.
\subsection{Properties of the self-adjointness} \label{SS:PropSA}
For a CPTP map $\mathcal{E}$, the self-adjointness $H(\mathcal{E})$ and the self-adjointness parameter $h(\mathcal{E})$ are defined by
\begin{align}
H(\mathcal{E}) &:= 1- \frac{d+1}{2d} \int d \varphi \bigl|\! \bigr| \mathcal{E}( \ketbra{\varphi}{\varphi}) - \mathcal{E}^{\dagger} ( \ketbra{\varphi}{\varphi}) \bigl|\! \bigr|_2^2,\\
h(\mathcal{E}) &:= \frac{d}{d-1} \int d \varphi \tr[\mathcal{E}'(\varphi) \mathcal{E}'^{\dagger}(\varphi)],\\
&= \frac{1}{d^2-1} \tr[ \tilde{L}_{\mathcal{E}}^2], \label{Eq:B1rr}
\end{align}
where $\mathcal{E}'(\rho) = \mathcal{E}(\rho- I/d)$, and the last line is shown in Appendix~\ref{App:CharaNoise}.
We first show the relation between $H(\mathcal{E})$ and $h(\mathcal{E})$, i.e., Eq.\,\eqref{Eq:66} in Subsec.\,\ref{SS:defdef}:
\begin{align}
H(\mathcal{E}) &= 1 - \frac{d^2-1}{d^2}\bigl( u(\mathcal{E}) - h(\mathcal{E}) \bigr) - \frac{d+1}{2d^2} | \alpha_{\mathcal{E}} |^2. \label{Eq:ice}
\end{align}
From the definition of $H(\mathcal{E})$, we have
\begin{multline}
\frac{2d}{d+1} \bigl(1- H(\mathcal{E})\bigr)\\
= \int d \varphi \biggl[ \tr[ \mathcal{E}( \ketbra{\varphi}{\varphi})^2] + \tr[ \mathcal{E}^{\dagger}( \ketbra{\varphi}{\varphi})^2] \\-2 \tr[\mathcal{E}( \ketbra{\varphi}{\varphi})\mathcal{E}^{\dagger}( \ketbra{\varphi}{\varphi})] \biggr]
\end{multline}
By rewriting $\mathcal{E}$ with $\mathcal{E}'$, the first term in the right-hand side is expressed in terms of the unitarity $u(\mathcal{E})$, such as
\begin{align}
&\int d \varphi \tr[ \mathcal{E}( \ketbra{\varphi}{\varphi})^2] = \frac{d-1}{d} u(\mathcal{E}) + \tr\bigl[ \mathcal{E} (I/d)^2 \bigr].
\end{align}
By using the swap operator $\mathbb{F}:= \sum_{\vec{n}} \sigma_{\vec{n}} \otimes \sigma_{\vec{n}}$, and the property that $\tr[MN]=\tr[\mathbb{F} (M \otimes N)]$ for any matrices $M$ and $N$, which is called a \emph{swap trick}, it follows that
\begin{align}
\tr\bigl[ \mathcal{E} (I/d)^2 \bigr] &= \tr\bigl[ \mathbb{F} \mathcal{E}(I/d)^{\otimes 2} \bigr]\\
&= \frac{1}{d}\sum_{\vec{n}} \braa{\sigma_{\vec{n}}} L_{\mathcal{E}} \kett{\sigma_{\vec{0}}}^2\\
&= \frac{1}{d} | \alpha_{\mathcal{E}} |^2 + \frac{1}{d},
\end{align}
which leads to
\begin{align}
&\int d \varphi \tr[ \mathcal{E}( \ketbra{\varphi}{\varphi})^2] = \frac{d-1}{d} u(\mathcal{E}) + \frac{1}{d} | \alpha_{\mathcal{E}} |^2 + \frac{1}{d}.
\end{align}
Similarly, we obtain
\begin{align}
&\int d \varphi \tr[ \mathcal{E}^{\dagger}( \ketbra{\varphi}{\varphi})^2] = \frac{d-1}{d} u(\mathcal{E}) + \frac{1}{d},
\end{align}
from the facts that $L_{\mathcal{E}^{\dagger}} = L_{\mathcal{E}}^{\dagger}$ and that $| \alpha_{\mathcal{E}^{\dagger}} | = 0$ for any TP map $\mathcal{E}$.
From the definition of $\mathcal{E}'$, it is straightforward to show that the self-adjointness parameter $h(\mathcal{E})$ is given by
\begin{align}
h(\mathcal{E}) &=\frac{1}{d-1}\biggl[ d \int \tr \bigl[ \mathcal{E}(\varphi) \mathcal{E}^{\dagger}(\varphi) \bigr] d\varphi - 1 \biggr]. \label{Eq:B12rr}
\end{align}
Combining these altogether, we arrive at
\begin{equation}
\frac{2d}{d+1} \bigl(1- H(\mathcal{E})\bigr)
=
\frac{2(d-1)}{d} \bigl( u(\mathcal{E}) - h(\mathcal{E}) \bigr) + \frac{1}{d}| \alpha_{\mathcal{E}} |^2,
\end{equation}
implying Eq.\,\eqref{Eq:ice}.\\
The self-adjointness parameter also satisfies the following properties. They are all shown in Appendix~\ref{App:CharaNoise}.
\begin{enumerate}
\item $-\frac{1}{d^2-1} \leq h(\mathcal{E}) \leq u(\mathcal{E})$.
\item $h(\mathcal{E}) = u(\mathcal{E})$ if and only if $\tilde{L}_{\mathcal{E}} = \tilde{L}_{\mathcal{E}}^{\dagger}$. For a unital noise $\mathcal{E}$, $h(\mathcal{E}) = u(\mathcal{E})$ if and only if the noise is self-adjoint ($\mathcal{E} = \mathcal{E}^{\dagger}$).
\item $h(\mathcal{E}) =-\frac{1}{d^2-1}$ if and only if $\tr[K_iK_j] = 0$ for any $i, j$, where $\{ K_i \}$ are the Kraus operators of $\mathcal{E}$.
\item the average gate fidelity $F(\mathcal{E})$ is bounded from above by $u(\mathcal{E})$ and $h(\mathcal{E})$:
\begin{equation}F(\mathcal{E}) \leq \frac{d-1}{d} \sqrt{\frac{h(\mathcal{E}) + u(\mathcal{E})}{2}} + \frac{1}{d}. \label{Eq:97779}\end{equation}
\end{enumerate}
\subsection{A general expression for the $t$-RB} \label{SS:generaltRB}
We here show that the expectation value $V^{(t)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas})$ in the $t$-RB has a general form of
\begin{equation}
V^{(t)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas})
= \sum_{\lambda} \tr \bigl[ \hat{A}_{\lambda} (\hat{C}_{\lambda}(\mathcal{E}))^m\bigr], \label{Eq:general2}
\end{equation}
where $\hat{A}_{\lambda}$ is a regular matrix depending on $O_{\rm ini}$ and $\mathcal{E}(O_{\rm meas})$, and $\hat{C}_{\lambda}(\mathcal{E})$ is a regular matrix depending only on $\mathcal{E}$. As we will see below, $\lambda$ labels the irreps of a $t$-copy representation of the unitary group, and the size of the matrices is equal to the multiplicity of each irrep.
The expectation value is defined by
\begin{multline}
V^{(t)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas}) :=\\
\mathbb{E}_{U_{\boldsymbol{i}}} \bigl[ \bigl(\tr \bigl[ O_{\rm meas} {\cal G}_{i_{m+1}} \circ {\cal G}_{i_m} \circ \dots \circ {\cal G}_{i_1}(O_{\rm ini}) \bigr] \bigr)^t \bigr],
\end{multline}
where $\mathbb{E}_{U_{\boldsymbol{i}}}$ is the average over all unitary sequences $U_{\boldsymbol{i}} \sim {\sf U}_{2t}^{\times m}$. Note that ${\cal G}_i = \mathcal{E} \circ \mathcal{U}_i$ and that $\mathcal{U}_i$ is the unitary channel defined by $\mathcal{U}_i(\rho) = U_i \rho U_i^{\dagger}$.
In terms of the Liouville representation, we have
\begin{multline}
\tr \bigl[ O_{\rm meas} {\cal G}_{i_{m+1}} \circ {\cal G}_{i_m} \circ \dots \circ {\cal G}_{i_1}(O_{\rm ini}) \bigr]\\ = \braa{O'_{\rm meas}} L_{\mathcal{U}'_m \circ \mathcal{E} \circ \mathcal{U}^{'\dagger}_m} \dots L_{\mathcal{U}'_1 \circ \mathcal{E} \circ \mathcal{U}^{'\dagger}_1} \kett{O_{\rm ini}},
\end{multline}
where we have used that ${\cal G}_i = \mathcal{E} \circ \mathcal{U}_i$, $\mathcal{U}'_n = \mathcal{U}_n \circ \mathcal{U}_{n-1} \circ \dots \circ \mathcal{U}_2 \circ \mathcal{U}_1$, and $O'_{\rm meas} = \mathcal{E}(O_{\rm meas})$.
Noticing the $t$-th power and the fact that each unitary is independently chosen from a unitary $2t$-design ${\sf U}_{2t}$, we obtain
\begin{align}
&V^{(t)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas})= \bigl(\braa{O'_{\rm meas}}^{\otimes t} \bigr) (L_{\rm av})^m \bigl( \kett{O_{\rm ini}}^{\otimes t} \bigr), \label{Eq:kore}
\end{align}
where $L_{\rm av}$ is defined by
\begin{align}
L_{\rm av} : &= \mathbb{E}_{U \sim {\sf U}_{2t}}[ (L_{\mathcal{U} \circ \mathcal{E} \circ \mathcal{U}^{\dagger}})^{\otimes t}],\\
&=\mathbb{E}_{U \sim {\sf U}_{2t}}[ (L_{\mathcal{U}} L_{\mathcal{E}} L_{\mathcal{U}^{\dagger}})^{\otimes t}]\\
&= \mathbb{E}_{U \sim {\sf H}}[ (L_{\mathcal{U}} L_{\mathcal{E}} L_{\mathcal{U}^{\dagger}})^{\otimes t}].
\end{align}
The last line follows since ${\sf U}_{2t}$ is an exact unitary $2t$-design.
To write down $L_{\rm av}$ explicitly, let us consider the tensor-$t$ Liouville representation given by
\begin{equation}
{\sf U}(d) \ni U \rightarrow L_{\mathcal{U}^{\otimes t}} \in GL({\cal K}),
\end{equation}
where $GL({\cal K})$ is the general linear group acting on the $d^q$-dimensional vector space ${\cal K}$ defined by
\begin{equation}
{\cal K} := {\rm span}\{ \bigotimes_{s = 1}^t \kett{\sigma_{\vec{n}_s}} : \vec{n}_s \in \{0,1,2,3 \}^{q}, s \in [1, t] \}. \label{Eq:repspace}
\end{equation}
We denote the irreducible decomposition by
\begin{equation}
{\cal K} = \bigoplus_{\lambda} {\cal K}_{\lambda}^{\oplus m_{\lambda}},
\end{equation}
where $\lambda$ labels the irreps, and $m_{\lambda}$ is the multiplicity of the irrep labeled by $\lambda$.
The key observation is that
\begin{equation}
\forall V \in {\sf U}(d), \quad [L_{\rm av}, L_{\mathcal{V}^{\otimes t}}] = 0,
\end{equation}
which simply follows from the unitary invariance of the Haar measure. This implies that $L_{\rm av} \in {\rm End}_{\sf U}({\cal K})$, where ${\rm End}_{\sf U}({\cal K})$ is a set of all endomorphisms of ${\cal K}$ that commute with the tensor-$t$ Liouville action of ${\sf U}(d)$.
It is well-known that ${\rm End}_{\sf U}({\cal K})$ is isomorphic to the direct sum of matrix algebras:
\begin{equation}
{\rm End}_{\sf U}({\cal K}) \simeq \bigoplus_{\lambda} M(m_{\lambda}, \mathbb{C}), \label{Eq:End}
\end{equation}
where $M(m_{\lambda}, \mathbb{C})$ is a set of all $m_{\lambda} \times m_{\lambda}$ matrices over $\mathbb{C}$.
Thus, the operator $L_{\rm av} \in {\rm End}_{\sf U}({\cal K})$ can be represented by a direct sum of matrices.
To obtain the explicit form of $L_{\rm av}$, let ${\cal K}_{\lambda}^{(1)} \oplus \dots \oplus {\cal K}_{\lambda}^{(m_{\lambda})}$ be a fixed decomposition of ${\cal K}_{\lambda}^{\oplus m_{\lambda}}$, and denote $\eta_{\lambda}^{p \rightarrow q}$ be the isomorphism from ${\cal K}_{\lambda}^{(p)}$ to ${\cal K}_{\lambda}^{(q)}$. We also denote by $\Pi_{\lambda}^{(p)}$ the projection onto ${\cal K}_{\lambda}^{(p)}$. Then, from the explicit form of the isomorphism, we have
\begin{equation}
L_{\rm av}
=
\sum_{\lambda} \sum_{p, q =1}^{m_{\lambda}} (\hat{C}_{\lambda}(\mathcal{E}))_{pq} \eta_{\lambda}^{p \rightarrow q} \Pi_{\lambda}^{(p)},\label{Eq:EndMat}
\end{equation}
where $\hat{C}_{\lambda}(\mathcal{E}) \in M(m_{\lambda}, \mathbb{C})$. Each element $\hat{C}_{\lambda}(\mathcal{E})$ is given by
\begin{align}
( \hat{C}_{\lambda} (\mathcal{E}) )_{pq} &= \tr[ L_{\rm av} \eta_{\lambda}^{q \rightarrow p} \Pi_{\lambda}^{(q)}],\\
&= \tr \bigl[ \mathbb{E}_{U \sim {\sf H}}[ (L_{\mathcal{U}} L_{\mathcal{E}} L_{\mathcal{U}^{\dagger}})^{\otimes t}]\eta_{\lambda}^{q \rightarrow p} \Pi_{\lambda}^{(q)} \bigr],\\
&= \mathbb{E}_{U \sim {\sf H}} \tr \bigl[ L_{\mathcal{E}}^{\otimes t} L_{\mathcal{U}^{\dagger}}^{\otimes t}\eta_{\lambda}^{q \rightarrow p} \Pi_{\lambda}^{(q)} L_{\mathcal{U}}^{\otimes t} \bigr],\\
&= \mathbb{E}_{U \sim {\sf H}} \tr \bigl[ L_{\mathcal{E}}^{\otimes t} \eta_{\lambda}^{q \rightarrow p} \Pi_{\lambda}^{(q)} \bigr],\\
&= \tr \bigl[ L_{\mathcal{E}}^{\otimes t} \eta_{\lambda}^{q \rightarrow p} \Pi_{\lambda}^{(q)} \bigr],
\end{align}
where we have used the irreducibility in the fourth line.
Consequently, it follows that
\begin{equation}
(L_{\rm av})^m
=
\sum_{\lambda} \sum_{p, q =1}^{m_{\lambda}} \bigl(\hat{C}_{\lambda}(\mathcal{E})^m)_{pq} \eta_{\lambda}^{p \rightarrow q} \Pi_{\lambda}^{(p)}.
\end{equation}
Substituting this into Eq.\,\eqref{Eq:kore}, we obtain
\begin{equation}
V^{(t)}(m, \mathcal{E}| O_{\rm ini}, O_{\rm meas})\\
=
\sum_{\lambda} \tr \bigl[ \hat{A}_{\lambda}( \hat{C}_{\lambda}(\mathcal{E}) )^m\bigr],
\end{equation}
where the $m_{\lambda} \times m_{\lambda}$ matrices $\hat{A}_{\lambda}$ are given by
\begin{equation}
( \hat{A}_{\lambda} )_{pq}
=
\braa{\mathcal{E}(O_{\rm meas})^{\otimes t}}
\eta_{\lambda}^{q \rightarrow p} \Pi_{\lambda}^{(q)}
\kett{O_{\rm ini}^{\otimes t}}.
\end{equation}
This completes the proof. \hfill $\blacksquare$
\subsection{The first-order RB} \label{SS:1RBdetail}
Let us briefly overview the $1$-RB using an exact unitary $2$-design, namely, the standard RB. We also explain how the result changes when the $2$-design is an approximate one rather than the exact one.
In the $1$-RB, the representation space is given by
\begin{equation}
{\cal K} = {\rm span} \bigl\{ \kett{\sigma_{\vec{n}}}: \vec{n} \in \{ 0,1,2,3\}^q \bigr \}.
\end{equation}
We need to find a irreducible decomposition of ${\cal K}$ under the action of a unitary group ${\sf U}(d)$ as $U \rightarrow L_{\mathcal{U}}$. The Liouville representation $L_{\mathcal{U}}$ is defined by $L_{\mathcal{U}}\kett{\rho} = \kett{U \rho U^{\dagger}}$.
Hence, ${\cal K}$ is irreducibly decomposed to
\begin{equation}
{\cal K} = {\cal K}_0 \oplus {\cal K}_1,
\end{equation}
where
\begin{align}
&{\cal K}_0={\rm span} \{ \kett{\sigma_{\vec{0}}} \},\\
&{\cal K}_1 = {\rm span}\{ \kett{\sigma_{\vec{n}}}: \vec{n} \in \{ 0,1,2,3\}^q, \vec{n} \neq \vec{0} \}.
\end{align}
Denoting by $\Pi_0$ and $\Pi_1$ projectors onto ${\cal K}_0$ and ${\cal K}_1$, respectively, we have
\begin{align}
L_{\rm av} &:= \mathbb{E}_{U \sim {\sf U}_2}[(L_{\mathcal{U} \circ \mathcal{E} \circ \mathcal{U}^{\dagger}})], \\
&= \Pi_{0} + f(\mathcal{E}) \Pi_{1}, \label{Eq:exactapp}
\end{align}
where $f(\mathcal{E})$ is the fidelity parameter. Note that $ {\sf U}_2$ is an exact unitary $2$-design.
We thus obtain that
\begin{equation}
V^{(1)}(m, \mathcal{E}| \Delta, O_{\rm meas}) = A_0 + A_1 f(\mathcal{E})^m, \label{Eq:korekoredetail}
\end{equation}
where $A_i = \braa{\mathcal{E}(O_{\rm meas})}\Pi_i \kett{O_{\rm ini}}$ for $i=0,1$.\\
When the $2$-design is an approximate one ${\sf U}^{(\epsilon)}_2$, Eq.\,\eqref{Eq:exactapp} holds only approximately. The degree of approximation depends on how we measure it, but we here assume that the design is $\epsilon$-approximate when Eq.\,\eqref{Eq:exactapp} holds up to $\epsilon$-approximation.
That is, we assume that
\begin{align}
L_{\rm av}^{(\epsilon)} &:= \mathbb{E}_{U \sim {\sf U}_2^{(\epsilon)}}[(L_{\mathcal{U} \circ \mathcal{E} \circ \mathcal{U}^{\dagger}})],\\
&= L_{\rm av} + \epsilon \Delta,
\end{align}
where $\Delta$ is some operator of $O(1)$.
Note that standard definitions of approximate designs require harder criteria (see, e.g., Ref.\,\cite{L2010}).
In this case, instead of Eq.\, \eqref{Eq:korekoredetail}, we have
\begin{multline}
V^{(1)}(m, \mathcal{E}| \Delta, O_{\rm meas})
=
A_0 + E+ A_1 f^m \\
+ \epsilon (m-2) (E_2 f^2 + E_1 f + E_0) + O(m^2 \epsilon^2), \label{Eq:koredamedetail}
\end{multline}
where
\begin{align}
&E_2 = \braa{\mathcal{E}(O_{\rm meas})}\Pi_0 \Delta \Pi_0 \kett{O_{\rm ini}},\\
&E_1 =\braa{\mathcal{E}(O_{\rm meas})}( \Pi_0 \Delta \Pi_1 + \Pi_1 \Delta \Pi_0) \kett{O_{\rm ini}},\\
&E_0 = \braa{\mathcal{E}(O_{\rm meas})}\Pi_1 \Delta \Pi_1 \kett{O_{\rm ini}},\\
&E = \braa{\mathcal{E}(O_{\rm meas})} \{\Pi_0, \Delta\} + f \{\Pi_1, \Delta\} \kett{O_{\rm ini}}.
\end{align}
Comparing Eqs.\,\eqref{Eq:korekoredetail} and~\eqref{Eq:koredamedetail}, we observe that using approximate unitary $2$-designs result in more complicated form or the fitting function.
\subsection{The second-order RB} \label{SS:2RBdetail}
We now focus on the $2$-RB. Although the representation space in this case is
\begin{equation}
{\cal K} = {\rm span} \bigl\{ \kett{\sigma_{\vec{n}_1 \otimes \vec{n}_2}} : \vec{n}_1, \vec{n}_2 \in \{0,1,2,3 \}^{q} \bigr\},\label{Eq:erccc}
\end{equation}
where we have used the notation that $\sigma_{\vec{n}_1 \otimes \vec{n}_2} = \sigma_{\vec{n}_1} \otimes \sigma_{\vec{n}_2}$,
it is not necessary to consider the whole space because we assume that the initial operator $\Delta$ is traceless. This, together with the fact that the noise map is trace-preserving, implies that the operator remains traceless during the whole process. We also observe that the whole process is symmetric under the exchange of the first and the second spaces, each labeled by $\vec{n}_1$ and $\vec{n}_2$ in Eq.\,\eqref{Eq:erccc}. Hence, in the analysis of the $2$-RB, the relevant space is only the \emph{traceless symmetric} subspace defined by
\begin{multline}
{\cal K}_{TS} := {\rm span}\{ \kett{\sigma_{\vec{n}_1 \otimes \vec{n}_2} + \sigma_{\vec{n}_2 \otimes \vec{n}_1}}\\
: \vec{n}_1, \vec{n}_2 \in \{0,1,2,3 \}^{q}, (\vec{n}_1, \vec{n}_2) \neq (\vec{0}, \vec{0}) \},
\end{multline}
where $\vec{0} = (0, \dots, 0)$.
The irreducible decomposition of ${\cal K}_{TS}$ can be obtained by an extensive use of the result in Ref.\,\cite{HWW2018} (see Appendix~\ref{App:Irreps}), based on which we explicitly compute $V^{(2)}(m, \mathcal{E}| \Delta, O_{\rm meas})$.
It turns out that the situation differs depending on whether $q=1$ or $q \geq 2$. We, hence, deal with the two cases separately.
\subsubsection{$2$-RB in a single-qubit system}
When $q = 1$, the irreducible decomposition of ${\cal K}_{TS}$ is given by
\begin{equation}
{\cal K}_{TS} = {\cal K}_0 \oplus {\cal K}_1,
\end{equation}
which is multiplicity-free. Here, ${\cal K}_0$ and ${\cal K}_1$ are
\begin{align}
&{\cal K}_0 := {\rm span} \{ \kett{\sigma_{1}^{\otimes 2} + \sigma_{2}^{\otimes 2} + \sigma_{3}^{\otimes 2}} \}, \label{Eq:1qv0}\\
&{\cal K}_1 := {\rm span} \{ \kett{S_{1,2}}, \kett{S_{1,3}}, \kett{S_{2,3}}, \notag \\
&\hspace{15mm} \kett{\sigma_1^{\otimes 2} - 2 \sigma_2^{\otimes 2} + \sigma_3^{\otimes 2}}, \kett{\sigma_1^{\otimes 2} - \sigma_3^{\otimes 2}} \}, \label{Eq:1qv1}
\end{align}
respectively, with $S_{n, m} := (\sigma_n \otimes \sigma_m + \sigma_m \otimes \sigma_n)/\sqrt{2}$. It is obvious that $d_0 := \dim {\cal K}_0 = 1$ and $d_1 := \dim {\cal K}_1 = 5$.
This decomposition implies that the expectation $V^{(2)}(m, \mathcal{E}| \Delta, O_{\rm meas})$ is in the form of
\begin{equation}
V^{(2)}(m, \mathcal{E}| \Delta, O_{\rm meas}) = A_0 C_0(\mathcal{E})^m + A_1 C_1(\mathcal{E})^m,
\end{equation}
where both $A_{\lambda}$ and $C_{\lambda}$ are given by
\begin{align}
&A_{\lambda} = \braa{\mathcal{E}(O_{\rm meas})^{\otimes 2}} \Pi_{\lambda} \kett{\Delta^{\otimes 2}} \label{Eq:Ai1q},\\
&C_{\lambda} (\mathcal{E}) = \frac{\tr[ \Pi_{\lambda} L_{\mathcal{E}}^{\otimes 2} ]}{\tr[\Pi_{\lambda}]},
\end{align}
with $\Pi_{\lambda}$ being the projections onto ${\cal K}_{\lambda}$. Since the projections can be explicitly constructed from Eqs.\,\eqref{Eq:1qv0} and~\eqref{Eq:1qv1}, we can compute $C_{\lambda}(\mathcal{E})$.
First, we have
\begin{align}
C_0(\mathcal{E}) &= \frac{1}{3} \sum_{n, m=1}^3 \bigl(\braa{\sigma_n} L_{\mathcal{E}} \kett{\sigma_m}\bigr)^2,\\
&=\frac{1}{3} \sum_{n=1}^3 \braa{\sigma_n} L_{\mathcal{E}}^{\dagger} L_{\mathcal{E}} \kett{\sigma_n},\\
&= u(\mathcal{E}). \label{Eq:unitarity}
\end{align}
For $C_1 (\mathcal{E})$, we start from the relation that
\begin{equation}
\Pi_1 = \Pi_{\rm sym} - \Pi_0 - \kettbraa{\sigma_0^{\otimes 2}}{\sigma_0^{\otimes 2}} - \sum_{n=1}^3\kettbraa{S_{0,n}}{S_{0,n}}, \label{Eq:1qubitRBstart}
\end{equation}
where $\Pi_{\rm sym}$ is the projection onto the symmetric subspace of ${\cal K}^{\otimes 2}$.
The projection $\Pi_{\rm sym}$ is also expressed by $(\mathbb{I} + \mathbb{F} )/2$. Here, $\mathbb{I}$ is the identity operator on ${\cal K}^{\otimes 2}$ and $\mathbb{F}$ is the swap operator on ${\cal K}^{\otimes 2}$ defined by $\sum_{n,m=0}^3 \kettbraa{\sigma_n}{\sigma_m} \otimes \kettbraa{\sigma_m}{\sigma_n}$. Using the swap trick, we have
\begin{align}
\tr[ \Pi_{\rm sym} L_{\mathcal{E}}^{\otimes 2}] =\frac{1}{2}\bigl( \tr[L_{\mathcal{E}}]^2 +\tr[L_{\mathcal{E}}^2] \bigr).
\end{align}
Moreover, from the direct calculations, we obtain
\begin{align}
&\tr[ \kettbraa{\sigma_0^{\otimes 2}}{\sigma_0^{\otimes 2}} L_{\mathcal{E}}^{\otimes 2}] = L_{00}^2,\\
&\tr[ \kettbraa{S_{0,n}}{S_{0,n}} L_{\mathcal{E}}^{\otimes 2}] = L_{00}L_{nn} + L_{0n}L_{n0},
\end{align}
where $L_{nm} = \braa{\sigma_n} L_{\mathcal{E}} \kett{\sigma_m}$.
Further using the relations
\begin{align}
&\tr[L_{\mathcal{E}}] = L_{00} + \tr[\tilde{L}_{\mathcal{E}}],\\
&\tr[L_{\mathcal{E}}^2] = L_{00}^2 + 2 \sum_{n=1}^3 L_{0n}L_{n0} + \tr[\tilde{L}_{\mathcal{E}}^2],
\end{align}
we obtain from Eq.\,\eqref{Eq:1qubitRBstart} that
\begin{align}
\tr[ \Pi_1 L_{\mathcal{E}}^{\otimes 2}]
&=
\frac{1}{2} \tr[\tilde{L}_{\mathcal{E}}^2] - u(\mathcal{E})+\frac{1}{2} \tr[\tilde{L}_{\mathcal{E}}]^2,\\
&=
\frac{3}{2} h(\mathcal{E}) - u(\mathcal{E})+\frac{9}{2} f(\mathcal{E})^2.
\end{align}
Altogether, we obtain
\begin{multline}
V(m, \mathcal{E}| \Delta, O_{\rm meas}) \\= A_0 u(\mathcal{E})^m +
A_1 \biggl(\frac{9}{10} f(\mathcal{E})^2 - \frac{1}{5}u(\mathcal{E}) + \frac{3}{10} h(\mathcal{E})\biggr)^m.
\end{multline}
\subsubsection{$2$-RB in a multi-qubit system}
For a multi-qubit system ($q \geq 2$), the traceless symmetric subspace is decomposed into four irreducible subspaces:
\begin{equation}
{\cal K}_{TS} = {\cal K}_0 \oplus {\cal K}_{\rm I} \oplus {\cal K}_{\rm II} \oplus {\cal K}_{\rm III},
\end{equation}
where ${\cal K}_0 = {\rm span} \{ \sum_{\vec{n} \neq \vec{0}} \kett{ \sigma_{\vec{n} \otimes \vec{n}}} \}$ and the others are given in Appendix~\ref{App:Irreps}. Each irrep is multiplicity-free.
We denote by $D_{\lambda}$ the dimension of each subspace, which are
\begin{align}
&D_0 = 1,\\
&D_{\rm I} = \frac{d^2(d-1)(d+3)}{4},\\
&D_{\rm II} = \frac{d^2(d + 1)(d-3)}{4},\\
&D_{\rm III} = d^2-1.
\end{align}
Since the decomposition is multiplicity-free, $V^{(2)}(m, \mathcal{E}| \Delta, O_{\rm meas})$ is a sum of four exponentially decaying functions. Furthermore, from the fact that ${\cal K}_0 ={\rm span}\{ \sum_{\vec{n} \neq \vec{0}} \kett{ \sigma_{\vec{n} \otimes \vec{n}}} \}$, we obtain that $C_0(\mathcal{E}) = u(\mathcal{E})$. Hence, we have
\begin{equation}
V^{(2)}(m, \mathcal{E}| \Delta, O_{\rm meas}) = A_0 u(\mathcal{E})^m
+
\sum_{\lambda={\rm I}, {\rm II}, {\rm III}} A_{\lambda} C_{\lambda}(\mathcal{E})^m,
\end{equation}
where
\begin{align}
&A_{\lambda} = \braa{\mathcal{E}(O_{\rm meas})^{\otimes 2}} \Pi_{\lambda} \kett{O_{\rm ini}^{\otimes 2}} \label{Eq:Ai1q},\\
&C_{\lambda} (\mathcal{E}) = \frac{\tr[ \Pi_{\lambda} L_{\mathcal{E}}^{\otimes 2} ]}{\tr[\Pi_{\lambda}]},
\end{align}
with $\Pi_{\lambda}$ being the projections onto the irrep ${\cal K}_{\lambda}$.
It is not clear whether each $C_{\lambda}(\mathcal{E})$ ($\lambda = {\rm I}, {\rm II}, {\rm III}$) has a clear physical meaning, such as $C_0(\mathcal{E}) = u(\mathcal{E})$ being the unitarity. However, a linear combination of them does. To see this, we use the relation that
\begin{multline}
\Pi_{\rm I} + \Pi_{\rm II} + \Pi_{\rm III} =\\
\Pi_{\rm sym} -\Pi_0
- \kettbraa{\sigma_{\vec{0}\otimes \vec{0}}}{\sigma_{\vec{0} \otimes \vec{0}}} - \sum_{\vec{n} \neq \vec{0}} \kettbraa{S_{\vec{0} \vec{n}}}{S_{\vec{0} \vec{n}}},
\end{multline}
where $S_{\vec{0} \vec{n}} := (\sigma_{\vec{0} \otimes \vec{n}} + \sigma_{\vec{n} \otimes\vec{0}})/\sqrt{2}$.
From this relation, we can show, by a calculation similar to the one-qubit case, that
\begin{multline}
\tr\bigl[ (\Pi_{\rm I} + \Pi_{\rm II} + \Pi_{\rm III} ) L_{\mathcal{E}}^{\otimes 2} \bigr]\\
=
\frac{(d^2 - 1)^2}{2} f(\mathcal{E})^2 - u(\mathcal{E}) + \frac{d^2-1}{2} h(\mathcal{E}).
\end{multline}
Since $\tr[\Pi_{\lambda} L_{\mathcal{E}}^{\otimes 2}] = \tr[\Pi_{\lambda}] C_{\lambda}(\mathcal{E}) = D_{\lambda} C_{\lambda}(\mathcal{E})$, we obtain
\begin{equation}
\sum_{\lambda={\rm I}, {\rm II}, {\rm III}} D_{\lambda} C_{\lambda}(\mathcal{E})
=
\frac{(d^2 - 1)^2}{2} f(\mathcal{E})^2 - u(\mathcal{E}) + \frac{d^2-1}{2} h(\mathcal{E}). \label{Eq:ervrcr}
\end{equation}
We finally note that Tab.\,\ref{Tab:coeff} in Subsec.\,\ref{SS:2RB} is obtained by constructing the orthonormal basis in each subspace ${\cal K}_{\rm I}$, ${\cal K}_{\rm II}$, and ${\cal K}_{\rm III}$ (see Appendix~\ref{App:Irreps}). We also assume that the noise $\mathcal{E}$ is weak, so that $\mathcal{E}(O_{\rm meas}) \approx O_{\rm meas}$. Based on this assumption, we have
\begin{equation}
A_{\lambda} \approx \braa{O_{\rm meas}^{\otimes 2}} \Pi_{\lambda} \kett{O_{\rm ini}^{\otimes 2}},
\end{equation}
enabling us to compute $A_{\lambda}$ for given initial and measurement operators.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{fig_result_1q_SRshift.pdf}
\caption{Numerical results for the single-qubit $2$-RB with various numbers of samplings for measurement and those for unitary sequences when the noise parameters are set to $p=q=0.02$. The dots correspond to the values of $V^{(2)}$ obtained from the given numbers of samplings for measurement and for unitary sequences, and the dashed lines are fitting results. Theoretical values are shown in the solid lines.}
\label{fig:one_qubit_decay}
\end{figure*}
\section{Experimental realization of $2$-RB} \label{S:NumExpDetail}
Based on the former sections, we explain in detail how we have experimentally implemented the $2$-RB and estimated the self-adjointness of the noise in the system.
In Subsec.\,\ref{SS:NumericalEvalDetail}, we provide the details of the numerical evaluation of the one- and two-qubit $2$-RB discussed in Sec.\,\ref{SS:NumericalEval}.
The details of experiment is given in Subsec.\,\ref{SS:ExperimentDetail}.
\subsection{Numerical analysis} \label{SS:NumericalEvalDetail}
\subsubsection{Single-qubit systems}
We explain the fitting procedure of the 2-RB on one-qubit systems in detail.
The noise we consider is given by the following CPTP map:
\begin{equation}
\mathcal{E}_1(\rho) = q e^{i\theta X} \rho e^{- i\theta X} + (1-q)((1-p) \rho + p X \rho X), \label{Eq:1qubitnoise}
\end{equation}
which is characterized by three parameters $p, q$, and $\theta$. We particularly choose $\theta$ as $p = \sin^2 \theta$ for the fidelity parameter $f(\mathcal{E}_1)$ to be independent of the coherence parameter $q$.
Using the Liouville representation, it is straightforward to compute the fidelity parameter, the unitarity, and the self-adjointness parameter of this noise. They are, respectively, given by
\begin{align}
&f(\mathcal{E}) = 1 - \frac{4}{3}p,\\
&u(\mathcal{E}) = 1 - \frac{8}{3}p(1-p)(1-q^2),\\
&h(\mathcal{E}) = 1- \frac{8}{3}p(1-p)(1+q^2).
\end{align}
As shown in Theorem~\ref{Thm:2RB}, $V^{(2)}$ for one qubit is
\begin{multline}
V^{(2)}(m, \mathcal{E}_1| \Delta, O_{\rm meas}) \\= A_0 u(\mathcal{E}_1)^m + A_1 \biggl( \frac{9}{10}f(\mathcal{E}_1)^2 - \frac{1}{5} u(\mathcal{E}_1) + \frac{3}{10} h(\mathcal{E}_1) \biggr)^m. \label{Eq:162}
\end{multline}
To obtain Fig.\,\ref{Fig:numerical_overview_1q} in Subsec.\,\ref{SS:NumericalEval}, we first estimate the fidelity parameter $f(\mathcal{E}_1)$ from the $1$-RB, i.e., by fitting $V^{(1)}(m, \mathcal{E}_1| \ket{0}\bra{0}, \ket{0}\bra{0})$ and then obtain $u(\mathcal{E}_1)$ and $h(\mathcal{E}_1)$ from the fitting results of $V^{(2)}(m, \mathcal{E}_1|Z, \ket{0}\bra{0})$ and the estimated value of $f$.
The fitting of $V^{(2)}$ is first performed based on Eq.\,\eqref{Eq:162} by regarding the coefficients $A_0, A_1$ and the two exponential decaying rates as free parameters.
Then, we subtract the first exponential curve of Eq.\,\eqref{Eq:162} from the data and carry out the fitting of the second exponential curve again where we consider $A_1$ and the base of the second exponential curve as free parameters. This procedure is redundant, but turns out to improve the accuracy of the fitting because, in most cases, the first exponential decaying rate is larger than the second one, and thus the second exponential curve is clearly visible in the region of long sequence length.
We also estimate how many measurements and unitary sequences suffice to obtain a good estimate of the noise parameters. In Fig.\,\ref{fig:one_qubit_decay}, we provide $V^{(2)}$ for various numbers of measurement and samplings of unitary sequences. We observe that $1000$ for both suffice to obtain a good estimate when $p = q = 0.02$, which corresponds to the gate fidelity $97.3\%$.
When the number of measurements is small, experimental values are positively biased from theoretical values with the infinite number of samplings (see the figures in the top line of Fig.\,\ref{fig:one_qubit_decay}).
This difference can be understood as follows. Let $\langle O_{\rm meas} \rangle_{O_{\rm ini}, \bm{i}, n}$ be the expectation value for a random sequence described by $\bm{i}$ with the finite number $n$ of measurements. We describe the expectation value and variance of this random variable averaged over all choices of the unitary sequence as $\mu$ and $\sigma^2/n$, respectively. Note that this mean value is independent of $n$, and this variance is inverse proportional to $n$ since $\langle O_{\rm meas} \rangle_{O_{\rm ini}, \bm{i}, n}$ is a linear combination of binomial distribution. $V^{(2)}$ is the expectation value of squared random variable $\langle O_{\rm meas} \rangle^2_{O_{\rm ini}, \bm{i}}$, and its expectation value over random sequences is derived as
\begin{align}
\mathbb{E}[\langle O_{\rm meas} \rangle^2_{O_{\rm ini}, \bm{i}}] &= \mathbb{E}[(\langle O_{\rm meas} \rangle_{O_{\rm ini}, \bm{i}} - \mu)^2] + \mu^2 \\
&= \sigma^2/n + \mu^2.
\end{align}
Therefore, an experimentally obtained value with a finite number of measurements is positively biased by $\sigma^2/n$. In practice, we can remove the effect of this bias by increasing the number of sampling $n$ and using the region satisfying $\mu^2 \ll \sigma^2/n$ for fitting.
\begin{figure}[tb!]
\centering
\includegraphics[width=.4\textwidth]{fig_result_1q_SPAM.pdf}
\caption{Numerical results for the single-qubit $2$-RB with SPAM errors for $p=q=0.02$. We have taken $1000$ measurements and $1000$ random unitary sequences. The dots represent the numerically obtained data, and the dashed lines are the fitting curves.}
\label{Fig:numerical_SPAM_1q}
\end{figure}
We finally check the robustness of the $2$-RB on single-qubit systems against SPAM errors. Although the $2$-RB is ideally SPAM-error free, the fitting may become harder with the existence of SPAM errors. Our analysis, however, reveals that this is unlikely the case.
Here, we model the state-preparation error $\eta_{\mathrm{prep}}$ as $\rho=\eta_{\mathrm{prep}}\ketbra{0}{0}+(1-\eta_{\mathrm{prep}})\ketbra{1}{1}$ and $\rho'=(1-\eta_{\mathrm{prep}})\ketbra{0}{0}+\eta_{\mathrm{prep}}\ketbra{1}{1}$, and measurement error $\eta_{\mathrm{meas}}$ as readout bit-flip error, i.e., the POVM is $\{\Pi_{x'}|x' \in \{0,1\} \}$ where $\Pi_{x'}=\sum_{x \in \{0,1\}}\eta_{\mathrm{meas}}(x'|x) \Pi_x$ and $\eta_{\mathrm{meas}}(x'|x)$ is conditional probability.
In this numerical experiments, we assume that $\eta=\eta_{\mathrm{prep}}=\eta_{\mathrm{meas}}(0|1)=\eta_{\mathrm{meas}}(1|0)$ and get 1000 samples and 1000 random sequences. We set the parameters $p = q= 0.02$.
The results are provided in Fig.\,\ref{Fig:numerical_SPAM_1q} and Tab.\,\ref{Tab:numerical_SPAM_1q}. The relative errors of estimates for $F, u$, and $H$ are within $5\%$ except for the estimate for $u$ when $\eta=0.3$, thus the 2-RB is likely to work well even in realistic situations with SPAM errors as expected from the analytical studies.
\begin{table}[t!]
\centering
\caption{The estimates of $F,u$,and $H$ from the numerics shown in Fig.\,\ref{Fig:numerical_SPAM_1q}. The parameter $\eta$ is for SPAM errors such as $\eta = \eta_{\mathrm{prep}}=\eta_{\mathrm{meas}}(0|1)=\eta_{\mathrm{meas}}(1|0)$. The theoretical values are shown at the bottom of the table.}
\begin{tabular}{c||r|r|r}
SPAM error $\eta$ & $F$ & $u$ & $H$ \\ \hline \hline
0.0 & $0.986(6)$ & $0.9979(5)$ & $0.92(5)$\\
0.001 & 0.986(5) & 0.9979(6) & 0.92(7)\\
0.003 & 0.986(5) & 0.9978(5) & 0.92(1) \\
0.01 & 0.987(0) & 0.9979(2) & 0.92(2) \\
0.03 & 0.986(5)& 0.9980(1)& 0.92(8)\\
0.1 & 0.986(5) & 0.9978(9)& 0.92(7) \\
0.3 & 0.986(6) & 0.9981(8)& 0.92(7) \\ \hline
& 0.9866 & 0.99793 & 0.9247
\end{tabular}
\label{Tab:numerical_SPAM_1q}
\end{table}
\subsubsection{Two-qubit systems}
\begin{figure*}[tbh!]
\centering
\includegraphics[width=\textwidth]{fig_result_2q_p001_rb_decay.pdf}
\caption{A step-by-step fitting process of $2$-RB for two-qubit systems are shown. The first line of the matrix are about the standard RB ($1$-RB). The second, third, and fourth lines show $2$-RB with two, three, and four exponentially decaying functions, respectively. The initial and measurement operators for $2$-RB are chosen according to Tab.\,\ref{Tab:coeff}. See the main text for detail.}
\label{fig:two_qubit_decay}
\end{figure*}
For two-qubit systems, we investigate the noise given by
\begin{align}
\mathcal{E}_2(\rho) &= q e^{i \theta (X\otimes X)} \rho e^{- i\theta (X \otimes X)} \nonumber \\
&+ (1-q)((1-p) \rho + p (X \otimes X) \rho (X\otimes X) ),
\end{align}
where we set $\theta$ to $p = \sin^2 \theta$.
From Theorem~\ref{Thm:2RB}, $V^{(2)}$ in this case shall be in the form of
\begin{equation}
V^{(2)}(m, \mathcal{E}_2| \Delta, O_{\rm meas}) = A_0 u(\mathcal{E}_2)^m + \sum_{\lambda={\rm I}, {\rm II}, {\rm III}} A_{\lambda} C_{\lambda}(\mathcal{E}_2)^m, \label{Eq:1623}
\end{equation}
and $C_{\lambda}(\mathcal{E}_2)$ satisfy
\begin{multline}
84 C_{\rm I}(\mathcal{E}) + 20C_{\rm II}(\mathcal{E}) + 15 C_{\rm III}(\mathcal{E}) \\
=
\frac{225}{2} f(\mathcal{E})^2 + \frac{15}{2} h(\mathcal{E}) - u(\mathcal{E}). \label{Eq:19}
\end{multline}
With this setting, theoretical values are derived as follows.
\begin{align}
&f(\mathcal{E}_2) = 1-\frac{16}{15}p,\\
&u(\mathcal{E}_2) = 1 - \frac{32}{15}p(1-p)(1-q^2),\\
&C_{\rm I}(\mathcal{E}_2) = 1 - \frac{4}{105}p \bigl(56-31p+14(1-p)q^2\bigr),\\
&C_{\rm II}(\mathcal{E}_2) = 1 - \frac{4}{15}p \bigl(8-5p-2(1-p)q^2 \bigr),\\
&C_{\rm III}(\mathcal{E}_2) = 1 - \frac{16}{15}p \bigl(2-p-(1-p)q^2 \bigr),
\end{align}
from which the theoretical value of the self-adjointness parameter $h(\mathcal{E})$ can be computed from Eq.\,\eqref{Eq:19}.
A step-by-step fitting process of two-qubit systems is shown in Fig.\,\ref{fig:two_qubit_decay} for $p=0.01$. The numbers of measurement and samplings of unitary sequences are both set to $10^4$.
The fidelity parameters $f(\mathcal{E}_2)$ can be obtained from $V^{(1)}(m, \mathcal{E}_1 | \ket{00}\bra{00}, \ket{00}\bra{00})$, which are plotted in the top line of the figure. In the figure, dashed lines are theoretical values, and the shaded area represents a standard deviation of each data. When $q$ is near to unity, standard deviations are large even when the sequence length increased. This is because the final quantum state is nearly pure state when $q \sim 1$, and probability distributions fluctuate randomly according to chosen random sequences. On the other hand, when $q \sim 0$, the final quantum state quickly converges to the maximally mixed state, and thus probability distribution becomes independent of the chosen random sequences.
We then fit four values $u, C_{\rm I}, C_{\rm II}, C_{\rm III}$ step by step where initial and measurement operators $\Delta$ and $O_{\rm meas}$ are chosen according to Tab.\,\ref{Tab:coeff}. The fitting results for several $q$ are shown in Tab.\,\ref{Tab:two_qubit_data}.
\begin{table}[t!]
\centering
\begin{tabular}{c||l|l|l|l}
$q$ & $u$ & $C_{\rm I}$ & $C_{\rm II}$ & $C_{\rm III}$ \\ \hline \hline
$0.00$ & $0.9792(7)$ & $0.9787(2)$ & $0.978869(7)$ & $0.97896(1)$ \\
& $0.97888$ & $0.97879$ & $0.9788000$ & $0.978773$ \\ \hline
$0.50$ & $0.98416(9)$ & $0.97744(3)$ & $0.98015(1)$ & $0.98150(2)$ \\
& $0.984160$ & $0.977465$ & $0.980120$ & $0.981413$ \\ \hline
$0.95$ & $0.997942(9)$ & $0.97401(3)$ & $0.98354(3)$ & $0.98830(5)$ \\
& $0.9979410$ & $0.974020$ & $0.983565$ & $0.988304$ \\ \hline
$1.00$ & $0.999994(6)$ & $0.97349(6)$ & $0.98414(5)$ & $0.98930(9)$ \\
& $1.0000000$ & $0.973505$ & $0.984080$ & $0.989333$
\end{tabular}
\caption{Fitting results for $p=0.01$ in the two-qubit 2-RB. In each cell, fitting results are written in the first line, and theoretical values are written in the second line.}
\label{Tab:two_qubit_data}
\end{table}
\begin{table}[t!]
\centering
\begin{tabular}{c||l|l|l}
$q$ & $F$ & $u$ & $H$ \\ \hline \hline
$0.00$ & $0.9920004(5)$ & $0.9792(7)$ & $0.999(2)$ \\
& $0.99200000$ & $0.97888$ & $1.0000$ \\ \hline
$0.50$ & $0.991998(3)$ & $0.98416(9)$ & $0.9902(3)$ \\
& $0.9920000$ & $0.984160$ & $0.99010$ \\ \hline
$0.95$ & $0.99203(1)$ & $0.997942(9)$ & $0.9630(5)$ \\
& $0.992000$ & $0.9979408$ & $0.96426$ \\ \hline
$1.00$ & $0.99198(1)$ & $0.999994(6)$ & $0.9610(8)$ \\
& $0.992000$ & $1.0000000$ & $0.96040$
\end{tabular}
\caption{Processed data for $p=0.01$ in the two-qubit 2-RB. In each cell, fitting results are written in the first line, and theoretical values are written in the second line.}
\label{Tab:two_qubit_data_process}
\end{table}
First, we obtain $u$ and $C_{\rm I}$ from $V^{(2)}_1 := V^{(2)}(m, \mathcal{E}_2 | ZZ, \ket{00}\bra{00})$. The obtained results are shown in the second line of Fig.\,\ref{fig:two_qubit_decay}. The sampled data points and fitting results are shown as blue points and dashed lines, respectively. These lines are linear combinations of two exponentially decaying function. Exponential decays with coefficient $u$ and $C_{\rm I}$ are shown as orange and green lines, respectively. While we can clearly see two exponential decays for coherent noise, i.e., in the case of $q \sim 1$, an exponential decay of $C_{\rm I}$ part becomes dominant when noise becomes probabilistic. Even in this case, the fitting results are still reliable as shown in Tab.\,\ref{Tab:two_qubit_data}.
Then, we estimate $C_{\rm II}$ from $V^{(2)}_2 := V^{(2)}(m, \mathcal{E}_2 | ZZ, ZZ)$ and $C_{\rm III}$ from $V^{(2)}_3 := V^{(2)}(m, \mathcal{E}_2 | \rho_-, \rho_-)$, where these initial and measurement operators are chosen from the third and fourth columns of Tab.\,\ref{Tab:coeff}. The obtained results are shown in the third and fourth lines of Fig.\,\ref{fig:two_qubit_decay}.
In each figure, numerical data is plotted as blue circles and fitting results are shown as dashed lines. In each fitting process, only a single exponentially decaying term is unknown in advance. We showed unfitted exponential decay as orange circles and fitting results as dashed lines. Although accuracy of orange data becomes not reliable when its value becomes much smaller than the others, we can fit $C_{\rm II}$ and $C_{\rm III}$ reliably.
We calculate the averaged fidelity $F(\mathcal{E})$, unitarity $u(\mathcal{E})$, and self-adjointness $H(\mathcal{E})$ from the fitting results. The processed values are shown in Tab.\,\ref{Tab:two_qubit_data_process}. We evaluate relative errors for all the plots with the same method as the single-qubit $2$-RB, and confirm that the relative errors are below $4\%$ for all the cases of $p=0.01, 0.02, 0.1, 0.2$ except the case when the theoretical value is exactly zero. While the relative errors of the self-adjointness become a few tens of percent in the case of $p=0.4$, we can say that reliable values can be obtained when the fidelities of operations are sufficiently high.
Note that, although standard deviations of the fitting results become large when $u$ is almost equal to $C_{\rm I}$, we confirmed that the fitted results are close to theoretical values even in such cases. Thus, we conclude that $2$-RB works also for two-qubit systems.
\subsection{Details of the experiments}\label{SS:ExperimentDetail}
In this section, we provide the details of the experiments in Sec.\,\ref{SS:Experiments}.
A superconducting qubit can be regarded as a sort of the LC resonant circuit, where a Josephson junction is an effective inductance, and has the Hamiltonian equivalent to that of the one-dimensional free particle trapped in anharmonic potential.
\begin{table}[t!]
\centering
\caption{The parameter fields of the qubits.}
\begin{tabular}{c||c|c|c|c|c}
& $\omega_q/2\pi$ & $\alpha/2\pi$ & $T_1$ & $T_2$ echo \\ \hline \hline
Q1 & $9.077~{\rm GHz}$ & $-328.9~{\rm MHz}$ & $9.724~{\rm \mu s}$ & $13.670~{\rm \mu s}$ \\
Q2 & $8.927~{\rm GHz}$ & $-419.9~{\rm MHz}$ & $12.634~{\rm \mu s}$ & $15.763~{\rm \mu s}$
\end{tabular}
\label{Tab:Experiment_parameter_field}
\end{table}
The parameter fields of the qubits are summarized in Tab.\,\ref{Tab:Experiment_parameter_field}.
All unitary gates required in the $2$-RB for a single-qubit case were implemented by two $R_X(\pi/2)$ gates and three $R_Z$ gates with an arbitrary rotation angle, which are implemented by the shaped microwave pulse (Half-DRAG) with the length $11.70~{\rm ns}$~\cite{lucero2010reduced} and the Virtual-$Z$ gates~\cite{mckay2017efficient}, respectively.
The pre-measured averaged gate fidelity of the single-qubit Clifford gate is $0.991$.
The single-shot qubit readout was performed via the impedance-matched Josephson parametric amplifier~\cite{doi:10.1063/1.4886408}, and the assignment fidelity of the readout was $0.943$.
\begin{figure}[tb!]
\centering
\includegraphics[width=.4\textwidth]{210212_fig_t1.pdf}
\caption{
T1 decay experiment.
The horizontal axis represents the delay time, the vertical axis represents the projected value of the IQ readout signal.
The blue and orange dots represents the experimental results of the T1 decay experiments with and without flip operation just before measurement, respectively.
The lines are fitting curves.
}
\label{Fig:experiment_t1}
\end{figure}
\begin{figure}[tb!]
\centering
\includegraphics[width=.4\textwidth]{210209_fig_ramsey_fft_2.pdf}
\caption{
Ramsey oscillation experiment.
The horizontal axis represents the detuning from the qubit eigenfrequency, and the vertical axis represents the power spectrum of the Ramsey oscillation.
The points connected by the blue line represent the experimental data, and the orange line represents the fitting curve.
}
\label{Fig:experiment_ramsey}
\end{figure}
As supplemental experiments, T1 decay and Ramsey oscillation were observed to clarify the background noise source of Q1.
The experimental results of the T1 decay is shown in Fig.\,\ref{Fig:experiment_t1}, where the horizontal and the vertical axes represent the delay time and the projected value of the IQ readout signal, respectively.
The blue and orange dots represent the experimental results with and without a flip operation just before measurement, respectively. The lines provide the fitting curves. As seen from the result, the T1 decay of Q1 follows exponential behavior, which is consistent with the expected behavior in isolated qubits.
The experimental results of the Ramsey oscillation is also given in Fig.\,\ref{Fig:experiment_ramsey}, where the horizontal and vertical axes represent the detuning from the qubit eigenfrequency and the power spectrum of the Ramsey oscillation, respectively.
The points connected by a blue line represent the experimental data, and the orange line provides a fitting curve.
In the fitting, we did not take the data in the small power spectrum region ($<1~\mathrm{V}^2$) into account. This is because the noise floor derived from the white noise is dominant there.
As seen from the result, the power spectrum of the Ramsey oscillation has no peaks other than the qubit eigenfrequency.
From the fitting curve, it was found that the detuning from the qubit eigenfrequency $\Delta$~($\mathrm{MHz}$) and the power spectrum of the Ramsey oscillation $PS(\Delta)$~($\mathrm{V^2}$) are related as
\begin{align}
PS(\Delta) \propto \Delta^{-2.004}.
\end{align}
This is consistent with the expected behavior when the transmon qubits are isolated well.
From the results of these supplemental experiments, we conclude that Q1 is not in the strong coupling regime with any noise source, which implies that the background noise of Q1 is time-independent. Thus, the requirements for $2$-RB are met in our experimental system.
\section{Summary and Discussions} \label{S:SD}
In this paper, we have provided an explicit constructions of \emph{exact} unitary $t$-designs for any $t$. In particular, quantum circuits for exact unitary $t$-designs on $N$ qubits have been provided for the first time. Our construction is inductive with respect to the number of qubits. Hence, all constructions obtained in this paper are inefficient when the number of qubits is large, implying that it is of practical use only when the size of the system is small.
As an application of exact unitary $2t$-designs on a small system, we have proposed the $t$-RB, which enables us to experimentally estimate higher-order properties of the noise on a quantum system.
Since the unitary designs are used in multiple times in a single run of the protocol, it is important for the design to be exact.
After providing a general scheme of the $t$-RB, we have studied the $2$-RB in detail.
It was shown that the $2$-RB reveals the self-adjointness of the noise, a new characterization of the noise that we argue to play an important role in QEC especially when decoders are based on applications of Pauli operators. Our results have been demonstrated numerically, which shows that the $2$-RB is experimentally tractable. We have then experimentally implemented the single-qubit $2$-RB on the superconducting qubit system. From the experimental results, we found that the characteristics of the background noise of a qubit changes depending on the presence of the interaction with the adjacent qubits.
Our results open a number of future problems. Regarding the implementation of $t$-designs, it is important to improve the efficiency. Despite that the inefficiency in our construction is likely to be intrinsic due to an inductive nature of the construction, the representation-theoretic method provides a way to searching more efficient ones. More specifically, the key in the construction is the relation between the representation of the whole unitary group and that of a certain subgroup of the unitary group. This indicates that finding the construction of exact unitary designs may be reduced to the problem of searching for a subgroup whose representation has a good relation to that of the whole unitary group.
It is also important to further develop the theory of the $t$-RB protocol. In this paper, we have analyzed only the $2$-RB in detail. It then turns out that self-adjointness of the noise can be revealed. It is of great interest to concretely investigate what characterization of the noise can be generally obtained from the $t$-RB. In the context of QEC, it is also important to comprehensively analyze quantitative relations between the self-adjointness and the feasibility of QEC.
Another promising future problem is to use exact higher-designs in the other RB-type protocols.
The $t$-RB is a straightforward generalization of the standard RB. However, there are numerous variant protocols~\cite{HHFFW2019}, most of which, if not all, are based on the Clifford group that is an exact unitary $2$-design. By extending such protocols to those with higher-designs, the noise on the system can be characterized in more detail.
\section{Acknowledgements}
The authors are grateful to H. Yamasaki and A. Darmawan for helpful discussions.
Y. Nakata is supported by JST, PRESTO (No. JPMJPR1865).
Y. Suzuki is supported by JST, PRESTO (No.\,JPMJPR1916).
Y. Suzuki, K. Heya, Z. Yan, K. Zuo, S. Tamate., Y. Tabuchi, and Y. Nakamura are supported by JST ERATO (No. JPMJER1601) and by MEXT Q-LEAP (No. JPMXS0118068682).
\bibliographystyle{unsrt}
|
\section{Introduction}
\label{introduction}
\IEEEPARstart{I}{n} the near future of robotic exploration, teams of robots are expected to perform complex decision-making tasks autonomously in extreme environments, where their motions are typically governed by nonlinear dynamics with external disturbances. For such operations to be successful, they need to compute optimal motion plans online while robustly guaranteeing convergence to the target trajectory, both with their limited onboard computational resources. Thus, this work aims to propose a learning-based robust motion planning and control algorithm that meets these challenging requirements.
\subsubsection*{Related Work}
Learning-based control designs have been an emerging area of research since the rise of neural networks and reinforcement learning~\cite{sutton,ndp}. Model-free approaches learn optimal policies using data obtained in real-world environments, making them robust but not suitable for situations where sampling large training datasets is difficult. Also, proving the robustness and stability properties of such data-driven systems is challenging in general, although some techniques do exist~\cite{boffi2020learning,9029986}. In contrast, model-based methods allow sampling as much data as we want to design the policies by, \eg{}, imitation learning~\cite{glas,9001182}, reinforcement learning~\cite{8593871,NIPS2017_766ebcd5}, or both~\cite{NIPS2016_cc7e2b87,8578338,7995721}. However, the learned controller could yield cascading errors in the real-world environment if its nominal model poorly represents the true underlying dynamics~\cite{cascading}.
Control theoretical approaches to circumvent such difficulties include robust tube-based motion planning~\cite{7989693,ccm,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,sun2021uncertaintyaware,10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355} equipped with a tracking control law for robustness and stability certificates. Among these are contraction theory-based robust control~\cite{ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,sun2021uncertaintyaware}, which tracks a target trajectory computed externally by existing motion planners, and thereby robustly keeps the system trajectories in a control invariant tube that satisfies given state constraints. Although these provable guarantees are promising, they still assume that the target trajectory can be computed online solving some planning problems, unlike the aforementioned learning frameworks.
\subsubsection*{Contributions}
\begin{figure*}
\centering
\includegraphics[width=170mm]{lagros_rev2.png}
\caption{Illustration of LAG-ROS: See Theorems~\ref{Thm:lagros_stability} -- \ref{Prop:sampling} and Algorithm~\ref{lagros_alg} for detailed explanation on each block of the LAG-ROS design. Note that LAG-ROS requires only one neural net evaluation to get $u_L$, and Theorem~\ref{Thm:lagros_stability} provides its robustness and stability guarantees.\label{lagrosdrawing}}
\vspace{-0.5em}
\end{figure*}
In this paper, we present Learning-based Autonomous Guidance with RObustness and Stability guarantees (LAG-ROS) as a novel way to bridge the gap between the learning-based and robust tube-based motion planners. In particular, while LAG-ROS requires one neural network evaluation to get its control input as in the learning schemes~\cite{8593871,NIPS2017_766ebcd5,9001182,glas,NIPS2016_cc7e2b87,8578338,7995721}, its contraction theory-based architecture still allows obtaining formal robustness and stability guarantees as in~\cite{ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,sun2021uncertaintyaware}. This framework depicted in Fig.~\ref{lagrosdrawing} is summarized as follows.
\begin{table*}
\caption{Comparison of the Proposed Method with the Learning-based and Robust Tube-based Motion Planners.} \label{learning_summary}
\vspace{-1.5em}
\begin{center}
\begin{tabular}{|l|l|l|l|}
\hline
Motion planning scheme & Policy to be learned & State tracking error $\|x-x_d\|$ & Computational load \\
\hline
\hline
\ref{itemFF} Learning-based motion planner~\cite{8593871,NIPS2017_766ebcd5,9001182,glas,NIPS2016_cc7e2b87,8578338,7995721} & $(x,o_{\ell},t) \mapsto u_d$ & Increases exponentially (Lemma~\ref{Lemma:naive_learning}) & One neural net evaluation \\
\ref{itemMP} Robust tube-based motion planner~\cite{10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355,ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,sun2021uncertaintyaware} & $(x,x_d,u_d,t) \mapsto u^*$ & Exponentially bounded (Theorem~\ref{ncm_clf_thm}) & Computation required to get $(x_d,u_d)$ \\
\ref{itemLAGROS} Proposed method (LAG-ROS) & $(x,o_{\ell},t) \mapsto u^*$ & Exponentially bounded (Theorem~\ref{Thm:lagros_stability}) & One neural net evaluation \\
\hline
\end{tabular}
\end{center}
\vspace{-2.0em}
\end{table*}
\subsubsection{Robustness and Stability Guarantees}
The theoretical foundation of LAG-ROS rests on contraction theory, which utilizes a contraction metric to characterize a necessary and sufficient condition of exponential incremental stability of nonlinear system trajectories~\cite{contraction}. The central result of this paper is that, if there exists a control law which renders a nonlinear system contracting, or equivalently, the closed-loop system has a contraction metric, then LAG-ROS trained to model the controller ensures the Euclidean distance between the target and controlled trajectories to be bounded exponentially with time, linearly in the learning error and size of perturbation. This property helps quantify how small the learning error should be in practice, giving some guidance in choosing design parameters of neural net training. We further show that such a contracting control law and its corresponding contraction metric can be designed explicitly via convex optimization, using the method of CV-STEM~\cite{mypaperTAC,ncm,nscm} to minimize a steady-state upper bound of the LAG-ROS tracking error.
\subsubsection{State Constraint Satisfaction}
We further exploit the computed bound on the tracking error in generating expert demonstrations for training, so that the learned policy will not violate given state constraints even with the learning error and external disturbances. In this phase, LAG-ROS learns the contracting control law independently of a target trajectory, making it implementable without solving any motion planning problems online unlike~\cite{10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355,ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,sun2021uncertaintyaware}. The performance of LAG-ROS is evaluated in cart-pole balancing~\cite{6313077} and nonlinear motion planning of multiple robotic agents~\cite{SCsimulator} in a cluttered environment, demonstrating that LAG-ROS indeed satisfies the formal exponential bound as in~\cite{ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,sun2021uncertaintyaware} with its computational load as low as that of existing learning-based motion planners~\cite{8593871,NIPS2017_766ebcd5,9001182,glas,NIPS2016_cc7e2b87,8578338,7995721}. In particular, LAG-ROS requires less than $0.1$s for computation in all of these tasks and achieves higher control performances and task success rates (see Sec.~\ref{Sec:performance}), when compared with the existing motion planners in Table~\ref{learning_summary} which outlines the differences of these schemes from our proposed method.
\subsubsection*{Notation}
For $x \in \mathbb{R}^n$ and $A \in \mathbb{R}^{n \times m}$, we let $\|x\|$, $\delta x$, and $\|A\|$, denote the Euclidean norm, infinitesimal variation of $x$, and induced 2-norm, respectively.
For a square matrix $A$, we use the notation $A \succ 0$, $A \succeq 0$, $A \prec 0$, and $A \preceq 0$ for the positive definite, positive semi-definite, negative definite, negative semi-definite matrices, respectively, and $\sym(A) = (A+A^{\top})/2$. Also, $I \in \mathbb{R}^{n\times n}$ denotes the identity matrix.
\section{Learning-based Robust Motion Planning with Guaranteed Stability (LAG-ROS)}
\label{Sec:lagros}
In this paper, we consider the following nonlinear systems with a controller $u \in \mathbb{R}^{m}$ ($f$ \& $B$ are known but $d$ is unknown):
\begin{align}
\label{dynamics}
\dot{x}(t) &= f(x(t),t)+B(x(t),t)u+d(x(t),t)\\
\label{dynamicsd}
\dot{x}_d(o_g,t) &= f(x_d(o_g,t),t)+B(x_d(o_g,t),t)u_d(x_d(o_g,t),o_g,t)
\end{align}
where $t\in \mathbb{R}_{\geq0}$, $f:\mathbb{R}^n\times\mathbb{R}_{\geq0} \mapsto \mathbb{R}^{n}$, $B:\mathbb{R}^n\times\mathbb{R}_{\geq0}\mapsto\mathbb{R}^{n\times m}$, $x:\mathbb{R}_{\geq0} \mapsto \mathbb{R}^{n}$ is the state trajectory of the true system \eqref{dynamics} perturbed by the bounded disturbance $d:\mathbb{R}^{n}\times\mathbb{R}_{\geq0} \mapsto \mathbb{R}^{n}$ \st{} $\sup_{x,t}\|d(x,t)\|=\bar{d}$, $o_g \in \mathbb{R}^g$ is a vector containing global environment information such as initial and terminal states, states of obstacles and other agents, etc., and $x_d:\mathbb{R}^g\times\mathbb{R}_{\geq0}\mapsto\mathbb{R}^{n}$ and $u_d:\mathbb{R}^n\times\mathbb{R}^g\times\mathbb{R}_{\geq0}\mapsto\mathbb{R}^{n}$ are the target trajectories given by existing motion planning algorithms, \eg{}, \cite{10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355,ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,sun2021uncertaintyaware}.
\subsection{Problem Formulation of LAG-ROS}
We seek to find $u$ that is computable with one neural network evaluation and guarantees exponential boundedness of $\|x-x_d\|$ in \eqref{dynamics} and \eqref{dynamicsd}, robustly against the learning error and external disturbances. The objective is thus not to develop new learning-based planners that compute $(x_d,u_d)$, but to augment them with formal robustness and stability guarantees. To this end, let us review the following existing planning techniques:
\begin{enumerate}[label={\color{caltechgreen}{(\alph*)}}]
\item Learning-based motion planner~\cite{glas} or~\cite{8593871,NIPS2017_766ebcd5,9001182,NIPS2016_cc7e2b87,8578338,7995721}:\\
$(x,o_{\ell}(x,o_g),t) \mapsto u_d(x,o_g,t)$, modeled by a neural network, where $o_{\ell}:\mathbb{R}^n\times\mathbb{R}^g\mapsto\mathbb{R}^{\ell}$ with $\ell \leq g$ is local environment information extracted from $o_g \in \mathbb{R}^g$~\cite{glas}.\label{itemFF}
\item Robust tube-based motion planner~\cite{7989693} or~\cite{10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355,ccm,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,sun2021uncertaintyaware}:\\
$(x,x_d,u_d,t) \mapsto u^*(x,x_d,u_d,t)$, where $u^*$ is a contraction theory-based tracking controller, \eg{}, in Theorem~\ref{ncm_clf_thm}. \label{itemMP}
\end{enumerate}
The robust tube-based motion planner~\ref{itemMP} ensures that the perturbed trajectories $x$ of \eqref{dynamics} stay in an exponentially bounded error tube around the target trajectory $x_d$ of \eqref{dynamicsd}~\cite{ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,sun2021uncertaintyaware} (see Theorem~\ref{ncm_clf_thm}). However, it requires the online computation of $(x_d,u_d)$ as an input to their control policy, which is not realistic for systems with limited computational resources.
The learning-based motion planner~\ref{itemFF} circumvents this issue by modeling the target policy $(x,o_{\ell},t) \mapsto u_d$ by a neural network. In essence, our approach, to be proposed in Theorem~\ref{Thm:lagros_stability}, is for providing~\ref{itemFF} with the contraction theory-based stability guarantees~\ref{itemMP}. We remark that~\ref{itemFF} can only assure the tracking error $\|x-x_d\|$ to be bounded by a function which exponentially increases with time, as to be shown in Lemma~\ref{Lemma:naive_learning} for comparison with LAG-ROS of Theorem~\ref{Thm:lagros_stability}.
\subsection{Stability Guarantees of LAG-ROS}
The approach of LAG-ROS bridges the gap between \ref{itemFF} and \ref{itemMP} by ensuring that the distance between the target and controlled trajectories to be exponentially bounded.
\begin{enumerate}[label={\color{caltechgreen}{(\alph*)}},start=3]
\item Proposed approach (LAG-ROS, see Fig.~\ref{lagrosdrawing}):\\
$(x,o_{\ell}(x,o_g),t) \mapsto u^*(x,x_d(o_g,t),u_d(x_d(o_g,t),o_g,t),t)$ with $o_{\ell}$ of~\ref{itemFF}, modeled by a neural network $u_L$ of Theorem~\ref{Thm:lagros_stability}, where $u^*$ of \ref{itemMP} is viewed as a function of $(x,o_{\ell},t)$.\label{itemLAGROS}
\end{enumerate}
\begin{theorem}
\label{Thm:lagros_stability}
Suppose that \eqref{dynamics} is controlled to track \eqref{dynamicsd} by the LAG-ROS $u_L=u_L(x,o_{\ell}(x,o_g),t)$, learned to satisfy
\begin{align}
\label{learning_error}
\|u_L-u^*(x,x_d(o_g,t),u_d(x_d(o_g,t),o_g,t),t)\| \leq \epsilon_{\ell},~\forall x,o_g,t
\end{align}
where $\epsilon_{\ell} \in [0,\infty)$ is the learning error, $u^*$ is the robust control policy of \ref{itemMP} (to be designed in Theorem~\ref{ncm_clf_thm}), and $(x_d,u_d)$ is given by the robust motion planner~\ref{itemMP} (to be discussed in Theorem~\ref{tube_lemma}). Now consider the following virtual system of $y(\mu,t)$ parameterized by $\mu\in[0,1]$, which has $y(\mu=1,t)=x$ of \eqref{dynamics} and $y(\mu=0,t)=x_d$ of \eqref{dynamicsd} as its particular solutions:
\begin{align}
\label{virtual_lagros}
\dot{y} = \zeta(y,x,x_d,u_d,t)+d_y(\mu,x,x_d,u_d,o_g,t)
\end{align}
where $\zeta$ is parameterized by $y$ to verify $\zeta|_{y=x}=f(x,t)+B(x,t)u^*(x,x_d,u_d,t)$ and $\zeta|_{y-x_d}=f(x_d,t)+B(x_d,t)u_d$ (see, \eg{}, \eqref{def:zeta}), $d_y =\mu(B(x,t)(u_L(x,o_{\ell}(x,o_g),t)-u^*(x,x_d,u_d,t))+d(x,t))$, and $x$, $x_d$, $u_d$, and $d$ are as defined in \eqref{dynamics} and \eqref{dynamicsd}. Note that $y(\mu=1,t)=x$ and $y(\mu=0,t)=x_d$ are indeed particular solutions of \eqref{virtual_lagros}. If $\exists \bar{b}\in [0,\infty)$ \st{} $\|B(x,t)\| \leq \bar{b},~\forall x,t$, and if $u^*$ satisfies the following contraction conditions~\cite{Ref:contraction3} with respect to $y$, for a contraction metric $M(y,x,x_d,u_d,t)=\Theta^{\top}\Theta \succ 0$ and $\alpha,\underline{\omega},\overline{\omega} \in (0,\infty)$:
\begin{align}
\label{lagros_contraction}
&\dot{M}+2\sym{}\left(M\frac{\partial \zeta}{\partial y}\right) \preceq -2\alpha M,~\forall y,x,x_d,u_d,t \\
\label{lagros_Mcon}
&\overline{\omega}^{-1}I \preceq M \preceq \underline{\omega}^{-1}I,~\forall y,x,x_d,u_d,t
\end{align}
then we have the following for $\mathtt{e}(t)=x(t)-x_d(o_g,t)$, $\forall o_g$:
\begin{align}
\label{robust_learning_bound}
\|\mathtt{e}(t)\| \leq \mathcal{R}(0)\sqrt{\overline{\omega}}e^{-\alpha t}+\frac{\bar{b}\epsilon_{\ell}+\bar{d}}{\alpha}\sqrt{\frac{\overline{\omega}}{\underline{\omega}}}(1-e^{-\alpha t}) = r_{\ell}(t)
\end{align}
where $\mathcal{R}(t)=\int^{x}_{x_d}\|\Theta\delta y\|$ for $M=\Theta^{\top}\Theta$, $x_d=x_d(o_g,t)$, and $u_d=u_d(x_d(o_g,t),o_g,t)$.
\end{theorem}
\begin{IEEEproof}
Let $V=\int^{x}_{x_d}\delta y^{\top}M\delta y = \int^{x}_{x_d}\|\Theta\delta y\|^2$ for $(x_d,u_d)$ in \eqref{robust_learning_bound}. Since $\|\partial d_y/\partial \mu\| \leq \bar{b}\epsilon_{\ell}+\bar{d}=\bar{d}_{\epsilon_{\ell}},~\forall \mu,x,o_g,t$, for $d_y$ in \eqref{virtual_lagros} with such $(x_d,u_d)$, the contraction condition \eqref{lagros_contraction} gives
\begin{align}
\dot{V} \leq& \int^{x}_{x_d}\delta y^{\top}\left(\dot{M}+2\sym{}\left(M\frac{\partial \zeta}{\partial y}\right)\right)\delta y+2\bar{d}_{\epsilon_{\ell}}\int^{x}_{x_d}\|M \delta y\| \\
\leq& -2\alpha V+({2\bar{d}_{\epsilon_{\ell}}}/{\sqrt{\underline{\omega}}})\mathcal{R}(t).
\end{align}
Since $d(\|\Theta\delta y\|^2)/dt = 2\|\Theta\delta y\|(d\|\Theta\delta y\|/dt)$, this implies that $\dot{\mathcal{R}} \leq -\alpha\mathcal{R}+{\bar{d}_{\epsilon_{\ell}}}/{\sqrt{\underline{\omega}}}$. Therefore, applying the comparison lemma~\cite[pp.102-103, pp.350-353]{Khalil:1173048} (\ie{}, if $\dot{v}_1 \leq h(v_1,t)$ for $v_1(0) \leq v_2(0)$ and $\dot{v}_2 =h(v_2,t)$, then $v_1(t) \leq v_2(t)$), along with the relation $\mathcal{R}(t) \geq \|\mathtt{e}(t)\|/\sqrt{\overline{\omega}}$, results in \eqref{robust_learning_bound}
\end{IEEEproof}
Theorem~\ref{Thm:lagros_stability} implies that the bound \eqref{robust_learning_bound} decreases linearly in the learning error $\epsilon_{\ell}$ and disturbance $\bar{d}$, and \eqref{dynamics} controlled by LAG-ROS is exponentially stable when $\epsilon_{\ell}=0$ and $\bar{d} = 0$, showing a great improvement over \ref{itemFF} which only gives an exponentially diverging bound as to be derived in Lemma~\ref{Lemma:naive_learning}~\cite{8593871,NIPS2017_766ebcd5,9001182,glas,NIPS2016_cc7e2b87,8578338,7995721}. This property permits quantifying how small $\epsilon_{\ell}=0$ should be to meet the required guidance and control performance, giving some intuition on the neural network architecture (see Sec.~\ref{sec_simulation}). Also, since we model $u^*$ by $u_L(x,o_{\ell},t)$ independently of $x_d$, it is indeed implementable without solving any motion planning problems online unlike the robust motion planners \ref{itemMP}~\cite{10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355,ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,sun2021uncertaintyaware}, as outlined in Table~\ref{learning_summary}. If we can sample training data of $u^*$ explicitly considering the bound \eqref{robust_learning_bound}, the LAG-ROS control enables guaranteeing given state constraints even with the learning error $\epsilon_{\ell}$ and external disturbance $d(x,t)$, as will be seen in Sec.~\ref{Sec:contraction}~and~\ref{imitation_learning}.
\begin{remark}
$\epsilon_{\ell}$ of \eqref{learning_error} can only be found empirically in practice, and thus we propose one way to generate training data with \eqref{robust_learning_bound} and use the test error of $u_L$ as $\epsilon_{\ell}$~(see Sec.~\ref{sec_robust_sampling} and Sec.~\ref{sec_simulation}). Note that models $f(x,t)$ learned by system identification for a more accurate description of the nominal dynamics, \eg{}, \cite{8794351}, is still utilizable in \eqref{virtual_lagros} as long as the learned system is contracting~\cite{boffi2020learning}, and the modeling error is bounded. Other types of perturbations, such as stochastic or parametric uncertainty, could be handled using~\cite{nscm,ancm}.
\end{remark}
To appreciate the importance of the guarantees in Theorem~\ref{Thm:lagros_stability}, let us additionally show that~\ref{itemFF}, which models $(x,o_{\ell},t) \mapsto u_d$, only leads to an exponentially diverging bound.
\begin{lemma}
\label{Lemma:naive_learning}
Suppose that $u$ of \eqref{dynamics} is learned to satisfy
\begin{align}
\label{learning_error0}
\|u(x,o_{\ell}(x,o_g),t)-u_d(x,o_g,t)\| \leq \epsilon_{\ell},~\forall x,o_g,t
\end{align}
for $u_d$ of \eqref{dynamicsd} with the learning error $\epsilon_{\ell} \in [0,\infty)$, and that $\exists \bar{b}$ \st{} $\|B(x,t)\| \leq \bar{b},~\forall x,t$. If $f_{c\ell}=f(x,t)+B(x,t)u_d(x,o_g,t)$ is Lipschitz, \ie{}, $\exists L_{f} \in [0,\infty)$ \st{} $\|f_{c\ell}(x_1,o_g,t)-f_{c\ell}(x_2,o_g,t)\| \leq L_f\|x_1-x_2\|,~\forall x_1,x_2 \in \mathbb{R}^n$, then we have the following bound:
\begin{align}
\label{naive_learning_error}
\|\mathtt{e}(t)\| \leq \|\mathtt{e}(0)\|e^{L_{f} t}+L_{f}^{-1}(\bar{b}\epsilon_{\ell}+\bar{d})(e^{L_{f} t}-1).
\end{align}
where $\mathtt{e} = x-x_d$, and $x$, $x_d$, and $\bar{d}$ are given in \eqref{dynamics} and \eqref{dynamicsd}.
\end{lemma}
\begin{IEEEproof}
Integrating \eqref{dynamics} and \eqref{dynamicsd} for $u$ in \eqref{learning_error0} yields $\|\mathtt{e}(t)\| \leq \|\mathtt{e}(0)\|+L_f\int_{0}^t\|\mathtt{e}(\tau)\|d\tau+(\bar{b}\epsilon_{\ell}+\bar{d})t$. Applying the Gronwall-Bellman inequality~\cite[pp. 651]{Khalil:1173048} gives
\begin{align}
\|\mathtt{e}(t)\| \leq& \|\mathtt{e}(0)\|+\bar{d}_{\epsilon_{\ell}}t+L_{f}\int_{0}^t(\|\mathtt{e}(0)\|+\bar{d}_{\epsilon_{\ell}}\tau)e^{L_{f}(t-\tau)}d\tau
\end{align}
where $\bar{d}_{\epsilon_{\ell}} = \bar{b}\epsilon_{\ell}+\bar{d}$. Thus, integration by parts results in the desired relation \eqref{naive_learning_error}.
\end{IEEEproof}
Lemma~\ref{Lemma:naive_learning} indicates that if there exists either a learning error $\epsilon_{\ell}$ or external disturbance $d$, the tracking error bound grows exponentially with time, and thus \eqref{naive_learning_error} becomes no longer useful for large $t$. Section~\ref{sec_simulation} demonstrates how the computed bounds of \eqref{robust_learning_bound} ($\lim_{t\to \infty}e^{-\alpha t} = 0$) and \eqref{naive_learning_error} ($\lim_{t\to \infty}e^{L_f t} = \infty$) affect the control performance in practice.
\section{Contraction Theory-based Robust and Optimal Tracking Control}
\label{Sec:contraction}
Theorem~\ref{Thm:lagros_stability} is subject to the assumption that we have a contraction theory-based robust tracking control law $u^*$, which satisfies \eqref{lagros_contraction} and \eqref{lagros_Mcon} for a given $(x_d,u_d)$. This section thus delineates one way to extend the method called ConVex Optimization-based Steady-state Tracking Error Minimization (CV-STEM)~\cite{mypaperTAC,ncm,nscm} to find a contraction metric $M$ of Theorem~\ref{Thm:lagros_stability}, which minimizes an upper bound of the steady-state error of \eqref{robust_learning_bound} via convex optimization. Minimizing the bound renders the tube-based planning of Theorem~\ref{tube_lemma} to be given in Sec.~\ref{imitation_learning} for sampling training data less conservative, resulting in a better optimal solution for $(x_d,u_d)$.
In addition, we modify the CV-STEM in~\cite{mypaperTAC} to derive a robust control input $u^*$ which also greedily minimizes the deviation of $u^*$ from the target $u_d$, using the computed contraction metric $M$ to construct a differential Lyapunov function $V = \delta y^{\top}M\delta y$. Note that $u^*$ is to be modeled by a neural network which maps $(x,o_{\ell},t)$ to $u^*$ implicitly accounting for $(x_d,u_d)$ as described in Theorem~\ref{Thm:lagros_stability}, although $u^*$ takes $(x,x_d,u_d,t)$ as its inputs (see Sec.~\ref{sec_robust_sampling}).
\subsection{Problem Formulation of CV-STEM Tracking Control}
For given $(x_d,u_d)$, we assume that $u^*$ of Theorem~\ref{Thm:lagros_stability} can be decomposed as $u^*(x,x_d,u_d,t) = u_d+K(x,x_d,u_d,t)(x-x_d)$, the generality of which is guaranteed by the following lemma.
\begin{lemma}
\label{u_equivalence_lemma}
Consider a general tracking controller $u$ defined as $u = k(x,x_d,u_d,t)$ with $k(x_d,x_d,u_d,t)=u_d$, where $k:\mathbb{R}^n\times\mathbb{R}^n\times\mathbb{R}^m\times\mathbb{R}_{\geq0}\mapsto\mathbb{R}^{m}$. If $k$ is piecewise continuously differentiable, then $\exists K:\mathbb{R}^n\times\mathbb{R}^n\times\mathbb{R}^m\times\mathbb{R}_{\geq0}\mapsto\mathbb{R}^{m\times n}$ \st{} $u = k(x,x_d,u_d,t) = u_d+K(x,x_d,u_d,t)(x-x_d)$.
\end{lemma}
\begin{IEEEproof}
We have $u=u_d+(k(x,z_d,t)-k(x_d,z_d,t))$ for $z_d=(x_d,u_d)$ due to $k(x_d,z_d,t)=u_d$. Since $k(x,z_d,t)-k(x_d,z_d,t)=\int_0^1(dk(c x+(1-c)x_d,z_d,t)/dc)dc$, choosing $K$ as $\int_0^1(\partial k(c x+(1-c)x_d,z_d,t)/\partial x)dc$ gives the desired relation.
\end{IEEEproof}
Lemma~\ref{u_equivalence_lemma} implies that designing optimal $k$ of $u^* = k(x,x_d,u_d,t)$ reduces to designing optimal $K(x,x_d,u_d,t)$ of $u^*=u_d+K(x,x_d,u_d,t)(x-x_d)$. When \eqref{dynamics} is controlled by the LAG-ROS $u_L$ of Theorem~\ref{Thm:lagros_stability} with such $u^*$, the virtual system of \eqref{dynamics} and \eqref{dynamicsd} which has $y=x$ and $y=x_d$ as its particular solutions can be given by \eqref{virtual_lagros}, where $\zeta$ is defined as follows:
\begin{align}
\label{def:zeta}
\zeta =& \dot{x}_d+(A(x,x_d,u_d,t)+B(x,t)K(x,x_d,u_d,t))(y-x_d)
\end{align}
where $A$ is the State-Dependent Coefficient (SDC) form of the dynamical system \eqref{dynamics} given by Lemma~2 of~\cite{nscm}, which verifies $A(x,x_d,u_d,t)(x-x_d) = f(x,t)+B(x,t)u_d-f(x_d,t)-B(x_d,t)u_d$. Note that $\zeta$ indeed satisfies $\zeta|_{y=x}=f(x,t)+B(x,t)u^*$ and $\zeta|_{y=x_d}=f(x_d,t)+B(x_d,t)u_d$ for such $A$, to have $y=x$ and $y=x_d$ as the particular solutions to \eqref{virtual_lagros}.
\subsection{CV-STEM Contraction Metrics as Lyapunov Functions}
The remaining task is to construct $M$ so that it satisfies \eqref{lagros_contraction} and \eqref{lagros_Mcon}. The CV-STEM approach suggests that we can find such $M$ via convex optimization to minimize an upper bound of \eqref{robust_learning_bound} as $t\to \infty$ when $\alpha$ of \eqref{lagros_contraction} is fixed. Theorem~\ref{ncm_clf_thm} proposes using the metric $M$ designed by the CV-STEM for a Lyapunov function, thereby augmenting $u^*$ with additional optimality to greedily minimize $\|u^*-u_d\|^2$ for $u_d$ in \eqref{dynamicsd}.
\begin{theorem}
\label{ncm_clf_thm}
Suppose that $f$ and $B$ are piecewise continuously differentiable, and let $B=B(x,t)$ and $A=A(x,x_d,u_d,t)$ in \eqref{def:zeta} for notational simplicity. Consider a contraction metric $M(x,x_d,u_d,t) = W(x,x_d,u_d,t)^{-1} \succ 0$ given by the following convex optimization (CV-STEM)~\cite{ncm,mypaperTAC,nscm} to minimize an upper bound on the steady-state tracking error of \eqref{robust_learning_bound}:
\begin{align}
\label{convex_opt_ncm}
&{J}_{CV}^* = \min_{\nu>0,\chi \in \mathbb{R},\bar{W}\succ 0} \frac{\bar{b}\epsilon_{\ell}+\bar{d}}{\alpha}\chi \text{~~\st{} \eqref{deterministic_contraction_tilde} and \eqref{W_tilde}}
\end{align}
with the convex constraints \eqref{deterministic_contraction_tilde} and \eqref{W_tilde} given as
\begin{align}
\label{deterministic_contraction_tilde}
&-\dot{\bar{W}}+2\sym{\left(A\bar{W}\right)}-2\nu BR^{-1}B^{\top} \preceq -2\alpha \bar{W},~\forall x,x_d,u_d,t \\
\label{W_tilde}
&I \preceq \bar{W}(x,x_d,u_d,t) \preceq \chi I,~\forall x,x_d,u_d,t
\end{align}
where $\alpha,\underline{\omega},\overline{\omega} \in (0,\infty)$, $\nu = 1/\underline{\omega}$, $\chi = \overline{\omega}/\underline{\omega}$, $\bar{W} = \nu W$, and $R=R(x,x_d,u_d,t)\succ0$ is a weight matrix on the control input. Suppose also that $u^*$ of Theorem~\ref{Thm:lagros_stability} is designed as $u^* = u_d+K^*(x,x_d,u_d,t)\mathtt{e}$, where $\mathtt{e}=x-x_d$, and $K^*$ is found by the following convex optimization for given $(x,x_d,u_d,t)$:
\begin{align}
\label{ncm_robust_control}
&K^* = \text{arg}\min_{K\in \mathbb{R}^{m\times n}}\|u-u_d\|^2 = \text{arg}\min_{K\in \mathbb{R}^{m\times n}}\|K(x,x_d,u_d,t)\mathtt{e}\|^2 \\
\label{stability_clf}
&\text{\st{} } \dot{M}+2\sym{}(MA+MBK(x,x_d,u_d,t)) \preceq -2\alpha M.
\end{align}
Then $M$ satisfies \eqref{lagros_contraction} and \eqref{lagros_Mcon} for $\zeta$ defined in \eqref{def:zeta}, and thus \eqref{robust_learning_bound} holds, \ie{}, we have the exponential bound on the tracking error $\|x-x_d\|$ when the dynamics \eqref{dynamics} is controlled by the LAG-ROS control input $u_L$ of Theorem~\ref{Thm:lagros_stability}. Furthermore, the problem \eqref{ncm_robust_control} is always feasible.
\end{theorem}
\begin{IEEEproof}
Since the differential dynamics of \eqref{virtual_lagros} with \eqref{def:zeta} is given as $\delta \dot{y}=(\partial \zeta/\partial y)\delta y = (A-BK)\delta y$, substituting this into \eqref{lagros_contraction} verifies that \eqref{lagros_contraction} and \eqref{stability_clf} are equivalent. For $\bar{K}=-R^{-1}B^{\top}M$, \eqref{stability_clf} can be rewritten as
\begin{equation}
\nu^{-1}M(-\dot{\bar{W}}+2\sym{}(A\bar{W})-\nu BR^{-1}B^{\top})M \preceq -2\alpha \nu^{-1}M\bar{W}M.
\end{equation}
Since this is clearly feasible as long as $M$ satisfies the condition \eqref{deterministic_contraction_tilde}, this implies that the problem \eqref{ncm_robust_control} is always feasible. Also, multiplying \eqref{lagros_Mcon} by $W_s$ \st{} $W = W_s^2$ with $W_s \succ 0$ from both sides gives \eqref{W_tilde}~\cite{mypaperTAC}. These facts indicate that the conditions \eqref{lagros_contraction} and \eqref{lagros_Mcon} are satisfied for $M$ and $u^*$ constructed by \eqref{convex_opt_ncm} and \eqref{ncm_robust_control}, respectively, and thus we have the exponential bound \eqref{robust_learning_bound} as a result of Theorem~\ref{Thm:lagros_stability}. Furthermore, the problem \eqref{convex_opt_ncm} indeed minimizes an upper bound of \eqref{robust_learning_bound} as $t \to \infty$ due to the relation $0\leq\sqrt{\overline{\omega}/\underline{\omega}} = \sqrt{\chi} \leq \chi$. We remark that \eqref{convex_opt_ncm} is convex as the objective is affine in $\chi$ and \eqref{deterministic_contraction_tilde} and \eqref{W_tilde} are linear matrix inequalities in terms of $\nu$, $\chi$, and $\bar{W}$.
\end{IEEEproof}
\begin{remark}
\label{optimal_remark}
\eqref{convex_opt_ncm} and \eqref{ncm_robust_control} are convex in terms of their respective decision variables and thus can be solved computationally efficiently~\cite[pp. 561]{citeulike:163662}. For systems with a known Lyapunov function (\eg{} Lagrangian systems~\cite[pp. 392]{Slotine:1228283}), we could simply use it to get robust tracking control $u^*$ in Theorem~\ref{ncm_clf_thm} without solving \eqref{ncm_robust_control}, although optimality may no longer be guaranteed in this case.
\end{remark}
\begin{remark}
The contraction metric construction itself can be performed using a neural network~\cite{ncm,nscm,ancm,chuchu}, leading to an analogous incremental stability and robustness results to those of Theorem~\ref{Thm:lagros_stability}~\cite{ancm,cdc_ncm}.
\end{remark}
\if0
\begin{corollary}
Consider the following control law:
\begin{align}
\label{clqqp_solution}
u = \begin{cases} u_d-\frac{\varphi_f+\varphi_g^{\top}u_d}{\varphi_g^{\top}\varphi_g}\varphi_g & \text{if $e^T(\dot{M}+2MA+2\alpha M)e > 0$}\\ u_d & \text{if $e^T(\dot{M}+2MA+2\alpha M)e \leq 0$}\end{cases}
\end{align}
where $\varphi_f = e^T\dot{M}e+2e^TM(Ae-Bu_d)+2\alpha e^{\top}Me$, $\varphi_g = 2B^{\top}Me$, $B=B(x,t)$, $A=A(x,x_d,t)$. Then \eqref{clqqp_solution} is the solution to \eqref{ncm_robust_control} of Theorem~\ref{ncm_clf_thm}, and thus satisfies the bound \eqref{dist_dist}.
\end{corollary}
\begin{IEEEproof}
Since \eqref{ncm_robust_control} is convex, the KKT conditions~\cite[pp. 244]{citeulike:163662} provide necessary and sufficient conditions for optimality, and \eqref{clqqp_solution} follows from its application to \eqref{ncm_robust_control}.
\end{IEEEproof}
\fi
\if0
\subsection{Dynamical Systems with Known Lyapunov Functions}
\label{sec_systems_w_known_lyapunov}
There are some science and engineering applications where we know a Lyapunov function candidate of a given nonlinear system, one of which is a feedback linearizable system~\cite{Isidori:1995:NCS:545735,Slotine:1228283,Khalil:1173048}. In this section, we illustrate how to construct robust tracking control for such systems, taking Lagrangian systems as an example.
Consider the following system with a bounded disturbance $d(x,t)$:
\begin{align}
\label{sto_lag_system}
{H}(q)\ddot{q}+{C}(q,\dot{q})\dot{q}+{G}(q)=u+d(x,t)
\end{align}
where $q:\mathbb{R}_{\geq0} \to \mathbb{R}^{n}$, $u:\mathbb{R}_{\geq0} \to \mathbb{R}^{n}$, ${H}:\mathbb{R}^n \to \mathbb{R}^{n\times n}$, ${C}:\mathbb{R}^n \times \mathbb{R}^n \to \mathbb{R}^{n\times n}$, ${G}:\mathbb{R}^n \to \mathbb{R}^{n}$, and $d:\mathbb{R}^n\times\mathbb{R}_{\geq 0} \to \mathbb{R}^{n}$ with $\bar{d}=\sup_{x,t}\|d(x,t)\| < +\infty$ and $x = [q^{\top},\dot{q}^{\top}]^{\top}$. We note that the matrix ${C}(q,\dot{q})$ is selected to make $\dot{{H}}-2{C}$ skew-symmetric, so we have a useful property \st{} $z^{\top}(\dot{{H}}-2{C})z=0,~\forall z\in\mathbb{R}^n$.
\begin{theorem}
\label{lag_clf_thm}
Suppose that \eqref{sto_lag_system} is controlled by
\begin{align}
\label{lag_con}
u &= u_n = {H}(q)\ddot{q}_r+{C}(q,\dot{q})\dot{q}_r+{G}(q)-{K_{\ell}}(t)s
\end{align}
where $\dot{q}_r = \dot{q}_d-\Lambda (q-q_d)$, $s=\dot{q}-\dot{q}_r$, ${K_{\ell}} \succ 0$, and $\Lambda\succ 0$. If ${K_{\ell}}$, $\Lambda$, $\varepsilon>0$ are selected so that $\exists\alpha>0$ \st{} $\begin{bmatrix}2{K_{\ell}} & -\varepsilon I \\ -\varepsilon I & 2 \Lambda\end{bmatrix} \preceq -2\alpha M$, where $M=\Theta^{\top}\Theta=\diag{}({H},\varepsilon I)$, then we have that
\begin{align}
\label{dist_dist_lag}
\|x-x_d\| \leq \frac{1}{\sqrt{\underline{p}}}\left(\mathcal{R}(0)e^{-\alpha t}+\frac{\bar{d}}{\alpha}\frac{1}{\sqrt{\underline{m}}}\right)=R_{\rm NCM}(t)
\end{align}
where $x_d=[q_d^{\top},\dot{q}_d^{\top}]^{\top}$, $\mathcal{R}(t)=\int^{\xi}_{0}\|\Theta(q(t))\delta y(t)\|$, $\xi=[s^{\top},e^{\top}]^{\top}$, $e=x-x_d$, $\underline{m}I \preceq M$, $P = \begin{bmatrix}\Lambda{H}\Lambda+\varepsilon I &\Lambda {H} \\ {H}\Lambda & {H}\end{bmatrix} \succ 0$, $\underline{p}I \preceq P$, and $y$ is the state of the virtual system of \eqref{sto_lag_system}, which has $[s^{\top},e^{\top}]^{\top}$ and $[0^{\top},0^{\top}]^{\top}$ as its particular solutions.
\end{theorem}
\begin{IEEEproof}
For the virtual system defined as ${H}\dot{y}_1+{C}y_1+{K_{\ell}}y_1=d_s(y,t)$ and $\dot{y}_2=y_1-\Lambda y_2$, we have $y = [y_1^{\top},y_2^{\top}]^{\top} = \xi = [s^{\top},e^{\top}]^{\top}$ and $y = 0$ as its particular solutions when $d_s(\xi,t) = d(x,t)$ and $d_s(0,t) = 0$. Thus using a differential Lyapunov function $V = \delta y^{\top}M\delta y$, we can show that $\dot{\mathcal{R}} \leq -\alpha \mathcal{R}+\bar{d}/\sqrt{\underline{m}}$ under the condition of this theorem~\cite{modular_stability,Slotine:1228283}. Since we have $\xi^{\top}M\xi=x^{\top}Px$ by definition, applying the comparison lemma~\cite[pp.102-103, pp.350-353]{Khalil:1173048} (\ie{}, if $\dot{v}_1 \leq h(v_1,t)$ for $v_1(0) \leq v_2(0)$ and $\dot{v}_2 =h(v_2,t)$, then $v_1(t) \leq v_2(t)$) yields the desired relation.
\end{IEEEproof}
\begin{remark}
The control parameters ${K_{\ell}}$ and $\Lambda$ in Theorem~\ref{lag_clf_thm} could also be optimized using the same technique in Theorem~\ref{ncm_clf_thm} or the one in Remark~\ref{optimal_remark}~\cite{mypaperTAC}.
\end{remark}
\fi
\section{Generating Expert Demonstrations}
\label{imitation_learning}
This section presents how to sample training data, \ie{}, the target trajectory $(x_d,u_d)$ of \eqref{dynamicsd} and corresponding perturbed state and CV-STEM robust control $(x,u^*)$, so that $x$ of \eqref{dynamics} controlled by the LAG-ROS control $u_L$ will stay in given admissible state space as long as we have $\|u_L-u^*\| \leq \epsilon_{\ell}$ as in \eqref{learning_error} of Theorem~\ref{Thm:lagros_stability}. To be specific, since $u^*$ of Theorem~\ref{ncm_clf_thm} solves \eqref{convex_opt_ncm} to obtain an optimal error tube around $x_d$, \ie{}, \eqref{robust_learning_bound} of Theorem~\ref{Thm:lagros_stability}, we exploit it in the tube-based motion planning~\cite{7989693} for generating training data, which satisfies given state constraints even under the existence of the learning error $\epsilon_{\ell}$ and disturbance $d$ in \eqref{robust_learning_bound} and \eqref{convex_opt_ncm}. Note that the learning error $\epsilon_{\ell}$ and disturbance upper bound $\bar{d}$ of Theorem~\ref{Thm:lagros_stability} are assumed to be selected \textit{a priori}~\cite{8794351}.
\subsection{Tube-based State Constraint Satisfaction in LAG-ROS}
Given the learning error $\epsilon_{\ell}$ and disturbance upper bound $\bar{d}$ of Theorem~\ref{Thm:lagros_stability}, we sample $(x_d,u_d)$ of \eqref{dynamicsd} by solving the following tube-based global motion planning problem, for each element in a training set containing randomized $o_g$ of \eqref{dynamics}:
\begin{align}
\label{tube_motion_plan}
&\min_{\substack{\bar{x}=\bar{x}(o_g,t)\\\bar{u}=\bar{u}(\bar{x},o_g,t)}}~\int_{0}^{T}c_1\|\bar{u}\|^2+c_2P(\bar{x},\bar{u},t)dt \\
&\text{\st{} $\dot{\bar{x}} = f(\bar{x},t)+B(\bar{x},t)\bar{u}$, $\bar{x}\in\bar{\mathcal{X}}(o_g,t)$, and $\bar{u} \in\bar{\mathcal{U}}(o_g,t)$}
\end{align}
where $c_1\geq0$, $c_2\geq0$, $P(\bar{x},\bar{u},t)$ is some performance-based cost function, $T>0$ is a given time horizon, $\bar{\mathcal{X}}$ is robust admissible state space defined as $\bar{\mathcal{X}}(o_g,t)=\{v(t)\in\mathbb{R}^n|\forall \xi(t) \in \{\xi(t)|\|v(t)-\xi(t)\|\leq r_{\ell}(t)\},~\xi(t)\in\mathcal{X}(o_g,t)\}$, $\mathcal{X}(o_g,t)$ is given admissible state space, $r_{\ell}(t)$ is the right-hand side of \eqref{robust_learning_bound} in Theorem~\ref{Thm:lagros_stability}, and $\bar{u} \in\bar{\mathcal{U}}(o_g,t)$ is an input constraint. The following theorem shows that the LAG-ROS control ensures the perturbed state $x$ of \eqref{dynamics} to satisfy state constraints $x \in \mathcal{X}$, due to the contracting property of $u^*$ in Theorem~\ref{ncm_clf_thm}.
\begin{theorem}
\label{tube_lemma}
If the solution of \eqref{tube_motion_plan} yields $(x_d,u_d)$ of \eqref{dynamicsd}, the LAG-ROS control $u_L$ of Theorem~\ref{Thm:lagros_stability}, where $M$ and $u^*$ are designed by Theorem~\ref{ncm_clf_thm}, ensures the perturbed solution $x(t)$ of \eqref{dynamics} to stay in the admissible state space $\mathcal{X}(o_g,t)$, \ie{}, $x(t) \in \mathcal{X}(o_g,t),\forall t$, even with the learning error $\epsilon_{\ell}$ of $u_L$ in \eqref{learning_error}.
\end{theorem}
\begin{IEEEproof}
Since $x_d \in \bar{\mathcal{X}}$, we have $\xi \in \mathcal{X}$ for all $\xi \in S(x_d) = \{\xi|\|x_d-\xi\|\leq r_{\ell}\}$ by the definition of $\bar{\mathcal{X}}$. Also, if we control \eqref{dynamics} by $u_L$, we have $\|x-x_d\| \leq r_{\ell}$, or equivalently, $x\in S(x_d)$. These two statements indicate that $x \in \mathcal{X}$.
\end{IEEEproof}
\begin{remark}
Theorem~\ref{tube_lemma} implies that if $x_d$ is sampled by \eqref{tube_motion_plan}, the perturbed trajectory \eqref{dynamics} controlled by LAG-ROS $u_L$ of Theorem~\ref{Thm:lagros_stability} will not violate the given state constraints as long as LAG-ROS $u_L$ of Theorem~\ref{Thm:lagros_stability} satisfies \eqref{learning_error}. This helps greatly reduce the need for safety control schemes such as~\cite{8405547}.
\end{remark}
\subsection{Learning Contraction Theory-based Robust Control}
\label{sec_robust_sampling}
To meet the learning error requirement \eqref{learning_error} for the sake of robustness and state constraint satisfaction proposed in Theorems~\ref{Thm:lagros_stability} and~\ref{tube_lemma}, we should also sample $x$ to get training data for the CV-STEM robust control inputs $u^*$ of Theorem~\ref{ncm_clf_thm}.
\begin{figure}
\centering
\includegraphics[width=80mm]{sample_rev.pdf}
\caption{Illustration of state sampling in robust bounded error tube.}
\label{samplingfig}
\vspace{-1.0em}
\end{figure}
\begin{theorem}
\label{Prop:sampling}
If \eqref{learning_error} of Theorem~\ref{Thm:lagros_stability} is satisfied $\forall x(t) \in S(x_d) = \{\xi\in\mathbb{R}^n|\|x_d(o_g,t)-\xi\|\leq r_{\ell}(t)\}$ and $\forall o_g,t$ instead of $\forall x,o_g,t$, then $u_L$ with such a constraint still guarantees \eqref{robust_learning_bound}. Also, if we always choose $x(0) = x_d(o_g,0)$ for the perturbed state $x$ in \eqref{dynamics}, the radius of the tube $S(x_d)$ is bounded $\forall o_g,t$.
\end{theorem}
\begin{IEEEproof}
Following the proof of Theorem~\ref{Thm:lagros_stability}, we still get \eqref{learning_error} for all $x(t) \in S(x_d(o_g,t))$, and \eqref{learning_error} implies that the perturbed state $x$ of \eqref{dynamics} indeed lies in $S(x_d(o_g,t))$. When $x(0) = x_d(o_g,0)$, \ie{}, $\mathcal{R}(0)=0$ in \eqref{robust_learning_bound}, $r_{\ell}(t)$ is independent of the state $x$ and thus the radius of the tube $S(x_d)$ is bounded.
\end{IEEEproof}
Using Theorem~\ref{Prop:sampling}, we can sample $x$ in the bounded error tube $S(x_d(o_g,t))$ with $x(0) = x_d(o_g,0)$ to obtain training data for the CV-STEM robust control $u^*$ by \eqref{ncm_robust_control} of Theorem~\ref{ncm_clf_thm}. These samples are to be used for training the LAG-ROS neural network, so we can satisfy the condition \eqref{learning_error} of Theorem~\ref{Thm:lagros_stability} for the pre-defined learning error $\epsilon_{\ell}$.
\begin{remark}
The localization method in~\cite{glas} allows extracting $o_{\ell}$ of~\ref{itemFF} by $o_g$ of \eqref{tube_motion_plan}, to render LAG-ROS applicable to different environments in a distributed way (see Sec.~\ref{sec_simulation}).
\end{remark}
The pseudocode for the offline construction of the LAG-ROS control of Theorem~\ref{Thm:lagros_stability} is presented in Algorithm~\ref{lagros_alg} (see Fig.~\ref{lagrosdrawing} for its visual description). Once we get $u_L(x,o_{\ell},t)$ by Algorithm~\ref{lagros_alg}, $u=u_L$ in \eqref{dynamics} can be easily computed by only evaluating the neural net $u_L$ for a given $(x,o_{\ell},t)$ observed at $(x,t)$, whilst ensuring robustness, stability, and state constraint satisfaction due to Theorems~\ref{Thm:lagros_stability}, \ref{ncm_clf_thm}, and \ref{tube_lemma}.
\begin{algorithm}
\SetKwInOut{Input}{Inputs}\SetKwInOut{Output}{Outputs}
\BlankLine
\Input{Random environment information $\{(o_{g})_i\}^{N}_{i=1}$ \\
Contraction metric $M$ of Theorem~\ref{ncm_clf_thm} \\
Motion planner $\mathcal{P}$ \\ Learning error $\epsilon_{\ell}$ and disturbance bound $\bar{d}$}
\Output{LAG-ROS control $u_L$ of Theorem~\ref{Thm:lagros_stability}}
\For{$i\leftarrow 1$ \KwTo $N$}{
Solve \eqref{tube_motion_plan} of Theorem~\ref{tube_lemma} for $(o_{g})_i$ using $\mathcal{P}$ and obtain a target trajectory and environment observation history $(x_d,u_d,o_g,o_{\ell},t)_i$ \\
Sample $D$ robust CV-STEM control $\{(x,u^*)_{ij}\}_{j=1}^D$ using Theorems~\ref{ncm_clf_thm} and~\ref{Prop:sampling} (see Fig.~\ref{samplingfig})
}
Model $(x,o_{\ell},t)_{ij} \mapsto u^*_{ij}$ by a neural net to satisfy $\|u_L-u^*\| \leq \epsilon_{\ell}$ as in \eqref{learning_error} of Theorem~\ref{Thm:lagros_stability} (see Fig.~\ref{lagrosdrawing})
\caption{LAG-ROS Algorithm}
\label{lagros_alg}
\end{algorithm}
\if0
\begin{remark}
Since \eqref{ncm_robust_control} is modeled by a neural network, using local observations in training also enables its decentralized implementation accounting for centralized global solutions~\cite{glas,NIPS2017_f22e4747}. This property will be demonstrated in Sec.~\ref{sec_simulation} to enable distributed implementation of $u_L$ in Theorem~\ref{Thm:lagros_stability}.
\end{remark}
\fi
\section{Simulation}
\label{sec_simulation}
Our proposed LAG-ROS framework is demonstrated using multiple motion planning problems under external disturbances ({\color{caltechgreen}\underline{\href{https://github.com/astrohiro/lagros}{https://github.com/astrohiro/lagros}}}). CVXPY~\cite{cvxpy} with the MOSEK solver~\cite{mosek} is used to solve optimization problems in Theorems~\ref{ncm_clf_thm}~and~\ref{tube_lemma} for sampling training data.
\subsection{Simulation Setup}
The maximum admissible control time interval is selected to be $\Delta t_{\rm max} = 0.1$(s). The computational time of each framework is measured for the Macbook Pro laptop (2.2 GHz Intel Core i7, 16 GB 1600 MHz DDR3 RAM), and each simulation result is the average of $50$ simulations for each random environment and disturbance realization.
\subsubsection{Neural Network Training}
\label{Sec:training}
We use a neural network $u_L$ with $3$ layers and $100$ neurons. The network is trained using stochastic gradient descent with training data sampled by Theorems~\ref{ncm_clf_thm}--\ref{Prop:sampling}, and the loss function is defined as $\|u_L-u^*\|$ to satisfy the learning error bound $\epsilon_{\ell}$ of Theorem~\ref{Thm:lagros_stability}. We use $50000$ training samples and select $\epsilon_{\ell} = 0.01$ for all the tasks, but these numbers can be modified accordingly depending on situations, considering the required performance for the error bound \eqref{learning_error} in Theorem~\ref{Thm:lagros_stability}. Note that we can guarantee \eqref{learning_error} only empirically, using the test error computed with a given test set as in standard neural network training.
\subsubsection{Environment Information} $o_g$ of \eqref{dynamicsd} is selected as initial states, target terminal states, and states of obstacles and other agents, if any. In Sec.~\ref{Sec:sc_planning}, the deep set framework~\cite{NIPS2017_f22e4747} is used to extract local information $o_{\ell}$ of \ref{itemFF} from $o_g$ as in~\cite{glas}.
\subsubsection{Performance Measure}
\label{Sec:performance}
The objective function of Theorem~\ref{tube_lemma} for sampling $(x_d,u_d)$ of \eqref{dynamicsd} is selected as $\int_0^T\|\bar{u}\|^2dt$. Since LAG-ROS is not for proposing a new trajectory optimization solver but for augmenting it with the formal guarantees of Theorem~\ref{Thm:lagros_stability}, one could use any other objective, \eg{}, information-based cost~\cite{doi:10.2514/6.2021-1103}, as long as $(x_d,u_d)$ is obtainable.
We define success of each task as the situation where the agent reaches, avoiding collisions, if any, a given target terminal state $x_f$ within a given time horizon $\mathcal{T} \geq T$, \ie{} $\exists t^* \in [0,\mathcal{T}]$ \st{} $\|x(t^*)-x_f\| \leq r_{\ell}(t^*)$ for $r_{\ell}$ in \eqref{robust_learning_bound}, where the value for $r_{\ell}$ is to be defined in the subsequent sections. The success rate is computed as the percentage of successful trials in the total $50$ simulations. Also, we evaluate the performance of each planner, \ref{itemFF}, \ref{itemMP}, and \ref{itemLAGROS} in Sec.~\ref{Sec:lagros}, by the objective function $\int_0^{t^*}\|u\|^2dt$ if the task is successfully completed, and $\int_0^{T}\|u\|^2dt$ otherwise, where $T$ is the nominal time horizon.
\subsubsection{External Disturbances}
As shown in Lemma~\ref{Lemma:naive_learning} and Theorem~\ref{Thm:lagros_stability}, the tracking error bound of learning-based motion planners~\ref{itemFF} increases exponentially with time, whilst it decreases exponentially for the proposed approach~\ref{itemLAGROS}. To make such a difference clear, we consider the situations where $d(x,t)$ of \eqref{dynamics} is non-negligible, and thus~\ref{itemFF} tends to fail the task of Sec~\ref{Sec:performance} due to Lemma~\ref{Lemma:naive_learning}. This is not to argue that \ref{itemFF}~\cite{8593871,NIPS2017_766ebcd5,9001182,glas,NIPS2016_cc7e2b87,8578338,7995721} should be replaced by~\ref{itemLAGROS}, but to imply that they can be improved further to have the guarantees of Theorem~\ref{Thm:lagros_stability}.
\subsubsection{Computational Complexity}
\label{Sec:computation}
Since~\ref{itemFF} and LAG-ROS~\ref{itemLAGROS} are implementable with one neural net evaluation at each $t$, their performance is also compared with~\ref{itemMP} which requires solving motion planning problems to get its control input. Its time horizon is selected to make the trajectory optimization~\cite{47710} solvable online considering the current computational power, for the sake of a fair comparison. We denote the computational time as $\Delta t$ in this section, and it should be less than the maximum control time interval $\Delta t_{\rm max}$, \ie{}, $\Delta t \leq \Delta t_{\rm max} = 0.1$(s).
\subsection{Cart-Pole Balancing}
\label{Sec:cartpole}
We first consider the cart-pole balancing task in Fig.~\ref{lagros_cp_dist_fig} to demonstrate the differences of \ref{itemFF} -- \ref{itemLAGROS} summarized in Table~\ref{learning_summary}. Its dynamics is given in~\cite{ancm,6313077}, and we use $g = 9.8$, $m_c = 1.0$, $m = 0.1$, $\mu_c = 0.5$, $\mu_p = 0.002$, and $l = 0.5$.
\subsubsection{LAG-ROS Training}
$T$, $\mathcal{T}$, and $x_f$ in Sec.~\ref{Sec:performance} are selected as $T=\mathcal{T}=9$(s) and $x_f=[p_f,0,0,0]^{\top}$ for $x$ in Fig.~\ref{lagros_cp_dist_fig}, where $p_f$ is a random terminal position at each episode. We let $\mathcal{R}(0)=0$ and $\bar{d}_{\epsilon} = \bar{b}\epsilon_{\ell}+\bar{d} = 0.75$ in \eqref{robust_learning_bound}, and train the neural net of Sec.~\ref{Sec:training} to have $(\bar{d}_{\epsilon}/{\alpha})\sqrt{\chi} = 3.15$ with $\alpha = 0.60$ using Algorithm~\ref{lagros_alg}. The performance of LAG-ROS \ref{itemLAGROS} is compared with the learning-based planner~\ref{itemFF} and robust tube-based planner~\ref{itemMP}.
\subsubsection{Simulation Results and Discussions}
Note that the robust tube-based motion planner~\ref{itemMP} (\ie{} $u^*$ of Theorem~\ref{Thm:lagros_stability}) is computable with $\Delta t \leq \Delta t_{\rm max}$ (see Sec.~\ref{Sec:computation}) in this case, and thus we expect that the performance of LAG-ROS~\ref{itemLAGROS} should be worse than that of~\ref{itemMP}, as~\ref{itemLAGROS} is a neural net model that approximates~\ref{itemMP}. The right-hand side of Fig.~\ref{lagros_cp_dist_fig} shows the tracking error $\|x-x_d\|$ of \eqref{dynamics} and \eqref{dynamicsd} averaged over $50$ simulations at each time instant $t$. It is still interesting to see that LAG-ROS of \ref{itemLAGROS} and $u^*$ of \ref{itemMP} indeed satisfies the exponential bound \eqref{robust_learning_bound} of Theorem~\ref{Thm:lagros_stability} given as
\begin{align}
r_{\ell}(t) = \mathcal{R}(0)\sqrt{\overline{\omega}}+(\bar{d}_{\epsilon}/{\alpha})\sqrt{\chi}(1-e^{-\alpha t}) = 3.15(1-e^{-0.60 t}) \nonumber
\end{align}
for all $t$ with a small standard deviation $\sigma$, unlike learning-based motion planner~\ref{itemFF} with a diverging bound~\eqref{naive_learning_error} and increasing deviation $\sigma$, as can be seen from Fig.~\ref{lagros_cp_dist_fig}. Contraction theory enables such quantitative analysis on robustness and stability of learning-based planners, which is one of the major advantages of our proposed technique.
\begin{figure}
\centerline{
\subfloat{\includegraphics[clip, width=25mm]{cartpole.pdf}}
\hfil
\subfloat{\includegraphics[clip, width=45mm]{cp_dist.pdf}}
}
\vspace{-1.3em}
\caption{Cart-pole balancing task: $x=[p,\theta,\dot{p},\dot{\theta}]^{\top}$, $x$ and $x_d$ are given in \eqref{dynamics} and \eqref{dynamicsd}, \ref{itemFF} -- \ref{itemLAGROS} are given in Sec.~\ref{Sec:lagros}, and $r_{\ell}$ is given in \eqref{robust_learning_bound}. The shaded area denotes the standard deviation ($+1\sigma$ and $-0.5\sigma$).}
\label{lagros_cp_dist_fig}
\vspace{-1.2em}
\end{figure}
\renewcommand{\arraystretch}{1.1}
\begin{table}
\caption{Control Performances for Cart-Pole Balancing ($\bar{d}_{\epsilon}=0.75$). \\
Note that the notations are as given in Sec.~\ref{Sec:performance}~and~\ref{Sec:computation}. \label{lagros_cp_tab}}
\vspace{-1.2em}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
& Success rate (\%) & $\int \|u(t)\|^2dt$ & $\Delta t$ (s) \\
\hline
\hline
\ref{itemFF} & $40.0$ & $4.79 \times 10^2$ & $3.01\times 10^{-2}$ \\
\ref{itemMP} & $100.0$ & $5.19 \times 10^2$ & $1.00\times 10^{-1}$ \\
\ref{itemLAGROS} & $100.0$ & $5.67 \times 10^2$ & $3.01\times 10^{-2}$ \\
\hline
\end{tabular}
\end{center}
\vspace{-2.5em}
\end{table}
\renewcommand{\arraystretch}{1.0}
Table~\ref{lagros_cp_tab} shows the control performance and computational cost of \ref{itemFF} -- \ref{itemLAGROS}, which is a good summary of their differences and trade-offs aforementioned in Sec.~\ref{Sec:lagros} and in Table~\ref{learning_summary}:
\begin{itemize}
\item \ref{itemFF} approximates $u_d$ of \eqref{dynamics}, and thus requires lower computational cost $\Delta t=0.03$(s) with a smaller objective value $\int \|u\|^2dt=479$, but robustness is not guaranteed (Lemma~\ref{Lemma:naive_learning}) resulting in a $40\%$ success rate.
\item \ref{itemMP} computes $u^*$ of Theorem~\ref{ncm_clf_thm}, and thus possesses robustness as in Fig.~\ref{lagros_cp_dist_fig} resulting in a $100\%$ success rate, but requires larger $\Delta t=0.1$(s) to compute $x_d$.
\item \ref{itemLAGROS} approximates $u^*$ independently of $x_d$, and thus possesses robustness of Theorem~\ref{Thm:lagros_stability} as in Fig.~\ref{lagros_cp_dist_fig} resulting in a $100\%$ success rate, even with $\Delta t = 0.03$(s) as small as that of \ref{itemMP}. It yields $9.2\%$ larger $\int \|u\|^2dt$ than \ref{itemMP} as it models $(x,o_{\ell},t) \mapsto u^*$, not $(x,o_{\ell},t)\mapsto u_d$.
\end{itemize}
It is demonstrated that LAG-ROS indeed possesses the robustness and stability guarantees of $u^*$ as in Theorem~\ref{Thm:lagros_stability}, unlike~\ref{itemFF}, with significantly lower computational cost than that of~\ref{itemMP} as expected.
\subsection{Multi-Agent Nonlinear Motion Planning}
\label{Sec:sc_planning}
The advantages of~\ref{itemLAGROS} demonstrated in Sec.~\ref{Sec:cartpole} are more appreciable in the problem settings where the robust tube-based motion planner~\ref{itemMP} is no longer capable of computing a global solution with $\Delta t \leq \Delta t_{\rm max}$ (see Sec.~\ref{Sec:computation}). We thus consider motion planning and collision avoidance of multiple robotic simulators~\cite{SCsimulator} in a cluttered environment with external disturbances, where the agents are supposed to perform tasks based only on local observations $o_{\ell}$ of \ref{itemFF}. Note that its nonlinear equation of motion is given in~\cite{SCsimulator} and all of its parameters are normalized to $1$.
\subsubsection{LAG-ROS Training}
We select $T$, $\mathcal{T}$, and $x_f$ in Sec.~\ref{Sec:performance} as $T=30$(s), $\mathcal{T}=45$(s), and $x_f=[p_{xf},p_{yf},0,0]^{\top}$, where $(p_{xf},p_{yf})$ is a random position in $(0,0) \leq (p_{xf},p_{yf}) \leq (5,5)$. We let $\mathcal{R}(0)=0$, and train the neural network of Sec.~\ref{Sec:training} by Algorithm~\ref{lagros_alg} to get $(\bar{d}_{\epsilon}/{\alpha})\sqrt{\chi} = 0.125$ with $\alpha = 0.30$ in \eqref{robust_learning_bound}. In particular, the following error tube \eqref{robust_learning_bound}:
\begin{align}
r_{\ell}(t) = (\bar{d}_{\epsilon}/{\alpha})\sqrt{\chi}(1-e^{-\alpha t}) = 0.125(1-e^{-0.30 t}) \nonumber
\end{align}
is used for sampling $(x_d,u_d)$ by Theorem~\ref{tube_lemma} with an input constraint $u_i\geq 0,~\forall i$ to avoid collisions with a random number of multiple circular obstacles ($0.5$m in radius) and of other agents, even under the learning error and disturbances. When training LAG-ROS, the input constraint is satisfied by using a ReLU function for the network output, and the localization technique~\cite{glas} is used to extract $o_{\ell}$ of~\ref{itemFF} from $o_g$ of \eqref{dynamics} for its distributed implementation, with the communication radius $2.0$m. Its performance is compared with~\ref{itemFF},~\ref{itemMP}, and a centralized planner~\ref{itemC} which is not computable with $\Delta t \leq \Delta t_{\rm max}$, where $\Delta t$ is given in Sec.~\ref{Sec:computation}:
\begin{enumerate}[label={\color{caltechgreen}{(\alph*)}},start=4]
\item Centralized robust motion planner:\\
$(x,x_d,u_d,t) \mapsto u^*$, offline centralized solution of \ref{itemMP}. \label{itemC}
\end{enumerate}
\subsubsection{Remarks on Sub-optimal Trajectories}
Multi-agent problems have sub-optimal solutions with optimal values close to the global optimum~\cite{8424034}, and thus~\ref{itemLAGROS} does not necessarily take the same $x_d$ as that of the centralized planner~\ref{itemC} in the presence of disturbances, as depicted in Fig.~\ref{lagros_sc_dist_fig}. However, it implicitly guarantees tracking to its own (sub-)optimal $x_d$ due to Theorem~\ref{Thm:lagros_stability}, which means we can still utilize their objective value $\int\|u\|^2dt$ and success rate to evaluate their performance.
\subsubsection{Implication of Simulation Results}
\label{Sec:implication}
Figure~\ref{lagros_sc_fig} shows one example of the trajectories of the motion planners \ref{itemFF} -- \ref{itemC}, under external disturbances with $\bar{d} = \sup_{x,t}\|d(x,t)\| = 0.4$.
\begin{itemize}
\item For~\ref{itemFF}, the tracking error accumulates exponentially with time due to $\bar{d}$ (Lemma~\ref{Lemma:naive_learning}), which necessitates the use of safety control for avoiding collisions~\cite{glas}. Such non-optimal control inputs also increase the error in \eqref{naive_learning_error} as \ref{itemFF}~does not possess any robustness guarantees.
\item For~\ref{itemMP}, robustness is guaranteed by Theorem~\ref{ncm_clf_thm} but can only obtain locally optimal $(x_d,u_d)$ of \eqref{dynamicsd}, as its time horizon has to be small enough to make the problem solvable within $\Delta t \leq \Delta t_{\rm max} = 0.1$(s), and the agent only has access to local information. This renders some agents stuck in local minima as depicted in Fig.~\ref{lagros_sc_fig}.
\item LAG-ROS \ref{itemLAGROS} tackles these two problems by providing formal robustness and stability guarantees of Theorems~\ref{Thm:lagros_stability} -- \ref{Prop:sampling}, whilst implicitly knowing the global solution (only from the local information $o_{\ell}$ as in~\cite{glas}) without computing it online. It satisfies the given state constraints due to Theorem~\ref{tube_lemma} as can be seen from Fig.~\ref{lagros_sc_fig}
\end{itemize}
\subsubsection{Simulation Results and Discussions}
Figure~\ref{lagros_sc_dist_fig} and Table~\ref{lagros_sc_tab} summarize the simulation results which corroborate the arguments of Sec.~\ref{Sec:implication} implied by Fig.~\ref{lagros_sc_fig}:
\begin{itemize}
\item \ref{itemFF} satisfies the requirement on the computational cost since we have $\Delta t \leq \Delta t_{\rm max} = 0.1$(s) as shown in Table~\ref{lagros_sc_tab}, but its success rate remains the lowest for all $\bar{d}=\sup_{x,t}\|d(x,t)\|$ due to the cumulative error \eqref{naive_learning_error} of Lemma~\ref{Lemma:naive_learning}, as can be seen in Fig.~\ref{lagros_sc_dist_fig} and Table~\ref{lagros_sc_tab}. Although its objective value is the smallest for $\bar{d} \leq 0.6$ since it models $u_d$, it gets larger for larger $\bar{d}$, due to the lack of robustness to keep $x$ around $x_d$.
\item \ref{itemMP}~has a success rate higher than that of~\ref{itemFF}, but still lower than $50$\% as it can only compute sub-optimal $x_d$ under the limited computational capacity. In fact, we have $\Delta t \geq \Delta t_{\rm max}$ in this case, which means it requires a slightly better onboard computer. Also, it uses excessive control effort larger than $10^3$ due to such sub-optimality.
\item \ref{itemLAGROS} achieves more than $90$\% success rates for all $\bar{d}$, and its objective value remains only $1.64$ times larger than that of the centralized planner \ref{itemC}, even without computing $(x_d,u_d)$ of \eqref{dynamicsd} online. Its computational cost is as low as that of \ref{itemFF}, satisfying $\Delta t \leq \Delta t_{\rm max} = 0.1$(s) while retaining a standard deviation smaller than~\ref{itemFF}.
\end{itemize}
These results imply that LAG-ROS indeed enhances learning-based motion planners~\ref{itemFF} with robustness and stability guarantees of contraction theory as in \ref{itemC} (and \ref{itemMP}), thereby bridging the technical gap between them.
\renewcommand{\arraystretch}{1.2}
\begin{table}
\caption{Computational time of each motion planner for Multi-agent Nonlinear Motion Planning.\label{lagros_sc_tab}}
\vspace{-1.5em}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
& \ref{itemFF} & \ref{itemMP} & \ref{itemLAGROS} & \ref{itemC} \\
\hline
\hline
$\Delta t$ (s) & $4.63\times 10^{-2}$ & $1.53\times 10^{-1}$ & $4.66\times 10^{-2}$ & $1.12\times 10^{3}$ \\
\hline
\end{tabular}
\end{center}
\vspace{-2.0em}
\end{table}
\renewcommand{\arraystretch}{1.0}
\begin{figure*}
\centering
\includegraphics[width=125mm]{sc_obs.pdf}
\vspace{-1.2em}
\caption{Trajectories for the learning-based planner~\ref{itemFF}, robust tube-based planner~\ref{itemMP}, LAG-ROS~\ref{itemLAGROS}, and offline centralized solution~\ref{itemC} ($\circ$: start, $\bullet$: goal).}
\label{lagros_sc_fig}
\vspace{-1.3em}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=80mm]{sc_obs_dist.pdf}
\vspace{-1.2em}
\caption{Control performances versus $\sup_{x,t}\|d(x,t)\|$ of \eqref{dynamics} for multi-agent nonlinear motion planning. Note that \ref{itemFF} -- \ref{itemC} are as given in Fig.~\ref{lagros_sc_fig}, and the performances are as defined in Sec.~\ref{Sec:performance}. The shaded area denotes the standard deviation ($\pm10^{-1}\sigma$).}
\label{lagros_sc_dist_fig}
\vspace{-1.5em}
\end{figure}
\if0
\renewcommand{\arraystretch}{1.1}
\begin{table}
\caption{Control Performances for Multi-agent Nonlinear Motion Planning (averaged over $50$ simulations with $\sup\|d\|=0.8$). \\
Note that computational time should be less than the control time interval $\Delta t = 0.1$(s). \label{lagros_sc_tab}}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
& Success rate (\%) & $\int\|u\|^2dt$) & $\Delta t$ (s) \\
\hline
\hline
\ref{itemFF} & $2.5641$ & $2.1787 \times 10^3$ & $8.9623\times 10^{-2}$ \\
\ref{itemMP} & $4.2628\times 10$ & $7.6912 \times 10^3$ & $2.2400\times 10^{-1}$ \\
\ref{itemLAGROS} & $9.0064\times 10$ & $2.1971 \times 10^3$ & $9.5559\times 10^{-2}$ \\
\ref{itemC} & $1.0000\times 10^2$ & $1.7910 \times 10^3$ & $1.2064\times 10^{3}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\renewcommand{\arraystretch}{1.0}
\begin{figure*}
\centerline{
\subfloat{\includegraphics[clip, width=120mm]{sc_obs.pdf}}
\hfil
\subfloat{\includegraphics[clip, width=75mm]{sc_obs_dist.pdf}}
}
\caption{Control performances versus $\sup_{x,t}\|d(x,t)\|$ for multi-agent nonlinear motion planning (averaged over $50$ simulations). \ref{itemFF} -- \ref{itemC} are as in Table~\ref{lagros_sc_tab}.}
\label{lagros_sc_dist_fig}
\end{figure*}
\subsection{Spacecraft Reconfiguration in Low Earth Orbit (LEO)}
We finally consider the problem of spacecraft reconfiguration in LEO as an example of systems equipped with a known Lyapunov function. The dynamical system is presented in~\cite{doi:10.2514/1.55705} accounting for $J_2$ and atmospheric drag perturbations. Since it can be expressed as a fully-actuated Lagrangian system~\cite{doi:10.2514/1.G000218}, we can design a Lyapunov function by its inertia matrix~\cite{Slotine:1228283} and \textcolor{red}{use it for LAG-ROS construction of Proposition~\ref{lagros_prop} due to Theorem~\ref{lag_clf_thm}.} The spacecraft communication radius for local observations $o_{\ell}$ is selected as $2.0$km, which prevents the use of global motion planners. The agents are expected to perform reconfiguration within $75$s under bounded disturbances ($\sup\|d\|=1.0$).
Table~\ref{lagros_leo_tab} shows the simulation results with external disturbances given in Fig.~\ref{lagros_leo_fig} for the robust tube-based motion planner, learning-based motion planner, LAG-ROS, and global motion planner~\cite{doi:10.2514/1.G000218}. LAG-ROS achieves only a $5.9$\% increase in the objective value even with the computational time less than $\Delta t = 0.1$(s). Again, a slightly better onboard computer is required for the robust tube-based planner as its computational time $1.5357\times 10^{-1}$s is greater than $\Delta t = 0.1$(s). Since LAG-ROS models the global robust optimal solution, it also attains more than a $28.4$\% decrease in the objective value when compared with the robust motion planner, even though the success rate of learning-based motion planning is only $2.6$\% due to external disturbances. These results indicate that LAG-ROS can be applied with known Lyapunov functions constructed utilizing a special structure of underlying dynamical systems~\cite{Slotine:1228283,ccm} for the sake of a learning-based robustness guarantee. These trends are the same for other sizes of disturbances as can be seen in Fig.~\ref{lagros_sc_dist_fig}: LAG-ROS keeps more than $98$\% success rate for all the disturbances when averaged over $50$ simulations, with $\mathcal{L}$2 control effort smaller than $2.4$ times that of the global solution, unlike the learning-based and robust tube-based motion planner.
\renewcommand{\arraystretch}{1.1}
\begin{table}
\caption{Control Performances for Multi-agent Spacecraft Reconfiguration in LEO (averaged over $50$ simulations with $\sup\|d\|=1.0$). \\
Note that computational time should be less than the control time interval $\Delta t = 0.1$(s). \label{lagros_leo_tab}}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
& Success rate (\%) & $\int\|u\|^2dt$ & $\Delta t$ (s) \\
\hline
\hline
\ref{itemFF} & $3.0000$ & $1.4636\times 10^2$ & $7.6204\times 10^{-2}$ \\
\ref{itemMP} & $1.0000\times 10^2$ & $2.1109 \times 10^2$ & $1.5357\times 10^{-1}$ \\
\ref{itemLAGROS} & $1.0000\times 10^2$ & $1.5119 \times 10^2$ & $8.3024\times 10^{-2}$ \\
\ref{itemC} & $1.0000\times 10^2$ & $1.4278 \times 10^2$ & $7.4810\times 10^{2}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\renewcommand{\arraystretch}{1.0}
\begin{figure*}
\centerline{
\subfloat{\includegraphics[clip, width=110mm]{leo.pdf}}
\hfil
\subfloat{\includegraphics[clip, width=70mm]{leo_dist.pdf}}
}
\caption{Spacecraft reconfiguration in LEO with baselines (a)~\cite{glas} \& (b)~\cite{tube_nmp} and for (c) LAG-ROS, where (d) shows the solution computed offline assuming global environment observations. LAG-ROS achieves the highest success rate with the control effort closest to that of the global solution.; Control performances versus $\sup_{x,t}\|d(x,t)\|$ for spacecraft reconfiguration in LEO (averaged over $50$ simulations). (a) -- (d)~are as given in Table~\ref{lagros_leo_tab}.}
\label{lagros_leo_dist_fig}
\end{figure*}
\fi
\section{Conclusion}
\label{sec_conclusion}
In this work, we propose a new learning-based motion planning framework, called LAG-ROS, with the formal robustness and stability guarantees of Theorem~\ref{Thm:lagros_stability}. It extensively utilizes contraction theory to provide an explicit exponential bound on the distance between the target and controlled trajectories, even under the existence of the learning error and external disturbances. Simulation results demonstrate that it indeed satisfies the bound in practice, thereby yielding consistently high success rates and control performances in contrast to the existing motion planners~\ref{itemFF} and~\ref{itemMP}. Note that other types of disturbances can be handled, using~\cite{nscm} for stochastic systems and~\cite{ancm} for parametric uncertain systems.
\subsubsection*{Acknowledgments}
This work was in part funded by the Jet Propulsion Laboratory, California Institute of Technology, and benefited from discussions with J. Castillo-Rogez, M. D. Ingham, and J.-J. E. Slotine.
\section{Introduction}
\label{introduction}
In the near future, teams of robotic explorers are expected to perform complex decision-making tasks autonomously in extreme environments, where their motions are typically governed by nonlinear dynamics with external disturbances. For such operations to be successful, they need to compute optimal motion plans while robustly guaranteeing convergence to the target trajectory, both with their limited onboard computational resources. This work thus aims to propose a learning-based robust motion planning and control algorithm that meets these challenging requirements.
\subsubsection*{Related Work}
Learning-based control designs have been an emerging area of research since the rise of neural networks and reinforcement learning~\cite{sutton,ndp}. Model-free approaches learn desired optimal policies using training data obtained in real-world environments, making them robust but not suitable for situations where sampling large training datasets is difficult. Also, proving the robustness and stability properties of such data-driven systems is challenging in general, although some techniques do exist~\cite{boffi2020learning,9029986}. In contrast, model-based methods allow sampling as much data as we want to design optimal policies by, \eg{}, imitation learning~\cite{glas,9001182}, reinforcement learning~\cite{8593871,NIPS2017_766ebcd5}, or both~\cite{NIPS2016_cc7e2b87,8578338,7995721}. However, the learned motion planner could yield cascading errors in the real-world environment if its nominal model poorly represents the true underlying dynamics~\cite{cascading}.
Control theoretical approaches to circumvent such difficulties include robust tube-based motion planning~\cite{7989693,ccm,mypaperTAC,ncm,nscm,chuchu,8814758,9303957,10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355} equipped with a tracking control law for robustness and stability certificates. Among these are contraction theory-based robust control~\cite{ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957}, derived to guarantee to track any feasible target trajectory computed externally by existing motion planners, thereby robustly keeping the system trajectories in a control invariant tube that satisfies given state constraints. Although these provable guarantees are promising, they still assume that the target trajectory can be computed online solving some motion planning problems, unlike the aforementioned learning frameworks.
{\color{caltechgreen}
\subsubsection*{Contributions}
\begin{figure*}
\centering
\includegraphics[width=170mm]{lagros_rev2.png}
{\color{caltechgreen}
\caption{Illustration of LAG-ROS: See Theorems~\ref{Thm:lagros_stability}--\ref{Prop:sampling} and Algorithm~\ref{lagros_alg} for detailed explanation on each block of the LAG-ROS design. Note that LAG-ROS requires only one neural net evaluation to get $u_L$, and Theorem~\ref{Thm:lagros_stability} provides its robustness and stability guarantees.\label{lagrosdrawing}}}
\vspace{-0.5em}
\end{figure*}
\HTnote{In this paper}{Reorganized to emphasize the novel parts of this manuscript, highlighted in Fig.~1}, we present Learning-based Autonomous Guidance with RObustness and Stability guarantees (LAG-ROS) as a novel way to bridge the gap between the learning-based and robust tube-based motion planners. In particular, while LAG-ROS requires only one neural network evaluation to get its control input as in the machine learning schemes~\cite{8593871,NIPS2017_766ebcd5,9001182,glas,NIPS2016_cc7e2b87,8578338,7995721}, its internal contraction theory-based architecture still allows obtaining formal robustness and stability guarantees as in~\cite{ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957}. This framework depicted in Fig.~\ref{lagrosdrawing} is summarized as follows.
\begin{table*}
\caption{{\color{caltechgreen}Comparison of the Proposed Method with the Learning-based and Robust Tube-based Motion Planners.} \label{learning_summary}}
\vspace{-1.5em}
\begin{center}
\begin{tabular}{|l|l|l|l|}
\hline
Motion planning scheme & Policy to be learned & State tracking error $\|x-x_d\|$ & Computational load \\
\hline
\hline
\ref{itemFF} Learning-based motion planner~\cite{8593871,NIPS2017_766ebcd5,9001182,glas,NIPS2016_cc7e2b87,8578338,7995721} & $(x,o,t) \mapsto u_d$ & Increases exponentially (Lemma~\ref{Lemma:naive_learning}) & One neural net evaluation \\
\ref{itemMP} Robust tube-based motion planner~\cite{10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355,ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957} & $(x,x_d,t) \mapsto u^*$ & Exponentially bounded (Theorem~\ref{ncm_clf_thm}) & Computation required to get $x_d$ \\
\ref{itemLAGROS} Proposed method (LAG-ROS) & $(x,o,t) \mapsto u^*$ & Exponentially bounded (Theorem~\ref{Thm:lagros_stability}) & One neural net evaluation \\
\hline
\end{tabular}
\end{center}
\vspace{-2.0em}
\end{table*}
\subsubsection{Robustness and Stability Certificates}
The theoretical foundation of LAG-ROS rests on contraction theory, which utilizes a contraction metric to characterize a necessary and sufficient condition of exponential incremental stability of nonlinear system trajectories~\cite{contraction}. The central result of this paper is that, if there exists a control law which renders a nonlinear system contracting, or equivalently, the closed-loop system has a contraction metric, then LAG-ROS trained to model the controller ensures the Euclidean distance between the target and controlled trajectories to be bounded exponentially with time, linearly in the learning error and size of perturbation. This property helps quantify how small the learning error should be in practice, giving some intuition on the number of samples and length of time in neural net training. We further show that such a contracting control law and its corresponding contraction metric can be designed explicitly via convex optimization, using the method of CV-STEM~\cite{mypaperTAC,ncm,nscm} to minimize a steady-state upper bound of the LAG-ROS tracking error.
\subsubsection{State Constraint Satisfaction}
We further exploit the computed bound on the tracking error in generating expert demonstrations for training, so that the learned policy will not violate given state constraints even with the learning error and external disturbances. In this phase, LAG-ROS learns the contracting control law independently of a target trajectory, making it implementable without solving any motion planning problems online unlike~\cite{10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355,ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957}. The performance of LAG-ROS is evaluated in cart-pole balancing~\cite{6313077} and nonlinear motion planning of multiple robotic agents~\cite{SCsimulator} in a cluttered environment, demonstrating that LAG-ROS indeed satisfies the formal exponential bound as in~\cite{ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957} with its computational load as low as that of existing learning-based motion planners~\cite{8593871,NIPS2017_766ebcd5,9001182,glas,NIPS2016_cc7e2b87,8578338,7995721}.} In particular, LAG-ROS requires less than $0.1$s for computation in all of these tasks and achieves higher control performances and task success rates (see Sec.~\ref{Sec:performance}), when compared with the existing learning-based and robust tube-based motion planners. Table~\ref{learning_summary} summarizes the differences of these motion planning schemes from our proposed method.
\subsubsection*{Notation}
For $x \in \mathbb{R}^n$ and $A \in \mathbb{R}^{n \times m}$, we let $\|x\|$, $\delta x$, and $\|A\|$, denote the Euclidean norm, infinitesimal variation of $x$, and induced 2-norm, respectively.
For a square matrix $A$, we use the notation $A \succ 0$, $A \succeq 0$, $A \prec 0$, and $A \preceq 0$ for the positive definite, positive semi-definite, negative definite, negative semi-definite matrices, respectively, and $\sym(A) = (A+A^{\top})/2$. Also, $I \in \mathbb{R}^{n\times n}$ denotes the identity matrix.
\section{{\color{caltechgreen}Learning-based Robust Motion Planning with Guaranteed Stability (LAG-ROS)}}
\label{Sec:lagros}
\HTnote{In this paper}{Table~1 is added to clarify the advantage of LAG-ROS}, we consider the following nonlinear dynamical systems with a controller $u \in \mathbb{R}^{m}$:
\begin{align}
\label{dynamics}
\dot{x} &= f(x,t)+B(x,t)u+d(x,t)\\
\label{dynamicsd}
\dot{x}_d &= f(x_d,t)+B(x_d,t)u_d(x_d,o(t),t)
\end{align}
where $t\in \mathbb{R}_{\geq0}$, $f:\mathbb{R}^n\times\mathbb{R}_{\geq0} \mapsto \mathbb{R}^{n}$, $B:\mathbb{R}^n\times\mathbb{R}_{\geq0}\mapsto\mathbb{R}^{n\times m}$, $x:\mathbb{R}_{\geq0} \mapsto \mathbb{R}^{n}$ is the state trajectory of the true dynamical system \eqref{dynamics} perturbed by the bounded disturbance $d:\mathbb{R}^{n}\times\mathbb{R}_{\geq0} \mapsto \mathbb{R}^{n}$ \st{} $\sup_{x,t}\|d(x,t)\|=\bar{d}$, $o:\mathbb{R}_{\geq0} \mapsto \mathbb{R}^{p}$ is an environment observation, $x_d \in \mathbb{R}^{n}$ and $u_d:\mathbb{R}^n\times\mathbb{R}^{p}\times\mathbb{R}_{\geq0}\mapsto\mathbb{R}^{n}$ are the target state trajectory and control policy given by existing motion planning algorithms, \eg{}, \cite{10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355,ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957}.
{\color{caltechgreen}
\subsection{Problem Formulation of LAG-ROS}
\HTnote{The problem}{Provided a high-level overview of LAG-ROS} of our interest is to find $u$ that is computable with one neural network evaluation and that guarantees exponential boundedness of $\|x-x_d\|$ in \eqref{dynamics} and \eqref{dynamicsd}, robustly against the learning error and external disturbances. The objective of this paper is thus not to develop new learning-based motion planning algorithms that compute $x_d$ but to augment them with formal robustness and stability guarantees. To this end, let us briefly review the following existing motion planning techniques and their inherent limitations:
\begin{enumerate}[label={\color{caltechgreen}{(\alph*)}}]
\item Learning-based motion planner~\cite{glas} (GLAS) or~\cite{8593871,NIPS2017_766ebcd5,9001182,NIPS2016_cc7e2b87,8578338,7995721}:\\
$(x,o,t) \mapsto u_d$, approximately. \label{itemFF}
\item Robust tube-based motion planner~\cite{7989693} or~\cite{10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355,ccm,mypaperTAC,ncm,nscm,chuchu,8814758,9303957}:\\
$(x,x_d,t) \mapsto u^*$, where $u^*$ is a robust tracking controller. \label{itemMP}
\end{enumerate}
The robust tube-based motion planner~\ref{itemMP} ensures that the perturbed trajectories $x$ of \eqref{dynamics} stay in an exponentially bounded error tube around the target trajectory $x_d$ of \eqref{dynamicsd}~\cite{ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957} (see Theorem~\ref{ncm_clf_thm}). However, it requires the online computation of $x_d$ as an input to their control policy, which is not realistic for systems with limited computational resources.
The learning-based motion planner~\ref{itemFF} circumvents this issue by modeling the target policy $(x,o,t) \mapsto u_d$ by a neural network. In essence, our approach, to be proposed in Theorem~\ref{Thm:lagros_stability}, is for providing~\ref{itemFF} with the contraction theory-based stability guarantees~\ref{itemMP}. We remark that~\ref{itemFF} can only assure the tracking error $\|x-x_d\|$ to be bounded by a function which exponentially increases with time, as to be shown in Lemma~\ref{Lemma:naive_learning} for comparison with LAG-ROS of Theorem~\ref{Thm:lagros_stability}.
\subsection{Stability Certificates of LAG-ROS}
The approach of LAG-ROS bridges the gap between \ref{itemFF} and \ref{itemMP} by ensuring that the distance between the target and controlled trajectories to be exponentially bounded.
\begin{enumerate}[label={\color{caltechgreen}{(\alph*)}},start=3]
\item Proposed approach (LAG-ROS):\\
$(x,o,t) \mapsto u^*$, modeled by a neural network $u_L(x,o,t)$ of Theorem~\ref{Thm:lagros_stability}, where $o$ is a vector containing environment observations~(see Fig.~\ref{lagrosdrawing}) and $u^*(x,x_d,t)$ of \ref{itemMP} is viewed as $u^*(x,x_d(x,o,t),t)$, which is a function of $(x,o,t)$.\label{itemLAGROS}
\end{enumerate}
\begin{theorem}
\label{Thm:lagros_stability}
\HTnote{Suppose}{Added Theorem~1 for formal stability and robustness guarantees of LAG-ROS with learning error} that \eqref{dynamics} is controlled to track \eqref{dynamicsd} by the LAG-ROS control policy $u_L(x,o,t)$, learned to satisfy
\begin{align}
\label{learning_error}
\|u_L(x,o,t)-u^*(x,x_d(x,o,t),t)\| \leq \epsilon_{\ell},~\forall x,o,t
\end{align}
where $\epsilon_{\ell} \in [0,\infty)$ is the learning error, $u^*$ is the target robust control law of \ref{itemMP} (to be designed in Theorem~\ref{ncm_clf_thm}), and $x_d$ is given by the robust motion planner~\ref{itemMP} (to be discussed in Theorem~\ref{tube_lemma}). Now consider the following virtual system which has $x$ of \eqref{dynamics} and $x_d$ of \eqref{dynamicsd} as its particular solutions:
\begin{align}
\label{virtual_lagros}
\dot{y} = \zeta(y,x,x_d,t)+d_y(y,t)
\end{align}
where $\zeta$ and $d_y$ are parameterized by $y$ to verify $\zeta(y=x,x,x_d,t)=f(x,t)+B(x,t)u^*$, $\zeta(y=x_d,x,x_d,t)=f(x_d,t)+B(x_d,t)u_d(x_d,o(t),t)$, $d_y(y=x,t)=B(x,t)(u_L-u^*)+d(x,t)$, and $d_y(y=x_d,t)=0$, and $x$, $x_d$, $u_d$, and $d$ are as defined in \eqref{dynamics} and \eqref{dynamicsd}. Note that $y=x$ and $y=x_d$ are indeed particular solutions of \eqref{virtual_lagros}. If $\exists \bar{b}\in [0,\infty)$ \st{} $\|B(x,t)\| \leq \bar{b},~\forall x,t$, and if $u^*$ is constructed to satisfy the following partial contraction conditions~\cite{Ref:contraction3} with respect to $y$, for a contraction metric $M(y,x,x_d,t)=\Theta^{\top}\Theta \succ 0$ and $\alpha,\underline{\omega},\overline{\omega} \in (0,\infty)$:
\begin{align}
\label{lagros_contraction}
&\dot{M}+2\sym{}\left(M\frac{\partial \zeta}{\partial y}\right) \preceq -2\alpha M,~\forall y,x,x_d,t \\
\label{lagros_Mcon}
&\overline{\omega}^{-1}I \preceq M \preceq \underline{\omega}^{-1}I,~\forall y,x,x_d,t
\end{align}
then we have the following bound for $\mathtt{e}=x-x_d$:
\begin{align}
\label{robust_learning_bound}
\|\mathtt{e}(t)\| \leq \mathcal{R}(0)\sqrt{\overline{\omega}}e^{-\alpha t}+\frac{\bar{b}\epsilon_{\ell}+\bar{d}}{\alpha}\sqrt{\frac{\overline{\omega}}{\underline{\omega}}}(1-e^{-\alpha t}) = r_{\ell}(t)
\end{align}
where $\mathcal{R}(t)=\int^{x}_{x_d}\|\Theta\delta y(t)\|$ for $M=\Theta^{\top}\Theta$.
\end{theorem}
\begin{proof}
Let $V=\int^{x}_{x_d}\delta y^{\top}M\delta y = \int^{x}_{x_d}\|\Theta\delta y\|^2$. Since $\|d_y(x,t)\| \leq \bar{b}\epsilon_{\ell}+\bar{d}=\bar{d}_{\epsilon_{\ell}}$ for $d_y$ in \eqref{virtual_lagros}, the relation \eqref{lagros_contraction} gives
\begin{align}
\dot{V} \leq& \int^{x}_{x_d}\delta y^{\top}\left(\dot{M}+2\sym{}\left(M\frac{\partial \zeta}{\partial y}\right)\right)\delta y+2\bar{d}_{\epsilon_{\ell}}\int^{x}_{x_d}\|M \delta y\| \\
\leq& -2\alpha V+\frac{2\bar{d}_{\epsilon_{\ell}}}{\sqrt{\underline{\omega}}}\int^{x}_{x_d}\|\Theta \delta y\|
\end{align}
Since $d(\|\Theta\delta y\|^2)/dt = 2\|\Theta\delta y\|(d\|\Theta\delta y\|/dt)$, this implies that $\dot{\mathcal{R}} \leq -\alpha\mathcal{R}+{\bar{d}_{\epsilon_{\ell}}}/{\sqrt{\underline{\omega}}}$. Therefore, applying the comparison lemma~\cite[pp.102-103, pp.350-353]{Khalil:1173048} (\ie{}, if $\dot{v}_1 \leq h(v_1,t)$ for $v_1(0) \leq v_2(0)$ and $\dot{v}_2 =h(v_2,t)$, then $v_1(t) \leq v_2(t)$), along with the relation $\mathcal{R}(t) \geq \|\mathtt{e}(t)\|/\sqrt{\overline{\omega}}$, results in \eqref{robust_learning_bound}
\end{proof}
Theorem~\ref{Thm:lagros_stability} implies that the bound \eqref{robust_learning_bound} decreases linearly in the learning error $\epsilon_{\ell}$ and disturbance $\bar{d}$, and \eqref{dynamics} controlled by LAG-ROS is exponentially stable when $\epsilon_{\ell}=0$ and $\bar{d} = 0$, showing a great improvement over \ref{itemFF} which only gives an exponentially diverging bound as to be derived in Lemma~\ref{Lemma:naive_learning}~\cite{8593871,NIPS2017_766ebcd5,9001182,glas,NIPS2016_cc7e2b87,8578338,7995721}. This property permits quantifying how small $\epsilon_{\ell}=0$ should be to meet the required performance of motion planning, giving some intuition on the neural network architecture (see Sec.~\ref{sec_simulation}). Also, since we model $u^*$ by $u_L(x,o,t)$ independently of $x_d$, it is indeed implementable without solving any motion planning problems online unlike robust motion planners \ref{itemMP}~\cite{10.1007/BFb0109870,tube_mp,tube_nmp,doi:10.1177/0278364914528132,9290355,ccm,7989693,mypaperTAC,ncm,nscm,chuchu,8814758,9303957}, as outlined in Table~\ref{learning_summary}. If we can sample training data of $u^*$ explicitly considering the bound \eqref{robust_learning_bound}, the LAG-ROS control enables guaranteeing given state constraints even with the learning error $\epsilon_{\ell}$ and external disturbance $d(x,t)$, as will be seen in Sec.~\ref{Sec:contraction}~and~\ref{imitation_learning}.
\begin{remark}
$\epsilon_{\ell}$ of \eqref{learning_error} can only be found empirically in practice, and thus we propose one way to generate training data with \eqref{robust_learning_bound} and use the test error of $u_L$ as $\epsilon_{\ell}$~(see Sec.~\ref{sec_robust_sampling} and Sec.~\ref{sec_simulation}). Note that models $f(x,t)$ learned by system identification for a more accurate description of the nominal dynamics, \eg{}, \cite{8794351}, is still utilizable in \eqref{virtual_lagros} as long as the learned system is contracting~\cite{boffi2020learning}, and the modeling error is bounded. Other types of perturbations, such as stochastic or parametric uncertainty, could be handled using~\cite{nscm,ancm}.
\end{remark}
\HTnote{To appreciate}{Added Lemma~1 for showing that existing learning schemes only have an exponentially diverging bound} the importance of the guarantees in Theorem~\ref{Thm:lagros_stability}, let us additionally show that~\ref{itemFF}, which models $(x,o,t) \mapsto u_d$, only leads to an exponentially diverging bound.
\begin{lemma}
\label{Lemma:naive_learning}
Suppose that $u$ of \eqref{dynamics} is learned to satisfy
\begin{align}
\label{learning_error0}
\|u(x,o,t)-u_d(x,o,t)\| \leq \epsilon_{\ell},~\forall x,o,t
\end{align}
for $u_d$ of \eqref{dynamicsd} with the learning error $\epsilon_{\ell} \in [0,\infty)$, and that $\exists \bar{b}$ \st{} $\|B(x,t)\| \leq \bar{b},~\forall x,t$. If $f_{c\ell}=f(x,t)+B(x,t)u_d$ is Lipschitz, \ie{}, $\exists L_{f} \in [0,\infty)$ \st{} $\|f_{c\ell}(x_1,t)-f_{c\ell}(x_2,t)\| \leq L_f\|x_1-x_2\|,~\forall x_1,x_2 \in \mathbb{R}^n$, then we have the following bound:
\begin{align}
\label{naive_learning_error}
\|\mathtt{e}(t)\| \leq \|\mathtt{e}(0)\|e^{L_{f} t}+L_{f}^{-1}(\bar{b}\epsilon_{\ell}+\bar{d})(e^{L_{f} t}-1).
\end{align}
where $\mathtt{e} = x-x_d$, and $x$, $x_d$, and $\bar{d}$ are given in \eqref{dynamics} and \eqref{dynamicsd}.
\end{lemma}
\begin{proof}
Integrating \eqref{dynamics} and \eqref{dynamicsd} for $u$ in \eqref{learning_error0} yields $\|\mathtt{e}(t)\| \leq \|\mathtt{e}(0)\|+L_f\int_{0}^t\|\mathtt{e}(\tau)\|d\tau+(\bar{b}\epsilon_{\ell}+\bar{d})t$. Applying the Gronwall-Bellman inequality~\cite[pp. 651]{Khalil:1173048} gives
\begin{align}
\|\mathtt{e}(t)\| \leq& \|\mathtt{e}(0)\|+\bar{d}_{\epsilon_{\ell}}t+L_{f}\int_{0}^t(\|\mathtt{e}(0)\|+\bar{d}_{\epsilon_{\ell}}\tau)e^{L_{f}(t-\tau)}d\tau
\end{align}
where $\bar{d}_{\epsilon_{\ell}} = \bar{b}\epsilon_{\ell}+\bar{d}$ and $\Delta \tau = \tau-t_0$. Thus, integration by parts results in the desired relation \eqref{naive_learning_error}.
\end{proof}
Lemma~\ref{Lemma:naive_learning} indicates that if there exists either a learning error $\epsilon_{\ell}$ or external disturbance $d$, the tracking error bound grows exponentially with time, and thus \eqref{naive_learning_error} becomes no longer useful for large $t$. Section~\ref{sec_simulation} demonstrates how the computed bounds of \eqref{robust_learning_bound} ($\lim_{t\to \infty}e^{-\alpha t} = 0$) and \eqref{naive_learning_error} ($\lim_{t\to \infty}e^{L_f t} = \infty$) affect the control performance in practice.}
\section{Contraction Theory-based Robust and Optimal Tracking Control}
\label{Sec:contraction}{\color{caltechgreen}
\HTnote{Theorem}{Clarified the role of Thm~2 for getting $u^*$, not $(x_d,u_d)$}~\ref{Thm:lagros_stability} is subject to the assumption that we have a contraction theory-based robust tracking control law $u^*$, which satisfies \eqref{lagros_contraction} and \eqref{lagros_Mcon} for a given $(x_d,u_d)$. This section thus delineates one way to extend the method called ConVex Optimization-based Steady-state Tracking Error Minimization (CV-STEM)~\cite{mypaperTAC,ncm,nscm} to find a contraction metric $M$ of Theorem~\ref{Thm:lagros_stability}, which minimizes an upper bound of the steady-state error of \eqref{robust_learning_bound} via convex optimization. Minimizing the bound renders the tube-based planning of Theorem~\ref{tube_lemma} to be given in Sec.~\ref{imitation_learning} for sampling training data less conservative, resulting in a better optimal solution for $(x_d,u_d)$.}
In addition, we modify the CV-STEM in~\cite{mypaperTAC} to derive a robust control input $u^*$ which also greedily minimizes the deviation of $u^*$ from the target $u_d$, using the computed contraction metric $M$ to construct a differential Lyapunov function $V = \delta y^{\top}M\delta y$. Note that $u^*$ is to be modeled by a neural network which maps $(x,o,t)$ to $u^*$ implicitly accounting for $(x_d,u_d)$ as described in Theorem~\ref{Thm:lagros_stability}, although $u^*$ takes $(x,x_d,t)$ as its inputs (see Sec.~\ref{sec_robust_sampling}).
\subsection{Problem Formulation of CV-STEM Tracking Control}
For given $(x_d,u_d)$, we assume that $u^*$ of Theorem~\ref{Thm:lagros_stability} can be decomposed as $u^* = u_d(x_d,o(t),t)+K(x,x_d,t)(x-x_d)$, the generality of which is guaranteed by the following lemma.
\begin{lemma}
\label{u_equivalence_lemma}
Consider a general tracking controller $u$ defined as $u = k(x,x_d,t)$ with $k(x_d,x_d,t)=u_d(x_d,o(t),t)$, where $k:\mathbb{R}^n\times\mathbb{R}^n\times\mathbb{R}_{\geq0}\mapsto\mathbb{R}^{m}$. If $k$ is piecewise continuously differentiable, then $\exists K:\mathbb{R}^n\times\mathbb{R}^n\times\mathbb{R}_{\geq0}\mapsto\mathbb{R}^{m\times n}$ \st{} $u = k(x_d,x_d,t) = u_d(x_d,o(t),t)+K(x,x_d,t)(x-x_d)$.
\end{lemma}
\begin{proof}
We have $u=u_d+(k(x,x_d,t)-k(x_d,x_d,t))$ due to $k(x_d,x_d,t)=u_d(x_d,o(t),t)$. Since $k(x,x_d,t)-k(x_d,x_d,t)=\int_0^1(dk(c x+(1-c)x_d,x_d,t)/dc)dc$, choosing $K$ as $\int_0^1(\partial k(c x+(1-c)x_d,x_d,t)/\partial x)dc$ gives the desired relation.
\end{proof}
Lemma~\ref{u_equivalence_lemma} implies that designing optimal $k$ of $u^* = k(x,x_d,t)$ reduces to designing optimal $K(x,x_d,t)$ of $u^*=u_d+K(x,x_d,t)(x-x_d)$. When \eqref{dynamics} is controlled by the LAG-ROS $u_L$ of Theorem~\ref{Thm:lagros_stability} with such $u^*$, the virtual system of \eqref{dynamics} and \eqref{dynamicsd} which has $y=x$ and $y=x_d$ as its particular solutions can be given by \eqref{virtual_lagros}, where $\zeta$ is defined as follows:
\begin{align}
\label{def:zeta}
\zeta(y,x,x_d,t) =& \dot{x}_d+(A(x,x_d,t)+B(x,t)K(x,x_d,t))(y-x_d)
\end{align}
where $A$ is the State-Dependent Coefficient (SDC) form of the dynamical system \eqref{dynamics} given by Lemma~2 of~\cite{nscm}, which verifies $A(x,x_d,t)(x-x_d) = f(x,t)+B(x,t)u_d-f(x_d,t)-B(x_d,t)u_d$. Note that $\zeta$ indeed satisfies $\zeta(x,t)=f(x,t)+B(x,t)u^*$ and $\zeta(x_d,t)=f(x_d,t)+B(x_d,t)u_d$ for such $A$, to have $y=x$ and $y=x_d$ as the particular solutions to \eqref{virtual_lagros}.
\subsection{CV-STEM Contraction Metrics as Lyapunov Functions}
The remaining task is to construct $M$ so that it satisfies \eqref{lagros_contraction} and \eqref{lagros_Mcon}. The CV-STEM approach suggests that we can find such $M$ via convex optimization to minimize an upper bound of \eqref{robust_learning_bound} as $t\to \infty$ when $\alpha$ of \eqref{lagros_contraction} is fixed. Theorem~\ref{ncm_clf_thm} proposes using the metric $M$ designed by the CV-STEM for a Lyapunov function, thereby augmenting $u^*$ with additional optimality to greedily minimize $\|u^*-u_d\|^2$ for $u_d$ in \eqref{dynamicsd}.
\begin{theorem}
\label{ncm_clf_thm}
Suppose that $f$ and $B$ are piecewise continuously differentiable, and let $B=B(x,t)$ and $A=A(x,x_d,t)$ in \eqref{def:zeta} for notational simplicity. Consider a contraction metric $M(x,x_d,t) = W(x,x_d,t)^{-1} \succ 0$ given by the following convex optimization (CV-STEM)~\cite{ncm,mypaperTAC,nscm} to minimize an upper bound on the steady-state tracking error of \eqref{robust_learning_bound}:
\begin{align}
\label{convex_opt_ncm}
&{J}_{CV}^* = \min_{\nu>0,\chi \in \mathbb{R},\bar{W}\succ 0} \frac{\bar{b}\epsilon_{\ell}+\bar{d}}{\alpha}\chi \text{~~\st{} \eqref{deterministic_contraction_tilde} and \eqref{W_tilde}}
\end{align}
with the convex constraints \eqref{deterministic_contraction_tilde} and \eqref{W_tilde} given as
\begin{align}
\label{deterministic_contraction_tilde}
&-\dot{\bar{W}}+2\sym{\left(A\bar{W}\right)}-2\nu BR^{-1}B^{\top} \preceq -2\alpha \bar{W},~\forall x,x_d,t \\
\label{W_tilde}
&I \preceq \bar{W}(x,x_d,t) \preceq \chi I,~\forall x,x_d,t
\end{align}
where $\alpha,\underline{\omega},\overline{\omega} \in (0,\infty)$, $\nu = 1/\underline{\omega}$, $\chi = \overline{\omega}/\underline{\omega}$, $\bar{W} = \nu W$, and $R=R(x,x_d,t)\succ0$ is a given weight matrix on the control input. Suppose also that $u^*$ of Theorem~\ref{Thm:lagros_stability} is given by $u^* = u_d(x_d,o(t),t)+K^*(x,x_d,t)\mathtt{e}$, where $\mathtt{e}=x-x_d$, and $K^*$ is given by the following convex optimization $(x,x_d,t)$:
\begin{align}
\label{ncm_robust_control}
&K^* = \text{arg}\min_{K\in \mathbb{R}^{m\times n}}\|u-u_d\|^2 = \text{arg}\min_{K\in \mathbb{R}^{m\times n}}\|K(x,x_d,t)\mathtt{e}\|^2 \\
\label{stability_clf}
&\text{\st{} } \dot{M}+2\sym{}(MA+MBK(x,x_d,t)) \preceq -2\alpha M.
\end{align}
Then $M$ satisfies \eqref{lagros_contraction} and \eqref{lagros_Mcon} for $\zeta$ defined in \eqref{def:zeta}, and thus \eqref{robust_learning_bound} holds, \ie{}, we have the exponential bound on the tracking error $\|x-x_d\|$ when the dynamics \eqref{dynamics} is controlled by the LAG-ROS control input $u_L$ of Theorem~\ref{Thm:lagros_stability}. Furthermore, the problem \eqref{ncm_robust_control} is always feasible.
\end{theorem}
\begin{proof}{\color{caltechgreen}
\HTnote{Since}{Added details to make it self-contained} the virtual dynamics of \eqref{virtual_lagros} with \eqref{def:zeta} is given as $\delta \dot{y}=(\partial \zeta/\partial y)\delta y = (A-BK)\delta y$, substituting this into \eqref{lagros_contraction} verifies that \eqref{lagros_contraction} and \eqref{stability_clf} are equivalent. For $\bar{K}=-R^{-1}B^{\top}M$, \eqref{stability_clf} can be rewritten as
\begin{equation}
\nu^{-1}M(-\dot{\bar{W}}+2\sym{}(A\bar{W})-\nu BR^{-1}B^{\top})M \preceq -2\alpha \nu^{-1}M\bar{W}M.
\end{equation}
Since this is clearly feasible as long as $M$ satisfies the condition \eqref{deterministic_contraction_tilde}, this implies that \eqref{ncm_robust_control} is always feasible. Also, multiplying \eqref{lagros_Mcon} by $W$ from both sides and then by $\nu$ gives \eqref{W_tilde}. These facts indicate that the conditions \eqref{lagros_contraction} and \eqref{lagros_Mcon} are satisfied for $M$ and $u^*$ constructed by \eqref{convex_opt_ncm} and \eqref{ncm_robust_control}, respectively, and thus we have the exponential bound \eqref{robust_learning_bound} as a result of Theorem~\ref{Thm:lagros_stability}. Furthermore, the problem \eqref{convex_opt_ncm} indeed minimizes an upper bound of \eqref{robust_learning_bound} as $t \to \infty$ due to the relation $0\leq\sqrt{\overline{\omega}/\underline{\omega}} = \sqrt{\chi} \leq \chi$. We remark that \eqref{convex_opt_ncm} is convex as the objective is affine in $\chi$ and \eqref{deterministic_contraction_tilde} and \eqref{W_tilde} are linear matrix inequalities in terms of $\nu$, $\chi$, and $\bar{W}$.}
\end{proof}
\begin{remark}
\label{optimal_remark}
\eqref{convex_opt_ncm} and \eqref{ncm_robust_control} are convex in terms of their respective decision variables and thus can be solved computationally efficiently~\cite[pp. 561]{citeulike:163662}. For systems with a known Lyapunov function (\eg{} Lagrangian systems~\cite[pp. 392]{Slotine:1228283}), we could simply use it to get robust tracking control $u^*$ in Theorem~\ref{ncm_clf_thm} without solving \eqref{ncm_robust_control}, although optimality may no longer be guaranteed in this case.
\end{remark}
\begin{remark}
The contraction metric construction itself can be performed using a neural network~\cite{ncm,nscm,ancm,chuchu}, leading to an analogous incremental stability and robustness results to those of Theorem~\ref{Thm:lagros_stability}~\cite{ancm,cdc_ncm}.
\end{remark}
\if0
\begin{corollary}
Consider the following control law:
\begin{align}
\label{clqqp_solution}
u = \begin{cases} u_d-\frac{\varphi_f+\varphi_g^{\top}u_d}{\varphi_g^{\top}\varphi_g}\varphi_g & \text{if $e^T(\dot{M}+2MA+2\alpha M)e > 0$}\\ u_d & \text{if $e^T(\dot{M}+2MA+2\alpha M)e \leq 0$}\end{cases}
\end{align}
where $\varphi_f = e^T\dot{M}e+2e^TM(Ae-Bu_d)+2\alpha e^{\top}Me$, $\varphi_g = 2B^{\top}Me$, $B=B(x,t)$, $A=A(x,x_d,t)$. Then \eqref{clqqp_solution} is the solution to \eqref{ncm_robust_control} of Theorem~\ref{ncm_clf_thm}, and thus satisfies the bound \eqref{dist_dist}.
\end{corollary}
\begin{proof}
Since \eqref{ncm_robust_control} is convex, the KKT conditions~\cite[pp. 244]{citeulike:163662} provide necessary and sufficient conditions for optimality, and \eqref{clqqp_solution} follows from its application to \eqref{ncm_robust_control}.
\end{proof}
\fi
\if0
\subsection{Dynamical Systems with Known Lyapunov Functions}
\label{sec_systems_w_known_lyapunov}
There are some science and engineering applications where we know a Lyapunov function candidate of a given nonlinear system, one of which is a feedback linearizable system~\cite{Isidori:1995:NCS:545735,Slotine:1228283,Khalil:1173048}. In this section, we illustrate how to construct robust tracking control for such systems, taking Lagrangian systems as an example.
Consider the following system with a bounded disturbance $d(x,t)$:
\begin{align}
\label{sto_lag_system}
{H}(q)\ddot{q}+{C}(q,\dot{q})\dot{q}+{G}(q)=u+d(x,t)
\end{align}
where $q:\mathbb{R}_{\geq0} \to \mathbb{R}^{n}$, $u:\mathbb{R}_{\geq0} \to \mathbb{R}^{n}$, ${H}:\mathbb{R}^n \to \mathbb{R}^{n\times n}$, ${C}:\mathbb{R}^n \times \mathbb{R}^n \to \mathbb{R}^{n\times n}$, ${G}:\mathbb{R}^n \to \mathbb{R}^{n}$, and $d:\mathbb{R}^n\times\mathbb{R}_{\geq 0} \to \mathbb{R}^{n}$ with $\bar{d}=\sup_{x,t}\|d(x,t)\| < +\infty$ and $x = [q^{\top},\dot{q}^{\top}]^{\top}$. We note that the matrix ${C}(q,\dot{q})$ is selected to make $\dot{{H}}-2{C}$ skew-symmetric, so we have a useful property \st{} $z^{\top}(\dot{{H}}-2{C})z=0,~\forall z\in\mathbb{R}^n$.
\begin{theorem}
\label{lag_clf_thm}
Suppose that \eqref{sto_lag_system} is controlled by
\begin{align}
\label{lag_con}
u &= u_n = {H}(q)\ddot{q}_r+{C}(q,\dot{q})\dot{q}_r+{G}(q)-{K_{\ell}}(t)s
\end{align}
where $\dot{q}_r = \dot{q}_d-\Lambda (q-q_d)$, $s=\dot{q}-\dot{q}_r$, ${K_{\ell}} \succ 0$, and $\Lambda\succ 0$. If ${K_{\ell}}$, $\Lambda$, $\varepsilon>0$ are selected so that $\exists\alpha>0$ \st{} $\begin{bmatrix}2{K_{\ell}} & -\varepsilon I \\ -\varepsilon I & 2 \Lambda\end{bmatrix} \preceq -2\alpha M$, where $M=\Theta^{\top}\Theta=\diag{}({H},\varepsilon I)$, then we have that
\begin{align}
\label{dist_dist_lag}
\|x-x_d\| \leq \frac{1}{\sqrt{\underline{p}}}\left(\mathcal{R}(0)e^{-\alpha t}+\frac{\bar{d}}{\alpha}\frac{1}{\sqrt{\underline{m}}}\right)=R_{\rm NCM}(t)
\end{align}
where $x_d=[q_d^{\top},\dot{q}_d^{\top}]^{\top}$, $\mathcal{R}(t)=\int^{\xi}_{0}\|\Theta(q(t))\delta y(t)\|$, $\xi=[s^{\top},e^{\top}]^{\top}$, $e=x-x_d$, $\underline{m}I \preceq M$, $P = \begin{bmatrix}\Lambda{H}\Lambda+\varepsilon I &\Lambda {H} \\ {H}\Lambda & {H}\end{bmatrix} \succ 0$, $\underline{p}I \preceq P$, and $y$ is the state of the virtual system of \eqref{sto_lag_system}, which has $[s^{\top},e^{\top}]^{\top}$ and $[0^{\top},0^{\top}]^{\top}$ as its particular solutions.
\end{theorem}
\begin{proof}
For the virtual system defined as ${H}\dot{y}_1+{C}y_1+{K_{\ell}}y_1=d_s(y,t)$ and $\dot{y}_2=y_1-\Lambda y_2$, we have $y = [y_1^{\top},y_2^{\top}]^{\top} = \xi = [s^{\top},e^{\top}]^{\top}$ and $y = 0$ as its particular solutions when $d_s(\xi,t) = d(x,t)$ and $d_s(0,t) = 0$. Thus using a differential Lyapunov function $V = \delta y^{\top}M\delta y$, we can show that $\dot{\mathcal{R}} \leq -\alpha \mathcal{R}+\bar{d}/\sqrt{\underline{m}}$ under the condition of this theorem~\cite{modular_stability,Slotine:1228283}. Since we have $\xi^{\top}M\xi=x^{\top}Px$ by definition, applying the comparison lemma~\cite[pp.102-103, pp.350-353]{Khalil:1173048} (\ie{}, if $\dot{v}_1 \leq h(v_1,t)$ for $v_1(0) \leq v_2(0)$ and $\dot{v}_2 =h(v_2,t)$, then $v_1(t) \leq v_2(t)$) yields the desired relation.
\end{proof}
\begin{remark}
The control parameters ${K_{\ell}}$ and $\Lambda$ in Theorem~\ref{lag_clf_thm} could also be optimized using the same technique in Theorem~\ref{ncm_clf_thm} or the one in Remark~\ref{optimal_remark}~\cite{mypaperTAC}.
\end{remark}
\fi
\section{Generating Expert Demonstrations}
\label{imitation_learning}
{\color{caltechgreen}
\HTnote{This section}{Clarified the role of Thm~3 for state constraint satisfaction} presents how to sample training data, \ie{}, the target trajectory $(x_d,u_d)$ of \eqref{dynamicsd} and corresponding perturbed state and CV-STEM robust control $(x,u^*)$, so that $x$ of \eqref{dynamics} controlled by the LAG-ROS control $u_L$ will stay in given admissible state space as long as we have $\|u_L-u^*\| \leq \epsilon_{\ell}$ (\eqref{learning_error} of Theorem~\ref{Thm:lagros_stability}). To be specific, since $u^*$ of Theorem~\ref{ncm_clf_thm} solves \eqref{convex_opt_ncm} to obtain an optimal bounded error tube around $x_d$, \ie{}, \eqref{robust_learning_bound} of Theorem~\ref{Thm:lagros_stability}, we exploit it in the tube-based motion planning~\cite{7989693} for generating training data, which satisfies given state constraints even under the existence of the learning error $\epsilon_{\ell}$ and disturbance $d$ in \eqref{robust_learning_bound} and \eqref{convex_opt_ncm}. Note that the learning error $\epsilon_{\ell}$ and disturbance upper bound $\bar{d}$ of Theorem~\ref{Thm:lagros_stability} are assumed to be selected \textit{a priori}~\cite{8794351}.}
\subsection{Tube-based State Constraint Satisfaction in LAG-ROS}
Given the learning error $\epsilon_{\ell}$ and disturbance upper bound $\bar{d}$ of Theorem~\ref{Thm:lagros_stability}, we sample $(x_d,u_d)$ of \eqref{dynamicsd} by solving the following tube-based motion planning problem:
\begin{align}
\label{tube_motion_plan}
&\min_{\bar{x}(t),\bar{u}(t)}~c_1\int_{0}^{T}\|\bar{u}(t)\|^2dt+c_2\int_0^TP(\bar{x}(t),\bar{u}(t),t)dt \\
&\text{\st{} $\dot{\bar{x}} = f(\bar{x},t)+B(\bar{x},t)\bar{u}$, $\bar{x}(t)\in\bar{\mathcal{X}}(t)$, and $\bar{u}(t) \in\bar{\mathcal{U}}(t)$}
\end{align}
where $c_1>0$, $c_2\geq0$, $P(\bar{x},\bar{u},t)$ is some performance-based cost function, $\bar{\mathcal{X}}$ is robust admissible state space defined as $\bar{\mathcal{X}}(t)=\{\bar{x}(t)\in\mathbb{R}^n|\forall \xi(t) \in \{\xi(t)|\|\bar{x}(t)-\xi(t)\|\leq r_{\ell}(t)\},~\xi(t)\in\mathcal{X}(t)\}$, $\mathcal{X}(t)$ is given admissible state space, $r_{\ell}(t)$ is the right-hand side of \eqref{robust_learning_bound} in Theorem~\ref{Thm:lagros_stability}, and $\bar{u}(t) \in\bar{\mathcal{U}}(t)$ is an input constraint. The following theorem shows that the LAG-ROS control ensures the perturbed state $x$ of \eqref{dynamics} to satisfy state constraints $x \in \mathcal{X}$, due to the contracting property of $u^*$ in Theorem~\ref{ncm_clf_thm}.{\color{caltechgreen}
\begin{theorem}
\label{tube_lemma}
\HTnote{If the solution}{Added LAG-ROS learning error to formally show that it is implementable only with one neural net evaluation} of \eqref{tube_motion_plan} yields $(x_d,u_d)$, the LAG-ROS control $u_L$ of Theorem~\ref{Thm:lagros_stability}, where $M$ and $u^*$ are designed by Theorem~\ref{ncm_clf_thm}, ensures the perturbed solution $x(t)$ of \eqref{dynamics} to stay in the admissible state space $\mathcal{X}(t)$, \ie{}, $x(t) \in \mathcal{X}(t),\forall t$, even with the learning error $\epsilon_{\ell}$ of $u_L$ in \eqref{learning_error}.
\end{theorem}
\begin{proof}
Since $x_d \in \bar{\mathcal{X}}$, we have $\xi \in \mathcal{X}$ for all $\xi \in S(x_d) = \{\xi|\|x_d-\xi\|\leq r_{\ell}\}$ by the definition of $\bar{\mathcal{X}}$. Also, if we control \eqref{dynamics} by $u_L$, we have $\|x-x_d\| \leq r_{\ell}$, or equivalently, $x\in S(x_d)$. These two statements indicate that $x \in \mathcal{X}$.
\end{proof}
\begin{remark}
Theorem~\ref{tube_lemma} indicates that if $x_d$ is sampled by \eqref{tube_motion_plan}, the perturbed trajectory \eqref{dynamics} controlled by LAG-ROS $u_L$ of Theorem~\ref{Thm:lagros_stability} will not violate the given state constraints as long as LAG-ROS $u_L$ of Theorem~\ref{Thm:lagros_stability} satisfies \eqref{learning_error}. This helps greatly reduce the need for safety control schemes, unlike~\cite{glas}.
\end{remark}}
\subsection{Learning Contraction Theory-based Robust Control}
\label{sec_robust_sampling}
To meet the learning error requirement \eqref{learning_error} for the sake of robustness and state constraint satisfaction proposed in Theorems~\ref{Thm:lagros_stability} and~\ref{tube_lemma}, we should view $x_d$ sampled by \eqref{tube_motion_plan} also as a function of $x$ as mentioned in \ref{itemLAGROS} of Sec.~\ref{Sec:lagros}, so that we can also sample perturbed states $x$ and associated CV-STEM robust control inputs $u^*$ of Theorem~\ref{ncm_clf_thm}.
\begin{figure}
\centering
\includegraphics[width=80mm]{sample_rev.pdf}
\caption{Illustration of state sampling in robust bounded error tube.}
\label{samplingfig}
\vspace{-1.0em}
\end{figure}
\begin{theorem}
\label{Prop:sampling}
For a given target trajectory $(x_d,u_d)$ at time $t$ sampled by solving \eqref{tube_motion_plan}, its associated perturbed state $x$ lies in $S(x_d) = \{\xi\in\mathbb{R}^n|\|x_d-\xi\|\leq r_{\ell}(t)\}$, where $r_{\ell}(t)$ is the right-hand side of \eqref{robust_learning_bound}, when the system \eqref{dynamics} is controlled by $u_L$ of Theorem~\ref{Thm:lagros_stability}. Thus, the CV-STEM control $u^*$ of Theorem~\ref{ncm_clf_thm} can be sampled by solving \eqref{ncm_robust_control} for such $(x,x_d,u_d)$, where $x \in S(x_d)$ (see Fig.~\ref{samplingfig}).
\end{theorem}
\begin{proof}
We have $x \in S(x_d)$ due to \eqref{robust_learning_bound}. Hence, $u^*$ can be sampled for a given $(x_d,u_d)$ with artificially perturbed states $x$ in $S(x_d)$ simply by solving \eqref{ncm_robust_control}.
\end{proof}
Using Theorem~\ref{Prop:sampling}, multiple CV-STEM robust control samples could be obtained for each target trajectory $(x_d,u_d)$ of Theorem~\ref{tube_lemma}. These samples are to be used for training the LAG-ROS neural network, so we can satisfy the condition \eqref{learning_error} of Theorem~\ref{Thm:lagros_stability} for the pre-defined learning error $\epsilon_{\ell}$.
\begin{remark}
The environment observations $o$ of \eqref{learning_error} are also sampled by \eqref{tube_motion_plan} as in~\cite{glas}, just to render LAG-ROS applicable to different environments with a single policy (see Sec.~\ref{Sec:sc_planning}).
\end{remark}
The pseudocode for the offline construction of the LAG-ROS control of Theorem~\ref{Thm:lagros_stability} is presented in Algorithm~\ref{lagros_alg} (see Fig.~\ref{lagrosdrawing} for its visual description). Once we get $u_L(x,o,t)$ by Algorithm~\ref{lagros_alg}, $u=u_L$ in \eqref{dynamics} can be easily computed by only evaluating the neural net $u_L$ for a given $(x(t),o(t),t)$ observed at time $t$, whilst ensuring robustness, stability, and state constraint satisfaction due to Theorems~\ref{Thm:lagros_stability}, \ref{ncm_clf_thm}, and \ref{tube_lemma}.
\begin{algorithm}
\SetKwInOut{Input}{Inputs}\SetKwInOut{Output}{Outputs}
\BlankLine
\Input{Random environments $\{\mathcal{E}_i\}^{N}_{i=1}$ \\
Contraction metric $M(x,x_d,t)$ of Theorem~\ref{ncm_clf_thm} \\
Motion planner $\mathcal{P}$ \\ Learning error $\epsilon_{\ell}$ and disturbance bound $\bar{d}$}
\Output{LAG-ROS control $u_L$ of Theorem~\ref{Thm:lagros_stability}}
\For{$i\leftarrow 1$ \KwTo $N$}{
Solve \eqref{tube_motion_plan} of Theorem~\ref{tube_lemma} for $\mathcal{E}_i$ using $\mathcal{P}$ and obtain a target trajectory and environment observation history $(x_d,u_d,o,t)_i$ \\
Sample $D$ robust CV-STEM control $\{(x,u^*)_{ij}\}_{j=1}^D$ by Theorem~\ref{ncm_clf_thm} using Theorem~\ref{Prop:sampling} (see Fig.~\ref{samplingfig})
}
Model $(x,o,t)_{ij} \mapsto u^*_{ij}$ by a neural net to satisfy $\|u_L-u^*\| \leq \epsilon_{\ell}$ as in \eqref{learning_error} of Theorem~\ref{Thm:lagros_stability} (see Fig.~\ref{lagrosdrawing})
\caption{LAG-ROS Algorithm}
\label{lagros_alg}
\end{algorithm}
\if0
\begin{remark}
Since \eqref{ncm_robust_control} is modeled by a neural network, using local observations in training also enables its decentralized implementation accounting for centralized global solutions~\cite{glas,NIPS2017_f22e4747}. This property will be demonstrated in Sec.~\ref{sec_simulation} to enable distributed implementation of $u_L$ in Theorem~\ref{Thm:lagros_stability}.
\end{remark}
\fi
\section{Simulation}
\label{sec_simulation}
Our proposed LAG-ROS framework is demonstrated using multiple motion planning problems under external disturbances ({\color{caltechgreen}\underline{\href{https://github.com/astrohiro/lagros}{https://github.com/astrohiro/lagros}}}). CVXPY~\cite{cvxpy} with the MOSEK solver~\cite{mosek} is used to solve optimization problems in Theorems~\ref{ncm_clf_thm}~and~\ref{tube_lemma} for sampling training data.
{\color{caltechgreen}
\subsection{Simulation Setup}
\HTnote{The maximum}{Added detailed explanations on simulation setup} admissible control time interval is selected to be $\Delta t_{\rm max} = 0.1$(s). The computational time of each framework is measured for the Macbook Pro laptop (2.2 GHz Intel Core i7, 16 GB 1600 MHz DDR3 RAM), and simulation results are the average of $50$ simulations for each random environment and disturbance realization.
\subsubsection{Neural Network Training}
\label{Sec:training}
We use a neural network $u_L$ with $3$ layers and $100$ neurons. The network is trained using stochastic gradient descent with training data sampled by Theorems~\ref{ncm_clf_thm}--\ref{Prop:sampling}, and the loss function is defined as $\|u_L-u^*\|$ to satisfy the learning error bound $\epsilon_{\ell}$ of Theorem~\ref{Thm:lagros_stability}. We use $50000$ training samples and select $\epsilon_{\ell} = 0.1$ for all the tasks, but these numbers can be modified accordingly depending on situations, considering the required performance for the error bound \eqref{learning_error} in Theorem~\ref{Thm:lagros_stability}. Note that we can guarantee \eqref{learning_error} only empirically, using the test error computed with a given test set as in standard neural network training.
\subsubsection{Performance Measure}
\label{Sec:performance}
We assume that the objective function of Theorem~\ref{tube_lemma} for sampling $(x_d,u_d)$ of \eqref{dynamicsd} is $\int_0^T\|\bar{u}\|^2dt$, where $T$ is a nominal time horizon. Since LAG-ROS is not for proposing a new trajectory optimization solver but for robustly tracking it with formal stability guarantees of Theorem~\ref{Thm:lagros_stability}, one could use any other objective functions as long as $(x_d,u_d)$ is obtainable.
We define success of each task as the situation where the agent reaches, avoiding collisions, if any, a given target terminal state $x_f$ within a given time horizon $\mathcal{T} \geq T$, \ie{} $\exists t^* \in [0,\mathcal{T}]$ \st{} $\|x(t^*)-x_f\| \leq \lim_{t\to\infty}r_{\ell}(t)$ for $r_{\ell}$ in \eqref{robust_learning_bound}, where the value for $r_{\ell}$ is to be defined in the subsequent sections. The success rate is computed as the percentage of successful trials in the total $50$ simulations. Also, we evaluate the performance of each motion planner, \ref{itemFF}, \ref{itemMP}, and \ref{itemLAGROS} in Sec.~\ref{Sec:lagros}, by the objective function $\int_0^{t^*}\|u\|^2dt$ if the task is successfully completed, and $\int_0^{T}\|u\|^2dt$ otherwise, where $T$ is the nominal time horizon.
\subsubsection{External Disturbances}
As shown in Lemma~\ref{Lemma:naive_learning} and Theorem~\ref{Thm:lagros_stability}, the tracking error bound of learning-based motion planners~\ref{itemFF} increases exponentially with time, whilst it decreases exponentially for the proposed approach~\ref{itemLAGROS}. To make such a difference clear, we consider the situations where $d(x,t)$ of \eqref{dynamics} is non-negligible, and thus~\ref{itemFF} tends to fail the task of Sec~\ref{Sec:performance} due to Lemma~\ref{Lemma:naive_learning}. This is not to argue that \ref{itemFF}~\cite{8593871,NIPS2017_766ebcd5,9001182,glas,NIPS2016_cc7e2b87,8578338,7995721} should be replaced by LAG-ROS~\ref{itemLAGROS}, but to imply that they can be improved further to have robustness and stability guarantees of Theorem~\ref{Thm:lagros_stability}.
\subsubsection{Computational Complexity}
\label{Sec:computation}
Since~\ref{itemFF} and LAG-ROS~\ref{itemLAGROS} are implementable only with one neural network evaluation at each time instant, their performance is also compared with~\ref{itemMP} which requires solving motion planning problems to get its control input. Its time horizon is selected to make the trajectory optimization~\cite{47710} solvable online considering the current computational power, for the sake of a fair comparison. We denote the computational time as $\Delta t$ in this section, and it should be less than the maximum control time interval $\Delta t_{\rm max}$, \ie{}, $\Delta t \leq \Delta t_{\rm max} = 0.1$(s).}
\subsection{Cart-Pole Balancing}
\label{Sec:cartpole}
We first consider the cart-pole balancing task in Fig.~\ref{lagros_cp_dist_fig} to demonstrate the differences of \ref{itemFF} -- \ref{itemLAGROS} summarized in Table~\ref{learning_summary}. Its dynamics is given in~\cite{ancm,6313077}, and we use $g = 9.8$, $m_c = 1.0$, $m = 0.1$, $\mu_c = 0.5$, $\mu_p = 0.002$, and $l = 0.5$.
\subsubsection{LAG-ROS Training}
$T$, $\mathcal{T}$, and $x_f$ in Sec.~\ref{Sec:performance} are selected as $T=\mathcal{T}=9$(s) and $x_f=[p_f,0,0,0]^{\top}$ for $x$ in Fig.~\ref{lagros_cp_dist_fig}, where $p_f$ is a random terminal position at each episode. We let $\mathcal{R}(0)=0$ and $\bar{d}_{\epsilon} = \bar{b}\epsilon_{\ell}+\bar{d} = 0.75$ in \eqref{robust_learning_bound}, and train the neural net of Sec.~\ref{Sec:training} to have $(\bar{d}_{\epsilon}/{\alpha})\sqrt{\chi} = 3.15$ with $\alpha = 0.60$ using Algorithm~\ref{lagros_alg}. The performance of LAG-ROS \ref{itemLAGROS} is compared with the learning-based planner~\ref{itemFF} and robust tube-based planner~\ref{itemMP}.
{\color{caltechgreen}
\subsubsection{Simulation Results and Discussions}
\HTnote{Note that}{Performed additional simulations and analysis to show that the bound on Theorem~1 is satisfied, and to demonstrate the benefits of LAG-ROS in Table~1} the robust tube-based motion planner~\ref{itemMP} (\ie{} $u^*$ of Theorem~\ref{Thm:lagros_stability}) is computable with $\Delta t \leq \Delta t_{\rm max}$ (see Sec.~\ref{Sec:computation}) in this case, and thus we expect that the performance of LAG-ROS~\ref{itemLAGROS} should be worse than that of~\ref{itemMP}, as~\ref{itemLAGROS} is a neural net model that approximates~\ref{itemMP}. The right-hand side of Fig.~\ref{lagros_cp_dist_fig} shows the tracking error $\|x-x_d\|$ of \eqref{dynamics} and \eqref{dynamicsd} averaged over $50$ simulations at each time instant $t$. It is still interesting to see that LAG-ROS of \ref{itemLAGROS} and $u^*$ of \ref{itemMP} indeed satisfies the exponential bound \eqref{robust_learning_bound} of Theorem~\ref{Thm:lagros_stability} given as
\begin{align}
r_{\ell}(t) = \mathcal{R}(0)\sqrt{\overline{\omega}}+(\bar{d}_{\epsilon}/{\alpha})\sqrt{\chi}(1-e^{-\alpha t}) = 3.15(1-e^{-0.60 t}) \nonumber
\end{align}
for all $t$ with a small standard deviation $\sigma$, unlike learning-based motion planner~\ref{itemFF} with a diverging bound~\eqref{naive_learning_error} and increasing deviation $\sigma$, as can be seen from Fig.~\ref{lagros_cp_dist_fig}. Contraction theory enables such quantitative analysis on robustness and stability of learning-based planners, which is one of the major advantages of our proposed technique.
\begin{figure}
\centerline{
\subfloat{\includegraphics[clip, width=25mm]{cartpole.pdf}}
\hfil
\subfloat{\includegraphics[clip, width=50mm]{cp_dist.pdf}}
}
\vspace{-0.9em}
\caption{{\color{caltechgreen}Cart-pole balancing task: $x=[p,\theta,\dot{p},\dot{\theta}]^{\top}$, $x$ and $x_d$ are given in \eqref{dynamics} and \eqref{dynamicsd}, \ref{itemFF} -- \ref{itemLAGROS} are given in Sec.~\ref{Sec:lagros}, and $r_{\ell}$ is given in \eqref{robust_learning_bound}. The shaded area denotes the standard deviation ($+1\sigma$ and $-0.5\sigma$).}}
\label{lagros_cp_dist_fig}
\end{figure}
\renewcommand{\arraystretch}{1.1}
\begin{table}
\caption{Control Performances for Cart-Pole Balancing ($\bar{d}_{\epsilon}=0.75$). \\
Note that the notations are as given in Sec.~\ref{Sec:performance}~and~\ref{Sec:computation}. \label{lagros_cp_tab}}
\vspace{-1.5em}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
& Success rate (\%) & $\int \|u(t)\|^2dt$ & $\Delta t$ (s) \\
\hline
\hline
\ref{itemFF} & $40.0$ & $4.79 \times 10^2$ & $3.01\times 10^{-2}$ \\
\ref{itemMP} & $100.0$ & $5.19 \times 10^2$ & $1.00\times 10^{-1}$ \\
\ref{itemLAGROS} & $100.0$ & $5.67 \times 10^2$ & $3.01\times 10^{-2}$ \\
\hline
\end{tabular}
\end{center}
\vspace{-2.0em}
\end{table}
\renewcommand{\arraystretch}{1.0}
Table~\ref{lagros_cp_tab} shows the control performance and computational cost of \ref{itemFF} -- \ref{itemLAGROS}, which is a good summary of their differences and trade-offs aforementioned in Sec.~\ref{Sec:lagros} and in Table~\ref{learning_summary}:
\begin{itemize}
\item \ref{itemFF} approximates $u_d$ of \eqref{dynamics}, and thus requires lower computational cost $\Delta t=0.03$(s) with a smaller objective value $\int \|u\|^2dt=479$, but robustness is not guaranteed (Lemma~\ref{Lemma:naive_learning}) resulting in a $40\%$ success rate.
\item \ref{itemMP} computes $u^*$ of Theorem~\ref{ncm_clf_thm}, and thus possesses robustness as in Fig.~\ref{lagros_cp_dist_fig} resulting in a $100\%$ success rate, but requires larger $\Delta t=0.1$(s) to compute $x_d$.
\item \ref{itemLAGROS} approximates $u^*$ independently of $x_d$, and thus possesses robustness of Theorem~\ref{Thm:lagros_stability} as in Fig.~\ref{lagros_cp_dist_fig} resulting in a $100\%$ success rate, even with $\Delta t = 0.03$(s) as small as that of \ref{itemMP}. It yields $9.2\%$ larger $\int \|u\|^2dt$ than \ref{itemMP} as it models $(x,o,t) \mapsto u^*$, not $(x,o,t)\mapsto u_d$.
\end{itemize}
It is demonstrated that LAG-ROS can utilize the robustness guarantees of $u^*$ as in Theorem~\ref{Thm:lagros_stability}, unlike~\ref{itemFF}, with significantly lower computational cost than that of~\ref{itemMP} as expected.}
\subsection{Multi-Agent Nonlinear Motion Planning}
\label{Sec:sc_planning}
The advantages of~\ref{itemLAGROS} demonstrated in Sec.~\ref{Sec:cartpole} are more appreciable in the problem settings where the robust tube-based motion planner~\ref{itemMP} is no longer capable of computing a global solution with $\Delta t \leq \Delta t_{\rm max}$ (see Sec.~\ref{Sec:computation}). We thus consider motion planning and collision avoidance of multiple robotic simulators~\cite{SCsimulator} in a cluttered environment with external disturbances, where the agents are supposed to perform tasks based only on local observations. Note that its nonlinear equation of motion is given in~\cite{SCsimulator} and all of its parameters are normalized to $1$.
\subsubsection{LAG-ROS Training}
We select $T$, $\mathcal{T}$, and $x_f$ in Sec.~\ref{Sec:performance} as $T=30$(s), $\mathcal{T}=45$(s), and $x_f=[p_{xf},p_{yf},0,0]^{\top}$, where $(p_{xf},p_{yf})$ is a random position in $(0,0) \leq (p_{xf},p_{yf}) \leq (5,5)$. {\color{caltechgreen}\HTnote{We}{Provided additional details of the simulation} let $\mathcal{R}(0)=0$, and train the neural network of Sec.~\ref{Sec:training} by Algorithm~\ref{lagros_alg} to get $(\bar{d}_{\epsilon}/{\alpha})\sqrt{\chi} = 0.125$ with $\alpha = 0.30$ in \eqref{robust_learning_bound}. In particular, the following bounded error tube \eqref{robust_learning_bound}:
\begin{align}
r_{\ell}(t) = (\bar{d}_{\epsilon}/{\alpha})\sqrt{\chi}(1-e^{-\alpha t}) = 0.125(1-e^{-0.30 t}) \nonumber
\end{align}
is used for sampling $(x_d,u_d)$ by Theorem~\ref{tube_lemma} with a input constraint $u_i\geq 0,~\forall i$ to avoid collisions with a random number of multiple circular obstacles ($0.5$m in radius) and of other agents, even under the learning error and disturbances. When training LAG-ROS, the input constraint is satisfied by using a ReLU function for the network output, and the localization technique~\cite{glas} is used to take $o$ of \eqref{learning_error} also as its network input for its distributed implementation, with the communication radius $2.0$m.} Its performance is compared with~\ref{itemFF},~\ref{itemMP}, and a centralized planner~\ref{itemC} which is not computable with $\Delta t \leq \Delta t_{\rm max}$, where $\Delta t$ is given in Sec.~\ref{Sec:computation}:
\begin{enumerate}[label={\color{caltechgreen}{(\alph*)}},start=4]
\item Centralized robust motion planner:\\
$(x,x_d,t) \mapsto u^*$, offline centralized solution of \ref{itemMP}. \label{itemC}
\end{enumerate}
{\color{caltechgreen}
\subsubsection{Remarks on Sub-optimal Trajectories}
Multi-agent problems have several sub-optimal solutions with optimal values close to the global optimum~\cite{8424034}, and thus~\ref{itemLAGROS} does not necessarily take the same $x_d$ as that of the centralized planner~\ref{itemC} when disturbances are acting on the system, as depicted in Fig.~\ref{lagros_sc_dist_fig}. However, it implicitly guarantees tracking to its own (sub-)optimal $x_d$ due to Theorem~\ref{Thm:lagros_stability}, which means we can still utilize their objective value $\int\|u\|^2dt$ and success rate to evaluate their performance.
\subsubsection{Implication of Simulation Results}
\label{Sec:implication}
\HTnote{Figure}{Re-performed simulations to demonstrate the robustness guarantees of Theorem~1 and to emphasize the benefits of LAG-ROS in Table~1}~\ref{lagros_sc_fig} shows one example of the trajectories of the motion planners \ref{itemFF} -- \ref{itemC}, under external disturbances with $\bar{d} = \sup_{x,t}\|d(x,t)\| = 0.4$.
\begin{itemize}
\item For~\ref{itemFF}, the tracking error accumulates exponentially with time due to $\bar{d}$ (Lemma~\ref{Lemma:naive_learning}), which necessitates the use of safety control for avoiding collisions~\cite{glas}. Such non-optimal control inputs also increase the error in \eqref{naive_learning_error} as \ref{itemFF}~does not possess any robustness guarantees.
\item For~\ref{itemMP}, robustness is guaranteed by Theorem~\ref{ncm_clf_thm} but can only obtain locally optimal $(x_d,u_d)$ of \eqref{dynamicsd}, as its time horizon has to be small enough to make the problem solvable within $\Delta t \leq \Delta t_{\rm max} = 0.1$(s), and the agent only has access to local information. This renders some agents stuck in local minima as depicted in Fig.~\ref{lagros_sc_fig}.
\item LAG-ROS \ref{itemLAGROS} tackles these two problems by providing formal robustness and stability guarantees of Theorems~\ref{Thm:lagros_stability} --~\ref{Prop:sampling}, whilst implicitly knowing the global solution (only from the local information as in~\cite{glas}) without computing it online. It satisfies the given state constraints due to Theorem~\ref{tube_lemma} as can be seen from Fig.~\ref{lagros_sc_fig}
\end{itemize}
\subsubsection{Simulation Results and Discussions}
Figure~\ref{lagros_sc_dist_fig} and Table~\ref{lagros_sc_tab} summarize the simulation results which corroborate the arguments of Sec.~\ref{Sec:implication} implied by Fig.~\ref{lagros_sc_fig}:
\begin{itemize}
\item \ref{itemFF} satisfies the requirement on the computational cost since we have $\Delta t \leq \Delta t_{\rm max} = 0.1$(s) as shown in Table~\ref{lagros_sc_tab}, but its success rate remains the lowest for all $\bar{d}=\sup_{x,t}\|d(x,t)\|$ due to the cumulative error \eqref{naive_learning_error} of Lemma~\ref{Lemma:naive_learning}, as can be seen in Fig.~\ref{lagros_sc_dist_fig} and Table~\ref{lagros_sc_tab}. Although its objective value is the smallest for $\bar{d} \leq 0.6$ since it models $u_d$, it gets larger for larger $\bar{d}$, due to the lack of robustness to keep $x$ around $x_d$.
\item \ref{itemMP}~has a success rate higher than that of~\ref{itemFF}, but still lower than $50$\% as it can only compute sub-optimal $x_d$ under the limited computational capacity. In fact, we have $\Delta t \geq \Delta t_{\rm max}$ in this case, which means it requires a slightly better onboard computer. Also, it uses excessive control effort larger than $10^3$ due to such sub-optimality.
\item \ref{itemLAGROS} achieves more than $90$\% success rates for all $\bar{d}$, and its objective value remains only $1.64$ times larger than that of the centralized planner \ref{itemC}, even without computing $(x_d,u_d)$ of \eqref{dynamicsd} online. Its computational cost is as low as that of \ref{itemFF}, satisfying $\Delta t \leq \Delta t_{\rm max} = 0.1$(s) while retaining a standard deviation smaller than~\ref{itemFF}.
\end{itemize}
These results imply that LAG-ROS indeed enhances learning-based motion planners~\ref{itemFF} with robustness and stability guarantees of contraction theory as in \ref{itemC} (and \ref{itemMP}), thereby bridging the technical gap between them.}
\renewcommand{\arraystretch}{1.2}
\begin{table}
\caption{Computational time of each motion planner for Multi-agent Nonlinear Motion Planning.\label{lagros_sc_tab}}
\vspace{-1.5em}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
& \ref{itemFF} & \ref{itemMP} & \ref{itemLAGROS} & \ref{itemC} \\
\hline
\hline
$\Delta t$ (s) & $4.63\times 10^{-2}$ & $1.53\times 10^{-1}$ & $4.66\times 10^{-2}$ & $1.12\times 10^{3}$ \\
\hline
\end{tabular}
\end{center}
\vspace{-2.0em}
\end{table}
\renewcommand{\arraystretch}{1.0}
\begin{figure*}
\centering
\includegraphics[width=120mm]{sc_obs.pdf}
\vspace{-1.0em}
\caption{{\color{caltechgreen}Trajectories for the learning-based planner~\ref{itemFF}, robust tube-based planner~\ref{itemMP}, LAG-ROS~\ref{itemLAGROS}, and offline centralized solution \ref{itemC} ($\circ$: start, $\bullet$: goal).}}
\label{lagros_sc_fig}
\vspace{-1.3em}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=80mm]{sc_obs_dist.pdf}
\vspace{-0.9em}
\caption{{\color{caltechgreen}Control performances versus $\sup_{x,t}\|d(x,t)\|$ of \eqref{dynamics} for multi-agent nonlinear motion planning. Note that \ref{itemFF} -- \ref{itemC} are as given in Fig.~\ref{lagros_sc_fig}, and the performances are as defined in Sec.~\ref{Sec:performance}. The shaded area denotes the standard deviation ($\pm10^{-1}\sigma$).}}
\label{lagros_sc_dist_fig}
\vspace{-1.5em}
\end{figure}
\if0
\renewcommand{\arraystretch}{1.1}
\begin{table}
\caption{Control Performances for Multi-agent Nonlinear Motion Planning (averaged over $50$ simulations with $\sup\|d\|=0.8$). \\
Note that computational time should be less than the control time interval $\Delta t = 0.1$(s). \label{lagros_sc_tab}}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
& Success rate (\%) & $\int\|u\|^2dt$) & $\Delta t$ (s) \\
\hline
\hline
\ref{itemFF} & $2.5641$ & $2.1787 \times 10^3$ & $8.9623\times 10^{-2}$ \\
\ref{itemMP} & $4.2628\times 10$ & $7.6912 \times 10^3$ & $2.2400\times 10^{-1}$ \\
\ref{itemLAGROS} & $9.0064\times 10$ & $2.1971 \times 10^3$ & $9.5559\times 10^{-2}$ \\
\ref{itemC} & $1.0000\times 10^2$ & $1.7910 \times 10^3$ & $1.2064\times 10^{3}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\renewcommand{\arraystretch}{1.0}
\begin{figure*}
\centerline{
\subfloat{\includegraphics[clip, width=110mm]{sc_obs.pdf}}
\hfil
\subfloat{\includegraphics[clip, width=65mm]{sc_obs_dist.pdf}}
}
\caption{Control performances versus $\sup_{x,t}\|d(x,t)\|$ for multi-agent nonlinear motion planning (averaged over $50$ simulations). \ref{itemFF} -- \ref{itemC} are as in Table~\ref{lagros_sc_tab}.}
\label{lagros_sc_dist_fig}
\end{figure*}
\subsection{Spacecraft Reconfiguration in Low Earth Orbit (LEO)}
We finally consider the problem of spacecraft reconfiguration in LEO as an example of systems equipped with a known Lyapunov function. The dynamical system is presented in~\cite{doi:10.2514/1.55705} accounting for $J_2$ and atmospheric drag perturbations. Since it can be expressed as a fully-actuated Lagrangian system~\cite{doi:10.2514/1.G000218}, we can design a Lyapunov function by its inertia matrix~\cite{Slotine:1228283} and \textcolor{red}{use it for LAG-ROS construction of Proposition~\ref{lagros_prop} due to Theorem~\ref{lag_clf_thm}.} The spacecraft communication radius for local observations is selected as $2.0$km, which prevents the use of global motion planners. The agents are expected to perform reconfiguration within $75$s under bounded disturbances ($\sup\|d\|=1.0$).
Table~\ref{lagros_leo_tab} shows the simulation results with external disturbances given in Fig.~\ref{lagros_leo_fig} for the robust tube-based motion planner, learning-based motion planner, LAG-ROS, and global motion planner~\cite{doi:10.2514/1.G000218}. LAG-ROS achieves only a $5.9$\% increase in the objective value even with the computational time less than $\Delta t = 0.1$(s). Again, a slightly better onboard computer is required for the robust tube-based planner as its computational time $1.5357\times 10^{-1}$s is greater than $\Delta t = 0.1$(s). Since LAG-ROS models the global robust optimal solution, it also attains more than a $28.4$\% decrease in the objective value when compared with the robust motion planner, even though the success rate of learning-based motion planning is only $2.6$\% due to external disturbances. These results indicate that LAG-ROS can be applied with known Lyapunov functions constructed utilizing a special structure of underlying dynamical systems~\cite{Slotine:1228283,ccm} for the sake of a learning-based robustness guarantee. These trends are the same for other sizes of disturbances as can be seen in Fig.~\ref{lagros_sc_dist_fig}: LAG-ROS keeps more than $98$\% success rate for all the disturbances when averaged over $50$ simulations, with $\mathcal{L}$2 control effort smaller than $2.4$ times that of the global solution, unlike the learning-based and robust tube-based motion planner.
\renewcommand{\arraystretch}{1.1}
\begin{table}
\caption{Control Performances for Multi-agent Spacecraft Reconfiguration in LEO (averaged over $50$ simulations with $\sup\|d\|=1.0$). \\
Note that computational time should be less than the control time interval $\Delta t = 0.1$(s). \label{lagros_leo_tab}}
\begin{center}
\begin{tabular}{|l|c|c|c|c|}
\hline
& Success rate (\%) & $\int\|u\|^2dt$ & $\Delta t$ (s) \\
\hline
\hline
\ref{itemFF} & $3.0000$ & $1.4636\times 10^2$ & $7.6204\times 10^{-2}$ \\
\ref{itemMP} & $1.0000\times 10^2$ & $2.1109 \times 10^2$ & $1.5357\times 10^{-1}$ \\
\ref{itemLAGROS} & $1.0000\times 10^2$ & $1.5119 \times 10^2$ & $8.3024\times 10^{-2}$ \\
\ref{itemC} & $1.0000\times 10^2$ & $1.4278 \times 10^2$ & $7.4810\times 10^{2}$ \\
\hline
\end{tabular}
\end{center}
\end{table}
\renewcommand{\arraystretch}{1.0}
\begin{figure*}
\centerline{
\subfloat{\includegraphics[clip, width=110mm]{leo.pdf}}
\hfil
\subfloat{\includegraphics[clip, width=70mm]{leo_dist.pdf}}
}
\caption{Spacecraft reconfiguration in LEO with baselines (a)~\cite{glas} \& (b)~\cite{tube_nmp} and for (c) LAG-ROS, where (d) shows the solution computed offline assuming global environment observations. LAG-ROS achieves the highest success rate with the control effort closest to that of the global solution.; Control performances versus $\sup_{x,t}\|d(x,t)\|$ for spacecraft reconfiguration in LEO (averaged over $50$ simulations). (a) -- (d)~are as given in Table~\ref{lagros_leo_tab}.}
\label{lagros_leo_dist_fig}
\end{figure*}
\fi
\section{Conclusion}
\label{sec_conclusion}
In this work, we propose a new learning-based motion planning framework, called LAG-ROS, with formal robustness and stability guarantees. It extensively utilizes contraction theory to provide an explicit exponential bound on the distance between the target and controlled trajectories, even under the existence of the learning error and external disturbances. Simulation results demonstrate that it indeed satisfies the bound in practice, thereby yielding consistently high success rates and control performances in contrast to the existing motion planners~\ref{itemFF} and~\ref{itemMP}. Note that other types of disturbances can be handled, using~\cite{nscm} for stochastic systems and~\cite{ancm} for parametric uncertain systems.
\subsubsection*{Acknowledgments}
This work was in part funded by the Jet Propulsion Laboratory, California Institute of Technology, and benefited from discussions with J. Castillo-Rogez, M. D. Ingham, and J.-J. E. Slotine. |
\section{Introduction}
Performing data analysis over large graphs with billions of edges is an important yet challenging task. Previous work has estimated the average distance in the graph between any two users of large social networking sites such as Facebook, resulting in 4.75 hops on average~\cite{boldi2012four}. Such insights on the structure of a graph are often difficult to investigate for the research community, as it is neither possible to obtain all the neighbourhood sets of each node in the social network due to the terms and conditions for users of the social network API, nor practical to process a complete graph of billions of nodes in terms of performance on a laptop. A practical solution is the emerging use of the graph estimation techniques that approximate the value of graph properties at the node level to reduce computation overhead.
In the literature, estimation algorithms propose to calculate neighbourhood information efficiently by scaling distributed computing to approximate massive graph data mining~\cite{palmer2002anf}. Boldi \& Vigna focus on approximating geometric centrality network metrics at scale using an algorithm based on HyperLogLog (HLL) counters and their HyperANF algorithm~\cite{boldi2011hyperanf}. This allows approximation of geometric centrality metrics at a very high speed and acceptable accuracy for certain use cases that do not require exact values.
\paragraph{Contribution}
In this work we firstly explore the utility of approximate neighbourhood information of a Decentralised Online Social Network (DOSN) with millions of edges. It is the first study to date with a DOSN dataset using estimation functions. We estimate the neighbourhood function of graph over the Mastodon DOSN data provided by~\cite{zignani2018follow}, namely the average shortest path length, with and without the use of approximated estimations. For the latter, we implement and employ the HLL algorithm using Python to provide the HLL counters that are used in HyperBall. We refer the reader to the HyperBall algorithm in Appendix~\ref{appendix:hb} for a more detailed explanation of the work from~\cite{boldi2013core}.
Firstly, regarding reproducibility of these results we provide our code in Python, which also implements the original algorithm for HyperBall~\cite{zignani2018follow} and there we apply our neighbourhood function. In all cases, we adjust the error bounds of the estimation technique using HLL~\cite{flajolet2008hyperloglog} counters. The expectation for our benchmark here is to see how a smaller graph, in the order of millions instead of billions, behaves in terms of performance first. Our results show error bounds on the order of 20-30\% in the most pessimistic cases of the network graph traversal (with exception of the level 0 of the traversal where only 1 node exists and thus error is always plausible because of the problem of small cardinalities when using HyperLoLog), while less error occurs otherwise in line with the theoretical foundations in the data approximation technique of HLL. Performance gains in large graphs are evident in Table~\ref{tab:performance}.
Secondly, our small-world results are according to previous work in the area, in which a random graph (R) should have a lesser average shortest path length than a social network graph (G) with community structure. The size of the community should balance the value of such network metrics among these two kind of graphs when the community size grows so that they become nearly the same~\cite{pattanayak2020lengthening} as we observe in Table~\ref{tab:small-world-performance}. We observe that even to a point in the Mastodon dataset, the R/G proportion is actually higher for the random graph. This shows that in the Mastodon dataset, the federated community structure is vastly growing to a size that actually interconnects communities themselves quite easily to reduce average shortest path length. From that second table of results, we confirm that the Mastodon DOSN shows the highest proportion among all the graphs probed.
\section{Methodology}
Estimation using sketching techniques such as HLL~\cite{flajolet2008hyperloglog} counters are a tool for cardinality estimation of large multisets. In short they can efficiently count the number of different elements in a stream using a single pass over it. A HLL counter uses a hash function producing $b$ bits that maps each element uniformly to one of $2^b$ possibilities. For any produced hash value, say $h$, we use $\text{Prefix}^t(h)$ to denote the first $t$ bits of h and $\text{Rest}^t(h)$ to denote the remaining bits after the first $t$ bits are removed. Further for any bit-string $x$, let $\text{LZ}(x)$ be the number of leading zeros in $x$. With these definitions in hand, we now present the functions for manipulating HLL counters in Algorithm~\ref{alg:hll}. In addition to functions that add an item to the count and extract the current value of the counters, HLL counters allow us to compute lossless cardinality of unions: the intuition is that between two compatible HLL counters, we can just take the maximum counter to obtain the lossless union of two counters, equivalent to counting the union of the two multi-sets at each of the two counters and thus with no added error. This is however not true for taking the minimum for intersections so we use another technique, namely MinHashes, to approach this challenge and obtain good values as well as shown later here.
\begin{algorithm}
\caption{HyperLogLog Counter Manipulation}
\label{alg:hll}
\begin{algorithmic}[1]
\State $h:D \rightarrow 2^b$, a hash function from the domain of items
\State $M$ an array of $m=2^t$ counters each initialised to $-\infty$
\State $\alpha_m$ is a constant that depends on the number of counters
\Function{AddItem}{$M$:counter, $x$:item}
\State $i \gets \text{Prefix}^t(h(x))$
\State $M[i] \gets \text{Max}\{M[i], \text{LZ}(\text{Rest}^t(h(x)))\}$
\EndFunction
\Function{GetCount}{$M$:counter}
\State $Z \gets \sum\limits_{i=0}^{m-1}2^{-M[i]}$
\State $E \gets \alpha_mm^2Z$
\If{$E \leq \frac{5}{2}m$}\label{corr_begin}
\State Let V be number of registers equal to 0
\If{$V \neq 0$}
\State $E \gets m\text{log}\frac{m}{V}$ \Comment{Small range corrections}
\EndIf
\EndIf\label{corr_end}
\State \textbf{return} $E$
\EndFunction
\Function{Union}{$M$:counter, $N$:counter}
\State x:hll\_counter
\For{i:=0 to m-1}
\State $x[i] \gets \text{Max}\{M[i], N[i]\}$
\EndFor
\State \textbf{Return} x
\EndFunction
\end{algorithmic}
\end{algorithm}
Most crucially for this paper, HLL counters provide tight statistical bounds on the measured cardinality. In the limit $n \rightarrow \infty$, the returned count (say $\hat{n}$) is an almost unbiased estimator~\cite{flajolet2008hyperloglog} of the true count (say $n$) with a relative standard deviation $\frac{\sigma}{n} \leq \frac{1.06}{\sqrt{m}}$. Increasing the number of counters therefore decreases the uncertainty in the measurement at the expense of space. Extending the example above, using 8 bits per counter and 128 counters means we can count items from a universe of $2^{256}$ distinct items with a relative standard deviation under $9.3\%$. The low space overheads and tight statistical bounds are the key features of HLL counters that we wish to exploit in this paper. HLL counters however have one flaw, the bounds on their accuracy only apply at high enough counts. One way to tackle this is to introduce small range correction~\cite{flajolet2008hyperloglog}: if any of the counters are found to be zero then the algorithm returns $m\text{log}(\frac{m}{V})$, where $V$ is the number of zero registers (lines \ref{corr_begin}--\ref{corr_end} of Algorithm~\ref{alg:hll}). Existing work on making HLL counters sound for small counts~\cite{hll_engineer} have recommended using a bias correction instead, and as in their appendix work we implement their recommended bias correction algorithm instead of linear counting at small ranges. We use the same bias correction at low counts for all datasets here, see equation~\ref{eq:precision} below.
\begin{equation}\label{eq:precision}
\overbrace{101\ldots 011}^{p\ \rm{bits}}\overbrace{001\ldots 110}^{64
- p\ \rm{bits}}
.
\end{equation}
\section{Results for Neighbourhood Estimation function}
%
The Mastodon DOSN dataset contains about 6.5 million direct edge relations and more than 566520 nodes. In our HyperBall implementation~\cite{boldi2013core} we choose the precision of 4 p-bits for HLL. The traversal of each dataset graph uses a depth as parameter to run with no less than 5 as to resemble the degrees of separation in a real social network graph (e.g., Facebook). We test our HyperBall implementation with two centralised social network datasets (Twitter, Facebook) from~\cite{mcauley2012learning}, and one decentralised social network dataset (Mastodon) for the very first time using a neighbourhood Estimation function.
\paragraph{Performance}
In Table~\ref{tab:performance} we observe that all of the datasets have naturally worse performance with a brute force approach. The results that use the Hyperball as expected show a considerable speedup in calculation of geometric Hyperballs in the order of minutes, compared to hours when using a brute force Breadth First Search (BFS) approach in a commodity Macbook Pro OS X Mojave with 4 cores CPU, 16GB of memory and 500GB SATA disk installed. In addition, we find a embarrassingly parallel number of tasks in the initialisation and computation of the counters for HyperBall, which provides a further speedup opportunity using the \emph{joblib} library as backend using threading with four embarrassingly parallel initialization tasks of the algorithm. This cuts the computation time of the Mastodon graph's HyperBall by half. Overall the theoretical probabilistic guarantees of the HyperBall seem to reduce times further for larger graphs as Mastodon, when approximations are based on the HLL counters.
\begin{table}
\centering
\caption{HyperBall computation times in (hh:mins:secs)}
\begin{adjustbox}{width=0.9\columnwidth}
\begin{tabular}{|c|c|c|c|c|c|}
\cline{4-6}
\multicolumn{3}{c}{}
&\multicolumn{1}{|c|}{Bfs} &\multicolumn{2}{|c|}{HyperBall} \\ \hline
& \multicolumn{1}{c|}{\textbf{Nodes}} &
\multicolumn{1}{c|}{\textbf{Edges}} &
\multicolumn{1}{c|}{\textbf{Sequential}} &
\multicolumn{1}{c|}{\textbf{Sequential}} &
\multicolumn{1}{c|}{\textbf{Parallel}}
\\
\hline
Twitter & 81306 & 2420766 & \ensuremath > 1:00:00.00 & 0:02:54.874483 & 0:02:46.414789\\
Facebook & 4039 & 88234 & 0:00:56.965906 & 0:00:08.249678 & 0:00:08.533745\\
Mastodon & 566520 & 6493563 & \ensuremath > 1:00:00.00 & 0:11:16.773455 & \textbf{0:05:29.926316} \\
\hline
\end{tabular}
\end{adjustbox}
\label{tab:performance}
\end{table}
\paragraph{Network metrics}
Armed with our implementation of the HyperBall algorithm we also study the \emph{path length probability distribution} and the \emph{average path length} over the same networks datasets. For the latter metric we apply our implementation of the approximated algorithm described in HyperBall. This allows us to compute the so called \emph{small world coefficient} for each of the networks we benchmark.
\begin{algorithm}
\caption{Average path length with HyperBall}
\label{alg:ppanf}
\begin{algorithmic}[1]
\State nr_paths: number of paths of length per node.
\State max\_t: maximum distance of HyperBall computations is equal to $b$.
\Function{HyperBall}{$G$:graph, $b$:radius of ball, $p$:hll\_c\_prec}
\State \textbf{return} $HB$
\EndFunction
\Function{NumNodesDistFrom}{$v$, $t$}
\If{t = 0}
\State \textbf{return} 1
\ElsIf{t >= get_max_t}
\State \textbf{return} 0
\Else{}
\State \textbf{return }{balls[v][t].size() - self.balls[v][t - 1].size()}
\EndIf
\EndFunction
\Function{average\_path\_length}{$G$}
\For{v $\in G$}
\For{$t \in$ 1..{max\_t}}
\State nr_paths[v] = $HB$.\textproc{NumNodesDistFrom}({$v$, $t$})
\EndFor
\EndFor
\EndFunction
\end{algorithmic}
\end{algorithm}
\paragraph{Small World}
\begin{table*}
\centering
\caption{Time difference for Algorithm~\ref{alg:ppanf} with NetworkX vs HyperBall computing vs approximating three network metrics.}
\begin{adjustbox}{width=0.9\textwidth}
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\cline{4-8}
\multicolumn{3}{c}{}
&\multicolumn{1}{|c}{NetworkX}
&\multicolumn{4}{|c|}{HyperBall} \\
\hline
& \multicolumn{1}{c|}{\textbf{Nodes}} &
\multicolumn{1}{c|}{\textbf{Edges}} &
\multicolumn{1}{c|}{\textbf{hh:mins:secs}} &
\multicolumn{1}{c|}{\textbf{(hh:mins:secs)}} &
\multicolumn{1}{c|}{\textbf{Avg. Shortest Path Length (R/G)}} &
\multicolumn{1}{c|}{\textbf{Clustering Coeff.(G/L)}} &
\multicolumn{1}{c|}{\textbf{Small World Coeff.(l - c)}}
\\
\hline
Twitter & 81306 & 2420766 & 0:03:21.698368 & 0:03:34.580287 & 0.9072928958767731 & 0.565311468612065 & 0.3415490228344126\\
Facebook & 4039 & 88234 & 0:00:12.997504 & \textbf{0:00:09.321404} & 0.8024902838225895 & 1.0171284634760704 & -0.3015614725029204 \\
Mastodon & 566520 & 6493563 & 0:15:28.511020 & 0:19:39.377784 & 1.11599944828774786 & $\approx{0.0}$ & 1.1159994828774786 \\
\hline
\end{tabular}
\end{adjustbox}
\label{tab:small-world-performance}
\end{table*}
From the computations of Table~\ref{tab:small-world-performance} we obtain the corresponding \emph{small world coefficient} by (i) computing a random graph with equivalent amount of nodes and edges to our input graph, and (ii) computing a lattice graph with same amount of vertices too. Once we have those values computed, now we can obtain the ratio $l$ of average path lengths among a random graph $R$ compared to our actual input graph $G$. If we then also compare the ratio of the average clustering coefficient among the input graph $G$ and the previously computed lattice graph $L$, we obtain the clustering coefficient of $G/L$ called $c$. We calculate the \emph{small world coefficients} as $l - c$. Usually the coefficient should be between -1 and 1 for indicating how strong the small world phenomenon occurs. A coefficient close to 0 indicates a strong influence of small world, negative values indicate regularity/lattice-like graphs, and positive values indicate a graph with more random characteristics. Our result indicates that the Mastodon dataset is of the latter sort of graph according to our results for both raw and approximated values, and due to the random graph effect explained in~\cite{evans2004complex} the clustering coefficient is close or zero indeed for a same random graph of the same number of edges and vertices and even in~\cite{zignani2018follow} it is said that clustering is $\approx{0.17}$ for large in degree nodes and larger for less degree nodes. Also, a fast growing community structure in such a federated social network is plausible and in line with a related preliminary measurements~\cite{complexnets2020}.
We observe that performance of the network properties in the Mastodon dataset improves again somehow. From the standpoint of Social Network Analysis, small-world networks have short average path length and high clustering coefficient (Facebook in Table~\ref{tab:small-world-performance}). On the other hand, an average shortest path length, shorter indicates that information propagates more easily than in a random or regular lattice-type social network as Twitter. For Mastodon this can be related to the relatively self-organising nature of the Mastodon platform among users. For Twitter we would expect a higher average path length as expressed in previous works in the matter~\cite{myers2014information}. However, their dataset at Twitter is more up to data that the one we use from an ego network of 2012 in~\cite{mcauley2012learning}.
Note that the times in Table~\ref{tab:small-world-performance} indicate the average path length computation, which is the main source of overhead for the small world coefficient we provide. Some minor shifts exist over repetition of our runs in Twitter and the Mastodon. We obtain performance times with a parallel initialisation of HLL counters for the HyperBall because it yields a more realistic result than initialising them so sequentially. In the benchmark, for the Mastodon dataset we list in the previous table we compare our approach using HyperANF for getting the path length against state-of-the-art library NetworkX (nx.average\_shortest\_path\_length(G)) that is expected to be more optimised than our prototype. Surprisingly, our approach still performs quite similar or better in some cases for the calculations listed in Table~\ref{alg:ppanf}.
The privacy-preserving version of the average path length above in Algorithm~\ref{alg:ppanf} eventually only needs to learn, as before, the absolute value $|B(v,t)|$ not $B(v,t)$ itself, as $|B(v,t)|$ is the \# of different elements in the summary connected by one edge to the vertex $v$ in question $B(v,t-1)$. Inside HyperBall this is done by first keeping a HLL counter $H(v)$ at each node $v \in V$. So $iff$ $B(v,t) = B(v, t-1)$ holds, then for each $(v,u) \in E, B(v,t) = B(v,t) \cup B(u,t-1)$.
\section{Towards Privacy-Preserving Community Detection}
Equipped with theoretical vision and empirical results now, we envision a future privacy preserving version of HyperBall for decentralised social network deployments in which network properties require to be exchanged using these set cardinalities or summary sets over insecure channels (encryption suffices) to perform federated learning tasks for instance.
A key operation in community detection is inferring strong ties inside communities by calculating the representative neighbourhood among nodes in a large graph to show the community they form. To provide a timely representation of a community and useful input to other applications as recommendation systems, this can either be done with exact state-of-the-art methods as Louvain~\cite{Louvain} that compare the density of edges in and out modularity, the Hyperball for community detection in~\cite{TriangleCountHyperBall} that estimates the size of those neighbourhoods, counts triangles (note our algorithm differs from theirs also for triangles not shown here). However, we additionally plan to intersect the HyperBalls from several communities whereas one can also potentially decide where to set the ``cut'' based~\cite{schaeffer2007graph} conductance, which yields a ratio of the number of edges connected to nodes outside the cluster in relation to the ones connected to inside nodes. Thus, measuring how well or strong is a community structure. Note this does not consider nor mention anything about privacy of the methods.
However, the HyperBall has a shortcoming as the underlying HLL that it uses can not compute intersections by itself on the fly with the neighbouring summaries or fingerprints alone. Fortunately we can use state-of-the-art estimation techniques based on~\cite{broder1998min} for intersecting two HyperBall. Our approach obtains that by computing the product of the \emph{HLL} with an approximation of the Jaccard similarity using MinHashes. Because MinHashes is approximately the intersection of the Jaccard of the two HyperBall divided by a fixed parameter $k$ (see equation~\ref{eq:minhash}), we are left with just an approximated value of the intersection among two sets in equation~\ref{eq:hll}.
Rather than dividing by the union which would be more costly, for using the MinHashes in the first equation, the result can be approximated as the Jaccard coefficient, which effectively requires just the intersection divided by the MinHashes parameter $k$. Also, this provides desirable privacy properties due to the way the counters and such summary set or fingerprint is built with HLL -- using just an encoding of inputs. It is easy to see from the decomposition below in formulae from equation~\cref{eq:cancelout,eq:minhash,eq:hll} that we can approximately cancel out unions from with the following multiplication of HLL and MinHashes:
\begin{equation}\label{eq:cancelout}
\frac{|\cap|}{\cancel{|\cup|}} \cdot {\cancel{|\cup|}}
\end{equation}
\begin{equation}\label{eq:minhash}
\frac{\left|h_k(A_i) \cap h_k(B_i)\right|}{k}
\end{equation}
\begin{equation}\label{eq:hll}
\left|\bigcap A_i \right| = J(A_1,\ldots,A_n)\cdot\left|\bigcup A_i\right| \approx
\rm{MinHash} \cdot \rm{HLL}
\end{equation}
Meanwhile, the mentioned MinHashes $H$, if $H'(v) = H(v)$ also holds, then for each $(v,u) \in E$, $H'(v)$ = $(H'(v) \cap H(u))$. And surprisingly, this can be done while reading edges in sequence.
In summary, the data structure is as follows: (i) 16384, 8 bit entries for the HLL counter. Each entry keeps track of the longest 0 sequence that has been produced for it. From this, we can easily estimate the size of the union of the 2 sets; (ii) 16384, 64 bit hash values in the MinHash table. For each element in the set we generate 16384 hash values with 16384 hash functions and store the smallest ones. Based on how many of these match comparing 2 sets we can estimate the Jaccard coefficient.
The benefit of our data structure is that it enables fast union with HLL and fast intersection with MinHashes. A disadvantage is that a large dataset may take a long time to generate so many hash values as it is the case in the Mastodon dataset.However, offline pre-computation of the HLL counters is a practical approach.
\section{Conclusion}
We use approximated computing techniques using summary sets (HyperBall, Minhashes) that can be applied to estimate properties in graphs resulting in algorithms that require only a bounded amount of data, e.g., just a HLL set per vertex rather than a list of all reachable vertices. Summary sets appear to have anonymity properties e.g., hard to deduce actually reachable neighbours given a HLL encoding of neighbours that are reachable in K steps (like in HyperBall) and adding the envisioned signed privacy guarantees in a future work with standard but wisely employed pairing based cryptography~\cite{BLS}. Therefore, moving towards our hypothesis that they HyperBall algorithm is a good candidate for privacy-preserving protocols over decentralised social networks.
Firstly, we may apply our proposal to social recommendation systems or abuse detection in an anonymous social network, where only summary sets are available for friend lists or detection algorithms locally at each node. Therefore, some key features will be the intersection of such lists of friends to make recommendations or detect unsolicited content, both in zero-knowledge to attaining a secure two party computation by applying existing protocols we have for cardinality estimation and possibly doing so even in a (de)centralised manner.
Secondly, we have explored the feasibility of estimating well-known measures of closeness as the \emph{path length probability distribution} and the \emph{average path length}, the latter using HyperBall for performance. Later we may also consider the variance-to-mean ratio of the shortest-paths distribution in the Mastodon dataset. These metrics are important in network geometry as explained in~\cite{boldi2011hyperanf} and serve as indicator for instance to distinguish structural differences among a social network from a web graph. To the best of our knowledge we are the first to compute them over a lists of ego networks and a decentralised network using approximated or sketching techniques that encode privacy properties.
\paragraph{Limitations}
A limitation we need to analyse is how the summary sets from the union of HLL counters at each vertex using the candidate set of Algorithm~\ref{alg:ppanf} are vulnerable they are to attacks from an honest but curious participant in the protocols/systems we will develop as said. However if estimated locally these computations are secure, and thus only the broadcasting of the resulting local values to calculate a global aggregated \emph{average path length} need to be protected for privacy and security reasons. Effectively, our network model from here will assume a decentralised computation of such network metrics in future DOSN deployments as Mastodon.
\paragraph{Future work}
In upcoming work we will show that it is possible to perform efficient community detection with estimation algorithms that use a representation of the neighbouring nodes when computing an intersections/union in a federated manner with privacy. Indeed we would envision the integration of this new work with our former~\cite{garcia2016privacy} in order to intersect fingerprints of the HyperBall in zero-knowledge under standard techniques for exchanging insights of a neighbourhood estimation function across nodes and use it to propagate trust.
|
\section{Introduction}
Efficient and accurate simulation of SDEs is important in many applications such as Monte Carlo sampling, data assimilation and predictive modeling (see e.g.,\cite{KP99,Maday2002,weinan2007heterogeneous,KMK03,legoll2010effective,CLMMT16}). In particular, ergodic systems often demand efficient and accurate simulation of both short-time dynamics and large-time statistics. Explicit schemes, while efficient and accurate for short-time, tend to miss the invariant measure in large-time simulations because of the accumulation of numerical error. In particular, for locally Lipschitz SDEs, they tend to be numerical unstable and may miss the invariant measure even for the small time-step (for example, the Euler-Maruyama scheme, because it destroys the Lyapunov structure \cite{roberts1996exponential,MSH02}) and require special treatments such as taming scheme under small time-step size \cite{Kloeden2012,Jentzen2015}.
Implicit schemes, on the other hand, are numerically stable and can accurately simulate the invariant measure when the time-step is small. However, they are computationally inefficient due to the limited time-step size and the costly implicit step.
We introduce ISALT, inference-based schemes adaptive to a large time-stepping statistical learning framework to construct schemes with large time-steps from data. When the data are generated from an implicit scheme, ISALT combines the advantages of both explicit and implicit schemes: it is as fast as an explicit scheme and is accurate as an implicit scheme in producing the invariant measure. The inference is done once for all and the inferred scheme can be used for general purpose simulations, either long trajectories or ensembles of trajectories with different initial distributions.
More specifically, we consider the large time-step approximation of the ergodic SDE with additive noise
\begin{equation}\label{eq:sde}
d\mX_t = f(\mX_t)dt + \sigma d{\bf B}_t; \,
\end{equation}
where the drift $f:\mathbb{R}^d\to\mathbb{R}^d$ is local-Lipschitz. Here ${\bf B}$ is a standard m-dimensional Brownian motion with $m\leq d$, the diffusion matrix $\sigma\in \mathbb{R}^{d\times m}$ has linearly independent columns, and they represent a degenerate noise when $m<d$. Our goal is to design an explicit scheme with large time-stepping so that it can efficiently and accurately simulate both short-time dynamics
and large-time statistics such as invariant measures.
We infer such explicit schemes with large time-stepping from offline data generated by an implicit scheme. Figure \ref{fig:schematic} shows the schematic plot of the procedure. The essential task is to approximate the infinite-dimensional discrete-time flow map. A major difficulty in a statistical learning approach is the curse of dimensionality (COD) when using generic basis functions. Our key contribution is to approximate the flow map by parametrization of numerical schemes, which provides informed basis functions, thus avoiding the COD by harnessing the rich information and structure in classical numerical schemes. We also introduce a scalable algorithm to compute the maximal likelihood estimator by least squares, which converges and is asymptotic normal as the data size increases (see Theorem \ref{thm_convEst}). Furthermore, we show that the inferred scheme, when it is a parametrization of an explicit scheme, has 1-step strong order as the explicit scheme.
\begin{figure}[htp!]
\centering \vspace{-1mm}
\includegraphics[width=0.7\textwidth]{./figures/scheme.pdf}
\vspace{-3mm} \caption{Schematic plot of inferring explicit scheme with a large time-step.}
\label{fig:schematic}
\end{figure}
In this study, we focus on learning approximate flow maps that use only the increments of the Brownian motion each time interval (that is, the function $F^\delta}%{\Delta t(X_{t_n}, \Delta B_{t_n})$ in Figure \ref{fig:schematic}). We explore the derivation of informed-basis functions from three types of classical numerical schemes: the Euler-Maruyama (EM) \cite{KP99}, a hybrid RK4 (fourth-order Runge-Kutta)\cite{HP06}, and an implicit stochastic split backward Euler (SSBE) \cite{MSH02}, and we denote the inferred schemes by IS-EM, IS-RK4 and IS-SSBE. We test them on three non-globally Lipschitz SDEs: the 1D double-well potential, a 2D multiscale gradient system, and the 3D stochastic Lorenz equation with degenerate noise. Numerical results show that the inferred schemes can tolerate time-steps ten to hundreds times larger than plain numerical schemes, and it reaches optimal accuracy in reproducing the invariant measure at a medium large time-step (see Figures \ref{fig:TVDPDF_LangLocal1D}, \ref{fig:TVDPDF_GradCouple}, \ref{fig:TVDPDF_Lorenz}, and \ref{fig:TVDPDF_Lorenz_x3}). Overall, IS-RK4 produces the most accurate invariant measures in all examples, particularly when the dynamics is dominated by the drift (e.g., the Lorentz system) because the plain RK4 provides a higher order approximation to the drift.
Discretization with large time-stepping for differential equations (SDEs, ODEs and PDEs) is a model reduction in time, part of the general problem of space-time model reduction (see e.g., \cite{weinan2007heterogeneous,KMK03,legoll2010effective,MH13,CL15,leiDatadrivenParameterization2016,Lu20Burgers,hudson2020coarse}). Since the large time-step prevents classical numerical approximations based on Taylor expansions, data-driven approaches have been the primary efforts and have witnessed many successes, including the time series approaches (see e.g., \cite{CL15,LLC17,LinLu20}) and deep learning methods that can efficiently solve high-dimensional PDEs and SDEs on rough space-time meshes (see e.g., \cite{han2018_SolvingHighdimensional,sirignano2018_DGMDeep,bar2019learning,Oosterlee2020,yang2020physics}), to name just a few.
In these approaches, the discrete-time models account for the effects of the unresolved dynamics in an averaged fashion through inference, thus leading to computationally efficient models for the effective dynamics \cite{legoll2010effective,Legoll2019,CL15,LLC16}.
The contribution of our ISALT is to provide a simple yet effective approach to achieve large time-stepping by combining inference with classical numerical schemes. In particular, the explicit parametric form in ISALT clearly identifies the connection between classical numerical schemes and the models inferred from data. It provides a ground for further understanding the fundamental issues of data-based reduced models, such as quantification of the approximation and optimality of the reduction in time or in space-time.
The exposition of our study proceeds as follows. We first summarize the notations in Table 1. After introducing a flow map view for numerical schemes, we introduce in Section 2 the ISALT framework, that is, the procedure and algorithm for inferring schemes adaptive to the large time-step from data. Section 3 presents the theoretical results on the convergence of the estimators. In Section 4, we test ISALT on the three typical non-globally Lipschitz SDEs. Section 5 concludes our main findings with an outlook of future research.
\vspace{-2mm}
\begin{table}[H]
\begin{center}
\caption{Notations
\label{tab:notation-Infer}
\begin{tabular}{ l l }
\toprule
Notation & Description \\ \hline
$\mX_t$ and ${\bf B}_t$ & true state process and original stochastic force \\
$f(\mX_t)$,\; $\sigma\in \mathbb{R}^{d\times m}$ with $m\le d$ & local-Lipschitz drift and diffusion matrix\\
$dt$ & time-step generating data \\
$\delta}%{\Delta t= \mathrm{Gap} \times dt$ & time-step for inferred scheme, $\mathrm{Gap}\in \{ 1, 2, 4, 10, 20, 40, 80,\ldots\}$ \\
$t_i = i\delta}%{\Delta t$ & discrete time instants of data \\
$\{\mX_{t_0:t_N}^{(m)}, {\bf B}_{t_0:t_N}^{(m)}\}_{m=1}^M$ & Data: $M$ independent paths of $\mX$ and ${\bf B}$ at discrete-times \\ \hline
$\mathcal{F}\left(\mX_{t_i},\, {\bf B}_{[t_{i}, \, t_{i+1})}, t_{i}, t_{i+1}\right)$
& true flow map representing $(\mX_{t_{i+1}}-\mX_{t_i})/\delta}%{\Delta t$\\
${F}^{\delta}(\mX_{t_n},\Delta {\bf B}_{t_n})$ & approximate flow map using only $\mX_{t_n}$ and $\Delta {\bf B}_{t_n} = {\bf B}_{t_{n+1}}-{\bf B}_{t_{n}} $\\
$\widetilde F^{\delta}\left(c^{\delta}, \mX_{t_n},\Delta {\bf B}_{t_n} \right)$ & parametric approximate flow map \\
$c^{\delta}=(c_0^{\delta},\dots,c_p^{\delta})=c_{0:p}^{\delta}$
& parameters to be estimated for the inferred scheme\\
$\eta_n$ and $\sigma_{\eta}^{\delta}$ & iid $N(0, I_d)$ and a diagonal matrix, representing regression residual\\ \hline
EM and IS-EM & plain Euler-Maruyama and inferred scheme (IS) parametrizing EM\\
HRK4 and IS-RK4 & plain hybrid RK4 and inferred scheme parametrizing RK4 \\
SSBE and IS-SSBE & split-step stochastic backward Euler and IS parametrizing it\\
\bottomrule
\end{tabular}
\end{center} \vspace{-2mm}
\end{table}
\section{Inference of explicit schemes from data}
Throughout this study, we assume that the SDE \eqref{eq:sde} is ergodic. Roughly speaking, the SDE is ergodic when (i) there is a Lyapunov function $V:\mathbb{R}^d\to\mathbb{R}^+$ such that $\lim_{|x|\to\infty}V(x) =\infty$ and $\sup_{|x|>R}\mathcal{A}V(x) \to\infty$ as $R\to\infty$, where $\mathcal{A}$ is the generator for \eqref{eq:sde} given by $\mathcal{A} g= \innerp{f,\nabla g} +\frac{1}{2} \sum_{i,j=1}^d [\sigma \sigma^\top]_{i,j} \partial_{x_i,x_j} g$ \cite{MSH02}; and (ii) $\mX$ satisfies a minority condition that ensures recurrence
\cite{MSH02,Khasminskii12}.
Our goal is to design a numerical scheme with a large time-step so that it can efficiently and accurately simulate both short-time dynamics
and large-time statistics such as invariant measures. This is of particular interest for SDEs with non-globally Lipschitz drift, because explicit schemes such as Euler-Maruyama often blow up or miss the invariant measure even if they are stable \cite{roberts1996exponential,MSH02} and implicit schemes are computationally costly while being accurate in large-time statistics.
We obtain explicit schemes with large time-steps through inference from offline data generated by an implicit scheme. The key is to approximate the flow map by parametrization of numerical schemes,
instead of using a generic basis,
to avoid the curse of dimensionality in the statistical learning of the flow map.
Toward the goal, we will first introduce the view that numerical schemes are approximations of the flow map, then we outline the framework of statistical learning of the flow map.
\subsection{A flow map view of numerical schemes}
A numerical scheme aims to approximate the discrete-time flow map of the stochastic process. More precisely, for a time-step $\delta}%{\Delta t> 0$, let $t_i= i\delta}%{\Delta t$ and denote $(\mX_{t_i}, i\geq 0)$ the process defined in \eqref{eq:sde} at discrete times. Based on the Markov property of $(\mX_t)$
a numerical scheme approximates the flow map
\begin{equation}\label{eq:flowmap}
\begin{aligned}
\mX_{t_{i+1}} - \mX_{t_{i}} =\int_{t_i}^{t_{i+1}}f(\mX_s)ds +\int_{t_i}^{t_{i+1}} \sigma d{\bf B}_s
& = \delta}%{\Delta t\mathcal{F}(\mX_{t_i}, {\bf B}_{[t_i,t_{i+1}]}, t_i, t_{i+1}) \\
& \approx \delta}%{\Delta t F^{\delta}%{\Delta t} (\mX_{t_i}, \Delta {\bf B}_{t_i})
\end{aligned}
\end{equation}
where $\mathcal{F}$ is a functional depending on $\mX_{t_i}$, the continuous trajectory $ {\bf B}_{[t_i,t_{i+1}]}$, $t_i$, and $t_{i+1}$. The simplest schemes approximate the functional by a function $F^{\delta}%{\Delta t} (\mX_{t_i}, \Delta {\bf B}_{t_i})$ on $\mathbb{R}^{2d}$, in which one represents $ {\bf B}_{[t_i,t_{i+1}]}$ by its increment on the interval $\Delta {\bf B}_{t_i} ={\bf B}_{t_{i+1}}-{\bf B}_{t_i} \sim \mathcal{N}(0, \delta}%{\Delta t I_d)$. Among many such schemes (\cite{KP99,Hu96,MSH02,mao2007stochastic,leimkuhler2016molecular}), we consider three simple and representative examples: the explicit Euler-Maruyama scheme (EM) scheme \cite{KP99}, the hybrid RK4 (HRK4) \cite{HP06},
and the split-step stochastic backward Euler (SSBE) \cite{MSH02}
\begin{equation}\label{eq:schemes}
\begin{aligned}
& \text{EM} & \mathbf{X}_{n+1} &= \mathbf{X}_n + f(\mathbf{X}_n) \delta}%{\Delta t + \sigma \Delta {\bf B}_n, \\
& \text{HRK4} & \mathbf{X}_{n+1} &= \mathbf{X}_n + \phi_1 (\mathbf{X}_n,\sigma \Delta {\bf B}_n) \delta}%{\Delta t + \sigma \Delta {\bf B}_n,\\
& \text{SSBE} & \mathbf{X}_{n+1}&= \mathbf{X}_* + \sigma \Delta {\bf B}_n, \text{ with } \mathbf{X}_*=\mathbf{X}_n + f(\mathbf{X}_*)\delta}%{\Delta t, \\
\end{aligned}
\end{equation}
where the term $ \phi_1$ is a standard RK4 step with the stochastic force treated as a constant input:
\begin{align*}
\phi_1(\mathbf{X}_n,\sigma \Delta {\bf B}_n): & = (k_1+2 k_2+2 k_3+k_4)/6, \text{ with } \\
k_1 & = f(\mathbf{X}_n)+\sigma \Delta {\bf B}_n/\delta}%{\Delta t,\\
k_2 &= f(\mathbf{X}_n + k_1\cdot \delta /2)+\sigma \Delta {\bf B}_n/\delta}%{\Delta t,\\
k_3 &= f(\mathbf{X}_n + k_2\cdot \delta /2)+\sigma \Delta {\bf B}_n/\delta}%{\Delta t,\\
k_4 &= f(\mathbf{X}_n + k_3\cdot \delta /2)+\sigma \Delta {\bf B}_n/\delta}%{\Delta t.
\end{align*}
Correspondingly, they approximate the flow map $\mathcal{F}(\mX_{t_i}, {\bf B}_{[t_i,t_{i+1}]}, t_i, t_{i+1})$ by
\begin{equation}\label{eq:flowmap_schemes}
\begin{aligned}
& \text{EM} & & F_{EM}^{\delta}%{\Delta t} (\mX_{t_i}, \Delta {\bf B}_{t_i})= f(\mX_{t_i}) + \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t, \\
& \text{HRK4} & & F_{RK4}^{\delta}%{\Delta t} (\mX_{t_i}, \Delta {\bf B}_{t_i})= \phi_1(\mX_{t_i}, \sigma \Delta {\bf B}_{t_i}) + \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t, \\
& \text{SSBE} & & F_{SSBE}^{\delta}%{\Delta t} (\mX_{t_i}, \Delta {\bf B}_{t_i})= (\mX_* - \mX_{t_i})/\delta}%{\Delta t + \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t, \text{ with } \mX_*=\mX_{t_i} + f(\mX_*)\delta}%{\Delta t. \\
\end{aligned}
\end{equation}
For short time simulation, these schemes are of strong order 1, i.e., the discrete approximation converges to the true solution trajectory-wisely in probability at order $\delta^{1}$ as the time-step vanishes, since the noise is additive \cite{Rum82,KP99,Hu96,MSH02}.
For large time simulation aiming to approximate the invariant measure, the explicit schemes can be problematic for local Lipschitz drifts and degenerate noises, for instance, the EM scheme may destroy the Lyapunov structure and fail to be ergodic for any choice of time-step \cite[Lemma 6.3]{MSH02}. The implicit scheme SSBE, on the other hand, is ergodic and produces accurate invariant measure when the time-step is sufficiently small \cite[Section 6]{MSH02}.
In many applications, it is desirable to have an efficient numerical scheme being accurate in both short-time and large-time. A drawback of an implicit scheme is its inefficiency: it has to solve a fixed point problem in the implicit step, which is computationally costly and limits the time-step size. Taking advantage of implicit schemes, we use them to generate data and learn
efficient explicit schemes with large time-steps from the data.
\subsection{Inference of a scheme from data} \label{sec:InferFramework}
We infer from data an explicit scheme that is accurate in both short-time dynamics and large-time statistics. It maintains the efficiency of explicit schemes while preserving the invariant measure as implicit schemes. The key idea is to learn an approximation of the flow map from data. To avoid the curse of dimensionality in the learning of the flow map, which is often high-dimensional and nonlinear, we derive parametric functions from the system and its numerical schemes. Roughly, the inference consists of four parts:
\begin{enumerate}
\item Generation of faithful data using an implicit scheme with a small time-step size;
\item Derivation of a parametric form to approximate the flow map, by extracting basis functions from the system and its numerical approximations;
\item Parameter estimation by maximal likelihood methods, which leads to a least squares problem when the parametric form is linear in the parameters;
\item Model selection: by cross-validation and convergence criteria.
\end{enumerate}
\paragraph{Data generation} We generate faithful data, consisting of trajectories of the process
at discrete times $\{t_i=i\delta\}$, by an accurate implicit scheme. That is, we first solve the system by an implicit scheme with a small time-step $\Delta t< \delta$, then we downsample the solution at the discrete times. We also save the trajectory data of the stochastic force $({\bf B}_t)$. Denote these trajectories by
\begin{equation}\label{eq:data}
\text{Data: } \{\mX_{t_0:t_N}^{(m)}, {\bf B}_{t_0:t_N}^{(m)}\}_{m=1}^M,
\end{equation}
where $N$ denotes the number of observing time grids and $M$ denotes the number of independent trajectories.
The initial conditions $\{ \mX_{t_0}^{(m)} \}_{m=1}^M$ are samples from either a long trajectory, which represents the invariant measure, or an initial distribution that helps to explore the distribution of the process.
\paragraph{Derivation of parametric form} The major difficulty in inference is the approximation of the flow map $\mathcal{F}(\mX_{t_i}, {\bf B}_{[t_i,t_{i+1}]}, t_i, t_{i+1})$, which is an infinite-dimensional functional. When using a non-parametric approach with the generic dictionary or basis functions, one encounters the well-known curse-of-dimensionality (COD): the size of the dictionary or basis functions increases exponentially as the dimension increases. Recent efforts on overcoming the COD include selecting adaptive-to-data basis functions in a nonparametric fashion \cite{jiang2020modeling}, assuming a low-dimensional interaction between the components of the state variable in the spirit of particle interactions \cite{LZTM19}, or or deep learning methods that approximate high dimensional functions through compositions of simple functions \cite{han2018_SolvingHighdimensional,sirignano2018_DGMDeep,bar2019learning,Oosterlee2020,yang2020physics}.
We take a semi-parametric approach: we avoid the COD by deriving parametric functions from the full system and its numerical schemes, which provide rich information about the flow map. In particular, we aim for parametric functions depending linearly on the parameters, so that the parameters can be estimated by least squares and our algorithm is scalable .
We focus on approximating the flow map $\mathcal{F}(\mX_{t_i}, {\bf B}_{[t_i,t_{i+1}]}, t_i, t_{i+1})$ by
the simplest functions $F^{\delta}%{\Delta t} (\mX_{t_i}, \Delta {\bf B}_{t_i})$, in a parametric form
\begin{equation} \label{Fdelta}
F^\delta}%{\Delta t (c^\delta}%{\Delta t; x,\xi) = \sum_{i=0}^p c_i^\delta}%{\Delta t \phi_i(x,\xi),
\end{equation}
with $\xi$ having the same distribution as $\Delta {\bf B}_{t_i}$.
Here $\phi_i:\mathbb{R}^{2d}\to \mathbb{R}^d$ are basis functions to be extracted from numerical schemes (see Section \ref{sec:derivePara}), and $\{c_i^\delta\}$ are the parameters to be estimated from data. That is, with $(\mathbf{X}_n,\xi_n)$ corresponding to $(\mX_{t_n},\Delta {\bf B}_{t_n}) $, we infer the following scheme
\begin{equation} \label{RM_paraForm}
\mathbf{X}_{n+1} = \mathbf{X}_{n} + \delta}%{\Delta t F^\delta}%{\Delta t (c^\delta}%{\Delta t; \mathbf{X}_n,\xi_n) +\delta}%{\Delta t \sigma_\eta \eta_n = \mathbf{X}_{n} + \delta}%{\Delta t\sum_{i=0}^p c_i^\delta}%{\Delta t \phi_i(\mathbf{X}_n,\xi_n) + \delta}%{\Delta t\sigma_\eta^\delta}%{\Delta t \eta_n,
\end{equation}
where we add $\{\sigma_\eta^\delta}%{\Delta t \eta_n\}$ to represent the residual of the regression. For convenience, we assume that $\{\eta_n\}$ is a sequence of iid Gaussian $N(0,I_d)$ random variables and is independent of $\{\xi_n\}$, and $\sigma_\eta^\delta}%{\Delta t$ is a diagonal matrix.
In view of statistical learning,
the function \eqref{Fdelta} approximates the flow map in the function space $\mathcal{H}=\mathrm{span}\{\phi_i(x,\xi)\}_{i=0}^p$, which is a subspace of $L^2(\mathbb{R}^{2d}, \mu\otimes \nu)$ with $\mu$ being the invariant measure of $\mX$ and $\nu\sim \mathcal{N}(0,\delta I_d)$ being the distribution of $\xi$ (which represents $\Delta {\bf B}_{t_i}$). We refer $\{\phi_i(\helen{x},\xi)\}$ as basis functions and will extract them from numerical scheme (see Section \ref{sec:derivePara}).
Here we focus on using only $\Delta {\bf B}_{t_n}$, but one can use more sample points of the trajectory ${\bf B}_{[t_n,t_{n+1}]}$ and extract terms from high-order approximations based on multiple stochastic integral \cite{Hu96}. We postpone this as future work.
\paragraph{Parameter estimation} We estimate the parameters by maximizing the likelihood for the model in \eqref{RM_paraForm} with the data $\{\mX_{t_0:t_N}^{(m)}, {\bf B}_{t_0:t_N}^{(m)}\}_{m=1}^M$:
\begin{align*}
l(c_{0:p}^\delta}%{\Delta t)
& = \frac{1}{M}\sum_{m=1}^M l(\mX_{t_0:t_N}^{(m)}, {\bf B}_{t_0:t_N}^{(m)} \mid c_{0:p}^\delta}%{\Delta t), \text{ where } \\
l(\mX_{t_0:t_N}, {\bf B}_{t_0:t_N} \mid c_{0:p}^\delta}%{\Delta t) & =
\frac{1}{N} \sum_{k=1}^d\sum_{n=0}^{N-1} \left[ \frac{|\mX_{t_{n+1}}^k -\mX_{t_{n}}^k - \delta}%{\Delta t F_k^\delta}%{\Delta t (c^\delta}%{\Delta t,\mX_{t_i},\Delta {\bf B}_{t_n})|^2 }{2\sigma_{k,\delta}%{\Delta t}^2} -\frac{1}{2} \log (2\pi (\sigma_{k,\delta}%{\Delta t})^2) \right],
\end{align*}
where $F_k^\delta}%{\Delta t$ is the $k$-th entry of the $\mathbb{R}^d$-valued function $F^\delta}%{\Delta t$ defined in \eqref{Fdelta}:
\[
F_k^\delta}%{\Delta t (c^\delta}%{\Delta t, \mX_{t_i},\Delta {\bf B}_{t_n}) = \sum_{i=0}^p c_{i,k}^\delta}%{\Delta t \phi_i^k(\mX_{t_n},\Delta {\bf B}_{t_n}).
\]
Noticing that the likelihood function is quadratic in the parameters $\big\{c_{i,k}^\delta\big\}_{i=0}^{p}$, we estimate them by least squares regression:
\begin{equation}\label{est_c}
\begin{aligned}
\widehat{c_{0:p,k}^{\delta,N,M}} & = (\widebar{A}^{N,M}_k)^+ \widebar{b}_k^{N,M}, \\
(\widehat{\sigma_{\eta,k}^{\delta}%{\Delta t,N,M} })^2 & =\frac{1}{N} \sum_{n=0}^{N-1} |\mX_{t_{n+1}}^k -\mX_{t_{n}}^k - \delta}%{\Delta t F_k^\delta}%{\Delta t (\widehat{c^{\delta}%{\Delta t,N,M} },\mX_{t_i},\Delta {\bf B}_{t_n})|^2,
\end{aligned}
\end{equation}
where $A^{+}$ denotes the pseudo-inverse of $A$, and the normal matrix $\widebar{A}^{N,M}_k$ and vector $ \widebar{b}_k^{N,M}$ are given by
\begin{equation}\label{LeatSquare_def}
\begin{aligned}
\widebar{A}^{N,M}_k(i,j) &=\frac{1}{MN} \sum_{m=1}^M\sum_{n=0}^{N-1} \phi_i^k(\mX_{t_n}^{k,(m)},\Delta {\bf B}_{t_n}^{k,(m)}) \phi_j^k(\mX_{t_n}^{k,(m)},\Delta {\bf B}_{t_n}^{k,(m)}),\quad i,j=0,\dots, p, \\
\widebar{b}_k^{N,M} (i) &= \frac{1}{MN} \sum_{m=1}^M\sum_{n=0}^{N-1} \frac{ \mX_{t_{n+1}}^{k,(m)}-\mX_{t_{n}}^{k,(m)}}{\delta}%{\Delta t} \phi_i^k(\mX_{t_n}^{k,(m)},\Delta {\bf B}_{t_n}^{k,(m)}) ,
\quad i=0,\dots,p.
\end{aligned}
\end{equation}
Here $\widehat{\sigma_{\eta,k}^{\delta}%{\Delta t,N,M} }$, the square root of the regression's residuals, provide the diagonal entries of $\sigma_\eta^\delta}%{\Delta t$.
The above least square regression is based on the assumption that the residual $\sigma^{\delta}\eta_n$ defined in \eqref{RM_paraForm} is Gaussian with uncorrelated entries. The entry-wise regression aims to reflect the dynamical scale difference between entries. One may improve the approximation by considering correlated entries or other distributions for the residual.
\paragraph{Model selection}
The parametric form in Eq.\eqref{RM_paraForm} has many freedoms underdetermined, particularly when we have multiple options for the parametric form, along with possible overfitting and redundancy in these options. We select the estimated scheme by the following criteria:
\begin{itemize}
\item Cross validation: the estimated scheme should be stable and can reproduce the distribution of the process, particularly the main dynamical-statistical properties. We will consider the marginal invariant densities and temporal correlations:
\begin{equation} \label{eq:stats}
\begin{aligned}
\text{Invariant density of $(\mX_t^k)$: } & p(z)dz = \mathbb{E} [\mathbf{1}_{ (z, z+ dz)}(\mX_{t_n}^{k})] \approx \frac{1}{N M}\sum_{m,n=1}^{M,N} \mathbf{1}_{ (z, z+ dz)}(\mX_{t_n}^{k,(m)}), \\
\text{Temporal correlations: } & C_k(h)= \mathbb{E} [\mX_{t_n+h}^{k} \mX_{t_n}^{k}] \approx \frac{1}{N M}\sum_{m,n=1}^{M,N} \mX_{t_n+h}^{k,(m)} \mX_{t_n}^{k,(m)}
\end{aligned}
\end{equation}
for $k=1,\ldots,d$.
\item Convergence of the estimators. If the model is perfect and the data are either independent trajectories or a long trajectory from an ergodic measure, the estimators should converge to the true values when the data size increases (see Theorem~\ref{thm:const_perfect_model}). While our parametric model is not perfect, the estimators should also converge when the data size increases (see Theorem~\ref{thm_convEst}) and highly oscillatory estimators indicate large misfits between the proposed model and data.
\end{itemize}
\subsection{Parametrization of numerical schemes} \label{sec:derivePara}
We derive parametric forms to approximate the flow map from numerical schemes. The numerical schemes provide informed basis functions for inference because of their error-controlled approximations to the flow map $\mathcal{F}(\mX_{t_i}, {\bf B}_{[t_i,t_{i+1}]},t_i,t_{i+1})$ in \eqref{eq:flowmap}. These basis functions can either be simply the terms in an explicit scheme or terms approximating the implicit schemes. One may view this approach as a parametrization of numerical schemes.
We focus on using only $\Delta {\bf B}_{t_i}$, the increment of ${\bf B}_{[t_i,t_{i+1}]}$, and seek parametric functions
$F^{\delta}%{\Delta t} (c^{\delta}%{\Delta t},\mX_{t_i}, \Delta {\bf B}_{t_i})$ (as in \eqref{Fdelta}) to approximate the flow map. This constraint has two advantages: first, it makes the inferred-scheme computationally efficient, because the inferred scheme will generate only two random numbers ($\xi_i, \eta_i$ in \eqref{RM_paraForm}) in each time step to represent the stochastic forces; second, it significantly reduces the function space of inference, from a functional depending on the path ${\bf B}_{[t_i,t_{i+1}]}$ to a function depending only on the increment. By starting from this simple setting, we hope to provide insight on the future design of schemes using multi-point noise by parametrizing high-order stochastic schemes (see e.g.\cite{Hu96,KP99,jentzen2010taylor}).
The flow maps \eqref{eq:flowmap_schemes} of the numerical schemes in \eqref{eq:schemes} provide three representative candidates for a parametric function $ \widetilde F^{\delta}%{\Delta t} (c^{\delta}%{\Delta t},\mX_{t_i}, \Delta {\bf B}_{t_i})$. The EM is an explicit one-step scheme, the RK4 is an explicit multi-step scheme, and the SSBE is an implicit one-step scheme. Linearly parametrizing them or their Ito-Taylor expansions, i.e., adding coefficients to the terms, we obtain parametric flow maps:
\begin{equation}\label{eq:paraFormEuler}
\begin{aligned}
& \text{EM} & \widetilde F_{EM}^{\delta}%{\Delta t} (c^{\delta}%{\Delta t};\mX_{t_i}, \Delta {\bf B}_{t_i}) &= c_0^{\delta}%{\Delta t} \mX_{t_i} + c_1^{\delta}%{\Delta t} f(\mX_{t_i}) + c_2^{\delta}%{\Delta t} \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t, \\
& \text{HRK4} & \widetilde F_{RK4}^{\delta}%{\Delta t} (c^{\delta}%{\Delta t};\mX_{t_i}, \Delta {\bf B}_{t_i}) &= c_0^{\delta}%{\Delta t} \mX_{t_i} + c_1^{\delta}%{\Delta t} \phi_1(\mX_{t_i}, \sigma \Delta {\bf B}_{t_i}) + c_2^{\delta}%{\Delta t} \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t, \\
& \text{SSBE} & \widetilde F_{SSBE}^{\delta}%{\Delta t} (c^{\delta}%{\Delta t};\mX_{t_i}, \Delta {\bf B}_{t_i}) &= c_0^{\delta}%{\Delta t} \mX_{t_i} + c_1^{\delta}%{\Delta t} \phi_1^{SSBE}(\mX_{t_i}) + c_2^{\delta}%{\Delta t} \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t,
\end{aligned}
\end{equation}
where the function
$\phi_1^{SSBE}$ is given by
\begin{equation}\label{phi1_BEs}
\begin{aligned}
\phi_1^{SSBE}(\mX_{t_i}) & = (I_d - \delta}%{\Delta t \nabla f(\mX_{t_i})) ^{-1} f(\mX_{t_i}).
\end{aligned}
\end{equation}
These terms are derived as follows.
\begin{itemize}
\item The parametric flow map $\widetilde F_{EM}^{\delta}%{\Delta t} (c^{\delta}%{\Delta t};\mX_{t_i}, \Delta {\bf B}_{t_i}) $ and $\widetilde F_{RK4}^{\delta}%{\Delta t} (c^{\delta}%{\Delta t};\mX_{t_i}, \Delta {\bf B}_{t_i}) $ come simply by adding coefficients to each term in $F_{EM}^{\delta}%{\Delta t} $ and $F_{RK4}^{\delta}%{\Delta t} $ of the Euler and RK4 schemes in \eqref{eq:flowmap_schemes}.
\item We introduced an extra linear term $c_0^{\delta}%{\Delta t} \mX_{t_i} $. When $f$ is nonlinear, it serves as a linear basis function, and it helps to data-adaptively adjust the linear stability of the inferred scheme.
\item The parametric flow maps $\widetilde F_{SSBE}^{\delta}%{\Delta t} (c^{\delta}%{\Delta t};\mX_{t_i}, \Delta {\bf B}_{t_i}) $
comes from parametrizing the terms in an approximation of $F_{SSBE}^{\delta}%{\Delta t} (\mX_{t_i}, \Delta {\bf B}_{t_i})$ in \eqref{eq:flowmap_schemes}. More precisely, by the mean-value theorem, there exists a state $\widetilde{\mX}_{t_i}$ depending on $\mX_*$ and $\mX_{t_i}$ such that
\begin{equation}\label{eq:res_SSBE}
\begin{aligned}
f(\mX_*) &= f(\mX_{t_i}) + \nabla f(\widetilde{\mX}_{t_i}) (\mX_*-\mX_{t_i}) \\
& = f(\mX_{t_i}) + \nabla f(\mX_{t_i}) (\mX_*-\mX_{t_i}) +R(\mX_*,\mX_{t_i},\nabla f),
\end{aligned}
\end{equation}
where $R(\mX_*,\mX_{t_i},\nabla f)= [\nabla f(\widetilde{\mX}_{t_i})-\nabla f(\mX_{t_i})] (\mX_*-\mX_{t_i}) $.
Then, by the definition of $\mX_*$ in the SSBE in \eqref{eq:flowmap_schemes}, we have
\begin{align*}
& \mX_*=\mX_{t_i}+ \delta}%{\Delta t f(\mX_*) = \mX_{t_i} + \delta}%{\Delta t [ f(\mX_{t_i}) + \nabla f(\mX_{t_i}) (\mX_{*}-\mX_{t_i})] +R(\mX_*,\mX_{t_i},\nabla f) \\
& \Rightarrow (\mX_*-\mX_{t_i}) = (I_d - \delta}%{\Delta t \nabla f(\mX_{t_i})) ^{-1} \delta}%{\Delta t f(\mX_{t_i}) +R(\mX_*,\mX_{t_i},\nabla f) .
\end{align*}
Thus, we have
\[
F_{SSBE}^{\delta}%{\Delta t} (\mX_{t_i}, \Delta {\bf B}_{t_i})= (I_d - \delta}%{\Delta t \nabla f(\mX_{t_i})) ^{-1} f(\mX_{t_i}) + \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t + R(\mX_*,\mX_{t_i},\nabla f).
\]
Assuming that $R(\mX_*,\mX_{t_i},\nabla f)$ is negligible, parametrizing the other terms, and adding $c_0^{\delta}%{\Delta t} \mX_{t_i} $, we obtain $\widetilde F_{SSBE}^{\delta}%{\Delta t}$ with $\phi_1^{SSBE}$ above. Note that when $f$ is globally Lipschitz (thus $|\nabla f|$ is bounded above), we have $\mathbb{E}[ |R(\mX_*,\mX_{t_i},\nabla f)|] \leq C \mathbb{E}[|\mX_*-\mX_{t_i}|^2]$, i.e., $R(\mX_*,\mX_{t_i},\nabla f)$ is an order smaller than $\mX_*-\mX_{t_i}$. However, when $f$ is non-globally Lipschitz (thus $|\nabla f|$ is unbounded ), $R(\mX_*,\mX_{t_i},\nabla f)$ may be non-negligible and require additional terms to account for its effect.
\end{itemize}
\begin{comment}
The parametric flow map $\widetilde F_{SBE}^{\delta}%{\Delta t} (c^{\delta}%{\Delta t};\mX_{t_i}, \Delta {\bf B}_{t_i}) $
comes similarly. From the definition of $\mX_*$ in the SBE in \eqref{eq:flowmap_schemes}, we have
\begin{align*}
& \mX_* =\mX_{t_i}+ \delta}%{\Delta t f(\mX_*) + \sigma \Delta {\bf B}_{t_i} = \mX_{t_i} + \delta}%{\Delta t [ f(\mX_{t_i}) + \nabla f(\mX_{t_i}) (\mX_{*}-\mX_{t_i})] + \text{O}(\delta}%{\Delta t^2) + \sigma \Delta {\bf B}_{t_i} \\
& \Rightarrow (\mX_*-\mX_{t_i}) = (I_d - \delta}%{\Delta t \nabla f(\mX_t)) ^{-1} [ \delta}%{\Delta t f(\mX_{t_i}) + \sigma \Delta {\bf B}_{t_i}] +\text{O}(\delta}%{\Delta t^2),
\end{align*}
Thus, the flow map $F_{SBE}^{\delta}%{\Delta t}$ can be written as
\begin{align*}
F_{SBE}^{\delta}%{\Delta t} (\mX_{t_i}, \Delta {\bf B}_{t_i}) & = f(\mX_*) + \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t \\
&= f(\mX_{t_i}) + \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t + \nabla f(\mX_{t_i}) (I_d - \delta}%{\Delta t \nabla f(\mX_t)) ^{-1} [ \delta}%{\Delta t f(\mX_{t_i}) + \sigma \Delta {\bf B}_{t_i}] +\text{O}(\delta}%{\Delta t^2).
\end{align*}
Neglecting $\text{O}(\delta}%{\Delta t^2)$, parametrizing the other terms, and adding $c_0^{\delta}%{\Delta t} \mX_{t_i} $, we obtain $\widetilde F_{SBE}^{\delta}%{\Delta t}$ with $\phi_3^{SBE}$ above.
\end{comment}
\bigskip
Putting the parametric flow maps in the form in \eqref{RM_paraForm}, the corresponding inferred schemes (IS) with these parametrized flow maps in \eqref{eq:paraFormEuler} are
\begin{equation}\label{eq:IS_Euler}
\begin{aligned}
& \text{IS-EM} & (\mX_{t_{i+1}} - \mX_{t_i})/\delta}%{\Delta t &= c_0^{\delta}%{\Delta t} \mX_{t_i} + c_1^{\delta}%{\Delta t} f(\mX_{t_i}) + c_2^{\delta}%{\Delta t} \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t + \sigma_\eta \eta_i, \\
& \text{IS-RK4: } & (\mX_{t_{i+1}} - \mX_{t_i})/\delta}%{\Delta t &=c_0^{\delta}%{\Delta t} \mX_{t_i} + c_1^{\delta}%{\Delta t} \phi_1 (\mX_{t_i},\sigma \Delta {\bf B}_{t_i})+
c_2^{\delta}%{\Delta t} \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t+\sigma_{\eta}\eta_{i} \\
& \text{IS-SSBE} & (\mX_{t_{i+1}} - \mX_{t_i})/\delta}%{\Delta t &= c_0^{\delta}%{\Delta t} \mX_{t_i} + c_1^{\delta}%{\Delta t} \phi_1^{SSBE}(\mX_{t_i}) + c_2^{\delta}%{\Delta t} \sigma \Delta {\bf B}_{t_i}/\delta}%{\Delta t + \sigma_\eta \eta_i.
\end{aligned}
\end{equation}
We point out that there are many other options for the parametric form. These three to-be-inferred schemes are typical: IS-EM and IS-RK4 are explicit schemes, and they will improve the statistical accuracy of the plain EM or RK4 by design (see Section \ref{sec:convImperfect}). IS-RK4 is based on a multi-step scheme which provides a high-order approximation of the drift, so it is likely to perform better than IS-EM when it is stable. The IS-SSBE comes from an implicit scheme, and is likely to inherit the stability.
\subsection{Algorithm}
The following algorithm summarizes that above procedure for the inference of a scheme.
\begin{algorithm}[H]
{\small
\caption{Inference-based reduced model with memory: detailed algorithm}\label{alg:main}
\begin{algorithmic}[1]
\Require{Full model; a high fidelity solver preserving the invariant measure. }
\Ensure{Estimated parametric scheme}
\State Generate data: solve the system with the high fidelity solver, which has a small time-step $dt$; down sample to get time series with $\delta}%{\Delta t= \mathrm{Gap}\times dt$. Denote the data, consisting of $M$ independent trajectories on $[0,N\delta}%{\Delta t]$, by $\{\mX_{t_0:t_N}^{(m)}, {\bf B}_{t_0:t_N}^{(m)}\}_{m=1}^M$ with $t_i= i\delta}%{\Delta t$.
\State Pick a parametric form approximating the flow map \eqref{eq:flowmap} as in \eqref{Fdelta}--\eqref{RM_paraForm}.
\State Estimate parameters $c_{0:p}^\delta}%{\Delta t$ and $\sigma_\eta$ as in \eqref{est_c}.
\State Model selection: run the inferred scheme for cross-validation, and test the consistency of the estimators.
\end{algorithmic}
}
\end{algorithm}
\section{Convergence of estimators}\label{sec:conv}
We consider the convergence of the estimators in sample size in two settings: perfect model and imperfect model. The perfect model setting aims to validate our algorithm, in the sense that the algorithm can yield consistent and asymptotically normal estimators.
The imperfect model setting is what we have in practice, and we show that our estimator converges to the (optimal) projection. In particular, we show that an inferred-scheme improves the statistical accuracy of its explicit counterpart.
For simplicity of notation, we assume that $d=1$ throughout this section. But the results also hold true entry-wisely for the system with $d>1$.
\subsection{Convergence of estimator for perfect model}
We denote the expectation of $\widebar{A}^{N,M}$ and $\widebar{b}^{N,M}$ in \eqref{LeatSquare_def} by $A$ and $b$:
\begin{equation}\label{Aexpectation}
\begin{aligned}
A= \mathbb{E}[\widebar{A}^{N,M}] &= \frac{1}{N} \sum_{n=0}^{N-1} \left( \mathbb{E}\left[ \langle \phi_i(\mX_{t_n}^{(m)}, \Delta {\bf B}_{t_n}^{(m)}), \phi_j(\mX_{t_n}^{(m)},\Delta {\bf B}_{t_n}^{(m)})\rangle_{\mathbb{R}^d} \right] \right)_{i,j}, \\
b = \mathbb{E}[\widebar{b}^{N,M}] & = \frac{1}{N} \sum_{n=0}^{N-1}(\mathbb{E}[ \langle \frac{ \mX_{t_{n+1}}^{(m)}-\mX_{t_{n}}^{(m)}}{\delta}%{\Delta t}, \phi_i(\mX_{t_n}^{(m)},\Delta {\bf B}_{t_n}^{(m)}) \rangle_{\mathbb{R}^d} ])_i.
\end{aligned}
\end{equation}
Here the expectation is with respect to the distribution filtration generated by the initial distribution and the Brownian motion.
\begin{assumption}\label{Assum0}
(a) Suppose that the data $\{\mX_{t_0:t_N}^{(m)}, {\bf B}_{t_0:t_N}^{(m)}\}_{m=1}^M$ are independent trajectories of the system \eqref{RM_paraForm} with $\{\mX_{t_0}^{(m)}\}_{m=1}^M$ sampled from the ergodic measure of $\mX$. (b) Suppose that the normal matrix $\widebar{A}^{N,M}$ in \eqref{LeatSquare_def} and its expectation in \eqref{Aexpectation}
are invertible. (c) Suppose that the flow map $\mathcal{F}^\delta}%{\Delta t$ in \eqref{eq:flowmap} is square integrable.
\end{assumption}
\begin{theorem}[Consistency and asymptotic normality for perfect model]\label{thm:const_perfect_model}
Under Assumption {\rm \ref{Assum0}},
the estimator in \eqref{est_c} converges to $c^{\delta}%{\Delta t}$ (the true parameter value) almost surely, and is asymptotically normal, when either $M\to \infty$ or $N\to \infty$:
\begin{equation}
\begin{aligned}
& \sqrt{M} (\widehat{c^{\delta}%{\Delta t, N,M}} - c^{\delta}%{\Delta t}) \xrightarrow[]{d} \mathcal{N}(0,\frac{1}{N}\sigma_\eta^2 A), \\
& \sqrt{N} (\widehat{c^{\delta}%{\Delta t, N,M}} - c^{\delta}%{\Delta t}) \xrightarrow[]{d} \mathcal{N}(0,\frac{1}{M}\sigma_\eta^2 A).
\end{aligned}
\end{equation}
\end{theorem}
\begin{proof}
By definition of $\widebar{b}^{N,M}$ in \eqref{LeatSquare_def} and the equation \eqref{RM_paraForm}, we have
\begin{align*}
\widebar{b}^{N,M}(i)
& = \frac{1}{MN} \sum_{m=1}^M\sum_{n=0}^{N-1} \langle \sum_{j=0}^p c^\delta}%{\Delta t_j \phi_j (\mX_{t_n}^{(m)},\Delta {\bf B}_{t_n}^{(m)})+ \sigma_\eta \eta_n^{(m)}, \phi_i(\mX_{t_n}^{(m)} ,\Delta {\bf B}_{t_n}^{(m)})\rangle_{\mathbb{R}^d} \\
& = \left( \widebar{A}^{N,M}c^\delta}%{\Delta t\right)(i) + \widebar{S}^{N,M} ,
\end{align*}
where in the second equality we used the definition of $\widebar{A}^{N,M}$ in \eqref{LeatSquare_def}, and we denote
\[
\widebar{S}^{N,M} = \frac{1}{M} \sum_{m=1}^M S^{N,(m)}, \, \text{ with } S^{N,(m)}= \frac{1}{N} \sum_{n=0}^{N-1} \langle \sigma_\eta \eta_n^{(m)}, \phi_i(\mX_{t_n}^{(m)} ,\Delta {\bf B}_{t_n}^{(m)})\rangle_{\mathbb{R}^d}.
\]
Note that $\eta_n$ is standard Gaussian and is independent of ${\bf B}_{t_n}$ and $\mX_{t_n}$. Then,
$S^{N,(m)}$
has mean zero and its covariance is
\[
\mathrm{Cov}(S^{N,(m)}) = \sigma_\eta^2 \frac{1}{N^2} \sum_{n,n'=0}^{N-1} \mathbb{E} \left[ \langle \eta_n^{(m)}, \phi_i(\mX_{t_n}^{(m)} ,\Delta {\bf B}_{t_n}^{(m)})\rangle_{\mathbb{R}^d} \langle \eta_{n'}^{(m)}, \phi_i(\mX_{t_{n'}}^{(m)} ,\Delta {\bf B}_{t_{n'}}^{(m)})\rangle_{\mathbb{R}^d} \right] = \frac{1}{N}\sigma_\eta^2 A.
\]
Thus, when $M\to \infty$, we have by the central limit theorem,
\begin{equation}\label{S_AN_M}
\sqrt{M} \frac{1}{M} \sum_{m=1}^M S^{N,(m)} \xrightarrow[]{d} \mathcal{N}(0, \frac{1}{N}\sigma_\eta^2 A);
\end{equation}
Furthermore, $S^{N,(m)}$ is a martingale with respect to the filtration generated by $\{\mX_{t_n}, {\bf B}_{t_n}, \eta_n\}$, and
when $N\to \infty$, we have by martingale central limit theorem \cite[Theorem 3.2]{hall2014martingale}
\begin{equation} \label{S_AN_N}
\sqrt{N} \frac{1}{M} \sum_{m=1}^M S^{N,(m)} \xrightarrow[]{d} \mathcal{N}(0, \frac{1}{M}\sigma_\eta^2 A ).
\end{equation}
We show first that, when $M\to \infty$ and for each fixed $N$, the estimator is consistent and asymptotically normal. Note that by the strong Law of Large Numbers, $ \widebar{A}^{N,M}\to A$ and $\widebar{b}^{N,M}\to b$ a.s.~as $M\to \infty$. Thus, $ (\widebar{A}^{N,M})^{-1} \to A^{-1}$ almost surely (using the fact that $A^{-1}-B^{-1} = A^{-1}(B-A)B^{-1}$, see \cite[page 22]{LMT20}). Then, $\widehat{c^{\delta}%{\Delta t, N,M}}=(\widebar{A}^{N,M})^{-1} \widebar{b}^{N,M} \to A^{-1}b$ almost surely, i.e. the estimator is consistent. Combining \eqref{S_AN_M} and the almost sure convergence of $(\widebar{A}^{N,M})^{-1} $, we obtain the asymptotic normality by noticing that
\[
\widehat{c^{\delta}%{\Delta t, N,M}}=(\widebar{A}^{N,M})^{-1} \widebar{b}^{N,M} = c^\delta}%{\Delta t + (\widebar{A}^{N,M})^{-1} \widebar{S}^{N, M} .
\]
When $N\to\infty$ and $M$ fixed, we obtain $ \widebar{A}^{N,M}\to A$ and $\widebar{b}^{N,M}\to b$ a.s.~by the ergodicity of the process. The consistency and asymptotic normality follows similarly by using \eqref{S_AN_N}.
\end{proof}
\subsection{Convergence of estimator for imperfect model} \label{sec:convImperfect}
In practice, the model is imperfect in our inferred scheme because we can rarely parametrize the flow map exactly. We show next that for an imperfect proposed model, the estimator converges to the projected coefficient of the flow map onto the function space spanned by the proposed basis in the ambient $L^2$ space.
Furthermore, we show that the inferred scheme improves the statistical accuracy of the explicit scheme that it parametrizes.
\begin{assumption}\label{Assum1}
(a) Suppose that the data $\{\mX_{t_0:t_N}^{(m)}, {\bf B}_{t_0:t_N}^{(m)}\}_{m=1}^M$ are independent trajectories of the system \eqref{eq:sde} with $\{\mX_{t_0}^{(m)}\}_{m=1}^M$ sampled from the ergodic measure of $\mX$. (b) Suppose that the normal matrix $\widebar{A}^{N,M}$ in \eqref{LeatSquare_def} and its expectation in \eqref{Aexpectation}
are invertible. (c) Suppose that the flow map $\mathcal{F}^\delta}%{\Delta t$ in \eqref{eq:flowmap} is square integrable.
\end{assumption}
The invertibility of the normal matrices $\widebar{A}^{N,M}$ and $A$ is crucial for our theory, and they lead to constraints on the basis functions. In practice, we can use it to guide the selection of basis functions and we recommend using pseudo-inverse and regularization when the normal matrix is close to singular.
With the notation $A$ and $b$ in \eqref{Aexpectation}, and assuming that $A$ is invertible, we define
\begin{equation}\label{eq:c_proj}
c^{\delta}%{\Delta t,\mathrm{proj}} := A^{-1} b.
\end{equation}
The following lemma shows that $c^{\delta}%{\Delta t,\mathrm{proj}} $ is the projection coefficients of the flow map $\mathcal{F}^\delta}%{\Delta t$.
\begin{lemma} \label{lemma:c_proj}
Under Assumption {\rm \ref{Assum1}}, the vector $c^{\delta}%{\Delta t,\mathrm{proj}} $ in \eqref{eq:c_proj} is the projection coefficient of the flow map $\mathcal{F}^\delta}%{\Delta t$ in \eqref{eq:flowmap} onto the space $\mathrm{span}\{\phi_i\}_{i=0}^p$ in $L^2(\mathbb{R}^{d}\times \Omega^\delta}%{\Delta t, \mu\otimes \nu)$ with $\mu$ being the invariant measure of $\mX$ and $(\Omega^\delta}%{\Delta t, \mathcal{B},\nu)$ being the canonical probability space for the Brownian motion $({\bf B}_t, t\in [0,\delta}%{\Delta t])$.
\end{lemma}
\begin{proof} Note that $\mathcal{F}^\delta}%{\Delta t_{t_n} = \frac{\mX_{t_{n+1}}-\mX_{t_{n}}}{\delta}%{\Delta t}$. Denote $\mathcal{F}^{\delta}%{\Delta t, m}_{t_n} = \frac{\mX_{t_{n+1}}^{(m)}-\mX_{t_{n}}^{(m)}}{\delta}%{\Delta t}$. By the definition of $b$ in \eqref{Aexpectation}, we have
\begin{align*}
b(i
& = \frac{1}{N} \sum_{n=0}^{N-1} \mathbb{E} \langle\mathcal{F}^{\delta}%{\Delta t,m}_{t_n}, \phi_i(\mX_{t_n}^{(m)} ,\Delta {\bf B}_{t_n}^{(m)})\rangle_{\mathbb{R}^d} = \mathbb{E} \langle\mathcal{F}^{\delta}%{\Delta t}_{t_n}, \phi_i(\mX_{t_n} ,\Delta {\bf B}_{t_n})\rangle_{\mathbb{R}^d},
\end{align*}
where the second equality follows from that $( \mX_{t_n}, \Delta {\bf B}_{t_n})$ is stationary (so does $\mathcal{F}^{\delta}%{\Delta t}_{t_n}$).
Denote by $c = (c_0,c_1,\ldots, c_p)^\top$ the projection coefficients of $\mathcal{F}^\delta}%{\Delta t_{t_n}$ to $\mathrm{span}\{\phi_i\}_{i=0}^p$, and write $\mathcal{F}^{\delta}%{\Delta t}_{t_n} = \sum_{i=0}^p c_i \phi_i + \mathcal{F} $ with $\mathcal{F}$ satisfying $\mathbb{E}[\innerp{\mathcal{F}, \phi_i}_{\mathbb{R}^d} ] =0$ for each $i=0,1,\ldots, p$. Then
\[
\mathbb{E}[\innerp{ \mathcal{F}^{\delta}%{\Delta t}_{t_n}, \phi_i}_{\mathbb{R}^d} ]= \sum_{j=0}^p c_j \mathbb{E}[\innerp{\phi_j, \phi_i}_{\mathbb{R}^d} ] = (A c)(i).
\]
Combining the above two equations, we obtain that $c^{\delta}%{\Delta t,\mathrm{proj}} = A^{-1}b = c$.
\end{proof}
We remark that because $\mathcal{F}^\delta}%{\Delta t(\mX_{t_i}, {\bf B}_{[t_i,t_{i+1}]}, t_i, t_{i+1})$ is a functional depending on the trajectory ${\bf B}_{[t_i,t_{i+1}]}$, the function space of projection, $L^2(\mathbb{R}^{d}\times \Omega^\delta}%{\Delta t, \mu\otimes \nu)$, has an infinite dimensional state space for $(\mX_{t_i}, {\bf B}_{[t_i,t_{i+1}]})$. When $\mathcal{F}^\delta}%{\Delta t_{t_n}$ depends only on $(\mX_{t_n}, \Delta {\bf B}_{t_n})$ (for instance, in the case of perfect model discussed in the previous section), the state space becomes finite dimensional and the function space is simplified to $L^2(\mathbb{R}^{2d}, \mu\otimes \nu)$ with $\nu\sim \mathcal{N}(0,\delta}%{\Delta t I_d)$.
\begin{theorem}[Convergence of the estimator] \label{thm_convEst}
In addition to Assumption {\rm \ref{Assum1}}, assume that $\mathbb{E}[|\mathcal{F}^\delta}%{\Delta t_{t_0}|^4] <\infty$ and $\mathbb{E}[|\phi_i(\mX_{t_0}, \Delta {\bf B}_{t_0}) |^4] <\infty$ for each $i=0,\dots,p$. Then, we have
\begin{itemize}
\item when $M\to \infty$ and $N$ fixed, the estimator in \eqref{est_c} converges to the projection coefficients $c^{\delta}%{\Delta t,\mathrm{proj}}$ in \eqref{eq:c_proj} a.s.~and is asymptotically normal:
\begin{equation}\label{AN_inM}
\begin{aligned}
& \sqrt{M} (\widehat{c^{\delta}%{\Delta t, N,M}} - c^{\delta}%{\Delta t,\mathrm{proj}}) \xrightarrow[]{d} \mathcal{N}(0,A^{-1}\Sigma^N (A^{-1})^\top),
\end{aligned}
\end{equation}
where the matrix $\Sigma^N$ is the covariance of
\[
\widetilde{b}^{N,m}(i) = \frac{1}{N}\sum_{n=0}^{N-1} b^{n,m},\, \text{ with } b^{n,m}= \langle \frac{ \mX_{t_{n+1}}^{(m)}-\mX_{t_{n}}^{(m)}}{\delta}%{\Delta t}, \phi_i(\mX_{t_n}^{(m)},\Delta {\bf B}_{t_n}^{(m)}) \rangle_{\mathbb{R}^d} .
\]
\item when $N\to \infty$ and $M$ fixed, the estimator in \eqref{est_c} also converges and is asymptotically normal
\begin{equation}\label{AN_inN}
\sqrt{N} (\widehat{c^{\delta}%{\Delta t, N,M}} - c^{\delta}%{\Delta t,\mathrm{proj}}) \rightarrow \mathcal{N}(0,\frac{1}{M}A^{-1}\Sigma (A^{-1})^\top),
\end{equation}
with $\Sigma = \lim_{N\to \infty}N\Sigma^N$, provided that for each $m$,
\begin{align*}
& \sum_{n=0}^\infty\mathbb{E}[b^{n,m} \mathbb{E}[b^{k,m} | \mathcal{M}_0] ] \, \text{converges for each $k\geq 0$ and} \\
& \lim_{N\to \infty} \sum_{k=K}^\infty \mathbb{E}[b^{k,m} \mathbb{E}[b^{N,m} | \mathcal{M}_0] ] =0 \text{ uniformly in $K$},
\end{align*}
where $\mathcal{M}_0$ denotes the filtration generated by the extended stationary process up to time $t_0$.
\end{itemize}
\end{theorem}
\begin{proof} When $M\to \infty$, by the strong Law of Larger Numbers, we have $ \widebar{A}^{N,M}\to A$ and $\widebar{b}^{N,M}\to b$ a.s.~as $M\to \infty$. Thus, $\widehat{c^{\delta}%{\Delta t, N,M}}=(\widebar{A}^{N,M})^{-1} \widebar{b}^{N,M} \to A^{-1}b= c^{\delta}%{\Delta t,\mathrm{proj}}$ a.s. according to Lemma~\ref{lemma:c_proj}. To prove the asymptotic normality, note that for each $m$, the random vector $\widetilde{b}^{N,m}$ with entries
\[
\widetilde{b}^{N,m}(i) = \frac{1}{N}\sum_{n=0}^{N-1} \langle \mathcal{F}^{\delta}%{\Delta t,m}_{t_n}, \phi_i(\mX_{t_n}^{(m)},\Delta {\bf B}_{t_n}^{(m)}) \rangle_{\mathbb{R}^d}
\]
has mean $\mathbb{E}[\widetilde{b}^{N,m}] =b$ and covariance $\Sigma^N$ with entries $\Sigma^N_{i,j} = \mathbb{E}[\widetilde{b}^{N,m}(i) \widetilde{b}^{N,m}(j)] - b(i)b(j)$.
Here the covariance exists because
\begin{align*}
\mathbb{E}[\widetilde{b}^{N,m}(i) \widetilde{b}^{N,m}(j)] &\leq \max_i \mathbb{E}[|\widetilde{b}^{N,m}(i)|^2] \leq \max_i \mathbb{E} [| \langle \mathcal{F}^{\delta}%{\Delta t,m}_{t_n}, \phi_i(\mX_{t_n}^{(m)},\Delta {\bf B}_{t_n}^{(m)}) \rangle_{\mathbb{R}^d}|^2] \\
& \leq ( \mathbb{E}[|\mathcal{F}^\delta}%{\Delta t_{t_0}|^4])^{1/2} \max_i (\mathbb{E}[|\phi_i(\mX_{t_0}, \Delta {\bf B}_{t_0}) |^4)^{1/2}.
\end{align*}
Then, $ \widebar{b}^{N,M} $ is the average
of $M$ iid samples $\{\widetilde{b}^{N,m}\} $, each of which has covariance $\Sigma^N$. Hence, by the central limit theorem, we have
\begin{align*}
\sqrt{M}(\widebar{b}^{N,M} - b) & \xrightarrow[]{d} \mathcal{N}(0,\Sigma^N).
\end{align*}
Combining with the facts that $ \widebar{A}^{N,M}\to A$ a.s.~and that these matrices are invertible, we obtain
\eqref{AN_inM}.
\bigskip
When $N\to \infty$, we obtain the convergence and asymptotic normality by ergodicity. First, by ergodicity, we have $ \widebar{A}^{N,M}\to A$ and $\widebar{b}^{N,M}\to b$ a.s.~as $M\to \infty$. Thus, $\widehat{c^{\delta}%{\Delta t, N,M}}=(\widebar{A}^{N,M})^{-1} \widebar{b}^{N,M} \to A^{-1}b= c^{\delta}%{\Delta t,\mathrm{proj}}$ almost surely. Next, to prove the asymptotic normality, note that by the central limit theorem for stationary processes \cite[Theorem 1]{heyde1974central}, we have $\Sigma = \lim_{N\to \infty}N\Sigma^N$ and
\begin{align*}
\sqrt{N}(\widebar{b}^{N,M} - b) & \xrightarrow[]{d} \mathcal{N}(0,\frac{1}{M}\Sigma).
\end{align*}
Then we obtain \eqref{AN_inN} by noting that $ \widebar{A}^{N,M}\to A$ a.s. as above.
\end{proof}
\bigskip
We show next that the parametrization by inference will lead to improvement to an explicit scheme, in the sense that the residual of the inferred scheme is not larger than the explicit scheme. In other words, the 1-step discretization error of the inferred scheme is not larger than the explicit scheme.
\begin{theorem}[Order of residual] \label{thm:order_res}
Assume that the parametric form in \eqref{RM_paraForm} comes from an explicit scheme, such as {\rm IS-EM} or {\rm IS-RK4} in \eqref{eq:IS_Euler} from its explicit scheme in \eqref{eq:flowmap_schemes}. Then, under Assumption {\rm \ref{Assum1}}, the inferred scheme's residual is smaller than the explicit scheme's. Specifically, the residual in \eqref{est_c} satisfies
\begin{equation}\label{eq:order_res}
\mathbb{E} \widehat{(\sigma^{N,M})^2} \leq \frac{2}{d} \mathbb{E}\Big[| \frac{\mX_{t_{n+1}}-\mX_{t_{n}}}{\delta}%{\Delta t} - F^\delta}%{\Delta t(\mX_{t_n}, \Delta {\bf B}_{t_n} )|^2 \Big],
\end{equation}
where $F^\delta}%{\Delta t(\mX_{t_n}, \Delta {\bf B}_{t_n} )$ denotes the flow map of the explicit scheme, such as $ F^\delta}%{\Delta t_{EM}$ or $ F^\delta}%{\Delta t_{RK4}$ in \eqref{eq:flowmap_schemes}. In other words, the inferred scheme has the same 1-step strong order as the explicit scheme it parametrizes.
\end{theorem}
\begin{proof}
Write the flow map of the explicit scheme in parametric form: $F^\delta}%{\Delta t(\mX_{t_n}, \Delta {\bf B}_{t_n} ) = F^\delta}%{\Delta t(c^*,\mX_{t_n}, \Delta {\bf B}_{t_n} )$ as in \eqref{Fdelta}. Then, since the estimator $\widehat{c^{\delta,N,M}}$ in \eqref{est_c} is the minimizer of the likelihood, we have
\begin{align*}
\widehat{(\sigma^{N,M})^2} & = \frac{2}{d\delta}%{\Delta t^2} \frac{1}{MN} \sum_{m=1}^M\sum_{n=0}^{N-1} |\mX_{t_{n+1}}^{(m)} -\mX_{t_{n}}^{(m)} - \delta}%{\Delta t F^\delta}%{\Delta t (\widehat{c^{\delta}%{\Delta t,N,M} },\mX_{t_n}^{(m)},\Delta {\bf B}_{t_n}^{(m)})|^2 \\
& \leq \frac{2}{d\delta}%{\Delta t^2} \frac{1}{MN} \sum_{m=1}^M\sum_{n=0}^{N-1} |\mX_{t_{n+1}}^{(m)} -\mX_{t_{n}}^{(m)} - \delta}%{\Delta t F^\delta}%{\Delta t (c^*,\mX_{t_n}^{(m)},\Delta {\bf B}_{t_n}^{(m)})|^2.
\end{align*}
Since the process $(\mX_{t_n},\Delta {\bf B}_{t_n})_n$ is stationary, we have
\begin{align} \label{Eresidual}
\mathbb{E} \widehat{(\sigma^{N,M})^2} & = \frac{2}{d\delta}%{\Delta t^2} \mathbb{E} |\mX_{t_{n+1}} -\mX_{t_{n}} - \delta}%{\Delta t F^\delta}%{\Delta t (\widehat{c^{\delta}%{\Delta t,N,M} },\mX_{t_i},\Delta {\bf B}_{t_n})|^2 \\
& \leq \frac{2}{d} \mathbb{E} |\frac{\mX_{t_{n+1}} -\mX_{t_{n}} }{\delta}%{\Delta t}- F^\delta}%{\Delta t (c^*, \mX_{t_i},\Delta {\bf B}_{t_n})|^2. \notag
\end{align}
Recall that $F^\delta}%{\Delta t(\mX_{t_n}, \Delta {\bf B}_{t_n} ) = F^\delta}%{\Delta t(c^*,\mX_{t_n}, \Delta {\bf B}_{t_n} )$
we have \eqref{eq:order_res}.
\end{proof}
\begin{remark}[Order of residual for IS-RK4 and IS-EM]\label{rmk:res}
Recall that either Euler-Maruyama scheme or the HRK4 schemes have
$ \mathbb{E} \big[|\frac{\mX_{t_{n+1}} -\mX_{t_{n}}}{\delta}%{\Delta t} - F^\delta}%{\Delta t (c^*, \mX_{t_i},\Delta {\bf B}_{t_n})|^2 \big]= O(\delta}%{\Delta t)$, which follows from Ito formula. Thus, for the inferred schemes IS-EM and IS-RK4 in \eqref{eq:flowmap_schemes}, we have $\mathbb{E} [\widehat{(\sigma^{N,M})^2}] = O(\delta}%{\Delta t) $.
Furthermore, by the Law of Large Numbers, we have $\widehat{(\sigma^{N,M})^2} \to \mathbb{E} \widehat{(\sigma^{N,M})^2} $ a.s.~when either $M\to \infty$ or $N\to \infty$. Thus, the estimator $\widehat{\sigma^{N,M}}= O(\delta}%{\Delta t^{1/2})$ a.s.~for large $N$ or $M$. However, IS-SSBE's residual is not controlled by SSBE, because SSBE is not in the parametric family of IS-SSBE and the neglected term $R$ in \eqref{eq:res_SSBE} may prevent the residual from decaying {\rm (see Figure \ref{fig:conv_GradCouple}(b))}.
\end{remark}
\begin{remark}\label{rmk:N_vs_M} The framework of inference-based scheme applies also for non-ergodic systems to achieve reduced-in-time models. The convergence of the parameters in Theorem~{\rm\ref{thm_convEst}} and Theorem~{\rm\ref{thm:order_res}} remain true when the sample size $M$ goes to infinity. Furthermore, one may accelerate the simulation of slowly converging ergodic systems by training the inference-based schemes iteratively in time. In this study, we focus on non-globally Lipschitz ergodic systems to highlight the ability of the inferred-scheme in producing long-term statistics.
\end{remark}
\begin{remark}[Optimal reduction in time]
We emphasize that our goal is to infer an explicit scheme with a relative large time-step for efficient simulation of non-globally Lipschitz ergodic systems. Theorem {\rm \ref{thm:order_res}} shows that the error in 1-step approximation (of the flow map $\mathcal{F}^\delta}%{\Delta t$ in \eqref{eq:flowmap}) decays at the time-step decreases. But a smaller residual due to a smaller time-step does not necessarily imply a better performance for the inferred scheme, because large error may be accumulated in the invariant measure even when the time-step is small (e.g., EM may have a wrong invariant measure). To improve the inferred scheme, we seek for an optimal time-step that balances the 1-step approximation error and the accumulation into the invariant measure. In our numerical examples in the next section, the inferred scheme performs (in the sense of reproducing the invariant measure and temporal correlation) the best when the time-step is moderately large. This is similar to the parameter estimation for homogenization of multiscale process {\rm \cite{PS07}}, where the sub-sampling rate must be between the two characteristic time scales of the SDE. One may view the inference as an averaging process through the sampling of the invariant measure and connect the error in invariant measure with sampling error, then one can find an optimal time step through a trade-off between the approximation error (i.e., numerical error) and the sampling error. We leave this as future work.
\end{remark}
\section{Examples}
\label{sec:example1}
In this section, we test three benchmark examples for each inference-based scheme proposed in \eqref{eq:IS_Euler} using two different parametric settings: $c_0$ excluded vs $c_0$ included, {so as to distinct the contribution of the linear term parametrized by $c_0$.}
Three non-globally Lipschitz examples are: a 1D equation with double-well potential; a 2D gradient system; and a 3D stochastic Lorenz system with degenerate noises.
In each of the examples, we generate data for inference by the Split Step Backward Euler (SSBE) scheme with very fine scale time step $\Delta t$. We infer schemes for different time step-sizes $\delta}%{\Delta t= \mathrm{Gap}\times \Delta t$ with 10 options for the time gap: $\mathrm{Gap} \in \{ 1, 2, 4, 10, 20, 40, 80, 120, 160,200\}$, which will be used to select optimal time gap and demonstrate the convergence order of the residual in Theorem \ref{thm:order_res}. The computations of inference schemes include 5 different options
(1) IS-EM with $c_0$ excluded;
(2) IS-RK4 with $c_0$ excluded;
(3) IS-RK4 with $c_0$ included;
(4) IS-SSBE with $c_0$ excluded;
and (5) IS-SSBE with $c_0$ included.
We assess the performance of these schemes by the accuracy of the reproduced invariant density (PDF) and the auto-correlations function (ACF), which are empirically computed from a long trajectory. The accuracy of the PDF are measured by the total variation distance (TVD) from the reference PDF from data.
Once we identify the best performing scheme for each example, we fix the inference settings and present the convergence of the estimators and the residuals with respect to the time $\mathrm{Gap}$ as well as the number of trajectories $M$.
\bigskip
In summary, we find from the examples that
\begin{itemize}
\item The inferred scheme has significantly stronger numerical stability than the plain schemes. The IS-RK4 and IS-SSBE exhibit better stability than IS-EM. In particular, they can tolerate time-steps that are significantly larger than the plain RK4 or SSBE. Specifically, we find the plain RK4 and SSBE (without inferred parameters) always blow up even when $\mathrm{Gap}=20$, whereas the inferred schemes are still stable when $\mathrm{Gap}$ is larger than 200, which improves the efficiency by an order of more than 10. We summarize the blow-up gap for plain verse inferred schemes for each example in the following table.
\begin{table}[htp!]
\centering \vspace{-1mm}
\begin{tabular}{|c|c|c|c|}
\hline
{}&{1D double-well}&{2D gradient system}&{3D Lorenz system}\\
\hline
Plain RK4 &$\mathrm{Gap}=20$ & $\mathrm{Gap}=20$ & $\mathrm{Gap}=10$\\
\hline
IS-RK4 &$\mathrm{Gap}>200$ & $\mathrm{Gap}>200$ & $\mathrm{Gap}>400$\\
\hline
Plain SSBE &$\mathrm{Gap}=40$ & $\mathrm{Gap}=40$ & $\mathrm{Gap}=20$\\
\hline
IS-SSBE &$\mathrm{Gap}>200$ & $\mathrm{Gap}>200$ & $\mathrm{Gap}>400$\\
\hline
\end{tabular}
\vspace{-3mm} \caption{Blow up time gap for each scheme: plain verse inferred.
\label{tab:blow_up_gap}}
\end{table}
\item The inferred scheme can reproduce the invariant measure accurately. Both IS-RK4 and IS-SSBE perform well when the stochastic force dominates the dynamics. But when the drift dominates the dynamics in the example of Lorenz system, IS-RK4 performs better than IS-SSBE, because it provides a better approximation to the drift than IS-SSBE.
\item The inferred scheme reproduces the invariant density the best when the time-step is medium large (with a time gap between $\mathrm{Gap}=80$ and $\mathrm{Gap}=160$), suggesting a balance between the approximation error of the flow map and the numerical error in simulating the invariant density. It is open to have an \textit{a-priori} estimate of the optimal time gap.
\end{itemize}
\subsection{1D double-well potential}
First consider an 1D SDE with a double-well potential \cite{MSH02}
\begin{equation}\label{example:1D_DoubleWell}
dX_t=- V'(X_t)dt+\sqrt{2/\beta} dB_t,
\end{equation}
with $V(x)=\frac{\mu}{4}(x^2-1)^2$.
The corresponding invariant measure is $\frac{1}{Z}\exp^{-\beta V(x)}$
where $Z$ being the normalizing constant $Z:=\int_{\mathbb{R}}\exp^{-\beta V(x)}dx$. We set $\mu=2$ and $\beta=1$.
We generate data by SSBE with a fine time-step $\Delta t=1e-3$. We first simulate a long trajectory on an interval $[0,T]$ with $X_0=1/2$ and $T=2000$ (i.e., two million time steps), which is found to be long enough to represent the invariant density (PDF). This long trajectory will also provide us the reference PDF and ACF, which are referred as the true values to be approximated.
Then we generate $M=1000$ trajectories on the time interval $[0,1000]$ with initial conditions sampled from the long trajectory. The data for inference are the $M$ trajectories of both the Brownian motion and the process $(X_{t})$ observed at discrete times $\{t_n=n\delta}%{\Delta t = n\mathrm{Gap} \times \Delta t\}$, as in \eqref{eq:data}.
The parameters of the schemes in \eqref{eq:IS_Euler} are then estimated by Algorithm \ref{alg:main} for each $\delta}%{\Delta t = \mathrm{Gap} \times \Delta t$.
Figure~\ref{fig:TVDPDF_LangLocal1D}(a) shows the TVD of the five schemes with time gaps $\mathrm{Gap}\in \{10, 20, 40, 80, 120, 160,200\}$.
Note that for every scheme, the TVD first decreases and then increases, reaching the smallest TVD when $\mathrm{Gap}=80$. This suggests that when the gap is small, the approximation error of the flow map (recall that the data are from an implicit scheme while the inferred schemes are explicit schemes) dominates the error in the invariant measure; when the gap is large, the numerical error of the inferred schemes dominates the TVD. A balance between the two errors is reached at the medium large time-step.
We first select the scheme that reproduces the invariant density with the smallest TVD. Overall, the IS-RK4 schemes perform the best and the inclusion of $c_0$ brings in negligible improvement. Thus, we select IS-RK4 without $c_0$ to demonstrate further results.
\begin{figure}[htp!]
\centering \vspace{-1mm}
\subfigure[TVD]{\includegraphics[width=0.30\textwidth]{./figures/TVD_LangLocal1D.pdf}}
\subfigure[PDF]{\includegraphics[width=0.32\textwidth]{./figures/PDF_LangLocal1D.pdf}}
\subfigure[ACF]{\includegraphics[width=0.32\textwidth]{./figures/ACF_LangLocal1D.pdf}}
\vspace{-3mm} \caption{Large-time statistics for 1D double-well potential. (a) TVD between the empirical invariant densities (PDF) of the inferred schemes and the reference PDF from data. (b) and (c): PDFs and ACFs comparison between the IS-RK4 with $c_0$ excluded and the reference data.
\label{fig:TVDPDF_LangLocal1D}}
\end{figure}
Figure~\ref{fig:TVDPDF_LangLocal1D} (b-c) show the PDFs and auto-correlation functions (ACFs) of IS-RK4 with $c_0$ excluded at three representative time gaps $\mathrm{Gap}\in\{10,80,200\}$, in comparison with those of the reference data and the plain RK4 with $\mathrm{Gap}=10$.
When $\mathrm{Gap}$ is small, that is $\mathrm{Gap}=10$, the IS-RK4 is close to the plain RK4, and both produce PDFs and ACFs with large errors. The PDF and ACF generated by IS-RK4 with $\mathrm{Gap}=80$ is the best among all used gaps, fitting the true PDF and ACF almost perfectly. Furthermore, when time gap is as large as $\mathrm{Gap}=200$, the IS-RK4 can still produce qualitative results with the feature of PDF (that is the double-well feature), whereas the plain RK4 scheme blows up when $\mathrm{Gap}=20$.
\begin{figure}[htp!]
\centering \vspace{-1mm}
\subfigure[Convergence of $c_1$ in sample size]{
\includegraphics[width=0.39\textwidth]{./figures/DW_conv_TM_RK4.pdf}
}
\subfigure[Coefficients and residuals]{
\includegraphics[width=0.58 \textwidth]{./figures/coeff_res_1DdoubleWell.pdf}}
\vspace{-3mm} \caption{1D double-well potential: Convergence of estimators in IS-RK4 with $c_0$ excluded. (a) The relative error of the estimator $\widehat{c_1^{\delta}%{\Delta t, N,M}}$ with $\delta}%{\Delta t = 80\times \Delta t$ converges at an order about $(MN)^{-1/2}$, matching Theorem \ref{thm_convEst}. (b) Left column: The coefficients depend on the time-step $\delta}%{\Delta t = \mathrm{Gap}\times \Delta t$, with $c_1$ being almost 1 and $c_2$ being close to linear in $\delta}%{\Delta t$ until $\delta}%{\Delta t>0.08$. The error bars, which are too narrow to be seen, are the standard deviations of the single-trajectory estimators from the $M$-trajectory estimator. Right column: The residual decays at an order $O(\delta}%{\Delta t^{1/2})$,
matching Theorem \ref{thm:order_res}. }
\label{fig:conv_LangLocal1D}
\end{figure}
We also test the convergence of the estimators in sample size and their dependence on the time-step, as well as the order of residual, aiming to confirm the theory in Section \ref{sec:conv}. Figure \ref{fig:conv_LangLocal1D}(a) shows that the relative error of $\widehat{c_1^{\delta}%{\Delta t, N,M}}$ converges at a rate about $(MN)^{-1/2}$ as the sample size $N$ or $M$ increases. Here we take the estimator from the largest sample size as the projection coefficient, and compute the relative error to it. Note that the estimator of $c_1$ is close to 1. Thus, the estimator $\widehat{c_1^{\delta}%{\Delta t, N,M}}$ converges at a rate about $(MN)^{-1/2}$, matching Theorem \ref{thm_convEst}. The convergence of the estimator of $c_2$ has similar convergence.
Figure \ref{fig:conv_LangLocal1D}(b) shows the dependence of the estimators on the time-step $\delta}%{\Delta t = \mathrm{Gap}\times \Delta t$. The coefficient $c_1$ is almost 1, while $c_2$ is close to linear in $\delta}%{\Delta t$. Furthermore, it also shows that the estimators from each single trajectory are close to the M-trajectory estimator, with small standard deviations represented by error bars that are too narrow to be seen. The residual decays at an order about $0.49$ with respect to $\delta}%{\Delta t$, closely matching the rate in Theorem~\ref{thm:order_res}.
\subsection{A 2D gradient system}
We now consider a 2D dissipative gradient system \cite{MSH02}
\begin{equation}\label{example:GradCouple}
d\mathbf{X}_t=-\nabla V(\mathbf{X}_t)dt+\sqrt{2/\beta} d\mathbf{B}_t,
\end{equation}
with $V(\mathbf{X})=V(x_1, x_2)=\exp\left(\frac{\mu_1}{2}x_1^2+\frac{\mu_2}{2} x_2^2\right)$.
The corresponding invariant measure is $\frac{1}{Z}exp^{-\beta V(x_1,x_2)}$
where $Z$ being the normalizing constant $Z:=\int_{\mathbb{R}^2}exp^{-\beta V(x_1,x_2)}dx_1dx_2$. We set $\mu_1=0.1$, $\mu_2=1$ and $\beta=2$. Because $\mu_2=10 \mu_1$, so $x_1$ is a slowly evolving variable compared to $x_2$ and the resulting dynamics displays a multi-scale feature. Consequently, we estimate parameters of the inferred schemes entry-wisely and we focus on the marginal invariant density of $x_1$.
We generate data by the SSBE scheme with $\Delta t=2e-3$ and time interval $[0, 2000] $ with total time steps $tN=1e6$. The rest setting and procedure are the same as the 1D double-well potential case.
Figure~\ref{fig:TVDPDF_GradCouple}(a) shows that IS-RK4 and IS-SSBE schemes have comparable TVD, and they reach the minimal TVD when $\mathrm{Gap} = 120$, where IS-EM blows put. They produce similar PDFs and ACFs, so we only present those of IS-SSBE with $c_0$ excluded. Figure~\ref{fig:TVDPDF_GradCouple}(b-c) show the PDFs and ACFs at representative time gaps $\mathrm{Gap}\in \{10, 80, 120, 200\}$. The findings are similar to those for the 1D double-well potential: (i) the performance of IS-SSBE first improves and then deteriorates as $\mathrm{Gap}$ increases; (ii) IS-SSBE can tolerate significantly larger time-step than the plain SSBE, {where the plain SSBE blows up due to the Newton-Raphson method used as the implicit solver,} {which can only tolerate a small time-step limited by the inversion (similar to \eqref{phi1_BEs}) in the Newton-Raphson method in the implicit solver.
}
\begin{figure}[htp!]
\centering \vspace{-1mm}
\subfigure[TVD]{\includegraphics[width=0.30\textwidth,height= 4cm]{./figures/TVD_GradCouple.pdf}}
\subfigure[PDF]{\includegraphics[width=0.32\textwidth,height= 4cm]{./figures/PDF_GradCouple.pdf}}
\subfigure[ACF]{\includegraphics[width=0.32\textwidth,height= 4cm]{./figures/ACF_GradCouple.pdf}}
\vspace{-3mm} \caption{Large-time statistics for the 2D gradient system. (a) TVD between the $x_1$ marginal invariant densities (PDF) of the inferred schemes and the reference PDF from data. (b) and (c): PDFs and ACFs comparison between IS-SSBE with $c_0$ excluded and the reference data. }
\label{fig:TVDPDF_GradCouple}
\end{figure}
The convergence of the estimators in sample size is also roughly of order $(MN)^{-1/2}$, as shown in Figure \ref{fig:conv_GradCouple}(a). Figure \ref{fig:conv_GradCouple}(b) shows that the estimators of $c_1$ and $c_2$ depend almost linearly on $\delta}%{\Delta t$. Also, $c_1$'s single-trajectory estimators have negligible standard deviations from the $M$-trajectory estimator, while $c_2$'s estimators have a persistent noticeable standard deviation. This suggests that IS-SSBE has large uncertainties in the stochastic force term (recall that $c_1$ and $c_2$ being the coefficients of the scaled drift and the stochastic force, see \eqref{eq:IS_Euler}). In the right column, the residual of IS-SSBE remains little changed when $\delta}%{\Delta t$ decreases, far from a decay rate $0.5$. This does not violate Theorem \ref{thm:order_res}, which is for parametrizations of explicit schemes. Instead, this highlights that the IS-SSBE is not a parametrization of the SSBE implicit scheme, and it has a flow map $\widetilde F_{SSBE}^{\delta}%{\Delta t}$ with distance to the true flow map $\mathbb{E} \big[|\frac{\mX_{t_{n+1}} -\mX_{t_{n}}}{\delta}%{\Delta t} - \widetilde F_{SSBE}^{\delta}%{\Delta t} (c, \mX_{t_i},\Delta {\bf B}_{t_n})|^2 \big]$ depending little on $\delta}%{\Delta t$. Such a feature may be helpful for further efforts on improving the parametric form.
\begin{figure}[htp!]
\centering \vspace{-1mm}
\subfigure[Convergence of $c_1$ in sample size]{
\includegraphics[width=0.38\textwidth]{./figures/2D_conv_TM_SSBE.pdf}}
\subfigure[Coefficients and residuals]{
\includegraphics[width=0.58 \textwidth]{./figures/coeff_res_2D_SSBE.pdf}}
\vspace{-3mm} \caption{2D gradient system: Convergence of estimators in IS-SSBE with $c_0$ excluded. (a) The relative error of the estimator $\widehat{c_1^{\delta}%{\Delta t, N,M}}$ with $\delta}%{\Delta t = 120 \Delta t$ converges at an order about $(MN)^{-1/2}$, matching Theorem \ref{thm_convEst}. (b) Left column: The estimators of $c_1, c_2$ are almost linear in $\delta}%{\Delta t$. Right column: The residual changes little as $\delta}%{\Delta t$ decreases, due to that IS-SSBE is not a parametrization of an explicit scheme (thus, Theorem~\ref{thm:order_res} does not apply).
\label{fig:conv_GradCouple} }
\end{figure}
Figure~\ref{fig:conv_GradCouple_RK4} shows the convergence of the estimator for IS-RK4. Similar to the 1D case, we observe a convergence rate $(MN)^{-1/2}$ in Figure~\ref{fig:conv_GradCouple_RK4}(a). Also, in Figure~\ref{fig:conv_GradCouple_RK4}(b), we observe almost $\delta}%{\Delta t$ independent estimators and the expected decay rate $O(\delta}%{\Delta t^{1/2})$ proved in Theorem \ref{thm:order_res}.
\begin{figure}[htp!]
\centering \vspace{-1mm}
\subfigure[Convergence of $c_1$ in sample size]{
\includegraphics[width=0.38\textwidth]{./figures/2D_conv_TM_RK4.pdf}}
\subfigure[Coefficients and residuals]{
\includegraphics[width=0.58 \textwidth]{./figures/coeff_res_2D_rk4.pdf}}
\vspace{-3mm} \caption{2D gradient system: Convergence of estimators in IS-RK4 with $c_0$ excluded. (a) The relative error of the estimator $\widehat{c_1^{\delta}%{\Delta t, N,M}}$ with $\delta}%{\Delta t = 120 \Delta t$ converges at an order about $(MN)^{-1/2}$, matching Theorem \ref{thm_convEst}. (b) Left column: The estimators of $c_1, c_2$ are constant for all $\delta}%{\Delta t$. Right column: The residual decays at an order $O(\delta}%{\Delta t^{1/2})$,
matching Theorem \ref{thm:order_res}.
\label{fig:conv_GradCouple_RK4} }
\end{figure}
\subsection{Stochastic Lorenz system with degenerate noise}
Consider next the 3D stochastic Lorenz system with degenerate noise \cite{MSH02}
\begin{equation}\label{example:Lorenz}
\begin{aligned}
dx_1 & =\sigma(x_2-x_1)dt+\sqrt{2/\beta}dB_1,\\
dx_2 & =\big( x_1(\gamma-x_3)-x_2\big)dt+\sqrt{2/\beta}dB_2,\\
dx_3& =(x_1x_2-bx_3)dt.
\end{aligned}
\end{equation}
We set $\sigma=10$, $\gamma=28$, $b=8/3$ and $\beta=1$.
This stochastic chaotic system is exponentially ergodic with a regular invariant measure because it is dissipative and hypoelliptic.
As before, we generate data by SSBE with $\Delta t=5e-4$ and a reference long trajectory with $tN=6e6$ time steps (or equivalently, on the time interval $[0,3000]$). We consider time gaps $\mathrm{Gap}\in \{20, 40, 80, 160, 240, 320,400\}$, so the maximal time-step is still $0.2$.
\begin{figure}[htp!]
\centering \vspace{-1mm}
\subfigure[TVD]{\includegraphics[width=0.30\textwidth,height= 4cm]{./figures/TVD_Lorenz_x1.pdf}}
\subfigure[PDF]{\includegraphics[width=0.32\textwidth,height= 4cm]{./figures/PDF_Lorenz_x1.pdf}}
\subfigure[ACF]{\includegraphics[width=0.32\textwidth,height= 4cm]{./figures/ACF_Lorenz_x1.pdf}}
\vspace{-3mm} \caption{Large-time statistics of $x_1$ for the stochastic Lorenz system. (a) TVD between the $x_1$ marginal invariant densities (PDF) of the inferred schemes and the reference PDF from data. (b) and (c): PDFs and ACFs comparison between IS-RK4 with $c_0$ included and the reference data.}
\label{fig:TVDPDF_Lorenz}
\end{figure}
Figure~\ref{fig:TVDPDF_Lorenz}(a) shows the TVD of the inferred schemes. This time, the IS-RK4 scheme performs significantly better than IS-SSBE schemes,
with relatively small TVD for most time gaps. This is due to the high-order approximation of RK4 to the drift, particularly when the drift dominates the dynamics (note that the state variable $x_1$ is at a scale of magnitude larger than the degenerate noise). The IS-RK4 with $c_0$ included performs the best and we select it for further demonstration of results.
Figure~\ref{fig:TVDPDF_Lorenz}(b-c) show the PDFs and ACFs at representative time gaps $\mathrm{Gap}\in \{20, 240, 320\}$. Since the plain RK4 blows up at $\mathrm{Gap}=10$, so we display the results from IS-EM instead. The findings are similar to those for the 1D double-well potential: (i) the performance of IS-RK4 first improves and then deteriorates as $\mathrm{Gap}$ increases; (ii) IS-RK4 can tolerate significantly larger time-step than the plain RK4.
\begin{figure}[htp!]
\centering \vspace{-1mm}
\subfigure[PDF]{\includegraphics[width=0.32\textwidth,height= 4.3cm]{./figures/PDF_Lorenz_x3.pdf}}
\subfigure[ACF]{\includegraphics[width=0.32\textwidth,height= 4.3cm]{./figures/ACF_Lorenz_x3.pdf}}
\vspace{-3mm} \caption{ACF and PDF of $x_3$ in the stochastic Lorenz system. Similar to the other examples, IS-RK4 (with $c_0$ included) reproduces the PDF and the ACF the best when the time-step is medium large, while plain RK4 and IS-EM blow up even when $\mathrm{Gap}=20$.}
\label{fig:TVDPDF_Lorenz_x3}
\end{figure}
Moreover, we also plot the PDF and ACF of $x_3$ in Figure~\ref{fig:TVDPDF_Lorenz_x3}. The dynamics of $x_3$ is the most challenging because there is no diffusive stochastic force acting on it and its ACF is highly oscillatory. As usual, the IS-RK4 can reproduce the PDF and ACF well, whereas the plain RK4 and IS-EM blow up even when the time-step is small. In particular, the IS-RK4 produce the periodic and decay feature of $x_3$'s ACF when the time-step is medium large, that is $\mathrm{Gap}=240$. We expect the best performance to be achieved at a gap between 120 to 240, and we postpone the study on the optimal time gap and other improvements in future work.
The IS-RK4 has convergence results mostly as expected. Figure~\ref{fig:conv_res_Lorenz_RK4}(a) shows that the estimators of $c_1$ for each entry of $(x_1,x_2,x_3)$ converge at an almost perfect rate $(NM)^{-1/2}$. Figure~\ref{fig:conv_res_Lorenz_RK4}(b) shows that the estimator of $c_0,c_1,c_2$ remain little varied until $\delta}%{\Delta t=0.12$ (i.e., $\mathrm{Gap}>240$) for each entry. It also shows that the residuals of all three entries decay at a rate slightly higher than $O(\delta}%{\Delta t^{1/2})$.
\begin{figure}[htp!]
\centering \vspace{-1mm}
\subfigure[Convergence of $c_1$ in sample size]{
\includegraphics[width=0.37\textwidth]{./figures/Lorenz_conv_TM_RK4Gap2to400_useNoise1YnOFF_M1000.pdf}}
\subfigure[Coefficients and residuals]{
\includegraphics[width=0.59 \textwidth]{./figures/coeff_res_Lorenz.pdf}}
\vspace{-3mm} \caption{The 3D stochastic Lorenz system: Convergence of estimators in IS-RK4 with $c_0$ included. (a) The relative error of the estimator $\widehat{c_1^{\delta}%{\Delta t, N,M}}$ with $\delta}%{\Delta t = 240 \Delta t = 0.12$ converges at order about $(MN)^{-1/2}$, matching Theorem \ref{thm_convEst}. (b) Left column: The estimators of $c_0,c_1, c_2$ are varies little until $\delta}%{\Delta t>0.12$. The vertical dash line is the optimal time gap. Right column: The residuals decay at orders slightly higher than $O(\delta}%{\Delta t^{1/2})$.
\label{fig:conv_res_Lorenz_RK4} }
\end{figure}
\section{Conclusions and outlook}
We have introduced a framework to infer schemes adaptive to large time-stepping (ISALT) from data for locally Lipschitz ergodic SDEs. We formulate it as a statistical learning problem, in which we learn an approximation to the infinite-dimensional discrete-time flow map. By deriving informed basis functions from classical numerical schemes, we obtain a low-dimensional parameter estimation problem, avoiding the curse of dimensionality in statistical learning.
Under mild conditions, we show that the estimator converges as the data size increases, and the inferred scheme has same a 1-step strong order as the explicit scheme it parametrizes. Thus, our algorithm comes with performance guarantee. Numerical tests on three non-globally Lipschitz examples confirm the theory. The inferred scheme can tolerate large time-steps and efficiently and accurately simulate the invariant measure.
Many fronts are left open for further investigation. (1) The optimal time-step. We have observed that the inferred schemes perform the best (producing the most accurate invariant measure) when the time-step is medium-large. This observation suggests a trade-off between the 1-step approximation error of the flow map and the accumulated numerical error in the invariant measure. Similar optimality in the medium range was observed in space-time model reduction \cite{Lu20Burgers} and in parameter estimation for multiscale diffusion \cite{PS07}. It is crucial to have a universal \emph{a priori} estimate on the optimal time-step, which can guide all data-driven model reduction approaches. (2) Multi-step noise. We focused on approximate flow maps that use only the increments of the Brownian motion. This limits the performance of the inferred scheme because we omit the details of the stochastic force. Thus, a multi-step noise provides the necessary information for further improvements, particularly when the noise is non-stationary \cite{Li_Duan21}. (3) Non-ergodic systems and/or space-time reduction. We expect to extend the framework of ISALT to simulate non-ergodic systems or achieve space-time reduction for high-dimensional nonlinear systems by extracting informed basis functions from the classical numerical scheme.
\bigskip
\noindent \textbf{Acknowledgments
FL is grateful for supports from NSF-DMS 1913243 and NSF-DMS 1821211. XL is grateful for supports from NSF DMS CAREER-1847770. FY is grateful for supports from AMS-Simons travel grants. FL would like to thank Kevin Lin for helpful discussions.
\bibliographystyle{alpha}
\let\oldbibliography\thebibliography
\renewcommand{\thebibliography}[1]{%
\oldbibliography{#1}%
\setlength{\itemsep}{0pt}%
}
|
\section{Introduction}
\label{sec:intro}
Let $\mathcal{P}$ be a set of $h$ pairwise disjoint polygonal obstacles with
a total of $n$ vertices in the plane. Let $\mathcal{F}$ denote the {\em free space}, i.e., the plane
minus the interior of the obstacles. Given two points $s$ and $t$ in $\mathcal{F}$, we consider the problem of
finding a Euclidean shortest path from $s$ to $t$ in $\mathcal{F}$. This is a
classical problem in computational geometry and has been studied
extensively,
e.g.,~\cite{ref:ChenCo13,ref:GhoshAn91,ref:HershbergerAn99,ref:MitchellA91,ref:MitchellSh96,ref:SharirOn86,ref:StorerSh94,ref:RohnertSh86,ref:GuibasOp89,ref:GuibasLi87,ref:HershbergerA91,ref:HershbergerCo94,ref:LeeEu84,ref:WangSh21}.
To solve the problem, two
methods are often used in the literature: the visibility graph and the continuous
Dijkstra. The visibility graph method is to first construct the
visibility graph of the vertices of $\mathcal{P}$ along with $s$ and $t$,
and then run Dijkstra's shortest path algorithm on the graph to find a
shortest $s$-$t$\ path. The best algorithms for constructing the
visibility graph run in $O(n\log n+K)$ time~\cite{ref:GhoshAn91} or in
$O(n+h\log^{1+\epsilon}h +K)$ time~\cite{ref:ChenA15} for any constant
$\epsilon>0$, where $K$ is the number of edges of the visibility graph. Because
$K=\Omega(n^2)$ in the worst case, the visibility graph method
inherently takes quadratic time. To deal with the case where $h$ is relatively small
comparing to $n$, a variation of the
visibility graph method was proposed that is to first construct a
so-called {\em tangent graph} and then find a shortest $s$-$t$\ path in
the graph. Using this method, a shortest $s$-$t$\ path can be found in $O(n+h\log h+K')$ time~\cite{ref:ChenCo15} after the free space $\mathcal{F}$ is triangulated, where $K'$ may be considered as the number of tangents among obstacles of
$\mathcal{P}$ and $K'=O(h^2)$. Note that triangulating $\mathcal{F}$ can be done
in $O(n\log n)$ time or in $O(n+h\log^{1+\epsilon}h)$ time for any small
$\epsilon>0$~\cite{ref:Bar-YehudaTr94}. Hence, the running time of
the above algorithm in~\cite{ref:ChenCo15} is still quadratic in the worst case.
Using the continuous Dijkstra method, Mitchell~\cite{ref:MitchellSh96} made a breakthrough and achieved the first subquadratic algorithm of $O(n^{3/2+\epsilon})$ time for any constant $\epsilon>0$. Also using the continuous Dijkstra approach plus a novel conforming subdivision of the free space, Hershberger and Suri~\cite{ref:HershbergerAn99} presented an algorithm of $O(n\log n)$ time and $O(n\log n)$ space; the running time is optimal when $h=\Theta(n)$ as $\Omega(n+h\log h)$ is a lower bound in the algebraic computation tree model (which can be obtained by a reduction from sorting; e.g., see Theorem~3~\cite{ref:deRezendeRe89} for a similar reduction). Recently, by modifying Hershberger and Suri's algorithm, Wang~\cite{ref:WangSh21} reduced the space to $O(n)$ while the running time is still $O(n\log n)$.\footnote{An unrefereed report
\cite{ref:InkuluA10} announced an algorithm based on the
continuous Dijkstra approach with $O(n+h\log h\log n)$ time and $O(n)$ space.}
All three continuous Dijkstra algorithms~\cite{ref:HershbergerAn99,ref:MitchellSh96,ref:WangSh21}
construct the {\em shortest path map}, denoted by $\mbox{$S\!P\!M$}(s)$,
for a source point $s$. $\mbox{$S\!P\!M$}(s)$ is of $O(n)$ size and can be used to
answer shortest path queries. By building a point location data
structure on $\mbox{$S\!P\!M$}(s)$ in additional $O(n)$ time~\cite{ref:EdelsbrunnerOp86,ref:KirkpatrickOp83}, given a query point $t$, the
shortest path length from $s$ to $t$ can be computed in $O(\log
n)$ time and a shortest $s$-$t$\ path can be output in time linear
in the number of edges of the path.
The problem setting for $\mathcal{P}$ is usually referred to as {\em
polygonal domains} or {\em polygons with holes} in the literature. The
problem in simple polygons is relatively
easier~\cite{ref:GuibasOp89,ref:GuibasLi87,ref:HershbergerA91,ref:HershbergerCo94,ref:LeeEu84}.
Guibas et al.~\cite{ref:GuibasLi87} presented an algorithm that can
construct a shortest path map in
linear time. For two-point shortest path query problem where both $s$
and $t$ are query points, Guibas and
Hershberger~\cite{ref:GuibasOp89,ref:HershbergerA91} built a data
structure in linear time such that each query can be answered in
$O(\log n)$ time. In contrast, the two-point query problem in
polygonal domains is much more challenging: to achieve $O(\log n)$
time queries, the current best result uses $O(n^{11})$
space~\cite{ref:ChiangTw99}; alternatively Chiang and
Mitchell~\cite{ref:ChiangTw99} gave a data structure of $O(n+h^5)$
space with $O(h\log n)$ query time. Refer to~\cite{ref:ChiangTw99} for
other data structures with trade-off between space and query time.
The $L_1$ counterpart of the problem where the path length is measured
in the $L_1$ metric also attracted much attention, e.g.,
\cite{ref:BaeL119,ref:ClarksonRe87,ref:ClarksonRe88,ref:MitchellAn89,ref:MitchellL192,ref:ChenCo19}.
For polygons with holes,
Mitchell~\cite{ref:MitchellAn89,ref:MitchellL192} gave an algorithm
that can build a shortest path map for a source point in
$O(n\log n)$ time and $O(n)$ space; for small $h$, Chen and Wang~\cite{ref:ChenCo19}
proposed an algorithm of $O(n+h\log
h)$ time and $O(n)$ space, after the free space is triangulated.
For simple polygons, Bae and Wang~\cite{ref:BaeL119} built a data
structure in linear time that can answer each two-point $L_1$ shortest path
query in $O(\log n)$ time. The two-point query problem in polygons with holes has also been studied~\cite{ref:ChenTw16,ref:ChenSh00,ref:WangA20}. To achieve $O(\log n)$ time queries, the current best result uses $O(n+h^2\log^3h/\log\log h)$ space~\cite{ref:WangA20}.
\subsection{Our result}
In this paper, we show that the problem of finding an Euclidean shortest path among obstacles in $\mathcal{P}$ is solvable in $O(n+h\log h)$ time and $O(n)$ space, after a triangulation of the free space $\mathcal{F}$ is given.
If the time for triangulating $\mathcal{F}$ is included and the triangulation algorithm in~\cite{ref:Bar-YehudaTr94} is used, then the total time of the algorithm is $O(n+h\log^{1+\epsilon}h)$, for any constant $\epsilon>0$.\footnote{If randomization is allowed, the algorithm of Clarkson, Cole, and Tarjan~\cite{ref:ClarksonRa92} can compute a triangulation in $O(n\log^* n+h\log h)$ expected time.}
With the assumption that the triangulation could be done in $O(n+h\log h)$ time, which has been an open problem and is beyond the scope of this paper, our result settles Problem 21 in The Open Problem Project~\cite{ref:TOPP}. Our algorithm actually constructs the shortest path map $\mbox{$S\!P\!M$}(s)$ for the source point $s$ in $O(n+h\log h)$ time and $O(n)$ space. We give an overview of our approach below.
The high-level scheme of our algorithm is similar to that for the $L_1$ case~\cite{ref:WangA20} in the sense that we first solve the {\em convex case} where all obstacles of $\mathcal{P}$ are convex and then extend the algorithm to the general case with the help of the extended corridor structure of $\mathcal{P}$~\cite{ref:ChenCo19,ref:ChenTw16,ref:ChenCo17,ref:ChenA15,ref:KapoorAn97,ref:MitchellSe95}.
\paragraph{The convex case.}
We first discuss the convex case. Let $\mathcal{V}$ denote the set of topmost, bottommost, leftmost, and rightmost vertices of all obstacles. Hence, $|\mathcal{V}|\leq 4h$. Using the algorithm of Hershberger and Suri~\cite{ref:HershbergerAn99}, we build a conforming subdivision $\mathcal{S}$ on the points of $\mathcal{V}$, without considering the obstacle edges. Since $|\mathcal{V}|=O(h)$, the size of $\mathcal{S}$ is $O(h)$. Then, we insert the obstacle edges into $\mathcal{S}$ to build a conforming subdivision $\mathcal{S}'$ of the free space.
The subdivision $\mathcal{S}'$ has $O(h)$ cells (in contrast, the conforming subdivision of the free space in~\cite{ref:HershbergerAn99} has $O(n)$ cells).
Unlike the subdivision in~\cite{ref:HershbergerAn99} where each cell is of constant size, here the size of each cell of $\mathcal{S}'$ may not be constant but its boundary consists of $O(1)$ transparent edges and $O(1)$ convex chains (each of which belongs to the boundary of an obstacle of $\mathcal{P}$). Like the subdivision in~\cite{ref:HershbergerAn99}, each transparent edge $e$ of $\mathcal{S}'$ has a well-covering region $\mathcal{U}(e)$. In particular, for each transparent edge $f$ on the boundary of $\mathcal{U}(e)$, the shortest path distance between $e$ and $f$ is at least $2\cdot\max\{|e|,|f|\}$. Using $\mathcal{S}'$ as a guidance, we run the continuous Dijkstra algorithm as in~\cite{ref:HershbergerAn99} to expand the wavefront, starting from the source point $s$.
A main challenge our algorithm needs to overcome (which is also a main difference between our algorithm and that in~\cite{ref:HershbergerAn99}) is that each cell in our subdivision $\mathcal{S}'$ may not be of constant size. One critical property our algorithm relies on is that the boundary of each cell of $\mathcal{S}'$ has $O(1)$ convex chains. Our strategy is to somehow treat each such convex chain as a whole. We also borrow some idea from the algorithm of Hershberger, Suri, and Y{\i}ld{\i}z~\cite{ref:HershbergerA13} for computing shortest paths among curved obstacles. To guarantee the $O(n+h\log h)$ time, some global charging analysis is used. In addition, the tentative prune-and-search technique of Kirkpatrick and Snoeyink~\cite{ref:KirkpatrickTe95} is applied to perform certain operations related to bisectors, in logarithmic time each. Finally, the techniques of Wang~\cite{ref:WangSh21} are utilized to reduce the space to $O(n)$. All these efforts lead to an $O(n+h\log h)$ time and $O(n)$ space algorithm to construct the shortest path map $\mbox{$S\!P\!M$}(s)$ for the convex case.
\paragraph{The general case.}
We extend the convex case algorithm to the general case where obstacles may not be convex. To this end, we resort to the extended corridor structure of $\mathcal{P}$, which was used before for reducing the time complexities from $n$ to $h$, e.g.,~\cite{ref:ChenCo19,ref:ChenTw16,ref:ChenCo17,ref:ChenA15,ref:KapoorAn97,ref:MitchellSe95}. The structure partitions the free space $\mathcal{F}$ into an ocean $\mathcal{M}$, $O(n)$ bays, and $O(h)$ canals.
Each bay is a simple polygon that shares an edge with $\mathcal{M}$. Each canal is a simple polygon that shares two edges with $\mathcal{M}$. But two bays or two canals, or a bay and a canal do not share any edge. A common edge of a bay (or canal) with $\mathcal{M}$ is called a gate. Thus each bay has one gate and each canal has two gates. Further,
$\mathcal{M}$ is bounded by $O(h)$ convex chains (each of which is on the boundary of an obstacle).
An important property related to shortest paths is that if both $s$ and $t$ are in $\mathcal{M}$, then any shortest $s$-$t$\ path must be in the union of $\mathcal{M}$ and all corridor paths, each of which is contained in a canal.
As the boundary of $\mathcal{M}$ consists of $O(h)$ convex chains, by incorporating all corridor paths, we can easily extend our convex case algorithm to computing $\mbox{$S\!P\!M$}(\mathcal{M})$,
the shortest path map $\mbox{$S\!P\!M$}(s)$ restricted to $\mathcal{M}$, i.e.,
$\mbox{$S\!P\!M$}(\mathcal{M})=\mbox{$S\!P\!M$}(s)\cap \mathcal{M}$. To compute the entire map $\mbox{$S\!P\!M$}(s)$,
we expand $\mbox{$S\!P\!M$}(\mathcal{M})$ to all bays and canals through their gates.
For this, we process each bay/canal individually. For each
bay/canal $C$, expanding the map into $C$ is actually a special case
of the (additively)-weighted geodesic Voronoi diagram problem on a
simple polygon where all sites are outside $C$ and can influence $C$
only through its gates. In summary, after a triangulation of $\mathcal{F}$
is given, building $\mbox{$S\!P\!M$}(\mathcal{M})$ takes $O(n+h\log h)$ time, and
expanding $\mbox{$S\!P\!M$}(\mathcal{M})$ to all bays and canals takes additional
$O(n+h\log h)$ time. The space of the algorithm is bounded by $O(n)$.
\paragraph{Outline.}
The rest of the paper is organized as follows. Section~\ref{sec:pre}
defines notation and introduces some concepts. Section~\ref{sec:convex} presents the algorithm for the convex case.
The general case is discussed in Section~\ref{sec:general}.
\section{Preliminaries}
\label{sec:pre}
For any two points $a$ and $b$ in the plane, denote by $\overline{ab}$
the line segment with $a$ and $b$ as endpoints; denote by
$|\overline{ab}|$ the length of the segment.
For any two points $s$ and $t$ in the free space $\mathcal{F}$, we use
$\pi(s,t)$ to denote a shortest path from $s$ to $t$ in $\mathcal{F}$. In the case
where shortest paths are not unique, $\pi(s,t)$ may refer to an
arbitrary one. Denote by $d(s,t)$ the length of $\pi(s,t)$;
we call $d(s,t)$ the {\em geodesic distance} between $s$ and $t$. For
two line segments $e$ and $f$ in $\mathcal{F}$, their {\em geodesic
distance} is defined to be the minimum geodesic distance between any
point on $e$ and any point on $f$, i.e., $\min_{s\in e, t\in
f}d(s,t)$; by slightly abusing the notation, we use $d(e,f)$ to denote
their geodesic distance.
For any path $\pi$ in the plane, we use $|\pi|$ to denote its
length.
For any compact region $A$ in the plane, let $\partial A$ denote its boundary. We use $\partial \mathcal{P}$ to denote the union of the boundaries of all obstacles of $\mathcal{P}$.
Throughout the paper, we use $s$ to refer to the source point. For
convenience, we consider $s$ as a degenerate obstacle in $\mathcal{P}$. We
often refer to the vertices of $\mathcal{P}$ as {\em obstacle vertices} and
refer to the edges of $\mathcal{P}$ as {\em obstacle edges}.
For any point $t\in \mathcal{F}$, we call the adjacent vertex of $t$ in $\pi(s,t)$ the {\em anchor} of $t$ in $\pi(s,t)$\footnote{Usually ``predecessor'' is used in the literature instead of ``anchor'', but here we reserve ``predecessor'' for other purpose.}.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=2.5in]{spm.pdf}
\caption{\footnotesize Illustrating the shortest path map $\mbox{$S\!P\!M$}(s)$. The solid (red) curves are walls and the
(blue) dotted segments are windows. The anchor of each cell is also shown with a black point.}
\label{fig:spm}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
The {\em shortest path map} $\mbox{$S\!P\!M$}(s)$ of $s$ is a decomposition of the free space $\mathcal{F}$ into maximal regions such that all points in each region $R$ have the same anchor~\cite{ref:HershbergerAn99,ref:MitchellA91} in their shortest paths from $s$; e.g., see Fig.~\ref{fig:spm}. Each edge of $\mbox{$S\!P\!M$}(s)$ is either an obstacle edge fragment or a {\em bisecting-curve}\footnote{This is usually called {\em bisector} in the literature. Here we reserve the term ``bisector'' to be used later.}, which is the locus of points $p$ with $d(s,u)+|\overline{pu}|=d(s,v)+|\overline{pv}|$ for two obstacle vertices $u$ and $v$. Each bisecting-curve is in general a hyperbola; a special case happens if one of $u$ and $v$ is the anchor of the other, in which case their bisecting-curve is a straight line. Following the notation in~\cite{ref:Eriksson-BiqueGe15}, we differentiate between two types of bisecting curves: {\em walls} and {\em windows}. A bisecting curve of $\mbox{$S\!P\!M$}(s)$ is a {\em wall} if there exist two topologically different shortest paths from $s$ to each point of the edge; otherwise (i.e., the above special case) it is a {\em window} (e.g., see Fig.~\ref{fig:spm}).
We make a general position assumption that for each obstacle vertex $v$, there is a unique shortest path from $s$ to $v$, and for any point $p$ in the plane, there are at most three different shortest paths from $s$ to $p$. The assumption assures that each vertex of $\mbox{$S\!P\!M$}(s)$ has degree at most three, and there are at most three bisectors of $\mbox{$S\!P\!M$}(s)$ intersecting at a common point, which is sometimes called a {\em triple point} in the literature~\cite{ref:Eriksson-BiqueGe15}.
A curve in the plane is {\em $x$-monotone} if its intersection with any vertical line is connected; the {\em $y$-monotone} is defined similarly. A curve is $xy$-monotone if it is both $x$- and $y$-monotone.
The following observation will be used throughout the paper without explicitly mentioning it again.
\begin{observation}\label{obser:hlogh}
$n+h\log n=O(n+h\log h)$.
\end{observation}
\begin{proof}
Indeed, if $h<n/\log n$, then $n+h\log n=\Theta(n)$, which is $O(n+h\log h)$; otherwise, $\log n=O(\log h)$ and $n+h\log n=O(n+h\log h)$.
\end{proof}
\section{The convex case}
\label{sec:convex}
In this section, we present our algorithm for the convex case where all obstacles of $\mathcal{P}$ are convex. The algorithm will be extended to the general case in Section~\ref{sec:general}.
For each obstacle $P\in \mathcal{P}$, the topmost, bottommost, leftmost, and rightmost vertices of $P$ are called {\em rectilinear extreme vertices}. The four rectilinear extreme vertices partition $\partial P$ into four portions and each portion is called an {\em elementary chain}, which is convex and $xy$-monotone. For technical reason that will be clear later, we assume that each rectilinear extreme vertex $v$ belongs to the elementary chain counterclockwise of $v$ with respect to the obstacle (i.e., $v$ is the clockwise endpoint of the chain; e.g., see Fig.~\ref{fig:elechain}).
We use {\em elementary chain fragment} to refer to a portion of an elementary chain.
\begin{figure}[t]
\begin{minipage}[t]{0.46\textwidth}
\begin{center}
\includegraphics[height=1.4in]{elechain.pdf}
\caption{\footnotesize Illustrating an elementary chain (the thick segments), which contains the vertex $v$ but not $u$.}
\label{fig:elechain}
\end{center}
\end{minipage}
\begin{minipage}[t]{0.54\textwidth}
\begin{center}
\includegraphics[height=1.4in]{pred.pdf}
\caption{\footnotesize Illustrating the predecessor.}
\label{fig:pred}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
We introduce some notation below that is similar in spirit to those
from~\cite{ref:HershbergerA13} for shortest paths among curved
obstacles.
Consider a shortest path $\pi(s,p)$ from $s$ to a point $p$ in the free space $\mathcal{F}$.
It is not difficult to see that $\pi(s,p)$ is a sequence of
elementary chain fragments and common tangents between obstacles of
$\mathcal{P}\cup \{p\}$. We define the {\em predecessor} of $p$, denoted
by $pred(p)$, to be the initial vertex of the last elementary
chain fragment in $\pi(s,p)$ (e.g., see Fig.~\ref{fig:pred} (a)). Note that since each rectilinear extreme vertex belongs to a single elementary chain, $pred(p)$ in $\pi(s,p)$ is unique. A special case happens if $p$ is a rectilinear extreme vertex and $\pi(s,p)$ contains a portion of an elementary chain $A$ clockwise of $p$.
In this case, we let $pred(p)$ be endpoint of the fragment of $A$ in $\pi(s,p)$ other than $p$ (e.g., see Fig.~\ref{fig:pred}~(b)); in this way, $pred(p)$ is unique in $\pi(s,p)$. Note that $p$ may still have multiple
predecessors if there are multiple shortest paths from $s$ to $p$.
Intuitively, the reason we define predecessors as above is to treat each elementary chain somehow as a whole, which is essential for reducing the runtime of the algorithm from $n$ to $h$.
The rest of this section is organized as follows. In Section~\ref{sec:subdivision}, we compute a conforming subdivision $\mathcal{S}'$ of the free space $\mathcal{F}$. Section~\ref{sec:notation} introduces some basic concepts and notation for our algorithm. The wavefront expansion algorithm is presented in Section~\ref{sec:algo}, with two key subroutines of the algorithm described in Section~\ref{sec:merge} and Section~\ref{sec:propagation}, respectively. Section~\ref{sec:time} analyzes the time complexity of the algorithm, where a technical lemma is proved separately in Section~\ref{sec:lemnumgen}. Using the information computed by the wavefront expansion algorithm, Section~\ref{sec:spm} constructs the shortest path map $\mbox{$S\!P\!M$}(s)$. The overall algorithm runs in $O(n+h\log h)$ time and $O(n+h\log h)$ space. Section~\ref{sec:space} reduces the space to $O(n)$ while keeping the same runtime, by using the techniques from Wang~\cite{ref:WangSh21}.
\subsection{Computing a conforming subdivision of the free space}
\label{sec:subdivision}
Let $\mathcal{V}$ denote the set of the rectilinear extreme vertices of all
obstacles of $\mathcal{P}$. Hence, $|\mathcal{V}|=O(h)$. Using the algorithm of
algorithm of Hershberger and Suri~\cite{ref:HershbergerAn99} (called
the {\em HS} algorithm), we build a conforming subdivision $\mathcal{S}$
with respect to the vertices of $\mathcal{V}$, without considering the
obstacle edges.
The subdivision $\mathcal{S}$, which is of size $O(h)$, is a quad-tree-style subdivision of the plane into $O(h)$ cells. Each cell of $\mathcal{S}$ is a square or a square annulus (i.e., an outer square with an inner square hole). Each vertex of $\mathcal{V}$ is contained in the interior of a square cell and each square cell contains at most one vertex of $\mathcal{V}$.
Each edge $e$ of $\mathcal{S}$ is axis-parallel and {\em well-covered},
i.e., there exists a set $\mathcal{C}(e)$ of $O(1)$ cells of $\mathcal{S}$ such
that their union $\mathcal{U}(e)$ contains $e$ with the following
properties: (1) the total complexity of all cells of $\mathcal{C}(e)$ is
$O(1)$ and thus the size of $\mathcal{U}(e)$ is $O(1)$; (2) for any edge $f$
of $\mathcal{S}$ that is on $\partial \mathcal{U}(e)$ or outside $\mathcal{U}(e)$, the
Euclidean distance between $e$ and $f$ (i.e., the minimum
$|\overline{pq}|$ among all points $p\in e$ and $q\in f$) is at least
$2\cdot\max\{|e|,|f|\}$; (3) $\mathcal{U}(e)$, which is called the {\em
well-covering region} of $e$, contains at most one vertex of $\mathcal{V}$.
In addition, each cell $c$ of $\mathcal{S}$ has $O(1)$ edges on its boundary with the
following {\em uniform edge property}:
the lengths of the edges on the boundary of $c$ differ by at most a
factor of $4$, regardless of whether $c$ is a square or square annulus.
The subdivision $\mathcal{S}$ can be computed in $O(h\log h)$ time and $O(h)$ space~\cite{ref:HershbergerAn99}.
Next we insert the obstacle edges into $\mathcal{S}$ to produce a conforming
subdivision $\mathcal{S}'$ of the free space $\mathcal{F}$. In $\mathcal{S}'$, there are two
types of edges: those introduced by the subdivision construction
(which are in the interior of $\mathcal{F}$ except possibly their endpoints)
and the obstacle edges; we call the former the {\em transparent edges}
(which are axis-parallel) and the latter the {\em opaque edges}. The
definition of $\mathcal{S}'$ is similar to the conforming subdivision of the
free space used in the HS algorithm. A main difference is that here
endpoints of each obstacle edge may not be in $\mathcal{V}$, a consequence of which is that each cell of
$\mathcal{S}'$ may not be of constant size (while each cell in the
subdivision of the HS algorithm is of constant size). However, each cell $c$
of $\mathcal{S}'$ has the following property that is critical to our algorithm: The boundary $\partial c$ consists of $O(1)$ transparent edges and $O(1)$ convex chains (each of which is a portion of an elementary chain).
More specifically, $\mathcal{S}'$ is a subdivision of $\mathcal{F}$ into $O(h)$ cells. Each cell of $\mathcal{S}'$ is one of the connected components formed by intersecting $\mathcal{F}$ with an axis-parallel rectangle (which is the union of a set of adjacent cells of $\mathcal{S}$) or a square annulus of $\mathcal{S}$. Each cell of $\mathcal{S}'$ contains at most one vertex of $\mathcal{V}$. Each vertex of $\mathcal{V}$ is incident to a transparent edge. Each transparent edge $e$ of $\mathcal{S}'$ is {\em well-covered}, i.e., there exists a set $\mathcal{C}(e)$ of $O(1)$ cells whose union $\mathcal{U}(e)$ contains $e$ with the following property: for each transparent edge $f$ on $\partial \mathcal{U}(e)$, the geodesic distance $d(e,f)$ between $e$ and $f$ is at least $2\cdot\max\{|e|,|f|\}$. The region $\mathcal{U}(e)$ is called the {\em well-covering region} of $e$ and contains at most one vertex of $\mathcal{V}$. Note that $\mathcal{S}'$ has $O(h)$ transparent edges.
Below we show how $\mathcal{S}'$ is produced from $\mathcal{S}$. The procedure is similar to that in the HS algorithm. We overlay the obstacle edges on top of $\mathcal{S}$ to obtain a subdivision $\mathcal{S}_{overlay}$. Because each edge of $\mathcal{S}$ is axis-parallel and all obstacle edges constitute a total of $O(h)$ elementary chains, each of which is $xy$-monotone, $\mathcal{S}_{overlay}$ has $O(h^2)$ faces. We say that a face of $\mathcal{S}_{overlay}$ is {\em interesting} if its boundary contains a vertex of $\mathcal{V}$ or a vertex of $\mathcal{S}$. We keep intact the interesting faces of $\mathcal{S}_{overlay}$ while deleting every edge fragment of $\mathcal{S}$ not on the boundary of any interesting cell. Further, for each cell $c$ containing a vertex $v\in \mathcal{V}$, we partition $c$ by extending vertical edges from $v$ until the boundary of $c$. This divides $c$ into at most three subcells. Finally we divide each of the two added edges incident to $v$ into segments of length at most $\delta$, where $\delta$ is the length of the shortest edge on the boundary of $c$. By the uniform edge property of $\mathcal{S}$, $\partial c$ has $O(1)$ edges, whose lengths differ by at most a factor of $4$; hence dividing the edges incident to $v$ as above produces only $O(1)$ vertical edges. The resulting subdivision is $\mathcal{S}'$.
As mentioned above, the essential difference between our subdivision $\mathcal{S}'$ and the one in the HS algorithm is that the role of an obstacle edge in the HS algorithm is replaced by an elementary chain. Therefore, each opaque edge in the subdivision of the HS algorithm becomes an elementary chain fragment in our case. Hence, by the same analysis as in the HS algorithm (see Lemma 2.2~\cite{ref:HershbergerAn99}), $\mathcal{S}'$ has the properties as described above and the well-covering region $\mathcal{U}(e)$ of each transparent edge $e$ of $\mathcal{S}'$ is defined in the same way as in the HS algorithm.
The following lemma computes $\mathcal{S}'$. It should be noted that although $\mathcal{S}'$ is defined with the help of $\mathcal{S}_{overlay}$, $\mathcal{S}'$ is constructed directly without computing $\mathcal{S}_{overlay}$ first.
\begin{lemma}\label{lem:subalgo}
The conforming subdivision $\mathcal{S}'$ can be constructed in $O(n+h\log h)$ time and $O(n)$ space.
\end{lemma}
\begin{proof}
We first construct $\mathcal{S}$ in $O(h\log h)$ time and $O(h)$ space~\cite{ref:HershbergerAn99}. In
the following we construct $\mathcal{S}'$ by inserting the obstacle edges
into $\mathcal{S}$. The algorithm is similar to that in the HS algorithm
(see Lemma 2.3~\cite{ref:HershbergerAn99}). The difference is that we
need to handle each elementary chain as a whole.
We first build a data structure so that for any query horizontal ray with origin in $\mathcal{F}$, the first obstacle edge of $\mathcal{P}$ hit by it can be computed in $O(\log n)$ time. This can be done by building a horizontal decomposition of $\mathcal{F}$, i.e., extend a horizontal segment from each vertex until it hits $\partial \mathcal{P}$. As all obstacles of $\mathcal{P}$ are convex, the horizontal decomposition can be computed in $O(n+h\log h)$ time and $O(n)$ space~\cite{ref:HertelFa85}. By building a point location data structure~\cite{ref:EdelsbrunnerOp86,ref:KirkpatrickOp83} on the horizontal decomposition in additional $O(n)$ time, each horizontal ray shooting can be answered in $O(\log n)$ time. Similarly, we can construct the vertical decomposition of $\mathcal{F}$ in $O(n+h\log h)$ time and $O(n)$ space so that each vertical ray shooting can be answered in $O(\log n)$ time.
The edges of $\mathcal{S}'$ are obstacle edges, transparent edges incident to the vertices of $\mathcal{S}$, and transparent edges subdivided on the vertical segments incident to the vertices of $\mathcal{V}$. To identify the second type of edges, we trace the boundary of each interesting cell separately. Starting from a vertex $v$ of $\mathcal{S}$, we trace along each edge incident to $v$. Using the above ray-shooting data structure, we determine whether the next cell vertex is a vertex of $\mathcal{S}$ or the first point on $\partial \mathcal{P}$ hit by the ray. As $\mathcal{S}$ has $O(h)$ edges and vertices, this tracing takes $O(h\log n)$ time in total. Tracing along obstacle edges is done by starting from each vertex of $\mathcal{V}$ and following each of its incident elementary chains. For each elementary chain, the next vertex is either the next obstacle vertex on $e$, where $e$ is the current tracing edge of the elementary chain, or the intersection of $e$ with a transparent edge of the current cell. Hence, tracing all elementary chains takes $O(n)$ time in total. The third type of edges can be computed in linear time by local operations on each cell containing a vertex of $\mathcal{V}$.
Finally, we assemble all these edges together to obtain an adjacency
structure for $\mathcal{S}'$. For this, we could use a plane sweep algorithm as in the
HS algorithm. However, that would take $O(n\log n)$ time as there are
$O(n)$ edges in $\mathcal{S}'$. To obtain an $O(n+h\log h)$ time algorithm,
we propose the following approach. During the above tracing of
elementary chains, we record the fragment of the chain that lies in a
single cell. Since each such portion may not be of constant size, we
represent it by a line segment connecting its two endpoints; note that
this segment does not intersect any other cell edges because the
elementary chain fragment is on the boundary of an obstacle (and thus
the segment is inside the obstacle). Then, we apply the plane sweep
algorithm to stitch all edges with each elementary chain fragment in a
cell replaced by a segment as above. The algorithm takes $O(h\log h)$ time and
$O(h)$ space. Finally, for each segment representing an elementary
chain portion, we locally replace it by the chain fragment in linear
time. Hence, this step takes $O(n)$ time altogether for all such
segments. As such, the total time for computing the adjacency
information for $\mathcal{S}'$ is $O(n+h\log n)$, which is $O(n+h\log h)$.
Clearly, the space complexity of the algorithm is $O(n)$.
\end{proof}
\subsection{Basic concepts and notation}
\label{sec:notation}
Our shortest path algorithm uses the continuous Dijkstra method. The algorithm initially generates a wavefront from $s$, which is a circle centered at $s$. During the algorithm, the wavefront consists of all points of $\mathcal{F}$ with the same geodesic distance from $s$ (e.g., see Fig.~\ref{fig:wavefront}). We expand the wavefront until all points of the free space are covered. The conforming subdivision $\mathcal{S}'$ is utilized to guide the wavefront expansion. Our wavefront expansion algorithm follows the high-level scheme as the HS algorithm. The difference is that our algorithm somehow considers each elementary chain as a whole, which is in some sense similar to the algorithm of Hershberger, Suri, and Y{\i}ld{\i}z~\cite{ref:HershbergerA13} (called the {\em HSY} algorithm). Indeed, the HSY algorithm considers each $xy$-monotone convex arc as a whole, but the difference is that each arc in the HSY algorithm is of constant size while in our case each elementary chain may not be of constant size. As such, techniques from both the HS algorithm and the HSY algorithm are borrowed.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=2.0in]{wavefront.png}
\caption{\footnotesize Illustrating the wavefront. The black region are obstacles. The green point is $s$. The red curves are bisecting curves of $\mbox{$S\!P\!M$}(s)$. The gray region is the free space that has been covered by the wavefront. The boundary between the white region and the grey region is the wavefront. The figure is generated using the applet at~\cite{ref:HershbergerGe14}.}
\label{fig:wavefront}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
We use $\tau$ to denote the geodesic distance from $s$ to all points in the wavefront. One may also think of $\tau$ as a parameter representing time. The algorithm simulates the expansion of the wavefront as time increases from $0$ to $\infty$.
The wavefront comprises a sequence of {\em wavelets}, each emanating from a {\em generator}. In the HS algorithm, a generator is simply an obstacle vertex. Here, since we want to treat an elementary chain as a whole, similar to the HSY algorithm, we define a generator as a couple $\alpha=(A,a)$, where $A$ is an elementary chain and $a$ is an obstacle vertex on $A$, and further a clockwise or counterclockwise direction of $A$ is designated for $\alpha$; $a$ has a weight $w(a)$ (one may consider $w(a)$ as the geodesic distance between $s$ and $a$).
We call $a$ the {\em initial vertex} of the generator $\alpha$.
We say a point $q$ is {\em reachable} by a generator $\alpha=(A,a)$ if one can draw a path in $\mathcal{F}$ from $a$ to $q$ by following $A$ in the designated direction to a vertex $v$ on $A$ such that $\overline{vq}$ is tangent to $A$ and then following the segment $\overline{vq}$ (e.g., see Fig.~\ref{fig:generator}). The (weighted) distance between the generator $\alpha$ and $q$ is the length of this path plus $w(a)$;
by slightly abusing the notation, we use $d(\alpha,q)$ to denote the distance.
From the definition of reachable points, the vertex $a$ partitions $A$ into two portions and only the portion following the designated direction is relevant (e.g., in Fig.~\ref{fig:generator}, only the portion containing the vertex $v$ is relevant). Henceforth, unless otherwise stated, we use $A$ to refer to its relevant portion only and we call $A$ the {\em underlying chain} of $\alpha$.
In this way the initial vertex $a$ becomes an endpoint of $A$.
For convenience, sometimes we do not differentiate $\alpha$ and $A$. For example, when we say ``the tangent from $q$ to $\alpha$'', we mean ``the tangent from $q$ to $A$''; also, when we say ``a vertex of $\alpha$'', we mean ``a vertex of $A$''.
\begin{figure}[h]
\begin{minipage}[t]{0.51\textwidth}
\begin{center}
\includegraphics[height=1.4in]{generator.pdf}
\caption{\footnotesize Illustrating a generator $\alpha=(A,a)$. $A$ consists of the thick segments, and is designated clockwise direction around the obstacle. $q$ is a reachable point through vertex $v$.}
\label{fig:generator}
\end{center}
\end{minipage}
\hspace{0.02in}
\begin{minipage}[t]{0.47\textwidth}
\begin{center}
\includegraphics[height=1.4in]{wavelet.pdf}
\caption{\footnotesize Illustrating the wavelet generated by $\alpha=(A,q)$, designated counterclockwise direction. The wavelet has three pieces each of which is a circular arc. }
\label{fig:wavelet}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
The wavefront can thus be represented by the sequence of generators of
its wavelets. A wavelet generated by a generator $\alpha=(A,a)$ at time $\tau$ is a
contiguous set of reachable points $q$ such that $d(\alpha,q)=\tau$ and
$d(\alpha',q)\geq \tau$ for all other generators $\alpha'$ in the
wavefront; we also say that $q$ is claimed by $\alpha$. Note that as
$A$ may not be of constant size, a wavelet may
not be of constant size either; it actually consists of a
contiguous sequence of circular arcs centered at the obstacle vertices
$A$ (e.g., see Fig.~\ref{fig:wavelet}). If a point $q$ is claimed by $\alpha$, then
$d(s,q)=d(\alpha,q)=\tau$ and the predecessor $pred(q)$ of $q$ is $a$; sometimes for convenience we also say that the generator $\alpha$ is the predecessor of $q$. If $q$ is on an elementary
chain $A'$ and the tangent from $q$ to $a$ is also tangent to $A'$,
then a new generator $(A',q)$ is added to the wavefront (e.g., see Fig.~\ref{fig:newgenerator}(a)). A special
case happens when $q$ is the counterclockwise endpoint of $A$ (and thus
$q$ does not belong to $A$); in this case, a new generator $\alpha'=(A',q)$
is also added, where $A'$ is the elementary chain that contains $q$ (e.g., see Fig.~\ref{fig:newgenerator}(b)).
\begin{figure}[h]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=2.2in]{newgenerator.pdf}
\caption{\footnotesize Illustrating a new generator $\alpha'=(A',q)$. (a) A general case where both $A$ and $A'$ are marked with thick segments. (b) A special case where $A$ is marked with solid (red) segments and $A'$ is marked with dashed (blue) segments. }
\label{fig:newgenerator}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
As $\tau$ increases, the points bounding the adjacent wavelets trace out the bisectors that form the edges of the shortest path map $\mbox{$S\!P\!M$}(s)$. The {\em bisector} between the wavelets of two generators $\alpha$ and $\alpha'$, denoted
by $B(\alpha,\alpha')$, consists of points $q$ with $d(\alpha,q)=d(\alpha',q)$. Note that since $\alpha$ and $\alpha'$ may not be of constant size, $B(\alpha,\alpha')$ may not be of constant size either. More specifically, $B(\alpha,\alpha')$ has multiple pieces each of which is a hyperbola defined by two obstacle vertices $v\in \alpha$ and $v'\in \alpha'$ such that the hyperbola consists of all points that have two shortest paths from $s$ with $v$ and $v'$ as the anchors in the two paths, respectively (e.g., see Fig.~\ref{fig:bisector}).
A special case happens if $\alpha'$ is a generator created by the wavelet of $\alpha$, such as that illustrated in Fig.~\ref{fig:newgenerator}(a), then $B(\alpha,\alpha')$ is the half-line extended from $q$ along the direction from $v$ to $q$ (the dashed segment in Fig.~\ref{fig:newgenerator}(a)); we call such a bisector an {\em extension bisector}.
Note that in the case illustrated in Fig.~\ref{fig:newgenerator}(b), $B(\alpha,\alpha')$, which is also an extension bisector, is the half-line extended from $q$ along the direction from $v$ to $q$ (the dashed segment in Fig.~\ref{fig:newgenerator}(b)), where $v$ is the obstacle vertex adjacent to $q$ in $A$.
\begin{figure}[h]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=2.4in]{bisector.pdf}
\caption{\footnotesize Illustrating the bisector $B(\gamma,\gamma')$ defined by two generators $\gamma=(A,q)$ and $\gamma=(A',q')$. The portion between $q$ and $q'$ is a hyperbola defined by $v$ and $v'$. }
\label{fig:bisector}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
A wavelet gets eliminated from the wavefront if the two bisectors bounding it intersect, which is called a {\em bisector event}. Specifically, if $\alpha_1$, $\alpha$, and $\alpha_2$ are three consecutive generators of the wavefront, the wavelet generated by $\alpha$ will be eliminated when $B(\alpha_1,\alpha)$ intersects $B(\alpha,\alpha_2)$; e.g., see Fig.~\ref{fig:bisectorintersection}.
Wavelets are also eliminated by collisions with obstacles and other wavelets in front of it.
If a bisector $B(\alpha,\alpha')$ intersects an obstacle, their intersection is also called a {\em bisector event}.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=1.4in]{bisectorintersection.pdf}
\caption{\footnotesize Illustrating the intersection of two bisectors. }
\label{fig:bisectorintersection}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
Let $\mbox{$S\!P\!M$}'(s)$ be the subdivision of $\mathcal{F}$ by the bisectors of all generators (e.g. see Fig.~\ref{fig:map}). The intersections of bisectors and intersections between bisectors and obstacle edges are vertices of $\mbox{$S\!P\!M$}'(s)$. Each bisector connecting two vertices is an edge of $\mbox{$S\!P\!M$}'(s)$, called a {\em bisector edge}.
As discussed before, a bisector, which consists of multiple hyperbola pieces, may not be of constant size. Hence, a bisector edge $e$ of $\mbox{$S\!P\!M$}'(s)$ may not be of constant size. For differentiation, we call each hyperbola piece of $e$, a {\em hyperbolic-arc}.
In addition, if the boundary of an obstacle $P$ contains more than one vertex of $\mbox{$S\!P\!M$}'(s)$, then the chain of edges of $\partial P$ connecting two adjacent vertices of $\mbox{$S\!P\!M$}'(s)$ also forms an edge of $\mbox{$S\!P\!M$}'(s)$, called a {\em convex-chain edge}.
\begin{figure}[h]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=2.5in]{map.pdf}
\caption{\footnotesize Illustrating the map $\mbox{$S\!P\!M$}'(s)$. The red curves are non-extension bisectors. The green and blue segments are extension bisectors, where the green and blue ones are Type~(a) and (b) as illustrated in Fig.~\ref{fig:newgenerator}. The predecessor in each cell is also shown with a black point. For example, all points in the cell containing $q$ have $a$ as their predecessor. }
\label{fig:map}
\end{center}
\end{minipage}
\end{figure}
\begin{figure}[h]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=2.5in]{fullmap.pdf}
\caption{\footnotesize Illustrating the map $\mbox{$S\!P\!M$}(s)$. The dashed segments are windows that are not in $\mbox{$S\!P\!M$}'(s)$; removing them becomes $\mbox{$S\!P\!M$}'(s)$. The anchor of each cell is also shown.}
\label{fig:fullmap}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
The following lemma shows that $\mbox{$S\!P\!M$}'(s)$ is very similar to $\mbox{$S\!P\!M$}(s)$. Refer to Fig.~\ref{fig:fullmap} for $\mbox{$S\!P\!M$}(s)$.
\begin{lemma}\label{lem:spm}
Each extension bisector edge of $\mbox{$S\!P\!M$}'(s)$ is a window of $\mbox{$S\!P\!M$}(s)$. The union of all non-extension bisector edges of $\mbox{$S\!P\!M$}'(s)$ is exactly the union of all walls of $\mbox{$S\!P\!M$}(s)$. $\mbox{$S\!P\!M$}'(s)$ can be obtained from $\mbox{$S\!P\!M$}(s)$ by removing all windows except those that are extension bisectors of $\mbox{$S\!P\!M$}'(s)$.
\end{lemma}
\begin{proof}
We first prove that each extension bisector edge $e$ of $\mbox{$S\!P\!M$}'(s)$ is a window of $\mbox{$S\!P\!M$}(s)$.
By definition, one endpoint of $e$, denoted by $u$, is an obstacle vertex, such that $e$ is a half-line extended from $u$ in the direction from $v$ to $u$ for another obstacle vertex $v$. Hence, for each point $p\in e$, there is a shortest $s$-$p$ path $\pi(s,p)$ such that $\overline{vp}$ is a segment of $\pi(s,p)$. As $u\in \overline{vp}$, $p$ must be on a window $w$ of $\mbox{$S\!P\!M$}(s)$. This proves that $e\subseteq w$. On the other hand, for any point $p\in w$, there is a shortest $s$-$p$ path $\pi(s,p)$ that contains $\overline{vp}$. Since $u\in \overline{vp}$, $p$ must be on the extension bisector edge $e$ of $\mbox{$S\!P\!M$}'(s)$, and thus $w\subseteq e$. Therefore, $e=w$. This proves that each extension bisector edge of $\mbox{$S\!P\!M$}'(s)$ is a window of $\mbox{$S\!P\!M$}(s)$.
We next prove that the union of all non-extension bisector edges of $\mbox{$S\!P\!M$}'(s)$ is exactly the union of all walls of $\mbox{$S\!P\!M$}(s)$.
\begin{itemize}
\item
We first show that the union of all non-extension bisector edges of $\mbox{$S\!P\!M$}'(s)$ is a subset of the union of all walls of $\mbox{$S\!P\!M$}(s)$.
Let $q$ be a point on a non-extension bisector $B(\alpha,\alpha')$ of two generators $\alpha=(A,a)$ and $\alpha=(A',a')$. Let $v$ and $v'$ be the tangent points on $A$ and $A'$ from $q$, respectively.
Hence, $q$ has two shortest paths from $s$, one containing $\overline{qv}$ and the other containing $\overline{qv'}$. Since $B(\alpha,\alpha')$ is not an extension bisector, $v'\not\in \overline{qv}$ and $v\not\in\overline{qv'}$. This means that the two paths are combinatorially different and thus $q$ is on a wall of $\mbox{$S\!P\!M$}(s)$. This proves that the union of all non-extension bisector edges of $\mbox{$S\!P\!M$}'(s)$ is a subset of the union of all walls of $\mbox{$S\!P\!M$}(s)$.
\item
We then show that the union of all walls of $\mbox{$S\!P\!M$}(s)$ is a subset of the union of all non-extension bisector edges of $\mbox{$S\!P\!M$}'(s)$.
Let $q$ be a point on a wall of $\mbox{$S\!P\!M$}(s)$. By definition, there are two obstacle vertices $v$ and $v'$ such that there are two combinatorially different shortest paths from $s$ to $q$ whose anchors are $v$ and $v'$, respectively. Further, since the two paths are combinatorially different and also due to the general position assumption, $v'\not\in \overline{qv}$ and $v\not\in \overline{qv'}$. Therefore, $q$ must be on the bisector edge of the two generators $\alpha$ and $\alpha'$ in $\mbox{$S\!P\!M$}'(s)$ whose underlying chains containing $v$ and $v'$, respectively. Further, since $v\not\in \overline{qv'}$ and $v'\not\in \overline{qv}$, the bisector of $\alpha$ and $\alpha'$ is not an extension bisector. This proves that the union of all walls of $\mbox{$S\!P\!M$}(s)$ is a subset of the union of all non-extension bisector edges of $\mbox{$S\!P\!M$}'(s)$.
\end{itemize}
The above proves that the first and second statements of the lemma. The third statement immediately follows the first two.
\end{proof}
\begin{corollary}\label{coro:size}
The combinatorial size of $\mbox{$S\!P\!M$}'(s)$ is $O(n)$.
\end{corollary}
\begin{proof}
By Lemma~\ref{lem:spm}, $\mbox{$S\!P\!M$}'(s)$ is a subset of $\mbox{$S\!P\!M$}(s)$. As the combinatorial size of $\mbox{$S\!P\!M$}(s)$ is $O(n)$~\cite{ref:HershbergerAn99,ref:MitchellA91}, the combinatorial size of $\mbox{$S\!P\!M$}'(s)$ is also $O(n)$.
\end{proof}
\begin{lemma}\label{lem:sizespm}
The subdivision $\mbox{$S\!P\!M$}'(s)$ has $O(h)$ faces, vertices, and edges. In particular, $\mbox{$S\!P\!M$}'(s)$ has $O(h)$ bisector intersections and $O(h)$ intersections between bisectors and obstacle edges.
\end{lemma}
\begin{proof}
By Lemma~\ref{lem:spm}, for any vertex of $\mbox{$S\!P\!M$}'(s)$ that is the intersection of two non-exenstion bisectors, it is also the intersection of two walls of $\mbox{$S\!P\!M$}(s)$, which is a {\em triple point}.
Let $\mbox{$S\!P\!M$}''(s)$ refer to $\mbox{$S\!P\!M$}'(s)$ excluding all extension bisectors.
We define a planar graph $G$ corresponding to $\mbox{$S\!P\!M$}''(s)$ as follows. Each obstacle defines a node of $G$ and each triple point also defines a node of $G$. For any two nodes of $G$, if they are connected by a chain of bisector edges of $\mbox{$S\!P\!M$}'(s)$ such that the chain does not contain any other node of $G$, then $G$ has an edge connecting the two nodes.
It is proved in~\cite{ref:WangQu19} that $G$ has $O(h)$ vertices, faces, and edges.
It is not difficult to see that a face of $\mbox{$S\!P\!M$}''(s)$ corresponds to a face of $G$ and thus the number of faces of $\mbox{$S\!P\!M$}''(s)$ is $O(h)$. For each bisector intersection of $\mbox{$S\!P\!M$}''(s)$, it must be a triple point and thus it is also a node of $G$. As $G$ has $O(h)$ nodes, the number of bisector intersections of $\mbox{$S\!P\!M$}''(s)$ is $O(h)$.
For each intersection $v$ between a bisector and an obstacle edge in $\mbox{$S\!P\!M$}''(s)$, $G$ must has an edge $e$ corresponding a chain of bisector edges and $v$ is the endpoint of the chain; we charge $v$ to $e$. As $e$ as two incident nodes of $G$, $e$ can be charged at most twice. Since $G$ has $O(h)$ edges, the number of intersections between bisectors and obstacle edges in $\mbox{$S\!P\!M$}''(s)$ is $O(h)$.
We next prove that the total number of extension bisector edges of $\mbox{$S\!P\!M$}'(s)$ is $O(h)$. For each extension bisection edge $e$, it belongs to one of the two cases $(a)$ and $(b)$ illustrated in Fig.~\ref{fig:newgenerator}. For Case~(b), one endpoint of $e$ is a rectilinear extreme vertex. Since each rectilinear extreme vertex can define at most two extension bisector edges and there are $O(h)$ rectilinear extreme vertices, the total number of Case (b) extension bisectors is $O(h)$. For Case (a), $e$ is an extension of a common tangent of two obstacles; we say that $e$ is {\em defined} by the common tangent. Note that all such common tangents are interior disjoint as they belong to shortest paths from $s$ encoded by $\mbox{$S\!P\!M$}'(s)$. We now define a planar graph $G'$ as follows. The node set of $G'$ consists of all obstacles of $\mathcal{P}$. Two obstacles have an edge in $G'$ if they have at least one common tangent that defines an extension bisector. Since all such common tangents are interior disjoint, $G'$ is a planar graph. Apparently, no two nodes of $G'$ share two edges and no node of $G'$ has a self-loop. Therefore, $G'$ is a simple planar graph. Since $G'$ has $h$ vertices, the number of edges of $G'$ is $O(h)$. By the definition of $G'$, each pair of obstacles whose common tangents define extension bisectors have an edge in $G'$. Because each pair of obstacles can define at most four common tangents and thus at most four extension bisectors and also because $G'$ has $O(h)$ edges, the total number of Case (a) extension bisectors is $O(h)$.
The bisector intersections of $\mbox{$S\!P\!M$}'(s)$ that are not vertices of $\mbox{$S\!P\!M$}''(s)$ involve extension bisectors of $\mbox{$S\!P\!M$}'(s)$. The intersections between bisectors and obstacle edges in $\mbox{$S\!P\!M$}'(s)$ that are not in $\mbox{$S\!P\!M$}''(s)$ also involve extension bisectors. Since all extension bisectors and all edges of $\mbox{$S\!P\!M$}''(s)$ are interior disjoint, each extension bisector can involve in at most two bisector intersections and at most two intersections between bisectors and obstacle edges. Because the total number of extension bisector edges of $\mbox{$S\!P\!M$}'(s)$ is $O(h)$, the number of bisector intersections involving extension bisectors is $O(h)$ and the number of intersections between extension bisectors and obstacle edges is also $O(h)$.
Therefore, comparing to $\mbox{$S\!P\!M$}''(s)$, $\mbox{$S\!P\!M$}'(s)$ has $O(h)$ additional vertices and $O(h)$ additional edges. Note that the number of convex-chain edges of $\mbox{$S\!P\!M$}'(s)$ is $O(h)$, for $\mbox{$S\!P\!M$}'(s)$ has $O(h)$ vertices.
The lemma thus follows.
\end{proof}
\begin{corollary}\label{coro:complexity}
There are $O(h)$ bisector events and $O(h)$ generators in $\mbox{$S\!P\!M$}'(s)$.
\end{corollary}
\begin{proof}
Each bisector event is either a bisector intersection or an intersection between a bisector and an obstacle edge. By Lemma~\ref{lem:sizespm}, there are $O(h)$ bisector intersections and $O(h)$ intersections between bisectors and obstacle edges. As such, there are $O(h)$ bisector events in $\mbox{$S\!P\!M$}'(s)$.
By definition, each face of $\mbox{$S\!P\!M$}'(s)$ has a unique generator.
Since $\mbox{$S\!P\!M$}'(s)$ has $O(h)$ faces by Lemma~\ref{lem:sizespm}, the total number of generators is $O(h)$.
\end{proof}
By the definition of $\mbox{$S\!P\!M$}'(s)$, each cell $C$ of $\mbox{$S\!P\!M$}'(s)$ has a unique generator $\alpha=(A,a)$, all points of the cell are reachable from the generator, and $a$ is the predecessor of all points of $C$ (e.g., see Fig.~\ref{fig:map}; all points in the cell containing $q$ have $a$ as their predecessor).
Hence, for any point $q\in C$, we can compute $d(s,q)=d(\alpha,q)$ by computing the tangent from $q$ to $A$. Thus, $\mbox{$S\!P\!M$}'(s)$ can also be used to answer shortest path queries. In face, given $\mbox{$S\!P\!M$}'(s)$, we can construct $\mbox{$S\!P\!M$}(s)$ in additional $O(n)$ time by inserting the windows of $\mbox{$S\!P\!M$}(s)$ to $\mbox{$S\!P\!M$}'(s)$, as shown in the lemma below.
\begin{lemma}\label{lem:spmconstruction}
Given $\mbox{$S\!P\!M$}'(s)$, $\mbox{$S\!P\!M$}(s)$ can be constructed in $O(n)$ time (e.g., see Fig.~\ref{fig:map} and Fig.~\ref{fig:fullmap}).
\end{lemma}
\begin{proof}
It suffices to insert all windows of $\mbox{$S\!P\!M$}(s)$ (except those that are also extension bisectors of $\mbox{$S\!P\!M$}'(s)$) to $\mbox{$S\!P\!M$}'(s)$. We consider each cell $C$ of $\mbox{$S\!P\!M$}'(s)$ separately. Let $\alpha=(A,a)$ be the generator of $C$.
Our goal is to extend each obstacle edge of $A$ along the designated direction of $A$ until the boundary of $C$. To this end, since all points of $C$ are reachable from $\alpha$, the cell $C$ is ``star-shaped'' with respect to the tangents of $A$ (along the designated direction) and
the extension of each obstacle edge of $A$ intersects the boundary of $C$ at most once. Hence, the order of the endpoints of these extensions is consistent with the order of the corresponding edges of $A$. Therefore, these extension endpoints can be found in order by traversing the boundary of $C$, which takes linear time in the size of $C$. Since the total size of all cells of $\mbox{$S\!P\!M$}'(s)$ is $O(n)$ by Corollary~\ref{coro:size}, the total time of the algorithm is $O(n)$.
\end{proof}
In light of Lemma~\ref{lem:spmconstruction}, we will focus on computing the map $\mbox{$S\!P\!M$}'(s)$.
\subsection{The wavefront expansion algorithm}
\label{sec:algo}
To simulate the wavefront expansion, we compute the wavefront passing
through each transparent edge of the conforming subdivision $\mathcal{S}'$.
As in the HS algorithm, since it is difficult to compute a true
wavefront for each transparent edge $e$ of $\mathcal{S}'$, a key idea
is to compute two one-sided wavefronts (called {\em
approximate wavefronts}) for $e$, each representing the wavefront
coming from one side of $e$. Intuitively, an approximate wavefront
from one side of $e$ is what the true wavefront would be if the
wavefront were blocked off at $e$ by considering $e$ as an
opaque segment (with open endpoints).
In the following, unless otherwise stated, a wavefront at a
transparent edge $e$ of $\mathcal{S}'$ refers to an approximate wavefront.
Let $W(e)$ denote a wavefront at $e$. To make the description concise,
as there are two wavefronts at $e$, depending on the context, $W(e)$
may refer to both wavefronts, i.e., the discussion on $W(e)$ applies
to both wavefronts. For example, ``compute the wavefronts $W(e)$''
means ``compute both wavefronts at $e$''.
For each transparent edge $e$ of $\mathcal{S}'$,
define $input(e)$ as the set of transparent edges on the boundary of
the well-covering region $\mathcal{U}(e)$, and define $output(e)=\{g\ |\ e\in
input(g)\}\cup input(e)$\footnote{We include $input(e)$ in $output(e)$
mainly for the proof of Lemma~\ref{lem:markcell}, which relies on
$output(e)$ having a cycle enclosing $e$.}. Because $\partial
\mathcal{U}(e')$ for each transparent edge $e'$ of $\mathcal{S}'$ has $O(1)$
transparent edges, both $|input(e)|$ and $|output(e)|$ are $O(1)$.
\paragraph{The wavefront propagation and merging procedures.}
The wavefronts $W(e)$ at $e$ are computed from the wavefronts at edges of
$input(e)$; this guarantees the correctness because $e$ is in $\mathcal{U}(e)$ (and thus
any shortest path $\pi(s,p)$ must cross some edge $f\in input(e)$ for any point $p\in e$).
After the wavefronts $W(e)$ at $e$ are computed, they will pass
to the edges of $output(e)$. Also, the geodesic distances
from $s$ to both endpoints of $e$ will be computed. Recall that
$\mathcal{V}$ is the set of rectilinear extreme vertices of all obstacles
and each
vertex of $\mathcal{V}$ is incident to a transparent edge of $\mathcal{S}'$. As such,
after the algorithm is finished, geodesic distances from $s$ to all
vertices of $\mathcal{V}$ will be available. The process of passing the
wavefronts $W(e)$ at $e$ to all edges $g\in output(e)$ is called the
{\em wavefront propagation procedure}, which will compute the
wavefront $W(e,g)$, where $W(e,g)$ is
the portion of $W(e)$ that passes to $g$ through the well-covering
region $\mathcal{U}(g)$ of $g$ if $e\in input(g)$ and through $\mathcal{U}(e)$
otherwise (in this case $g\in input(e)$);
whenever the procedure is invoked on $e$, we say that $e$ is {\em
processed}. The wavefronts $W(e)$ at $e$ are constructed by
merging the wavefronts $W(f,e)$ for edges $f\in input(e)$; this procedure is called the
{\em wavefront merging procedure}.
\paragraph{The main algorithm.}
The transparent edges of $\mathcal{S}'$ are processed in a rough time order.
The wavefronts $W(e)$ of each transparent edge $e$ are constructed at the time $\tilde{d}(s,e)+|e|$, where $\tilde{d}(s,e)$ is the minimum geodesic distance from $s$ to the two endpoints of $e$.
Define $covertime(e)=\tilde{d}(s,e)+|e|$. The value $covertime(e)$ will be
computed during the algorithm. Initially, for each edge $e$ whose
well-covering region $\mathcal{U}(e)$ contains $s$, $W(e)$ and $covertime(e)$ are computed
directly (and set $covertime(e)=\infty$ for all other edges);
we refer to this step as the {\em initialization step}, which will be elaborated below.
The algorithm maintains a timer $\tau$ and processes the
transparent edges $e$ of $\mathcal{S}'$ following the order of
$covertime(e)$.
The main loop of the algorithm works as follows.
As long as $\mathcal{S}'$ has an unprocessed transparent edge, we do the following.
First, among all unprocessed transparent edges, choose the one $e$
with minimum $covertime(e)$ and set $\tau=covertime(e)$. Second, call
the
wavefront merging procedure to construct the wavefronts $W(e)$ from
$W(f,e)$ for all edges $f\in input(e)$ satisfying
$covertime(f)<covertime(e)$; compute $d(s,v)$ from $W(e)$ for each
endpoint $v$ of $e$. Third, process $e$, i.e., call the wavefront
propagation procedure on $W(e)$ to compute $W(e,g)$ for all edges
$g\in output(e)$; in particular, compute the time $\tau_g$ when the
wavefronts $W(e)$ first encounter an endpoint of $g$ and set
$covertime(g)= \min\{covertime(g),\tau_g+|g|\}$.
The details of the wavefront merging procedure and the wavefront propagation procedure will be described in Section~\ref{sec:merge} and Section~\ref{sec:propagation}, respectively.
\paragraph{The initialization step.} We provide the details on the initialization step. Consider a transparent edge $e$ whose well-covering region $\mathcal{U}(e)$ contains $s$. To compute $W(e)$, we only consider straight-line paths from $s$ to $e$ inside $\mathcal{U}(e)$. If $\mathcal{U}(e)$ does not have any island inside, then the points of $e$ that can be reached from $s$ by straight-line paths in $\mathcal{U}(e)$ form an interval of $e$ and the interval can be computed by considering the tangents from $s$ to the elementary chains on the boundary of $\mathcal{U}(e)$. Since $\mathcal{U}(e)$ has $O(1)$ cells of $\mathcal{S}'$, $\partial \mathcal{U}(e)$ has $O(1)$ elementary chain fragments and thus computing the interval on $e$ takes $O(\log n)$ time. If $\mathcal{U}(e)$ has at least one island inside, then $e$ may have multiple such intervals. As $\mathcal{U}(e)$ is the union of $O(1)$ cells of $\mathcal{S}'$, the number of islands inside $\mathcal{U}(e)$ is $O(1)$. Hence, $e$ has $O(1)$ such intervals, which can be computed in $O(\log n)$ time altogether. These intervals form the wavefront $W(e)$, i.e., each interval corresponds to a wavelet with generator $s$. From $W(e)$, the value $covertime(e)$ can be immediately determined. More specifically, for each endpoint $v$ of $e$, if one of the wavelets of $W(e)$ covers $v$, then the segment $\overline{sv}$ is in $\mathcal{U}(e)$ and thus $d(s,v)=|\overline{sv}|$; otherwise, we set $d(s,v)=\infty$. In this way, we find an upper bound for $\tilde{d}(s,e)$ and set $covertime(e)$ to this upper bound plus $|e|$.
\paragraph{The algorithm correctness.}
At the time $\tilde{d}(s,e)+|e|$, all edges $f\in input(e)$ whose wavefronts
$W(f)$ contribute a wavelet to $W(e)$ must have already been
processed. This is due to the property of the well-covering regions of
$\mathcal{S}'$ that $d(e,f)\geq 2\cdot \max\{|e|,|f|\}$ since $f$ is on
$\partial\mathcal{U}(e)$. The proof is the same as that of
Lemma~4.2~\cite{ref:HershbergerAn99}, so we omit it. Note that this
also implies that the geodesic distance $d(s,v)$ is correctly computed
for each endpoint $v$ of $e$. Therefore, after the algorithm is
finished, geodesic distances from $s$ to endpoints of all
transparent edges of $\mathcal{S}'$ are correctly computed.
\paragraph{Artificial wavelets.}
As in the HS algorithm, to limit the interaction between wavefronts from different sides of each transparent edge $e$, when a wavefront propagates across $e$, i.e., when $W(e)$ is computed in the wavefront merging procedure, an {\em artificial wavelet} is generated at each endpoint $v$ of $e$, with weight $d(s,v)$. This is to eliminate a wavefront from one side of $e$ if it arrives at $e$ later than the wavefront from the other side of $e$.
\paragraph{Topologically different paths.}
In the wavefront propagation procedure to pass $W(e)$ to all edges $g\in output(e)$, $W(e)$ travels through the cells of the well-covering region $\mathcal{U}$ of either $g$ or $e$. Since $\mathcal{U}$ may not be simply connected (e.g., the square-annulus), there may be multiple topologically different shortest paths between $e$ and $g$ inside $\mathcal{U}$; the number of such paths is $O(1)$ as $\mathcal{U}$ is the union of $O(1)$ cells of $\mathcal{S}'$. We propagate $W(e)$ in multiple components of $\mathcal{U}$, each corresponding to a topologically different shortest path and defined by the particular sequence of transparent edges it crosses. These wavefronts are later combined in the wavefront merging step at $g$. This also happens in the initialization step, as discussed above.
\paragraph{Claiming a point.}
During the wavefront merging procedure at $e$, we have a set of wavefronts $W(f,e)$ that reach $e$ from the same side for edges $f\in input(e)$. We say that a wavefront $W(f,e)$ {\em claims} a point $p\in e$ if $W(f,e)$ reaches $p$ before any other wavefront. Further, for each wavefront $W(f,e)$, the set of points on $e$ claimed by it forms an interval (the proof is similar to that of Lemma~4.4~\cite{ref:HershbergerAn99}).
Similarly, a wavelet of a wavefront {\em claims} a point $p$ of $e$ if the wavelet reaches $p$ before any other wavelet of the wavefront, and the set of points on $e$ claimed by any wavelet of the wavefront also forms an interval. For convenience, if a wavelet claims a point, we also say that the generator of the wavelet claims the point.
\bigskip
Before presenting the details of the wavefront merging procedure and the wavefront propagation procedure in the next two subsections, we first discuss the data structure (for representing elementary chains, generators, and wavefronts) and a monotonicity property of bisectors, which will be used later in our algorithm.
\subsubsection{The data structure}
We use an array to represent each elementary chain. Then, a generator $\alpha=(A,a)$ can be represented by recording the indices of the two end vertices of its underlying chain $A$. In this way, a generator takes $O(1)$ additional space to record and binary search on $A$ can be supported in $O(\log n)$ time (e.g., given a point $p$, find the tangent from $p$ to $A$).
In addition, we maintain the lengths of the edges in each elementary chain so that given any two vertices of the chain, the length of the sub-chain between the two vertices can be obtained in $O(1)$ time. All these preprocessing can be done in $O(n)$ time and space for all elementary chains.
For a wavefront $W(e)$ of one side of $e$, it is a list of generators $\alpha_1,\alpha_2,\ldots$ ordered by the intervals of $e$ claimed by these generators. Note that these generators are in the same side of $e$. Formally, we say that a generator $\alpha$ is in one side of $e$ if the initial vertex of $\alpha$ lies in that side of the supporting line of $e$.
We maintain these generators by a balanced binary search tree so that the following operations can be supported in logarithmic time each. Let $W$ be a wavefront with generators $\alpha_1,\alpha_2,\ldots,\alpha_k$.
\begin{description}
\item[Insert] Insert a generator $\alpha$ to $W$.
In our algorithm, $\alpha$ is inserted either in the front of $\alpha_1$ or after $\alpha_k$.
\item[Delete] Delete a generator $\alpha_i$ from $W$, for any $1\leq i\leq k$.
\item[Split] Split $W$ into two sublists at some generator $\alpha_i$ so that the first $i$ generators form a wavefront and the rest form another wavefront.
\item[Concatenate] Concatenate $W$ with another list $W'$ of generators so that all generators of $W$ are in the front of those of $W'$ in the new list.
\end{description}
We will show later that each wavefront involved in our algorithm has $O(h)$ generators. Therefore, each of the above operation can be performed in $O(\log h)$ time.
We make the tree fully persistent by path-copying~\cite{ref:DriscollMa89}. In this way, each operation on the tree will cost $O(\log h)$ additional space but the older version of the tree will be kept intact (and operations on the old tree can still be supported).
\subsubsection{The monotonicity property of bisectors}
\begin{lemma}\label{lem:intersection}
Suppose $\alpha_1=(A_1,a_1)$ and $\alpha_2=(A_2,a_2)$ are two generators on the same side of an axis-parallel line $\ell$ (i.e., $a_1$ and $a_2$ are on the same side of $\ell$; it is possible that $\ell$ intersects $A_1$ and $A_2$). Then, the bisector $B(\alpha_1,\alpha_2)$ intersects $\ell$ at no more than one point.
\end{lemma}
\begin{proof}
Consider a generator $\alpha=(A,a)$. Recall that $a$ is an endpoint of $A$. Let $a'$ be the other endpoint of $A$.
We define $R(\alpha)$ as the set of points in the plane that are reachable from $\alpha$ (without considering any other obstacles), and we call it the {\em reachable region} of $\alpha$.
The reachable region $R(\alpha)$ is bounded by $A$, a ray $\rho(a)$ with origin $a$, and another ray $\rho(a')$ with origin $a'$ (e.g., see Fig.~\ref{fig:reachableregion}).
Specifically, $\rho(a)$ is along the direction from $z$ to $a$, where $z$ is the anchor of $a$ in the shortest path from $s$ to $a$;
$\rho(a')$ is the ray from the adjacent obstacle vertex of $a'$ on $A'$ to $a'$.
We claim that the intersection $\ell\cap R(\alpha)$ must be an interval of $\ell$. Indeed, since $A$ is $xy$-monotone, without loss of generality, we assume that $a'$ is to the northwest of $a$. Let $\rho'(a)$ be the vertically downward ray from $a$ (e.g., see Fig.~\ref{fig:reachableregion10}). Observe that the concatenation of $\rho(a')$, $A$, and $\rho'(a)$ is $xy$-monotone; let
$R'$ be the region of the plane to the right of their concatenation. Since the boundary of $R'$ is $xy$-monotone and $\ell$ is axis-parallel, $R'\cap \ell$ must be an interval of $\ell$. Notice that $\rho(a)$ must be in $R'$ and thus it partitions $R'$ into two subregions, one of which is $R(\alpha)$. Since $\ell$ intersects $\rho(a)$ at most once and $R'\cap \ell$ is an interval, $R(\alpha)\cap \ell$ must also be an interval.
\begin{figure}[t]
\begin{minipage}[t]{0.49\textwidth}
\begin{center}
\includegraphics[height=1.2in]{reachableregion.pdf}
\caption{\footnotesize Illustrating the reachable region $R(\alpha)$ of a generator $\alpha=(A,a)$.}
\label{fig:reachableregion}
\end{center}
\end{minipage}
\hspace{0.02in}
\begin{minipage}[t]{0.49\textwidth}
\begin{center}
\includegraphics[height=1.5in]{reachableregion10.pdf}
\caption{\footnotesize Illustrating the ray $\rho'(a)$ and the region $R'$.}
\label{fig:reachableregion10}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
According to the above claim, both $\ell\cap R(\alpha_1)$ and $\ell\cap R(\alpha_2)$ are intervals of $\ell$.
Let $I$ denote the common intersection of the two intervals. Since $B(\alpha_1,\alpha_2)\subseteq R(\alpha_1)$ and $B(\alpha_1,\alpha_2)\subseteq R(\alpha_2)$, we obtain that $B(\alpha_1,\alpha_2)\cap \ell\subseteq I$.
\begin{figure}[h]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=1.4in]{monotone.pdf}
\caption{\footnotesize Illustrating the proof of Lemma~\ref{lem:intersection}.}
\label{fig:monotone}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
Assume to the contrary that $B(\alpha_1,\alpha_2)\cap \ell$ contains two points, say, $q_1$ and $q_2$. Then, both points are in $I$. Let $v_1$ and $u_1$ be the tangents points of $q_1$ on $A_1$ and $A_2$, respectively.
Let $v_2$ and $u_2$ be the tangents points of $q_2$ on $A_1$ and $A_2$, respectively (e.g., see Fig.~\ref{fig:monotone}). As $I=\ell\cap R(\alpha_1)\cap R(\alpha_2)$ and $I$ contains both $q_1$ and $q_2$, if we move a point $q$ from $q_1$ to $q_2$ on $\ell$, the tangent from $q$ to $A_1$ will continuously change from $\overline{q_1v_1}$ to $\overline{q_2v_2}$ and the tangent from $q$ to $A_2$ will continuously change from $\overline{q_1u_1}$ to $\overline{q_2u_2}$. Since $A_1$ and $A_2$ are in the same side of $\ell$, either $\overline{q_1u_1}$ intersects $\overline{q_2v_2}$ in their interiors or $\overline{q_1v_1}$ intersects $\overline{q_2u_2}$ in their interiors; without loss of generality, we assume that it is the former case. Let $p$ be the intersection of $\overline{q_1u_1}$ and $\overline{q_2v_2}$ (e.g., see Fig.~\ref{fig:monotone}). Since $q_1\in B(\alpha_1,\alpha_2)$, points of $\overline{q_1u_1}$ other than $q_1$ have only one predecessor, which is $a_2$. As $p\in \overline{q_1u_1}$ and $p\neq q_1$, $p$ has only one predecessor $a_2$. Similarly, since $q_2\in B(\alpha_1,\alpha_2)$ and $p\in \overline{q_2v_2}$, $a_1$ is also $p$'s predecessor. We thus obtain a contradiction.
\end{proof}
\begin{corollary}\label{coro:monotone}
Suppose $\alpha_1=(A_1,a_1)$ and $\alpha_2=(A_2,a_2)$ are two generators both below a horizontal line $\ell$. Then, the portion of the bisector $B(\alpha_1,\alpha_2)$ above $\ell$ is $y$-monotone.
\end{corollary}
\begin{proof}
For any horizontal line $\ell'$ above $\ell$, since both generators are below $\ell$, they are also below $\ell'$. By Lemma~\ref{lem:intersection}, $B(\alpha_1,\alpha_2)\cap \ell'$ is either empty or a single point. The corollary thus follows.
\end{proof}
\subsection{The wavefront merging procedure}
\label{sec:merge}
In this section, we present the details of the wavefront merging procedure. Given all contributing wavefronts $W(f,e)$ of $f\in input(e)$ for $W(e)$, the goal of the procedure is to compute $W(e)$.
The algorithm follows the high-level scheme of the HS algorithm (i.e., Lemma 4.6~\cite{ref:HershbergerAn99}) but the implementation details are quite different.
We only consider the wavefronts $W(f,e)$ and $W(e)$ for one side of $e$ since the algorithm for the other side is analogous. Without loss of generality, we assume that $e$ is horizontal and all wavefronts $W(f,e)$ are coming from below $e$. We describe the algorithm for computing the interval of $e$ claimed by $W(f,e)$ if only one other wavefront $W(f',e)$ is present. The common intersection of these intervals of all such $f'$ is the interval of $e$ claimed by $W(f,e)$. Since $|input(e)|=O(1)$, the number of such $f'$ is $O(1)$.
We first determine whether the claim of $W(f,e)$ is to the left or right of that of $W(f',e)$. To this end,
depending on whether both $W(f,e)$ and $W(f',e)$ reach the left endpoint $v$ of $e$, there are two cases. Note that the intervals of $e$ claimed by $W(f,e)$ and $W(f',e)$ are available from $W(f,e)$ and $W(f',e)$; let $I_f$ and $I_{f'}$ denote these two intervals, respectively.
\begin{itemize}
\item
If both $I_f$ and $I_{f'}$ contain $v$, i.e., both $W(f,e)$ and $W(f',e)$ reach $v$, then we compute the (weighted) distances from $v$ to the two wavefronts. This can be done as follows. Since $v\in I_f$, $v$ must be reached by the leftmost generator $\alpha$ of $W(f,e)$. We compute the distance $d(\alpha,v)$ by computing the tangent from $v$ to $\alpha$ in $O(\log n)$ time. Similarly, we compute $d(\alpha',v)$, where $\alpha'$ is leftmost generator of $W(f',e)$. If $d(\alpha, v)\leq d(\alpha',v)$, then the claim of $W(f,e)$ is to the left of that of $W(f',e)$; otherwise, the claim of $W(f,e)$ is to the right of that of $W(f',e)$.
\item
If not both $I_f$ and $I_{f'}$ contain $v$, then the order of the left endpoints of $I_f$ and $I_{f'}$ will give the answer.
\end{itemize}
Without loss of generality, we assume that the claim of $W(f,e)$ is to the left of that of $W(f',e)$. We next compute the interval $I$ of $e$ claimed by $W(f,e)$ with respect to $W(f',e)$. Note that the left endpoint of $I$ is the left endpoint of $I_f$. Hence, it remains to find the right endpoint of $I$, as follows.
Let $\ell_e$ be the supporting line of $e$.
Let $\alpha$ be the rightmost generator of $W(f,e)$ and let $\alpha'$ be the leftmost generator of $W(f',e)$.
Let $q_1$ be the left endpoint of the interval on $e$ claimed by $\alpha$ in $W(f,e)$, i.e., $q_1$ is the intersection of the bisector $B(\alpha_1,\alpha)$ and $e$, where $\alpha_1$ is the left neighboring generator of $\alpha$ in $W(f,e)$. Similarly, $q_2$ be the right endpoint of the interval on $e$ claimed by $\alpha'$ in $W(f',e)$, i.e., $q_2$ is the intersection of $e$ and the bisector $B(\alpha',\alpha_1')$, where $\alpha_1'$ is the right neighboring generator of $\alpha'$ in $W(f',e)$.
Let $q_0$ be the intersection of the bisector $B(\alpha,\alpha')$ and $e$. We assume that the three points $q_i$, $i=0,1,2$ are available and we will discuss later that each of them can be computed in $O(\log n)$ time by a {\em bisector-line intersection operation} given in Lemma~\ref{lem:bl-intersection}. If $q_0$ is between $q_1$ and $q_2$, then $q_0$ is the right endpoint of $I$ and we can stop the algorithm. If $q_0$ is to the left of $q_1$, then we delete $\alpha$ from $W(f,e)$. If $q_0$ is to the right of $q_2$, then we delete $\alpha'$ from $W(f',e)$. In either case, we continue the same algorithm by redefining $\alpha$ or $\alpha'$ (and recomputing $q_i$ for $i=0,1,2$).
Clearly, the above algorithm takes $O((1+k)\log n)$ time, where $k$ is the number of generators that are deleted. We apply the algorithm on $f$ and other $f'$ in $input(e)$ to compute the corresponding intervals for $f$. The common intersection of all these intervals is the interval of $e$ claimed by $W(f,e)$. We do so for each $f\in input(e)$, after which $W(e)$ is obtained. Since the size of $input(e)$ is $O(1)$, we obtain the following lemma.
\begin{lemma}\label{lem:merge}
Given all contributing wavefronts $W(f,e)$ of edges $f\in input(e)$ for $W(e)$, we can compte the interval of $e$ claimed by each $W(f,e)$ and thus construct $W(e)$ in $O((1+k)\log n)$ time, where $k$ is the total number of generators in all wavefronts $W(f,e)$ that are absent from $W(e)$.
\end{lemma}
\begin{lemma}{\em\bf (Bisector-line intersection operation)}\label{lem:bl-intersection}
Each bisector-line intersection operation can be performed in $O(\log n)$ time.
\end{lemma}
\begin{proof}
Given two generators $\alpha_1=(A_1,a_1)$ and $\alpha_2=(A_2,a_2)$ below a horizontal line $\ell$, the goal of the operation is to compute the intersection between the bisector $B(\alpha_1,\alpha_2)$ and $\ell$. By Lemma~\ref{lem:intersection}, $B(\alpha_1,\alpha_2)\cap \ell$ is either empty or a single point.
We apply a prune-and-search technique from Kirkpatrick and Snoeyink~\cite{ref:KirkpatrickTe95}. To avoid the lengthy background explanation, we follow the notation in~\cite{ref:KirkpatrickTe95} without definition.
We will rely on Theorem~3.6 in~\cite{ref:KirkpatrickTe95}. To do so,
we need to define a decreasing function $f$ and an increasing function $g$.
We first compute the intersection of $\ell$ and the reachable region $R(\alpha_1)$ of $\alpha_1$, as defined in the proof of Lemma~\ref{lem:intersection}. This can be easily done by computing the intersections between $\ell$ and the boundary of $R(\alpha_1)$ in $O(\log n)$ time. Let $I_1$ be their intersection, which is an interval of $\ell$ as proved in Lemma~\ref{lem:intersection}. Similarly, we compute the intersection $I_2$ of $\ell$ and the reachable region $R(\alpha_2)$ of $\alpha_2$. Let $I=I_1\cap I_2$.
For each endpoint of $I$, we compute its tangent point on $A_1$ (along its designated direction) to determine the portion of $A_1$ whose tangent rays intersect $I$ and let $A$ denote the portion. Similarly, we determine the portion $B$ of $A_2$ whose tangent rays intersect $I$. All above takes $O(\log n)$ time.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=1.7in]{bisectorline.pdf}
\caption{\footnotesize Illustrating the definitions of $f(a)$ and $g(b)$.}
\label{fig:bisectorline}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
We parameterize over $[0,1]$ each of the two convex chains $A$ and
$B$ in clockwise order, i.e., each value of $[0,1]$ corresponds to a slope of a tangent at a point on
$A$ (resp., $B$). For each point $a$ of $A$, we define $f(a)$ to be the parameter of the point $b\in B$ such that the tangent ray of $A$ at $a$ along the designated direction and the tangent ray of $B$ at $b$ along the designated direction intersect at a point on the bisector $B(\alpha_1,\alpha_2)$ (e.g., see Fig.~\ref{fig:bisectorline} left); if the tangent ray at $a$ does not intersect $B(\alpha_1,\alpha_2)$, then define $f(a)=1$. For each point $b$ of $B$, we define $g(b)$ to be the parameter of the point $a\in A$ such that the tangent ray of $A$ at $a$ and the tangent ray of $B$ at $b$ intersect at a point on the line $\ell$ (e.g., see Fig.~\ref{fig:bisectorline} right).
One can verify that $f$ is a continuous decreasing function while $g$ is a continuous increasing function (the tangent at an obstacle vertex of $A$ and $B$ is not unique but the issue can be handled~\cite{ref:KirkpatrickTe95}). The fixed-point of the
composition of the two functions $g \cdot f$ corresponds to
the intersection of $\ell$ and $B(\alpha_1,\alpha_2)$,
which can be computed by applying the
prune-and-search algorithm of Theorem 3.6~\cite{ref:KirkpatrickTe95}.
As both chains $A$ and $B$ are represented by arrays (of size $O(n)$),
to show that the algorithm can be implemented in $O(\log n)$ time, it suffices to show that given
any $a\in A$ and any $b\in B$, we can determine whether $f(a)\geq b$ in $O(1)$
time and determine whether $g(b)\geq a$ in $O(1)$ time.
To determine whether $f(a)\geq b$, we do the following.
We first find the intersection $q$ of the tangent ray of $A$ at
$a$ and the tangent ray of $B$ at $b$. Then, $f(a)\geq b$ if and only if $d(\alpha_1,q)\leq d(\alpha_2,q)$. Notice that $d(\alpha_1,q)=w(a_1)+|\widehat{a_1a}|+|\overline{aq}|$, where $|\widehat{a_1a}|$ is the length of the portion of $A_1$ between $a_1$ and $a$. Recall that we have a data structure on each elementary chain $C$ such that given any two vertices on $C$, the length of the sub-chain of $C$ between the two vertices can be computed in $O(1)$ time. Using the data structure,
$|\widehat{a_1a}|$ can be computed in constant time. Since $w(a_1)$ is already known, $d(\alpha_1,q)$ can be computed in constant time. So is $d(\alpha_2,q)$.
In the case where the two tangent rays do not intersect, either the tangent ray of $A$ at $a$ intersects the backward extension of the tangent ray of $B$ at $b$ or the tangent ray of $B$ at $b$ intersects the backward extension of the tangent ray of $A$ at $a$. In the former case $f(a)\leq b$ holds while in the latter case $f(a)\geq b$ holds. Hence, whether $f(a)\geq b$ can be determined in constant time.
To determine whether $g(b)\geq a$, we do the following. Find the intersection $p_a$ between $\ell$ and the tangent ray of $A$ at $a$ and the intersection $p_b$ between $\ell$ and the tangent ray of $B$ at $b$. If $p_a$ is to the left of $p_b$, then $g(b)\geq a$; otherwise $g(b)\leq a$.
Note that by the definition of $A$, the tangent ray at any point of $A$ intersects $I$; the same is true for $B$. Hence, whether $g(b)\geq a$ can be determined in constant time.
The above algorithm returns a point $q$ in $O(\log n)$ time. If the intersection of $\ell$ and $B(\alpha_1,\alpha_2)$ exists, then $q$ is the intersection. Because we do not know whether the intersection exists, we finally validate $q$ by computing $d(\alpha_1,q)$ and $d(\alpha_2,q)$ in $O(\log n)$ time as well as checking whether $q\in \ell$. The point $q$ is valid if and only if $d(\alpha_1,q)=d(\alpha_2,q)$ and $q\in \ell$.
\end{proof}
\subsection{The wavefront propagation procedure}
\label{sec:propagation}
In this section, we discuss the wavefront propagation procedure, which
is to compute the wavefront $W(e,g)$ for all transparent edges $g\in output(e)$ based on
$W(e)$. Consider a transparent edge $g\in output(e)$. The wavefront
$W(e,g)$ refers to the portion of $W(e)$ that arrives at $g$ through
the well-covering region $\mathcal{U}(g)$ of $g$ if $e\in
input(g)$ and through $\mathcal{U}(e)$ otherwise (in the latter case $g\in
input(e)$). We will need to handle the bisector events, i.e., the
intersections between bisectors and the intersections between
bisectors and obstacle edges.
The HS algorithm processes the bisector events in {\em temporal} order,
i.e., in order of the simulation time $\tau$. The HSY algorithm
instead proposes a simpler
approach that processes the events in {\em spatial order}, i.e., in
order of their geometric locations. We will adapt the HSY's
spacial-order method.
Recall that each wavefront $W(e)$ is represented by a list of generators, which are
maintained in the leaves of a fully-persistent balanced binary search
tree $T(e)$. We further assign each generator a ``next bisector event'', which
is the intersection of its two bounding bisectors (it is set to null if the two bisectors do not intersect). More
specifically, for each bisector $\alpha$, we assign it
the intersection of the two bisectors $B(\alpha_l,\alpha)$ and
$B(\alpha,\alpha_r)$, where $\alpha_l$ and $\alpha_r$ are $\alpha$'s
left and right neighboring generators in $W(e)$, respectively; we store the intersection at the leaf for $\alpha$. Our algorithm
maintains a variant that the next bisector event for each generator in
$W(e)$ has already been computed and stored in $T(e)$.
We further endow the tree $T(e)$ with additional node-fields so that each internal node stores a value
that is equal to the minimum (resp., maximum) $x$-coordinate (resp., $y$-coordinate) among all bisector
events stored at the leaves of the subtree rooted at the node. Using these extra values, we
can find from a query range of generators the generator whose bisector event has the minimum/maximum $x$- or $y$-coordinate in logarithmic time.
The propagation from $W(e)$ to $g$ through $\mathcal{U}$ is done cell by cell, where $\mathcal{U}$ is either $\mathcal{U}(e)$ or $\mathcal{U}(g)$. We start propagating $W(e)$ to the adjacent cell $c$ of $e$ in $\mathcal{U}$ to compute the wavefront passing through all edges of $c$. Then by using the computed wavefronts on the edges of $c$, we recursively run the algorithm on cells of $\mathcal{U}$ adjacent to $c$. As $\mathcal{U}$ has $O(1)$ cells, the propagation passes through $O(1)$ cells. Hence, the essential ingredient of the algorithm is to propagate a single wavefront, say, $W(e)$, across a single cell $c$ with $e$ on its boundary.
Depending on whether $c$ is an empty rectangle, there are two cases.
\subsubsection{$c$ is an empty rectangle}
We first consider the case where $c$ is an empty rectangle, i.e., there is no island inside $c$ and $c$ does not intersect any obstacle. Without loss of generality, we assume that $e$ is an edge on the bottom side of $c$, and thus all generators of $W(e)$ are below $e$. Our goal is to compute $W(e,g)$, i.e., the generators of $W(e)$ claiming $g$, for all other edges $g$ of $c$. Our algorithm is similar to the HSY algorithm in the high level but the low level implementations are quite different. The main difference is that each bisector in the HSY algorithm is of constant size while this is not the case in our problem. Due to this, it takes constant time to compute the intersection of two bisectors in the HSY algorithm while in our problem this costs $O(\log n)$ time.
The technical crux of the algorithm is to process the intersections in $c$ among the bisectors of generators of $W(e)$. Since all generators of $W(e)$ are below $e$, their bisectors in $c$ are $y$-monotone by Corollary~\ref{coro:monotone}. This is a critical property our algorithm relies on. Due to the property, we only need to compute $W(e,g)$ for all edges $g$ on the left, right, and top sides of $c$.
Another helpful property is that since we propagate $W(e)$ through $e$ inside $c$, if a generator of $\alpha$ of $W(e)$ claims a point $q\in c$, then the tangent from $q$ to $\alpha$ must cross $e$; we refer it as the {\em propagation property}. Due to this property, the points of $c$ claimed by $\alpha$ must be to the right of the tangent ray from the left endpoint of $e$ to $\alpha$ (the direction of the ray is the from the tangent point to the left endpoint of $e$), as well as to the left of the tangent ray from the right endpoint of $e$ to $\alpha$ (the direction of the ray is the from the tangent point to the right endpoint of $e$). We call the former ray the {\em left bounding ray} of $\alpha$ and the latter the {\em right bounding ray} of $\alpha$.
As such, for the leftmost generator of $W(e)$, we consider its left bounding ray as its left bounding bisector; similarly, for the rightmost generator of $W(e)$, we consider its right bounding ray as its right bounding bisector.
Starting from $e$, we use a horizontal line segment $\ell$ to sweep $c$ upwards until its top side. At any moment during the algorithm, the algorithm maintains a subset $W(\ell)$ of generators of $W(e)$ for $\ell$ by a balanced binary search tree $T(\ell)$; initially $W(\ell)=W(e)$ and $T(\ell)=T(e)$.
Let $[x_1,x_2]\times[y_1,y_2]$ denote the coordinates of $c$. Using the extra fields on the nodes of the tree $T(\ell)$, we compute a maximal prefix $W_1(\ell)$ (resp., $W_2(\ell)$) of generators of $W(\ell)$ such that the bisector events assigned to all generators in it have $x$-coordinates less than $x_1$ (resp., larger than $x_2$). Let $W_m(\ell)$ be the remaining elements of $W(\ell)$. By definition, the first and last generators of $W_m(\ell)$ have their bisector events with $x$-coordinates in $[x_1,x_2]$. As all bisectors are $y$-monotone in $c$, the lowest bisector intersection in $c$ above $\ell$ must be the ``next bisector event'' $b$ associated with a generator in $W_m(\ell)$,
which can be found in $O(\log n)$ time using the tree $T(\ell)$. We advance $\ell$ to the $y$-coordinate of $b$ by removing the generator $\alpha$ associated with the event $b$. Finally, we recompute the next bisector events for the two neighbors of $\alpha$ in $W(\ell)$.
Specifically, let $\alpha_l$ and $\alpha_r$ be the left and right neighboring generators of $\alpha$ in $W(\ell)$, respectively. We need to compute the intersection of the two bounding bisectors of $\alpha_l$,
and update the bisector event of $\alpha_l$ to this intersection. Similarly, we need to compute the intersection of the bounding bisectors of $\alpha_r$,
and update the bisector event of $\alpha_r$ to this intersection. Lemma~\ref{lem:bb-intersection} below shows that each of these bisector intersections can be computed in $O(\log n)$ time by a bisector-intersection operation, using the tentative prune-and-search technique of Kirkpatrick and Snoeyink~\cite{ref:KirkpatrickTe95}.
Note that if $\alpha$ is the leftmost generator, then $\alpha_r$ becomes the leftmost after $\alpha$ is deleted, in which case we compute the left bounding ray of $\alpha_r$ as its left bounding generator.
If $\alpha$ is the rightmost generator, the process is symmetric.
\begin{lemma}{\em\bf (Bisector-bisector intersection operation)}\label{lem:bb-intersection}
Each bisector-bisector intersection operation can be performed in $O(\log n)$ time.
\end{lemma}
\begin{proof}
We are given a horizontal line $\ell$ and three generators $\alpha_1$, $\alpha_2$, and $\alpha_3$ below $\ell$ such that they claim points on $\ell$ in this order. The goal is to compute the intersection of the two bisectors $B(\alpha_1,\alpha_2)$ and $B(\alpha_2,\alpha_3)$ above $\ell$, or determine that such an intersection does not exist. Using the tentative prune-and-search technique of Kirkpatrick and Snoeyink~\cite{ref:KirkpatrickTe95}, we present an $O(\log n)$ time algorithm.
To avoid the lengthy background explanation, we follow the notation
in~\cite{ref:KirkpatrickTe95} without definition.
We will rely on Theorem 3.9 in~\cite{ref:KirkpatrickTe95}. To this
end, we need to define three continuous and decreasing functions $f$, $g$,
and $h$. We define them in a way similar to Theorem~4.10 in~\cite{ref:KirkpatrickTe95} for finding a point equidistant to three convex polygons. Indeed, our problem may be considered as a weighted
case of their problem because each point in the underlying chains of the generators has a
weight that is equal to its weighted distance from its generator.
Let $A$, $B$, and $C$ be the underlying chains of $\alpha_1$, $\alpha_2$, and $\alpha_3$, respectively.
We parameterize over $[0,1]$ each of the three convex chains $A$,
$B$, and $C$ from one end to the other in
clockwise order, i.e.,
each value of $[0,1]$ corresponds to a slope of a tangent at a point on
the chains. For each point $a$ of $A$, we define
$f(a)$ to be the parameter of the point $b\in B$ such that the tangent ray
of $A$ at $a$ (following the designated direction) and the tangent ray of $B$ at $b$ intersect at a point on the bisector $B(\alpha_1,\alpha_2)$ (e.g., see Fig.~\ref{fig:bisectorline} left); if the tangent ray at $a$ does not intersect $B(\alpha_1,\alpha_2)$, then define $f(a)=1$.
In a similar manner, we define $g(b)$ for $b\in B$ with respect to $C$ and define $h(c)$ for $c\in C$
with respect to $A$. One can verify that all three functions are
continuous and decreasing (the tangent at an obstacle vertex of the chains is not unique but the issue can be handled~\cite{ref:KirkpatrickTe95}). The fixed-point of the
composition of the three functions $h\cdot g \cdot f$ corresponds to
the intersection of $B(\alpha_1,\alpha_2)$ and $B(\alpha_2,\alpha_3)$, which can be computed by applying the tentative prune-and-search algorithm of Theorem~3.9~\cite{ref:KirkpatrickTe95}.
To guarantee that the algorithm can be implemented in $O(\log n)$ time, since each of the chains $A$, $B$, and $C$ is represented by an array, we need to show that given
any $a\in A$ and any $b\in B$, we can determine whether $f(a)\geq b$ in $O(1)$
time. This can be done in the same way as in the proof of Lemma~\ref{lem:bl-intersection}.
Similarly, given any $b\in B$ and $c\in C$, we can determine whether $g(b)\geq c$ in $O(1)$ time, and given any $c\in C$ and $a\in A$, we can determine whether $h(c)\geq a$ in $O(1)$ time.
Therefore, applying Theorem 3.9~\cite{ref:KirkpatrickTe95} can compute the intersection of $B(\alpha_1,\alpha_2)$ and $B(\alpha_2,\alpha_3)$ in $O(\log n)$ time.
The above algorithm is based on the assumption that the intersection of the two bisectors exists. However, we do not know whether that is true or not. To determine it, we finally validate the solution as follows. Let $q$ be the point returned by the algorithm. We compute the distances $d(\alpha_i,q)$ for $i=1,2,3$. The point $q$ is a true bisector intersection if and only if the three distances are equal. Finally, we return $q$ if and only if $q$ is above $\ell$.
\end{proof}
The algorithm finishes once $\ell$ is at the top side of $c$. At this moment, no bisector events of $W(\ell)$ are in $c$. Finally, we run the following {\em wavefront splitting step} to split $W(\ell)$ to obtain $W(e,g)$ for all edges $g$ on the left, right, and top sides $c$. Our algorithm relies on the following observation.
Let $\zeta$ be the union of the left, top, and right sides of $c$.
\begin{lemma}
The list of generators of $W(\ell)$ are exactly those in $W(\ell)$ claiming $\zeta$ in order.
\end{lemma}
\begin{proof}
It suffices to show that during the sweeping algorithm whenever a bisector $B(\alpha_1,\alpha_2)$ of two generators $\alpha_1$ and $\alpha_2$ intersects $\zeta$, it will never intersect $\partial c$ again. Let $q$ be such an intersection. Let $\zeta_l$, $\zeta_t$, and $\zeta_r$ be the left, top, and right sides of $c$, respectively.
If $q$ is on $\zeta_t$, then since both $\alpha_1$ and $\alpha_2$ are below $e$, they are also below $\zeta_t$. By Lemma~\ref{lem:intersection}, $B(\alpha_1,\alpha_2)$ will not intersect the supporting line of $\zeta_t$ again and thus will not intersect $\partial c$ again.
If $q$ is on $\zeta_l$, then we claim that both generators $\alpha_1$ and $\alpha_2$ are to the right of the supporting line $\ell_l$ of $\zeta_l$. Indeed, since both generators claim $q$, the bounding rays (i.e., the left bounding ray of the leftmost generator of $W(\ell)$ and the right bounding ray of the rightmost generator of $W(\ell)$ during the sweeping algorithm) guarantee the propagation property: the tangents from $q$ to the two generators must cross $e$. Therefore, both generators must be to the right of $\ell_l$. By Lemma~\ref{lem:intersection}, $B(\alpha_1,\alpha_2)$ will not intersect the supporting line of $\zeta_l$ again and thus will not intersect $\partial c$ again.
If $q$ is on $\zeta_r$, the analysis is similar to the above second case.
\end{proof}
In light of the above lemma, our wavefront splitting step algorithm for computing $W(e,g)$ of all edges $g\in \zeta$ works as follows. Consider an edge $g\in \zeta$. Without loss of generality, we assume that the points of $\zeta$ are clockwise around $c$ so that we can talk about their relative order.
Let $p_l$ and $p_r$ be the front and rear endpoints of
$g$, respectively. Let $\alpha_l$ and $\alpha_r$ be the generators of
$W(\ell)$ claiming $p_l$ and $p_r$, respectively. Then all generators
of $W(\ell)$ to the left of $\alpha_l$ including $\alpha_l$ form
the wavefront for all edges of $\zeta$ in the front of $g$; all generators of $W(\ell)$ to
the right of $\alpha_r$ including $\alpha_r$ form the wavefront for all edges of $\zeta$ after $g$; all generators of $W(\ell)$ between $\alpha_l$ and $\alpha_r$ including $\alpha_l$ and $\alpha_r$ form
$W(e,g)$. Hence, once $\alpha_l$ and $\alpha_r$ are known, $W(e,g)$ can be obtained
by splitting $W(\ell)$ in $O(\log n)$ time. It remains to compute $\alpha_l$ and $\alpha_r$.
Below, we only discuss how to compute the generator $\alpha_l$ since $\alpha_r$ can be computed analogously.
Starting from the root $v$ of $T(\ell)$, we determine the intersection $q$ between $B(\alpha_1,\alpha_2)$ and $\zeta$, where $\alpha_1$ is the rightmost generator in the left subtree of $v$ and $\alpha_2$ is the leftmost generator of the right subtree of $v$.
If $q$ is in the front of $g$ on $\zeta$, then we proceed to the right subtree of $v$; otherwise, we proceed to the left subtree of $v$.
It is easy to see that the runtime of the algorithm is bounded by $O(\eta\cdot \log n)$ time, where $\eta$ is the time for computing $q$. In the HSY algorithm, each bisector is of constant size and an oracle is assumed to exist that can compute $q$ in $O(1)$ time. In our problem, since a bisector may not be of constant size, it is not clear how to bound $\eta$ by $O(1)$. But $\eta$ can be bounded by $O(\log n)$ using the bisector-line intersection operation in Lemma~\ref{lem:bl-intersection}. Thus, $\alpha_l$ can be computed in $O(\log^2 n)$ time. However, this is not sufficient for our purpose, as this would lead to an overall $O(n+h\log^2 h)$ time algorithm. We instead use the following {\em binary search plus bisector tracing} approach.
During the wavefront expansion algorithm, for each pair of neighboring generators $\alpha=(A,a)$ and $\alpha'=(A',a')$ in a wavefront (e.g., $W(e)$), we maintain a special point $z(\alpha,\alpha')$ on the bisector $B(\alpha,\alpha')$. For example, in the above sweeping algorithm, whenever a generator $\alpha$ is deleted from $W(\ell)$ at a bisector event $b=B(\alpha_l,\alpha)\cap B(\alpha,\alpha_r)$, its two neighbors $\alpha_l$ and $\alpha_r$ now become neighboring in $W(\ell)$. Then, we initialize $z(\alpha_l,\alpha_r)$ to $b$ (the tangent points from $b$ to $\alpha_l$ and $\alpha_r$ are also associated with $b$). During the algorithm, the point $z(\alpha_l,\alpha_r)$ will move on $B(\alpha,\alpha')$ further away from the two defining generators $\alpha$ and $\alpha'$ and the movement will trace out the hyperbolic-arcs of the bisector. We call $z(\alpha,\alpha')$ the {\em tracing-point} of $B(\alpha,\alpha')$.
Our algorithm maintains a variant that the tracing point of each bisector of $W(\ell)$ is below the sweeping line $\ell$ (initially, the tracing point of each bisector of $W(e)$ is below $e$).
With the help of the above $z$-points, we compute the generator $\alpha_l$ as follows. Like the above algorithm, starting from the root $v$ of $T(\ell)$, let $\alpha_1$ and $\alpha_2$ be the two generators as defined above.
To compute the intersection $q$ between $B(\alpha_1,\alpha_2)$ and $\zeta$, we trace out the bisector $B(\alpha_1,\alpha_2)$ by moving its tracing-point $z(\alpha_1,\alpha_2)$ upwards (each time trace out a hyperbolic-arc of $B(\alpha_1,\alpha_2)$) until the current tracing hyperbolic-arc intersects $\zeta$ at $q$.
If $q$ is in the front of $e$ on $\zeta$, then we proceed to the right subtree of $v$; otherwise, we proceed to the left subtree of $v$.
After $W(e,g)$ is obtained, we compute $W(e,g')$ for other edges $g'$ on $\zeta$ using the same algorithm as above. For the time analysis, observe that each bisector hyperbolic-arc will be traced out at most once in the wavefront splitting step for all edges of $\zeta$ because the tracing point of each bisector will never move ``backwards''.
This finishes the algorithm for propagating $W(e)$ through the cell $c$. Except the final wavefront splitting step, the algorithm runs in $O((1+h_c)\log n)$ time, where $h_c$ is the number of bisector events in $c$. Because $c$ has $O(1)$ edges, the wavefront splitting step takes $O(\log n + n_c)$ time, where $n_c$ is the number of hyperbolic-arcs of bisectors that are traced out.
\subsubsection{$c$ is not an empty rectangle}
We now discuss the case in which the cell $c$ is not an empty rectangle. In this case, $c$ has a square hole inside or/and the boundary of $c$ contains obstacle edges.
Without loss of generality, we assume that $e$ is on the bottom side of $c$.
If $c$ contains a square hole, then we partition $c$ into four subcells by cutting $c$ with two lines parallel to $e$, each passing through an edge of the hole. If $c$ has obstacle edges on its boundary, recall that these obstacles edges belong to $O(1)$ convex chains (each of which is a fragment of an elementary chain); we further partition $c$ by additional edges parallel to $e$, so that each resulting subcell contains at most two convex chains, one the left side and the other on the right side. Since $\partial c$ has $O(1)$ convex chains, $O(1)$ additional edges are sufficient to partition $c$ into $O(1)$ subcells as above. Then, we propagate $e$ through the subcells of $c$, one by one. In the following, we describe the algorithm for one such subcell. By slightly abusing the notation, we still use $c$ to denote the subcell with $e$ on its bottom side.
Since $\partial c$ has obstacle edges, the propagation algorithm becomes more complicated. As in the HSY algorithm, comparing with the algorithm for the previous case, there are two new bisector events.
\begin{itemize}
\item
First, a bisector may intersect a convex chain (and thus intersect an obstacle). The HSY algorithm does not explicitly compute these bisector events because such an oracle is not assumed to exist. In our algorithm, however, because the obstacles in our problem are polygonal, we can explicitly determine these events without any special assumption. This is also a reason that the high-level idea of our algorithm is simpler than the HSY algorithm.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=1.7in]{creategenerator10.pdf}
\caption{\footnotesize Illustrating the creation of a new generator at $q'$.}
\label{fig:creategenerator10}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
\item
Second, new generators may be created at the convex chains. We still sweep a horizontal line $\ell$ from $e$ upwards. Let $W(\ell)$ be the current wavefront at some moment during the algorithm. Consider two neighboring generators $\alpha_1$ and $\alpha_2$ in $W(\ell)$ with $\alpha_1$ on the left of $\alpha_2$. We use $\zeta_l$ to denote the convex chain on the left side of $c$. Let $q'$ be the tangent point on $\zeta_l$ of the common tangent between $\zeta_l$ and $\alpha_1$ and let $q''$ be the tangent point on $\alpha_1$ (e.g., see Fig.~\ref{fig:creategenerator10}). If $d(\alpha_1,q')<d(\alpha_2,q')$, then a new generator $\alpha$ on $\zeta_l$ with initial vertex $q'$ and weight equal to $d(\alpha_1,q')$ is created (designated counterclockwise direction) and inserted into $W(\ell)$ right before $\alpha_1$. The bisector $B(\alpha,\alpha_1)$ is the ray emanating from $q'$ and extending away from $q''$. The region to the left of the ray has $\alpha$ as its predecessor. When the sweeping line $\ell$ is at $q'$, all wavelets in $W(\ell)$ to the left of $\alpha_1$ have already collided with $\zeta_l$ and thus the first three generators of $W(\ell)$ are $\alpha$, $\alpha_1$, and $\alpha_2$.
\end{itemize}
In what follows, we describe our sweeping algorithm to propagate $W(e)$ through $c$.
We begin with an easier case where only the left side of $c$ is a convex chain, denoted by $\zeta_l$ (and the right side is a vertical transparent edge, denoted by $\zeta_r$). We use $\zeta_t$ to denote the top side of $c$, which is a transparent edge.
As in the previous case, we sweep a line $\ell$ from $e$ upwards until the top side $\zeta_t$.
During the algorithm, we maintain a list $W(\ell)$ of generators by a balanced binary search tree $T(\ell)$. Initially, $W(\ell)=W(e)$ and $T(\ell)=T(e)$.
We compute the intersection $q$ of the convex chain $\zeta_l$ and the bisector $B(\alpha_1,\alpha_2)$, for the leftmost bisectors of $\alpha_1$ and $\alpha_2$ of $W(\ell)$. We call it the {\em bisector-chain intersection operation}. The following lemma shows that this operation can be performed in $O(\log n)$ time.
\begin{lemma}{\em\bf (Bisector-chain intersection operation)}\label{lem:bc-intersection}
Each bisector-chain intersection operation can be performed in $O(\log n)$ time.
\end{lemma}
\begin{proof}
We are given a convex chain $\zeta_l$ above a horizontal line $\ell$ and two generators $\alpha_1=(A_1,a_1)$ and $\alpha_2=(A_2,a_2)$ below $\ell$ such that they claim points on $\ell$ in this order. The goal is to compute the intersection of $B(\alpha_1,\alpha_2)$ and $\zeta_l$, or determine that they do not intersect. In the following, using the tentative prune-and-search technique of Kirkpatrick and Snoeyink~\cite{ref:KirkpatrickTe95}, we present an $O(\log n)$ time algorithm.
To avoid the lengthy background explanation, we follow the notation
in~\cite{ref:KirkpatrickTe95} without definition.
We will rely on Theorem 3.9 in~\cite{ref:KirkpatrickTe95}. To this
end, we need to define three continuous and decreasing functions $f$, $g$,
and $h$.
Suppose $q$ is the intersection of $B(\alpha_1,\alpha_2)$ and $\zeta_l$. Let $p_1$ and $p_2$ be the tangent points from $q$ to $A_1$ and $A_2$, respectively. Then, $\overline{qp_1}$ (resp., $\overline{qp_2}$) does not intersect $\zeta_l$ other than $q$. We determine the portion $C$ of $\zeta_l$ such that for each point $p\in C$, its tangent to $A_1$ does not intersect $\zeta_l$ other than $p$. Hence, $q\in C$. $C$ can be determined by computing the common tangents between $\zeta_l$ and $A_1$, which can be done in $O(\log n)$ time~\cite{ref:GuibasCo91,ref:OvermarsMa81}. Also, we determine the portion $B$ of $A_2$ such that the tangent ray at any point of $B$ must intersect $C$. This can be done by computing the common tangents between $C$ and $A_2$ in $O(\log n)$ time~\cite{ref:GuibasCo91,ref:OvermarsMa81}. Let $A=A_1$.
We parameterize over $[0,1]$ each of the three convex chains $A$, $B$, and $C$ from one end to the other in
clockwise order, i.e.,
each value of $[0,1]$ corresponds to a slope of a tangent at a point on
the chains $A$ and $B$, while each value of $[0,1]$ corresponds to a point of $C$.
For each point $a$ of $A$, we define
$f(a)$ to be the parameter of the point $b\in B$ such that the tangent ray
of $A$ at $a$ (following the designated direction of $\alpha_1$) and the tangent ray of $B$ at $b$ intersect at a point on the bisector $B(\alpha_1,\alpha_2)$ (e.g., see Fig.~\ref{fig:bisectorline} left); if the tangent ray at $a$ does not intersect $B(\alpha_1,\alpha_2)$, then define $f(a)=1$.
For each point $b$ of $B$, define $g(b)$ to be the parameter of the point $c\in C$ such that $\overline{cb}$ is tangent to $B$ at $b$ (e.g., see Fig.~\ref{fig:bcintersection} left); note that by the definition of $B$, the tangent ray from any point of $B$ must intersect $C$ and thus such a point $c\in C$ must exist.
For each point $c\in C$, define $h(c)$ to be the parameter of the point of $a\in A$ such that $\overline{ac}$ is tangent to $A$ at $a$ (e.g., see Fig.~\ref{fig:bcintersection} right); note that by the definition of $C$, such a point $a\in A$ must exist and $\overline{ac}$ does not intersect $C$ other than $c$. One can verify that all three functions are
continuous and decreasing (the tangent at an obstacle vertex of the chains is not unique but the issue can be handled~\cite{ref:KirkpatrickTe95}). The fixed-point of the
composition of the three functions $h\cdot g \cdot f$ corresponds to
the intersection $q$ of $B(\alpha_1,\alpha_2)$ and $\zeta_l$, which can be computed by applying the tentative prune-and-search algorithm of Theorem 3.9~\cite{ref:KirkpatrickTe95}.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=1.5in]{bcintersection.pdf}
\caption{\footnotesize Illustrating the definitions of $g(b)$ and $h(c)$.}
\label{fig:bcintersection}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
To make sure that the algorithm can be implemented in $O(\log n)$ time, since each convex chain is part of an elementary chain and thus is represented by an array, it suffices to show the following: (1) given
any $a\in A$ and any $b\in B$, whether $f(a)\geq b$ can be determined in $O(1)$ time; (2) given any $b\in B$ and any $c\in C$, whether $g(b)\geq c$ can be determined in $O(1)$ time; (3) given any $c\in C$ and any $a\in A$, whether $h(c)\geq a$ can be determined in $O(1)$ time. We prove them below.
Indeed, for (1), it can be done in the same way as in the proof of
Lemma~\ref{lem:bl-intersection}.
For (2), $g(b)\geq c$ if and only if $c$ is to the right of the tangent ray of $B$ at $b$, which can be easily determined in $O(1)$ time.
For (3), $h(c)\geq a$ if and only $c$ is to the right of the tangent ray of $A$ at $a$, which can be easily determined in $O(1)$ time.
Therefore, applying the tentative prune-and-search technique in Theorem 3.9~\cite{ref:KirkpatrickTe95} can compute $q$ in $O(\log n)$ time.
Note that the above algorithm is based on the assumption that the intersection of $B(\alpha_1,\alpha_2)$ and $\zeta_l$ exists. However, we do not know whether this is true or not. To determine that, we finally validate the solution as follows. Let $q$ be the point returned by the algorithm. We first determine whether $q\in \zeta_l$. If not, then the intersection does not exist. Otherwise, we further compute the two distances $d(\alpha_i,q)$ for $i=1,2$ in $O(\log n)$ time. If the distances are equal, then $q$ is the true intersection; otherwise, the intersection does not exist.
\end{proof}
If the intersection $q$ of $\zeta_l$ and $B(\alpha_1,\alpha_2)$ does not exist, then we compute the tangent between $\zeta_l$ and $\alpha_1$, which can be done in $O(\log n)$ time~\cite{ref:OvermarsMa81}; let $q'$ be the tangent point at $\zeta_l$.
Regardless whether $q$ exists or not, we compute the lowest bisector intersection $b$ in $c$ above $\ell$ in the same way as in the algorithm for the previous case where $c$ is an empty rectangle. Depending on whether $q$ exists or not, we proceed as follows. For any point $p$ in the plane, let $y(p)$ denote the $y$-coordinate of $p$.
\begin{enumerate}
\item
If $q$ exists, then depending on whether $y(q)\leq y(b)$, there are two subcases.
If $y(q)\leq y(b)$, then we process the bisector event $q$: remove $\alpha_1$ from $W(\ell)$ and then recompute $q$, $q'$, and $b$. Otherwise, we process the bisector event at $b$ in the same way as in the previous case and then recompute $q$, $q'$, and $b$.
\item
If $q$ does not exist, then depending on whether $y(b)\leq y(q')$, there are two subcases. If $y(b)\leq y(q')$, then we process the bisector event at $b$ in the same way as before and then recompute $q$, $q'$, and $b$.
Otherwise, we insert a new generator $\alpha=(A,q')$ to $W(\ell)$ as the leftmost generator, where $A$ is the fragment of the elementary chain containing $\zeta_l$ from $q'$ counterclockwise to the end of the chain, and $\alpha$ is designated the counterclockwise direction of $A$ and the weight of $q'$ is $d(\alpha_1,q')$; e.g., see Fig.~\ref{fig:creategenerator10}. The ray from $q'$ in the direction from $q''$ to $q'$ is the bisector of $\alpha$ and $\alpha_1$, where $q''$ is the tangent point on $\alpha_1$ of the common tangent between $\alpha_1$ and $\zeta_l$. We initialize the tracing-point $z(\alpha,\alpha_1)$ of $B(\alpha,\alpha_1)$ to $q'$. Finally, we recompute $q$, $q'$, and $b$.
\end{enumerate}
Once the sweep line $\ell$ reaches the top side $\zeta_t$ of $c$, the algorithm stops. Finally, as in the previous case, we run a wavefront
splitting step. Because the left side $\zeta_l$ consists of obstacle
edges, we split $W(\ell)$ to compute $W(e,g)$ for all transparent edges $g$ on the top side $\zeta_t$ and the right side $\zeta_r$ of $c$.
The algorithm is the same as the previous case.
The above discusses the case that only the left side $\zeta_l$ of $c$ is a
convex chain. For the general case where both the left and right
sides of $c$ are convex chains, the algorithm is similar. The
difference is that we have to compute a point $p$ corresponding to $q$
and a point $p'$ corresponding to $q'$ on the right side $\zeta_r$ of
$c$. More specifically, $p$ is the intersection of
$B(\alpha_2',\alpha_1')$ with $\zeta_r$, where $\alpha_2'$ and $\alpha_1'$
are the two rightmost generators of $W$. If $p$ does not exist, then we
compute the common tangent between $\zeta_r$ and $\alpha_1'$, and $p'$ is the tangent
point on $\zeta_r$.
In the following, if $q$ does not exist, we let $y(q)$ be
$\infty$; otherwise, $q'$ is not needed and we let $y(q')$ be
$\infty$. We apply the same convention to $p$ and $p'$. We define $b$
as the bisector event in the same way as before. In each step, we
process the lowest point $r$ of $\{q,q',b,p,p'\}$. If $r$
is $q$ or $p$, we process it in the same way as before for $q$.
If $r$ is $q'$ or $p'$, we process it in the same way as before for $q'$.
If $r$ is $b$, we process it in the same way as before.
After processing $r$, we recompute the five points. Each step
takes $O(\log n)$ time.
After the sweep line $\ell$ reaches the top side $\zeta_t$ of $c$,
$W(\ell)$ is $W(e,\zeta_t)$ for the top side $\zeta_t$ of $c$ because both the
left and right sides of $c$ are obstacle edges.
Finally, we run the wavefront splitting step on $W(\ell)$ to compute the wavefronts $W(e,g)$ for all transparent edges $g$ on $\zeta_t$.
In summary, propagating $W(e)$ through $c$ takes $O((1+h_c)\cdot
\log n+n_c)$ time, where $h_c$ is the number of bisector events (including both the bisector-bisector intersection events and the bisector-obstacle intersection events) and $n_c$ is the number of hyperbolic-arcs of bisectors that are traced out in the wavefront splitting step.
We use the following lemma to summarize the algorithm for both cases (i.e., regardless whether $c$ is an empty rectangle or not).
\begin{lemma}\label{lem:propagation}
Suppose $W(e)$ is a wavefront on a transparent edge of a cell $c$ of the subdivision $\mathcal{S}'$. Then, $W(e)$ can be propagated through $c$ to all other transparent edges of $c$ in $O((1+h_c)\log n+n_c)$ time, where $h_c$ is the number of bisector events (including both the bisector-bisector intersection events and bisector-obstacle intersection events) and $n_c$ is the number of hyperbolic-arcs of bisectors that are traced out in the wavefront splitting step.
\end{lemma}
\subsection{Time analysis}
\label{sec:time}
In this section, we show that the running time of our wavefront expansion algorithm described above is bounded by $O(n+h\log h)$.
For this and also for the purpose of constructing the shortest path map $\mbox{$S\!P\!M$}'(s)$ later in Section~\ref{sec:spm}, as in the HS algorithm, we mark generators in the way that if a generator $\alpha$ is involved in a true bisector event of $\mbox{$S\!P\!M$}(s)$ (either a bisector-bisector intersection or a bisector-obstacle intersection) in a cell $c$ of the subdivision $\mathcal{S}'$, then $\alpha$ is guaranteed to be in a set of marked generators for $c$ (but a marked generator for $c$ may not actually participate in a true bisector event in $c$).
The generator marking rules are presented below, which are consistent with those in the HS algorithm.
\paragraph{Generator marking rules:}
\begin{enumerate}
\item
For any generator $\alpha=(A,a)$, if its initial vertex $a$ lies in a cell $c$, then mark $\alpha$ in $c$.
\item
Let $e$ be a transparent edge and let $W(e)$ be a wavefront coming from some generator $\alpha$'s side of $e$.
\begin{enumerate}
\item
If $\alpha$ claims an endpoint $b$ of $e$ in $W(e)$, or if it would do so except for an artificial wavefront, then mark $\alpha$ in all cells $c$ incident to $b$.
\item
If $\alpha$'s claim in $W(e)$ is shortened or eliminated by an artificial wavelet, then mark $\alpha$ for $c$, where $c$ is the cell having $e$ as an edge and on $\alpha$'s side of $e$.
\end{enumerate}
\item
Let $e$ and $g$ be two transparent edges with $g\in output(e)$. Mark a generator $\alpha$ of $W(e)$ in both cells having $e$ as an edge if one of the following cases happens:
\begin{enumerate}
\item
$\alpha$ claims an endpoint of $g$ in $W(e,g)$;
\item
$\alpha$ participates in a bisector event either during the wavefront propagation procedure for computing $W(e,g)$ from $W(e)$, or during the wavefront merging procedure for computing $W(g)$. Note that $\alpha$ is also considered to participate in a bisector event if its claim on $g$ is shortened or eliminated by an artificial wavelet.
\end{enumerate}
\item
If a generator $\alpha$ of $W(e)$ claims part of an obstacle edge during the wavefront propagation procedure for propagating $W(e)$ toward $output(e)$ (this includes the case in which $\alpha$ participates in a bisector-obstacle intersection event), then mark $\alpha$ in both cells having $e$ as an edge.
\end{enumerate}
Note that each generator may be marked multiple times and each mark applies to one instance of the generator.
\begin{lemma}\label{lem:numgen}
The total number of marked generators during the algorithm is at most $O(h)$.
\end{lemma}
Because the proof of Lemma~\ref{lem:numgen} is quite technical and lengthy, we devote the entire Section~\ref{sec:lemnumgen} to it. In the rest of this subsection, we use Lemma~\ref{lem:numgen} to show that the running time of our wavefront expansion algorithm is bounded by $O(n+h\log h)$. Our goal is to prove the following lemma.
\begin{lemma}\label{lem:algocom}
The wavefront expansion algorithm runs in $O(n+h\log h)$ time and space.
\end{lemma}
First of all, by Lemma~\ref{lem:subalgo}, constructing the conforming subdivision $\mathcal{S}'$ can be done in $O(n+h\log h)$ time and $O(n)$ space.
The wavefront expansion algorithm has two main subroutines: the wavefront merging procedure and the wavefront propagation procedure.
The wavefront merging procedure is to construct $W(e)$ based on $W(f,e)$ for the edges $f\in input(e)$. By Lemma~\ref{lem:merge}, this step takes $O((1+k)\log n)$ time, where $k$ is the total number of generators in all wavefronts $W(f,e)$ that are absent from $W(e)$. According to the algorithm, if a generator $\alpha$ is absent from $W(e)$, it must be deleted at a bisector event. Thus, $\alpha$ must be marked by Rule~3(b). Due to Lemma~\ref{lem:numgen}, the total sum of $k$ in the entire algorithm is $O(h)$. As such, the wavefront merging procedure in the entire algorithm takes $O(h\log n)$ time in total.
The wavefront propagation procedure is to compute $W(e,g)$ by propagating $W(e)$ to all edges $g\in output(e)$ either through $\mathcal{U}(g)$ or $\mathcal{U}(e)$.
By Lemma~\ref{lem:propagation}, the running time of the procedure is $O((1+h_c)\log n+n_c)$ time, where $h_c$ is the number of bisector events (including both the bisector-bisector intersection events and bisector-obstacle intersection events) and $n_c$ is the number of hyperbolic-arcs of bisectors that are traced out in the wavefront splitting step.
For each bisector-bisector intersection event, at least one involved generator is marked by Rule~3(b). For each bisector-obstacle intersection event, at least one involved generator is marked by Rule~4. Hence, by Lemma~\ref{lem:numgen}, the total sum of $h_c$ in the entire algorithm is $O(h)$. In addition, Lemma~\ref{lem:bisectorvertices} below shows that the total sum of $n_c$ in the entire algorithm is $O(n)$. Therefore, the wavefront propagation procedure in the entire algorithm takes $O(n+h\log n)$ time in total.
\begin{lemma}\label{lem:bisectorvertices}
The total number of traced hyperbolic-arcs of the bisectors in the entire algorithm is $O(n)$.
\end{lemma}
\begin{proof}
First of all, notice that each extension bisector consists of a single hyperbolic-arc, which is a straight line. As each generator is marked by Rule~1, by Lemma~\ref{lem:numgen}, the total number of generators created in the algorithm is $O(h)$. Since each generator can define at most one extension bisector, the number of hyperbolic-arcs on extension bisectors is at most $O(h)$. In the following, we focus on hyperbolic-arcs of non-extension bisectors. Instead of counting the number of traced hyperbolic-arcs, we will count the number of their endpoints.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=1.7in]{tracepoint.pdf}
\caption{\footnotesize Illustrating the definitions of $u$, $v$, $q$, and $o$.}
\label{fig:tracepoint}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
Consider a hyperbolic-arc endpoint $o$ that is traced out. According to our algorithm, $o$ is traced out during the wavefront propagation procedure for propagating $W(e)$ to compute $W(e,g)$ for some transparent edge $e$ and $g\in output(e)$. Suppose $o$ belongs to a non-extension bisector $B(\alpha,\alpha')$ of two generators $\alpha$ and $\alpha'$ in $W(e)$. Then, $o$ must be defined by an obstacle edge $\overline{uv}$ of either $\alpha$ or $\alpha'$, i.e., the ray $\rho(u,v)$ emanating from $v$ along the direction from $u$ to $v$ (which is consistent with the designated direction of the generator that contains $\overline{uv}$) hits $B(\alpha,\alpha')$ at $o$ (e.g., see Fig~\ref{fig:tracepoint}). Without loss of generality, we assume that $\overline{uv}$ belongs to $\alpha$. In the following, we argue that $\overline{uv}$ can define $O(1)$ hyperbolic-arc endpoints that are traced out during the entire algorithm (a hyperbolic-arc endpoint defined by $\overline{uv}$ is counted twice is it is traced out twice), which will prove Lemma~\ref{lem:bisectorvertices} as there are $O(n)$ obstacle edges in total.
We first discuss some properties. Since $o\in B(\alpha,\alpha')$, both $\alpha$ and $\alpha'$ claim $o$. As both generators are in $W(e)$, the ray $\rho(u,v)$ must cross $e$, say, at a point $q$ (e.g., see Fig~\ref{fig:tracepoint}). Because $o$ is traced out when we propagate $W(e)$ to compute $W(e,g)$, $\overline{oq}$ must be in either $\mathcal{U}(g)$ or $\mathcal{U}(e)$, i.e., $\overline{oq}\subseteq \mathcal{U}(e)\cup \mathcal{U}(g)$. We call $(e,g)$ the {\em defining pair} of $o$ for $\overline{uv}$.
According to our algorithm, during the propagation from $W(e)$ to $g$, $\overline{uv}$ defines only one hyperbolic-arc endpoint, because it is uniquely determined by the wavefront $W(e)$. As such, to prove that $\overline{uv}$ can define $O(1)$ hyperbolic-arc endpoints that are traced out during the entire algorithm, it suffices to show that there are at most $O(1)$ defining pairs for $\overline{uv}$. Let $\Pi$ denote the set of all such defining pairs.
We prove $|\Pi|=O(1)$ below.
For each pair $(e',g')\in \Pi$, according to the above discussion, $\overline{uv}$ and $(e',g')$ define a hyperbolic-arc endpoint $o'$ such that $o'$ is on the ray $\rho(u,v)$ and $\overline{vo'}$ crosses $e'$ at a point $q'$. Without loss of generality, we assume that $(e,g)$ is a pair that minimizes the length $|\overline{vq'}|$. Let $W(e')$ refer to the wavefront at $e'$ whose propagation to $g'$ traces out $o'$.
We partition $\Pi$ into two subsets $\Pi_1$ and $\Pi_2$, where $\Pi_1$ consists of all pairs $(e',g')$ of $\Pi$ such that $\overline{qo}$ intersects $e'$ and $\Pi_2=\Pi\setminus \Pi_1$. Since $\overline{oq}\subseteq \mathcal{U}(e)\cup \mathcal{U}(g)$, each well-covering region contains $O(1)$ cells, and $|output(e')|=O(1)$ for each transparent edge $e'$, the size of $\Pi_1$ is $O(1)$.
For $\Pi_2$, we further partition it into two subsets $\Pi_{21}$ and
$\Pi_{22}$, where $\Pi_{21}$ consists of all pairs $(e',g')$ of
$\Pi_2$ such that $e$ is in the well-covering region $\mathcal{U}(e')$ of
$e'$ or $e'\in \mathcal{U}(e)\cup \mathcal{U}(g)$, and $\Pi_{22}=\Pi_2\setminus \Pi_{21}$. Since $e$ is in
$\mathcal{U}(e')$ for a constant number of transparent edges $e'$, each of $\mathcal{U}(e)$ and $\mathcal{U}(g)$ contains $O(1)$ cells, and $|output(e')|=O(1)$ for each $e'$, it holds that $|\Pi_{21}|=O(1)$.
In the following, we argue that $\Pi_{22}=\emptyset$, which will prove $|\Pi|=O(1)$.
Assume to the contrary that $|\Pi_{22}|\neq \emptyset$ and let $(e',g')$ be a pair of $\Pi_{22}$. Since $(e',g')\in \Pi_2$, by the definition of $\Pi_2$, $e'$ does not intersect $\overline{oq}$. By the definition of $e$, $\overline{vq}\setminus\{q\}$ does not intersect $e'$. Recall that $q'$ is the intersection of $e'$ and $\rho(u,v)$. Therefore, the points $v$, $q$, $o$, $q'$, and $o'$ appear on the ray $\rho(u,v)$ in this order (e.g., see Fig~\ref{fig:tracepoint10}). Further, since $(e',g')\in \Pi_{22}$, by the definition of $\Pi_{22}$, $e$ is not in $\mathcal{U}(e')$ and $e'$ is not in $\mathcal{U}(e)\cup \mathcal{U}(g)$.
\begin{figure}[t]
\begin{minipage}[t]{0.49\textwidth}
\begin{center}
\includegraphics[height=2.2in]{tracepoint10.pdf}
\caption{\footnotesize Illustrating the definitions of $u$, $v$, $q$, and $o$.}
\label{fig:tracepoint10}
\end{center}
\end{minipage}
\hspace{0.02in}
\begin{minipage}[t]{0.49\textwidth}
\begin{center}
\includegraphics[height=1.8in]{tracepoint20.pdf}
\caption{\footnotesize The dashed (red) path is $\pi_{W(e)}(s,v)$, which crosses $e'$ at $b$.}
\label{fig:tracepoint20}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
Without loss of generality, we assume that $e'$ is horizontal and the
wavefront $W(e')$ is from below $e'$ (thus $v$ is below $e'$ while
$o'$ is above $e'$). Let $\mathcal{F}'$ be the modified free space by replacing $e'$ with an opaque edge
of open endpoints.
Since the generator in $W(e')$ that contains $v$ claims $q'$, $\pi'(s,q')$ is a shortest path from $s$ to $q'$ in $\mathcal{F}'$, where $\pi'(s,q')$ is the path following the
wavefront $W(e')$. Since $v$ is in the generator of $W(e')$ that
claims $q'$, $v$ is the anchor of $q'$ in $\pi'(s,q')$, i.e., the edge
of the path incident to $q'$ is $\overline{vq'}$. Let $\pi'(s,v)$ be
the sub-path of $\pi'(s,q')$ between $s'$ and $v$. Then, $\pi'(s,v)$ is a shortest path from $s$ to $v$ in $\mathcal{F}'$.
Recall that $v$ is in the generator $\alpha$ of $W(e)$. Let $\pi_{W(e)}(s,v)$ be the path from $s$ to $v$ following $W(e)$. We claim that $|\pi'(s,v)|=|\pi_{W(e)}(s,v)|$. Assume to the contrary this is not true. Then, either $|\pi'(s,v)|<|\pi_{W(e)}(s,v)|$ or $|\pi_{W(e)}(s,v)|<|\pi'(s,v)|$.
\begin{itemize}
\item
If $|\pi'(s,v)|<|\pi_{W(e)}(s,v)|$, then since $\pi_{W(e)}(s,v)$ is a shortest path from $s$ to $v$ in the modified free space by considering $e$ as an opaque edge of open endpoints, $\pi'(s,v)$ must cross the interior of $e$. This means that $\pi'(s,q')=\pi'(s,v)\cup \overline{vq'}$ crosses $e$ twice. Because $\pi'(s,q')$ is a shortest path in $\mathcal{F}'$ and $e\in \mathcal{F}'$, it cannot cross $e$ twice, a contradiction.
\item
If $|\pi_{W(e)}(s,v)|<|\pi'(s,v)|$, then since $\pi'(s,v)$ is a shortest path from $s$ to $v$ in $\mathcal{F}'$, the path $\pi_{W(e)}(s,v)$ cannot be in $\mathcal{F}'$ and thus must cross the interior of $e'$, say, at a point $b$ (e.g., see Fig~\ref{fig:tracepoint10}). Let $\pi_{W(e)}(s,b)$ be the sub-path of $\pi_{W(e)}(s,v)$ between $s$ and $b$. Let $\pi_{W(e)}(b,q')$ be a shortest path from $b$ to $q'$ along $e'$ by considering it as an opaque edge with open endpoints. Note that if $b$ and $q'$ are on different sides of $e'$, then $\pi_{W(e)}(b,q')$ must be through an open endpoint of $e'$. It is not difficult to see that $|\pi_{W(e)}(b,q')|\leq |e'|$. Let $\pi_{W(e)}(s,q')$ be the concatenation of $\pi_{W(e)}(s,b)$ and $\pi_{W(e)}(b,q')$. Hence, $\pi_{W(e)}(s,q')$ is a path in $\mathcal{F}'$. Since $\pi'(s,q')$ is a shortest path from $s$ to $q'$ in $\mathcal{F}'$, it holds that $|\pi'(s,q')|\leq |\pi_{W(e)}(s,q')|$.
Notice that $|\pi_{W(e)}(s,q')|=|\pi_{W(e)}(s,b)|+|\pi_{W(e)}(b,q')|\leq |\pi_{W(e)}(s,v)|+|e'|< |\pi'(s,v)|+|e'|$.
On the other hand, $|\pi'(s,q')|=|\pi'(s,v)|+|\overline{vq'}|\geq |\pi'(s,v)|+|\overline{qq'}|$. We claim that $|\overline{qq'}|\geq 2|e'|$. Indeed, since $e$ is outside $\mathcal{U}(e')$, $q\in e$, and $q'\in e'$, $\overline{qq'}$ must cross $\partial \mathcal{U}(e')$ at a point $b'$. By the property of well-covering regions of $\mathcal{S}'$, $|\overline{b'q'}|\geq 2|e'|$. Since $|\overline{qq'}|\geq |\overline{b'q'}|$, we obtain $|\overline{qq'}|\geq 2|e'|$. In light of the claim, we have $|\pi'(s,q')| \geq |\pi'(s,v)|+2|e'|>|\pi'(s,v)|+|e'|> |\pi_{W(e)}(s,q')|$. But this incurs contradiction since $|\pi'(s,q')|\leq |\pi_{W(e)}(s,q')|$.
\end{itemize}
Therefore, $|\pi'(s,v)|=|\pi_{W(e)}(s,v)|$ holds.
As $q'\not\in \overline{qo}$, we define $p$ as a point on
$\overline{oq'}\setminus\{o\}$ infinitely close to $o$ (e.g., see Fig.~\ref{fig:tracepoint30}). Hence, $p\in
\overline{vq'}$. Since $\pi'(s,q')=\pi'(s,v)\cup \overline{vq'}$ is a shortest
path from $s$ to $q'$ in $\mathcal{F}'$, $\pi'(s,v)\cup \overline{vp}$ is
a shortest path from $s$ to $p$ to $\mathcal{F}'$.
Recall that $o$ is on the non-extension bisector $B(\alpha,\alpha')$ of two generators $\alpha$
and $\alpha'$ in $W(e)$ and $v$ is on $\alpha$. Let $v'$ be the anchor
of $o$ in $\alpha'$ (e.g., see Fig.~\ref{fig:tracepoint30}). Since $|\pi'(s,v)|=|\pi_{W(e)}(s,v)|$, we have
$|\pi'(s,v)|+|\overline{vo}|=|\pi_{W(e)}(s,v)|+|\overline{vo}|=|\pi_{W(e)}(s,v')|+|\overline{v'o}|$, where
$\pi_{W(e)}(s,v')$ is the path from $s$ to $v'$ following $W(e)$.
By the definition of $p$ and because $B(\alpha,\alpha')$ is a non-extension bisector, it holds that
$|\pi_{W(e)}(s,v)|+|\overline{vp}|>|\pi_{W(e)}(s,v')|+|\overline{v'p}|$. As $|\pi'(s,v)|=|\pi_{W(e)}(s,v)|$, we have $|\pi'(s,v)|+|\overline{vp}|>|\pi_{W(e)}(s,v')|+|\overline{v'p}|$.
Since $\pi'(s,v)\cup \overline{vp}$ is a shortest path from $s$ to $p$ in
$\mathcal{F}'$,
$\pi_{W(e)}(s,v')\cup \overline{v'p}$ cannot be a path from $s$ to $p$ in $\mathcal{F}'$.
Therefore, $\pi_{W(e)}(s,v')\cup \overline{v'p}$ must intersect the interior of $e'$.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=2.0in]{tracepoint30.pdf}
\caption{\footnotesize The dashed (red) path is $\pi_{W(e)}(s,v')$, which crosses $e'$ at $b$.}
\label{fig:tracepoint30}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
We claim that $\overline{v'p}$ cannot intersect $e'$. Indeed, since $\overline{v'p}$ is covered by the wavefront $W(e)$ when $W(e)$ propagates to $g$ in either $\mathcal{U}(e)$ or $\mathcal{U}(g)$, $\overline{v'p}$ is in $\mathcal{U}(e)\cup \mathcal{U}(g)$. Since $(e',g')\in \Pi_{22}$ and by the definition of $\Pi_{22}$, $e'$ is not in $\mathcal{U}(e)\cup \mathcal{U}(g)$. Therefore, $\overline{v'p}$ cannot intersect $e'$.
The above claim implies that $\pi_{W(e)}(s,v')$ must intersect the interior of $e'$ at a point, say, $b$ (e.g., see Fig.~\ref{fig:tracepoint30}). Let $\pi_{W(e)}(s,b)$ be the sub-path of $\pi_{W(e)}(s,v')$ between $s$ and $b$.
Let $\pi_{W(e)}(b,q')$ be a path from $b$ to $q'$ along $e'$ by considering $e'$ as an opaque edge of open endpoints. As discussed above, $|\pi_{W(e)}(b,q')|\leq |e'|$. Hence, $\pi_{W(e)}(s,b)\cup \pi_{W(e)}(b,q')$ is a path in $\mathcal{F}'$, whose length is at most $|\pi_{W(e)}(s,b)|+|e'|\leq |\pi_{W(e)}(s,v')|+|e'|$.
On the other hand,
$|\pi_{W(e)}(s,v')|+|\overline{v'o}|+|\overline{oq'}|=|\pi'(s,v)|+|\overline{vo}|+|\overline{oq'}|=|\pi'(s,v)|+|\overline{vq'}|$. Since $(e',g')\in \Pi_{22}$ and by the definition of $\Pi_{22}$, $e$ is outside $\mathcal{U}(e')$. Because $q'\in e'$ and $\overline{ov'}$ crosses $e$ at a point $q''$, by the property of well-covering regions of $\mathcal{S}'$, $|\overline{q''o}|+|\overline{oq'}|\geq 2|e'|$.
Therefore, we obtain
\begin{align*}
|\pi'(s,v)|+|\overline{vq'}| & =|\pi_{W(e)}(s,v')|+|\overline{v'o}|+|\overline{oq'}|\\
& \geq |\pi_{W(e)}(s,v')|+|\overline{q''o}|+|\overline{oq'}|\\
& \geq |\pi_{W(e)}(s,v')|+2|e'|\geq |\pi_{W(e)}(s,b)| + 2|e'|\\
& >|\pi_{W(e)}(s,b)| + |e'|\geq |\pi_{W(e)}(s,b)| + |\pi_{W(e)}(b,q')|.
\end{align*}
This means that $\pi_{W(e)}(s,b)\cup \pi_{W(e)}(b,q')$ is a path from $s$ to $q'$ in $\mathcal{F}'$ that is shorter than the path $\pi'(s,v)\cup \overline{vq'}$. But this incurs a contradiction as $\pi'(s,v)\cup \overline{vq'}$ is a shortest path from $s$ to $q'$ in $\mathcal{F}'$.
This completes the proof of the lemma.
\end{proof}
In summary, the total time of the wavefront expansion algorithm is $O(n+h\log n)$, which is $O(n+h\log h)$.
For the space complexity of the algorithm, since each wavefront $W(e)$ is maintained by a persistent binary tree, each bisector event costs additional $O(\log n)$ space. As discussed above, the total sum of $k+h_c$ in the entire algorithm, which is the total number of bisector events, is $O(h)$. Hence, the space cost by persistent binary trees is $O(h\log n)$. The space used by other parts of the algorithm is $O(n)$. Hence, the space complexity of the algorithm is $O(n+h\log n)$, which is $O(n+h\log h)$. This proves Lemma~\ref{lem:algocom}.
\subsection{Proving Lemma~\ref{lem:numgen}}
\label{sec:lemnumgen}
In this section, we prove Lemma~\ref{lem:numgen}. The proof follows the same strategy in the high level as that in the HS algorithm, although many details are different. We start with the following lemma.
\begin{lemma}\label{lem:pairs}
Suppose $\Pi$ is the set of pairs $(e,B)$ of transparent edges $e$ and bisectors $B$ such that $B$ crosses $e$ in the wavefront $W(e)$ of $e$ during our wavefront expansion algorithm but the same crossing does not occur in $\mbox{$S\!P\!M$}'(s)$. Then $|\Pi|=O(h)$.
\end{lemma}
\begin{proof}
Let $\alpha_1=(A_1,a_1)$ and $\alpha_2=(A_2,a_2)$ be the two generators of $B$.
Recall that the well-covering region $\mathcal{U}(e)$ of $e$ is the union of $O(1)$ cells of $\mathcal{S}'$ and each cell has $O(1)$ elementary chain fragments on its boundary. Hence, $\mathcal{U}(e)$ has $O(1)$ convex chains on its boundary.
Recall also that $W(e)$ is computed by merging all contributing wavefronts $W(f,e)$ for $f\in input(e)$ in the wavefront merging procedure, and $W(f,e)$ is computed by propagating $W(f)$ to $e$ through $\mathcal{U}(e)$ in the wavefront propagation procedure.
We first argue that the number of pairs $(e,B)$ of $\Pi$ in the case where at least one of $\alpha_1$ and $\alpha_2$ is in $\mathcal{U}(e)$ is $O(h)$. Indeed, according to our wavefront propagation procedure, if a subcell $c\in \mathcal{U}(e)$ is an empty rectangle, then no new generators will be created when $W(f)$ is propagating through $c$; otherwise, although multiple generators may be created, at most two (one on each side of $c$) are in the wavefront existing $c$. As $\mathcal{U}(e)$ has $O(1)$ cells and each cell may be partitioned into $O(1)$ subcells during the wavefront propagation procedure, only $O(1)$ generators of $W(e)$ are inside $\mathcal{U}(e)$. Since each generator of $W(e)$ may define at most two bisectors in $W(e)$ and the total number of transparent edges of $\mathcal{S}'$ is $O(h)$, the number of pairs $(e,B)$ of $\Pi$ such that at least one generator of $B$ is in $\mathcal{U}(e)$ is at most $O(h)$.
In the following, we assume that both $\alpha_1$ and $\alpha_2$ are outside $\mathcal{U}(e)$, i.e., their initial vertices $a_1$ and $a_2$ are outside $\mathcal{U}(e)$. Let $q$ be the intersection of $B$ and $e$. Let $\pi'(a_1,q)$ denote the path from $q$ to $a_1$ following the tangent from $q$ to $A_1$ and then to $a_1$ along $A_1$; define $\pi'(a_2,q)$ similarly. Clearly, both $\alpha_1$ and $\alpha_2$ claim $q$ in $W(e)$.
Since $\alpha_1$ is outside $\mathcal{U}(e)$, $\alpha_1$ must be in $W(f_1)$ for some transparent edge $f_1$ of $\partial \mathcal{U}(e)$ so that $\alpha_1$ of $W(e)$ is from $W(f_1,e)$.
Since $\alpha_1$ is outside $\mathcal{U}(e)$, $\pi'(a_1,q)$ must intersects $f_1$, say, at point $q_1$ (e.g., see Fig.~\ref{fig:pair10}), and $\pi'(q_1,q)$ is inside $\mathcal{U}(e)$, where $\pi'(q_1,q)$ is the sub-path of $\pi'(a_1,q)$ between $q_1$ and $q$. Also, $f_1$ is processed (for the wavefront propagation procedure) earlier than $e$ because $W(f_1)$ contributes to $W(e)$.
We claim that $q_1$ is claimed by $\alpha_1$ in $\mbox{$S\!P\!M$}'(s)$. Assume to the contrary this is not true. Then,
let $\pi(s,q_1)$ be a shortest path from $s$ to $q_1$ in the free space $\mathcal{F}$.
\begin{itemize}
\item
If $\pi(s,q_1)$ does not intersect $e$, then $\pi(s,q_1)$ is in the modified free space $\mathcal{F}'$ by replacing $e$ with an opaque edge of open endpoints. Since $\alpha_1$ claims $q$ on $e$, $\pi=\pi_{W(e)}(s,a_1)\cup \pi'(a_1,q)$ is a shortest path from $s$ to $q$ in the modified free space $\mathcal{F}'$, where $\pi_{W(e)}(s,a_1)$ is the path from $s$ to $a_1$ following the wavefront $W(e)$. Since $q_1$ is in $\pi$ and $q_1$ is not claimed by $\alpha_1$ in $\mbox{$S\!P\!M$}'(s)$, if we replace the portion between $s$ and $q_1$ in $\pi$ by $\pi(s,q_1)$, we obtain a shorter path from $s$ to $q$ in $\mathcal{F}'$ than $\pi$. But this incurs a contradiction since $\pi$ is a shortest path from $s$ to $q$ in $\mathcal{F}'$.
\item
If $\pi(s,q_1)$ intersects $e$, say, at a point $b$, then since $f_1\in \partial \mathcal{U}(e)$ and thus $d(f_1,e)\geq 2\cdot \max\{|f_1|,|e|\}$ by the properties of the well-covering regions of $\mathcal{S}'$, we show below that $e$ must be processed earlier than $f_1$, which incurs a contradiction because $f_1$ is processed earlier than $e$.
Indeed, $covertime(e)=\tilde{d}(s,e)+|e|\leq d(s,b)+\frac{1}{2}\cdot |e| + |e|=d(s,b)+\frac{3}{2}\cdot |e|$. On the other hand, $covertime(f_1)=\tilde{d}(s,f_1)+|f_1|\geq d(s,q_1)-\frac{1}{2}\cdot |f_1| + |f_1|=d(s,b)+d(b,q_1)+\frac{1}{2}\cdot |f_1|$. Since $f_1\in \partial \mathcal{U}(e)$, $b\in e$, and $q_1\in f_1$, $d(b,q_1)\geq d(f_1,e)\geq 2\cdot |e|$. Hence, we obtain $covertime(f_1)>covertime(e)$ and thus $e$ must be processed earlier than $f_1$.
\end{itemize}
Therefore, $q_1$ must be claimed by $\alpha_1$ in $\mbox{$S\!P\!M$}(s)$.
We define $f_2$, $q_2$, and $\pi'(q,q_2)$ analogously with respect to $\alpha_2$. Similarly, we can show that $q_2$ is claimed by $\alpha_2$ in $\mbox{$S\!P\!M$}'(s)$.
\begin{figure}[t]
\begin{minipage}[t]{0.49\textwidth}
\begin{center}
\includegraphics[height=2.0in]{pair10.pdf}
\caption{\footnotesize Illustrating $q$ and $q_1$.}
\label{fig:pair10}
\end{center}
\end{minipage}
\hspace{0.02in}
\begin{minipage}[t]{0.49\textwidth}
\begin{center}
\includegraphics[height=2.2in]{pair20.pdf}
\caption{\footnotesize Illustrating the psuedo-triangle $\triangle(q,q_1,q_2)$ (the shaded region).}
\label{fig:pair20}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
For each $f\in \partial \mathcal{U}(e)$, the generators of $W(f)$ that are also in $W(e)$ may not form a single subsequence of the generator list of $W(e)$, but they must form a constant number of (maximal) subsequences. Indeed, since $\mathcal{U}(e)$ is the union of $O(1)$ cells of $\mathcal{S}'$, the number of islands in $\mathcal{U}(e)$ is $O(1)$. Thus, the number of topologically different paths from $f$ to $e$ in $\mathcal{U}(e)$ is $O(1)$; each such path will introduce at most one subsequence of generators of $W(e)$ that are also from $W(f)$. Therefore, the generators of $W(f)$ that are also in $W(e)$ form $O(1)$ subsequences of the generator list of $W(e)$.
Since $\partial\mathcal{U}(e)$ has $O(1)$ transparent edges, the generator list of $W(e)$ can be partitioned into $O(1)$ subsequences each of which is from $W(f)$ following topologically the same path for a single transparent edge $f$ of $\partial\mathcal{U}(e)$. Due to this property, we have the following observation: the number of pairs of adjacent generators of $W(e)$ that are from different subsequences is $O(1)$.
Due to the above observation, the number of pairs of edges $f_1$ and $f_2$ on $\partial \mathcal{U}(e)$ in the case where $f_1\neq f_2$, or $f_1=f_2$ but $\pi'(q,q_1)$ and $\pi'(q,q_2)$ are topologically different in $\mathcal{U}(e)$ is only $O(1)$. Therefore, the total number of pairs $(e,B)$ of $\Pi$ in that case is $O(h)$. In the following, it suffices to consider the case where $f_1= f_2$, and $\pi'(q,q_1)$ and $\pi'(q,q_2)$ are topologically the same in $\mathcal{U}(e)$; for reference purpose, we use $\Pi'$ to denote the subset of pairs $(e,B)$ of $\Pi$ with the above property. Below we prove $|\Pi'|=O(h)$.
Let $f=f_1=f_2$. Since $\pi'(q,q_1)$ and $\pi'(q,q_2)$ are topologically the same in $\mathcal{U}(e)$,
the region bounded by $\pi'(q,q_1)$, $\pi'(q,q_2)$, and $\overline{q_1q_2}$ must be in $\mathcal{U}(e)$; we call the above region a {\em pseudo-triangle}, for both $\pi'(q,q_1)$ and $\pi'(q,q_2)$ are convex chains, and we use $\triangle(q,q_1,q_2)$ to denote it (e.g., see Fig.~\ref{fig:pair20}).
Because $(e,B)$ is not an incident pair in $\mbox{$S\!P\!M$}'(s)$, the point $q$ must be claimed by a different generator $\alpha$ in $\mbox{$S\!P\!M$}'(s)$, which must be from the side of $e$ different than $\alpha_1$ and $\alpha_2$. We have proved above that $q_1$ is claimed by $\alpha_1$ and $q_2$ is claimed by $\alpha_2$ in $\mbox{$S\!P\!M$}'(s)$. Hence, there must be at least one bisector event in $\mbox{$S\!P\!M$}'(s)$ that lies in the interior of the pseudo-triangle $\triangle(q,q_1,q_2)$\footnote{This is also due to that for any point $p\in \overline{q_1q_2}$, the shortest path $\pi(s,p)$ cannot intersect $e$; this can be proved by a similar argument to the above for proving that $q_1$ is claimed by $\alpha_1$ in $\mbox{$S\!P\!M$}'(s)$. This observation implies that $\alpha$ cannot claim any points on $f$ in $\mbox{$S\!P\!M$}'(s)$.}. We charge the early demise of $B$ to any one of these bisector events in $\triangle(p,q_1,q_2)$.
Note that the path $\pi'(q,q_1)$ (resp., $\pi'(q,q_2)$) is in a shortest path from $s$ to $q$ following the wavefront $W(e)$ in the modified free space by replacing $e$ with an opaque edge of open endpoints. Hence, if $\Pi'$ has other pairs $(e,B')$ whose first element is $e$, then the corresponding paths $\pi'(q,q_1)$ and $\pi'(q,q_2)$ of all these pairs are disjoint, and thus the corresponding pseudo-triangles $\triangle(q,q_1,q_2)$ are also disjoint in $\mathcal{U}(e)$. Hence, each bisector event of $\mbox{$S\!P\!M$}'(s)$ inside $\triangle(q,q_1,q_2)$ is charged at most once for all pairs of $\Pi'$ that have $e$ as the first element. Since each cell of $\mathcal{S}'$ belongs to the well-covering region $\mathcal{U}(e)$ of $O(1)$ transparent edges $e$, each bisector event of $\mbox{$S\!P\!M$}'(s)$ is charged $O(1)$ times for all pairs of $\Pi'$. Because $\mbox{$S\!P\!M$}'(s)$ has $O(h)$ bisector events by Corollary~\ref{coro:complexity}, the number of pairs of $\Pi'$ is at most $O(h)$.
This completes the proof of the lemma.
\end{proof}
Armed with Lemma~\ref{lem:pairs}, we prove the subsequent five lemmas, which together lead to Lemma~\ref{lem:numgen}.
\begin{lemma}\label{lem:rule3b}
The total number of marked generators by Rule~2(a) and Rule~3 is $O(h)$.
\end{lemma}
\begin{proof}
Suppose $\alpha$ is a generator marked by Rule~2(a) for a transparent edge $e$. Then, $\alpha$ must be the first or last non-artificial generator of $W(e)$. Hence, at most two generators of $W(e)$ can be marked by Rule~2(a). As $\mathcal{S}'$ has $O(h)$ transparent edges, the total number of generators marked by Rule~2(a) is $O(h)$.
Suppose $\alpha$ is a generator marked by Rule~3(a) for the two transparent edges $(e,g)$ with $g\in output(e)$. Since $\alpha$ claims an endpoint of $g$, $\alpha$ must be the first or last generator of $W(e,g)$. Hence, at most two generators of $W(e,g)$ can be marked by Rule~3(a). Since $|output(e)|=O(1)$, at most $O(1)$ generators can be marked for the pairs of transparent edges with $e$ as the first edge. Since $\mathcal{S}'$ has $O(h)$ transparent edges, the total number of generators marked by Rule~3(a) is $O(h)$.
Suppose $\alpha$ is a generator marked by Rule~3(b) for the two transparent edges $(e,g)$ with $g\in output(e)$. Note that $\alpha$ is a generator in $W(e)$. We assume that $\alpha$ is not the first or last non-artificial generators of $W(e)$; the first and last non-artificial generators of $W(e)$, countered separately, sum to $O(h)$ for all transparent edges $e$ of $\mathcal{S}'$.
Let $\alpha_1$ and $\alpha_2$ be the two neighboring generators of $\alpha$ in $W(e)$. Thus, both $\alpha_1$ and $\alpha_2$ are non-artificial.
We assume that the bisectors $B(\alpha_1,\alpha)$ and $B(\alpha,\alpha_2)$ intersect $e$ in $\mbox{$S\!P\!M$}'(s)$; by Lemma~\ref{lem:pairs}, there are only $O(h)$ bisector and transparent edge intersections that appear in some approximate wavefront but not in $\mbox{$S\!P\!M$}'(s)$.
Since $\alpha$ is marked by Rule~3(b), at least one of $B(\alpha_1,\alpha)$ and $B(\alpha,\alpha_2)$ fails to reach the boundary of $D(e)$ during the algorithm, where $D(e)$ is the union of cells through which $W(e)$ is propagated to all edges $g'\in output(e)$. Note that $D(e)\subseteq \mathcal{U}(e)\cup \bigcup_{g'\in output(e)}\mathcal{U}(g')$, which contains $O(1)$ cells of $\mathcal{S}'$ as $|output(e)|=O(1)$ and the well-covering region of each transparent edge is the union of $O(1)$ cells of $\mathcal{S}'$; also, each cell of $D(e)$ is within a constant number of cells of $e$. Without loss of generality, we assume that $B(\alpha_1,\alpha)$ does not reach the boundary of $D(e)$ and a bisector event on $B(\alpha_1,\alpha)$ is detected during the algorithm. The detected bisector event on $B(\alpha_1,\alpha)$ also implies that an actual bisector event in $\mbox{$S\!P\!M$}'(s)$ happens to $B(\alpha_1,\alpha)$ no later than the detected event; we charge the marking of $\alpha$ to that bisector actual endpoint (i.e., a vertex) in $\mbox{$S\!P\!M$}'(s)$, and the endpoint is in $D(e)$ because $B(\alpha_1,\alpha)$ intersects $e$ in $\mbox{$S\!P\!M$}'(s)$.
Since each cell of $D(e)$ is within a constant number of cells of $e$, each cell of $\mathcal{S}'$ belongs to $D(e')$ for a constant number of transparent edges $e'$ of $\mathcal{S}'$. Therefore, each vertex of $\mbox{$S\!P\!M$}'(s)$ is charged $O(1)$ times. Since $\mbox{$S\!P\!M$}'(s)$ has $O(h)$ vertices by Lemma~\ref{lem:sizespm}, the total number of generators marked by Rule~3(b) is $O(h)$.
\end{proof}
\begin{lemma}\label{lem:rule1}
The total number of marked generators by Rule 1 is $O(h)$.
\end{lemma}
\begin{proof}
According to our algorithm, generators are only created during the wavefront propagation procedure, which is to propagate $W(e)$ to compute $W(e,g)$ for all edges $g\in output(e)$ through $\mathcal{U}$, where $\mathcal{U}$ is $\mathcal{U}(e)$ or $\mathcal{U}(g)$. The procedure has two cases depending on whether a cell $c$ of $\mathcal{U}$ is an empty rectangle. If yes, then no generators will be created in $c$. Otherwise, $c$ may be partitioned into $O(1)$ subcells, each of which may have a convex chain on its left side and/or its right side. Let $c$ be such a subcell. Without loss of generality, we assume that the current wavefront $W$ is propagating in $c$ from bottom to top. We consider the generators created on the left side $\zeta_l$ of $c$.
According to our algorithm, a generator on $\zeta_l$ is created by the leftmost generator of the current wavefront $W$. As such, if the leftmost wavelet of $W$ does not create a generator, then no generator on $\zeta_l$ will be created. Otherwise, let $\alpha$ be a generator created on $\zeta_l$, which becomes the leftmost generator of $W$ at the point of creation. Let $\alpha'$ be the right neighbor of $\alpha$ in $W$. According to our algorithm, if $\alpha$ does not involve in any bisector event in $c$, i.e., the bisector $B(\alpha,\alpha')$ does not intersect any other bisector in $c$ during the propagation of $W$ in $c$, then no more new generators will be created on $\zeta_l$. Otherwise, we charge the creation of $\alpha$ to the bisector event involving $B(\alpha,\alpha')$; each such event can be charged at most twice (one for a generator on $\zeta_l$ and the other for a generator created on the right side of $c$). Recall that for each bisector event a generator is marked by Rule~3(b).
In light of the above discussion, since each cell of $\mathcal{S}'$ belongs to the well-covering region $\mathcal{U}(e')$ of $O(1)$ transparent edges $e'$, the total number of generators marked by Rule~1 is $O(h)+O(k)$, where $k$ is the total number of generators marked by Rule~3(b), which is $O(h)$ by Lemma~\ref{lem:rule3b}. The lemma thus follows.
\end{proof}
\begin{lemma}\label{lem:rule4}
The total number of marked generators by Rule 4 is $O(h)$.
\end{lemma}
\begin{proof}
Let $\alpha$ be the generator and $e$ be the transparent edge specified in the rule statement. According to Rule~4, $\alpha$ is marked because it claims part of an obstacle edge during the wavefront propagation procedure for propagating $W(e)$ to compute $W(e,g)$ for edges $g\in output(e)$.
As in the proof of Lemma~\ref{lem:rule3b}, define $D(e)$ as the union of cells through which $W(e)$ is propagated to all edges $g\in output(e)$. Recall that $D(e)$ contains $O(1)$ cells of $\mathcal{S}'$, and each cell of $D(e)$ is within a constant number of cells of $e$, which implies that each cell of $\mathcal{S}'$ belongs to $D(e')$ for a constant number of transparent edges $e'$ of $\mathcal{S}'$.
First of all, for the case where $\alpha$ is a generator created during the wavefront propagation procedure, the total number of such marked generators is no more than the total number of marked generators by Rule~1, which is $O(h)$ by Lemma~\ref{lem:rule1}.
Second, for the case where $\alpha$ is the first or last generator in $W(e)$, the total number of such generators is clearly $O(h)$ since $\mathcal{S}'$ has $O(h)$ transparent edges $e$.
Third, for the case where $\alpha$ claims a rectilinear extreme vertex, the total number of such marked generators is $O(h)$. To see this, since $D(e)$ has $O(1)$ cells of $\mathcal{S}'$ and each cell contains at most one rectilinear extreme vertex, $D(e)$ contains $O(1)$ rectilinear extreme vertices. Therefore, at most $O(1)$ generators will be marked in $D(e)$. Since each cell of $\mathcal{S}'$ belongs to $D(e')$ for $O(1)$ transparent edges $e'$ of $\mathcal{S}'$ and $\mathcal{S}'$ contains $O(h)$ transparent edges $e$, the total number of marked generators in this case is $O(h)$.
Finally, any Rule~4 marked generator $\alpha$ that does not belong to any of the above cases has the following property: $\alpha$ does not claim a rectilinear extreme vertex, and is not the first or last non-artificial generator in $W(e)$, and is not a generator created in $D(e)$. Let $\alpha'$ be a neighbor of $\alpha$ in $W(e)$. Due to the above property, $\alpha'$ is a non-artificial generator. We can assume that $B(\alpha',\alpha)$ intersects $e$ in $\mbox{$S\!P\!M$}'(s)$; by Lemma~\ref{lem:pairs}, there are only $O(h)$ bisector and transparent edge intersections that appear in some approximate wavefront but not in $\mbox{$S\!P\!M$}'(s)$. Hence, in $\mbox{$S\!P\!M$}'(s)$, $B(\alpha',\alpha)$ terminates in $D(e)$, either on an obstacle edge or in a bisector event before reaching an obstacle edge. In either case, we charge the mark of $\alpha$ at $e$ to this endpoint of $B(\alpha',\alpha)$ in $\mbox{$S\!P\!M$}'(s)$, which is vertex of $\mbox{$S\!P\!M$}'(s)$. Because each cell of $\mathcal{S}'$ belongs to $D(e')$ for a constant number of transparent edges $e'$ of $\mathcal{S}'$, each vertex of $\mbox{$S\!P\!M$}'(s)$ is charged at most $O(1)$ times. As $\mbox{$S\!P\!M$}'(s)$ has $O(h)$ vertices by Lemma~\ref{lem:sizespm}, the total number of marked generators by Rule~4 is $O(h)$.
\end{proof}
\begin{lemma}
The total number of marked generators by Rule 2(b) is $O(h)$.
\end{lemma}
\begin{proof}
Let $\alpha$ be the generator and $e$ be the transparent edge specified in the rule statement.
First of all, the total number of marked generators in the case where $\alpha$ is in $\mathcal{U}(e)$ is $O(h)$, because the number of Rule~1 marked generators is $O(h)$ by Lemma~\ref{lem:rule1} and $\mathcal{U}(e)$ has $O(1)$ cells. In the following, we consider the other case where $\alpha$ is outside $\mathcal{U}(e)$.
Since $\alpha$ is outside $\mathcal{U}(e)$, there is a transparent edge $f$ on $\partial \mathcal{U}(e)$ (i.e., $f\in input(e)$), such that $\alpha$ is in $W(f)$ and it is in $W(e)$ because $W(f)$ is propagated to $e$ through $\mathcal{U}(e)$. Since $\alpha$'s claim is shortened or eliminated by an artificial wavefront, there must be a bisector event involving $\alpha$ during the computation of $W(f,e)$ from $W(f)$. Hence, $\alpha$ is also marked by Rule~3(b). We charge $\alpha$ to this Rule~3(b) mark for $f$. Since there are $O(h)$ generators marked by Rule~3(b) by Lemma~\ref{lem:rule3b}, the total number of marked generators of Rule~2(b) is $O(h)$.
\end{proof}
\subsection{Computing the shortest path map $\mbox{$S\!P\!M$}(s)$}
\label{sec:spm}
In this section, we compute the shortest path map $\mbox{$S\!P\!M$}(s)$. To this end, we need to mark generators following our rules during our wavefront merging and propagation procedures. As the total number of all marked generators is $O(h)$, marking these generators do not change the running time of our algorithm asymptotically. In the following, we show that $\mbox{$S\!P\!M$}(s)$ can be constructed in $O(n+h\log h)$ time with the help of the marked generators. In light of Lemma~\ref{lem:spmconstruction}, we will focus on constructing $\mbox{$S\!P\!M$}'(s)$.
We first show in Section~\ref{sec:markenough} that the marked generators are sufficient in the sense that if a generator participates in a true bisector event of $\mbox{$S\!P\!M$}'(s)$ in a cell $c$ of $\mathcal{S}'$, then $\alpha$ must be marked in $c$. Then, we present the algorithm for constructing $\mbox{$S\!P\!M$}'(s)$, which consists of two main steps.
The first main step is to identify all vertices of $\mbox{$S\!P\!M$}'(s)$ and the second one is to compute the edges of $\mbox{$S\!P\!M$}'(s)$ and assemble them to obtain $\mbox{$S\!P\!M$}'(s)$. The first main step is described in Section~\ref{sec:first} while the second one is discussed in Section~\ref{sec:second}.
\subsubsection{Correctness of the marked generators}
\label{sec:markenough}
In this section we show that if a generator participates in a bisector event of $\mbox{$S\!P\!M$}'(s)$ in a cell $c$ of $\mathcal{S}'$, then $\alpha$ must be marked in $c$. Our algorithm in the next section for computing $\mbox{$S\!P\!M$}'(s)$ relies on this property. Our proof strategy again follows the high-level structure of the HS algorithm. We first prove the following lemma.
\begin{lemma}\label{lem:markcell}
Let $\alpha$ be a generator in an approximate wavefront $W(e)$ for some transparent edge $e$. Suppose there is a point $p\in e$ that is claimed by $\alpha$ in $W(e)$ but not in $\mbox{$S\!P\!M$}'(s)$ (because the approximate wavefront from the other side of $e$ reaches $p$ first). Then, $\alpha$ is marked in the cell $c$ on $\alpha$'s side of $e$.
\end{lemma}
\begin{proof}
Assume to the contrary that $\alpha$ is not marked in $c$. Then, $\alpha$ must have two neighboring non-artificial generators $\alpha_1$ and $\alpha_2$ in $W(e)$ since otherwise Rule~2 would apply.
Also, because all transparent edges of $\partial \mathcal{U}(e)$ are in $output(e)$, the two bisectors $B(\alpha_1,\alpha)$ and $B(\alpha,\alpha_2)$ must exist the well-covering region $\mathcal{U}(e)$ of $e$ through the same transparent edge $g\in \partial \mathcal{U}(e)$, since otherwise Rule~3 or 4 would apply.
Further, the region $R$ bounded by $B(\alpha_1,\alpha)$, $B(\alpha,\alpha_2)$, $e$, and $g$ must be a subset of $\mathcal{U}(e)$. Indeed, if $R$ contains an island not in $\mathcal{U}(e)$, then $\alpha$ would claim an endpoint of a boundary edge of the island, in which case Rule~3 would apply.
Let $\alpha'=(A,a)$ be the true predecessor of $p$ in $\mbox{$S\!P\!M$}'(s)$.
Without loss of generality, we assume that $e$ is horizontal and
$\alpha$ is below $e$ and thus $\alpha'$ is above $e$.
Let $\pi'(a,p)$ be the path from
$p$ along its tangent to $A$ and then following $A$ to $a$.
We first consider the case where $\alpha'$ is not in the interior of
the well-covering region $\mathcal{U}(e)$, i.e., the initial vertex $a$ is
not in the interior of $\mathcal{U}(e)$. Since $R$ is a subset of of
$\mathcal{U}(e)$ without non-$\mathcal{U}(e)$ islands, $a$ is not in the
interior of $R$ and thus $\pi'(a,p)$ intersects $\partial R$, say,
at a point $q$. In the following, we argue that $\alpha$ has been
involved in a bisector event detected by our algorithm, and thus
marked in $c$. Let $g'$ be the portion of $g$ between its intersections with
$B(\alpha_1,\alpha)$ and $B(\alpha,\alpha_2)$. Depending on whether
$q\in g'$, there are two subcases.
\begin{itemize}
\item
If $q\in g'$ (e.g., see Fig.~\ref{fig:markcell}), then $\tilde{d}(s,g)\leq d(s,a)+|\pi'(a,q)|+|g|/2$, where
$\pi'(a,q)$ is the sub-path of $\pi'(a,p)$ between $a$ and $q$. Hence,
the time $\tau_g$ when artificial wavefronts originating from the
endpoints of $g$ cover $g$ is no later than $\tilde{d}(s,g)+|g|\leq
d(s,a)+|\pi'(a,q)|+3|g|/2$. Because $g\in \partial \mathcal{U}(e)$,
$|\overline{pq}|\geq 2\cdot \max\{|g|,|e|\}$. Hence, $\tau_g\leq
d(s,a)+|\pi'(a,q)|+3|g|/2<
d(s,a)+|\pi'(a,q)|+|\overline{qp}|=d(s,a)+|\pi'(a,p)|<d_{W(e)}(s,p)$, where
$d_{W(e)}(s,p)$ is the length of the path from $s$ to $p$ following the
wavefront $W(e)$; the last inequality holds because $\alpha'$ is the
true predecessor of $p$ while $\alpha$ is not.
\begin{figure}[t]
\begin{minipage}[t]{0.49\textwidth}
\begin{center}
\includegraphics[height=2.1in]{markcell.pdf}
\caption{\footnotesize Illustrating the case where $q\in g'$.}
\label{fig:markcell}
\end{center}
\end{minipage}
\hspace{0.02in}
\begin{minipage}[t]{0.49\textwidth}
\begin{center}
\includegraphics[height=2.1in]{markcell10.pdf}
\caption{\footnotesize Illustrating the case where $q\in B(\alpha_1,\alpha)$.}
\label{fig:markcell10}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
On the other hand, the time $\tau_e$ when the wavefront $W(e)$
reaches an endpoint of $e$ cannot be earlier than $d_{W(e)}(s,p)-|e|$. Hence,
the wavelet from $\alpha$ cannot reach $g$ earlier than
$d_{W(e)}(s,p)-|e|+d(e,g)\geq d_{W(e)}(s,p)-|e|+2|e|\geq d_{W(e)}(s,p)+|e|$,
which is larger than $\tau_g$ since $\tau_g<d_{W(e)}(s,p)$ as proved
above. Therefore, by the time the wavelet from
$\alpha$ reaches $g$, the artificial wavelets of $g$ have already
covered $g$, eliminating the wavelet from $\alpha$ from reaching $g$.
Thus, $\alpha$ must be marked by Rule~3(b).
\item
If $q\not\in g'$, then $q$ is on one of the bisectors
$B(\alpha_1,\alpha)$ and $B(\alpha,\alpha_2)$. Let $\pi(s,a)$ be a shortest path from $s$ to $a$ and let $\pi(s,p)=\pi(s,a)\cup \pi'(a,p)$, which is a shortest path from $s$ to $p$.
If $\pi(s,p)$ intersects $g$, then we can use the same analysis as above to show that the wavelet from $\alpha$ will be eliminated from reaching $g$ by the artificial wavelets of $g$ and thus $\alpha$ must be marked by Rule~3(b).
In the following, we assume that $\pi(s,p)$ does not intersect $g$.
Without of loss of generality, we assume that $q\in B(\alpha_1,\alpha)$ (e.g., see Fig.~\ref{fig:markcell10}). Since
$\pi'(a,p)$ is a subpath of the shortest path $\pi(s,p)$, every point of
$\pi'(a,p)$ has $\alpha'$ as its predecessor. As $q\in
\pi'(a,p)$, the predecessor of $q$ is $\alpha'$.
Let $\pi(s,q)$ be the subpath of $\pi(s,p)$ between $s$ and $q$, which is a shortest path. Since $\pi(s,p)$ does not intersect $g$, $\pi(s,q)$ does not intersect $g$. Hence, $\pi(s,q)$ is a shortest path in the modified free space $\mathcal{F}'$ by replacing $g$ with an opaque edge of open endpoints. Therefore, during the wavefront propagation procedure for computing $W(e,g)$ or during the wavefront merging procedure for computing $W(g)$, the point $q$, which is on the bisector $B(\alpha_1,\alpha)$, must be claimed by $\alpha'$. Hence, a bisector event must be detected for
$B(\alpha_1,\alpha)$ during the computation of $W(e,g)$ or $W(g)$. In
either case, $\alpha$ must be marked by Rule~3(b).
\end{itemize}
We next consider the case where $\alpha'$ lies inside $\mathcal{U}(e)$, i.e., its initial vertex $a$ is inside $\mathcal{U}(e)$.
If $a$ is not between the two bisectors $B(\alpha_1,\alpha)$ and $B(\alpha,\alpha_2)$, then $\pi'(a,p)$ must intersect one of the bisectors and thus we can use a similar argument as above second case to show that $\alpha$ must be marked. Otherwise, $a$ is in the region $R$.
This implies that not all points in $R$ are claimed by $\alpha$ when $W(e)$ is propagating to $g$, and therefore, a bisector event involving $\alpha$ must happen during the wavefront propagation procedure to propagate $W(e)$ to compute $W(e,g)$ and thus $\alpha$ is marked by Rule~3(b).
\end{proof}
With the help of the preceding lemma, we prove the following lemma.
\begin{lemma}\label{lem:markcorrect}
If a generator $\alpha$ participates in a bisector event of $\mbox{$S\!P\!M$}'(s)$ in a cell $c$ of $\mathcal{S}'$, then $\alpha$ must be marked in $c$.
\end{lemma}
\begin{proof}
If a bisector has an endpoint on an obstacle edge of $c$, it either
emanates from an obstacle vertex $a$ on the edge (i.e., the bisector
is an extension bisector) or defined by two
generators that claim part of the opaque edge. In the first case, $a$
is the initial vertex of a new created generator in $c$ and thus the generator
is marked by Rule~1. In the second case, both generators are marked by
Rule~4.
Let $\alpha$ be a generator that participates in a bisector event of
$\mbox{$S\!P\!M$}'(s)$ in a cell $c$ of $\mathcal{S}'$. Assume to the contrary that $\alpha$ is not marked for
$c$. Then by Rule~2(a) there must be transparent edges $e$ and $f$ on
the boundary of $c$ such that both $W(e)$ and $W(f)$ contain the three
consecutive generators $\alpha_1$, $\alpha$, and $\alpha_2$. Without
loss of generality, we assume that $W(e)$ enters $c$ and $W(f)$ leaves
$c$. Let $R$ be the region of $c$ bounded by $e$, $f$, and the two bisectors
$B(\alpha_1,\alpha)$ and $B(\alpha,\alpha_2)$; e.g., see Fig.~\ref{fig:markcorrect}. The region $R$ must be a subset of $c$. Indeed, if $R$ contains an island not in $c$, then $\alpha$ would claim an endpoint of a boundary edge of the island, in which case Rule~3 would apply.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=2.2in]{markcorrect.pdf}
\caption{\footnotesize Illustrating the proof of Lemma~\ref{lem:markcorrect}.}
\label{fig:markcorrect}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
Since $\alpha$ participates in a bisector event of $\mbox{$S\!P\!M$}'(s)$ in $c$, at least one
point $p$ in $R$ is not claimed by $\alpha$ in $\mbox{$S\!P\!M$}'(s)$. Let
$\alpha'=(A,a)$ be the true predecessor of $p$. Note that the initial vertex $a$ must be outside $R$ since otherwise a bisector event involving $\alpha$ must happen when $W(e)$ is propagating through $c$ and thus $\alpha$ would be marked by Rule~3(b). Let $\pi'(a,p)$ be the path from
$p$ along its tangent to $A$ and then following $A$ to $a$.
Since $a$ is outside $R$ and $p$ is inside $R$, $\pi'(a,p)$ must intersect the boundary of $R$.
Because $\alpha_1$, $\alpha$, and $\alpha_2$ are three consecutive
generators of $W(e)$, no generator other than $\alpha$ on the same side of $e$ as $\alpha$
claims any point of $R$. Thus, $\pi'(a,p)$ does not cross $e$. Let
$b_1$ and $b_2$ be the intersections of $f$ with $B(\alpha_1,\alpha)$ and $B(\alpha,\alpha_2)$, respectively. Then, if $\alpha$ claims both $b_1$ and $b_2$ in $\mbox{$S\!P\!M$}'(s)$, then $\pi'(a,p)$ cannot cross either bisector on $\partial R$, and thus it must cross $\overline{b_1b_2}$, say, at a point $q$ (e.g., see Fig.~\ref{fig:markcorrect}). Note that $q$ satisfies the
hypothesis of Lemma~\ref{lem:markcell} and thus $\alpha$ is marked for
$c$. It $\alpha$ does not claim either $b_1$ or $b_2$, then that point
satisfies the hypothesis of Lemma~\ref{lem:markcell} and thus $\alpha$
is marked for $c$.
\end{proof}
\subsubsection{Computing the vertices of $\mbox{$S\!P\!M$}'(s)$}
\label{sec:first}
In this section, we compute the vertices of $\mbox{$S\!P\!M$}'(s)$. The next section will compute the edges of $\mbox{$S\!P\!M$}'(s)$ and assemble them to obtain $\mbox{$S\!P\!M$}'(s)$.
\paragraph{Computing active regions.}
Because the approximate wavefronts are represented by persistent trees, after the above wavefront expansion algorithm finishes, the approximate wavefronts $W(e)$ for all transparent edges $e$ of $\mathcal{S}'$ are still available. Also, for each cell $c$ and each transparent edge $e$ of $c$, a set of marked generators in $W(e)$ are known. Using these marked generators, we first break $c$ into {\em active} and {\em inactive} regions such that no vertices of $\mbox{$S\!P\!M$}'(s)$ lie in the inactive regions. Note that each unmarked generator does not participate in a bisector event in $\mbox{$S\!P\!M$}'(s)$ by Lemma~\ref{lem:markcorrect}. If $\alpha_1$ and $\alpha_2$ are neighboring generators on $\partial c$ such that one of them is marked while the other is unmarked, their bisector belongs to $\mbox{$S\!P\!M$}'(s)$. Therefore, all such generators are disjoint and they together partition $c$ into regions such that each region is claimed only by marked generators or only by unmarked generators; the former regions are active while the latter are inactive. We can compute these active regions as follows.
Since the wavefronts $W(e)$ for all transparent edges $e$ of $c$ are available, the list of all generators ordered along the boundary of $c$ is also available. For each pair of adjacent generators $\alpha_1$ and $\alpha_2$, if one of them is marked and the other is unmarked, then we explicitly compute the portion of their bisector $B(\alpha_1,\alpha_2)$ in $c$. We describe the details of this step below.
First of all, observe that $\alpha_1$ and $\alpha_2$ must be from $W(e)$ for the same transparent edge $e$ of $c$, since otherwise each generator must claim an endpoint of their own transparent edge thus must have been marked by Rule~2(a).
Without loss of generality, we assume that $e$ is horizontal and both $\alpha_1$ and $\alpha_2$ are below $e$. Note that we cannot afford computing the entire bisector $B(\alpha_1,\alpha_2)$ and then determining its portion in $c$ because the running time would be proportional to the size of $B(\alpha_1,\alpha_2)$, i.e., the number of hyperbolic-arcs of $B(\alpha_1,\alpha_2)$. Instead, we first compute the intersection $b$ of $B(\alpha_1,\alpha_2)$ and $e$, which can be done in $O(\log n)$ time by the bisection-line intersection operation in Lemma~\ref{lem:bl-intersection}. Note that $b$ is one endpoint of the portion of $B(\alpha_1,\alpha_2)$ in $c$. To determine the other endpoint $b'$, we do the following. Our goal is to compute $b'$ in $O(|B_c|+\log n)$ time, with $B_c=B(\alpha_1,\alpha_2)\cap c$. To this end, we could trace $B_c$ in $c$ from $b$, and for each hyperbolic-arc of $B_c$, we determine whether it intersects $\partial c$. Recall that $\partial c$ consists of $O(1)$ transparent edges and at most $O(1)$ convex chains. To achieve the desired runtime, we need to determine whether each hyperbolic-arc of $B_c$ intersects $\partial c$ in $O(1)$ time. However, it is not clear to us whether this is possible as the size of each convex chain may not be of $O(1)$ size. To circumvent the issue, we use the following strategy. Before tracing $B_c$, we first compute the intersection between $B(\alpha_1,\alpha_2)$ and each convex chain on $\partial c$, which can be done in $O(\log n)$ time by the bisector-chain intersection operation in Lemma~\ref{lem:bc-intersection}. Among all intersections, let $b''$ be the closest one to $\alpha$. Then, we start to trace $B_c$ from $b$, for each hyperbolic-arc $e'$, we determine whether $e'$ intersects each of the transparent edges of $\partial c$. If not, we further check whether $e'$ contains $b''$. If $b''\not\in e'$, then $e'$ is in $c$ and we continue the tracing; otherwise, $b'$ is $b''$ and we stop the algorithm. If $e'$ intersects at least one of the transparent edges of $\partial c$, then among all such intersections as well as $b''$, $b'$ is the one closest to $\alpha$, which can be determined in $O(\log n)$ time by computing their tangents to $\alpha$.
In this way, computing the portion of $B(\alpha_1,\alpha_2)$ in $c$ can be done in $O(\log n+n_c(\alpha_1,\alpha_2))$ time, where $n_c(\alpha_1,\alpha_2)$ is the number of hyperbolic-arcs of $B(\alpha_1,\alpha_2)$ in $c$.
In this way, all active regions of $c$ can be computed in $O(h_c\log
n+n_c)$ time, where $h_c$ is the total number of marked generators in the
wavefronts $W(e)$ of all transparent edges $e$ of $c$ and $n_c$ is the
total number of hyperbolic-arcs of the bisector boundaries of these
active regions.
\paragraph{Computing the vertices of $\mbox{$S\!P\!M$}'(s)$ in each active region.}
In what follows, we compute the vertices of $\mbox{$S\!P\!M$}'(s)$ in each active region $R$ of $c$.
The boundary $\partial R$ consists of $O(1)$ pieces, each of which is
a transparent edge fragment, an elementary chain fragment, or a bisector in
$\mbox{$S\!P\!M$}'(s)$. Unlike the HS algorithm, where each of these pieces is of
$O(1)$ size, in our problem both an elementary chain fragment and a bisector
may not be of constant size. Let $e$ be a transparent edge of $R$.
Without loss of generality, we assume that $e$ is horizontal and $R$
is locally above $e$. Without considering other wavefronts, we use
$W(e)$ to partition $R$ into subregions, called {\em Voronoi faces}, such
that each face has a unique predecessor in $W(e)$. We use $Vor(e)$
to denote the partition of $R$. Note that if $\alpha$ is the
predecessor of a Voronoi face, then for each point $p$ in the face, its
tangent to $\alpha$ must cross $e$ and we assign $p$ a {\em weight}
that is equal to $d(\alpha,p)$.
Also, it is possible that these faces together may not cover the entire $R$; for those points
outside these faces, their weights are $\infty$.
We now discuss how to compute the partition $Vor(e)$. To this
end, we can use our wavefront propagation procedure to propagate
$W(e)$ inside $R$. To apply the algorithm, one issue
is that the boundary of $R$ may contain bisectors, which consists of hyperbolic-arcs instead of polygonal segments. To circumvent the issue, an observation is that the bisectors of $W(e)$ do not intersect
the bisectors on $\partial R$. Indeed, for each bisector $B$ of $\partial R$, one of its defining generators is unmarked and thus does not participate in any bisector event in $c$, and therefore, $B$ does not intersect any bisector in $c$. In light of the observation, we can simply apply the
wavefront propagation procedure to propagate $W(e)$ to $c$ instead of
$R$ to partition $c$ into Voronoi faces, denoted by $Vor_c(e)$. The above
observation implies that each bisector on the boundary of $R$ must lie
in the same face of $Vor_c(e)$. Hence, we can simply cut $Vor_c(e)$ to
obtain $Vor(e)$ using the bisectors on the boundary of $R$. When we
apply the wavefront propagation procedure to propagate $W(e)$ to $c$,
here we add an initial step to compute the intersection of $e$ and $B(\alpha,\alpha')$ for each
pair of neighboring generators $\alpha$ and $\alpha'$ of $W(e)$ and
use it as the initial tracing-point $z(\alpha,\alpha')$. Since each such
intersection can be computed in $O(\log n)$ time by the bisector-line
intersection operation in Lemma~\ref{lem:bl-intersection}, this initial step takes $O(h_{e}\log n)$ time, where $h_e$ is the number of generators of $W(e)$.
Hence, the total time for propagating $W(e)$ into $c$ to obtain
$Vor_c(e)$ is $O(h_e\log n+n_B(e))$, where $n_B(e)$ is the number of hyperbolic-arcs of the bisectors of $W(e)$ in $c$.
After having $Vor_c(e)$, cutting it to obtain $Vor(e)$ can be easily done in $O(h_e+n_B(e)+n_R)$ time, where $n_R$ is the number of hyperbolic-arcs on the bisectors of $\partial R$.
Therefore, the total time for computing the partition $Vor(e)$ is
$O(h_e\log n+n_B(e)+n_R)$ time.
Note that since the bisectors of $\partial R$ do not intersect any bisectors involving the generators of $W(e)$, all bisector events of $W(e)$ in $c$ are actually in $R$.
After having the partition $Vor(e)$ for all transparent edges $e$ of
$R$, we can now compute vertices of $\mbox{$S\!P\!M$}'(s)$ in $R$. Consider a transparent edge $e$ of $\partial R$. We process it as
follows. For each transparent edge $f$ of $\partial R$ other than $e$,
we merge the two partitions $Vor(e)$ and $Vor(f)$ by using the merge step
from the standard divide-and-conquer Voronoi diagram algorithm to
compute the sub-region of $R$ closer to $W(e)$ than to $W(f)$. This can be
done in $O(h_e+n_B(e)+n_R+h_f+n_B(f))$ time by finding a finding a curve
$\gamma$ in $R$ that consists of points equal to $W(e)$ and $W(f)$,
and the algorithm is similar to the HS algorithm.
Intersecting the results for all such $f$ produces the region $R(e)$ claimed by $W(e)$ in $R$.
Intersecting $R(e)$ with $Vor(e)$ gives the vertices of $\mbox{$S\!P\!M$}'(s)$ in $R$ that $W(e)$ contributes.
Repeating the above for all transparent edges $e$ of $\partial R$ gives the vertices of $\mbox{$S\!P\!M$}'(s)$ in $R$.
Since $\partial R$ has $O(1)$ transparent edges, the total time is $O(h_R\log n+n_B(R)+n_R)$, where $h_R$ is the number of generators in all wavefronts of all transparent edges of $\partial R$ and $n_B(R)$ is the total number of hyperbolic-arcs of the bisectors all wavefronts of all transparent edges of $\partial R$.
We do the above for all active regions $R$ of $c$, and then all
vertices of $\mbox{$S\!P\!M$}'(s)$ in $c$ can be computed. The total time is $O(h_c\log
n+n_R(c)+n_B(c))$, where $h_c$ is the total number of marked generators in the
wavefronts $W(e)$ of all transparent edges $e$ of $c$, $n_R(c)$ is the
total number of hyperbolic-arcs of the bisector boundaries of the
active regions of $c$, and $n_B(c)$ is the total number of hyperbolic-arcs of the bisectors of the wavefronts of all transparent edges of $c$.
Processing all cells $c$ of $\mathcal{S}'$ as above gives all vertices of
$\mbox{$S\!P\!M$}'(s)$. For the running time, the total sum of $n_R(c)$ among all cells $c\in \mathcal{S}'$ is $O(n)$ because each hyperbolic-arc of a bisector on the boundary of any active region also appears in $\mbox{$S\!P\!M$}'(s)$, whose total size is $O(n)$ by Corollary~\ref{coro:size}. The total sum of $h_c$ among all cells $c$ is $O(h)$ by Lemma~\ref{lem:numgen}.
For $n_B(c)$, Lemma~\ref{lem:bisectorvertices} is essentially a proof that the total sum of $n_B(c)$ over all cells $c$ is $O(n)$. To see this, since $c$ is a cell incident to $e$, the wavefront $W(e)$ will be propagated into $c$ during the wavefront propagation procedure to propagate $W(e)$ to edges of $output(e)$, and thus the hyperbolic-arcs of the bisectors of $W(e)$ are counted in the proof analysis of Lemma~\ref{lem:bisectorvertices}.
In summary, the total time for computing all vertices of $\mbox{$S\!P\!M$}'(s)$ is $O(n+h\log n)$, which is $O(n+h\log h)$.
\subsubsection{Constructing $\mbox{$S\!P\!M$}'(s)$}
\label{sec:second}
With all vertices of $\mbox{$S\!P\!M$}'(s)$ computed above, as in the HS
algorithm, we next compute the edges of $\mbox{$S\!P\!M$}'(s)$ separately and then assemble
them to obtain $\mbox{$S\!P\!M$}'(s)$. One difference is that the HS algorithm
uses a standard plane
sweep algorithm to assemble all edges to obtain $\mbox{$S\!P\!M$}(s)$, which takes
$O(n\log n)$ time as there are $O(n)$ edges in $\mbox{$S\!P\!M$}(s)$. In order to
achieve the desired $O(n+h\log h)$ time bound, here instead we propose a
different algorithm.
We first discuss how to compute the edges of $\mbox{$S\!P\!M$}'(s)$, given the vertices of $\mbox{$S\!P\!M$}'(s)$.
Recall that each vertex $v$ of $\mbox{$S\!P\!M$}'(s)$ is either an intersection
between a bisector and an obstacle edge, or an intersection of
bisectors. By our general position assumption, in the former case, $v$
is in the interior of an obstacle edge; in the latter case, $v$ is the
intersection of three bisectors, i.e., a {\em triple point}.
During our above algorithm for computing vertices of $\mbox{$S\!P\!M$}'(s)$, we
associate each vertex $v$ with the two generators of the bisector that
contains $v$ (more specifically, we associate $v$ with the initial
vertices of the generators).
We create a list of all vertices of $\mbox{$S\!P\!M$}'(s)$, each identified by a
{\em key} consisting of its two generators. If a vertex $v$ is a
triple point, then we put it in the list for three times (each time
with a different pair of generators); if $v$ is a bisector-obstacle
intersection, then it is put in the list once.
We now sort all vertices of $\mbox{$S\!P\!M$}'(s)$ with their keys; by traversing the sorted list, we can group together all vertices belong to the same bisector.
This sorting takes $O(h\log n)$ time as there are $O(h)$ vertices in
$\mbox{$S\!P\!M$}'(s)$.
We take all $\mbox{$S\!P\!M$}'(s)$-vertices in the same bisector and sort them
along the bisector determined by their weighted distances from the
generators of the bisector (each of these distances can be computed in
additional $O(\log n)$ time by computing the tangents from the vertex
to the generators). These sorting takes $O(h\log n)$ time altogether. The
above computes the bisector edges $e$ of $\mbox{$S\!P\!M$}'(s)$ that connect two adjacent
vertices. In fact, each edge $e$ is only implicitly determined in the
sense that the hyperbolic-arcs of $e$ are not explicitly computed.
In addition, for each obstacle $P$, we sort all vertices of $\mbox{$S\!P\!M$}'(s)$
on $\partial P$ to compute the convex-chain edges of $\mbox{$S\!P\!M$}'(s)$
connecting adjacent vertices of $\mbox{$S\!P\!M$}'(s)$ on $\partial P$.
This sorting can be easily done in $O(n+h\log n)$ time for all
obstacles.
For each vertex $v$ of $\mbox{$S\!P\!M$}'(s)$, the above computes its adjacent
vertices. As
discussed above, due to the general position assumption, $v$ has at
most three adjacent vertices. We next define a set $E(v)$ of at most three points for $v$.
For each adjacent vertex $u$ of $v$ in $\mbox{$S\!P\!M$}'(s)$, let $e(v,u)$ be the edge of $\mbox{$S\!P\!M$}'(s)$ connecting them.
$e(v,u)$ is either a bisector edge or a convex-chain edge. In the former case, we refer to each hyperbolic-arc of $e(v,u)$ as a {\em piece} of $e(v,u)$; in the latter case, we refer to each obstacle edge of $e(v,u)$ as a {\em piece}. If we traverse $e(v,u)$ from $v$ to $u$, the endpoint $u'$ of the first piece (incident to $v$) is added to $E(v)$; we define $h(u')$ to be $u$. Since $v$ is adjacent to at most three vertices in $\mbox{$S\!P\!M$}'(s)$, $|E(v)|\leq 3$. In fact, $E(v)$ is exactly the set of vertices adjacent to $v$ in the shortest path map $\mbox{$S\!P\!M$}(s)$.
With all edges of $\mbox{$S\!P\!M$}'(s)$ and the sets $E(v)$ of all vertices $v$ of $\mbox{$S\!P\!M$}'(s)$ computed above,
we construct the doubly-connected-edge-list (DCEL) of $\mbox{$S\!P\!M$}'(s)$, as follows.
For convenience, we assume that there is a bounding box that contains all obstacles so that no bisector edge of $\mbox{$S\!P\!M$}'(s)$ will go to infinity and thus each face of $\mbox{$S\!P\!M$}'(s)$ is bounded.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=1.5in]{traceedge.pdf}
\caption{\footnotesize Illustrating $v$, $u$, $u'$, $v'$, $b'$, and $b$.}
\label{fig:traceedge}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
For each vertex $v$ of $\mbox{$S\!P\!M$}'(s)$, we store its coordinate in the DCEL data structure. As
$|E(v)|\leq 3$, there are at most three faces in $\mbox{$S\!P\!M$}'(s)$ incident
to $v$, we construct them one by one. For each point $u'$ in $E(v)$, we
construct the face $F$ clockwise (with respect to $v$) incident to the
edge $e(v,u)$ of $\mbox{$S\!P\!M$}'(s)$ connecting $v$ and $u=h(u')$, as follows (e.g., see Fig.~\ref{fig:traceedge}).
We first trace out the edge $e(v,u)$, which is either a bisector edge or a convex chain edge.
In the former case, we compute the hyperbolic-arcs of $e(v,u)$, one
at a time, until we reach $u$, and add them to the DCEL data structure. Each
hyperbolic-arc can be computed in constant time using the two
generators of the bisector. In the latter case, we trace out the
obstacle edges of $e(v,u)$, one at a time, and add them to the DCEL data
structure. Let $v'$ be the last endpoint of the piece of $e(v,u)$ containing $u$ (e.g., see Fig.~\ref{fig:traceedge}). Note that $v'$ is in the set $E(u)$.
Let $b'$ be the first point of $E(u)$ counterclockwise around $u$ after $v'$
(e.g., see Fig.~\ref{fig:traceedge}). Let $b=h(b')$, which is a vertex
of $\mbox{$S\!P\!M$}'(s)$ adjacent to $u$.
Hence, the edge $e(u,b)$ of $\mbox{$S\!P\!M$}'(s)$ connecting $u$ to $b$ is incident to the
face $F$. We trace out the edge $e(u,b)$ in the same way as above. When we
reach $b$, we continue to trace the next edge of $F$.
Since each face of $\mbox{$S\!P\!M$}'(s)$ is bounded, eventually we
will arrive back to the vertex $v$ again\footnote{We could lift the assumption that each face of $\mbox{$S\!P\!M$}'(s)$ is bounded in the following way. During the above algorithm for constructing $F$, if $F$ is unbounded, then we will reach a bisector edge that extends to the infinity. If that happens, then we construct other edges of $F$ from the other direction of $v$. More specifically, starting from the first point of $E(v)$ clockwise around $v$ after $u'$, we trace out the edges of $F$ in the same way as above, until we reach a bisector edge that extends to the infinity, at which moment all edges of $F$ are constructed.}. This finishes the construction of the
face $F$. We do the same for all other faces of $\mbox{$S\!P\!M$}'(s)$, after
which the DCEL data structure for $\mbox{$S\!P\!M$}'(s)$ is constructed.
For the running time, since each edge of $\mbox{$S\!P\!M$}'(s)$ is traced at
most twice, by Corollary~\ref{coro:size}, the total time of the above procedure for constructing the DCEL data structure is $O(n)$.
In summary, $\mbox{$S\!P\!M$}'(s)$ can be computed in $O(n+h\log h)$ time. By
Lemma~\ref{lem:spmconstruction}, the shortest path map $\mbox{$S\!P\!M$}(s)$ can be built
in additional $O(n)$ time.
\subsection{Reducing the space to $O(n)$}
\label{sec:space}
The above provides an algorithm for computing $\mbox{$S\!P\!M$}(s)$ in $O(n+h\log h)$ time and $O(n+h\log h)$ space. In this subsection, we discuss how to reduce to the space to $O(n)$, using the technique given in~\cite{ref:WangSh21}.
The reason that the above algorithm needs $O(n+ h\log h)$ space is two-fold. First, it uses fully persistent binary trees (with the path-copying method) to represent wavefronts $W(e)$. Because there are $O(h)$ bisector events in the wavefront expansion algorithm and each event costs $O(\log n)$ additional space on a persistent tree, the total space of the algorithm is $O(n+h\log n)$. Second, in order to construct $\mbox{$S\!P\!M$}(s)$ after the wavefront expansion algorithm, the wavefronts $W(e)$ of all transparent edges $e$ of $\mathcal{S}'$ are needed, which are maintained in those persistent trees.
We resolve these two issues in the following way.
\subsubsection{Reducing the space of the wavefront expansion algorithm}
We still use persistent trees to represent wavefronts. However, as there are $O(h)$ bisector events in total in the algorithm, we divide the algorithm into $O(\log h)$ phases so that each phase has no more than $h/\log h$ events. The total additional space for processing the events using persistent trees in each phase is $O(h)$. At the end of each phase, we ``reset'' the space of the algorithm by only storing a ``snapshot'' of the algorithm (and discarding all other used space) so that (1) the snapshot contains sufficient information for the subsequent algorithm to proceed as usual, and (2) the total space of the snapshot is $O(h)$.
Specifically, we make the following changes to the wavefront propagation procedure, which is to compute the wavefronts $W(e,g)$ for all edges $g\in output(e)$ using the wavefront $W(e)$.
We now maintain a counter $count$ to record the number of bisector events that have been processed so far since the last space reset; $count=0$ initially. Consider a wavefront propagation procedure on the wavefront $W(e)$ of a transparent edge $e$. The algorithm will compute $W(e,g)$ for all edges $g\in output(e)$, by propagating $W(e)$. We apply the same algorithm as before. For each bisector event, we first do the same as before. Then, we increment $count$ by one. If $count< h/\log h$, we proceed as before (i.e., process the next event). Otherwise, we have reached the end of the current phase and will start a new phase. To do so, we first reset $count=0$ and then reset the space by constructing and storing a snapshot of the algorithm (other space occupied by the algorithm is discarded), as follows.
\begin{enumerate}
\item
Let $g$ refer to the edge of $output(e)$ whose $W(e,g)$ is currently being computed in the algorithm.
We store the tree that is currently being used to compute $W(e,g)$ right after the above event. To do so, we can make a new tree by copying the newest version of the current persistent
tree the algorithm is operating on. The size of the tree is bounded by $O(h)$. We will use this tree to ``resume'' computing $W(e,g)$ in the subsequent algorithm.
\item
For each $g'\in output(e)\setminus\{g\}$ whose $W(e,g')$
has been computed, we store the tree for $W(e,g')$. We will use the tree to
compute the wavefronts $W(g')$ of $g'$ in the subsequent algorithm.
\item
We store the tree for the wavefront $W(e)$. Note that the tree may have many versions due to processing the events and we only keep its original version for $W(e)$. Hence, the size of the tree is $O(h)$. This tree will be used in the subsequent algorithm to compute $W(e,g')$ for those edges $g'\in output(e)\setminus\{g\}$ whose $W(e,g')$ have not been computed yet.
\item
We check every transparent edge $e'$ of $\mathcal{S}'$ with $e'\neq e$. If $e'$ has been
processed (i.e., the wavefront propagation procedure has been called
on $W(e')$) and there is an edge $g'\in output(e')$ that has {\em not} been processed, we know that $W(e',g')$ has been computed and is available; we store the tree for $W(e',g')$. We will use the tree to compute the wavefronts $W(g')$ of $g'$ in the subsequent algorithm.
\end{enumerate}
We refer to the wavefronts stored in the algorithm as the {\em snapshot}; intuitively, the snapshot contains all wavelets in the forefront of the wavelet expansion.
By the same analysis as in~\cite{ref:WangSh21}, we can show that the snapshot contains sufficient information for the subsequent algorithm to proceed as usual and the total space of the snapshot is $O(h)$.
The above discusses our changes to the wavefront propagation procedure. For the wavefront merging procedure, which is to construct $W(e)$ from $W(f,e)$ for the edges $f\in input(e)$,
notice that we do not need the old versions of $W(f,e)$ anymore after $W(e)$ is constructed. Therefore, it is not necessary to use the path-copying method to process each event in the procedure.
Hence, the total space needed in the wavefront merging procedure in the entire algorithm is $O(n)$.
\subsubsection{Reducing the space of constructing $\mbox{$S\!P\!M$}'(s)$}
For the second issue of constructing $\mbox{$S\!P\!M$}'(s)$, our algorithm relies on the wavefronts $W(e)$ for all transparent edges $e$, which are maintained by persistent trees.
Due to the space-reset, our algorithm does not maintain the wavefronts anymore, and thus we need to somehow restore these wavefronts in order to construct $\mbox{$S\!P\!M$}'(s)$. To this end, a key observation is that by marking a total of $O(h)$ additional wavelet generators it is possible to restore all historical wavefronts that are needed for constructing $\mbox{$S\!P\!M$}'(s)$. In this way, $\mbox{$S\!P\!M$}'(s)$ can be constructed in $O(n + h\log h)$ time and $O(n)$ space.
More specifically, our algorithm considers each cell $c$ of $\mathcal{S}'$ individually. For each cell $c$, the algorithm has two steps. First, compute the active regions of $c$. Second, for each active region $R$, compute the vertices of $\mbox{$S\!P\!M$}'(s)$ in $R$. For both steps, our algorithm utilizes the wavefronts $W(e)$ of the transparent edges $e$ on the boundary of $c$. Due to the space reset, the wavefronts $W(e)$ are not available anymore in our new algorithm. We use the following strategy to resolve the issue.
First, to compute the active regions in $c$, we need to know the bisectors defined by an unmarked generator $\alpha$ and a marked generator $\alpha'$ in the wavefronts $W(e)$ of the transparent edges $e$ of $c$. We observe that $\alpha$ is a generator adjacent to $\alpha'$ in $W(e)$. Based on this observation, we slightly modify our wavefront expansion algorithm so that it also marks the neighbors of the generators that are marked in our original algorithm and we call them {\em newly-marked} generators (the generators marked in our original algorithm are called {\em originally-marked} generators). If a generator is both newly-marked and originally-marked, we consider it as originally-marked. As each generator has two neighbors in a wavefront, the total number of marked generators is still $O(h)$. The newly-marked generators and the originally-marked generators are sufficient for computing all active regions of each cell $c$. Indeed, the active regions are decomposition of $c$ by the bisectors of adjacent generators with one originally-marked and the other newly-marked in $W(e)$ of the transparent edges $e$ of $c$.
Second, to compute the vertices of $\mbox{$S\!P\!M$}'(s)$ in each active region $R$ of $c$, we need to restore the wavefronts $W(e)$ of the transparent edges $e$ on the boundary of $R$. To this end, we observe that $W(e)$ consists of exactly the originally-marked generators that claim $e$. Consequently, the same algorithm as before can be applied to construct $\mbox{$S\!P\!M$}'(s)$. When computing the partition $Vor(e)$ of $R$, we propagate $W(e)$ using the wavefront propagation procedure, which uses persistent trees to represent $W(e)$. Since here we do not need to keep the old versions of $W(e)$ any more, we can use an ordinary tree (without the path-copying method) to represent $W(e)$. In this way, processing each bisector event only introduces $O(1)$ additional space. Hence, constructing $\mbox{$S\!P\!M$}'(s)$ takes $O(n+h\log h)$ time and $O(n)$ space.
\section{The general case}
\label{sec:general}
In this section, we extend our algorithm for the convex case in Section~\ref{sec:convex} to the general case where obstacles of $\mathcal{P}$ may not be convex. To this end, we resort to an extended corridor structure of $\mathcal{P}$, which has been used to solve various problems in polygonal domains~\cite{ref:ChenCo19,ref:ChenTw16,ref:ChenCo17,ref:ChenA15,ref:KapoorAn97,ref:MitchellSe95}. The structure decomposes the free space $\mathcal{F}$ into three types of regions: an {\em ocean} $\mathcal{M}$, $O(h)$ {\em canals}, and $O(n)$ {\em bays}. The details are given below.
\subsection{The extended corridor structure}
Let $\mbox{$T\!r\!i$}(\mathcal{P})$ denote an arbitrary triangulation of $\mathcal{P}$.
Let $G(\mathcal{P})$ be the (planar) dual graph of $\mbox{$T\!r\!i$}(\mathcal{P})$,
i.e., each node of $G(\mathcal{F})$ corresponds to a triangle in
$\mbox{$T\!r\!i$}(\mathcal{P})$ and each edge connects two nodes of $G(\mathcal{P})$
corresponding to two triangles sharing a triangulation diagonal of $\mbox{$T\!r\!i$}(\mathcal{P})$.
We compute a {\em corridor graph} $G$, as follows. First, repeatedly remove every degree-one node from $G(\mathcal{P})$ until no such node remains. Second, repeatedly remove every degree-two node from
$G(\mathcal{P})$ and replace its two incident edges by a single edge
until no such node remains. The resulting graph is $G$ (e.g., see Fig.~\ref{fig:triangulation}), which has
$O(h)$ faces, nodes, and edges~\cite{ref:KapoorAn97}. Each node of
$G$ corresponds to a triangle in $\mbox{$T\!r\!i$}(\mathcal{P})$, which is called a
{\em junction triangle} (e.g., see Fig.~\ref{fig:triangulation}).
The removal of all junction triangles results in $O(h)$
{\em corridors} (defined below), and each corridor
corresponds to an edge of $G$.
\begin{figure}[t]
\begin{minipage}[t]{0.47\linewidth}
\begin{center}
\includegraphics[totalheight=1.5in]{triangulation.pdf}
\caption{\footnotesize
Illustrating a triangulation of $\mathcal{P}$ with two obstacles.
There are two junction triangles indicated by the large dots inside
them, connected by three solid (red) curves. Removing the two
junction triangles results in three corridors.}
\label{fig:triangulation}
\end{center}
\end{minipage}
\hspace*{0.04in}
\begin{minipage}[t]{0.52\linewidth}
\begin{center}
\includegraphics[totalheight=1.6in]{corridor.pdf}
\caption{\footnotesize
Illustrating an open hourglass (left) and a
closed one (right) with a corridor path connecting the apices
$x$ and $y$ of the two funnels. The dashed segments are diagonals.
The paths $\pi_C(a,b)$ and $\pi_C(e,f)$ are marked by thick solid
curves. A bay $bay(\overline{cd})$ with gate $\overline{cd}$ (left)
and a canal $canal(x,y)$ with gates $\overline{xd}$ and $\overline{yz}$
(right) are also shown.} \label{fig:corridor}
\end{center}
\end{minipage}
\vspace*{-0.15in}
\end{figure}
The boundary of a corridor $C$ consists of four parts (see
Fig.~\ref{fig:corridor}): (1) A boundary portion of an obstacle, from an obstacle vertex $a$ to an obstacle vertex $b$; (2) a triangulation
diagonal of a junction triangle from $b$ to an obstacle vertex $e$; (3) a boundary portion of an obstacle from $e$ to an obstacle vertex $f$; (4) a diagonal of a
junction triangle from $f$ to $a$. The two diagonals $\overline{be}$
and $\overline{af}$ are called the {\em doors} of $C$, and the other two parts of the boundary of $C$ are the two {\em sides} of $C$. Note that $C$ is a simple polygon. A point is in the {\em interior} of $C$ if it is in $C$ excluding the two doors.
Let $\pi_C(a,b)$ (resp., $\pi_C(e,f)$) be the shortest path from $a$ to $b$
(resp., $e$ to $f$) in $C$. The region $H_C$ bounded by
$\pi_C(a,b), \pi_C(e,f)$, $\overline{be}$, and
$\overline{fa}$ is called an {\em hourglass}, which is {\em open} if
$\pi_C(a,b)\cap \pi_C(e,f)=\emptyset$ and {\em closed} otherwise (see
Fig.~\ref{fig:corridor}). If $H_C$ is open, then both $\pi_C(a,b)$ and
$\pi_C(e,f)$ are convex chains and called the {\em sides} of
$H_C$; otherwise, $H_C$ consists of two ``funnels''
and a path $\pi_C=\pi_C(a,b)\cap \pi_C(e,f)$ joining the two apices of the
funnels, called the {\em corridor path} of $C$.
Each side of a funnel is also convex.
Let $\mathcal{M}$ be the union of all $O(h)$ junction triangles, open hourglasses, and funnels.
We call $\mathcal{M}$ the {\em ocean}, whose boundary
$\partial\mathcal{M}$ consists of $O(h)$ convex chains that are sides of open hourglasses and funnels.
The other space of $\mathcal{P}$, i.e., $\mathcal{P}\setminus\mathcal{M}$, is further
partitioned into two types of regions: {\em bays} and {\em canals}, defined as follows.
Consider the hourglass $H_C$ of a corridor $C$.
If $H_C$ is open (see Fig.~\ref{fig:corridor}), then $H_C$
has two sides. Let $S_1$ be a side of $H_C$.
The obstacle vertices on $S_1$ all lie on the same
side of $C$. Let $c$ and $d$ be any two consecutive
vertices on $S_1$ such that $\overline{cd}$ is
not an obstacle edge of $\mathcal{P}$ (e.g., see Fig.~\ref{fig:corridor} left). The region enclosed
by $\overline{cd}$ and the boundary portion of $C$ between $c$ and
$d$ is called a {\em bay}, denoted by
$bay(\overline{cd})$. We call
$\overline{cd}$ the {\em gate} of $bay(\overline{cd})$.
If $H_C$ is closed, let $x$ and $y$ be the two apices
of the two funnels. Consider two consecutive vertices $c$ and $d$ on
a side of a funnel such that $\overline{cd}$ is not an obstacle edge of $\mathcal{P}$.
If $c$ and $d$ are on the same side of the corridor $C$, then $\overline{cd}$ also
defines a bay. Otherwise, either $c$ or $d$ is a funnel
apex, say, $c=x$, and we call $\overline{xd}$ a {\em canal gate} at $x=c$
(e.g., see Fig.~\ref{fig:corridor} right).
Similarly, there is also a canal gate
at the other funnel apex $y$, say $\overline{yz}$.
The region of $C$ between the two canal gates
$\overline{xd}$ and $\overline{yz}$
is the {\it canal} of $H_C$, denoted by $canal(x,y)$.
Each bay or canal is a simple polygon.
All bays and canals together constitute the space $\mathcal{P}\setminus\mathcal{M}$.
Each vertex of $\partial\mathcal{M}$ is a vertex of $\mathcal{P}$ and
each edge of $\partial\mathcal{M}$ is either an edge of $\mathcal{P}$ or a gate
of a bay/canal. Gates are common boundaries between $\mathcal{M}$ and bays/canals.
After $\mathcal{P}$ is triangulated, $\mathcal{M}$ and all bays and canals can be obtained in $O(n)$ time~\cite{ref:KapoorAn97}.
The reason that the extended corridor structure can help find a shortest path is the following. Suppose we want to find a shortest $s$-$t$\ path for two points $s$ and $t$. We consider $s$ and $t$ as two special obstacles and build the extended corridor structure. If a shortest $s$-$t$\ path $\pi(s,t)$ contains a point in the interior of a corridor $C$, then $\pi(s,t)$ must cross both doors of $C$ and stay in the hourglasses of $C$, and further, if the hourglass is closed, then its corridor path must be contained in $\pi(s,t)$. In fact, $\pi(s,t)$ must be in the union of the ocean $\mathcal{M}$ and all corridor paths~\cite{ref:KapoorAn97}.
In light of the above properties, we propose the following algorithm.
Let $s$ be a given source point. By considering $s$ as a special obstacle of $\mathcal{P}$, we construct the extended corridor structure of $\mathcal{P}$. Consider any query point $t$, which may be in the ocean $\mathcal{M}$, a bay $bay(\overline{cd})$, or a canal $canal(x,y)$.
\begin{itemize}
\item
If $t\in \mathcal{M}$, then the union of $\mathcal{M}$ and all corridor paths contains a shortest $s$-$t$\ path. To handle this case, we will build a shortest path map $\mbox{$S\!P\!M$}(\mathcal{M})$ in $\mathcal{M}$ with respect to the union of $\mathcal{M}$ and all corridor paths. In face, $\mbox{$S\!P\!M$}(\mathcal{M})$ is exactly the portion of $\mbox{$S\!P\!M$}(s)$ in $\mathcal{M}$, i.e., $\mbox{$S\!P\!M$}(s)\cap \mathcal{M}$. To build $\mbox{$S\!P\!M$}(\mathcal{M})$, a key observation is that the boundary $\partial \mathcal{M}$ consists of $O(h)$ convex chains. Therefore, we can easily adapt our previous algorithm for the convex case. However, the algorithm needs to be modified so that the corridor paths should be taken into consideration. Intuitively, corridor paths provide certain kind of ``shortcuts'' for wavefronts to propagate.
\item
If $t$ is in a bay $bay(\overline{cd})$, then any shortest $s$-$t$\ path must cross its gate $\overline{cd}$. To handle this case, we will extend $\mbox{$S\!P\!M$}(\mathcal{M})$ into $bay(\overline{cd})$ through the gate $\overline{cd}$ to construct the shortest path map in $bay(\overline{cd})$, i.e., the portion of $\mbox{$S\!P\!M$}(s)$ in $bay(\overline{cd})$, $\mbox{$S\!P\!M$}(s)\cap bay(\overline{cd})$.
\item
If $t$ is in a canal $canal(x,y)$, then any shortest $s$-$t$\ path must cross one of the two gates of the canal. To handle this case, we will extend $\mbox{$S\!P\!M$}(\mathcal{M})$ into $canal(x,y)$ through the two gates to construct the shortest path map in $canal(x,y)$, i.e., the portion of $\mbox{$S\!P\!M$}(s)$ in $canal(x,y)$, $\mbox{$S\!P\!M$}(s)\cap canal(x,y)$.
\end{itemize}
In the following, we first describe our algorithm for constructing $\mbox{$S\!P\!M$}(\mathcal{M})$ in Section~\ref{sec:ocean}. We then expand $\mbox{$S\!P\!M$}(\mathcal{M})$ into all bays in Section~\ref{sec:bay} and expand $\mbox{$S\!P\!M$}(\mathcal{M})$ into all canals in Section~\ref{sec:canal}. The algorithm for the canal case utilizes the bay case algorithm as a subroutine.
\subsection{Constructing the shortest path map in the ocean $\mbox{$S\!P\!M$}(\mathcal{M})$}
\label{sec:ocean}
As the boundary of $\mathcal{M}$ consists of $O(h)$ convex chains, we can
apply and slightly modify our algorithm for the convex case. To do so, for each
convex chain of $\partial\mathcal{M}$, we define its rectilinear extreme
vertices in the same way as before. Let $\mathcal{V}$ be the set of the rectilinear extreme vertices of
all convex chains. Hence, $|\mathcal{V}|=O(h)$. In addition, to incorporate the
corridor paths into the algorithm, we include the endpoints of each corridor path in $\mathcal{V}$. As there are $O(h)$ corridor paths, the size of $\mathcal{V}$ is still bounded by $O(h)$. Note that each corridor path endpoint is also an endpoint of a convex chain of $\partial \mathcal{M}$. We construct the conforming
subdivision $\mathcal{S}$ based on the points of $\mathcal{V}$
and then insert the convex chains of $\mathcal{M}$ into
$\mathcal{S}$ to obtain $\mathcal{S}'$. The algorithm is essentially the same as
before. In addition, we make the following changes to $\mathcal{S}'$, which is mainly for incorporating the corridor paths into our wavefront expansion algorithm, as will be clear later.
Let $v$ be an endpoint of a corridor path $\pi$. Since $v$ is in $\mathcal{V}$, $v$ is incident to $O(1)$ transparent edges in $\mathcal{S}'$. For each such transparent edge $e$, if $|\pi|<2 \cdot |e|$, then we divide $e$ into two sub-edges such that the length of the one incident to $v$ is equal to $|\pi|/2$; for each sub-edge, we set its well-covering region the same as $\mathcal{U}(e)$. Note that this does not affect the properties of $\mathcal{S}'$. In particular, each transparent edge $e$ is still well-covered. This change guarantees the following property: for each corridor path $\pi$, $|\pi|\geq 2\cdot |e'|$ holds, where $e'$ is any transparent edge of $\mathcal{S}'$ incident to either endpoint of $\pi$. For reference purpose, we refer to it as the {\em corridor path length property}.
Next we apply the wavefront expansion algorithm. Here we need to
incorporate the corridor paths into the algorithm. Intuitively, each
corridor path provides a ``shortcut'' for the wavefront, i.e., if a
wavelet hits an endpoint of a corridor path, then the wavelet will
come out of the corridor path from its other endpoint but with a delay
of distance equal to the length of the corridor path. More details are
given below.
Since all corridor path endpoints are in $\mathcal{V}$, they are vertices of transparent edges of $\mathcal{S}'$.
Consider an endpoint $v$ of a corridor path $\pi$. Let $u$ be the other endpoint of $\pi$.
Recall that the wavefront propagation procedure for $W(e)$ is to propagate $W(e)$ to compute $W(e,g)$ for all edges $g\in output(e)$. In addition to the previous algorithm for the procedure, we also propagate $W(e)$ through the corridor path $\pi$ to $u$. This is done as follows. Recall that when $e$ is processed, since $v$ is an endpoint of $e$, the weighted distance of $v$ through $W(e)$ is equal to $d(s,v)$. Hence, the wavefront $W(e)$ hits $u$ through $\pi$ at time $d(s,v)+|\pi|$. We then update $covertime(g)= \min\{covertime(g),d(s,v)+|\pi|+|g|\}$, for each transparent edge $g$ incident to $u$. We also set the wavefront $W(e,g)$ consisting of the only wavelet with $u$ as the generator with weight equal to $d(s,v)+|\pi|$. Since there are $O(1)$ transparent edges $g$ incident to $u$, the above additional step takes $O(1)$ time, which does not change the time complexity of the overall algorithm asymptotically. The corridor path length property assures that if $W(e)$ contributes to a wavefront $W(g)$ at $g$, then $e$ must be processed earlier than $g$. This guarantees the correctness of the algorithm.
In this way, we can first construct a decomposition $\mbox{$S\!P\!M$}'(\mathcal{M})$ of $\mathcal{M}$ in $O(n+h\log h)$ time and $O(n)$ space, where $\mbox{$S\!P\!M$}'(\mathcal{M})$ is defined similarly as $\mbox{$S\!P\!M$}'(s)$ in Section~\ref{sec:convex}. Then, by a similar algorithm as that for Lemma~\ref{lem:spmconstruction}, $\mbox{$S\!P\!M$}(\mathcal{M})$ can be obtained in additional $O(n)$ time.
\subsection{Expanding $\mbox{$S\!P\!M$}(\mathcal{M})$ into all bays}
\label{sec:bay}
We now expand $\mbox{$S\!P\!M$}(\mathcal{M})$ into all bays in $O(n+h\log h)$ time and $O(n)$ space. In fact, we expand $\mbox{$S\!P\!M$}'(\mathcal{M})$ to the bays.
We process each bay individually. Consider a bay $bay(\overline{cd})$ with gate $\overline{cd}$.
Without loss of generality, we assume that $\overline{cd}$ is
horizontal, $c$ is to the left of $d$, and $bay(\overline{cd})$ is locally above
$\overline{cd}$.
Let $v_1,v_2,\ldots,v_{m}$ be the vertices of $\mbox{$S\!P\!M$}'(\mathcal{M})$ on
$\overline{cd}$ ordered from left to right (e.g., see Fig.~\ref{fig:bayextend}). Let $c=v_0$ and
$d=v_{m+1}$. Hence, each $\overline{v_iv_{i+1}}$ is claimed by a
generator $\alpha_i=(A_i,a_i)$ for all $i=0,1,\ldots,m$.
Let $b_i$ and $c_i$ be the tangent points on $A_{i-1}$ and $A_i$ from
$v_i$, respectively, for each $i=1,2,\ldots,m$ (e.g., see Fig.~\ref{fig:bayextend}). For $v_0$, only $c_0$
is defined; for $v_{m+1}$, only $b_{m+1}$ is defined.
Observe that for any point $p\in \overline{v_iv_{i+1}}$, which is
claimed by $\alpha_i$, its tangent point on $A_i$ must be on the
portion of $A_i$ between $c_i$ and $b_{i+1}$ and we use $A_i'$ to
denote that portion. So with respect to $bay(\overline{cd})$, we use
$\alpha_i'=(A_i',a_i')$ to
refer to the generator, where $a_i'$ refers to the one of $c_i$ and
$b_{i+1}$ that is closer to $a_i$. Hence, for any point $t\in bay(\overline{cd})$,
any shortest path $\pi(s,t)$ from $s$ to $t$ must be via one of the
generators $\alpha'_i$ for $i=0,1,\ldots,m$.
Consider the region $R$ bounded by $A'_i$ for all $i\in [0,m]$, the
tangents from $v_i$ to their generators for all $i\in [0,m+1]$, and
the boundary of the bay excluding its gate. Notice that $R$ is a
simple polygon. For any point $t\in bay(\overline{cd})$, the above observation implies any shortest $s$-$t$\ path $\pi(s,t)$ is the
concatenation of a shortest path from $s$ to a generator initial vertex $a_i'$ and
the shortest path from $a_i'$ to $t$ in $R$.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=2.7in]{bayextend.pdf}
\caption{\footnotesize Illustrating $bay(\overline{cd})$ and the generators. The thick segments on obstacles are $A_i'$, $i=0,1,\ldots,m$.}
\label{fig:bayextend}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
According to the above discussion, expanding $\mbox{$S\!P\!M$}'(\mathcal{M})$ into $bay(\overline{cd})$
is equivalent to the following weighted geodesic Voronoi diagram
problem in a simple polygon: Partition $R$ into cells with respect to
the point sites $a_0',a_1',\ldots,a_m'$ (with weights equal to their
geodesic distances to $s$) such that all points in the same cell have the same closest site.
Let $n_b$ be the number of vertices in $bay(\overline{cd})$ (the subscript ``b'' represents ``bay''). Let $n_g$ be the total number of obstacle vertices in $A_i'$ for all $i\in [0,m]$ (the subscript ``g'' represents ``generator'').
Note that $v_i$ for all $i=1,\ldots,m$ are also vertices of $R$.
Hence, the number of vertices of $R$ is $n_b+n_g+m$.
The above problem can be solved in $O(m\log m+n_b+n_g)$ time by the
techniques of Oh~\cite{ref:OhOp19}. Indeed, given $m'$ point sites in a simple polygon $P'$ of $n'$ vertices, Oh~\cite{ref:OhOp19} gave an algorithm that can compute the geodesic Voronoi diagram of the sites in $P'$ in $O(n'+m'\log m')$ time and $O(n'+m')$ space. Although the point sites in Oh's problem do not have weights, our problem is essentially an intermediate step of Oh's algorithm because all weighted point sites in our problem are on one side of $\overline{cd}$. Therefore, we can run Oh's algorithm from ``the middle'' and solve our problem in $O(m\log m+n_b+n_g)$ and $O(n_b+n_g)$ space. In fact, our problem is a special case of Oh's problem because there are no sites in $bay(\overline{cd})$. For this reason, we propose our own algorithm to solve this special case and the algorithm is much simpler than Oh's algorithm; our algorithm also runs in $O(m\log m+n_b+n_g)$ and $O(n_b+n_g+m)$ space. This also makes our paper more self-contained.
Before presenting the algorithm, we analyze the total time for processing all bays.
Since $\mbox{$S\!P\!M$}'(\mathcal{M})$ has $O(h)$ vertices, the total sum of $m$ for all bays is $O(h)$. The total sum of $n_b$ for all bays is at most $n$. Notice that the obstacle edges on $A_i'$ are disjoint for different bays, and thus the total sum of $n_g$ for all bays is $O(n)$. Hence, expanding $\mbox{$S\!P\!M$}'(\mathcal{M})$ to all bays takes $O(n+h\log h)$ time and $O(n)$ space in total.
The above actually only considers the case where the gate $\overline{cd}$ contains at least one vertex of $\mbox{$S\!P\!M$}'(\mathcal{M})$. It is possible that no vertex of $\mbox{$S\!P\!M$}'(\mathcal{M})$ is on $\overline{cd}$, in which case the entire gate is claimed by one generator $\alpha$ of $\mbox{$S\!P\!M$}'(\mathcal{M})$. We can still define the region $R$ in the same way. But now $R$ has only one weighted site and thus the geodesic Voronoi diagram problem becomes computing a shortest path map in the simple polygon $R$ for a single source point. This problem can be solved in $O(n_b+n_g)$ time~\cite{ref:GuibasLi87}; note that $m=0$ in this case. Hence, the total time for processing all bays in this special case is $O(n)$.
\subsubsection{Solving the special weighted geodesic Voronoi diagram problem}
We present an algorithm for the above special case of the weighted geodesic Voronoi diagram problem and the algorithm runs in $O(m\log m+n_b+n_g)$ and $O(n_b+n_g+m)$ space.
If we consider all generators $\alpha_i$ for $i=0,1,\ldots,m$ as a wavefront at $\overline{cd}$, denoted by $W(\overline{cd})$, then our algorithm is essentially to propagate $W(\overline{cd})$ inside $bay(\overline{cd})$. To this end, we first triangulate $bay(\overline{cd})$ and will use the triangulation to guide the wavefront propagation. Each time we propagate the wavefront through a triangle. In the following, we first discuss how to propagate $W(\overline{cd})$ through the triangle $\triangle cda$ with $\overline{cd}$ as an edge and $a$ as the third vertex. This is a somewhat special case as $\triangle cda$ is the first triangle the wavefront will propagate through; later we will discuss the general case but the algorithm is only slightly different.
Recall that each convex chain $A_i'$ is represented by an array and the generator list $W(\overline{cd})$ is represented by a balanced binary search tree $T(W(\overline{cd}))$. We build a point location data structure on the triangulation of $bay(\overline{cd})$ in $O(n_b)$ time~\cite{ref:EdelsbrunnerOp86,ref:KirkpatrickOp83}, so that given any query point $p$, we can determine the triangle that contains $p$ in $O(\log n_b)$ time.
We begin with computing the intersection of the adjacent bisectors $B(\alpha_{i-1},\alpha_i)$ and $B(\alpha_i,\alpha_{i+1})$ for all $i=1,2,\ldots, m-1$. Each intersection can be computed in $O(\log n_g)$ time by the bisector-bisector intersection operation in Lemma~\ref{lem:bb-intersection}. Computing all intersections takes $O(m\log n_g)$ time. For each intersection $q$, called a {\em bisector event}, we use the point location data structure to find the triangle of the triangulation that contains $q$ and store $q$ in the triangle.
Since all generators are outside $bay(\overline{cd})$, by Corollary~\ref{coro:monotone}, all bisectors are monotone with respect to the direction orthogonal to $\overline{cd}$. Our algorithm for propagating the wavefront through $\triangle cda$ is based on this property.
We sort all bisector events in $\triangle cda$ according to their perpendicular distances to the supporting line of $\overline{cd}$. Then, we process these events in the same way as in our wavefront propagation procedure. Specifically, for each bisector event $q$ of two bisectors $B(\alpha',\alpha)$ and $B(\alpha,\alpha'')$, we remove $\alpha$ from the generator list. Then, we compute the intersection $q'$ of $B(\alpha',\alpha'')$ with $B(\alpha_1',\alpha')$, where $\alpha_1'$ is the other neighboring bisector of $\alpha'$ than $\alpha$, and we use the point location data structure to find the triangle that contains $q'$ and store $q'$ in the triangle. If $q'\in \triangle cda$, then we insert it to the bisector event sorted list of $\triangle cda$. We do the same for $B(\alpha',\alpha'')$ and $B(\alpha'',\alpha''_1)$, where $\alpha''_1$ is the other neighbor of $\alpha''$ than $\alpha$. After all events in $\triangle cda$ are processed, we split the current wavefront $W$ at the vertex $a$. To this end, we first find the generator $\alpha^*$ of $W$ that claims $a$. For this, we use the same algorithm as before, i.e., binary search plus bisector tracing. So we need to maintain a tracing-point for each bisector as before (initially, we can set $v_i$ as the tracing-point for $B(\alpha_{i-1},\alpha_i)$, $i=1,2,\ldots,m$).
The correctness of the above algorithm for finding the generator $\alpha^*$ relies on the property of the following lemma.
\begin{lemma}\label{lem:intersectionbay}
The bisector $B(\alpha,\alpha')$ intersects $\overline{ad}\cup \overline{ac}$ at most once for any two bisectors of $\alpha$ and $\alpha'$ of $W(\overline{cd})$.
\end{lemma}
\begin{proof}
Note that we cannot apply the result of Lemma~\ref{lem:intersection} since to do so we need to have $\overline{ad}$ and $\overline{ac}$ parallel to $\overline{cd}$. But the proof is somewhat similar to that for Lemma~\ref{lem:intersection}.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=2.0in]{intersectionbay.pdf}
\caption{\footnotesize Illustrating the proof of Lemma~\ref{lem:intersectionbay}.}
\label{fig:intersectionbay}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
Assume to the contrary that $B(\alpha,\alpha')$ intersects $\overline{ad}\cup \overline{ac}$ at two points, $q_1$ and $q_2$. Let $A$ and $A'$ be the underlying arcs of $\alpha$ and $\alpha'$, respectively. Let $v_1$ and $u_1$ be the tangents points of $q_1$ on $A$ and $A'$, respectively (e.g., see Fig.~\ref{fig:intersectionbay}).
Let $v_2$ and $u_2$ be the tangents points of $q_2$ on $A$ and $A'$, respectively.
Since both $A$ and $A'$ are on one side of $\overline{cd}$ while $\overline{ad}\cup \overline{ac}$ is on the other side, if we move a point $q$ from $q_1$ to $q_2$ on $\overline{ad}\cup \overline{ac}$, the tangent from $q$ to $A$ will continuously change from $\overline{q_1v_1}$ to $\overline{q_2v_2}$ and the tangent from $q$ to $A'$ will continuously change from $\overline{q_1u_1}$ to $\overline{q_2u_2}$. Therefore, either $\overline{q_1u_1}$ intersects $\overline{q_2v_2}$ in their interiors or $\overline{q_1v_1}$ intersects $\overline{q_2u_2}$ in their interiors; without loss of generality, we assume that it is the former case. Let $p$ be the intersection of $\overline{q_1u_1}$ and $\overline{q_2v_2}$ (e.g., see Fig.~\ref{fig:intersectionbay}). Since $q_1\in B(\alpha,\alpha')$, points of $\overline{q_1u_1}$ other than $q_1$ have only one predecessor, which is $\alpha'$. As $p\in \overline{q_1u_1}$ and $p\neq q_1$, $p$ has only one predecessor $\alpha'$. Similarly, since $q_2\in B(\alpha,\alpha')$ and $p\in \overline{q_2v_2}$, $\alpha$ is also $p$'s predecessor. We thus obtain a contradiction.
\end{proof}
After $\alpha^*$ is found, depending on whether $\alpha^*$ is the first or last generator of $W$, there are three cases.
\begin{enumerate}
\item
If $\alpha^*$ is not the first or last generator of $W$, then we split $W$ into two wavefronts, one for $\overline{ca}$ and the other for $\overline{ad}$. To do so, we first split the binary tree $T(W)$ that represents the current wavefront $W$ at $\alpha^*$. Then,
we do binary search on $A^*$ to find the tangent point from $a$, where $A^*$ is the underlying chain of $\alpha^*$. We also split $\alpha^*$ into two at the tangent point of $A^*$, i.e., split $A^*$ into two chains that form two generators, one for $\overline{ac}$ and the other for $\overline{ad}$ (e.g., see Fig.~\ref{fig:generatorsplit}). As $A^*$ is represented by an array, splitting $\alpha^*$ can be performed in $O(1)$ time by resetting the end indices of the chains in the array.
This finishes the propagation algorithm in $\triangle acd$. The above splits $W$ into two wavefronts, one for $\overline{ac}$ and the other for $\overline{ad}$; we then propagate the wavefronts through $\overline{ac}$ and $\overline{ad}$ recursively.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=1.6in]{generatorsplit.pdf}
\caption{\footnotesize Splitting the generator $\alpha^*$. Assume that $A^*$ is the convex chain from $p_2$ to $p_3$ with $p_2$ as the initial vertex of the generator. $\overline{ap_1}$ is tangent to $A^*$ at $p_1$. After the split, the chain from $p_3$ to $p_1$ becomes a generator with $p_1$ as the initial vertex and the chain from $p_1$ to $p_2$ becomes another generator with $p_2$ as the initial vertex. }
\label{fig:generatorsplit}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
\item
If $\alpha^*$ is the first generator of $W$, then $\alpha^*$ must be $\alpha_0$, i.e., the leftmost generator of $W(\overline{cd})$. In this case, we do not need to split $W$. But we still need to split the generator $\alpha_0$ at the tangent point $p_0$ of $\alpha_0$ from $a$. To find the tangent point $p_0$, however, this time we do not use binary search as it is possible that we will need to do this for $\Omega(n_b)$ vertices of $bay(\overline{cd})$, which would take $\Omega(n_b\log n_g)$ time in total. Instead, we use the following approach. Recall that the vertex $c=v_0$ connects to $\alpha_0$ by a tangent with tangent point $c_0$ (e.g., see Fig.~\ref{fig:bayextend}), and $c_0$ is an endpoint of $A_0'$. We traverse $A'_0$ from $c_0$ to $b_0$, i.e., the other endpoint of $A'_0$; for each vertex, we check whether it is the tangent from $a$. In this way, $p_0$ can be found in time linear in the number of vertices of $A_0'$ between $c_0$ and the tangent point $p_0$ (e.g., see Fig.~\ref{fig:simplecase}). After that, we still split $\alpha_0$ into two generators at $p_0$; one is the only generator for $\overline{ac}$ and the other becomes the first generator of the wavefront for $\overline{ad}$.
\begin{figure}[t]
\begin{minipage}[t]{\textwidth}
\begin{center}
\includegraphics[height=1.9in]{simplecase.pdf}
\caption{\footnotesize Illustrating the case for propagating the one-generator wavefront to $R(\overline{ac})$. }
\label{fig:simplecase}
\end{center}
\end{minipage}
\vspace{-0.15in}
\end{figure}
For $\overline{ad}$, we propagate its wavefront through $\overline{ad}$ recursively. For $\overline{ac}$, to propagate its wavefront through $\overline{ac}$, since the wavefront has only one generator, we can simply apply the linear time shortest path map algorithm for simple polygons~\cite{ref:GuibasLi87}; we refer to this as the {\em one-generator case}. Indeed, $\overline{ac}$ partitions $bay(\overline{cd})$ into two sub-polygons and let $R(\overline{ac})$ denote the one that does not contain $\triangle cda$ (e.g., see Fig.~\ref{fig:simplecase}). Hence, all points of $R(\overline{ac})$ are claimed by the only generator for $\overline{ac}$, whose underlying chain $A'$ is the sub-chain of $A_0'$ from $c_0$ to $p_0$ and whose initial vertex is $p_0$.
Consider the region $R'(\overline{ac})$ bounded by $\overline{cc_0}$, $A'$, $\overline{p_0a}$, and the boundary of $R(\overline{ac})$ excluding $\overline{ac}$. It is a simple polygon with a single weighted source $p_0$. Therefore, our problem is equivalent to computing the shortest path map in $R'(\overline{ac})$ with respect to the source point $p_0$, which can be done in $O(|R(\overline{ac})|+|A'|)$ time~\cite{ref:GuibasLi87}, where $|R(\overline{ac})|$ and $|A'|$ are the numbers of vertices of $R(\overline{ac})$ and $|A'|$, respectively.
\item
If $\alpha^*$ is the last generator of $W$, the algorithm is symmetric to the above second case.
\end{enumerate}
The above describes our algorithm for propagating the wavefront $W(\overline{cd})$ through the first triangle $\overline{cda}$. Next, we discuss the general case where we propagate a wavefront $W$ of more than one generator through an arbitrary triangle of the triangulation of $bay(\overline{cd})$. For the sake of notational convenience, we consider the problem of propagating the wavefront $W(\overline{ad})$ at $\overline{ad}$ through $\overline{ad}$ into the region $R(\overline{ad})$, where $R(\overline{ad})$ is the one of the two sub-polygons of $bay(\overline{cd})$ partitioned by $\overline{ad}$ that does not contain $\triangle cda$. Let $\triangle adb$ be the triangle in $R(\overline{ad})$ with $\overline{ad}$ with as an edge, i.e., $b$ is the third vertex of the triangle. We describe the algorithm for propagating $W(\overline{ad})$ through $\triangle adb$. The algorithm is actually quite similar as before, with an additional {\em event-validation} step.
We first sort all bisector events in $\triangle adb$, following their perpendicular distances to the supporting line of $\overline{ad}$. Then we process these events as before. One difference is that we now need to check whether each event is still valid. Specifically, for each event $q$, which is associated with three generators $\alpha$, $\alpha'$, and $\alpha''$, i.e., $q$ is the intersection of the bisectors $B(\alpha,\alpha')$ and $B(\alpha',\alpha'')$, we check whether all three generators are still in the current wavefront $W$ \footnote{Note that the order of the generators in $W$ must be consistent with their initial index order, $\alpha_0,\alpha_1,\ldots,\alpha_m$. Hence, checking whether a generator is in $W$ can be easily done in $O(\log m)$ time by using the generator indices.}. If not (this is possible if one of the three generators was deleted before, e.g., when the wavefront was propagated through $\triangle cda$), then $q$ is not valid and we ignore this event; otherwise $q$ is still valid and we process it in the same way as before.
The above algorithm is based on the assumption that $\overline{ad}$ is a triangulation diagonal. If it is an obstacle edge, then the wavefront $W(\overline{ad})$ stops at $\overline{ad}$. Notice that each bisector of the wavefront $W(\overline{ad})$ must intersect $\overline{ad}$. For each bisector of the wavefront, starting from its current tracing-point, we trace it out until the current traced hyperbolic-arc intersects $\overline{ad}$.
The algorithm stops once all triangles are processed as above.
\paragraph{Time analysis.}
We now analyze the time complexity. As discussed before, the initial step for computing the intersections of adjacent bisectors of the wavefront $W(\overline{cd})$ and locating the triangles containing them together takes $O(m\log (n_g+n_b))$ time.
During the entire algorithm, each traced bisector hyperbolic-arc belongs to the shortest path map in $bay(\overline{cd})$, i.e., the portion of $\mbox{$S\!P\!M$}(s)$ in $bay(\overline{cd})$, whose size is $O(n_b+n_g+m)$.
Hence, the total time on the bisector tracing in the entire algorithm is $O(n_b+n_g+m)$.
For the one-generator case where only one generator is available for $\overline{ac}$, the time for processing the sub-polygon $R(\overline{ac})$ is $O(|R(\overline{ac})|+|A'|)$. Notice that all such sub-polygons $R(\overline{ac})$ in the one-generator case are interior disjoint. Hence, the total sum of their sizes is $O(n_b)$. Also, all such generator underlying chains $A'$ are also interior disjoint, and thus the total sum of their sizes is $O(n_g)$. Therefore, the overall time for processing the one-generator case sub-polygons is $O(n_g+n_b)$.
For the general case of processing a triangle $\triangle$ of the triangulation, the total time for processing all events is $O(m''\log (n_g+n_b))$\footnote{More specifically, sorting all events takes $O(m''\log m'')$ time. For each event, removing a generator from $W$ takes $O(\log m)$, finding the intersections of new adjacent bisectors takes $O(\log n_g)$ time, and then locating the triangles containing the intersections takes $O(\log n_b)$ time. Note that $m''\leq m\leq n_g$.}, where $m''$ is the number of events in $\triangle$, both valid and invalid.
Each valid or invalid event is computed either in the initial step or after a generator is deleted. The total number of bisector events in the former case in the entire algorithm is at most $m-1$. The total number in the latter case in the entire algorithm is no more than the number of generators that are deleted in the algorithm, which is at most $m$ because once a generator is deleted it will never appear in any wavefront again. Hence, the total time for processing events in the entire algorithm is $O(m\log (n_g+n_b))$.
Once all events in $\triangle$ are processed, we need to find the generator $\alpha^*$ of the current wavefront $W$ that claims the third vertex $b$ of the triangle, by binary search plus bisector tracing.
The time is $O(\log m)$ plus the time for tracing the bisector hyperbolic-arcs. Hence, excluding the time for tracing the bisector hyperbolic-arcs, which has been counted above, the total time for this operation in the entire algorithm is $O(m'\log m)$, where $m'$ is the number of triangles the algorithm processed for the case where $\alpha^*$ is not the first or last generator. We will show later in Lemma~\ref{lem:mprime} that $m'=O(m)$.
After $\alpha^*$ is found, depending on whether $\alpha^*$ is the first or last generator of $W$, there are three cases. If $\alpha^*$ is not the first or last generator of $W$, then we find the tangent from $b$ to $\alpha^*$ by binary search in $O(\log n_g)$ time and split both $W$ and $\alpha$; otherwise, we find the tangent by a linear scan on $\alpha^*$ and only need to split $\alpha^*$. Splitting $W$ takes $O(\log m)$ time while splitting a generator only takes $O(1)$ time as discussed before. Therefore, the total time for splitting generators in the entire algorithm is $O(n_b)$, as there are $O(n_b)$ triangles in the triangulation.
If $\alpha^*$ is either the first or last generator of $W$, then a one-generator case subproblem will be produced and the time of the linear scan for finding the tangent is $O(|A'|)$, where $A'$ is the sub-chain of $\alpha^*$ that belongs to the one-generator case subproblem. As discussed above, all such $A'$ in all one-generator case subproblems are interior disjoint, and thus the total time on the linear scan in the entire algorithm is $O(n_g)$. Therefore, the total time for finding the tangent point and splitting $W$ is $O(m'\log n_g+n_b)$ as $m\leq n_g$.
\begin{lemma}\label{lem:mprime}
$m'\leq m-1$.
\end{lemma}
\begin{proof}
Initially $W=W(\overline{cd})$, which consists of $m$ generators. Each
split operation splits a
wavefront $W$ at a generator $\alpha^*$ into two wavefronts each of which has at least
two generators (such that both wavefronts contain $\alpha^*$). More
specifically, if a wavefront of size $k$ is split, then one wavefront
has $k'$ generators and the other has $k-k'+1$ generators with $k'\geq 2$ and $k-k'+1\geq 2$. The value $m'$ is equal to the number of all
split operations in the algorithm.
We use a tree structure $T$ to
characterize the split operations. The root of $T$ corresponds to
the initial generator sequence $W(\overline{cd})$. Each split on a
wavefront $W$ corresponds to an internal node of $T$ with two children
corresponding to the two subsequences of $W$ after the split. Hence, $m'$ is equal to the number of internal nodes of $T$. In the worst case $T$ has $m$ leaves, each
corresponding to two generators $\alpha_i$ and $\alpha_{i+1}$ for
$i=0,1,\ldots,m$. Since each internal node of $T$ has two children and
$T$ has at most $m$ leaves, the number of internal nodes of $T$ is at
most $m-1$. Therefore, $m'\leq m-1$.
\end{proof}
With the preceding lemma, the total time of the algorithm is
$O(m\log (n_g+n_b)+n_g+n_b)$, which is $O(m\log m+n_g+n_b)$ by similar analysis as Observation~\ref{obser:hlogh}. The space is
$O(n_g+n_b+m)$ as no persistent data structures are used.
\subsection{Expanding $\mbox{$S\!P\!M$}'(\mathcal{M})$ into all canals}
\label{sec:canal}
Consider a canal $canal(x,y)$ with two gates $\overline{xd}$ and $\overline{yz}$. The goal is to expand the map $\mbox{$S\!P\!M$}'(\mathcal{M})$ into $canal(x,y)$ through the two gates to obtain the shortest path map in the canal, denoted by $\mbox{$S\!P\!M$}(canal(x,y))$.
The high-level scheme of the algorithm is similar in spirit to that for the $L_1$ problem~\cite{ref:ChenCo19}.
The algorithm has three main steps. First, we
expand $\mbox{$S\!P\!M$}'(\mathcal{M})$ into $canal(x,y)$ through the gate $\overline{xd}$, by applying our algorithm for bays. Let $\mbox{$S\!P\!M$}_1(canal(x,y))$
denote the map of $canal(x,y)$ obtained by the algorithm. Second,
we expand $\mbox{$S\!P\!M$}(\mathcal{M})$ into $canal(x,y)$ through the gate
$\overline{yz}$ by a similar algorithm as above; let $\mbox{$S\!P\!M$}_2(canal(x,y))$ denote the map of $canal(x,y)$
obtained by the algorithm. Third, we merge the two maps
$\mbox{$S\!P\!M$}_1(canal(x,y))$ and $\mbox{$S\!P\!M$}_2(canal(x,y))$ to obtain $\mbox{$S\!P\!M$}(canal(x,y))$. This
is done by using the merge step from the standard divide-and-conquer
Voronoi diagram algorithm to compute the region closer to the
generators at $\overline{xd}$ than those at $\overline{yz}$.
We will provide more details for this step below and we will show that this step can be done in time linear in the total size of the two maps $\mbox{$S\!P\!M$}_1(canal(x,y))$ and $\mbox{$S\!P\!M$}_2(canal(x,y))$. Before doing so, we analyze the complexities of the algorithm.
The first step takes $O(n+h\log
h)$ time for all canals. So is the second step. The third step takes
linear time in the total size of $\mbox{$S\!P\!M$}_1(canal(x,y))$ and
$\mbox{$S\!P\!M$}_2(canal(x,y))$. Since the total size of the two maps over all canals
is $O(n)$, the total time of the third step for all canals is $O(n)$.
In summary, the time for computing the shortest path maps in all
canals is $O(n+h\log h)$ and the space is $O(n)$.
In the following, we provide more details for the third step of the algorithm.
Recall that $x$ and $y$ are the two endpoints of the corridor path $\pi$ in $canal(x,y)$. It is possible that the shortest $s$-$x$ path $\pi(s,x)$ contains $y$ or the shortest $s$-$y$ path $\pi(s,y)$ contains $x$. To determine that, we can simply check whether $d(s,x)+|\pi|=d(s,y)$ and whether $d(s,y)+|\pi|=d(s,x)$. Note that both $d(s,x)$ and $d(s,y)$ are available once $\mbox{$S\!P\!M$}(\mathcal{M})$ is computed.
We first consider the case where neither $\pi(s,x)$ contains $y$ nor $\pi(s,y)$ contains $x$.
In this case, there must be a point $p^*$ in $\pi$ such that $d(s,x)+|\pi(x,p^*)|=d(s,y)+|\pi(y,p^*)|$, where $\pi(x,p^*)$ (resp., $\pi(y,p^*)$) is the subpath of $\pi$ between $x$ (resp., $y$) and $p^*$. We can easily find $p^*$ in $O(|\pi|)$ time. To merge the two maps $\mbox{$S\!P\!M$}_1(canal(x,y))$ and $\mbox{$S\!P\!M$}_2(canal(x,y))$ to obtain $\mbox{$S\!P\!M$}(canal(x,y))$, we find a dividing curve $\Gamma$ in $canal(x,y)$ such that $W(\overline{xd})$ claims all points of $canal(x,y)$ on one side of $\Gamma$ while $W(\overline{yz})$ claims all points of $canal(x,y)$ on the other side of $\Gamma$, where $W(\overline{xd})$ is the set of generators of $\mbox{$S\!P\!M$}'(\mathcal{M})$ claiming $\overline{xd}$ (one may consider $W(\overline{xd})$ is a wavefront) and $W(\overline{yz})$ is the set of generators of $\mbox{$S\!P\!M$}'(\mathcal{M})$ claiming $\overline{yz}$. The curve $\Gamma$ consists of all points in $canal(x,y)$ that have equal weighted distances to $W(\overline{xd})$ and $W(\overline{yz})$. Therefore, the point $p^*$ must be on $\Gamma$. Starting from $p^*$, we can trace $\Gamma$ out by walking simultaneously in the cells of the two maps $\mbox{$S\!P\!M$}_1(canal(x,y))$ and $\mbox{$S\!P\!M$}_2(canal(x,y))$.
The running time is thus linear in the total size of the two maps.
We then consider the case where either $\pi(s,x)$ contains $y$ or $\pi(s,y)$ contains $x$. Without loss of generality, we assume the latter case. We first check whether $\pi(s,p)$ contains $x$ for all points $p$ on the gate $\overline{yz}$. To do so, according to the definitions of corridor paths and gates of canals, for any point $p\in \overline{yz}$, its shortest path to $x$ in $canal(x,y)$ is the concatenation of the corridor path $\pi$ and $\overline{yp}$.
Hence, it suffices to check whether $d(s,z)=d(s,y)+|\overline{yz}|$.
\begin{itemize}
\item
If yes, then all points of $\overline{yz}$ are claimed by the generators of $W(\overline{xd})$ (in fact, they are claimed by $x$ because for any point $q\in \overline{xd}$ and any point $p\in \overline{yz}$, their shortest path in $canal(x,y)$ is the concatenation of $\overline{qx}$, $\pi$, and $\overline{yp}$). Hence, all points in $canal(x,y)$ are claimed by $W(\overline{xd})$ and $\mbox{$S\!P\!M$}(canal(x,y))$ is $\mbox{$S\!P\!M$}_1(canal(x,y))$.
\item
Otherwise, some points of $\overline{yz}$ are claimed by $W(\overline{xd})$ while others are claimed by $W(\overline{yz})$. As in the above case, we need to find a dividing curve $\Gamma$ consisting of all points with equal weighted distances to $W(\overline{xd})$ and $W(\overline{yz})$. To this end, we again first find a point $p^*$ in $\Gamma$. For this, since $\pi(s,y)$ contains $x$, $y$ is claimed by $W(\overline{xd})$. On the other hand, since $d(s,z)\neq d(s,y)+|\overline{yz}|$, $z$ is claimed by $W(\overline{yz})$. Therefore, $\overline{yz}$ must contains a point $p^*\in \Gamma$. Such a point $p^*$ can be found by traversing $\overline{yz}$ simultaneously in the cells of both $\mbox{$S\!P\!M$}_1(canal(x,y))$ and $\mbox{$S\!P\!M$}_2(canal(x,y))$. After $p^*$ is found, we can again trace $\Gamma$ out by walking simultaneously in the cells of $\mbox{$S\!P\!M$}_1(canal(x,y))$ and $\mbox{$S\!P\!M$}_2(canal(x,y))$. The running time is also linear in the total size of the two maps.
\end{itemize}
\subsection{Wrapping things up}
\begin{theorem}
Suppose $\mathcal{P}$ is a set of $h$ pairwise disjoint polygonal obstacles with a total of $n$ vertices in the plane and $s$ is a source point. Assume that a triangulation of the free space is given. The shortest path map $\mbox{$S\!P\!M$}(s)$ with respect to $s$ can be constructed in $O(n+h\log h)$ time and $O(n)$ space.
\end{theorem}
\begin{proof}
Using the triangulation, we decompose the free space $\mathcal{F}$ into an ocean $\mathcal{M}$, canals, and bays in $O(n)$ time~\cite{ref:KapoorAn97}. Then, the shortest path map $\mbox{$S\!P\!M$}(\mathcal{M})$ in the ocean $\mathcal{M}$ can be constructed in $O(n+h\log h)$ time and $O(n)$ space. Next, $\mbox{$S\!P\!M$}(\mathcal{M})$ can be expanded into all bays and canals in additional $O(n+h\log h)$ time and $O(n)$ space. The shortest path map $\mbox{$S\!P\!M$}(s)$ is thus obtained.
\end{proof}
The current best algorithms can compute a triangulation of the free space in $O(n\log n)$ time or in $O(n+h\log^{1+\epsilon}h)$ time for any small $\epsilon>0$~\cite{ref:Bar-YehudaTr94}. If all obstacles of $\mathcal{P}$ are convex, then the triangulation can be done in $O(n+h\log h)$ time~\cite{ref:HertelFa85}.
After $\mbox{$S\!P\!M$}(s)$ is computed, by building a point location data structure~\cite{ref:EdelsbrunnerOp86,ref:KirkpatrickOp83} on $\mbox{$S\!P\!M$}(s)$ in additional $O(n)$ time, given a query point $t$, the shortest path length from $s$ to $t$ can be computed in $O(\log n)$ time and a shortest $s$-$t$\ path can be produced in time linear in the number of edges of the path.
\begin{corollary}
Suppose $\mathcal{P}$ is a set of $h$ pairwise disjoint polygonal obstacles with a total of $n$ vertices in the plane and $s$ is a source point. Assume that a triangulation of the free space is given. A data structure of $O(n)$ space can be constructed in $O(n+h\log h)$ time and $O(n)$ space, so that given any query point $t$, the shortest path length from $s$ to $t$ can be computed in $O(\log n)$ time and a shortest $s$-$t$\ path can be produced in time linear in the number of edges of the path.
\end{corollary}
\footnotesize
\bibliographystyle{plain}
|
\section{Introduction and Summary}
Nambu-Goldstone theorems are some of our strongest non-perturbative constraints on the dynamics of Quantum Field Theories (QFTs).
Let us review the setup for continuous internal (non space-time) symmetries in a QFT in $d$ space-time dimensions.
By Noether's theorem, there exists a charge $Q$ which is an extended operator depending topologically on a co-dimension 1 surface $\Sigma$. If in some state $|\Omega\rangle$ we have for some local operator $O(0)$
$$\langle\Omega|[Q,O(0)]|\Omega\rangle\neq0~,$$
then it can be shown by a deformation of $\Sigma$ that, roughly speaking, $\langle\Omega|j_0 (x) O(0)|\Omega\rangle$
cannot decay faster than $1/|x|^{d-1}$ as we take $|x|\to \infty$. This algebraic decay of a correlation function implies gapless excitations of $\ket{\Omega}$ in infinite volume. Furthermore, under various additional assumptions about the nature of the state $\ket{\Omega}$, the existence of an ordinary massless boson (or a superfluid mode) can be established. Incidentally, if $\langle\Omega|j_0 (x) O(0)|\Omega\rangle\neq0$ holds true, this means that also $\langle\Omega|O^\dagger (x) O(0)|\Omega\rangle$ must decay algebraically at most. And under some assumptions it cannot decay faster than $1/|x|^{d-2}$.\footnote{The argument for this invokes inserting a complete set of states and recalling that the matrix elements of $j_\mu$ are suppressed by a factor of momentum at small momentum. It would be nice to understand to what extent this argument about the matrix elements is general. \label{fn1}} This is unacceptable in $d=2$ since it leads to a violation of clustering, due to the connected correlator not decaying. Hence no such states $|\Omega\rangle$ can exist in $d=2$, which is the familiar statement of the Coleman-Mermin-Wagner theorem~\cite{Mermin:1966fe,Coleman:1973ci}.
In finite volume (in the absence of boundaries) no state $ |\Omega\rangle$ can have the property that
$\langle\Omega|[Q,O(0)]|\Omega\rangle\neq0$. This is simply because we can diagonalize $Q$ in finite volume. Thus, the phenomenon of symmetry breaking is really due to the infinite volume limit. When symmetry breaking occurs, the Hilbert space of the finite volume theory becomes closer and closer to a direct sum of sub-Hilbert spaces which do not communicate via the action of {\it local} operators. When we take the infinite volume limit we only keep one of these sub-Hilbert spaces. But these sub-Hilbert spaces can still communicate by the action of extended operators such as $Q$. This is why symmetry breaking may occur in infinite volume and this is why the notion of superselection sectors exists. These comments will be important below.
The situation for space-time symmetries is, in principle, similar. There are however some interesting differences. Given the energy-momentum (EM) tensor of the theory, $T_{\mu\nu}$, we can construct the space-time symmetries of the infinite volume theory living in $\mathbb{R}^{d-1,1}$ from the conserved currents
$${\mathcal{J}}^{(\xi)}_\mu=\xi^\nu T_{\mu\nu}~,$$
where $\xi$ is a Killing vector satisfying as usual $\partial^{(\nu}\xi^{\mu)}=0$. The currents ${\mathcal{J}}^{(\xi)}_\mu$ of course lead to the usual translations, rotations, and boosts. We denote these charges by $Q^{(\xi)}$.
There can exist a state $|\Omega\rangle$ and a local operators $O(0)$ (which may or may not have spin indices, which we suppress for now) such that
\begin{equation}\label{Comm}\langle\Omega|[Q^{(\xi)},O(0)]\Omega\rangle\neq0~.\end{equation}
For constant $\xi$, namely the translation symmetry, this can be easily achieved in any state which is not translationally invariant. Similarly, for rotations the commutator would be generally nonzero in a non-isotropic state for any operator $O$ with spin indices.
A general discussion of various allowed symmetry breaking patterns can be found in~\cite{Nicolis:2015sra}. A rather common situation is the spontaneous breaking of boost symmetry in states $|\Omega\rangle$ which are homogeneous and isotropic. This is one of the focal points of our paper.\footnote{There are important consequences of the spontaneous breaking of boost symmetry also in states which break the spatial translation symmetry. See for example~\cite{Dubovsky:2012sh,Aharony:2013ipa}. Here our focus is on translationally invariant, isotropic states.}
From dimensional analysis, a nonzero commutator for the boost Killing vector~\eqref{Comm} (which we can take to be $\xi^1=x^0$, $\xi^0=x^1$, with the rest of the components vanishing) means that correlators of some components of the EM tensor and $O$ decay not faster than $1/|x|^{d}$. This does not lead to any particular problems in $d=2$ and hence there is no obstruction for the spontaneous breaking of boost symmetry in two space-time dimensions. We will indeed see some examples later.
The boost symmetry differs conceptually from ordinary (non space-time) symmetries in that it cannot be preserved at finite volume. Indeed, if we compactify space while keeping time intact, the symmetry between space and time is manifestly destroyed. Therefore, the spontaneous breaking of boost symmetry does not necessarily mean that super-selection sectors must arise!
Thus, the boost symmetry Nambu-Goldstone theorem does lead to an algebraic decay and hence a gapless spectrum of excitations of $|\Omega\rangle$, but it does not imply super-selection sectors since there is no sense in diagonalizing the boost symmetry in compact space.
As we have argued above, the boost symmetry Nambu-Goldstone theorem leads to a somewhat faster decay of correlation functions compared to the case of spontaneous breaking of ordinary global symmetries. Therefore, one-particle excitations are not necessary, but rather, composite massless excitations could play a role. This was realized in the beautiful recent paper \cite{Alberte:2020eil}.
While there is no boost symmetry in compact space, it is still true that if we focus our attention on a small enough patch of our compact space, there is an {\it approximate} boost symmetry. Depending on the state of the system, this approximate boost symmetry may appear to be spontaneously broken. The algebraic decay and gapless excitations in the flat space limit therefore entail some constraints on the spectrum of the finite volume theory. The nature of these constraints can be understood on dimensional grounds. Take space to be a hypercube $V=L^{d-1}$ where $L$ is the length and $V$ is the volume. Then to see a state with finite energy density $\epsilon$ in a small patch of the hypercube we have to start from a state with total energy $E= \epsilon V$. We assume that on distances $\epsilon^{-1/d}\ll \Delta x \ll L$ we see an algebraic decay. Indeed, $\epsilon^{-1/d}$ is a length scale in the infinite volume theory and since we see an algebraic decay in infinite volume it must be true that there is an algebraic decay at finite volume in the range
$\epsilon^{-1/d}\ll \Delta x \ll L$.
An algebraic decay in the range $\epsilon^{-1/d}\ll \Delta x \ll L$ is not to be taken for granted. It means that the gap above our state, $E_\text{gap}$, must be much smaller than $\epsilon^{1/d}$. So the energy of the excitations of $|\Omega\rangle$ should scale as $\epsilon V+ E_\text{gap}$ with $ E_\text{gap}\ll \epsilon^{1/d}$. Clearly, for the consistency of an algebraic decay, the gap has to be arbitrarily smaller than $\epsilon^{1/d}$. This means that the gap has to go to zero as $L\to\infty$, in agreement with the gapless nature of the excitations in infinite volume. We can roughly speaking say that therefore
$$E_\text{gap}\sim L^{-\gamma}~,$$
with $\gamma>0$. We can think of $\gamma$ as a certain critical exponent that measures how fast the gap closes in finite space as we increase the volume.
Of course, the most natural choice is $\gamma=1$ but this does not follow from our general considerations. That $\gamma=1$ is natural can be motivated based on Wilsonian considerations. Let us take space to be infinite. The deep infrared theory should be a fixed point of the renormalization group (perhaps in the Lifshitz sense) and hence there is a scale-free effective field theory description of the gapless degrees of freedom that give rise to the algebraic correlators at infinite volume. In that case, $\gamma=1$ follows from dimensional analysis. The above argument assumes that the long
correlators are captured by some scale-free infrared theory where the energy density is a cutoff scale. In practice, it could be that the gap is even smaller than what is predicted by dimensional analysis if the low-energy modes have additional degeneracy (which in finite volume is only approximate). Indeed, consider states that are generic in the Eigenstate Thermalization Hypothesis (ETH) sense (see~\cite{deutsch2018eigenstate} for a recent review and references). Those states have a macroscopic entropy and hence an exponentially small gap.
Our bound on the gap holds in all theories and appropriate states, regardless of whether the states are generic.
Furthermore, one can already at this stage say a few words about the density of these low-lying states. For instance, if there was just one such state with energy above $|\Omega\rangle$ scaling like $L^{-\gamma}$ then the correlator in the range $\epsilon^{-1/d}\ll\Delta x\ll L$ would have behaved non-algebraically.
The same argument holds for any finite collection of states. We therefore need {\it infinitely} many states that become gapless as $L\to \infty$.
The breaking of the boost symmetry is not a rare phenomenon. As we will see, in a unitary theory, any state that in the infinite volume limit has a non-vanishing energy density will break the boost symmetry. Therefore, the above constraints on the spectrum hold for quite generic states in finite volume.
Our purpose here is to apply these ideas to Conformal Field Theories (CFTs). CFTs can be studied on the cylinder $$S^{d-1}\times \mathbb{R}~,$$ where the radius of the sphere is $R$. The energy spectrum is related to the spectrum of scaling dimensions
\begin{equation}\label{cylinder} E={\Delta\over R}~.\end{equation}
We consider states with a nontrivial macroscopic limit, i.e. states with nonzero energy density. This means that we must take $\Delta_\Omega= s_{d-1}\epsilon \,R^{d}$ with $\epsilon$ the energy density and $ s_{d-1}$ the volume of the unit $S^{d-1}$. So as we take the macroscopic limit, we are discussing states that correspond to operators with a large scaling dimension.
From our general considerations above we found that there are infinitely many states with energy going to zero as $R^{-\gamma}$. This means that there are infinitely many states with energy below
\begin{equation}\label{scalingIntro}
E= s_{d-1}\epsilon \,R^{d-1}+c\, \epsilon^{(1-\gamma)/d}R^{-\gamma}~,
\end{equation}
where $c$ is some dimensionless constant and the power of $\epsilon$ in the second term is adjusted so that the result makes sense dimensionally.
This translates to having operators with scaling dimensions
\begin{equation}\label{scalingi} \Delta=\Delta_\Om +c\,\Delta_\Om^{1-\gamma\over d}~. \end{equation}
This means that the gap around heavy operators with scaling dimension $\Delta_\Om$ is at most $\Delta_\Om^{1-\gamma\over d}$ with $\gamma>0$.
As we have explained above, with some additional physical input it follows that $\gamma=1$ (or larger) and hence the gap around heavy operators scales like $O(1)$.
This bound on the scaling dimension gap around heavy operators should hold very generally and it does not require the genericity or typicality that is assumed in ETH. (Several of the applications we will study here are in fact concerned with large charge ground states, which are very atypical states.)
In CFTs on the cylinder~\eqref{cylinder} the gap constraint~\eqref{scalingi} may appear trivial given that there are descendant states. (For a review of CFTs on the cylinder see for instance~\cite{Simmons-Duffin:2016gjk}.) Indeed, for any primary state $|\Omega\rangle$ there is a family of states $|\partial^N \Omega\rangle$ (where the index contractions in the derivatives are not explicitly displayed) with scaling dimension $\Delta_\Om+N$ and hence energy $E=\Delta_\Om/R+N/R$ for any non-negative integer $N$.
There is therefore an interesting twist in the story: we can ask if the descendant states are those responsible for the low-energy theorems. The question of whether the descendant states are those responsible for the low-energy theorems is the essential new question addressed in this paper.
We will see that, perhaps surprisingly, the answer is negative. One must have new {\it primary} states with dimension~\eqref{scalingi}. Furthermore, in all examples we study, $\gamma=1$ for these new primary states, in agreement with the general arguments. The boost symmetry Nambu-Goldstone theorem is not sufficient to completely determine the low lying spectrum of primary excitations of $|\Omega\rangle$ and indeed the structure of excitations differs in different examples we study.
A situation where the existence of new low-lying primaries as above is a nontrivial constraint arises in the study of ground states at fixed, large charge $Q$ for some $U(1)$ symmetry. From general considerations (which however do not apply in mean field theory) one expects~\cite{Hellerman:2015nra} (for a review and more references see~\cite{Gaume:2020bmp})
$$\Delta\sim Q^{d\over d-1}~.$$
These ground states are very non-generic heavy states and the constraints arising from the spontaneous breaking of boost symmetry are satisfied quite differently in different examples we study.
Sometimes the excitations needed for the boost Nambu-Goldstone theorem should be considered as a Regge trajectory of (primary) excitations of $|\Omega\rangle$ -- the Regge trajectories appearing in these cases are reminiscent of~\cite{Jafferis:2017zna}. This occurs in the superfluid case and 2d, where these are one-particle states, and in mean field theory, where the excitations are to be thought of as a particle and a (zero momentum) hole in the Bose-Einstein condensate. Sometimes we find that the primary excitations are two-particle states, as in the free Fermi surface.
The case of 2d CFTs provides an interesting testing ground for our considerations. As briefly alluded to above, spontaneous breaking of boost symmetry is possible in 2d. The states responsible for the algebraic decay of correlators are the Virasoro (but not global conformal) descendants of the heavy state. Hence, we do not find any constraint on the gap above a heavy operator. Nevertheless, the relevant Virasoro descendants form a Regge trajectory. We also examine the fate of the large charge effective field theory of~\cite{Hellerman:2015nra} in 2d. We find that for compact CFTs, it just becomes the free compact scalar representation of the $u(1)\times u(1)$ Ka\v{c}-Moody algebra. And the theory only makes sense for $c=1$. For situations where the $U(1)$ symmetry does not get enhanced to a current algebra, the effective theory is nontrivial and it describes a single compact boson with arbitrary conformal anomaly. It resembles the effective string theory of Polchinski-Strominger~\cite{Polchinski:1991ax}.
To put the present paper in context, let us point out that there has been a monumental effort in recent years to understand the spectrum of scaling dimensions in CFTs. The numerical bootstrap constraints on the low scaling dimension operators are beautifully reviewed in~\cite{Simmons-Duffin:2016gjk} and~\cite{Chester:2019wfx}, among others. There is then a complementary effort to understand the scaling dimensions of various special heavy operators, e.g. the ground states at large fixed charge, as reviewed in~\cite{Gaume:2020bmp}, large fixed spin~\cite{Komargodski:2012ek,Fitzpatrick:2012yx}, and various combinations of large charge and large spin~\cite{Cuomo:2017vzg,Cuomo:2019ejv}. The present paper is aimed at understanding the general constraints from the spontaneous breaking of boost symmetry. This is relevant for the study of heavy operators quite generally. But except for some brief discussion of the hydrodynamic regime, our focus here will be exclusively on the large charge ground states.
The outline of the paper is as follows. The paper consists of two parts. In Part~\ref{part1} we present an abstract argument for the bound on the gap in the operator spectrum of CFTs around heavy operators. In section~\ref{sec:NGthm} we write the precise consequences of the spontaneous breaking of boost symmetry (reviewing and very slightly extending~\cite{Alberte:2020eil}) and dilatation symmetry. This results in some constraints on the low momentum, low frequency behavior of Energy-Momentum correlation functions. We show how these constraints are satisfied in states that obey the assumptions of hydrodynamics: the ballistic sound mode in hydrodynamics exists essentially because of the spontaneous breaking of boost symmetry.
In section~\ref{sec:CFT}, we discuss in more detail CFTs on the cylinder and review how the operator-state correspondence can be used to relate matrix elements on the cylinder and the CFT data. We explain some important properties of conformal blocks in heavy states and conclude that descendants cannot play a role in the Nambu-Goldstone theorem for boosts.
Making the additional assumption that heavy state $\ket{\Om}$ is the ground state of some sector of the theory, we present an argument that the Nambu-Goldstone theorem implies a bound on the gap in the CFT operator spectrum.
In Part~\ref{part2} we examine a variety of examples, show that our bound is obeyed and identify special sets of operators that saturate the Nambu-Goldstone theorem. In section~\ref{sec:superfluid} we discuss the case of the superfluid. We identify in detail the primary states that are responsible for the Nambu-Goldstone theorem.
In section~\ref{sec:FreeScalar} we repeat the exercise for mean field theory. In section~\ref{sec:FreeFermions} we repeat it for a free Fermi surface. In section~\ref{sec:boostbreak} we show that in $d=2$ the Nambu-Goldstone sum rule is satisfied by a Regge trajectory of one-particle-like states in the Verma module. We also discuss the role of the large charge effective field theory in 2d.
Three appendices contain technical details about Green's functions in nontrivial states, an extension of the free fermion discussion to $d=4$, and, finally, a derivation of some contact terms in correlation functions of the EM tensor.
\part{Bounding the Gap in the Operator Spectrum }\label{part1}
\section{The Boost Nambu-Goldstone Theorem}\label{sec:NGthm}
While Nambu-Goldstone theorems for internal symmetries are thoroughly understood, the case of spacetime symmetries provides us with new lessons to this day. One difference explained in the introduction, is that breaking of the boost symmetry does not lead to superselection sectors. Another difference is that while it is impossible to break continuous internal symmetries in $d=2$, boosts can be spontaneously broken in $d=2$. Finally, as mentioned above, in \cite{Alberte:2020eil} it was shown that the boost Nambu-Goldstone theorem can be saturated by multiparticle states; this is a novel phenomenon.\footnote{ See \cite{Rothstein:2017twg,Rothstein:2017niq} for a different, but equivalent discussion of the role of multiparticle states in closing the symmetry algebra of the theory. Ultimately, Landau's derivation of the relation between couplings and the Fermi velocity in Fermi liquid theory contains the same insights.}${}^,$\footnote{It would be nice to understand whether this can happen for internal symmetries away from the vacuum state -- see footnote \ref{fn1}.}
In this section we review the formulation of the boost Nambu-Goldstone theorem in translationally invariant energy eigenstates as a sum rule as discussed in \cite{Alberte:2020eil}, and extend it
to dilatations in CFTs. We demonstrate how these sum rules are obeyed in generic ETH states, i.e.~by hydrodynamics. Throughout the paper we use mostly minus signature, $g^{\mu\nu}={\rm diag}(+1,-1,-1,\dots)$.
The logic of the Nambu-Goldstone theorem requires us finding an order parameter for which $\bra{\Om}\de_{K^i}{\cal O}(t,x)\ket{\Om}\neq 0$, where $K^i$ is a boost in the $i$th direction. Let us choose ${\cal O}=T^{0j}$, for which we have
\es{SumRule}{
\bra{\Om}\de_{K^i}T^{0j}(t,x)\ket{\Om}&=i\bra{\Om}[K^i,T^{0j}(t,x)]\ket{\Om}\\
&=\bra{\Om}T^{00}(t,x)\de^{ij}+T^{ij}(t,x)\ket{\Om}\\
&=\le(\ep+P\ri)\de^{ij}\,.
}
$\epsilon$ is the energy density and $P$ is the pressure.
We can rewrite it as a sum rule obeyed by the following correlator
\es{SumRuleB}{
\le(\ep+P\ri)\de^{ij}&=i\bra{\Om}[K^i,T^{0j}(0)]\ket{\Om}\\
&=-i\int d^{d-1}x \ x^i\, \expval{[T^{00}(t,x),T^{0j}(0)]}\,,
}
which we conveniently rewrite in momentum space as (see appendix \ref{app:GreensFunctions} for our notation and definition of the Green's functions)\footnote{The Fourier transform is defined through $$f(\om,k)=\int d^dx \ e^{i(\om t-\vec{k}\cdot\vec{x})}f(t,x)\,.$$}
\es{SumRule2}{
2\pi \de(\om)\le(\ep+P\ri)\de^{j}_i&=\lim_{k\to 0}\,{\p\ov \p k^i} G^{(comm)}_{T^{00},T^{0j}}(\om,k)\,,
}
where $G^{(comm)}_{T^{00},T^{0j}}(\om,k)$ is the commutator Green's function, or by using \eqref{Grels},
\es{SumRule3}{
\de(\om)\le(\ep+P\ri)\de^{j}_i&=\lim_{k\to 0}\,{\p\ov \p k^i} \rho_{T^{00},T^{0j}}(\om,k)\,.
}
($ \rho_{T^{00},T^{0j}}$ stands for the spectral density.)
Our task is then to find states whose contribution to the spectral density saturates the above sum rule. As usual, the Nambu-Goldstone theorem is a constraint on the low-frequency behavior of Green's functions.
\subsection{The Example of Hydrodynamics}
In the standard treatment of hydrodynamics, we obtain the retarded correlators $G^{(R)}_{AB}(\om)$, see e.g.~\cite{Kovtun:2012rj}. We want to extract the spectral density from this data. Using time reflection symmetry to deduce $G^{(R)}_{T^{00},T^{0j}}(\om,k)=G^{(R)}_{T^{0j},T^{00}}(\om,k)$, using \eqref{Grels} we end up with the formula
\es{rhoRet}{
\rho_{T^{00},T^{0j}}(\om,k)={1\ov \pi}{\rm Im}\le[G^{(R)}_{T^{00},T^{0j}}(\om,k)\ri]\,.
}
We then calculate
\es{GRhydro}{
G^{(R)}_{T^{00},T^{0j}}(\om,k)&=\le(\ep+P\ri)\, {\om k^j\ov \om^2-(c_s k)^2+i \ga_s \om k^2}\\
\rho_{T^{00},T^{0j}}(\om,k)&={\le(\ep+P\ri)\ov \pi}\, {\ga_s\om^2 k^2 k^j\ov (\om^2-(c_s k)^2)^2+(\ga_s \om k^2)^2}\,.
}
If we plot this function for fixed $k$, we see two peaks at $\om=\pm c_s k$. As we decrease $k$ they get narrower, but also get closer. To understand this better, let us examine them in a new variable $\hat\om\equiv \om/(c_s k)$, in terms of which, we have
\es{rhodetails}{
\rho_{T^{00},T^{0j}}(\om,k)&={\le(\ep+P\ri)\ov \pi c_s}\, {\tilde k^j\, \hat\om^2 \ov (\hat\om^2-1)^2+\tilde k^2\, \hat\om^2 }\,,\\
\tilde k&\equiv {\ga_s k\ov c_s}\,.
}
Now it is a standard result that for small $\tilde k$
\es{rhodetails2}{
{1\ov \pi}\,{\tilde k^j\, \hat\om^2 \ov (\hat\om^2-1)^2+\tilde k^2\, \hat\om^2 }&\approx {\tilde k^j\hat\om^2\ov 2 \tilde k} \le[\de(\hat\om-1)+\de(\hat\om+1)\ri]\\
&={\tilde k^j\ov 2 \tilde k} \le[\de(\hat\om-1)+\de(\hat\om+1)\ri]\\
&={c_s k^j \ov 2} \le[\de(\om-c_s k)+\de(\om+c_s k)\ri]\\
&\approx {c_s k^j }\, \de(\om)\,.
}
Putting the factors together, we learn that for small $k$
\es{GRhydro3}{
\rho_{T^{00},T^{0j}}(\om,k)&\approx {\le(\ep+P\ri)k^j} \de(\om)\,.
}
Plugging this into the RHS of \eqref{SumRule3}, we verify that the equation indeed holds.
One wonders what states gave us the saturation of the sum rule. Hydrodynamics differs from the other EFTs we study in Part~\ref{part2} in that the sound mode is a collective excitation, and there is no preferred set of states $\ket{\Om'}$ that gives rise to the above spectral density. In harmony with ETH, all states $\ket{\Om'}$ close in energy to $\ket{\Om}$ contribute (see~\cite{Delacretaz:2020nit} for a detailed analysis). Yet, one can say that it is the ballistic pole in hydrodynamics which is responsible for the spontaneously broken boost symmetry.
\subsection{The Energy Density Spectral Function}\label{sec:energySumrule}
One can argue from conservation that $\rho_{T^{00},T^{00}}(\om,k)$ should go to $- {\le(\ep+P\ri)}\, k^2 \de'(\om)$ at very small momentum. One argument goes as follows: we have determined in \eqref{SumRule3} that for $k\ll \om$:
\es{Smallk}{
\rho_{T^{00},T^{0j}}(\om,k)\approx \le(\ep+P\ri)k^j\de(\om)\,,
}
hence the Green's function according to \eqref{GABSpect}
\es{SmallkGreen}{
G_{T^{00},T^{0j}}(\om,k)&=\int_{-\infty}^\infty d\om'\ {\rho_{T^{00},T^{0j}}(\om',k)\ov \om -\om'}\\
&\approx \le(\ep+P\ri){k^j\ov \om }
}
Conservation of the stress tensor gives
\es{cons}{
G_{T^{00},T^{00}}(\om,k)=-{k_j\ov \om}\,G_{T^{00},T^{0j}}(\om,k)\approx \le(\ep+P\ri){k^2\ov \om^2 }\,,
}
which is reproduced by
\es{t00t00}{\rho_{T^{00},T^{00}}(\om,k)\approx - {\le(\ep+P\ri)}\, k^2 \de'(\om)\,.}
Let us see how this comes about in hydrodynamics.
\es{GRhydro2}{
G^{(R)}_{T^{00},T^{00}}(\om,k)&=\le(\ep+P\ri)\, {k^2\ov \om^2-(c_s k)^2+i \ga_s \om k^2}\\
\rho_{T^{00},T^{00}}(\om,k)&={\le(\ep+P\ri)\ov \pi}\, {\ga_s\om k^4\ov (\om^2-(c_s k)^2)^2+(\ga_s \om k^2)^2}\\
&\approx {\le(\ep+P\ri)\ov c_s^2}\, {c_s k \ov 2} \le[\de(\om-c_s k)-\de(\om+c_s k)\ri]\\
&\approx - {\le(\ep+P\ri)}\, k^2 \de'(\om)\,.
}
Note that $- \de'(\om)$ is positive for $\om>0$, as required for the spectral density.
\subsection{Additional Sum Rules in CFTs}
Let us ask what sum rules follow from the breaking of boosts $K^i$, dilatation $D$, and special conformal transformations $S_\mu$ in a homogeneous and isotropic finite energy density CFT state $\ket{\Om}$, which we imagine as the macroscopic limit of some CFT state on $S^{d-1}$ that corresponds to a heavy scalar operator.
The idea is to find an operator for which $\bra{\Om}\de{\cal O}(t,x)\ket{\Om}\neq 0$, where $\de{\cal O}$ is the transformation of the operator under the generator $K^i,\, D,\, S^i$.
Let us first consider $K^i$. We have already worked out the case ${\cal O}=T^{0j}$, and obtained the sum rule \eqref{SumRule3}.
Similarly for ${\cal O}=J^{j}$, by using $\de_{K^i}J^{j}=J^0 \de^{ij}$, we get
\es{SumRuleJ}{
\de(\om)\rho\,\de^{j}_i&=\lim_{k\to 0}\,{\p\ov \p k^i} \rho_{T^{00},J^{j}}(\om,k)\,,
}
where $\rho$ is the charge density.
Next we ask about $D$, for which we have $\de_{D}{\cal O}=\De_{\cal O} {\cal O}$. We then write
\es{SumRuleD}{
\De_{\cal O}\bra{\Om} \sO\ket{ \Om} &=i\expval{[D,{\cal O}(0)]}\\
&=i\int d^{d-1}x \ x^\mu\, \expval{[T_{\mu 0}(t,x),{\cal O}(0)]}\\
&=-i\int d^{d-1}x \ x^i\, \expval{[T^{i 0}(t,x),{\cal O}(0)]}\,.
}
This is then almost identical to \eqref{SumRuleB}, and we get
\es{SumRuleD2}{
&\text{for ${\cal O}=T^{00}$:} \qquad \de(\om)d\ep=\lim_{k\to 0}\,{\p\ov \p k^i} \rho_{T^{i0},T^{00}}(\om,k)\\
&\text{for ${\cal O}=J^{0}$:} \qquad \de(\om)(d-1)\rho=\lim_{k\to 0}\,{\p\ov \p k^i} \rho_{T^{i0},J^0}(\om,k)\,.
}
The right hand side of the first equation here is known from \eqref{SumRule3} to be equal to\\ $ (d-1)(\ep+P)\de(\om)$, so we only learn that $\ep=(d-1)P$, which is true in CFT. The second equation is a new sum rule. We will check it for the superfluid in section~\ref{sec:superfluid}.
Finally we discuss $S_\mu$. Since for a primary ${\cal O}$, $[S_\mu,{\cal O}]=0$ we have to work with descendants. Let us consider the variation
\es{Svar}{
\de_{S_\mu}\le(i[P_\nu,{\cal O}]\ri)&=-[S_\mu,[P_\nu,{\cal O}]]\\
&=[[P_\nu,S_\mu],{\cal O}]\\
&=-2i[\eta_{\mu\nu} D-M_{\mu\nu},{\cal O}]\\
&=-2\eta_{\mu\nu}\de_D {\cal O}+2\de_{M_{\mu\nu}}{\cal O}\,,
}
where in the second line we used the Jacobi identity and $[S^\mu,{\cal O}]=0$, while in the third the conformal algebra. Since the last line is a linear combination of terms we have already considered, we conclude that we do not learn any new constraint from the sum rule associated to the breaking of $S_\mu$.
In appendix~\ref{app:contact} we derive identities for the contact terms in certain products of the energy-momentum tensor. The constraints above follow from these contact terms.
\section{Conformal Field Theory }\label{sec:CFT}
\subsection{General Considerations}
The purpose of this section is to explore the implication of sum rules for the spectrum of CFTs. The sum rules are obeyed by a state with finite energy and/or charge density. The strategy is to start from a state of a CFT on $ S^{d-1}\times \R $. The state on the cylinder $ S^{d-1}\times \R $ corresponds to an operator of the CFT in the plane with scaling dimension $\Delta_\Omega$. The energy $E$ of the state is given by
$$E={\Delta_\Omega\ov R}\,,\quad \epsilon=\frac{\Delta_\Omega}{s_{d-1}R^{d}}\,,$$
where $\epsilon$ is the energy density and $s_{d-1}$ is the volume of unit sphere $S^{d-1}$. Now we take the \textit{macroscopic limit} \cite{Lashkari:2016vgj,Jafferis:2017zna} by considering a family of operators with $\Delta_\Omega \to\infty$ and take $R\to\infty$ while keeping $\epsilon$ fixed.\footnote{The scale invariance of CFTs imply that $R$ is just a convenient auxiliary parameter. We could set $R=1$ and discuss the macroscopic limit equally well: it would involve zooming in onto a small patch of the cylinder in correlation functions.} Likewise the theory may have a conserved $U(1)$ symmetry and we can construct states with fixed charge density
$\rho=Q/(s_{d-1}R^{d-1})$.
In particular, we consider correlators of light operators of the form $\langle \Omega| \sO\sO|\Omega\rangle$ on the cylinder and take the required limit. Since we are considering a family of operators in the macroscopic limit, we are actually looking at a family of correlators, as we take $R\to\infty$. One of the underlying assumptions is that the correlators lead to a nice function of $\Delta_\Omega$ and/or $Q$ so that it makes sense to take the limit. In this limit the positions of the light operators $ \sO$ remain fixed on the cylinder.
The question that we are interested in is what states are responsible for saturating the sum rules discussed in section~\ref{sec:NGthm}. The sum rules are obtained as limits of two point functions on the cylinder, which we can rewrite by inserting a complete set of states
\es{CompSet}{
\bra{\Omega}\sO\sO\ket{\Omega}=\sum_{\#}\langle \Omega| \sO | \# \rangle \langle \#|\sO|\Omega\rangle\,.
}
In the limits prescribed by the sum rule, only certain states $|\#\rangle$ give contributions. For example as we discuss in Part~\ref{part2}, in the infinite volume limit, for the superfluid the $|\#\rangle$'s contributing are one particle states, while for free fermions they are particle-hole states~\cite{Alberte:2020eil}. If we trace them back to the states of CFT on the cylinder at finite $R$ and then via radial quantization to the operators on the plane, they correspond to certain operators appearing in the $s$-channel conformal block decomposition of the four point correlator $\langle \Omega(0) \sO (z) \sO(1) \Omega(\infty) \rangle$ (by $s$-channel we always mean the channel where we consider the OPE of $\Omega(0) \sO (z)$). Every $s$-channel conformal block sums up the contribution of a primary and its descendants in \eqref{CompSet}.
In our arguments below, we assume that the the states $\ket{\Om}$ are ground states in some sector of the theory. This translates to the fact that in the $S$ channel expansion of $\langle \Omega| \sO\sO|\Omega\rangle$, we only have operators with scaling dimension larger than $\Omega$. In cases where it is possible to construct EFT to describe the correlator in these heavy states, the heavy state effectively acts like vacuum for the EFT modes. We note that in the intuitive discussion of the Introduction, we did not need to make this assumption, and we expect that our conclusions remain true for arbitrary heavy scalar operators; it would be interesting to generalize our arguments to such states.
At the end of the day, we will be interested in spinning conformal blocks where $\sO$ carries spin index. Nonetheless in the macroscopic limit, the difference between scalar blocks and spinning blocks is inconsequential. So, let us illustrate the basic concepts of conformal blocks using scalar operators $\sO$
\begin{equation}\label{eqs:4point}
\langle \Omega(0) \sO (z) \sO(1) \Omega(\infty) \rangle= (z\bar z)^{-\frac{1}{2} (\Delta_\Omega+\Delta_{\sO})} \sum |C_{\Omega\sO\Delta}|^2\mathcal{G}_{\Delta_\Omega+\Delta,\ell} (z,\bar z)\,,
\end{equation}
where $\mathcal{G}_{\Delta_\Omega+\Delta,\ell} (z,\bar z)$ is the conformal block. Here we have parametrized the scaling dimension of the operators appearing in the intermediate channel in a way such that $\Delta$ denotes the excitation over the state $\Omega$ i.e $C_{\Omega\sO\Delta}$ is the OPE coefficient involving the operator $\Omega$, $\sO$ and the operator appearing in the $s$-channel with scaling dimension $\Delta_\Omega+\Delta$. This is a convenient choice as one of the main results of our paper involves putting a bound on this gap compared to $\Delta_\Omega$, i.e putting a bound on $\Delta$. We will come back to the discussion of gap in due time.
Starting from \eqref{eqs:4point}, one can transform to the cylinder and write down the correlator of $\sO$ in the heavy state $|\Omega\rangle$.
\begin{equation}\label{eqs:cyl}
\langle \Omega| \sO(\tau,\vec{n}_1) \sO(0,\vec{n}_2) | \Omega\rangle= \sum |C_{\Omega\sO\Delta}|^2g_{\Delta,\ell} (z,\bar z)\,,
\end{equation}
where we
defined the $\sO$ operator on the cylinder by conformally transforming it from the plane. Furthermore, we have defined $g_{\Delta,\ell}$ as
\es{gdef}{
g_{\Delta,\ell} (z,\bar z)\equiv (z\bar z)^{- \Delta_{\Omega}/2} \mathcal{G}_{\Delta_\Omega+\Delta,\ell} (z,\bar z)\,.
}
Note, the LHS of \eqref{eqs:cyl} is defined on the cylinder and thus the cross ratio $z,\bar z$ on the RHS should be understood as a function of cylinder coordinates. In particular, the conformally transformed operators $\sO$ are inserted at $(\tau,\vec{n}_1)$ and $(0,\vec{n}_2)$ with $\vec{n}_1\cdot\vec{n}_2=\cos\theta$. The cross ratio $z,\bar z$ is related to $\tau$ and $\theta$ in following way
\begin{equation}
\sqrt{z\bar z}=e^{\tau/R}\,,\quad \ \frac{z+\bar z}{2\sqrt{z\bar z}}=\cos\theta\,.
\end{equation}
The limit $R\to\infty$ is taken in a way, so that $\theta R\equiv x$ and $\tau$ are kept fixed and identified with the coordinate in the macroscopic limit. The macroscopic limit reads in terms of $u=\tau+ i x$ and $\bar u=\tau-ix$
\begin{equation}\label{Mlim}
z=1+\frac{u}{R}\,,\quad \bar{z}=1+\frac{\bar u}{R}\,, \quad R\to\infty\,, \quad u,\bar u\ \text{fixed}\,.
\end{equation}
In what follows, we will be establishing that in the macroscopic limit, the descendants are suppressed i.e the conformal blocks take a very simple form. This will be followed by the discussion on the implication of this suppression for the spectrum of primaries appearing in the $s$-channel. Along the way, we will explore how and which of these primaries survive the macroscopic limit and eventually saturate the sum rule.
\subsection{Supression of Descendants}
In this subsection we show that the primaries dominate the $s$-channel expansion in the macroscopic limit. We will first consider the heavy operator limit which amounts to $\Delta_\Omega\to\infty$. (This limit is not the macroscopic limit since we do not yet scale the coordinates as required in the macroscopic limit. The suppression of descendants in this limit was studied in~\cite{Jafferis:2017zna}.) Afterwards, we will keep the energy/charge density fixed and take $R\to\infty$ limit, i.e. the macroscopic limit.
Let us start with the blocks showing up in $\expval{\Omega \sO \sO \Omega}$, where $\Omega$ is the heavy primary with $\Delta_\Omega\to\infty$ and $\sO$ has $O(1)$ scaling dimension. We write the block as an expansion in Gegenbauer polynomials $C_\ell\equiv C_\ell^{(d/2-1)}\le(z+\bar{z}\ov 2 \sqrt{z\bar{z}}\ri)$:
\es{blocks}{
g_{\De,\ell}(z,\bar{z})&=\sum_{m,n} r_{m,n} \,(z\bar{z})^{(\Delta+m+n)/2} C_{\ell+m-n}\,,\\
r_{00}&=1\,,\\
r_{10}&={1\ov 2(2\ell+d-2)}\,{(\ell+1)(\De+\De_\sO+\ell)^2\ov \De_\Omega+\De+\ell}\,,\\
r_{01}&={1\ov 2(2\ell+d-2)}\,{(\ell+d-3)(\De+\De_\sO-\ell-d+2)^2\ov \De_\Omega+\De-\ell-d+2}\,\\
\vdots
}
(The sum is restricted to $\ell+m-n\geq 0$.) From the recursion relations of \cite{Dolan:2003hv}, one can show that
\es{desccoeff}{
r_{m,n}&={\widetilde r_{m,n}\ov \Delta_\Omega^{m+n}}+\dots\,,\\
\widetilde r_{m,n}&\approx m! \,n! \,, \quad\text{for $1\ll m,n\ll \Delta_\Omega$,}
}
where there are some powers of $m,\,n$ that we suppressed. We conclude that the contribution of the $k=m+n$ level descendants is suppressed by $1/\Delta_\Omega^k$. This suggests that we should be able to approximate the conformal block by the first term ($m=n=0$) in~\eqref{blocks}. Since we will be eventually interested in setting $z,\bar z$ close to $1$ as in~\eqref{Mlim} we need to make sure that the descendants corresponding to $m,n \gtrsim \Delta_\Omega$ continue to be suppressed. It can be shown again from the recursion relations that the contribution from $m,n \gg \Delta_\Omega$ is suppressed compared to that of the primary. Instead of going though a meticulous analysis of this kind, we use the knowledge of the $4d$ blocks to demonstrate the absence of anomalously large resummation effects.
The $4d$ conformal blocks are known in closed form~\cite{Dolan:2000ut}:
\es{4dblock}{
\mathcal{G}_{\de,\ell}(z,\bar z)&=(-1)^{\ell}\frac{z\bar z}{z-\bar z}\left[k_{\de+\ell}(z)k_{\de-\ell-2}(\bar z)-k_{\de-\ell-2}(z)k_{\de+\ell}(\bar z)\right]\\
k_{\beta}(z)&=z^{\beta/2} {}_2F_{1}\left(\frac{\beta-\Delta_{12}^{-}}{2},\frac{\beta-\Delta_{12}^{-}}{2},\beta;z \right)
}
As a consistency check, we recover \eqref{blocks}. Let us now investigate the macroscopic limit.
Let us first set $\Delta=O(1),\,\ell=O(1)$, and use $z= 1+{u\over R}$. We obtain using the definition \eqref{gdef}
\begin{equation}\label{nearvacBlock}
g_{\Delta,\ell} (z,\bar z)= (z\bar z)^{-\Delta_{\Omega}/2}\mathcal{G}_{\Delta_\Omega+\De,\ell}(z,\bar z)=(\ell+1)+{(\ell+1)(2\De-\De_\sO)(u+\bar{u})\over 2 R}+\dots\,.
\end{equation}
In order for our analysis to apply for the macroscopic limit we need to simultaneously take $z= 1+{u\over R}$ along with keeping the energy density fixed, in other words, we must study
the double scaling limit of the conformal blocks:
\begin{equation}\label{scaling}
z\equiv 1+{u\over R}\,,\quad \De={\cal E} R\,,\quad \ell\equiv p\, R \,,\quad \Delta_\Omega\equiv \ep\, R^d~.
\end{equation}
Then $w$ becomes a coordinate in the macroscopic limit, ${\cal E}$ is a fixed $O(1)$ coefficient (whose interpretation is the energy of the intermediate state above the energy of $|\Omega\rangle$), $p$ is the modulus of the momentum of the intermediate state and $\ep$ is the energy density in $|\Omega\rangle$. The reason that we scale $\Delta$ and $\ell$ as above is not immediately obvious but it will soon become clear that this leads to an interesting macroscopic limit.
We get
\es{DoubleScaledBlock}{
g_{\Delta,\ell} (z,\bar z)=R\,\exp\le({{\cal E}\ov 2}(u+\bar{u})\ri)\,{2\sinh\left({ p\ov 2} (u- \bar{u})\right)\over u-\bar{u}}+\dots\,.
}
Noting that in 4d,
\es{Gegenbauer}{
C_\ell^{(1)}\le(\cos(x)\ri)={\sin\le((\ell+1)x\ri)\ov \sin(x)}
}
and that
\es{Ratio}{
{z+\bar{z}\ov 2 \sqrt{z\bar{z}}}=\cos\le({u-\bar{u}\ov 2iR}\ri)+\dots\,,
}
we realize that in the two limits discussed above
\es{GegenLimit}{
C_\ell^{(1)}\le({z+\bar{z}\ov 2 \sqrt{z\bar{z}}}\ri)=\begin{cases}
(\ell+1)+\dots\qquad\quad &\text{for $\ell=O(1)$\,,}\\
R\,{2\sinh\left({ p\ov 2} (u- \bar{u})\right)\over u-\bar{u}}\qquad\quad &\text{for $\ell=O(R)$\,.}
\end{cases}
}
These are exactly the leading pieces of \eqref{nearvacBlock} and \eqref{DoubleScaledBlock}. Hence we find that the conclusion that the primaries dominate the blocks is indeed correct, i.e. it is not spoiled by resummation effect.
These conclusions remain true for spinning blocks, which are obtained by differential operators acting on scalar blocks \cite{Costa:2011mg,Costa:2011dw}.
Before we proceed it is worth noting that from~\eqref{GegenLimit} we readily see that the interesting intermediate blocks have $\ell=O(R)$ and $\Delta=O(R)$, which makes sense, since these correspond to intermediate states with finite momentum and frequency in infinite volume.
Thus far we have found that the contributions in the macroscopic limit are solely due to the primary operators in the intermediate channel and hence the macroscopic conformal blocks are exceedingly simple~\eqref{DoubleScaledBlock}. The factor $\exp\le({{\cal E}\ov 2}(u+\bar{u})\ri)$ is simply the time dependence that follows from translating the operators in energy eigenstates.
The spatial dependence $\sim{\sinh\left({ p\ov 2} (u - \bar{u})\right)\over u -\bar{u}}$ contains the absolute value of the momentum $p=|\vec p|$. It reflects the sum over all momentum eigenstate with fixed $|\vec p|$. Using the fact that $\sO$ is a scalar operator, the matrix elements have a simple dependence on $\vec p$ and the Gegenbauer polynomial arises through an integral of the form $C_{\ell=pR}^{(d/2-1)}(\cos(x/R))\sim \int d^{d-1} k \ \delta(|\vec k|-p)\,e^{i \vec k\cdot \vec x}$ (for details see \eqref{GegenLimiti}).
\subsection{An alternative argument}
For completeness let us present an intuitive argument for the suppression $\Delta_\Omega^{-k}$ found in \eqref{desccoeff} that also applies to spinning blocks.
One can argue for this suppression by computing matrix elements on the cylinder directly in the $\Delta_\Omega\to\infty$ limit. This computation can be done in a straightforward way even for spinning operators. For example, let us derive that the first descendant is indeed suppressed. To proceed, we recall that
\begin{equation}\label{eqs:three}
\begin{aligned}
\langle \Omega'(x) T_{\alpha\beta} (z) \Omega(y) \rangle= \frac{(y-z)^{\Delta-d}}{(y-x)^{2\Delta_\Omega+\Delta-d}} f(x,z)\,,
\end{aligned}
\end{equation}
where $f(x,z)$ is the OPE coefficient $C_{\Omega T_{00}\Omega'}$ times a kinematical function that carries the spin indices and is independent of the operator dimensions. $\Omega'$ is the primary with scaling dimension $\Delta_\Omega+\Delta$ appearing in the internal channel. We assume $\Delta$ scales slower than $\Delta_\Omega$, hence $\Omega'$ and $\Omega$ has the same scaling dimension to leading order in $\Delta_\Omega$. From \eqref{eqs:three} it follows that
\begin{equation}\label{eqs:threeder}
\begin{aligned}
\langle \partial_\mu\Omega'(x) T_{\alpha\beta} (z) \Omega(y) \rangle= \frac{2\Delta_\Omega (x-y)_\mu (y-z)^{\Delta-d}}{(y-x)^{2\Delta_\Omega+\Delta-d+2}} f(x,z) + \frac{(y-z)^{\Delta-d}}{(y-x)^{2\Delta_\Omega+\Delta-d}} \partial_\mu f(x,z) \,.
\end{aligned}
\end{equation}
Now we map \eqref{eqs:three} and \eqref{eqs:threeder} from the plane onto the cylinder ($T_{00}(z(\tau, \theta))$ should be understood as the operator on the cylinder and $T_{rr} (z)$ should be understood as an operator on the plane) via
\begin{equation}
\begin{aligned}
\langle \Omega' | T_{00}(z(\tau, \theta)) | \Omega\rangle &= \left(\frac{R}{r}\right)^{-d} \lim_{y\to\infty} y^{2\Delta_\Omega}\langle \Omega'(0) T_{rr} (z) \Omega(y) \rangle\sim \left(\frac{R}{r}\right)^{-d} f(0,z)\,,\\
\langle \partial_\mu\Omega' | T_{00}(z(\tau, \theta)) | \Omega\rangle &= \left(\frac{R}{r}\right)^{-d} \lim_{y\to\infty} y^{2\Delta_\Omega} \langle \partial_\mu\Omega'(0) T_{rr} (z) \Omega(y) \rangle\sim \left(\frac{R}{r}\right)^{-d} \partial_\mu f(0,z)\,,
\end{aligned}
\end{equation}
and it follows from the above
\begin{equation}
\begin{aligned}
& \frac{\langle \partial_\mu \Omega' | T_{00} | \Omega\rangle}{\langle\Omega' | T_{00} | \Omega\rangle} \simeq O(1)\,,
\end{aligned}
\end{equation}
as the ratio of function $f(x,z)$ and its derivative with respect to $x$ at $x=0$ is order one. Note that we are not making any assumption on the OPE coefficient $C_{\Omega T_{00}\Omega'}$, since it cancels out in the ratio.
Now the contribution to the correlator coming from an s-channel conformal block corresponding to a primary $\Omega'$ is given by
\begin{equation}
\langle \Omega| T^{00} T^{00} |\Omega\rangle \ni \sum_{\alpha,\beta=\Omega',P\Omega',PP\Omega'\cdots} \langle \Omega |T^{00}|\alpha\rangle \mathcal{N}_{\alpha\beta}^{-1} \langle\beta|T^{00}|\Omega\rangle\,,
\end{equation}
where $N_{\alpha\beta}=\langle\alpha | \beta\rangle$. We follow the notation of \cite{Simmons-Duffin:2016gjk} and by $P\Omega'$ we mean the operator $\partial \Omega'$. In particular, for the first descendant $P\Omega'$, we have
\begin{equation}
N_{P_\mu \Omega', P_\nu \Omega'}=\langle \Omega' | K_\mu P_\nu | \Omega' \rangle \simeq (\Delta_\Omega+\De)\delta_{\mu\nu} \,,
\end{equation}
where we used the commutation relation of $K$ and $P$. Altogether, we find that the contribution coming from the first descendant is suppressed, i.e.
\begin{equation}
\langle \Omega |T^{00} |\partial \Omega'\rangle \,\mathcal{N}_{P \Omega', P \Omega'}^{-1} \langle \partial \Omega'|T^{00}|\Omega \rangle \simeq\langle \partial_\mu \Omega' | T^{00}| \Omega\rangle|^2 \Delta_\Omega^{-1} \simeq \langle\Omega' | T_{00} | \Omega\rangle\Delta_\Omega^{-1} \,.
\end{equation}
A similar argument implies that the level $k$ descendant is suppresed by $\Delta_\Omega^{-k}$. The factors $m!\, n!$ in the conformal block in \eqref{desccoeff} can be accounted for from the number of distinct states at level $k=m+n$ with spin $\ell+m-n$.
\subsection{Implication for the Gap}
\label{subsec:gap}
In this subsection, we combine the Nambu-Goldstone boost sum rules, the macroscopic limit of CFTs and the form of conformal blocks in this limit to derive constraints on the gap in the CFT operator spectrum. We will be studying the correlator
\es{FocusCorr}{
\langle \Omega| T_{00}T_{00}|\Omega\rangle\,,
}
use the $s$-channel decomposition, and aim to put a bound on the gap $\Delta=\Delta_{\Omega'}-\Delta_\Omega$, where $\ket{\Omega'}$ is an excited state (above the ``vacuum'' state $\ket{\Omega}$) exchanged in the correlator. The sum rule corresponding to the correlator in \eqref{FocusCorr} was given in \eqref{t00t00} and in CFTs takes the form
\begin{equation}\label{eqs:sumrulet00}
{\rho_{T^{00}T^{00}} (\om,p)\ov p^2}= -\frac{d\epsilon}{d-1}\delta'(\omega)+O(p)\,.
\end{equation}
We chose to focus on a spectral density of two identical operators, since every contribution to such a spectral density is positive definite for $\om>0$. The obvious next step is to express the LHS of \eqref{eqs:sumrulet00} with CFT data from the cylinder.
To this end, we write the Euclidean correlator (for $\tau>0$) in the large charge limit at finite $R$ using the above results for conformal blocks:\footnote{We note that by mapping from the $\abs{z}<1$ region of plane to the cylinder, we naturally get $\tau<0$ and the radial (or cylinder time) ordering $T_{00}(0,\vec{n}_2)T_{00}(\tau,\vec{n}_1)$. Here we instead wrote the answer for $\tau>0$.}
\es{eqs:T00T00}{
\langle \Omega| T_{00}(\tau,\vec{n}_1) T_{00}(0,\vec{n}_2) |\Omega \rangle& =\epsilon^2\left[1+ \sum_\ell {\#\, \ell!\ov \De_\Omega^\ell} \, e^{-\ell |\tau|/R}\,C^{(d/2-1)}_{\ell}(\cos\theta)\right]\\
&\quad+\sum_{\Delta}{C_{\Omega T_{00}\De}^2\ov R^{2d}} \,\le[e^{-\Delta |\tau|/R}\, C^{(d/2-1)}_{\ell(\De)}(\cos\theta)+O\le(1/ \De_\Omega\ri)\ri]\,,
}
where in the first line we wrote the contribution from the conformal block of $\Omega$ itself (schematically including the descendants and using that its OPE coefficient $C_{\Omega T_{00}0}/R^{d}=\bra{\Omega}T_{00}\ket{\Omega}=\ep$), while in the second we included the blocks corresponding to other exchanged operators (and only wrote explicitly the contributions of the primaries of dimension $\De_\Omega+\De$ and spin $\ell(\De)$).
While the expression~\eqref{eqs:T00T00} is correct in the macroscopic limit, for the macroscopic limit to actually exist, one must make some assumptions about the OPE coefficients $C_{\Omega T_{00}\De}^2$. (A similar logic holds in the discussion of ETH in CFT~\cite{Lashkari:2016vgj}.)
Indeed, the most important exchanged operators which will contribute in the macroscopic limit are clearly such that $\Delta\sim R$ and $\ell\sim R$. In addition, there is a factor of $R^{d-3}$ from the Gegenbauer polynomials as in~\eqref{GegenLimit} (where $d=4$). Let us introduce the density of states by per unit energy and per unit momentum
\es{rhoDef}{
\rho(\omega,p; R)&=\sum_{\sO}\de\le(\om-{\De_\sO\ov R}\ri)\,\de\le(p-{\ell_\sO\ov R}\ri)\,,
}
This is a finite volume, un-smeared object. In terms of $\rho(\omega,p; R)$ we obtain an expression for the macroscopic limit of the the correlator~\eqref{eqs:T00T00}:
\es{eqs:T00T00M}{
\langle \Omega| T_{00}(\tau,x) T_{00}(0) |\Omega \rangle& = \epsilon^2+\int d\omega\, dp\ \rho(\omega,p) {C_{\Omega T_{00}\De}^2\ov R^{d+3}} e^{-\omega |\tau|}\, F_p(x)\,,
}
where $F_p(x)$ is the Gegenbauer polynomial with the $R$-dependence stripped out (we will soon write a concrete expression for it in any dimension; in $d=4$ this can be read out from~\eqref{GegenLimit} and we find $F_p(x) = {\sin(px)/ x}$).
For the macroscopic limit to exist the combination $K(\omega,p)\equiv\rho(\omega,p) \,{C_{\Omega T_{00}\De}^2/ R^{d+3}}$ must become $R$-independent in the appropriate sense.\footnote{To make the required averaging procedure precise we introduce the smearing:
\es{rhoCSmeared}{
K(\omega,p)&={1\ov 4 \delta^2}\,\sum_{\sO\in \sI(\om,p;R,\de) } {C_{\Omega T_{00}\De}^2\ov R^{d+3}}\,,\\
\sI(\om,p;R,\de)&\equiv\le\{\sO\,\Big\vert \,{\De_\sO\ov R}\in(\om-\de,\om+\de)\,,\, {\ell_\sO\ov R}\in(p-\de,p+\de)\ri\}\,,
}
where the set $\sI$ contains operators whose energy and momentum in the macroscopic limit agrees with $\om,\, p$ respectively. Here we have introduced an infinitesimal window width $\de$; $K(\omega,p)$ should be independent of $\de$ to leading order in the macroscopic limit, and hence we do not include $\de$ as its argument.}
Finally, to give an expression for $F_p(x)$ that is valid in any dimension we use the integral representation of the Gegenbauer polynomials:
\es{GegenLimiti}{
\lim_{R\to\infty}C^{(d/2-1)}_{pR}\left(\cos (x/R)\right)&=\lim_{R\to\infty} {\Ga\le(pR+d-2\ri)\ov 2^{d-3}\Ga\le(d-2\ov 2\ri)^2\Ga(pR+1)}\\
&\quad\times \int_0^\pi d\vartheta\ \sin^{d-3}(\vartheta)\, \le(\cos (x/R)+i\sin (x/R)\cos(\vartheta) \ri)^{pR}\\
&={(pR)^{d-3}\ov 2^{d-3}\Ga\le(d-2\ov 2\ri)^2}\int_0^\pi d\vartheta\ \sin^{d-3}(\vartheta)\, e^{i p x \cos(\vartheta)}\\
&={1\ov 2^{d-2}\pi^{(d-2)/2}\Ga\le(d-2\ov 2\ri)}\,{R^{d-3}\ov p}\int d^{d-1}k \ \de\le(\abs{\vec{k}}-p\ri) \,e^{i \vec{k}\cdot \vec{x}}\,.
}
The last relation is very intuitive: if we zoom in onto a small patch of the sphere, we get plane waves with fixed $|\vec p|$, averaged over all directions. In $d=4$ this precisely agrees with~\eqref{DoubleScaledBlock}.
Therefore,
\begin{equation}\label{Fform}F_p(\vec x)={1\ov 2^{d-2}\pi^{(d-2)/2}\Ga\le(d-2\ov 2\ri)p}\,\int d^{d-1}k \ \de\le(\abs{\vec{k}}-p\ri) \,e^{i \vec{k}\cdot \vec{x}}\end{equation}
The expression~\eqref{Fform} allows us to rewrite~\eqref{eqs:T00T00M} in a simpler form, after doing the $p$ integral
\es{eqs:T00T00Mi}{
\langle \Omega| T_{00}(\tau,x) T_{00}(0) |\Omega \rangle& = \epsilon^2+{1\ov 2^{d-2}\pi^{(d-2)/2}\Ga\le(d-2\ov 2\ri)}\int d\omega \,d^{d-1}k\ K(\omega,k) e^{-\omega |\tau|} \,{e^{i \vec{k}\cdot \vec{x}}\over k}\,.
}
It is now straightforward to obtain an expression for the spectral density in the macroscopic limit
\es{KeyForm}{
\rho_{T^{00}T^{00}} (\om,p)\sim {1\over p}\left(K(\omega,p)-K(-\omega,p)\right)~.
}
We must now make contact with \eqref{eqs:sumrulet00}. We learn that, at the very least, $K(\omega,p)$ must have support at $\omega=0$ and $p=0$. This means that there must be operators with $\Delta_\text{gap}/R\to 0$ as $R\to\infty$. In more conventional CFT terms, if we denote the dimension of the ground state by $\Delta_\Omega$, we conclude that the gap $\Delta_\text{gap}$ must be smaller than $\Delta_\Omega^{1/d}$ in the sense that there must be operators with \begin{equation}\label{boundg}\Delta_\text{gap}/\Delta_\Omega^{1/d}\to 0\end{equation}
as $\Delta_\Omega\to \infty$. If we in addition recall the physical reasoning advocated in the introduction, namely, that there is a scale invariant low energy theory describing the massless excitations in the small $\omega$ regime, we would conclude that $\Delta_\text{gap}=O( 1)$.
In addition, we can constrain the angular momentum of these operators in a similar fashion since the sum rule \eqref{eqs:sumrulet00}
implies that we need to take infinitesimal $p$. This means that the angular momentum must satisfy $$\ell /\Delta_\Omega^{1/d} \to 0$$ and by a similar effective theory reasoning it follows that it is in fact $O(1)$.
This implies a surprisingly small primary gap of $O(1)$ (for dimension and angular momentum) around any heavy state. One can say more about the density of operators or OPE coefficients using \eqref{eqs:sumrulet00} if one makes additional assumptions about the density or OPE coefficients separately. Incidentally, since the density $\rho(\omega,p; R)$ cannot decay as $R\to\infty$ (because the number of participating operators cannot go to zero) we get a rather general bound on the OPE coefficient ${C_{\Omega T_{00}\De}^2 }= O(R^{d+3})$. (This bound can be strengthened by requiring a more realistic density of states.)
Recalling that the sum rule \eqref{eqs:sumrulet00} is linear in $\ep$, we can refine this bound slightly. Since the macroscopic limit is a double scaling limit, we can choose $\ep\sim \Delta_\Omega/R^d$ at will, while $R,\,\Delta_\Omega\to \infty$. We can then resolve $R^{d+3}$ as $\Delta_\Omega R^{3} \le(\Delta_\Omega/R^d\ri)^\al$, and show that $\al=0$ by contradiction: were $\al>0$, we take $\ep$ large, while for $\al<0$, we take $\ep$ small to derive a violation of the sum rule. Hence we conclude
\es{BoundOnC}{
{C_{\Omega T_{00}\De}^2 }= O(\Delta_\Omega R^{3})~.}
\part{Examples}\label{part2}
\section{Superfluid Phase}\label{sec:superfluid}
In the superfluid phase the ground state is homogenous, isotropic, has a finite charge density, and breaks $U(1)$ spontaneously. As in the previous section, we imagine this state to be the
macroscopic limit of a family of large charge states on the cylinder $S^{d-1}\times \R $. These states correspond to scalar operators of the underlying CFT.
In the $Q\to\infty$ limit, there is a separation of energy scales; $\rho^{-1/(d-1)}$ (with $\rho\sim {Q/R^{d-1}}$ the charge density) is a UV scale while the IR scale is given by $R$. For distances much larger than the UV scale but much less than $R$, the system is described by an effective field theory with $1/Q$ being the expansion parameter \cite{Hellerman:2015nra,Monin:2016jmo}. As we will see the UV scale is precisely related to $\epsilon^{-1/d}$ (where $\epsilon$ is the energy density) mentioned in the introduction while $R$ plays the role of $L$. In the infinite volume limit, the state with finite energy density breaks $SO(d+1,1)\times U(1)$ down to a $SO(d)$ and a linear combination of $U(1)$ and time translation. The action of the effective field theory can be constructed in the CCWZ way \cite{PhysRev.177.2239,PhysRev.177.2247} in terms of a field $\chi$ and its fluctuation $\pi$ around the symmetry breaking saddle. The field $\pi$ is identified with the Goldstone mode, corresponding to the aforementioned spontaneous breaking.
In this section, we will elaborate on the general ideas described in the previous section using the explicit example of superfluid EFT, which captures the large charge sector of an underlying CFT. The aim is to identify the primary states that saturate the sum rule for broken boost symmetries and discuss the connection to what we have found based on general arguments in the previous section.
\subsection{General Consideration}
The effective field theory is described by the Euclidean action \cite{Hellerman:2015nra,Monin:2016jmo}
\begin{equation}\label{eqs:schi}
S=-c_1\int\ d^dx\ \sqrt{g}\ \abs{\partial\chi}^d+ \cdots+ i \int\ d^dx\ \sqrt{g}\ \rho\dot{\chi}
\end{equation}
(Here we assume that the underlying CFT preserves parity, hence the parity violating terms of~\cite{Cuomo:2021qws} are forbidden.)
The saddle point that describes a ground state with finite homogenous charge density $\rho$ is given by
\begin{equation}
\begin{aligned}
\chi &= -i \mu \tau +\pi\\
c_1 d\mu^{d-1}&= \rho
\end{aligned}
\end{equation}
Expanding around this saddle, we obtain the effective action for the Goldstone field $\pi$
\begin{equation}\label{spi}
S_\pi = \frac{d(d-1)}{2}c_1\mu^{d-2}\int d^dx \sqrt{g}\left(\dot{\pi}^2+\frac{1}{d-1}\partial_i\pi\partial^i\pi \right)+\cdots\,.
\end{equation}
For the purpose of examining the sum rules, we will eventually be interested in the correlator involving the stress-energy tensor (in particular the components $T_{00}$, $T_{0i}$) and current $J_i$. From \eqref{eqs:schi}, we obtain the stress-energy tensor (below we consider the Minkowski signature metric and we use $\tau=i t$)\footnote{Note, in Euclidean signature, we have Euclidean stress energy tensor defined as $T_{\mu\nu}=\frac{2}{\sqrt{g}}\frac{\delta S}{\delta g^{\mu\nu}}$. The Minkowski $T$ is related via $T^{\tau\tau}=-T^{00}$ and $T^{\tau i}=-iT^{0i}$. In the text, we use the index $0$ and work with the Minkowski $T$ operator. The same applies for $J$. \label{fn:Wick} }
\begin{equation}\label{eqs:T}
\begin{aligned}
T_{00}&= \epsilon + i \frac{d\epsilon}{\mu} \frac{d\pi}{d\tau}+ \cdots\,, \qquad T_{0i}= \frac{d\epsilon}{(d-1)\mu} \partial_i\pi + \cdots\,,\qquad T_{ij}=-\eta_{ij}\frac{\epsilon}{d-1}+\cdots\,,
\end{aligned}
\end{equation}
where we used that the (leading order in $\mu$) energy density is
\begin{equation}\label{eqs:densitymu}
\epsilon= c_1(d-1)\mu^d\,.
\end{equation}
The two-point correlator of $\pi$ on the cylinder in the large $Q$ limit is given by
\begin{equation}\label{prop}
D(\tau,x)=\frac{\frac{\mu^2}{s_{d-1}R^{d-1}}}{d\epsilon}\left[-|\tau|+\sum_{\ell=1}\frac{2\ell+d-2}{d-2} \frac{e^{-\omega_\ell |\tau|/R}}{2\omega_\ell/R}C^{d/2-1}_{\ell}(\cos\theta)\right]\,,
\end{equation}
where $\theta=\arccos(\vec{n}_1\cdot\vec{n}_2)$ and $\tau$ are the angle and time separation between insertion of two $\pi$ fields respectively. Here $\omega_\ell=\sqrt{\ell(\ell+d-2)/(d-1)}$~.
Another relevant quantity for our purpose is the current $J_\mu$ corresponding to $U(1)$ that acts as a shift symmetry on the field $\chi$:
\begin{equation}\label{eqs:J}
\begin{aligned}
& J_0 = \rho+i\frac{(d-1)\rho}{\mu}\frac{d\pi}{d\tau}\,, \qquad J_i = \frac{\rho}{\mu}\partial_i\pi\,.
\end{aligned}
\end{equation}
In rest of this section, we will be heavily using \eqref{eqs:T}, \eqref{prop}, \eqref{eqs:J}.
\subsection{Sum Rules}
The three major sum rules that we have been discussing in this paper involves looking at $\langle T_{00}T_{00}\rangle$, $\langle T_{00}T_{0i}\rangle$ and $ \langle T_{00}J_k\rangle$ in the large charge state. For now, we will focus on the two point correlator of $T_{00}$ in the large charge state and will perform the analysis in detail by working on the cylinder and then taking the macroscopic limit while tracking the set of states, that are eventually going to saturate the corresponding sum rule. We will come back to the other sum rules involving $\langle T_{00}T_{0i}\rangle$ and $\langle T_{00}J_k\rangle$ later and verify them by working directly in the macroscopic limit without performing the computation on the cylinder.
\subsection*{$\langle T_{00}T_{00}\rangle$}
The two point correlator in $T_{00}$ in the large charge state is given by for $\tau>0$
\begin{equation}\label{eqs:T00T00s}
\begin{aligned}
\langle T_{00}(\tau) T_{00}(0) \rangle& =\epsilon^2\left[1+ \frac{d^2}{2(d-2)}\frac{1}{\Delta_\Omega} e^{-|\tau|/R}C^{d/2-1}_{1}(\cos\theta)+\cdots\right]\\
&\quad+ \frac{d\epsilon}{s_{d-1}R^{d-1}}\left[\sum_{\ell=2}\left(\frac{2\ell+d-2}{2(d-2)} \right)\left(\frac{\omega_\ell}{R}\right) e^{-\omega_\ell |\tau|/R}C^{d/2-1}_{\ell}(\cos\theta)+\cdots\right]
\end{aligned}
\end{equation}
For $\tau<0$, the right hand side provides us with $\langle T_{00}(0) T_{00}(\tau) \rangle$. We will keep this mind and for brevity use $\langle T_{00} T_{00} \rangle$. We can compare the above with \eqref{eqs:T00T00}. In the first line on the RHS, we have two terms corresponding to the contribution coming from exchange of $\Omega$ and its descendant. We have omitted the contributions coming from higher descendants. They are suppressed in $\Delta_\Omega\to \infty $ limit. In the previous section, we have argued that there is no cumulative effect coming from considering all the descendants together. In the second line, we have a contribution from a single Regge trajectory, one primary for each given integer $\ell \geq 2$ with the scaling dimension $\Delta_\Omega+\Delta$, where we have $\Delta=\sqrt{\ell(\ell+d-2)/(d-1)}$~. Here we denote $\Delta$ as $\omega_\ell$ since we have single Regge trajectory and sum over spin $\ell$ suffices. The OPE coefficients can be read off as
\begin{equation}\label{OPE}
C^{2}_{\Omega T_{00}\Delta}=\frac{d\Delta_\Omega}{s_{d-1}^2}\left(\frac{2\ell+d-2}{2(d-2)} \right)\sqrt{\ell(\ell+d-2)/(d-1)}~.
\end{equation}
In what follows we will show that the states with $\ell\sim pR$ contribute to the macroscopic limit and then in the $p\to0$ limit saturate the sum rule. To obtain the macroscopic limit we therefore only need to take the macroscopic limit of the OPE coefficients \eqref{OPE} and we obtain for the spectral density
\begin{equation}\label{eq:spec}
\rho_{T^{00}T^{00}} (\om,p)=\frac{d \epsilon }{2 \sqrt{d-1}} \, p\left[\delta\left(\omega-\frac{p}{\sqrt{d-1}}\right)-\delta\left(\omega+\frac{p}{\sqrt{d-1}}\right)\right]\,.
\end{equation}
In the $p\to 0$ limit \eqref{eq:spec} yields
\begin{equation}
\rho_{T^{00}T^{00}} (\om,p) \underset{p\to 0}{\simeq} -\frac{d \epsilon }{ d-1}\, p^2\delta'(\om)=-(\epsilon+P)\delta'(\om)\,,
\end{equation}
and the sum rule \eqref{t00t00} is satisfied. We note that $p\to 0$ limit means that $\ell=o(R)$, thus it is clear that the states with $\Delta=\ell=o(R)$ from the Regge trajectory saturate the sum rule. The OPE coefficients for these states
\es{OPEsuperfluid}{
C^{2}_{\Omega T_{00}\Delta}=O\left(\Delta_\Omega \ell^2\right)=o\left(\Delta_\Omega R^2\right)\,.
}
Comparing with the bound from \eqref{BoundOnC} in sec.~\ref{subsec:gap}, we see that this is down by a factor of $1/R$. The reason for this is that that there are $o(R)$ states in the superfluid in this kinematical regime, whereas the bound \eqref{BoundOnC} allows only one state (as a worst case scenario). \\
\subsection*{The rest of the sum rules}
The two-point correlators of $T_{0\mu}$ and $J_\mu$ in the macroscopic limit can be found from the macroscopic limit of the two-point correlator of the $\pi$ field:
\begin{equation}\label{eqs:propmac}
\langle \pi(\tau,\vec{x}) \pi(0) \rangle_{\text{macro}}= {\mu^2\ov d(d-2)s_{d-1}\ep}\, \frac{\sqrt{d-1}}{\le(\frac{1}{d-1}\tau^2+\vec{x}\cdot\vec{x}\ri)^{d/2-1}}\,.
\end{equation}
From \eqref{eqs:T} and \eqref{eqs:J} we immediately realize that all we need to know for these computations is that for $\tau>0$
\es{eqs:propmac2}{
\p_\tau\p_i\langle \pi(\tau,\vec{x}) \pi(0) \rangle_{\text{macro}}&= {\mu^2\ov s_{d-1}\ep}\,\frac{\frac{\tau}{\sqrt{d-1}} x_i}{\le(\frac{1}{d-1}\tau^2+\vec{x}\cdot\vec{x}\ri)^{d/2+1}}\\
&\equiv h_i(x)\,.
}
In terms of this correlator:
\es{AllOthers}{
\langle T_{0i}(x) T_{00} (0)\rangle_{\text{macro}}&=-i{d^2\ep^2\ov(d-1)\mu^2}\,h_i(x)\,,\\
\langle T_{0i}(x) J_{0} (0)\rangle_{\text{macro}}&=\langle T_{00}(x) J_{i} (0)\rangle_{\text{macro}}=-i{d\ep\rho\ov \mu^2}\,h_i(x)\,.
}
The Fourier transformed function $\widetilde{h}(\omega,\vec{p})$ is defined through:
\begin{equation}
\label{eq:FourierWightman}
\widetilde{h}_i(\omega,\vec{p})=\int d^d x \ e^{i\omega \tau -i\vec{p}\cdot\vec{x}}\,h_i(x)\,.
\end{equation}
By rescaling $\tau=\sqrt{d-1}\,\tau',\, \om=\om'/\sqrt{d-1}$, we get a standard Lorentz invariant integral
\es{eq:FourierWightman2}{
\widetilde{h}_i\le({\omega'\ov \sqrt{d-1}},\vec{p}\ri)&={\sqrt{d-1} \,\mu^2\ov s_{d-1}\ep} \int d\tau'\, d\vec{ x} \ e^{i\omega' \tau' -i\vec{p}\cdot\vec{x}}\,\frac{{\tau'} x_i}{\le({\tau'}^2+\vec{x}\cdot\vec{x}\ri)^{d/2+1}}\\
&={\sqrt{d-1} \,\pi \mu^2\ov 2d\ep}\frac{\partial}{\partial\omega'}\frac{\partial}{\partial p^i}\left[ \Theta(\omega')\Theta({\om'}^2-p^2)\le({\om'}^2-p^2\ri)\right]\,.
}
We use \eqref{commWightman} and \eqref{Grels} to write
\es{rhoFromWightman}{
\rho_{A,B}(\omega,\vec{p})=\frac{1}{2\pi}\left(\tilde{G}_{A,B}(\omega,\vec{p})-\tilde{G}^*_{A,B}(-\omega,-\vec{p})\right)\,,
}
and get the following expressions for the spectral densities from \eqref{eq:FourierWightman2}:
\es{MacroSumRuleSuperfluid}{
\rho_{T_{0i}T_{00}}(\om,p)&=\frac{1}{2}(\epsilon+P)p_i \left[\Theta(\omega)\delta\left(\omega-\frac{p}{\sqrt{d-1}}\right)+\Theta(-\omega)\delta\left(\omega+\frac{p}{\sqrt{d-1}}\right)\right]\\
&\underset{p\to 0}{\simeq} (\epsilon+P)p_i \delta(\omega)\,,\\
\rho_{T_{00}J_{i}}(\om,p)&=\rho_{T_{0i}J_{0}}(\om,p)\underset{p\to 0}{\simeq} \rho p_i \delta(\omega)\,.
}
Hence the sum rules \eqref{SumRule3}, \eqref{SumRuleJ}, and \eqref{SumRuleD2} are all satisfied.
We end this section with a remark about three-dimensional parity violating superfluids~\cite{Cuomo:2021qws}. The ground state contains vortices, but it is homogenous and isotropic, hence our sum rules apply. The low lying excitations are phonons (the $\pi$ excitations that we have been studying) and other {\it softer} vortex excitations with spin $\ell$ and a gap of $\Delta\sim \ell(\ell+1)Q^{-3/2}$ above the ground state. (For parity preserving superfluids, the excitations with spin $\ell$ above the ground state have a gap $\Delta\sim \ell(\ell+1)$. Vortex excitations in parity preserving fluids only appear for $\ell > \sqrt{Q}$ \cite{Cuomo:2017vzg, Cuomo:2019ejv}.
A natural question is what saturates the sum rules in parity violating superfluids. The answer is that it is still the phonons, because \cite{Cuomo:2021qws} found that the OPE coefficients involving the phonon modes, the ground state $\Omega$ and $T_{00}$ (or other relevant components of $T$) stay unchanged compared to the parity preserving one. Thus the softer vortex modes should not contribute in the macroscopic limit and should not play any role in the sum rule. This is corroborated by the fact that there are only finitely many vortices on the cylinder, hence they disappear in the macroscopic limit. It would be nice to check these claims explicitly.
\section{Free Scalar}
\label{sec:FreeScalar}
In this section, we study the saturation of the sum rules associated with broken boosts and scale invariance described in section \ref{sec:NGthm} for the case of a free relativistic complex scalar field in a finite charge density state in dimension $d > 2$.
The free scalar field does not lead to a state with finite energy density in the macroscopic limit due to the flat directions arising from the shift symmetry of the free scalar action. Indeed, due to these flat directions, $\Delta\sim Q$ rather than $\Delta\sim Q^{d/(d-1)}$. This means that the effective field theory description of~\eqref{eqs:schi} is inappropriate, and should be replaced by the approach~\cite{Hellerman:2017veg,Hellerman:2017sur,Hellerman:2018xpi,Bourget:2018obm,Grassi:2019txd,Beccaria:2020azj}. In fact sometimes these two types of effective theories are connected~\cite{Sharon:2020mjs}. In our analysis of the boost symmetry realization on the large charge states in free field theory, we will not use an effective theory approach, rather, we will pursue a more straightforward analysis of the correlation functions.
\subsection{General Consideration}
The Euclidean two-point correlation function in a theory of a free complex scalar field is given by (we use the normalization of \cite{Osborn:1993cr}):
\begin{equation}
\label{eq:PropScalarOsborn}
G(x-y) \equiv \langle \bar{\phi}(x)\phi(y)\rangle = \frac{1}{(d-2)s_{d-1}}\,\frac{1}{|x-y|^{2\Delta_\phi}},
\end{equation}
where $ s_{d-1} = {2\pi^{\frac{d}{2}}}/{\Gamma(\frac{d}{2})}$ and $\Delta_\phi = \frac{d-2}{2}$. We define the lightest operator of charge $Q$ by the following:
\begin{equation}
\label{eq:DefOQForBosons}
\mathcal{O}_{Q} \equiv \frac{\left((d-2)s_{d-1}\right)^{Q/2}}{\sqrt{Q!}}\bar{\phi}^Q.
\end{equation}
Its scaling dimension scales like $\Delta_Q\sim Q$. Using radial quantization, the ground state of charge $Q$ on the cylinder is:
$$\vert\Omega\rangle \equiv \mathcal{O}_{Q}(0) |0\rangle .$$
For the purpose of examining the sum rules, we will eventually be interested in the correlators involving the $U(1)$ current $J^\mu$ and the stress-tensor $T^{\mu\nu}$.
On the plane, they are given by the following expressions:
\begin{align}
&T^{\mu\nu} = \partial^\mu\bar{\phi}\partial^\nu\phi +\partial^\nu\bar{\phi}\partial^\mu\phi-g^{\mu\nu}g^{\alpha\beta}\partial_\alpha\bar{\phi}\partial_\beta\phi + T^{\mu\nu}_{\text{imp}}, \qquad T^{\mu\nu}_{\text{imp}} = \chi\left[g^{\mu\nu}\partial^2-\partial^\mu\partial^\nu\right]\bar{\phi}\phi,\label{eq:BosonicStressTensor}\\
&J^\mu = i \left(\partial^\mu\bar{\phi}\phi-\bar{\phi}\partial^\mu\phi\right),
\end{align}
where $\chi = \frac{d-2}{2(d-1)}$ is the coefficient of the curvature coupling term ${\cal R}\bar{\phi}\phi$ (with ${\cal R}$ the Ricci scalar). In the next subsection, we will argue that the sum rules associated with the broken boosts and dilatations are not affected by the improvement term, and therefore we can ignore it when calculating correlation functions for the purpose of verifying the sum rules (see subsection \ref{subsubsec:FlatSpaceCorr} for further details). For details about Wick rotating tensors to Minkowski signature see footnote~\ref{fn:Wick}.
\subsection{ Sum Rules}
\label{subsec:SumRules}
We would like to check how the sum rules described in section \ref{sec:NGthm} are satisfied for the case of the free bosonic field theory. For this purpose, one can take the following strategy: first, calculate the correlators which are of interest for the saturation of the sum rules on the plane. Then, map it onto the cylinder $ S^{d-1}\times \R $ and take the radius of the $S^{d-1}$ sphere to infinity. This last step amounts to taking the infinite volume limit.
In practice, taking the macroscopic limit (as described in subsection \ref{subsubsec:MacroLimitScalar}) directly on the flat space correlators is equivalent to performing the procedure described above.
In what follows, we will be using $x^\mu,y^\mu$ as the coordinates on the plane.
On the cylinder, we want to evaluate
$$\langle \Omega| J^{i}(0)T^{\tau\tau}(\tau,\theta_j) | \Omega\rangle\,,$$
where $(\tau, \theta_j)$ refer to the coordinate on the cylinder. We can choose a coordinate system where we have only one angle $\theta$. This correlator is related to a four-point correlator on the plane
$$\langle \Omega(0) T^{rr} (z,\bar z) J^{i}(1) \Omega(\infty)\rangle\,$$
by conformal transformation where we have
\begin{equation}
r=e^{\tau/R}\,,\quad z\bar z=r^2 \,,\quad \frac{z+\bar z}{\sqrt{z\bar z}}=2\cos\theta \,.
\end{equation}
We follow the usual convention (where $x^{d}$ denotes the Euclidean time on the plane)
$$z=x^d+ i x^{1}\,, \ \bar z=x^d- i x^{1}\,.$$
In the macroscopic limit, as we let $z\to 1$, $T^{rr}$ on the plane essentially becomes $T^{dd}$.
Therefore, we will leverage that and on the plane we will calculate
$$\langle \Omega(0) T^{dd} (z,\bar z) J^{i}(1) \Omega(\infty)\rangle.$$
We deal with the right hand side of the sum rule in a similar way, going from the plane to the cylinder.
The organization of this subsection is as follows: in subsection \ref{subsubsec:FlatSpaceCorr}, we calculate the correlation functions on the plane. In subsection \ref{subsubsec:MacroLimitScalar}, we define the macroscopic limit associated with the free bosonic theory. In subsection \ref{subsubsec:SaturationOfSumRulesScalar}, we calculate the spectral density and show that the sum rules are satisfied.
\subsubsection{Flat Space Correlators}
\label{subsubsec:FlatSpaceCorr}
The free field correlators are obtained using Wick contractions.
We concentrate on
$$\langle \mathcal{O}_Q(0) T^{\mu\nu}(x)J^\rho(y)\mathcal{O}_{-Q}(\infty)\rangle\,.$$
Eventually, when taking the macroscopic limit, we will choose to work a particular configuration where $T,J$ lie in the $(x^{d}, x^1)$ plane as mentioned previously and get the macroscopic correlator and then covariantize to obtain the correlator for arbitrary insertion points of $T$ and $J$.
We first evaluate the contribution to $$\langle \mathcal{O}_Q(0) T^{\mu\nu}(x)J^\rho(y)\mathcal{O}_{-Q}(\infty)\rangle $$
from the non-improved stress energy tensor. We find that this results in:
\begin{equation}
\label{eq:CorrelationFunction1Scalars}
\begin{aligned}
&\langle \mathcal{O}_Q(0) T^{\mu\nu}(x)J^\rho(y)\mathcal{O}_{-Q}(\infty)\rangle \\
& \ni iQ\left[ H^{\mu\rho}(x-y)F^\nu(x)+H^{\nu\rho}(x-y)F^\mu(x)-g^{\mu\nu}H^{\alpha\rho}(x-y)F_\alpha(x)\right]\,.
\end{aligned}
\end{equation}
Here the functions $F$ and $H$ are given by:
\begin{equation}
\begin{aligned}
F^\mu(x)\equiv \frac{x^\mu}{s_{d-1}|x|^{d}}\,,\quad
H^{\mu\nu}(x) \equiv \frac{1}{s_{d-1}|x|^{d}} \left(\eta^{\mu\nu}-\frac{dx^\mu x^\nu}{x^2} \right) \,.
\end{aligned}
\end{equation}
We can further evaluate the contribution to the correlator coming from $T_{\text{imp}}$
\begin{equation}
\begin{aligned}
&\langle \mathcal{O}_Q(0) T_{\text{imp}}^{\mu\nu}(x)J^\rho(y)\mathcal{O}_{-Q}(\infty)\rangle \\
& = i\chi Q \left(g^{\mu\nu}\partial^2-\partial^\mu\partial^\nu \right) \left[F^\rho(x-y)\left(G(y)-G(x) \right)+F^\rho(y)G(x-y)+(Q-1)F^\rho(y)G(x) \right].
\end{aligned}
\end{equation}
Here $G$ is given by \eqref{eq:PropScalarOsborn}. This correlator has a finite macroscopic limit, but at the end of the day this does not contribute to the sum rules. The reason is that the improvement terms drop from the expressions for the charges and the sum rules stem from looking at the correlators of the charge generator and an operator that plays the role of an order parameter. In what follows, we will thus work with the non-improved $T_{\mu\nu}$ (unless otherwise stated) in order to show the saturation of sum rules.
The same conclusion holds for the case of the free scalar field even when considering the sum rule associated with the broken dilatations \eqref{SumRuleD2}.
Finally, the right hand side of the sum rule requires us to find
\begin{equation}
\label{eq:RHSOfWardIdentity}
\langle \mathcal{O}_Q(0) J^{d}(z=1,\bar z=1) \mathcal{O}_{-Q}(\infty)\rangle=-\frac{iQ}{s_{d-1}}\,.
\end{equation}
Once we are equipped with the expression for the free space correlator, we take the macroscopic limit, which is the subject of the next subsection.
\subsubsection{Macroscopic Limit}
\label{subsubsec:MacroLimitScalar}
By the state/operator correspondence, the operator $\mathcal{O}_Q$ in equation \eqref{eq:DefOQForBosons} describes a state with charge density $\rho$
on the cylinder $ S^{d-1}\times \R $. This can be seen by noting that
\begin{equation}
\langle \Omega\vert J^0\vert \Omega\rangle_{\text{cyl}}=R^{1-d}\frac{\langle \mathcal{O}_{Q}(0)J^0(1)\mathcal{O}_{-Q}(\infty)\rangle}{\langle\mathcal{O}_{Q}(0)\mathcal{O}_{-Q}(\infty)\rangle}=\rho\,.
\end{equation}
In the macroscopic limit, the charge density is kept finite as we take $R\to\infty$ limit. The scaling dimension associated with the operator $\mathcal{O}_Q$ satisfy $\Delta_Q \sim Q \sim R^{d-1}$. The energy density reads:
\begin{equation}
\epsilon = \frac{\Delta_H}{s_{d-1}R^d}\to 0.
\end{equation}
In the macroscopic limit, we take:
\begin{equation}
\label{eq:XmuYmuMacroLim}
x^\mu= y^\mu + \frac{u^\mu}{R}\,,\qquad \text{where}\ R\to\infty,\, \text{and} \ u^\mu\ \text{is fixed}.
\end{equation}
We also set $y^\mu=\delta^{\mu}_0$. In terms of $z,\bar{z}$ we have:\footnote{In this section we use the somewhat unfortunate notation $u=\tau+i u^1$, and when we refer to a spacetime point, we always write $u^\mu$. (In previous sections we instead used $u=\tau+i x^1$, but here $x^\mu$ is already reserved for coordinates on the plane before mapping to the cylinder and taking the macro limit. }
\begin{equation}
z = 1+\frac{u}{R}, \qquad \bar{z}=1+\frac{\bar{u}}{R}.
\end{equation}
We note that the convention of taking the macroscopic limit is different compared to \cite{Jafferis:2017zna}. We made this choice in order to ensure that there is no parity transformation implemented while taking the macroscopic limit. Altogether, the macroscopic limit of the correlation functions which are of interest for the saturation of the sum rules is given by:
\begin{equation}
\langle \Omega \vert J^i(0)T^{\tau\tau}(u^\mu)\vert\Omega\rangle_{\text{mac}} \equiv \lim_{R\to\infty} \left(R^{1-2d}\frac{\langle \Omega(0) T^{dd}(x)J^i(1)\Omega(\infty)\rangle}{\langle\Omega(0)\Omega(\infty)\rangle}\bigg|_{x^\mu=\delta^{\mu}_d+u^\mu/R} \right),
\end{equation}
and
\begin{equation}
\langle \Omega\vert J^\tau\vert \Omega\rangle_{\text{mac}} \equiv\lim_{R\to\infty} \left(R^{1-d}\frac{\langle \Omega(0)J^d(1)\Omega(\infty)\rangle}{\langle\Omega(0)\Omega(\infty)\rangle}\bigg|_{x^\mu=\delta^\mu_d+u^\mu/R} \right).
\end{equation}
The correlators associated with the numerator on the right-hand side in both equations above correspond to the correlators given by equations \eqref{eq:CorrelationFunction1Scalars} and \eqref{eq:RHSOfWardIdentity}.
Both limits are well-behaved. Setting $y^\mu=\delta^\mu_d$, the resulting expressions (in Euclidean signature) in the macroscopic limit after covariantizing are:
\begin{equation}
\begin{aligned}
\label{eq:MacroCorrFunctT00Ji}
\langle \Omega \vert J^i(0) T^{\tau\tau}(u^{\mu})\vert\Omega\rangle_{\text{mac}}
=-i\frac{\rho d}{
s_{d-1}}\frac{\tau {u}^i}{{u}^{d+2}}
\end{aligned}
\end{equation}
and
\begin{equation}
\begin{aligned}
\label{eq:ChargeDensityExpr}
&\langle \Omega\vert J^\tau\vert \Omega\rangle_{\text{mac}} =-i\rho\,.
\end{aligned}
\end{equation}
One can see from the last equation that the macroscopic limit has been taken in a way so that the charge density $\rho$ is constant.
\subsubsection{Saturation of the Sum Rules}
\label{subsubsec:SaturationOfSumRulesScalar}
The sum rules are inherently statements in Minkowski signature. For the purpose of evaluating the sum rules, we evaluate the following Wightman correlation function by doing proper analytic continuation of \eqref{eq:MacroCorrFunctT00Ji} and recalling $T^{00}=-T^{\tau\tau}$:
\begin{equation}
\label{eq:WightmanDef}
G_{T^{00},J^i}(u^0,\vec{u}) \equiv \langle \Omega \vert T^{00}(u^0,\vec{u})J^i(0)\vert\Omega\rangle\,.
\end{equation}
Here we have defined $(u^0,\vec{u})$ to denote the Minkowski coordinates. The proper analytic continuation is achieved by letting $\tau=iu^0+\epsilon$ and then take $\epsilon\to 0^{+}$.
We find that
\begin{equation}
\label{eq:GT00Ji}
G_{T^{00},J^i}(u^0,\vec{u}) = -\frac{\rho d}{
s_{d-1}}\frac{u^0{u}^i}{{u}^{d+2}}.
\end{equation}
We have already computed the spectral density from (a very close analog of) this Green's function in section~\ref{sec:superfluid}, see \eqref{eq:FourierWightman2} and \eqref{rhoFromWightman}. Plugging into those formulas we get
\begin{equation}
\begin{aligned}
\label{eq:ScalarSpectralDensity}
\rho_{T^{00},J^i}(\omega,\vec{p})&= \frac{1}{2}\rho\, p^{i} \left[\Theta(\omega)\delta(\omega-p)+\Theta(-\omega)\delta(\omega+p)\right]\\
& \underset{p \to 0}{\simeq} \rho p^i \delta(\omega) ,
\end{aligned}
\end{equation}
Thus, we find that the sum rule \eqref{SumRuleJ} associated with the broken boosts symmetry is satisfied.
As for the sum rule associated with the broken scale invariance, we note that in the macroscopic limit, without including the improvement term in the stress-tensor \eqref{eq:BosonicStressTensor}, the following happens to be true for the free scalar
$$\langle\Omega| T^{0i} J^{0} |\Omega\rangle=\langle\Omega| T^{00} J^{i} |\Omega\rangle, $$
where $T_{\mu\nu}$ above refers to the stress-tensor \eqref{eq:BosonicStressTensor} without the improvement part $T_{\text{imp}}$. Together with the analysis shown at the beginning of subsection \ref{subsubsec:FlatSpaceCorr}, this immediately tells us that the sum rule associated with the broken scale symmetry \eqref{SumRuleD2} is satisfied. The extra factor of $(d-1)$ in the sum rule comes from the contraction of spatial indices in the expression for $x_iT^{0i}$.
\subsection{Sum Rules from the Cylinder Vantage Point}
\label{subsec:SumRulesFromTheCylinderVantage}
In this subsection, we identify the states that are responsible for saturation of the sum rules discussed in the previous subsection. To this effect, we would like to study a sum rule in analogy to the $T^{00}T^{00}$ sum rule that was studied in the superfluid case in section \ref{sec:superfluid}. In the free scalar case, however, the energy density vanishes in the macroscopic limit and the $T^{00}T^{00}$ spectral density hence vanishes in the $p\to 0$ limit.
Instead, we study the sum rule associated with the $T^{00}J^{0}$ correlator, which reads:
\begin{equation}
\rho_{T^{00}J^{0}}(\om,p)\underset{p\to 0}{\simeq} - \rho p^2 \delta'(\omega) \, .
\end{equation}
The above equation follows from combining the conservation of $J$ with the $T^{00}J^{i}$ sum rule just as in section~\ref{sec:energySumrule}.
In the large $Q$ limit, we can write for $\tau>0$
\begin{equation}\label{eqs:T00J0free}
\begin{aligned}
\langle \Omega|T_{00} (\tau) J_{0}(0) |\Omega \rangle& =\frac{Q^2}{s_{d-1}^2R^{2d-1}}\\
&\quad + \underbrace{\frac{Q}{s_{d-1}^2R^{2d-1}}\left[\sum_{\ell=1}\frac{\ell^2}{(d-2)} e^{-\ell |\tau|/R}C^{(d/2-1)}_{\ell}(\cos\theta)\right]}_{\text{terms for the sum rule}}+\cdots\,.
\end{aligned}
\end{equation}
The RHS of the above expression gives $\langle \Omega|J_{0}(0) T_{00} (\tau) |\Omega \rangle$ for $\tau<0$.
This equation should be thought of as an analogue of \eqref{eqs:T00T00s} valid in the superfluid for the free scalar case. Here, the dots indicate terms that are not important for reproducing the sum rule
in the macroscopic limit, i.e to reproduce the $p\to 0$ behavior of spectral density. Now it is easy to realize that the sum rule is saturated by states living on the Regge trajectory $\omega_\ell=\ell$. The calculation proceeds exactly the same way as in the superfluid case, with the only difference being that here $\omega_\ell=\ell$ for all $\ell$. Once again, the states $\ell\sim p R$ become important in the infinite volume limit.
Now let us understand in detail how this single Regge trajectory on the cylinder comes about from the previous calculation of the four-point correlator on the plane via Wick contraction. Schematically, we have the following type of contractions in the correlator
\begin{equation}
\langle \wick{ \c6{\phi}^{Q-2} \c3{\phi}\c2 {\phi} \vert \c2{\partial \bar\phi(x)} \c5{\partial \phi(x)} \c3 {\bar\phi(y)}\c4{\partial \phi(y)} \vert\c4{\bar\phi} \c5{\bar\phi} \,\c6{\bar\phi}^{Q-2}} \rangle\sim Q^2\,, \nonumber
\end{equation}
which gives the first line of \eqref{eqs:T00J0free},
and
\begin{equation}
\langle \wick{ \c3{\phi}^{Q-1} \c2{\phi}\vert \c2{\partial\bar\phi(x)}\c2{\partial\phi(x)} \c2{\partial \bar \phi(y)} \c2{\phi(y)} \vert \c2{\bar\phi}\,\c3{\bar \phi}^{Q-1} }\rangle\sim Q\,, \nonumber
\end{equation}
which gives the second line of \eqref{eqs:T00J0free}.
The contractions that yield result proportional to $Q$ survive in the macroscopic limit and eventually a part of it
\es{importantPart}{
& \frac{Q}{R^{2d-3}}\partial_\tau^2 G(\tau,\theta)
}
is responsible for saturating the sum rules. $G(\tau,\theta)$ is the free scalar propagator on the cylinder:
\es{CylinderProp}{
&G(\tau,\theta)=\frac{1}{(d-2)s_{d-1}} \left(1+e^{-2|\tau|/R}+2e^{-|\tau|/R}\cos\theta\right)^{-\frac{d-2}{2}} \,.
}
The underbraced term in \eqref{eqs:T00J0free} comes precisely from the expansion of $\partial_\tau^2 G(\tau,\theta)$ in terms of Gegenbauer polynomials.
As a final remark, let us understand the single Regge trajectory in terms of single or multiparticle states. Of course, it is clear that the relevant states would be labelled by a single number $p\sim \ell R$. The question is whether this corresponds to single or multiparticle states. To proceed, recall that the relevant contribution arises from contracting one leg of $T$ with the bra $\langle\Omega|$ and one leg of $J$ with ket $|\Omega\rangle$, and then contracting the leftover leg of $T$ with left over leg of $J$. We can think of breaking apart the $TJ$ Wick contraction and inserting a complete set of states. To make this notion more precise, we start by noting that the state $|\Omega\rangle $ defines a Bose-Einstein condensate on the cylinder. One can view it as created by the zero angular momentum modes ($a_0^\dagger$) of the scalar field $\phi$, i.e. $|\Omega\rangle\sim(a_0^\dagger)^Q|0\rangle$, where $|0\rangle$ is the true vacuum.
On the cylinder, we can annihilate states with zero angular momentum and charge $1$ from this condensate, and create a particle with angular momentum $\ell$ and charge $1$ on top of it. This resembles a bit the case of a particle-hole pair in the theory of Fermi surface (studied in section \ref{sec:Fermi}), albeit one important difference: the particle-hole excitations on top of the Fermi surface are labelled by two numbers, the angular momentum of the particle and the angular momentum of the hole, both of which can take non-zero values. In the case of the free scalar field, however, the hole carries zero angular momentum. Thus, the particle-hole pairs are labelled by a single number and form the Regge trajectory as discussed above. Note that we need a particle-hole pair as opposed to a single particle excitation on the cylinder since
$\langle \Omega | T a^{\dagger}_{\ell}a_0 |\Omega\rangle$ is non-zero whereas $\langle \Omega | T a^{\dagger}_{\ell}|\Omega\rangle =0 $ due to the charge conservation on the cylinder. Nonetheless, in the macroscopic limit, both $|\Omega\rangle$ and $a_0|\Omega\rangle$ define the same state with equal and finite charge density. Thus, as we take $R\to\infty$ limit, the aforementioned particle-hole pair on the cylinder (that consists of a hole carrying zero angular momentum) behaves like a single particle excitation, labelled by a single momentum vector $\vec{p}$ in the infinite volume theory. This is similar to the behavior described in section 5 of \cite{Alberte:2020eil}, in the context of the free massive particle.
\section{Free Fermions}\label{sec:Fermi}
\label{sec:FreeFermions}
We study systems of free fermionic field theories at finite charge density $\rho$ and energy density $\epsilon$. In subsection \ref{subsec:FreeFermionsLargeCharge} we address the large charge sector of these models in three and four spacetime dimensions. In subsection \ref{subsec:RelativisticFermiGas} we consider the relativistic Fermi gas in four dimensions, calculate the spectral density, and show the saturation of the sum rules as described in section \ref{sec:NGthm} for the broken symmetries.
\subsection{Free Fermions at Large Charge}
\label{subsec:FreeFermionsLargeCharge}
In this subsection we consider free fermionic field theories with a global $U(1)$ symmetry in $d=3$ and $d=4$ dimensions. We denote the lightest operator of charge $Q$ by $\mathcal{O}_Q$, its dimension being $\Delta_Q$, and focus on the large $Q$ limit.
In addition, we restrict the discussion to cases in which the ground state is homogeneous and isotropic. Under this assumption, and in the limit of large charge, the lightest operator of charge $Q$ can be constructed using simple counting arguments. Similar arguments can be found in \cite{Shenker:2011zf,Aharony:2015pla}, where the leading order term in the expansion of $\Delta_Q$ was calculated for the $d=3$ case.
Our main findings in this subsection are:
\begin{enumerate}
\item There is no $Q^0$ term in the expansion for $\Delta_Q$.
\item The energy difference between the first excited state and the ground state is $O(1)$:
\begin{equation}
\Delta_{+1,Q}- \Delta_Q = O(1).
\end{equation}
\end{enumerate}
We emphasize that the analysis described in this subsection holds only under the assumption of a homogeneous and isotropic ground state. Towards the end of this subsection, we make a comment regarding cases in which this is not the situation.
We start with the $d=3$ case. The fermionic field $\psi$ is a two component complex Grassmann spinor. Under global $U(1)$ symmetry it transforms as:
\begin{equation}
\label{eq:GlobalU1TransDef}
\psi \to e^{i\alpha}\psi.
\end{equation}
As a result of the Fermi statistics, operators of the form $\psi^n$ vanish for all $n>2$.\footnote{We use the notation $\psi^2=\ep^{ab}\psi_a\psi_b$.} Therefore, in order to construct operators of charge $Q$ under the transformation \eqref{eq:GlobalU1TransDef}, one necessarily has to include fermions dressed with derivatives, in order to construct a product with more than two fermions. The resulting operator $\mathcal{O}_Q$ is therefore expected to be of the form ${\psi^2(\partial\psi)^2(\partial^2\psi)^2}_{\cdots}\,$.
The equation of motion reduces the number of independent physical degrees of freedom. Hence, without loss of generality, we can eliminate $\partial_2\psi$, as it is linearly dependent on the other derivatives of $\psi$.
We define by $\mathcal{D}^n$ an operator which consists of $n$ spacetime derivatives of the following form:
\begin{equation}
\mathcal{D}^n \equiv \partial_{0}^{n_0}\partial_1^{n_1}, \qquad \text{where } n_0+n_1=n.
\end{equation}
Note that $n$ therefore represents the total number of derivatives of type $\partial_0$ and $\partial_1$.
The operator $\mathcal{O}_Q$ will consist of multiplication of all the possible terms of the form $\prod_{k=0}^n (\partial_0^{n-k}\partial_1^k\psi)^2$, where $n$ takes all integer values between $0$ to a maximal value that is determined by the requirement of having a charge $Q$.
Note that for each value of $n$, there are $n+1$ different terms that contain $n$ derivatives (from either $\partial_0$ type, $\partial_1$ type, or a mixed combination). The fermions consist of two-complex components Grassmann fermions, hence each such term can be taken with a power of two at most.
Altogether, for large $Q$ the operator $\mathcal{O}_Q$ takes the following form:
\begin{equation}
\label{eq:OperatorConstruction3d}
\mathcal{O}_Q \propto (\psi)^2(\mathcal{D}_1\psi)^4\cdots (\mathcal{D}_{n_{\text{max}}}\psi)^{2n_{\text{max}}+2},
\end{equation}
where $n_{\text{max}}$ is determined by the condition that the total number of fermions in the operator $\mathcal{O}_Q$ is equal to $Q$:
\begin{equation}
\label{eq:3dQuadraticEquation}
Q = (n_{\text{max}}+1)(n_{\text{max}}+2).
\end{equation}
It is important to note that we are not obtaining all integer values of $Q$ through this construction, since $n_{\text{max}}$ is an integer. We discuss the operators for $Q$'s that cannot be produced through \eqref{eq:3dQuadraticEquation} later in this section: they have spin and hence do not correspond to homogeneous states on the cylinder.
The total number of derivatives $n_{\text{der}}$ that appear in the operator \eqref{eq:OperatorConstruction3d} is given by:
\begin{equation}
\label{eq:3dNumberOfDerivatives}
n_{\text{der}} = \frac{2}{3}n_{\text{max}}(n_{\text{max}}+1)(n_{\text{max}}+2)= \frac{2}{3}Q\, n_{\text{max}}.
\end{equation}
Solving the quadratic equation \eqref{eq:3dQuadraticEquation} for $n_{\text{max}}$ and plugging it into \eqref{eq:3dNumberOfDerivatives}, one finds the total number of derivatives associated with the operator $\mathcal{O}_Q$ to be given by:
\begin{equation}
\label{eq:3dCalcNder}
n_{\text{der}} = \frac{2}{3}Q^{\frac{3}{2}}-Q+\frac{\sqrt{Q}}{12}+O(Q^{-\frac{1}{2}}).
\end{equation}
The dimension $\Delta_Q$ associated with the operator $\mathcal{O}_Q$ is given by $
\Delta_Q = Q\Delta_{\psi}+1\cdot n_{\text{der}}\,$,
where $\Delta_\psi=1$ is the dimension of the fermionic field in $d=3$. Thus, we get:
\begin{equation}
\label{eq:3dDimensionOfOper}
\Delta_Q = \frac{2}{3}Q^{\frac{3}{2}}+\frac{1}{12}\sqrt{Q} +O(Q^{-\frac{1}{2}}).
\end{equation}
Next, we turn to consider the case of a Weyl fermion in $d=4$ dimensions.\footnote{The same analysis can be automatically extended to the case of a Dirac fermion.} The operator $\mathcal{O}_Q$ will consist of multiplication of all the possible terms of the form $\prod_{k,l=0}^n (\partial_0^{n-k-l}\partial_1^k\partial_2^l\psi)^2$, where again $n$ takes all possible integer values between $0$ to a maximal value that depends on $Q$.\footnote{Similar to the $d=3$ case, without loss of generality, we can set $\partial_3\psi$ as the term which linearly depends on the others using the equations of motion.}
For each value of $n$, there are $\frac{(n+2)(n+1)}{2}$ different terms, each can be taken with a power of $2$ at most. This yields the following expression for the large charge operator $\mathcal{O}_Q$:
\begin{equation}
\label{eq:OpConstruction4d}
\mathcal{O}_Q \propto (\psi)^2(\mathcal{D}_1\psi)^6(\mathcal{D}_2\psi)^{12}\cdots (\mathcal{D}_{n_{\text{max}}}\psi)^{(n_{\text{max}}+1)(n_{\text{max}}+2)}.
\end{equation}
From the condition that the operator carries a charge $Q$ under the global $U(1)$ symmetry, one finds the following relation for $n_{\text{max}}$:
\begin{equation}
\label{eq:Eqfornmax4d}
Q = \frac{1}{3}\left( n_{\text{max}}+3\right)\left(n_{\text{max}}+2\right)\left(n_{\text{max}}+1\right).
\end{equation}
The total number of derivatives in the operator
\eqref{eq:OpConstruction4d} reads:
\begin{equation}
\label{eq:4dNumOfDer}
n_{\text{der}}= \frac{1}{4}\left(n_{\text{max}}+3 \right)\left( n_{\text{max}}+2\right) \left(n_{\text{max}}+1\right)n_{\text{max}}=\frac{3}{4}n_{\text{max}}Q.
\end{equation}
Using equation \eqref{eq:Eqfornmax4d}, we get:
\begin{equation}
\label{eq:ExprforNmax}
n_{\text{der}}= \frac{3^{\frac{4}{3}}}{4}Q^{\frac{4}{3}}-\frac{3}{2}Q+\frac{1}{4\cdot 3^{\frac{1}{3}}}Q^{\frac{2}{3}}+O(Q^{-\frac{2}{3}}).
\end{equation}
The dimension $\Delta_Q$ is given by $\Delta_Q = Q\Delta_{\psi}+ n_{\text{der}}= \frac{3}{2}Q+n_{\text{der}}$. Using equation \eqref{eq:ExprforNmax}, we find the following expression for the scaling dimension $\Delta_Q$:
\begin{equation}
\label{eq:DeltaQ4d}
\Delta_Q = \frac{3^{\frac{4}{3}}}{4}Q^{\frac{4}{3}}+\frac{1}{4\cdot 3^{\frac{1}{3}}}Q^{\frac{2}{3}}+O(Q^{-\frac{2}{3}}),
\end{equation}
Note that as in the $d=3$ case, there is no $Q^0$ term in the expansion for $\Delta_Q$.
Excited states correspond to particle-hole excitations.
The lowest order excitation corresponds to removing a single fermion from the Fermi surface and replacing it with an excited fermion, with an energy slightly above the Fermi energy. In the language of operators, this problem translates to removing a single fermion with $n_{\text{max}}$ derivatives from the operator \eqref{eq:OperatorConstruction3d} (or \eqref{eq:OpConstruction4d} in the $d=4$ case), and replacing it with a fermion that carries $n_{\text{max}}+1$ derivatives. The resulting operator, which we denote by $\mathcal{O}_{+1,Q}$, corresponds to the next-to-lightest operator that carries the same charge $Q$.
Following \eqref{eq:3dNumberOfDerivatives} (or \eqref{eq:4dNumOfDer} in the $d=4$ case), the total number of derivatives such an operator contains is shifted by $+1$ compared to the number of derivatives associated with the operator $\mathcal{O}_Q$.
Hence, the energy difference between the lowest energy excitation to the ground state energy satisfies:
\begin{equation}
\Delta_{+1,Q}- \Delta_Q = O(1).
\end{equation}
Let us make a comment regarding cases in which the ground state is not homogeneous and isotropic. In terms of energy levels on the cylinder $ S^{d-1}\times \R $, this corresponds to cases in which the outermost energy shell is not fully occupied. For simplicity, we focus on $d=3$ dimensions. $\Delta_Q$ is then given by:
\begin{equation}
\label{eq:DeltaQNotFilledShell}
\Delta_Q = \sum_{j=\frac{1}{2}}^{j_{\text{max}}-1}\left(2j+1\right)\varepsilon_j+\delta Q \varepsilon_{j_{\text{max}}},
\end{equation}
where $\varepsilon_j \equiv j+\frac{1}{2}$ are the energy eigenvalues on the sphere, and the $(2j+1)$ factor above represents the degeneracy. $\delta Q$ represents the particles in the outermost, not necessarily filled energy shell (as described in figure \ref{fig:EnergyLevels3d}) and it is related to the charge $Q$ by:
\begin{equation}
\label{eq:QForNotFilledShell}
Q = \sum_{j=\frac{1}{2}}^{j_{\text{max}}-1} \left( 2j+1\right) + \delta Q.
\end{equation}
\begin{figure}[!h]
\centering
\includegraphics[scale=0.6]{EnergyLevelsFermi.jpg}
\caption{An illustration describing the energy shells. For each $j<j_{\text{max}}$, there are $2j+1$ occupied states in each shell. At the outermost shell, which corresponds to the $j_{\text{max}}$ shell, there are $\delta Q$ occupied states, where $0\leq \delta Q \leq 2j_{\text{max}}+1$. For $\delta Q = 0$ or $\delta Q= 2j_{\text{max}}+1$, the outermost shell is filled and we recover the homogenous and isotropic ground state with an associated $\Delta_Q$ that is given by equation \eqref{eq:3dDimensionOfOper}. } \label{fig:EnergyLevels3d}
\end{figure}
In general, $\delta Q$ can take any integer value in the range between $0$ to $2j_{\text{max}}+1$, where the latter corresponds to the case in which the outermost shell is filled and is associated with $j_{\text{max}}$, while the former corresponds to the case in which the outermost shell is the $j_{\text{max}}-1$ shell and it is also filled (see figure \ref{fig:EnergyLevels3d}). Note that for $\delta Q =0$ or $\delta Q = 2j_{\text{max}}+1$ we simply recover the homogeneous and isotropic ground state scenario described above: $Q$ is then such that equation \eqref{eq:3dQuadraticEquation} is satisfied with an integer value of $n_{\text{max}}$, as defined above, and using the two equations \eqref{eq:DeltaQNotFilledShell}, \eqref{eq:QForNotFilledShell} one can reproduce the result \eqref{eq:3dDimensionOfOper} for $\Delta_Q$ based on the counting arguments.
\begin{figure}[!h]
\centering
\subfloat[]{\label{fig:Fermion3dDeltaQ}{\includegraphics[width=0.35\textwidth]{figure_3d_deltaQ_Q.jpg}}} \qquad \qquad
\subfloat[]{\label{fig:Fermion3dEnergyFlactuations}{\includegraphics[width=0.55\textwidth]{figure_3d_energy_flactuations.jpg}}}
\caption{An illustration describing the behavior of the scaling dimension for a general value of $Q$. Figure \ref{fig:Fermion3dDeltaQ} describes $\Delta_Q$ as a function of $Q$. In figure \ref{fig:Fermion3dEnergyFlactuations}, the blue dashed line shows $\Delta_Q-\frac{2}{3}Q^{3/2}$ as a function of $Q$, while the orange line describes $\frac{\sqrt{Q}}{12}$. The points in which the blue dashed line meets the orange line correspond to cases in which $Q$ is such that equation \eqref{eq:3dQuadraticEquation} is satisfied with integer values of $n_{\text{max}}$. }
\label{fig:Fermion3dFig}
\end{figure}
In figure \ref{fig:Fermion3dFig}, we refer to the general $Q$ case. One can see, from figure
\ref{fig:Fermion3dEnergyFlactuations}, that $\Delta_Q$ meets the value \eqref{eq:3dDimensionOfOper} that is associated with a homogeneous and isotropic ground state only for specific values of $Q$. These values correspond to the cases in which the outermost shell is filled, as described above. It is also interesting to notice that the fluctuations which appear in the graph of the difference between the $\Delta_Q$ to the leading order term in equation \eqref{eq:3dDimensionOfOper}, $\Delta_Q-\frac{2}{3}Q^{3/2}$ (described in figure \ref{fig:Fermion3dEnergyFlactuations}), possess an amplitude of order $O(\sqrt{Q})$. We learn that $\Delta_Q$ is not an analytic function of $Q$, if we try to define it for by using all integer $Q$'s. It is however analytic, if we only consider $Q$'s that correspond to completely filled shells, and analytically continue from these points.
The same analysis can be extended to other dimensions as well. In appendix
\ref{app:DeltaQFreeFermion}, we discuss the $d=4$ case. We find that similar to the $d=3$ case, outside the scope of a homogeneous and isotropic ground state there are fluctuations in the difference between $\Delta_Q$ to the leading order term of \eqref{eq:DeltaQ4d}, these fluctuations are of order $O(Q^{2/3})$ (which is the same order in $Q$ as the next to leading order term in \eqref{eq:DeltaQ4d}), hence $\Delta_Q$ is not analytic in $Q$. In addition, we show that for a general number of spacetime dimensions $d$ the leading behavior of $\Delta_Q$ in the large charge limit is given by:
\begin{equation}
\label{eq:LargeChargeGeneralDLeadingOrder}
\Delta_Q\underset{Q\to\infty}{\simeq} \frac{ 2^{\frac{1-\lceil d/2\rceil }{d-1}}}{d\Gamma(d-1)} \left[\Gamma(d)Q \right]^{d/(d-1)} \,,
\end{equation}
where by $\lceil d/2\rceil $ we refer to the ceiling function of $d/2$.
\subsection{Relativistic Fermi Gas}
\label{subsec:RelativisticFermiGas}
In \cite{Alberte:2020eil}, non-relativistic Fermi-liquid theories were studied and were shown to satisfy the sum rule associated with the broken boost symmetry by a particle-hole continuum. We extend this analysis to the case of a relativistic Fermi gas, a state of matter that consists of many non-interacting fermions. We show that similar to the non-relativistic case, the sum rules associated with the broken symmetries are satisfied by particle-hole states.
As in \cite{Alberte:2020eil}, we are interested in cases in which the ground state of the theory itself breaks boosts, while preserving spacetime translational and spatial rotational invariance. The ground state of the theory consists of fermionic particles that occupy all momentum states with momenta $|\vec{p}| \leq p_F$, where $p_F$ is the corresponding Fermi momentum. It is therefore taken to be a tensor product of single-particle momentum eigenstates:
\es{eq:FLGroundState}{
\vert \text{GS}\rangle \equiv \mathcal{N} \prod_s \prod_{|\vec{p}\leq p_F} |\vec{p},s\rangle, \qquad |\vec{p},s\rangle\equiv c_{\vec{p}}^{s\dagger}|0\rangle,
}
where $c_{\vec{p}}^{s\dagger}$ is a fermionic creation operator creating a single particle state of momentum $\vec{p}$ and spin $s$. The constant $\mathcal{N}$ is a normalization constant and chosen such that $\langle \text{GS}\vert \text{GS}\rangle=1$.
From the anti-commutation relations, $\lbrace c_{\vec{p}}^r\,, c_{\vec{q}}^{s\dagger} \rbrace = (2\pi)^3\delta^{(3)}(\vec{p}-\vec{q})\,\delta^{rs} $, it is clear that the following properties of the ground state \eqref{eq:FLGroundState} hold:
\es{}{
c_{\vec{p}}^{s\dagger}&\vert \text{GS}\rangle=0, \qquad |\vec{p}|\leq p_F, \\
c_{\vec{p}}^s\,&\vert \text{GS}\rangle =0, \qquad |\vec{p}|> p_F.
}
The first line above simply represents Pauli's exclusion principle, while the second line states that one cannot annihilate a state which is not already contained in the Fermi ground state \eqref{eq:FLGroundState}.
Particle-hole states are defined defined by \cite{Alberte:2020eil}:
\begin{equation}
\label{eq:ParticleHoleState}
\vert \chi\rangle = c_{\vec{p_1}}^{r\dagger}\, c_{\vec{p_2}}^{r'}\vert \text{FL}\rangle,
\end{equation}
where $p_2 \leq p_F$ and $ p_1 >p_F$. The annihilation operator $c_{\vec{p_2}}^{r'}$ creates a hole with momentum $\vec{p}_2$ and spin $r'$, while the creation operator $c_{\vec{p_1}}^{s\dagger}$ creates a fermionic particle with momentum $\vec{p}_1$ and spin $r$.
The total momentum associated with the particle-hole state $\vert\chi\rangle$ is given by the difference $\vec{p}=\vec{p}_1-\vec{p}_2$. The energy associated with such a state reads $E(\vec{p},\vec{p_2})= E_{p_1}-E_{p_2}$, where $E_{p_i}=|\vec{p}_i|\equiv p_i$ is the energy of a single particle state with momentum $\vec{p}_i$. \\
In this subsection, we study the saturation of the sum rules associated with the broken boosts and dilatation for the case of relativistic free Dirac fermions in $d=4$ dimensions in flat spacetime.
\subsubsection{Matrix Elements}
The action of a free, massless Dirac fermion in $d=4$ dimensions is given by:
\begin{equation}
\label{eq:DiracAction}
S = \int d^4x \, i\bar{\psi}\slashed{\partial}\psi\, ,
\end{equation}
where $\bar{\psi}=\psi^\dagger\gamma^0$, $\slashed{\p}=\gamma^\mu\partial_\mu$.
The fermion can be written in terms of modes expansion:
\es{}{
&\psi(x) = \int \frac{d^3p}{(2\pi)^3}\,\frac{1}{\sqrt{2\omega_p}}\sum_s \left( c_{\vec{p}}^s\, u_{\vec{p}}^s\,e^{-ipx} + d_{\vec{p}}^{s\dagger}v_{\vec{p}}^s\,e^{ipx} \right),\\
&\bar{\psi}(x) = \int \frac{d^3p}{(2\pi)^3}\,\frac{1}{\sqrt{2\omega_p}}\sum_s \left( d_{\vec{p}}^s \,\bar{v}_{\vec{p}}^s\,e^{-ipx} + c_{\vec{p}}^{s\dagger}\bar{u}_{\vec{p}}^s\,e^{ipx} \right),
}
where $c_{\vec{p}}^s\,,c_{\vec{p}}^{s\dagger}$ and $d_{\vec{p}}^s\,,d_{\vec{p}}^{s\dagger}$ are the creation and annihilation operators of fermionic particles and anti-particles (respectively). They satisfy the following anti-commutation relations:
\begin{equation}
\label{eq:AntiCommRel}
\lbrace c_{\vec{p}}^r\,, c_{\vec{q}}^{s\dagger} \rbrace = \lbrace d_{\vec{p}}^r\,, d_{\vec{q}}^{s\dagger} \rbrace = (2\pi)^3\delta^{(3)}(\vec{p}-\vec{q})\,\delta^{rs}.
\end{equation}
The spinors $u_{\vec{p}}^s$ and $v_{\vec{p}}^s$ represent the solutions of the massless Dirac equation. They satisfy:
\begin{align}
&\sum_s u_{\vec{p}}^s\, \bar{u}_{\vec{p}}^s = \sum_s v_{\vec{p}}^s \, \bar{v}_{\vec{p}}^s = \gamma \cdot p \,, \label{eq:SpinorRelations1}\\
&\sum_s \bar{u}_{\vec{p}}^s\gamma^\mu u_{\vec{p}}^s = 4p^\mu,\label{eq:SpinorRelations2}\\
&\sum_{r,r'} \bar{u}_{\vec{p}_2}^{r'}\gamma^\mu u_{\vec{p}_1}^r\bar{u}_{\vec{p}_1}^r\gamma^\nu u_{\vec{p_2}}^{r'} = 4\left(p_2^\mu\,p_1^\nu +p_2^\nu\, p_1^\mu -g^{\mu\nu}p_1\cdot p_2 \right). \label{eq:SpinorRelations3}
\end{align}
The stress-tensor is given by:
\begin{equation}
T^{\mu\nu}(x) = \frac{i}{4}\left[ \bar{\psi}\gamma^\mu\partial^\nu\psi -\partial^\nu\bar{\psi}\gamma^\mu\psi + (\mu \leftrightarrow \nu ) \right] -\eta^{\mu\nu}\mathcal{L}.
\end{equation}
We consider the system \eqref{eq:DiracAction} in the ground state described by \eqref{eq:FLGroundState}.
The energy density $\epsilon$ and pressure $P$ are defined as the vacuum expectation values of $T^{00}$ and $T^{ij}$ (respectively) with respect to the ground state \eqref{eq:FLGroundState} using:
\begin{equation}
\epsilon\equiv \langle \text{GS}\vert T^{00} \vert\text{GS}\rangle \,,\quad\ P\delta^{ij}\equiv \langle \text{GS}\vert T^{ij} \vert\text{GS}\rangle \,.
\end{equation}
Here we have secretly used the fact that ground state is isotropic to pull out the factor $\delta^{ij}$ in defining the pressure $P$. Using $\langle \text{GS}\vert c_{\vec{p}}^{s\dagger}c_{\vec{q}}^{s'} \vert\text{GS}\rangle = (2\pi)^3\delta^{(3)}(\vec{p}-\vec{q})$,
we find:
\es{}{
&\epsilon = -\frac{1}{2}\int \frac{d^3q}{(2\pi)^3}\frac{q_i}{E_q}\sum_s \left( \bar{u}_{\vec{q}}^s\,\gamma^i u_{\vec{q}}^s \right) = \frac{p_F^4}{4\pi^2}\,, \\
& P\delta^{ij} = \frac{1}{2}\int \frac{d^3q}{(2\pi)^3}\frac{1}{2E_q}\sum_s\left( q^i\bar{u}_{\vec{q}}^s\gamma^ju_{\vec{q}}^s+q^j\bar{u}_{\vec{q}}^s\gamma^iu_{\vec{q}}^s\right) \quad \implies \quad P = \frac{1}{3}\frac{p_F^4}{4\pi^2}\, .
}
In the last step, we have used \eqref{eq:SpinorRelations2} and integrated over a sphere of radius $p_F$.
Using the anti-commutation relations \eqref{eq:AntiCommRel}, as well as the definitions of the ground state \eqref{eq:FLGroundState} and the particle-hole state \eqref{eq:ParticleHoleState}, one can show that the only nontrivial identity involving creation-anihilation operator is given by:
\begin{align}
& \langle \text{FL} \vert c_{\vec{p}}^{s\dagger} \, c_{\vec{q}}^{s'} \vert\chi \rangle = (2\pi)^6\delta^{s'r}\delta^{r's}\delta^{(3)}(\vec{q}-\vec{p}_1)\delta^{(3)}(\vec{p}-\vec{p}_2)\,.
\end{align}
We define the following matrix elements:
\begin{equation}
\mathcal{T}^{\mu\nu}(x) \equiv \langle \text{FL} \vert T^{\mu\nu}(x) \vert\chi \rangle \, .
\end{equation}
A straightforward calculation yields:
\es{}{
\mathcal{T}^{00}(0)& = \frac{1}{4}\frac{1}{\sqrt{E_{p_1}E_{p_2}}} \left( E_{p_1}+E_{p_2}\right)\bar{u}_{\vec{p}_2}^{r'}\gamma^0 u_{\vec{p}_1}^r,\\
\mathcal{T}^{0i}(0) &= \frac{1}{8}\frac{1}{\sqrt{E_{p_1}E_{p_2}}} \left[\left( E_{p_1}+E_{p_2}\right)\bar{u}_{\vec{p}_2}^{r'}\gamma^i u_{\vec{p}_1}^r+\left(p_1^i+p_2^i\right)\bar{u}_{\vec{p}_2}^{r'}\gamma^0 u_{\vec{p}_1}^r\right] .
}
Using the relation \eqref{eq:SpinorRelations3}, we get:
\begin{align}
\label{eq:T00T0iNotSimplified}
\mathcal{T}^{00}\mathcal{T}^{0i^*}(\vec{p}_1,\vec{p}_2) &= \frac{1}{8}\frac{1}{E_{p_1}E_{p_2}}\left(E_{p_1}+E_{p_2} \right)^2\left(E_{p_2}p_1^i+E_{p_1}p_2^i \right)\\
& +\frac{1}{8}\frac{1}{E_{p_1}E_{p_2}}\left(E_{p_1}+E_{p_2} \right)\left(p_1^i+p_2^i \right)\left(E_{p_2}E_{p_1}+\vec{p}_1\cdot\vec{p}_2 \right), \nonumber\\
\label{eq:T00T00NotSimplified}
\mathcal{T}^{00}\mathcal{T}^{00^*}(\vec{p}_1,\vec{p}_2) &= \frac{1}{4}\frac{(E_{p_1}+E_{p_2})^2}{E_{p1}E_{p_2}}\left(E_{p_1}E_{p_2}+\vec{p}_1\cdot\vec{p}_2\right),
\end{align}
where we have defined $\mathcal{T}^{\mu\nu}\mathcal{T}^{\rho\sigma^*}(\vec{p}_1,\vec{p}_2) \equiv\sum_{r,r'}\mathcal{T}^{\mu\nu}(0)\mathcal{T}^{\rho\sigma^*}(0)$.
\subsubsection{Saturation of the Sum Rules}
Expanding \eqref{eq:T00T0iNotSimplified}, \eqref{eq:T00T00NotSimplified} in small $\vec{p}$ (where $\vec{p}\equiv \vec{p}_1-\vec{p}_2$), we get:
\begin{equation}
\label{eq:T00T0iExpan}
\mathcal{T}^{00}\mathcal{T}^{0i^*}(\vec{p},\vec{p}_2) = p_2p^i+\frac{(\vec{p}\cdot\vec{p}_2)p_2^i}{p_2}+2p_2p_2^i+O(p^2),
\end{equation}
and:
\begin{equation}
\label{eq:T00T00Expan}
\mathcal{T}^{00}\mathcal{T}^{00^*}(\vec{p},\vec{p}_2) = 2p_2^2+2\vec{p}_2\cdot\vec{p}+O(p^2).
\end{equation}
In order to evaluate the spectral density we need to integrate over $\vec{p}_2$. Note that in the limit of small $\vec{p}$, the energy associated with the state of momentum $\vec{p}$ is given by:
\begin{equation}\label{eq:energy}
E(\vec{p},\vec{p}_2) = E_{p_1}-E_{p_2} = p\cos(\theta)+O(p^2),
\end{equation}
where $\theta$ is the angle between the vectors $\vec{p}_2$ and $\vec{p}$.
\begin{figure}[!h]
\centering
\subfloat[]{\label{fig:FermiFig1a}{\includegraphics[width=0.25\textwidth]{Fermi_Sphere_integral_fig1.jpg}}} \qquad \qquad
\subfloat[]{\label{fig:FermiFig1b}{\includegraphics[width=0.25\textwidth]{Fermi_Sphere_integral_fig2.jpg}}}
\caption{An illustration describing the momentum space notation corresponding to the integral in the expression for the spectral density \eqref{eq:GeneralSpectralDensFromMatrixFer} for a specific configuration of $\vec{p}_1,\vec{p}_2$: as $p\to 0$, $p_1 \to p_F^+$, and $p_2+p\cos\theta \approx p_F$. Thus the variable $\delta p_2\equiv p_2-p_F$ ranges from $-p\cos\theta$ to $0$ in $p\to0$ limit. This will be useful in evaluation \eqref{eqs:int}.
Figure \ref{fig:FermiFig1a} describes a Fermi surface of radius $p_F$ in this specific configuration in momentum space, while figure \ref{fig:FermiFig1b} zooms in on the triangle $AOB$. }
\label{fig:FermiFig}
\end{figure}
The spectral function $\rho_{T^{00}T^{0\mu}}$ to the leading order is given by the following :
\begin{equation}
\label{eq:GeneralSpectralDensFromMatrixFer}
\rho_{T^{00}T^{0\mu}}(\omega,\vec{p})= \mathcal{G}_{T^{00}T^{0\mu}}(\omega,\vec{p})\mp \mathcal{G}^{*}_{T^{00}T^{0\mu}}(-\omega,-\vec{p}),
\end{equation}
where the $\mp$ sign takes the values of $-$ for $\mu=0$ and $+$ for spatial indices $\mu=i$, and $\mathcal{G}_{T^{00}T^{0\mu}}(\omega,\vec{p})$ involves the following integral (see figure \ref{fig:FermiFig}):
\begin{equation}
\begin{aligned}\label{eqs:int}
&\mathcal{G}_{T^{00}T^{0\mu}}(\omega,\vec{p}) \equiv\frac{1}{(2\pi)^3} \int d^3p_2\ \delta\left(\omega-E(\vec{p},\vec{p}_2)\right) \mathcal{T}^{00}\mathcal{T}^{0\mu*}(\vec{p},\vec{p}_2) \\\
&= \left[\frac{ p_F^2}{(2\pi)^2}\int_{0}^{\pi/2} d\theta\ \sin\theta\ \delta(p\cos\theta-\omega) \mathcal{T}^{00}\mathcal{T}^{0\mu*}(p,p_F,\cos\theta) \int_{-p\cos\theta}^0d\delta p_2\right]+\cdots\\
&=\frac{\chi_{[0,p]}(\om)}{4\pi^2}\frac{\omega p_F^2 }{p} \mathcal{T}^{00}\mathcal{T}^{0\mu*}(p,p_F, \omega/p).
\end{aligned}
\end{equation}
Note that $\mathcal{T}^{00}\mathcal{T}^{0\mu*}(\vec{p},\vec{p}_2)$ is a function of $p, p_2$ and the angle between the vectors, i.e.~of $\cos\theta$. At this point we change the integration variable from $p_2$ to $\delta p_2=p_2-p_F$ and from fig.~\ref{fig:FermiFig}, we read off the limit of integral in $p\to 0$ limit. For the same reason, we have kept only the leading order terms in $\delta p_2$ of $\mathcal{T}^{00}\mathcal{T}^{0\mu*}(\vec{p},\vec{p}_2)$ in the second line, which amounts to replacing $p_2$ with $p_F$ and we made the angle dependence explicit. We also use \eqref{eq:energy} inside the delta function, which subsequently sets $\cos\theta=\omega/p$ leading to the term $\mathcal{T}^{00}\mathcal{T}^{0\mu*}(p,p_F, \omega/p) \equiv \mathcal{T}^{00}\mathcal{T}^{0\mu*}(p,p_2, \cos\theta)\vert_{p_2=p_F,\,\cos\theta=\omega/p}$. The function $\chi_{[0,p]}(\om)$ is the characteristic function of the interval $[0,p]$.
We start with the calculation of $\rho_{T^{00}T^{00}}(\omega,p)$. From \eqref{eq:T00T00Expan}, we read: $
\mathcal{T}^{00}\mathcal{T}^{0\mu*}\left(p,p_F,\omega/p\right)= 2p_F^2+ 2p_F \omega $, then:
\begin{equation}
\begin{aligned}
\rho_{T^{00}T^{00}}(\omega,\vec{p})
&= \frac{\chi_{[-p,p]}(\om)}{2\pi^2}\frac{\omega p_F^4 }{p} +\cdots \underset{p\to 0}{\simeq}-(\epsilon+p)p^2\delta'(\omega)\,,
\end{aligned}
\end{equation}
as it should be, in accordance with \eqref{cons}.
Next, we turn to calculate $\rho_{T^{00}T^{0i}}$. For this purpose, it is convenient to define $\vec{p}=p\hat{z}$, and keep $\vec{p_2}$ arbitrary. From \eqref{eq:T00T0iExpan}, we read:
$\mathcal{T}^{00}\mathcal{T}^{0z*}(p,p_F,\omega/p) = p_Fp(1+\omega^2/p^2)+2p_F^2\omega/p\,$. Plugging it into the expression for the spectral density \eqref{eq:GeneralSpectralDensFromMatrixFer}, after covariantizing the result for $\rho_{T^{00}T^{0z}}$, we find the following expression for the spectral density $\rho_{T^{00}T^{0i}}$:
\begin{equation}
\begin{aligned}
\rho_{T^{00}T^{0i}}(\omega,\vec{p})\underset{p\to 0}{\simeq} \frac{\chi_{[-p,p]}(\om)}{2\pi^2}\,\frac{\omega^2 p_F^4 }{p^3}p^{i} =\chi_{[-p,p]}(\om) \frac{3(\epsilon+P)}{2}\,\frac{\omega^2 }{p^3}p^{i}\,.
\end{aligned}
\end{equation}
Using the above result, it is straightforward to check the saturation of the sum rules. One finds:
\begin{equation}
\begin{aligned}
\label{eq:SimplExprSumRule}
\frac{\partial \rho_{T^{00}T^{0i}}}{\partial p^j}(\omega,\vec{p})=\underset{p\to 0}{\simeq} (\epsilon+P)\delta(\omega)\delta^i_j,
\end{aligned}
\end{equation}
where we have used
$$\chi_{[-p,p]}(\om)
\left(\frac{\omega^2}{p^3}\right) =\left[\Theta(\omega+p)-\Theta(\omega-p)\right]
\left(\frac{\omega^2}{p^3}\right) \underset{p\to 0}{\simeq} \frac{2}{3} \delta(\omega)\,.$$
Therefore, the sum rule associated with the broken boosts \eqref{SumRule3} is satisfied.
Using \eqref{eq:SimplExprSumRule} one easily finds:
\begin{equation}
\frac{\partial \rho_{T^{00}T^{0i}}}{\partial p^i}(\omega,\vec{p})\underset{p\to 0}{\simeq}\,3(\epsilon+P)\delta(\omega) = 4\epsilon\delta(\omega)\,,
\end{equation}
thus, the sum rule \eqref{SumRuleD2} associated with the broken dilatations is satisfied (with $d=4$).
While in this section we have been studying the CFT of a free fermion, the analysis is in fact applicable to any (possibly interacting) CFT state around which the effective theory is a free Fermi surface. To our knowledge it is not presently known if a such a free Fermi surface is a natural end point under the RG evolution around heavy states. It would be interesting to investigate it along the lines of~\cite{Polchinski:1992ed,Shankar1994}.
\section{2d CFTs at Large Charge} \label{sec:boostbreak}
\subsection{Boost Breaking in 2d CFTs}
We will consider 2d CFTs on the cylinder with circle of radius $R$:
$ds^2=d\tau^2+R^2(d\theta)^2$, with $\theta\simeq \theta+2\pi$.
The advantage of the 2d setup is that we can construct the correlation functions of the EM tensor explicitly and verify the existence of the large volume (macroscopic) limit. The low-energy states responsible for the boost Nambu-Goldstone theorem can be also identified. Remarkably, many of the things we find are similar to the superfluid discussion in section~\ref{sec:superfluid}.
\subsection*{TJ correlator}
We take an arbitrary state $|\Omega\rangle$ which corresponds to a spinless highest weight state in the Verma module with dimension $\Delta$. Let $\Phi$ be some primary and consider first the four-point function in flat space ($h_\Phi,\bar h_\Phi, h_\Omega,\bar h_\Omega$ stand for the obvious scaling dimensions and we assume $h_\Omega=\bar h_\Omega$.)
\begin{equation}\label{eq:TPhi}
\langle\Omega(0)T(z) \Phi(1) \Omega(\infty)\rangle={Ch_\Omega\over z^2}+{Ch_\Phi\over z(z-1)^2}~,
\end{equation}
where
\begin{equation}
\langle\Omega(0) \Phi(1) \Omega(\infty)\rangle=C~.
\end{equation}
In order to transform this to the cylinder with a circle of radius $R$ we need to plug $z= e^{u/R}$ and take $T(z)\to R^2 z^{-2} T(u)$ and $\Phi(1)\to R^{ \Delta_\Phi} \Phi(0)$.\footnote{The transformation of the EM tensor $T(z)\to R^2 z^{-2} T(u)$ is missing a constant -- the famous ground state energy from the Schwartzian. This is unimportant for us because we are considering the correlation functions in heavy states and hence we drop this constant. } Therefore the following Euclidean correlation function is found on the cylinder with coordinate $u=\tau+iR\theta$:
$$\langle\Omega|T(u) \Phi(0) |\Omega\rangle={Ch_\Omega\over R^{2+\Delta_\Phi}}+{Ch_\Phi\over 2R^{2+\Delta_\Phi} (\cosh(u/R)-1)}~.$$
The constant piece is necessary to account for the propagation of the state $|\Omega\rangle$.
Let us now investigate the macroscopic limit. For concreteness, we take $\Delta_\Phi=1$ and $\Phi$ to be a conserved current $(1,0)$ operator. $C/(\pi R)=\rho$ is the charge density in the state $|\Omega\rangle$ while $h_\Omega/(\pi R^2)=\epsilon$ is the energy density. Evidently, to achieve a nontrivial macroscopic limit, we need to take $C$ to scale as a positive power of $R$ -- i.e., we change the state $\Omega$ as a function of $R$ such that $C\over R$ is finite in the $R\to\infty$ limit. This is the same as having a constant charge density. Secondly, the constant ${Ch_\Omega\over R^{3}}$
is proportional to $\rho\epsilon$ which we should also hold fixed. Then the macroscopic limit becomes
$$\langle\Omega|T(u)J(0) |\Omega\rangle={\pi^2}\rho\epsilon+{\pi \rho\over u^2}~.$$
Similarly,
$$\langle\Omega|T(u)\bar J(0) |\Omega\rangle={\pi^2}\rho\epsilon~.$$
In components this becomes (assuming $C$ is real) at separated points in Euclidean signature (where $T=\pi(T_{tt}-T_{tx})$, $\bar T=\pi(T_{tt}+T_{tx})$, similary $J=\pi(J_t-J_x)$, $\bar J=\pi(J_t+J_x)$):\footnote{To conform with both the conventions of the 2d CFT literature and the usual notion of $T_{\mu\nu}$ in higher dimensions, we use the definitions $T\equiv -{4\pi\ov\sqrt{g}} \,{\de S\ov \de g^{uu}}$ and $T_{\mu\nu}\equiv {2\ov\sqrt{g}} \,{\de S\ov \de g^{\mu\nu}}$.}
\begin{equation}
\begin{aligned}
&\langle\Omega|T_{tt}(\tau,x) J_t(0)|\Omega\rangle=\rho\epsilon+\frac{\rho}{2\pi}{(\tau^2-x^2)\over (\tau^2+x^2)^2}\,,\\
&\langle\Omega|T_{xt}(\tau,x) J_x(0)|\Omega\rangle=\frac{\rho}{2\pi}{\tau^2-x^2\over (\tau^2+x^2)^2}\,,\\
&\langle\Omega|T_{tt}(\tau,x) J_x(0)|\Omega\rangle=-i\frac{\rho}{\pi}{\tau x\over (\tau^2+x^2)^2}\,,\\
&\langle\Omega|T_{xt}(\tau,x) J_t(0)|\Omega\rangle=-i\frac{\rho}{\pi}{\tau x\over (\tau^2+x^2)^2}\,.
\end{aligned}
\end{equation}
We do not write the matrix elements containing $T_{xx}$ as they are the same up to a sign as those containing $T_{tt}$.
Analytic continuation to Minkowski signature is implemented in the equations above by setting $\tau=it \pm \epsilon$ with $\epsilon\to0^{+}$. The $\pm$ correspond to different (Minkowski) time ordering of the operators:
\begin{equation}
\begin{aligned}
&\langle\Omega|T_{tt}(t,x) J_x(0)|\Omega\rangle=- \frac{\rho}{2\pi}{tx\over x^2+t^2}\left({\partial\over \partial x}{1\over t+x-i\epsilon}-{\partial\over \partial x}{1\over t-x-i\epsilon}\right)\\
&\langle\Omega| J_x(0)T_{tt}(t,x)|\Omega\rangle=- \frac{\rho}{2\pi}{tx\over x^2+t^2}\left({\partial\over \partial x}{1\over t+x+i\epsilon}-{\partial\over \partial x}{1\over t-x+i\epsilon}\right)
\end{aligned}
\end{equation}
Now we use the familiar identity (with $\epsilon\to 0^+$ assumed): ${1\over u+i\epsilon} - {1\over u-i\epsilon} =-2\pi i \delta (u)$ to find an expression for the commutator in position space:
\begin{equation}
\langle\Omega|[T_{tt}(t,x), J_x(0)]|\Omega\rangle= -i\rho{tx\over x^2+t^2}\left(\delta'(t+x)+\delta'(t-x)\right)~.
\end{equation}
The support of the commutators on the light-cone is of course due to the non-dissipative nature of the excitations pertinent to this problem. In frequency and momentum space we find
\begin{equation}
\begin{aligned}
\langle\Omega|[T_{tt}(\om,k), J_x(0)]|\Omega\rangle&= \int_{-\infty}^{\infty} dt\,dx\ e^{i\om t-i kx} \langle\Omega|[T_{tt}(t,x), J_x(0)]|\Omega\rangle \\
&= -\pi \rho k (\delta(\om+k)+\delta(\om-k))~\,.
\end{aligned}
\end{equation}
The spectral density is given by
\es{TttJxFunny}{
\rho_{T_{tt}J_x}(\om, k) =-\frac{1}{2} \rho k (\delta(\om+k)+\delta(\om-k)) \underset{k\to0}{\simeq} -\rho k \delta(\om)\,,
}
as required by the sum rule \eqref{SumRuleJ} for the boost. (We remind the reader that in our conventions $g_{xx}=-1$, hence $k_x=-k^x=-k$, which explains the sign in \eqref{TttJxFunny}.)
Similarly, one can show that
\begin{equation}
\rho_{T_{xt}J_t}(\om, k)=- \frac{1}{2} \rho k (\delta(\om+k)+\delta(\om-k)) \underset{k\to0}{\simeq} -\rho k \delta(\om)\,,
\end{equation}
and verify the sum rule \eqref{SumRuleD2} for the dilatation.
\subsection*{TT correlator}
The correlator involving the energy momentum tensor is given by
\begin{equation}
\langle\Omega(0)T(z) T(1) \Omega(\infty)\rangle={h^2_\Omega\over z^2}+{2h_\Omega\over z(z-1)^2}+{c/2\over (z-1)^4}~.
\end{equation}
This leads to an amplitude on the circle of radius $R$:
\es{ToRep}{
\langle\Omega|T(u) T(0) |\Omega\rangle={h^2_\Omega\over R^{4}}+{h_\Omega\over R^{4} (\cosh(u/R)-1)}+{c\over 8R^4(\cosh(u/R)-1)^2}~.
}
The macroscopic limit requires to keep $h_\Omega/R^2\equiv \pi\epsilon$ fixed and $c$ fixed. Then we obtain in the macroscopic limit:
\es{Aim}{
\langle\Omega|T(u) T(0) |\Omega\rangle={\pi^2\epsilon^2}+{2\pi\epsilon\over u^2}+{c\over 2u^4}~.
}
Similarly
$$\langle\Omega|T(u) \bar T(0) |\Omega\rangle={\pi^2\epsilon^2}~.$$
We can now extract the energy-density correlator with itself as before by inserting $T=-2\pi T_{uu}=-\pi(T_{\tau \tau}-iT_{\tau x})=\pi (T_{tt}-T_{tx})$ and expanding to find:
\begin{equation}
\langle\Omega|T_{tt}(\tau,x) T_{tt}(0) |\Omega\rangle=\epsilon^2+\frac{2\epsilon (\tau^2-x^2)}{ \pi(x^2+\tau^2)^2}+\frac{c}{4\pi^2}\frac{(x^4+\tau^4-6x^2\tau^2)}{4\pi^2(x^2+\tau^2)^4}
\end{equation}
The commutator is then given by (similar to the calculation of $TJ$ commutator)
\begin{equation}\langle\Omega|[T_{tt}(t,x), T_{tt}(0)] |\Omega\rangle= i\epsilon\left(\delta'(x+t)-\delta'(x-t)\right) -i \frac{c}{24\pi}\left(\delta'''(x+t)-\delta'''(x-t)\right)\,.
\end{equation}
This can be now transformed to frequency and momentum space to find
\begin{equation}
\begin{aligned}
\langle\Omega|[T_{tt}(\om,k), T_{tt}(0)] |\Omega\rangle&= \int_{-\infty}^{\infty} dt\, dx\ e^{i\om t-i kx} \langle\Omega|[T_{tt}(t,x), T_{tt}(0)]|\Omega\rangle \\
&=2\pi\epsilon k\left(\delta(\om-k)-\delta(\om+k)\right) +c\pi k^3 /6\left(\delta(\om-k)-\delta(\om+k)\right)~.
\end{aligned}
\end{equation}
The spectral density is given by
\es{spectralsmallk}{
\rho_{T_{tt}T_{tt}} \underset{k\to 0}{\simeq} -2\epsilon k^2\delta'(\omega) - \frac{ k^4}{6}\le[c\,\delta'(\omega)+2\ep\, \delta'''(\omega)\ri]+\cdots\,.
}
The sum rule \eqref{t00t00} is saturated by the first term, as $\epsilon+P=2\epsilon$ in $2$ dimension.
To understand which terms contribute to the sum rule we must study in detail the intermediate states by inserting a complete set of states. It is obvious that the only states $\langle\Omega'|$ for which $\langle\Omega'|T_{tt}(0) |\Omega\rangle\neq0$ are in the Verma module of $|\Omega\rangle$ (left or right descendants but not both). The usual basis of states $L_{-N}^{n_N}\cdots L_{-1}^{n_1}|\Omega\rangle$ is inconvenient to use since it is not orthonormal and the matrix elements are difficult to compute. Instead we use the oscillator basis of~\cite{zamolodchikov1986two}, nicely reviewed in~\cite{Besken:2019bsu} (we are using their notations) and we start by computing the wave function
\es{Ubasis}{
\langle U| T(u) |\Omega\rangle&={1\over R^2}\sum_{n\leq 0}e^{-nu/R} \langle U| L_n |\Omega\rangle ={1\over R^2}\sum_{n\leq 0}e^{-nu/R}{\cal L}_{-n}\cdot 1 \\
&={1\over R^2}\left[h-\sum_{k=2}^\infty e^{-ku/R}\sum_{p=1}^{k-1}p(k-p)u_{p}u_{k-p}+2\sum_{k=1}^\infty e^{-ku/R} k(\mu k -i\lambda) u_k\right]~.}
where $u$ is our usual coordinate on the cylinder of radius $R$ and $u_k$ is an infinite set of variables collectively denoted $U$. $\mu,\lambda$ are related to $h_\Omega$ and $c$ via $c=1+24\mu^2$, $h_\Omega=\lambda^2+\mu^2$. In this basis the monomials are orthogonal with norm
$$(1,1)=1~,\quad (u_k,u_l)=\delta_{k,l}S_{k,1}~,\quad (u_k^2,u_l^2)=\delta_{k,l}S_{k,2}~,\quad (u_ku_p,u_qu_{l})=
\delta_{k,q}\delta_{p,l} S_{k,1}S_{l,1}+\delta_{k,l}\delta_{p,q} S_{k,1}S_{l,1}~. $$
In the last term we assumed that the four indices are not all the same. We used $S_{k,j}=(2k)^{-j}\Gamma(j+1)$. We can unify the third and fourth formulas and write:
$$(1,1)=1~,\quad (u_k,u_l)=\delta_{k,l}{1\over 2k}~,\quad (u_ku_p,u_qu_{l})=
\left(\delta_{k,q}\delta_{p,l}++\delta_{k,l}\delta_{p,q}\right) {1\over 4kl}~. $$
Inserting a complete set of states, we find that only need to insert ``single particle'' and ``two particle'' states:
$$\langle\Omega|T(u) T(0) |\Omega\rangle=\langle\Omega|T(u)|\Omega\rangle\langle\Omega|T(0) |\Omega\rangle+\sum_{k=1} 2k\langle\Omega|T(u)|u_k\rangle\langle u_k|T(0) |\Omega\rangle$$ $$+4\sum_{k< l}kl \langle\Omega|T(u)|u_ku_l\rangle\langle u_ku_l|T(0) |\Omega\rangle+2\sum_{k}k^2 \langle\Omega|T(u)|u_k^2\rangle\langle u_k^2 |T(0) |\Omega\rangle~.$$
The first term on the right hand side gives $h/R^4$. The second term gives $${4\over R^4} \sum_{k=1}^\infty 2k e^{-ku/R} k(\mu k -i\lambda) {1\over 2k} k(\mu k +i\lambda) {1\over 2k}={2\over R^4} \sum_{k=1}^\infty e^{-ku/R} k(\mu^2 k^2 +\lambda^2)~.$$
This is easily summed to give
$${1\over R^4}(\lambda^2+\mu^2){1\over (\cosh(u/R)-1)}+{3\over R^4} \mu^2 {1\over (\cosh(u/R)-1)^2 }~.$$
This accounts for almost the whole answer in \eqref{ToRep}, except that the coefficient of the second term above is off. (In the full answer it is $c/(8 R^4)=(3\mu^2+1/8)/R^4$ instead of $3\mu^2/R^4$ that we obtained from the one particle exchange.)
One can check that the difference is made up by the two-particle states. As we have seen above in~\eqref{Aim}--\eqref{spectralsmallk}, the first term,
${1\over R^4}(\lambda^2+\mu^2){1\over (\cosh(u/R)-1)}$ is enough to saturate the boost Nambu-Goldstone theorem in the macroscopic limit. Therefore, the boost Nambu-Goldstone theorem is saturated by one-particle states $|\Omega\rangle$ and $|u_k\rangle$.
Note that of these states, only $|u_1\rangle$ is an ordinary conformal descendant of $|\Omega\rangle$ (indeed, it is proportional to the action of ${\cal L}_{-1}$ on $|\Omega\rangle$).
In frequency space the commutator directly on the cylinder as a result of these one-particle exchanges is (we denote by $p=n/R$, with $n$ an integer, the momentum on the circle of radius $R$)
\begin{equation}\label{sd}\langle\Omega|[T(\om,p),T(0)] |\Omega\rangle= {2n\over R^4} (\mu^2n^2+\lambda^2)\, \delta\le(\om-{n\ov R}\ri)+\cdots
\end{equation}
All the states that appear in the intermediate channel have energy and momentum that are related by $\om_n=p_n=n/R$. This is because they are excitations of the ground state given by the action of a holomorphic EM tensor.
As always, the spectral density, being a sum of delta functions, needs some smearing before it can be written in the infinite volume limit. By contrast, correlation functions for $u\ll R$ land themselves to a nice macroscopic limit more directly.
From the spectral density~\eqref{sd} we see that any one individual state, even if it has $n\sim R$ and $\lambda^2\sim R^2$ as required in the macroscopic limit, leads to a vanishing spectral density in the macroscopic limit (the contribution is suppressed as $1/R$). As was discussed in detail in section~\ref{subsec:gap}, we have to smear over a band of states centered around $n\sim R$ and $\lambda^2\sim R^2$ to recover the correct result.
Note the nice analogy to the superfluid effective theory and mean field theory: in \eqref{eqs:T00T00s}
and \eqref{eqs:T00J0free} we have contributions from what is analogous to one-particle states in the Verma module
$|u_n\rangle$, of which $|u_1\rangle$ is a conformal descendant of the ground state. These are sufficient to reproduce the boost sum rule. The contribution comes from the states with energy $n\sim R$ while the ground state has energy $\sim R^2$.
\subsection{A Compact Boson Effective Theory}
It is tempting to try and reproduce the above results with an effective theory. One candidate is the superfluid EFT of section~\ref{sec:superfluid} specialized to $d=2$, which could apply to the lowest energy state at fixed (large) $U(1)$ charge. In CFTs with a discrete operator spectrum it is a fundamental result \cite{Affleck:1985jc} that the $U(1)$ symmetry is enhanced to a $u(1)\times u(1)$ Ka\v{c}-Moody algebra, and the Energy-Momentum tensor of the full theory decomposes to two separately conserved Energy-Momentum tensors.
Below we present an effective theory in $d=2$ with the following properties:
\begin{itemize}
\item The $U(1)$ symmetry is of course not spontaneously broken -- it is in the usual ``log-ordered'' phase which is common in $d=2$.
\item It has one compact boson but it allows for an energy-momentum tensor with {\it arbitrary} central charge.
\item The $U(1)$ symmetry can be promoted to a Ka\v{c}-Moody symmetry only if the central charge is $c=1$.
\end{itemize}
Besides describing the compact boson large charge limit, which is somewhat trivial, the theory we present is potentially interesting for situations
where the $U(1)$ symmetry of a CFT does not enhance to Ka\v{c}-Moody. Such CFTs must have a continuous spectrum\footnote{The simplest such example is a noncompact complex scalar $\Phi$, with the $U(1)$ symmetry rotating around the origin of the target space $\mathbb{C}$ and the associated current
\es{J}{
J_\mu=i\le(\Phi^\dagger\p_\mu\Phi-\Phi\p_\mu\Phi^\dagger\ri)
} which does not get enhanced to a Ka\v{c}-Moody symmetry.} and one might worry that the large charge limit would be necessarily more complicated.
A more realistic and interesting application for our EFT is to describe superfluids with a boundary, or equivalently, the large charge limit of 3d boundary CFT (BCFT). Indeed, the aforementioned shift in the central charge of a single compact boson will be crucial for matching the boundary trace anomaly of the superfluid. For some literature on boundary trace anomalies see~\cite{Henningson:1999xi, Schwimmer:2008yh,Solodukhin:2015eca,Jensen:2015swa,Herzog:2015ioa,Herzog:2017kkj,Wang:2021mdq}.
We leave the development of this direction to future work.
\subsubsection{First Look at the Large Charge Effective Theory }
Let us start with the EFT \eqref{eqs:schi} specialized to $d=2$:\footnote{By $d^2x$ we mean $d\tau dx=\frac12 du d\bar{u}$.}
\begin{equation}\label{leadingterm}
S_0={\kappa\ov \pi} \int d^2x\ \partial \varphi\bar \partial \varphi~,
\end{equation}
where $\varphi$ is a compact scalar $\varphi\sim \varphi+2\pi$.
We expand the theory around
$\varphi= -i\mu \tau+\pi$ such that $\kappa\mu/(2\pi)=\rho$ and $\kappa \mu^2/(4\pi)=\epsilon$ and try to match the correlators we found from the theory for the fluctuations: $\kappa/\pi \int d^2x\ \partial\pi\bar\partial\pi$. The stress tensor of the theory $S_0$ in \eqref{leadingterm} is $T=-\ka (\p\varphi)^2$, which in terms of the fluctuations takes the form
\es{leadingT}{
T={\kappa\mu^2\ov 4}+i\kappa\mu \partial\pi-\kappa (\partial\pi)^2\,.
}
The stress tensor two point function can be computed using the propagator
\es{PropCorr}{
\langle\partial \pi(u)\partial\pi(0)\rangle = -{1\over 2\kappa u^2}\,,
}
and we get
\es{TTleading}{
\langle T(u)T(0)\rangle ={\pi^2\epsilon^2}+{2\pi\epsilon\over u^2}+{1\over 2 u^4} \,.
}
The first two pieces are exactly right but the third one is not. This is because our effective theory has central charge 1 instead of $c$. We have so far merely reproduced the known result that the compact boson has central charge 1.
If we want to reproduce \eqref{TTleading} exactly, we face the question of how to make a single compact boson $\varphi$ have $c\neq 1$, which seems at first sight impossible. A similar in spirit problem arises in the quantization of the effective string~\cite{Polchinski:1991ax} and the solution here is similar. We are allowed to add singular terms to the effective action since we are anyhow expanding around a nontrivial background:
\es{HigherDer}{
S_1={\beta\ov \pi} \int d^2x\ {\partial^2\varphi\bar\partial^2\varphi\over \partial\varphi\bar\partial\varphi}~.
}
$\beta$ will turn out to be proportional to the shift in central charge. Expanded about the superfluid solution we find that this leads to a contribution to the effective action of the fluctuations
\es{HigherDer2}{
S_1=-{4\beta\over \pi\mu^2} \int d^2x\ \partial^2\pi\bar\partial^2\pi+O(1/\mu^{4})\,.
}
On the one hand, as will be discussed in the next section, $S_1$ is somewhat trivial; one manifestation of this is that it does not lead to a modification of the propagator \eqref{PropCorr} up to $O(1/\mu^{4})$ (this is true since $S_1$ vanishes on shell).
On the other hand, the conformal symmetry of the action $S=S_0+S_1$ is modified: with precision $O(1/\mu^{3})$ the deformed action $S$ is invariant under the corrected conformal transformations
\es{CorrectedTF}{
\de \varphi(u,\bar{u})=\lam(u)\p \varphi-{\beta\ov2\ka}\,{\p^2\lam(u)\ov {\p}\varphi}\,,
}
(there is an independent antiholomorphic copy of the symmetry) and the stress tensor that generates this symmetry is
\es{CorrectedEMT}{
T&=-{\ka}(\p\varphi)^2+\beta\,{\p\varphi\,\p^3\varphi-(\p^2\varphi)^2\ov(\p\varphi)^2}+O(1/\mu^{3})\\
&={\ka \mu^2\ov 4}+i\ka\mu\p\pi-\ka(\p\pi)^2+{2i\beta\ov \mu}\,\p^3\pi+O(1/\mu^{2})\,.
}
The stress tensor two-point function can now be straightforwardly computed using Wick contractions using the propagator \eqref{PropCorr}. We get
\es{TTsubleading}{
\langle T(u)T(0)\rangle =\eqref{TTleading}+{12\beta\over u^4} +O(1/\mu)\,.
}
If we set $\beta=(c-1)/24$, we recover \eqref{Aim} to $O(1/\mu)$.
To work out the predictions of the EFT to higher orders, we need a more systematic approach, which we turn to next.
We will also see below that there is no holomorphic current, i.e.~a weight $(1,0)$ primary operator of the Virasoro symmetry generated by the deformed $T$ of \eqref{CorrectedEMT} unless $\beta=0$.
\subsubsection{Systematic Development of the Effective Theory}
There is a systematic procedure to construct all terms allowed by symmetry in the effective action. We define the Weyl invariant metric $\hat{g}_{\mu\nu}\equiv g_{\mu\nu} \abs{\p \varphi}^2$, where $\abs{\p \varphi}^2\equiv-g^{\mu\nu}\p_\mu\varphi\p_\nu\varphi$. Then in the derivative expansion we can write the following terms:
\es{derivexp}{
S_\text{deriv}={\kappa\ov \pi} \int d^2 x\ \sqrt{\hat{g}}\le[1+ \al_{4,1} \hat{R}^2+\al_{4,2} \hat\nabla^\mu\hat\nabla^\nu\hat{R} \,\p_\mu \varphi\p_\nu \varphi+\dots\ri]\,,
}
where $\alpha_{k,i}$ is the coefficient of the $i$th term at $k$th derivative order. We used the leading order equation of motion $\nabla^2 \varphi=0$, that in $2d$ the Riemann tensor has only one independent component, $\hat{R}$, and that $\int d^2 x\sqrt{\hat{g}}\hat{R}$ is a topological invariant, the Euler characteristic of the manifold to reduce the number of terms in \eqref{derivexp}. There is one famous term that is missing from $S_\text{deriv}$, since it is not a local Weyl invariant in itself, but transforms with a shift that is a total derivative. The Wess-Zumino term takes the form~\cite{Polyakov:1981rd,Komargodski:2011xv}:
\es{derivexp2}{
S_\text{WZ}&=\al_2 \int d^2 x\ \sqrt{{g}}\le[(\p_\mu\tau)^2+\tau R\ri]\,,\\
\tau&\equiv -\log \abs{\p \varphi}\,.
}
Note that $\tau$ here is a composite dynamical field (as opposed to a background field, which is the more common case in the literature).
Let us now take $g_{\mu\nu}$ to be flat. In complex coordinates the leading order equation of motion is $\p\bar\p\varphi=0$, and by dropping terms proportional to it, we realize that $S_\text{WZ}=S_1$ from \eqref{HigherDer} with $\beta=\pi\al_2/4$. So we make contact with the considerations in the previous section. The equation of motion
also implies that {\it on shell} $\varphi(u,\bar{u})=\chi(u)+\bar\chi(\bar u)$ and
\es{FlatHat}{
d\hat{s}^2&=\hat{g}_{\mu\nu}dx^\mu dx^\nu\\
&=-\p\chi\bar\p\bar\chi \, g_{\mu\nu}dx^\mu dx^\nu\\
&=-\p\chi\bar\p\bar\chi \, dz d\bar{z}\\
&=-d\chi d\bar\chi\,,
}
i.e.~$\hat{g}_{\mu\nu}$ is flat. This then implies that on-shell all curvature invariants that we used to build $S_\text{deriv}$ in \eqref{derivexp} vanish. In addition, as we have seen before, $S_\text{WZ}$ is also a total derivative modulo the equations of motion. The EFT is nontrivial despite the action having no terms which are nonzero on-shell. This is because the EM tensor could (and should) receive various corrections.
Since all the higher terms in the effective action vanish on-shell, we can use a powerful general result in EFTs that there exists a field redefinition $\varphi\to\tilde \varphi$, that makes the action quadratic.
\es{derivexp3}{
S={\kappa\ov \pi} \int d^2 x\ \p\tilde \varphi\bar\p\tilde\varphi\,.
}
We can then expand around the superfluid background by taking $\tilde \varphi=-i\mu\tau+\tilde \pi$.
The relation between $\pi$ and $\tilde \pi$ is (see also \cite{Drummond:2004yp}):
\es{FieldRedef}{
\pi=\tilde \pi-{2\beta\ov\ka \mu^2}\, \p\bar\p \tilde \pi+O(1/\mu^{3})\,.
}
The symmetry transformation of \eqref{CorrectedTF} and the corresponding stress tensor \eqref{CorrectedEMT} become
\es{CorrectedTF2}{
\de \tilde \pi(u,\bar{u})=&-{i\mu\ov2}\,\lam(u)-{i\beta\ov2\ka\mu}\,{\p^2\lam(u)}\\
&+\lam(u)\p \tilde \pi+O(1/\mu^{2})\,,\\
T(u)=&{\ka \mu^2\ov 4}+i\ka\mu\p\tilde\pi+{2i\beta\ov \mu}\,\p^3\tilde\pi\\
&-\ka(\p\tilde\pi)^2+O(1/\mu^{2})\,,
}
where in the first lines we collected terms that shift $\tilde \pi$ and hence their generators are linear in $\tilde \pi$, while the second lines correspond to conformal transformations. (In the expression of $T$ we dropped terms proportional to the equation of motion.) To $O(\mu^0)$ the symmetry is just a combination of the shift and conformal symmetry of the free compact boson, and correspondingly $T$ is just a sum of the conventional current and stress tensor. At higher orders the symmetry and its generator becomes more exotic. To the order we wrote down formulas the computation of the stress tensor correlator is identical in the $\tilde\pi$ and $\pi$ variables, but the introduction of $\tilde\pi$ streamlines the computation at higher orders.
In summary, we are faced with the problem of constructing a tensor in a derivative expansion from a free scalar governed by the action $S=\kappa/\pi \int d^2 x\ \p\tilde \pi\bar\p\tilde\pi$. The stress tensor is supposed to obey the OPE
\es{TTOPE}{
T(u)T(0)={c\ov 2 u^4}+{2\ov u^2}\,T(0)+{1\ov u}\,\p T(0)+\text{regular}\,,
}
which can be achieved order by order in $1/\mu$. The remarkable fact is that $c$ is tunable.\footnote{There is another stress tensor $T=\ka(\p\phi)^2+V\p^2\tilde\phi$ (that of the linear dilaton CFT) that produces a tunable central charge $c=1+6V^2/\ka$ from a free scalar action for the noncompact scalar $\phi$; our setup with a compact $\varphi$ is different as a linear dilaton term is forbidden. } We could have started from this formulation of the problem, but for physical intuition and to make contact with the literature, we took a detour.
Using the {\it Mathematica} package OPEdefs \cite{Thielemans:1991uw}, by imposing the stress tensor OPE, we found that the first few orders of the stress tensor are:
\es{CorrectedEMT2}{
T(u)=&{\ka \mu^2\ov 4}+i\ka\mu\p\tilde\pi-\ka(\p\tilde\pi)^2+{1 \ov \mu}\,\le[{i(c-1)\ov 12}\,\p^3\tilde\pi+\ga_1\, \p\tilde\pi \p^2\tilde\pi \ri]\\
&+{1 \ov \mu^2}\,\le[{(c-1)\ov 6}\,\le((\p^2\tilde\pi)^2+\p\tilde\pi\p^3\tilde\pi\ri)+\ga_2 \,\p^4\tilde\pi+\ga_3\, (\p\tilde\pi)^2 \p^2\tilde\pi \ri]+O(1/\mu^3)\,,
}
where $\ga_i$ are arbitrary coefficients (they are not quite Wilson coefficients, since, as we remarked, the action does not admit terms beyond the free kinetic term). We now attempt to construct a $(1,0)$ holomorphic primary in the $1/\mu$ expansion by imposing the OPEs:
\es{JOPE}{
T(u)j(0)&={1\ov u^2}\, j(0)+{1\ov u}\, j'(0)+\text{regular}\,,\\
j(u)j(0)&={\ka\ov 2u^2}+\text{regular}\,.
}
We succeed to $O(\mu^0)$, but at $O(1/\mu)$ the most general Ansatz
\es{JAnsatz}{
j={\ka\mu\ov 2}+i\ka \p\tilde\pi +{1 \ov \mu}\,\le[\lam\,\p^3\tilde\pi+\ga_1\, \p\tilde\pi \p^2\tilde\pi \ri]+O(1/\mu^2)
}
leads to a contradiction with \eqref{JOPE}: the OPE with $T$ wants to set $\lam=-i(c-1)/12$, while the OPE with $j$ to $\lam=0$. (Of course these are consistent for $c=1$.)
This is how the general theorem of \cite{Affleck:1985jc} manifests itself in our concrete computation. The absence of a holomorphic current in effective string theory is due to similar reasons~\cite{Polchinski:1991ax}.
\subsubsection{Ground State Energy}
While there are undetermined coefficients in the stress tensor~\eqref{CorrectedEMT2}, it turns out that the vacuum energy is universal in this 2d EFT to all orders in the large charge expansion.
The argument consists of two simple steps. First we note that while the conformal transformations implemented by $T$ in \eqref{CorrectedEMT2} are exotic (as displayed in \eqref{CorrectedTF2}), $L_0+\bar{L}_0$ generates ordinary time translations on the cylinder. This can be seen either from the transformation law it generates for constant $\lam(u)$ in \eqref{CorrectedTF2}, or by noticing that all higher order terms in $1/\mu$ are total derivatives,
\es{Ttotder}{
T(u)=&{\ka \mu^2\ov 4}+i\ka\mu\p\tilde\pi-\ka(\p\tilde\pi)^2+\p \psi\,,\\
\psi\equiv&{1 \ov \mu}\,\le[{i(c-1)\ov 12}\,\p^2\tilde\pi+{\ga_1\ov 2}\, (\p\tilde\pi)^2 \ri]+{1 \ov \mu^2}\,\le[{(c-1)\ov 6}\,\p\tilde\pi \p^2\tilde\pi+\ga_2 \p^3\tilde\pi+{\ga_3\ov 3}\, (\p\tilde\pi)^3 \ri]+O(1/\mu^3)\,.
}
The cylinder partition function in the fixed $Q$ sector can be written as
\es{cylZ}{
Z_Q[\beta]&=\Tr_\text{$Q$ fixed}\le[e^{-{\beta\ov R}\le(L_0+\bar L_0-{c\ov 12}\ri)}\ri]\\
&=\int_\text{$Q$ fixed} D\tilde\varphi\ e^{-S_\text{free}[\tilde\varphi]}\,.
}
While our manipulations above were in classical field theory, we can argue that keeping track of the Jacobian of the field redefinition $\varphi \to \tilde \varphi$ would not change the conclusion that there exists a field redefinition that makes the theory free. The Jacobian is a local Weyl invariant functional of $\varphi$, hence can be exponentiated and written in terms of $\hat{g}_{\mu\nu}$. Since we have written all these terms in the action \eqref{derivexp}, keeping track of the Jacobian only changes coefficients in the action. Then there must exist a field redefinition that makes the action free. This argument is reminiscent of the classic argument of \cite{David:1988hj,Distler:1988jt}.
In the $\beta\to\infty$ limit from the representation as a trace, we see that we pick up the ground state energy $\exp\le[-{\beta\ov R}\le(\De_Q-{c\ov 12}\ri) \ri]$. Evaluating the path integral at fixed charge gives
\es{Zpath}{
Z_Q[\beta]&\to \exp\le[-\beta \le({\ka \mu^2 R\ov 2}-{1\ov 12R}\ri) \ri]\,,
}
where the second term is the Casimir energy of the free real scalar. Using the relation \eqref{eqs:densitymu} with $c_1=\ka/(4\pi)$, we conclude that
\es{DimPred}{
\De_Q={Q^2\ov 2 \ka}+{c-1\ov 12}+O(e^{-Q})\,.
}
We hope that this prediction can be tested in a situation where our effective theory would apply (e.g. in the context of BCFT).
\pagebreak
We note that there exists another method for computing the dimension of the lowest dimension large charge scalar, analogous to the approach of~\cite{Drummond:2004yp}. This method gives the same result as~\eqref{DimPred}.
\section*{Acknowledgements}
We thank Gabriel Cuomo, Bruno Le Floch, Petr Kravchuk and Cumrun Vafa for very useful discussions. ZK, MM, and ARM are supported in part by the Simons Foundation grant 488657 (Simons Collaboration on the Non-Perturbative Bootstrap) and the BSF grant no. 2018204. The work of ARM was also supported in part by the Zuckerman-CHE STEM Leadership Program. SP acknowledges the support from DOE grant DE-SC0009988.
|
\section{Introduction}
How to properly handle uncertainty of given information has been a hot topic. A variety of theories have been proposed to extract useful information from uncertainty, such as $D$ numbers \cite{Xiao2019a, LiubyDFMEA, Deng2019}, fuzzy set theory \cite{Xue2020entailmentIJIS, Zadeh1965, 6608375, 8944285}, $Z$ number \cite{Liu2019b} and so on \cite{Yager2014, Pan2020, Yager2014a}. On the basis of proposed theories, lots of corresponding applications have been made, like target recognition \cite{Pan2020}, pattern classification \cite{Liu2019, Song2018, articledd}, decision making \cite{Song2019a, Yager2009, Han2019, fei2019mcdm} and so on \cite{Fei2019, articlesdsdsd,Li2020generateTDBF}.
Before processing uncertain information, it is necessary to measure level of uncertainty of given information. One of the efficient way is utilizing entropy, the most representative work is Shannon entropy \cite{inbook}, which reflects a volume of information, namely amount of uncertainty contained in provided information. Although some related entropies also have been proposed, such as interval-valued entropy \cite{articleYager}, motion entropy \cite{articleChakraborty} and non-additive entropy \cite{articleTsallis}, the Shannon entropy is still regarded as the most effective tool to judge conditions of given information. In order to describe every incident systematically, an efficient theory, Dempster-Shafer evidence theory (D-S evidence theory) \cite{book, Dempster1967Upper}, is developed to offer a more adaptable solution to handle relations and conflicts among propositions existing in frames of discernment. Utilizing Dempster-Shafer evidence theory, as a generalization of Shannon entropy, Deng entropy \cite{Deng2020ScienceChina} is proposed to measure conditions of uncertainty of frames of discernment, which considers constitution of each proposition and build a bridge on measuring method between informatics and D-S evidence theory.
However, all of the previous work ignore a crucial factor which is the order of confirmation of propositions. What should be pointed out is that an ascertainment of a proposition has a effect on values of other propositions contained in the same frame of discernment. Some worthy work has considered the difference between sets and orderly sets \cite{articleSunberg, articleCheng} which firstly notice the influences of orders of propositions. Therefore, it is necessary to take sequences of propositions into consideration in D-S evidence theory. Nevertheless, how to measure degrees of uncertainty of an ordinal frame of discernment is still an open issue. To address this problem, an ordinal relative belief entropy is proposed, which considers effects brought by sequences of propositions in frames of discernment. Compared with previously proposed entropies, it indicates fluctuations of uncertainty of the whole system of discernment.
This paper is organized as follows. Some related concepts are briefly introduced in the section of preliminaries. In the next section, the definition of proposed entropy is explained in details. Some numerical examples are offered to verify correctness and validity of the proposed entropy. In the last, the whole passage is summarized in the section of conclusion.
\section{Preliminaries}
In this section, some related concepts are briefly introduced. Lots of relative meaningful researches have been made, such as generalized D-S structures \cite{yager2019generalized, yager1987dempster} and fuzzy linguistic approach \cite{Liu2020a}.
\subsection{Dempster-Shafer evidence theory}
Dempster-Shafer evidences theory (D-S theory) \cite{book, Dempster1967Upper} has been widely accepted in handling multi-source information. Due to its prominent effectiveness, the theory has been applied in many fields, such as complex mass function \cite{Xiao2019b, Xiao2020b}, Bayesian network \cite{Jiang2019Znetwork}, negation theory \cite{Zhang2020a, Li2020, Luo2019negation} and so on \cite{Luo2020vectorIJIS, Gao2019b, Deng2020InformationVolume}.
\begin{myDef}(Frame of discernment)\end{myDef}
Let $\Theta$ be a non-empty and mutually exclusive set which is consist of $n$ random values. Then, set $\Theta$ is named as frame of discernment which is defined as \cite{book, Dempster1967Upper}:
\begin{equation}
\Theta = \{P_{1}, P_{2}, ..., P_{i}, ..., P_{n}\}
\end{equation}
And the power set of $\Theta$ composed of $2^{n}$ distinctive elements is dented as $2^{\Theta}$ defined as:
\begin{equation}
2^{\Theta} = \{\emptyset, \{P_{1}\}, \{P_{2}\}, ..., \{P_{1}, P_{2}\}, ..., \{P_{1}, P{2}, ..., P_{n}\}, ..., \Theta\}
\end{equation}
\begin{myDef}(Mass function)\end{myDef}
On the base of the definition of frame of discernment, a mapping $m$ is also a mass function from $2^{\Theta}$ to $[0, 1]$ which is defined as \cite{book, Dempster1967Upper}:
\begin{equation}
m: 2^{\Theta} \rightarrow [0, 1]
\end{equation}
Besides, the mass function also satisfies:
\begin{equation}
\sum_{A \in 2^{\Theta}}m(A) = 1
\end{equation}
\begin{equation}
m(\emptyset) = 0
\end{equation}
The mass function is also named as basic probability assignment (BPA) according to D-S evidence theory. If $A$ is a proposition existing in the frame of discernment, then the value of proposition $A$ indicates the support provided by evidences to proposition $A$. Besides, if $m(A) > 0$, A is also called as a focal element.
Due to effectiveness of D-S evidence theory in handling uncertainties, the theory is also extended to the field of quantum \cite{articlequantum} which alleviates uncertainties at a completely new level.
\subsection{Some entropy theories}
To figure out what is truly contained in uncertainties, it is necessary to find effective ways to measure the degree of uncertainties. Driven by this kind of motivation, entropy theory has been widely utilized to offer a solution to solve this problem \cite{ pan2020Probability, xiao2019multi}.
\begin{myDef}(Dubois $\&$ Prade’s weighted Hartley entropy)\end{myDef}
When a frame of discernment is given, Dubois $\&$ Prade’s weighted Hartley entropy is defined as \cite{articlefuzzy}:
\begin{equation}
E_{dp}(m) = \sum_{A \in 2^{\Theta}} m(A)log_{2}|A|
\end{equation}
Where m represents a mass function developed on the basis of the definition of a frame of discernment and $A$ is a proposition existing in the frame. Besides, the value of $m(A)$ denotes the level of support to the proposition $A$ and $|A|$ indicates the cardinality of the proposition.
\begin{myDef}(Deng entropy)\end{myDef}
When a frame of discernment is given, Deng entropy is defined as \cite{Deng2020ScienceChina}:
\begin{equation}
E_{d}(m) = -\sum_{A \in 2^{\Theta}} m(A)log_{2}\frac{m(A)}{2^{|A|} - 1}
\end{equation}
Where m is a mass function according to the definition of frame of discernment and $A$ is a focal element. The cardinality of proposition $A$ is denoted as $|A|$. Besides, the mass of $m(A)$ represents the support offered by evidences to proposition $A$.
\section{Proposed ordinal relative belief entropy}
However, entropies have been widely utilized to measure uncertainty of a system \cite{yanhy2020entropy, Abellan2017, WEI20114273, zhang2020extension}, but none of them take orders of propositions into consideration. The proposed entropy regards the orders of proposition as a crucial factor in determining the degree of uncertainty to present actual conditions, because occurrences of things are involved with their sequences.
\begin{myDef}(Ordinal frame of discernment)\end{myDef}
All of the elements contained in the ordinal frame of discernment are associated with an order. For a ordinal set $\Theta_{order} = \{P_{1}^{1}, P_{2}^{2}, ..., P_{n}^{n}\}$, the subscript represents the number of propositions existing in the frame of discernment and the superscript denotes the sequence of propositions. Then, the properties all the elements satisfy are listed as follows:
\begin{itemize}
\item The proposition $P_{1}^{1}$ is supposed to be ascertained before $P_{2}^{2}$ and the rest of propositions can be confirmed in the same manner.
\item All the propositions existing in the frame of discernment only have the relations mentioned in the last point. Except for that, the definitions of the propositions remains consistent with the ones defined in the classic D-S evidence theory.
\item In the process of determining every proposition, uncertainty of the whole system is further confirmed.
\end{itemize}
\textbf{Note: }Subscripts are not supposed to be the same as superscripts. In order to simplify the process of calculation, a constraint that subscript is equal to superscript is made.
\textbf{Case 1: }Assume three propositions $P_{1}^{1}, P_{2}^{2}, P_{3}^{3}$ are contained in an ordinal frame of discernment which can be presented as $\Theta_{order} = \{\{P_{1}^{1}\}, \{P_{2}^{2}\}, \{P_{3}^{3}\}\}$. Compared to classic D-S evidence theory, the difference is that the proposition $P_{1}^{1}$ must come first, then $P_{2}^{2}$ is confirmed and $P_{3}^{3}$ is ascertained in the last. Proposition $P_{2}^{2}$ or $P_{3}^{3}$ is not allowed to take the first place. Similarly, Proposition $P_{1}^{1}$ is also not expected to take the place of $P_{2}^{2}$ or $P_{3}^{3}$. If the order is disorganized, then the whole system may obtain a completely new status.
\begin{myDef}Distributed weights for proposed entropy\end{myDef}
Because the frame of discernment is ordinal, the uncertainty of systems is further confirmed with every step of ascertaining a proposition. It is necessary to regard the degrees of importance of determining the whole system of propositions are different. Therefore, assume the number of propositions existing in the ordinal frame of discernment is $n$ and the sequence of a proposition is represented by $m$. Then the process of obtaining the modified values is defined as:
\begin{enumerate} [(1)]
\item Weights of propositions is $n - m + 1$, the calculation is expressed as:
\begin{equation}
Weight_{P_{i}^{j}} = n - m_{P_{i}^{j}} +1
\end{equation}
\item Original mass of propositions is denoted as $Mass_{P_{i}^{j}}$. Contemporary intermediate values of propositions are expressed as:
\begin{equation} Value_{P_{i}^{j}} = Mass_{P_{i}^{i}} \times Weight_{P_{i}^{j}}\end{equation}
\item The step of normalization of intermediate values is expressed as:
\begin{equation} Value_{P_{i}^{j}}^{Final} = \frac{Value_{P_{i}^{j}}}{\sum_{i = 1 = j}^{n}Value_{P_{i}^{j}}}\end{equation}
\end{enumerate}
\begin{myDef}(Relative belief entropy)\end{myDef}
It is impractical to utilize previously proposed entropy to measure the level of uncertainty of ordinal frames of discernment. Because all of the entropies consider every proposition as an independent unit, which is acceptable in classic frame of discernment. However, determination of every proposition in ordinal frame of discernment affects conditions and values of propositions which are expected to be ascertained in the following. According to conditions described, it is necessary to manifest relations among propositions in the expression of proposed entropy instead of considering each of them solely and ignoring effects single individual has on the whole. Therefore, a relative belief entropy is proposed to solve the problem mentioned above and it is expressed as:
\begin{equation}
U(P_{i}^{j},P_{a}^{b}) = Value_{P_{i}^{j}}^{Final} \times ln^{((\frac{Value_{P_{i}^{j}}^{Final}}{2^{|P_{i}^{j}|} - 1}) / (\frac{Value_{P_{a}^{b}}^{Final}}{2^{|P_{a}^{b}|} - 1}) + e)}
\end{equation}
The denominators of each minimum division expression is dependent on the number of elements contained in propositions, which is also called the cardinality of a proposition. For example, proposition $\{P_{1}^{1}, P_{2}^{2}, P_{3}^{3}\}$ has three elements, so the value of corresponding denominator is $2^{|P_{1}^{1}, P_{2}^{2}, P_{3}^{3}|} - 1 = 2^{3} - 1 = 7$.
\textbf{Case 2: }Assume there are three propositions $P_{1}^{1}, P_{2}^{2}, P_{3}^{3}$ and the ordinal frame of discernment is defined as $\{\{P_{1}^{1}\}, \{P_{2}^{2}\}, \{P_{3}^{3}\}\}$. Besides, their values after modification are listed as: $Value_{P_{1}^{1}}^{Final}= \frac{1}{3}, Value_{P_{2}^{2}}^{Final}= \frac{1}{3}, Value_{P_{3}^{3}}^{Final}= \frac{1}{3}$. Then, with respect to proposition $P_{1}^{1}$, the mass of relative belief entropy can expressed as:
$U(P_{1}^{1},P_{2}^{2}) = Value_{P_{1}^{1}}^{Final} \times ln^{((\frac{Value_{P_{1}^{1}}^{Final}}{2^{|P_{1}^{1}|} - 1}) / (\frac{Value_{P_{2}^{2}}^{Final}}{2^{|P_{2}^{2}|} - 1}) + e)} = \frac{1}{3}$,
$U(P_{1}^{1},P_{3}^{3}) = Value_{P_{1}^{1}}^{Final} \times ln^{((\frac{Value_{P_{1}^{1}}^{Final}}{2^{|P_{1}^{1}|} - 1}) / (\frac{Value_{P_{3}^{3}}^{Final}}{2^{|P_{3}^{3}|} - 1}) + e)} = \frac{1}{3}$.
\begin{myDef}(Individual ordinal relative belief entropy)\end{myDef}
In the ordinal frame of discernment, an order exists among propositions, so the calculation of a value of relative belief entropy of a certain proposition is unidirectional. For instance, for proposition $P_{1}^{1}$, the mass about $P_{1}^{1} \rightarrow P_{2}^{2}$ and $P_{1}^{1} \rightarrow P_{3}^{3}$ can be calculated. Besides, with respect to proposition $P_{3}^{3}$, only the mass about $P_{2}^{2} \rightarrow P_{3}^{3}$ can be calculated. The calculation of individual ordinal relative belief entropy is defined as:
\begin{equation}
IU(P_{i}^{j},P_{a}^{b}) = \sum_{b = j+ 1}^{n} U(P_{i}^{j},P_{a}^{b})
\end{equation}
Where n is the sequence of the last proposition. The individual ordinal relative belief entropy can be also considered as a staged entropy, because only one proposition is confirmed at this stage and conditions of other propositions are not taken into account. An example is provided to offered the process of calculation.
\textbf{Note: }As for the last proposition, the value of entropy of it is regarded as 0. Because when all the preceding propositions is determined, the whole system of frame of discernment is also determined. In other words, relations exist among propositions and an ascertainment of a proposition affects situations of other propositions.
\textbf{Case 3: }Assume there are three propositions $P_{1}^{1}, P_{2}^{2}, P_{3}^{3}$ and the ordinal frame of discernment is defined as $\{\{P_{1}^{1}\}, \{P_{2}^{2}\}, \{P_{3}^{3}\}\}$. Besides, their values after modification are listed as: $Value_{P_{1}^{1}}^{Final}= \frac{1}{3}, Value_{P_{2}^{2}}^{Final}= \frac{1}{3}, Value_{P_{3}^{3}}^{Final}= \frac{1}{3}$. And the process of calculation of values of propositions is expressed as:
$IU(P_{1}^{1},P_{a}^{b}) = \sum_{b = j+ 1}^{n} U(P_{1}^{1},P_{a}^{b}) = Value_{P_{1}^{1}}^{Final} \times ln^{((\frac{Value_{P_{1}^{1}}^{Final}}{2^{|P_{1}^{1}|} - 1}) / (\frac{Value_{P_{2}^{2}}^{Final}}{2^{|P_{2}^{2}|} - 1}) + e)} + Value_{P_{1}^{1}}^{Final} \times ln^{((\frac{Value_{P_{1}^{1}}^{Final}}{2^{|P_{1}^{1}|} - 1}) / (\frac{Value_{P_{3}^{3}}^{Final}}{2^{|P_{3}^{3}|} - 1}) + e)} \\ = \frac{1}{3} + \frac{1}{3} = \frac{2}{3}$
$IU(P_{2}^{2},P_{a}^{b}) = \sum_{b = j+ 1}^{n} U(P_{2}^{2},P_{a}^{b}) = Value_{P_{2}^{2}}^{Final} \times ln^{((\frac{Value_{P_{2}^{2}}^{Final}}{2^{|P_{2}^{2}|} - 1}) / (\frac{Value_{P_{3}^{3}}^{Final}}{2^{|P_{3}^{3}|} - 1}) + e)} = \frac{1}{3}$
$IU(P_{3}^{3},P_{a}^{b}) = 0$
Because the frame of discernment is ordinal, the process of calculation is unidirectional which also conforms to the definition of ordinal frame of discernment. Every value of ordinal relative belief entropy manifest uncertainty at the stage of confirming current propositions. With increase of the number of propositions confirmed, degree of uncertainty of the whole system is further confirmed which is presented by the values of ordinal relative belief entropy. The proposed entropy properly measures conditions of every component in the whole system.
\begin{myDef}(Integral ordinal relative belief entropy)\end{myDef}
The integral ordinal relative belief entropy is a synthesis of individual ordinal relative belief entropy which measures conditions of the whole system. In other words, integral ordinal relative belief entropy considers every stage of the complete frame. The process of calculation is defined as:
\begin{equation}
INU(P_{i}^{j},P_{a}^{b}) = \sum_{j = 1}^{n} IU(P_{i}^{j},P_{a}^{b})
\end{equation}
\textbf{Case 4: }Assume there are three propositions $P_{1}^{1}, P_{2}^{2}, P_{3}^{3}$ and the ordinal frame of discernment is defined as $\{\{P_{1}^{1}\}, \{P_{2}^{2}\}, \{P_{3}^{3}\}\}$. Besides, their values after modification are listed as: $Value_{P_{1}^{1}}^{Final}= \frac{1}{3}, Value_{P_{2}^{2}}^{Final}= \frac{1}{3}, Value_{P_{3}^{3}}^{Final}= \frac{1}{3}$. And the process of calculation of values of propositions is expressed as:
$INU(P_{i}^{j},P_{a}^{b}) = \sum_{j = 1}^{n} IU(P_{i}^{j},P_{a}^{b}) = IU(P_{1}^{1},P_{a}^{b}) + IU(P_{2}^{2},P_{a}^{b}) + IU(P_{3}^{3},P_{a}^{b}) = 1$
Then, this is a final description of level of uncertainty about the system.
\subsection{Measurement about classic frame of discernment using ordinal relative belief entropy}
In the last section, a new entropy is proposed to measure degree of uncertainty of ordinal frame of discernment \cite{ deng2021fuzzymembershipfunction} which manifests an important role order plays in influencing degrees of uncertainty of the whole system \cite{article5555}. However, the proposed entropy can be also utilized to describe conditions about a classic frame of discernment considering every kind of order of a certain frame. For example, if there are three propositions existing in a frame of discernment, then every place corresponding to an order can be held by each of propositions. Therefore, it is necessary to find out all kinds of combinations and get an average utilizing sums of values from different combinations. Therefore, a measurement of classic frame of discernment is made. The process of calculation is defined as:
\begin{enumerate} [(1)]
\item Find out all kinds of the possible combinations.
\item Get a sum of values from combinations using definition of ordinal relative belief entropy.
\item Calculate an average of the sum according to the number of kinds of combinations to get a final mass.
\end{enumerate}
\textbf{Case 5: }Assume there are three propositions $P_{1}, P_{2}, P_{3}$ and the frame of discernment is defined as $\{\{P_{1}\}, \{P_{2}\}, \{P_{3}\}\}$. Besides, their values are listed as: $P_{1}= \frac{1}{6}, P_{2}= \frac{1}{2}, P_{3}= \frac{1}{3}$. In this case, in order to conform to usage of previously proposed entropy, the superscript is concealed to focus on effects brought by sequences. All the possible combinations are listed below:
$\{P_{1}, P_{2}, P_{3}\}$, $\{P_{1}, P_{3}, P_{2}\}$, $\{P_{2}, P_{1}, P_{3}\}$, $\{P_{2}, P_{3}, P_{1}\}$, $\{P_{3}, P_{1}, P_{2}\}$, $\{P_{3}, P_{2}, P_{1}\}$.
In all, there are six different combinations, which also means there are six different ordinal frame of discernment. Using ordinal relative belief entropy, their values calculated are listed below:
$INU_{Frame_{1}}(P_{i}^{j},P_{a}^{b}) = 1.6624$, $INU_{Frame_{2}}(P_{i}^{j},P_{a}^{b}) = 1.3267$, $INU_{Frame_{3}}(P_{i}^{j},P_{a}^{b}) = 2.9388$,
$INU_{Frame_{4}}(P_{i}^{j},P_{a}^{b}) = 3.1569$, $INU_{Frame_{5}}(P_{i}^{j},P_{a}^{b}) = 2.0190$, $INU_{Frame_{6}}(P_{i}^{j},P_{a}^{b}) = 2.6049$.
\textbf{Note: }All of the superscript is concealed to be consistent to environment of application.
Because the number of combinations is six, the final description about the unordered frame of discernment is the sum of six values of $INU$ divides six.
$INU_{Final} = \frac{INU_{Frame_{1}}(P_{i}^{j},P_{a}^{b})+ INU_{Frame_{2}}(P_{i}^{j},P_{a}^{b}) + INU_{Frame_{3}}(P_{i}^{j},P_{a}^{b}) + INU_{Frame_{4}}(P_{i}^{j},P_{a}^{b}) + INU_{Frame_{5}}(P_{i}^{j},P_{a}^{b}) + INU_{Frame_{6}}(P_{i}^{j},P_{a}^{b})}{6} \\= 2.1181$
Utilizing two entropies mentioned above, two degrees of uncertainty about frame of discernment provided are calculated. The results are listed in Table \ref{1}.
\begin{table}[htbp]
\centering
\caption{Results of three kinds of entropies}
\begin{tabular}{ccc}
\toprule
Dubois $\&$ Prade’s weighted Hartley entropy&Deng entropy&Proposed entropy \\
\midrule
0&1.0113&2.1181 \\
\bottomrule
\label{1}
\end{tabular}
\end{table}
In this case, the effectiveness of proposed ordinal entropy in measuring degrees of uncertainty of classic frame of discernment is verified, which considers every kind of order to obtain a proper description about unordered frame of discernment.
\section{Numerical examples}
In this section, some examples are provided to verify effectiveness of proposed entropy in illustrating how an order of a ordinal frame of discernment affects levels of uncertainty of whole systems.
\textbf{Example 1: }Assume there are three propositions $P_{1}^{1}, P_{2}^{2}, P_{3}^{3}$ and the ordinal frame of discernment is defined as $\{\{P_{1}\}, \{P_{2}\}, \{P_{3}\}\}$. Besides, their values after modification are listed as: $Mass_{P_{1}^{1}}= \frac{1}{3}, Mass_{P_{2}^{2}}= \frac{1}{3}, Mass_{P_{3}^{3}}= \frac{1}{3}$. The process of calculation with respect to the sequence $\{P_{1}^{1}, P_{2}^{2}, P_{3}^{3}\}$ is given as:
First, obtain weights corresponding to propositions.
$Weight_{P_{1}^{1}} = n - m_{P_{1}^{1}} +1 = 3$, $Weight_{P_{2}^{2}} =2$, $Weight_{P_{3}^{3}} = 1$
Second, get values of propositions after modification utilizing weights.
$Value_{P_{1}^{1}} = Mass_{P_{1}^{1}} \times Weight_{P_{1}^{1}} = 1$, $Value_{P_{2}^{2}} = \frac{2}{3}$, $Value_{P_{3}^{3}} = \frac{1}{3}$
Third, gain values after normalization.
$Value_{P_{1}^{1}}^{Final} = \frac{Value_{P_{1}^{1}}}{\sum_{i = 1 =j}^{n}Value_{P_{i}^{j}}} = \frac{1}{2}$, $Value_{P_{1}^{1}}^{Final} = \frac{1}{3}$, $Value_{P_{1}^{1}}^{Final} = \frac{1}{6}$
Forth, obtain the final value according to definition of ordinal relative belief entropy.
$INU(P_{i}^{j},P_{a}^{b}) = \sum_{j = 1}^{n} IU(P_{i}^{j},P_{a}^{b}) = 2.1087$
Then, 2.1087 is the value of measurement of degree of uncertainty of the provided ordinal frame of discernment, which takes the factor of order into consideration.
\textbf{Example 2: }Assume there are three propositions $P_{1}, P_{2}, P_{3}$. Besides, their values are listed as: $Mass_{P_{1}}= \frac{1}{4}, Mass_{P_{2}}= \frac{1}{6}, Mass_{P_{3}}= \frac{7}{12}$. In this example, all six cases are taken into account and the measurement of the ordinal frame of discernment in different situations are listed in Table \ref{2}.
\begin{table}[!htbp]
\begin{center}
\caption{Results of six conditions}
\begin{tabular}{lcccccc}
\toprule
\multicolumn{1}{m{2cm}}{Sequence}
&\multicolumn{3}{m{4cm}}{\centering $IU(P_{i}^{j},P_{a}^{b})$}&\multicolumn{1}{m{2.5cm}}{\centering $INU(P_{i}^{j},P_{a}^{b})$}&\multicolumn{1}{m{2.8cm}}{Deng entropy \cite{Deng2020ScienceChina}}&\multicolumn{1}{m{2.8cm}}{DP entropy \cite{articlefuzzy}}\\
\midrule
$P_{1}^{1}, P_{2}^{2}, P_{3}^{3}$&1.3456&0.2381&0&1.5838&1.3844&0\\
$P_{1}^{1}, P_{2}^{3}, P_{3}^{2}$&1.1480&1.2734&0&2.4214&1.3844&0\\
$P_{1}^{2}, P_{2}^{1}, P_{3}^{3}$&0.8170&0.4023&0&1.2194&1.3844&0\\
$P_{1}^{2}, P_{2}^{3}, P_{3}^{1}$&0.7037&1.2170&0&1.9208&1.3844&0\\
$P_{1}^{3}, P_{2}^{1}, P_{3}^{2}$&3.1927&0.3607&0&3.5535&1.3844&0\\
$P_{1}^{3}, P_{2}^{2}, P_{3}^{1}$&3.2621&0.1998&0&3.4619&1.3844&0\\
\bottomrule
\end{tabular}
\label{2}
\end{center}
\end{table}
It can be easily concluded that the mass obtained by proposed ordinal relative belief entropy is changing according to the sequence of proposition. On the contrary, values of the two previously proposed entropy do not change no matter how the sequence of propositions is. It is necessary to take orders of propositions into consideration due to inner connections among propositions, which means any occurrence of certain incident has an influence on evaluations of other ones.
\textbf{Example 3: }Assume there are three propositions $P_{1}, P_{2}, P_{3}$. Besides, their values are listed as: $Mass_{P_{1}}= \frac{1}{2}, Mass_{P_{2}}= \frac{5}{12}, Mass_{P_{3}}= \frac{1}{12}$. In this example, all six cases are taken into account and the measurement of the ordinal frame of discernment in different situations are listed in Table \ref{3}.
\begin{table}[!htbp]
\begin{center}
\caption{Results of six conditions}
\begin{tabular}{lcccccc}
\toprule
\multicolumn{1}{m{2cm}}{Sequence}
&\multicolumn{3}{m{4cm}}{\centering $IU(P_{i}^{j},P_{a}^{b})$}&\multicolumn{1}{m{2.5cm}}{\centering $INU(P_{i}^{j},P_{a}^{b})$}&\multicolumn{1}{m{2.8cm}}{Deng entropy \cite{Deng2020ScienceChina}}&\multicolumn{1}{m{2.8cm}}{DP entropy \cite{articlefuzzy}}\\
\midrule
$P_{1}^{1}, P_{2}^{2}, P_{3}^{3}$&2.8174&0.8769&0&3.6943&1.3250&0\\
$P_{1}^{1}, P_{2}^{3}, P_{3}^{2}$&3.0993&0.0909&0&3.1903&1.3250&0\\
$P_{1}^{2}, P_{2}^{1}, P_{3}^{3}$&2.2783&1.1524&0&3.4308&1.3250&0\\
$P_{1}^{2}, P_{2}^{3}, P_{3}^{1}$&2.5932&0.0970&0&2.6902&1.3250&0\\
$P_{1}^{3}, P_{2}^{1}, P_{3}^{2}$&0.3431&0.9796&0&1.3228&1.3250&0\\
$P_{1}^{3}, P_{2}^{2}, P_{3}^{1}$&0.3376&0.7111&0&1.0488&1.3250&0\\
\bottomrule
\end{tabular}
\label{3}
\end{center}
\end{table}
It can be also easily obtained that sequence plays an important role in affecting values of measurement of uncertainty. Meanwhile, previously proposed can not present these kinds of differences.
\textbf{Example 4: }Assume there are three propositions $P_{1}, P_{2}, P_{3}, P_{4_{P_{1},P_{2}}}$. And $P_{4_{P_{1},P_{2}}}$ represents the union set of $P_{1}, P_{2}$. Besides, their values are listed as: $Mass_{P_{1}}= \frac{4}{13}, Mass_{P_{2}}= \frac{3}{13}, Mass_{P_{3}}= \frac{5}{13}, Mass_{P_{4_{P_{1},P_{2}}}}= \frac{1}{13}$. In this example, all six cases are taken into account and the measurement of the ordinal frame of discernment in different situations are listed in Table \ref{4}.
\begin{table}[!htbp]
\begin{center}
\caption{Results of twenty-four conditions}
\begin{tabular}{lccccccc}
\toprule
\multicolumn{1}{m{2cm}}{Sequence}
&\multicolumn{4}{m{5cm}}{\centering $IU(P_{i}^{j},P_{a}^{b})$}&\multicolumn{1}{m{2.5cm}}{\centering $INU(P_{i}^{j},P_{a}^{b})$}&\multicolumn{1}{m{2.8cm}}{Deng entropy \cite{Deng2020ScienceChina}}&\multicolumn{1}{m{2.8cm}}{DP entropy \cite{articlefuzzy}}\\
\midrule
$P_{1}^{1}, P_{2}^{2}, P_{3}^{3}, P_{4_{P_{1},P_{2}}}^{4}$&3.0632& 1.1694&0.9688&0&5.2015&1.8262&0.0769\\[5pt]
$P_{1}^{1}, P_{2}^{2}, P_{3_{P_{1},P_{2}}}^{4},P_{4}^{3}$&3.2833&1.2077&0.0654&0&4.5565&1.8262&0.0769\\[5pt]
$P_{1}^{1}, P_{2}^{3}, P_{3}^{2}, P_{4_{P_{1},P_{2}}}^{4}$&2.9225&2.1779&0.4785&0&5.5790&1.8262&0.0769\\[5pt]
$P_{1}^{1}, P_{2}^{3}, P_{3_{P_{1},P_{2}}}^{4},P_{4}^{2}$&2.9787&2.1963&0.0599&0&5.2350&1.8262&0.0769\\[5pt]
$P_{1}^{1}, P_{2_{P_{1},P_{2}}}^{4},P_{3}^{2}, P_{4}^{3}$&3.4086&0.2130&0.2731&0&3.8948&1.3250&0.0769\\[5pt]
$P_{1}^{1}, P_{2_{P_{1},P_{2}}}^{4},P_{3}^{3},P_{4}^{2}$&2.8173&0.2266&0.5626&0&3.6065&1.8262&0.0769\\[5pt]
$P_{1}^{2}, P_{2}^{1}, P_{3}^{3}, P_{4_{P_{1},P_{2}}}^{4}$&2.1720&1.7218&0.9965&0&4.8904&1.8262&0.0769\\[5pt]
$P_{1}^{2}, P_{2}^{1}, P_{3_{P_{1},P_{2}}}^{4}, P_{4}^{3} $&2.3137&1.8053&0.0676&0&4.1866&1.8262&0.0769\\[5pt]
$P_{1}^{2}, P_{2}^{3},P_{3}^{1}, P_{4_{P_{1},P_{4}}}^{4}$&2.1179&2.2458&0.7300&0&2.8004&1.8262&0.0769\\[5pt]
$P_{1}^{2}, P_{2}^{3}, P_{3_{P_{1},P_{2}}}^{4},P_{4}^{1},$&2.1936&2.3156&0.0642&0&4.5735&1.8262&0.0769\\[5pt]
$P_{1}^{2}, P_{2_{P_{1},P_{2}}}^{4},P_{3}^{1}, P_{4}^{3} $&2.4691&0.2267&0.4179&0&3.1138&1.8262&0.0769\\[5pt]
$P_{1}^{2}, P_{2_{P_{1},P_{2}}}^{4},P_{3}^{3},P_{4}^{1} $&2.3993&0.2197&0.5697&0&3.1887&1.8262&0.0769\\[5pt]
$P_{1}^{3},P_{2}^{1}, P_{3}^{2}, P_{4_{P_{1},P_{2}}}^{4}$&3.8036&1.6023&0.46631&0&5.8723&1.8262&0.0769\\[5pt]
$P_{1}^{3},P_{2}^{1}, P_{3_{P_{1},P_{2}}}^{4},P_{4}^{2}, $&3.8947&1.6008&0.0583&0&5.5538&1.8262&0.0769\\[5pt]
$P_{1}^{3},P_{2}^{2},P_{3}^{1}, P_{4_{P_{1},P_{2}}}^{4}$&3.8885&1.1221&0.6916&0&5.7024&1.8262&0.0769\\[5pt]
$P_{1}^{3},P_{2}^{2},P_{3_{P_{1},P_{2}}}^{4},P_{4}^{1} $&4.0737&1.1286&0.0605&0&5.2629&1.8262&0.0769\\[5pt]
$P_{1}^{3},P_{2_{P_{1},P_{2}}}^{4},P_{3}^{1}, P_{4}^{2} $&4.1261&0.1906&0.3961&0&4.7129&1.8262&0.0769\\[5pt]
$P_{1}^{3},P_{2_{P_{1},P_{2}}}^{4},P_{3}^{2} ,P_{4}^{1} $&4.2224&0.1952&0.2617&0&4.6794&1.8262&0.0769\\[5pt]
$P_{1_{P_{1},P_{2}}}^{4},P_{2}^{1}, P_{3}^{2}, P_{4}^{3}, $&0.4758&1.4152&0.3034&0&2.1945&1.8262&0.0769\\[5pt]
$P_{1_{P_{1},P_{2}}}^{4},P_{2}^{1}, P_{3}^{3},P_{4}^{2}$&0.4932&0.9880&0.4501&0&1.9314&1.8262&0.0769\\[5pt]
$P_{1_{P_{1},P_{2}}}^{4},P_{2}^{2},P_{3}^{1}, P_{4}^{3}$&0.4932&0.9880&0.4501&0&1.9314&1.8262&0.0769\\[5pt]
$P_{1_{P_{1},P_{2}}}^{4},P_{2}^{2},P_{3}^{3},P_{4}^{1}$&0.4765&0.9630&0.6119&0&2.0514&1.8262&0.0769\\[5pt]
$P_{1_{P_{1},P_{2}}}^{4},P_{2}^{3},P_{3}^{1}, P_{4}^{2} $&0.4324&1.7840&0.4489&0&2.6654&1.3250&0.0769\\[5pt]
$P_{1_{P_{1},P_{2}}}^{4},P_{2}^{3},P_{3}^{2},P_{4}^{1}$&0.4450&1.8202&0.2978&0&2.5630&1.8262&0.0769\\[5pt]
\bottomrule
\end{tabular}
\label{4}
\end{center}
\end{table}
It can easily concluded that the values obtained from proposed entropy are changing from the orders of confirmations of propositions. On the contrary, when applying classic entropies to measure the condition of ordinal frame of discernment, the mass obtained does not change with the variation of orders.
\textbf{Example 5: }Assume there are three propositions $P_{1}, P_{2}, P_{3_{P_{1}, P_{2}}}$. Besides, their values after modification are listed as: $Mass_{P_{1}}= \frac{6}{17}, Mass_{P_{2}}= \frac{4}{17}, Mass_{P_{3_{P_{1}, P_{2}}}}= \frac{7}{17}$. In this example, all six cases are taken into account and the measurement of different situations are listed in Table \ref{5}.
\begin{table}[!htbp]
\begin{center}
\caption{Results of six conditions}
\begin{tabular}{lcccccc}
\toprule
\multicolumn{1}{m{2cm}}{Sequence}
&\multicolumn{3}{m{4cm}}{\centering $IU(P_{i}^{j},P_{a}^{b})$}&\multicolumn{1}{m{2.5cm}}{\centering $INU(P_{i}^{j},P_{a}^{b})$}&\multicolumn{1}{m{2.8cm}}{Deng entropy \cite{Deng2020ScienceChina}}&\multicolumn{1}{m{2.8cm}}{DP entropy \cite{articlefuzzy}}\\
\midrule
$P_{1}^{1}, P_{2}^{2}, P_{3_{P_{1}, P_{2}}}^{3}$&2.1534&0.4402&0&2.5936&1.5485&0.4117\\[5pt]
$P_{1}^{1}, P_{2_{P_{1}, P_{2}}}^{3}, P_{3}^{2}$&2.0867&1.0039&0&3.0906&1.5485&0.4117\\[5pt]
$P_{1}^{2}, P_{2}^{1}, P_{3_{P_{1}, P_{2}}}^{3}$&1.3065&0.7981&0&2.1046&1.5485&0.4117\\[5pt]
$P_{1}^{2}, P_{2_{P_{1}, P_{2}}}^{3}, P_{3}^{1}$&1.2898&0.9948&0&2.2846&1.5485&0.4117\\[5pt]
$P_{1_{P_{1}, P_{2}}}^{3}, P_{2}^{1}, P_{3}^{2}$&2.5047&0.7982&0&3.3029&1.5485&0.4117\\[5pt]
$P_{1_{P_{1}, P_{2}}}^{3}, P_{2}^{2}, P_{3}^{1}$&2.5544&0.4353&0&2.9898&1.5485&0.4117\\[5pt]
\bottomrule
\end{tabular}
\label{5}
\end{center}
\end{table}
The proposed ordinal relative belief entropy manifest effects brought by orders of propositions in an ordinal frame of discernment. Previously proposed entropies can not indicate changes brought by orders.
\section{Conclusion}
In practical situations, everything comes in an order. Previously proposed work dose not take this factor into consideration. The main contribution of this paper is that a completely new entropy to measure degrees of uncertainty of ordinal frame of discernment is proposed. Compared with classic entropies, it is able to manifest influences brought by sequence of propositions. Without doubt, the proposed entropy provides a reasonable idea in improving results of disposing uncertainties to better to be adaptive to real life. Numerical examples also proves the efficiency and rationality of the proposed ordinal relative belief entropy.
\section*{Acknowledgment}
\bibliographystyle{elsarticle-num}
|
\section{Introduction}
Magnetic space groups (MSGs)\cite{bradley1968magnetic} describe the symmetry of lattices where spins are magnetically ordered.
Magnetic ordering necessarily breaks time-reversal symmetry, and, as the order parameter is a vector, usually also breaks some point-group symmetries (rotation and reflection).
In many magnetically ordered materials, specially those having antiferromagnetism, there are a special type of composite symmetries: a group element $m=g\cdot{T}$ is the composition of a space-group (SG) symmetry $g$ and time-reversal symmetry $T$.
Consider for example an anti-ferromagnetic Neel order polarized along $z$ with propagation vector $\mathbf{Q}=(\pi/a,0,0)$, where $a$ is the lattice constant.
The lattice translation by one unit cell along $x$, $\{E|100\}$, is broken, and time reversal, $T$, is also broken, while their composition $\{E|100\}\cdot{T}$ remains a symmetry.
(Here we use $E$ to represent the identity $3\times3$ matrix, representing the trivial element of $O(3)$.)
According to the types of $g$ in $m=g\cdot{T}$, the MSGs are classified into four types:
if $g$ does not exist (so that $m$ does not exist), the magnetic group is type-I; if $g$ is identity, $\{E|000\}$, type-II; if $g$ involves a nontrivial point group operation, type-III; and if $g$ is a pure lattice translation, $\{E|klm\}$, the group is type-IV.
By this definition, type-II MSGs contain time reversal, and hence describe nonmagnetic materials.
In this work, we focus on type-I, III, IV MSGs\footnote{the ``MSGs'' we use in this work means type-1, 3, and 4 MSGs, and the non-magnetic type-2 MSGs are excluded.}.
MSGs are also the symmetries of the effective Hamiltonians describing elementary excitations, such as magnons and electrons, that move within a magnetically ordered lattice.
In this work, we focus on magnetic materials in which coherent quasiparticle fermion excitations form band(s) within a finite range of the Fermi energy, and study the band topology of these fermions.
Our theory can in principle be applied to any magnetic materials where the notion of electron-like quasiparticles is valid, at least near the Fermi energy.
Itinerant magnets\cite{shimizu1981itinerant}, heavy-fermion metals\cite{stewart1984heavy}, and doped Mott insulators\cite{lee2006doping} that maintain a magnetically ordered state, are considered to belong to this large class of materials.
The interplay between symmetry and topology has been a focus of modern condensed-matter research\cite{hasan2010colloquium,qi2011topological,chiu2016classification,armitage2018weyl}.
For a given symmetry group and a nonzero gap, all band Hamiltonians are grouped into \emph{equivalence classes}, where two Hamiltonians in the same (different) class(es) can(not) be smoothly deformed into each other, while maintaining both the gap and the symmetry.
Each equivalence class is denoted by a unique set of integers called the \emph{topological invariants}\cite{thouless1982quantized, haldane1983nonlinear, affleck1987rigorous, wen1990topological, kitaev2001unpaired, kane2005z, moore2007topological, fu2007topological}, the forms and types of which only depend on the symmetry group and dimensionality.
How many distinct equivalence classes exist for a given symmetry group in a given dimension, and what are the topological invariants for each class?
This is the question we call the problem of \emph{topological classification}.
The theory of topological classification for time-reversal and particle-hole symmetries has been done in all dimensions using the K-theory\cite{schnyder2008classification, kitaev2009periodic, ryu2010topological}; the classification problem for a single spatial symmetry plus time reversal in three dimensions has been solved, usually heuristically, for several symmetries\cite{fu2007topological3d,hsieh2012topological,teo2008surface,wang2016hourglass,shiozaki2014topology,song2017d,fang2019new,turner2010entanglement,hughes2011inversion,song2018quantitative,khalaf2018symmetry}; and the classification problem for arbitrary SGs plus time reversal in three dimensions have been attempted using either the ``real-space recipe'' argument\cite{song2018quantitative,song2019topological,song2020real} or the ``double-strong-topological-insulator construction''\cite{fang2019new,khalaf2018symmetry} argument more recently.
The classification problem for MSGs begins with the theory of axion insulators protected by space-inversion symmetry without time reversal\cite{qi2008topological,li2010dynamical,hughes2011inversion,qi2011topological,turner2012quantized,yue2019symmetry,zhang2019topological,otrokov2019prediction, rienks2019large}, followed by the theory of antiferromagnetic topological insulators\cite{mong2010antiferromagnetic,fang2013topological,liu2013antiferromagnetic,zhang2015topological}, again followed by the discovery of several topological invariants protected by wallpaper groups\cite{liu2014topological,shiozaki2017topological}.
A more systematic attempt is made in Ref.\cite{elcoro2020magnetic}, where the layer-construction (LC) method reveals a number of new topological states.
In this work, we use the \emph{real-space recipe} method developed in \cite{song2019topological,song2020real}.
Effectively, this method converts the problem of topological classification into a ``LEGO puzzle'', where one tries to find distinct ways one can build an ``edgeless'' construction using some given pieces.
This ``LEGO puzzle'' is then further transformed into finding all independent integer solutions of a set of linear equations on a $\mathbb{Z}_n$ ring.
This method not only yields a complete topological classification of gapped bands in each of the 1421 groups, but also gives us, for each nontrivial equivalence class, one explicit and microscopic construction that we call the ``topological crystal'' (TC)\cite{song2019topological}.
We emphasize that our method finds more topological classes than the layer-construction method, because LCs are a type of TCs, while not all TCs are LCs.
(In fact, we show that at least 553 of the 1421 MSGs have at least one topological state that cannot be layer-constructed.)
If the classification gives us the labels (topological invariants) for the equivalence classes into which gapped states are put in, the \emph{topological diagnosis} then tells us to which equivalence class a specific, given material (Hamiltonian) belongs.
Ideally, a diagnosis scheme computes the topological invariants of that Hamiltonian, and by comparing these values with the labels on the equivalence classes, one puts the Hamiltonian into the right one.
However, topological invariants are notoriously difficult to compute\cite{fang2015new,wang2016hourglass}, and for some, we do not even have the explicit expressions in terms of the wave functions of the bands\cite{song2017d,fang2019new}.
Fortunately, if we relax the requirement of ideal diagnosis to approximate diagnosis, the story is completely changed.
An approximate diagnosis uses partial information on the wave function, and in return gives us partial information on the topological invariants, not invariants themselves.
For example, an approximate diagnosis for systems with $n$-fold rotation symmetry yields the invariant (Chern number) modulo $n$ by using only the rotation eigenvalues at several high-symmetry momenta\cite{hughes2011inversion,fang2012bulk}.
Recently, the theory of symmetry-based indicators (SIs)\cite{po2017symmetry} and that of topological quantum chemistry\cite{bradlyn2017topological}, enhanced by the full mapping from indicators to topological invariants\cite{song2018quantitative}, give birth to a fast approximate diagnosis scheme.
This fast-diagnosis scheme has been applied to a large number of non-magnetic materials\cite{zhang2019catalogue,vergniory2019complete,tang2019comprehensive}.
In this work, we extend the above diagnosis scheme to magnetic materials that can be well-characterized in terms of band structures. We find the explicit formulas for all SIs in terms of band representations (which in part differ from previous works\cite{ono2018unified, elcoro2020magnetic}); and we calculate the values of indicators for each TC in every MSG.
Since each gapped state can be adiabatically continued to a TC, this result in fact yields the mapping from invariants to indicators.
In this calculation, we find that certain indicator values are never taken in any TC, and hence can only indicate nodal band structures
\cite{fang2003anomalous, wan2011topological, xu2011chern, wang2016time, kubler2016weyl, chang2016room, yang2017topological, liu2018giant, wang2018large, belopolski2019discovery, morali2019fermi, liu2019magnetic, nie2017topological, tang2016dirac, hua2018dirac, kim2018large, nie2019topological, zou2019study}.
These nodes are ``evasive'' as they are away from any high-symmetry points or lines.
In fact, we show that all these nodal indicators indicate Weyl nodes at generic momenta.
In a numerical diagnosis, the indicators are by far easier to obtain than the invariants, because the former only depend on the band representations at fewer than or equal to eight momenta and the latter depend on the valence-band wave functions in the entire Brillouin zone (BZ).
Therefore, an inverse mapping from indicators to invariants/nodes is generated using a script, provided along with the paper at \href{https://github.com/yjiang-iop/MSG_mapping}{https://github.com/yjiang-iop/MSG\_mapping}.
\section{Classification}
\subsection{General scheme}
The basic idea of the real-space recipe is that all topological crystalline insulators (TCIs) can be adiabatically deformed into a special form of real-space constructions called the TCs, such that classifying TCIs is equivalent to classifying TCs.
TCs are real-space patterns built from topological pieces in lower dimensions, which were first applied to non-magnetic SGs in order to obtain the full classifications of non-magnetic TCIs\cite{song2019topological}.
In this work, we apply the real-space recipe for the construction and the classification of all gapped topological states protected by MSGs.
To start with, we build a structure of cell complex by using MSG symmetries, including both unitary and anti-unitary ones, and partition the 3D space into finite 3D regions called asymmetric units (AUs) that fill the whole 3D space without overlaps.
In fact, all AUs are symmetry-related and can be generated by choosing one AU and then copying it using MSG operations.
AUs are also called 3-cells, and the 2D faces where they meet are 2-cells.
Similarly, 1D lines where 2-cells intersect are 1-cells, and the endpoints of 1-cells are 0-cells.
To construct TCs in 3D, we should take account of all $d$-dimensional topological building blocks with $d \leq 3$.
For each cell, its local symmetry group is defined as the collection of symmetries that keep every point of the cell unchanged.
The local symmetry group of a cell determines the onsite symmetry class (Altland-Zirnbauer class\cite{altland1997nonstandard}) of the Hamiltonian on that cell.
For MSGs, the effective symmetry class of a cell is always class A or class AI.
Note that although some cells have mirror plane as onsite symmetry, the states on them can be divided into two sectors by mirror eigenvalues, each of which belongs to class A.
A 2-cell may have local symmetry group generated by $M\cdot T$, and a 1-cell local symmetry group generated by $C_2\cdot T$.
Since $(M\cdot T)^2=(C_2\cdot T)^2=+1$, those cells belong to symmetry class AI.
According to the tenfold way results\cite{ryu2010topological}, systems of class AI have trivial classification in all dimension $\le 3$, and systems of class A have nontrivial classification in 2D and trivial classification in 0D, 1D, and 3D, which means only 2D topological building blocks need to be considered in the construction of TCs for MSG.
Furthermore, there are two types of 2-cells depending on whether they coincide with mirror planes. If coincide, they can be decorated with mirror Chern insulators characterized by two $\mathbb{Z}$ numbers, i.e., two mirror Chern numbers for $\pm i$ mirror sectors\footnote{Due to the absence of TRS in MSGs, the ``mirror Chern insulator'' we use here do not require two mirror Chern numbers to be opposite with each other, i.e., $C^{+}_m \neq C^{-}_m$ generically, which are different with the mirror Chern insulators in non-magnetic SGs with $C^{+}_m=-C^{-}_m$\cite{teo2008surface,hsieh2012topological}.}, and if not, with Chern insulators characterized by one $\mathbb{Z}$ number, i.e., the Chern number.
Therefore, there are only two types of building blocks for our real-space recipe, i.e., Chern insulators and mirror Chern insulators.
Having building blocks in hand, we next enumerate all topological inequivalent decorations on the cell complexes in MSGs.
Before proceeding, note that the building blocks themselves form a finitely generated Abelian group, e.g., $\mathbb{Z}$ for Chern insulators and $\mathbb{Z}^2$ for mirror Chern insulators.
Therefore, the TCs built from them form a linear space with integer coefficients, such that two TCs can be ``added'' to obtain another TC, and there exists a maximal set of linearly independent TCs (the generators) for each MSG.
As a result, one just needs to obtain the generators to describe the full set of TCs.
As TCs are supposed to be fully gapped topological states, all the boundary states contributed by 2D building blocks (Chern insulators and mirror Chern insulators) should cancel with each other on each 1-cell, leading to fully gapped states inside the bulk, a condition known as the ``gluing condition''\cite{song2019topological} or ``no-open-edge condition''\cite{song2020real}.
After this procedure, we obtain a set of generators that form an Abelian group $\mathbb{Z}^n$.
However, this is generally not the final classification, because some generators will reduce from $\mathbb{Z}$-type to $\mathbb{Z}_2$-type after a process of subtracting topological trivial elements called ``bubble-equivalence''\cite{song2019topological,song2020real}.
The final classification can be expressed as a quotient group Ker/Img, a structure resembling group (co)homology, where Ker stands for the linear space of TCs satisfying no-open-edge condition and Img for the space of bubble-equivalence.
These final classifications of MSGs have the form $\mathbb{Z}^n\times\mathbb{Z}_2^l$ and can be found in Appendix.\ref{AppendixN}.
Lastly, we compare the real-space constructions for non-magnetic SGs and MSGs.
First, observe that their building blocks are different.
In non-magnetic SGs, due to the time-reversal symmetry (TRS), the building blocks are two-dimensional topological insulators and mirror Chern insulators with $C^{+}_m = - C^{-}_m$, while in MSGs, the building blocks are Chern insulators and mirror Chern insulators with independent $C^{+}_m$ and $C^{-}_m$\footnote{in the following context we use the notation $C^{\pm}_{m}$ for both real-space mirror Chern numbers and momentum-space mirror Chern numbers, and the specific meaning of $C^{\pm}_m$ can be inferred from the context}.
More significantly, for non-magnetic SGs, one can obtain most of the TCs by LCs, which involve only layered 2D TIs and mirror Chern insulators as building blocks\cite{song2018quantitative}, with only 12 non-magnetic SGs having states beyond LCs\cite{song2019topological}.
However, non-layer constructions (non-LCs) exist widely in MSGs, and constructing with layers only may lose a number of TCs.
By contrast, the real-space recipe employing the structure of the cell complex automatically includes both LCs and non-LCs, giving the complete collection of TCs, hence the complete classification of gapped topological states.
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{surface_state.pdf}
\caption{\label{surface_states}Surface states of symmetries with non-trivial topological invariants in MSGs, with the surface terminations preserving corresponding symmetries. These surface states can be 1D chiral hinge modes and 2D surface modes, with 2D modes being either slope-like chiral surface modes in (1),(6),(10), or Dirac cones in (2),(3),(4),(5). More details can be found in Appendix.\ref{AppendixC}.}
\end{figure*}
\subsection{Topological invariants}
TCIs are characterized by crystalline-symmetry-protected topological invariants.
Previous works have looked into some of the topological invariants protected by magnetic crystalline symmetries, with the earliest one being the axion insulators with inversion invariant\cite{turner2012quantized,hughes2011inversion}, followed by the anti-ferromagnetic topological insulators protected by anti-unitary translations\cite{mong2010antiferromagnetic,fang2013topological}.
In this work, we exhaustively enumerate all topological invariants in MSGs.
Formally, having nontrivial symmetry-protected topological invariants means a topological state cannot be smoothly deformed into a trivial state when the symmetry is preserved.
Specially, if a TCI can not be adiabatically connected to an atomic insulator with the symmetry operation $g$ preserved, we say it has nontrivial $g$-invariant.
As all TCIs are adiabatically connected to TCs, we can utilize TCs to derive all invariants protected by MSGs symmetries.
In our real-space recipe, given a symmetry operation $g$ alone, if nontrivial TCs compatible with $g$ can be constructed, then we say $g$ can protect nontrivial topological invariants.
More specifically, each independent TC corresponds to an independent invariant, and they have the same group structure, e.g., a $\mathbb{Z}_2$ TC owns a $\mathbb{Z}_2$ invariant.
To find which symmetry operations in MSGs can protect nontrivial invariants and what kinds of the invariants they protect, we take account of all of MSG symmetries one by one, including translation, rotation, inversion, mirror, rotoinversion ($S_n$), screw, glide, and those combined with TRS such as $C_n \cdot T$, etc.
We consider when each of them is present alone, what TCs can be constructed.
If no TC exists, this symmetry has only trivial invariant; if there exist TCs, this symmetry must protect nontrivial topological invariants, and we further decide how many independent TCs and whether they are $\mathbb{Z}$-type or $\mathbb{Z}_n$ type by checking if the TCs can be smoothly deformed into trivial states after multiplying $n$ times.
In this way, we know all the MSG symmetries that can singly protect nontrivial topological invariants, with each symmetry hosting at most one invariant, being either $\mathbb{Z}_2$-type or $\mathbb{Z}$-type.
We summarize the results in Table.\ref{table_invariants}, and classify these invariants into three types:
\begin{itemize}
\item Trivial invariants: $C_n$ rotations and anti-unitary improper point group symmetries in MSGs cannot host nontrivial decorations.
Take $C_2$ for example.
When there is only a $C_2$ symmetry alone, consider a 2D plane that passes the $C_2$-axis, which is partitioned into two 2-cells by the axis.
Chern insulators cannot be decorated on the two 2-cells, because their chiral edge modes on the $C_2$-axis are $C_2$-related, and as such are in the same direction and cannot gap out each other.
This is different from the non-magnetic case where $C_2$ has a $\mathbb{Z}_2$ invariant\cite{song2019topological, song2018quantitative}, as the building blocks in non-magnetic SGs are 2D TIs and two $C_2$-related helical edge states can cancel with each other.
\item $\mathbb{Z}$ invariants: Unlike non-magnetic MSGs where only mirror symmetries have $\mathbb{Z}$ invariant, here in MSGs, as the building blocks are both $\mathbb{Z}$-type, we have many other $\mathbb{Z}$ invariants, including the invariant of unitary translation, unitary screw, and anti-unitary glide.
\item $\mathbb{Z}_2$ invariants:
A large proportion of MSG symmetries protect $\mathbb{Z}_2$ invariants, with many of them unique to MSGs, such as anti-unitary translations/rotations/screws.
In Appendix.\ref{AppendixD}, we define a special type of decoration called ``$Z_2$ decoration'' which has zero weak invariants and all $\mathbb{Z}_2$ invariants bond together and equal to 1 (for mirror Chern numbers, we can define $C_m^++C_m^-\bmod 2$ as a $\mathbb{Z}_2$ invariant), and can be seen as an axion insulator with the axion angle $\theta=\pi$. For $Z_2$ decorations, all these $\mathbb{Z}_2$ invariants merge into one $\mathbb{Z}_2$ invariant, i.e., the ``axion invariant'', whose definition can be taken as the 3D magnetoelectric polarization $P_3$ according to Ref.\cite{qi2008topological, qi2011topological,fang2012bulk}.
\end{itemize}
\begin{table}
\centering
\begin{tabular}{cc}
\hline\hline
\textbf{MSG Symmetries} & \textbf{Invariant type}
\\
\hline
unitary rotation $ C_n $& \multirow{3}{*}{Trivial} \\
\makecell[c]{anti-unitary improper point group \\
symmetries $P\cdot T$, $M\cdot T$, $S_n\cdot T$} &
\\
\hline
unitary improper $ S_n$, $P$, $\{M|\frac{1}{2}\}$ & \multirow{3}{*}{$\mathbb{Z}_2$} \\
anti-unitary translation $ \{ E|t\}\cdot T$ & \\
anti-unitary proper $C_n\cdot T$, $\{C_n|t\}\cdot T$ & \\
\hline
unitary translation $\{ E | R\}$ & \multirow{4}{*}{$\mathbb{Z}$} \\
unitary mirror $M$ & \\
unitary screw $ \{ C_n | t\} $ & \\
anti-unitary glide $ \{ M | \frac{1}{2} \}\cdot T$ & \\
\hline\hline
\end{tabular}
\caption{\label{table_invariants} MSG symmetries and their corresponding invariant types. We use the Seitz Symbol $\{O|t\}$ to represent symmetries with nonzero translations, where $R$ denotes a lattice translation and $t$ a fractional translation. The invariants of the unitary screw and anti-unitary glide are bound to the translation invariant, with their values being $\frac{1}{n}$ and $\frac{1}{2}$ of the translation invariant, respectively.}
\end{table}
Each nontrivial invariant has its distinct anomalous surface state due to the bulk-boundary correspondence.
Because the topological invariants together with their surface states can be superimposed, we only need to derive the surface state for every single invariant, and the surface states of all TCIs can be readily known from their invariants.
Among these surface states in MSGs, some have already been discussed in previous works, including those protected by $C_2 \cdot T$
\cite{shiozaki2014topology,fang2015new, ahn2019symmetry}, $C_4\cdot T$\cite{schindler2018higher}, glide\cite{fang2015new,shiozaki2015z,shiozaki2016topology,kim2019glide,kim2020glide}, and anti-unitary translation\cite{mong2010antiferromagnetic,fang2013topological}, etc. (for spinless invariants including $C_n\cdot T$ and anti-unitary translation, see Ref.\cite{liu2013antiferromagnetic,zhang2015topological}), while some are first proposed in this work, such as the one protected by anti-unitary glide.
We plot these surface states in Fig.\ref{surface_states}, and more details can be found in Appendix.\ref{AppendixC}.
As an example shown in Fig.\ref{surface_states}(3), on a cylinder geometry, the surface states protected by $C_n \cdot T, n = 2,4,6$ have a single Dirac cone on the top surface, and $n$ chiral hinge modes related to each other under $ C_n\cdot T$ on the side surface.
Although we find all the topological invariants protected by MSG symmetries, we have not derived their explicit formulas in k-space, which could be complicated.
Instead, we compute these topological invariants in real-space for the TCs we built using a unified method as follows\cite{song2019topological}.
Given an MSG $M$ and an element $g\in M$, first choose a generic point $r$ inside an arbitrary AU, then draw a path connecting $r$ and its image point $g \cdot r$, under the only constraint that the path does not cross any 1-cells and 0-cells. The corresponding invariant $\delta(g)$ is determined by the decorated 2-cells that the path crosses, i.e., the total Chern number or mirror Chern number accumulated through the path. The invariants thus calculated are well-defined and do not depend on the choice of the generic point $r$ or the path.
The full listing of topological invariants of the TCs will prove to be useful in the second part of this work, where we use SIs to diagnose topological states, and TCs function as an intermediate to connect SIs and invariants.
We take one example to show the correspondence between invariants and surface states for a given MSG.
The type-III MSG 3.3 $P2'$ has three independent generators of TCs.
One of them is protected by $C_{2y}\cdot T$ with nonzero invariant $\delta(C_{2y}\cdot T) = 1$, while the other two are protected by the translation symmetries in $x$ and $z$ directions, with nonzero weak invariant $\delta( \{ E | 100 \} ) = 1$ and $\delta{\{ E | 001\} } = 1 $, respectively.
For the first decoration, the surface state protected by $C_{2y}\cdot T$ has been described before, as shown in Fig.\ref{surface_states}(3), while the other two translation decorations have one chiral surface mode on the 2D surface preserving the translation symmetry in $x/z$ direction, as shown in Fig.\ref{surface_states}(1).
\subsection{Non-layer constructions}
The abundance of non-LCs distinguishes the TCs in MSGs from those in non-magnetic SGs, where most of the decorations are LCs.
In fact, LCs are just a special type of TCs where 2D planes are uniformly decorated, while non-LCs contain non-uniform decorations or incomplete 2D planes.
We use type-1 MSG $Pmmm$ as a representative to show the characters of non-LCs.
$Pmmm$ has three orthogonal mirrors $M_x, M_y$, and $M_z$, which do not commute with each other due to the spin rotation.
Placing mirror Chern insulator layers with $ ( C^{+}_m, C^{-}_m ) = ( 1,-1 ) $ on any of the six mirror planes, i.e., $x,y,z = 0,\frac{1}{2}$, forms 6 independent LCs.
However, one can still construct a distinct non-layer decoration falling outside the linear space of these 6 LCs.
As shown in Fig.\ref{Pmmm_nonLC}, this non-layer
decoration is constructed by sewing small patches of mirror Chern insulators with $ ( C^{+}_m, C^{-}_m ) = (1, 0)$ or $ ( 0, -1 ) $ on each mirror 2-cell, with adjacent patches having opposite mirror Chern numbers, making each mirror plane non-uniform.
Each 1-cell is shared by 4 patches, with patches in different directions contributing opposite chiral edge modes, canceling with each other and satisfying the no-open-edge condition.
This non-LC is distinct from LCs, where each mirror plane is decorated with a uniform, infinite-sized mirror Chern insulator.
The decorated 2-cells of this non-LC are all pinned on the mirror planes, preventing it to be deformed into an LC.
\begin{figure}
\centering
\includegraphics[width=0.2\textwidth]{Pmmm_nonLC.pdf}
\caption{\label{Pmmm_nonLC}The non-LC of $Pmmm$, where the arrows denote the directions of the chiral edge modes on the 2-cells. Note we only plot the 2-cells around the origin point inside the unit cell, and omit the six side-surfaces for simplicity, which are in fact all decorated. Assume the 8 inversion centers have coordinate $\bm{r}=\frac{1}{2}\mathbf{a}_1\delta_1+\frac{1}{2}\mathbf{a}_2\delta_2+\frac{1}{2}\mathbf{a}_3\delta_3$, where $\delta_i=0,1$ and $\mathbf{a}_i$ is the lattice vector. The 2-cells around an inversion center have the same decoration as shown in the figure if $\sum_i\delta_i$ is even, while the 2-cells have opposite directional edge modes if $\sum_i\delta_i$ is odd.}
\end{figure}
Although this non-LC seems complicated, we observe that it can be connected to the time-reversal strong TI (STI) from the perspective of surface states.
On the one hand, this non-LC has mirror Chern numbers $(C_{m,k_i=0}^+,C_{m,k_i=0}^-,C_{m,k_i=\pi}^+,C_{m,k_i=\pi}^-)=(1,-1,0,0), i=x,y,z$, which lead to a nontrivial mirror-protected surface state with a single Dirac cone on $k_i= 0$ in the 2D surface BZ.
On the other hand, an STI put on $Pmmm$ lattice also has a single Dirac cone on each surface.
By the principle of bulk-edge correspondence, we conclude that the STI with $Pmmm$ symmetry can be adiabatically deformed into this non-LC by breaking the TRS while preserving crystalline symmetries.
Note that for $Pmmm$, despite the absence of TRS, the three mirrors anti-commuting with each other also enforce the energy bands to form Kramer-like pairs, i.e., twofold degeneracy with the same parity and opposite mirror eigenvalues.
Therefore, the three mirror symmetries still pin the surface Dirac cone on the mirror invariant lines when the TRS is broken.
As a result, we can use the tight-binding models of STIs, for example, the 3D Bernevig-Hughes-Zhang (BHZ) model\cite{bernevig2006quantum}, to describe this non-LC.
As mentioned before, unlike non-magnetic SGs where only 12 SGs have non-LCs\cite{song2019topological}, non-LCs exist widely in MSGs.
For instance, all the supergroups of MSG 47.249 $Pmmm$ can host a non-LC as one of its TCI classification generators, and similar for MSG 25.57 $Pmm2$, 84.51 $P4_2/m$, 10.44 $P2'/m$, 75.3 $P4'$, 6.21 $P_am$, and 75.5 $P_C4$, thus the total number of MSGs that have non-LCs is at least 553, i.e., the number of the supergroups of these MSGs, including MSGs with different Bravais lattices.
As argued in Ref.\cite{song2019topological}, STI is compatible with all crystalline symmetries, thus the abundance of non-LCs in MSGs can be understood in a way that they can be obtained by inducing magnetism in STIs in different lattices.
We remark that the non-LCs in these MSGs, when being a ``$Z_2$ decoration'', are all axion insulators. As STIs are also axion insulators, when the TRS is broken, there still exist other unitary improper or anti-unitary proper symmetries that have non-trivial $\mathbb{Z}_2$ invariants, preserving the quantized $\pi$ axion angle.
\section{Diagnosis}
Symmetry-based indicator is a powerful tool for diagnosing topological states\cite{bradlyn2017topological, po2017symmetry, kruthoff2017topological} and has been applied in both non-magnetic SGs\cite{song2018quantitative, song2018diagnosis, khalaf2018symmetry} and MSGs\cite{ono2018unified, watanabe2018structure, elcoro2020magnetic, bouhon2020topological}, which leads to the discovery of a significant number of new topological materials\cite{zhang2019catalogue, vergniory2019complete, tang2019comprehensive, xu2020high}.
SI theory in MSGs has been partly tackled in previous works, with the SI group structures and part of the SI expressions given.
However, the mappings between SIs to topological invariants in MSGs have not been fully investigated.
In this work, we derive the explicit formulas of all indicators, and their quantitative mappings to topological invariants if correspond to gapped states, and possible Weyl point configurations if correspond to gapless states.
These two different correspondences to gapped and gapless states are found thanks to the TCI classifications, which include all the gapped TCI states, and states having nonzero SI while not belonging to any gapped state must be Weyl semimetals.
\subsection{Explicit expression of indicators}
Among 1421 MSGs, 688 of them have nontrivial SIs.
Despite the seemingly large number, SIs in all MSGs can be induced from 16 generating MSGs, which we list in Table.\ref{generatingSI}, with only 10 corner cases to be discussed later.
\begin{table}
\centering
\begin{tabular}{c|c|c}
\hline\hline
\textbf{MSG} & $\bm{X}_{\text{BS}}$ & \textbf{SI} \\\hline
$P\overline{1}$ & $\mathbb{Z}_{2,2,2,4}$ & $z_{2P,1},z_{2P,2},z_{2P,3},z_{4P}$ \\\hline
$Pmmm$ & $\mathbb{Z}_{2,2,2,4}$ & $z_{2P,1}',z_{2P,2}',z_{2P,3}',z_{4P}'$ \\\hline
$Pn,n=2,3,4,6$ & $\mathbb{Z}_n$ & $z_{nC}$ \\\hline
$Pn/m,n=2,3,4,6$ & $\mathbb{Z}_{n,n,n}$ & $z_{nm,0}^+,z_{nm,0}^-,z_{nm,\pi}^+$ \\\hline
$P\overline{4}$ & $\mathbb{Z}_{2,2,4}$ & $z_{2,S_4}, z_{2,\text{Weyl}}, z_{4C}$ \\\hline
$P4/mmm$ & $\mathbb{Z}_{2,4,8}$ & $z_{2P,1}^\prime, z_{4m,\pi}^+, z_8$ \\\hline
$P6/mmm$ & $\mathbb{Z}_{6,12}$ & $z_{6m,\pi}^+, z_{12}$ \\\hline
$Pnc'c',n=2,4,6$ & $\mathbb{Z}_{n}$ & $z_{nC}^\prime$ \\
\hline\hline
\end{tabular}
\caption{\label{generatingSI}Generating SIs and generating MSGs in MSGs. We use a simplified notation to represent the SI group, i.e., $\mathbb{Z}_{n_1,n_2,\cdots}=\mathbb{Z}_{n_1}\times\mathbb{Z}_{n_2}\times\cdots$. Among these generating MSGs, only $Pnc'c'$ are type-3 MSGs, while all the others are type-1. The definition and interpretation of these SIs can be found in Appendix.\ref{AppendixG}.}
\end{table}
Most of the SIs can be expressed in terms of some topological invariants.
In fact, the idea of diagnosing nontrivial topology using only the symmetry data on high-symmetry points (HSPs) in the BZ stems from the Fu-Kane formula\cite{fu2007topological} for non-magnetic SGs with inversion symmetry, and in the following works\cite{hughes2011inversion,turner2012quantized, fang2012bulk} $C_n$ rotation eigenvalues are used to calculate the Chern number modulo $n$, i.e., the $z_{nC}$ indicator in Table.\ref{generatingSI}.
When there exist mirror planes perpendicular to the rotation axis, the $z_{nC}$ indicator can be applied to two mirror sectors which give the $z_{nm,0/\pi}^\pm$ indicator, where $\pm$ represents two mirror sectors.
There are also three type-3 generating MSGs $Pnc'c',n=2,4,6$ which have anti-unitary glide symmetries that make all irreducible corepresentations (coirreps) twofold degenerate (with the same $C_n$ eigenvalues) on the $k_z=\pi$ plane, allowing the definition of a new set of SIs $z_{nC}^\prime$ using the number of degenerate pairs, the value of which corresponds to the Chern number divided by two modulo n of the 2D BZ.
The interpretations of other generating SIs are left in Appendix.\ref{AppendixG}.
The word ``generating'' means the SIs in other MSGs can be expressed in terms of these generating SIs.
Some times, the ``generation'' involves the reduction of the indicator group, e.g. from $\mathbb{Z}_4$ to $\mathbb{Z}_2$ by taking only the even numbers in $\mathbb{Z}_4$.
Moreover, the interpretation of the SI, generated from the same indicator, can be different in different MSGs.
For example, the $z_{nC}$ indicator originally represents the Berry phase of a closed loop in the 2D BZ as shown in Appendix.\ref{AppendixK}, and it becomes the Chern number when the state is gapped.
However, as shown later, in some other groups, this indicator if nonzero indicates gapless topological states, and its value represents the number of Weyl points.
For another example, the $z_{4P}^\prime$ indicator is adopted in some MSGs where the coirreps are not doubly degenerate with the same parity.
In these cases, the interpretation of these indicators may change which takes case-by-case examination, and the SI expressions are only effective and may become invalid when, for example, the origin point is changed.
Thus we fix the coordinate system to avoid these problems by adopting the Bilbao convention\cite{aroyo2006bilbao1,aroyo2006bilbao2,aroyo2011crystallography} when calculating the indicators.
To sum up, the indicator formulas in generating SIs can be used to express all SIs in other MSGs, except 10 corner-case MSGs which have SI formulas given independently using coirreps in Appendix.\ref{AppendixH}.
Here we show an example of MSG 42.222 $Fm'm'2$, which has a $\mathbb{Z}_2$ indicator with expression:
\begin{equation}
z_{2,42.222} = N(\overline{\Gamma}_3) - N(\overline{A}_3) \bmod 2
\end{equation}
where $N(\bar{K}_i)$ represents the number of coirrep $\bar{K}_i$.
This MSG has TCI classification $\mathbb{Z}$, the generator of which has weak invariants $\delta_w=(1,1,0)$ and all other invariants equal zero.
The SI $z_{2,42.222}=1$ represents this generator as well as all the odd number copies of it.
Note that this indicator is different from the choice in Ref.\cite{elcoro2020magnetic}.
\subsection{Computation of indicators for topological crystals}
In this section, we show how one can calculate the SIs for a given TC.
As most of the SIs correspond to some topological invariants, including $z_{2P,i},z_{4P},z_{2P}^\prime,z_{nC},z_{nm,0/\pi}^\pm,z_{2,S_4}$, and $z_{nC}^\prime$, their values can be determined directly.
For example, the $\mathbb{Z}_{2}\times\mathbb{Z}_{2}\times\mathbb{Z}_{2}\times\mathbb{Z}_{4}$ indicators defined in $P\bar{1}$ have straightforward meanings:
\begin{equation}
z_{2P,i}=\delta_{w,i}\bmod 2,\ \ z_{4P}=2\delta(P)
\end{equation}
where $\delta_{w,i}(i=1,2,3)$ is the weak invariant and $\delta(P)$ is the inversion invariant.
Note that $z_{4P}=1,3$ represent Weyl semimetals\cite{hughes2011inversion, turner2012quantized}.
However, there also exist many SIs that cannot be read directly from the invariants, including $z_{4P}^\prime$, $z_8$, and $z_{12}$.
To determine their value for a TC, we need to find a compatible set of coirreps for the TC and then calculate the SIs of the coirrep set.
The result does not, however, depend on which particular set we choose as long as it is compatible with the TC.
This is because a given set of invariants, that is, a given TC, can only correspond to one possible set of SIs.
Therefore, we only need to find one compatible set of coirreps for the TC and calculate its SI.
Finding compatible coirreps for an LC is simple, but to a non-LC TC can be challenging.
Fortunately, as shown before that many of the non-LCs in MSGs are adiabatically connected to STIs, the BHZ model describing STIs can be used, which has symmetries of SG 221 plus the TRS, a supergroup to many MSGs.
As a result, we can adjust the model parameters such that the BHZ model and the non-LCs share the same set of invariants, and then compute the indicators for the BHZ model, with details in Appendix.\ref{AppendixL}.
The mappings between invariants and SIs only need to be derived for the generating MSGs shown in Table.\ref{generatingSI}.
Other MSGs, except the corner cases and Weyl states, either are the supergroups of these generating MSGs or have a different Bravais lattice, and their mappings can be induced from the mappings of the generating MSGs.
For corner cases, we derive their mappings case-by-case in Appendix.\ref{AppendixH}.
\subsection{The mapping from indicators to invariants}
So far we have obtained the mapping from invariants to indicators for topological gapped states protected by all MSGs, listed in Appendix.\ref{AppendixO}.
A practically more useful mapping is its inverse, one from indicators to invariants.
This is because, on one hand, the indicators only depend on the coirreps at HSPs, and as such are far easier to obtain in first-principles calculations.
On the other hand, the invariants, as shown in the previous section, are directly related to the topological surface states that may be detected experimentally.
A mapping from indicator to invariants hence links first-principles calculation to experimental observables.
These mappings from SIs to invariants are ``one-to-many'' in general, in which a given SI set may be mapped to multiple invariant sets.
Assume an MSG has $m$ linearly independent TCs, which have invariant sets $V_1, V_2, ..., V_m$ and SI sets $S_1, S_2, ..., S_m$.
Given an arbitrary SI set $S$ from the SI group, to find all the possible invariant sets corresponding to it, one needs to solve the linear equation $\sum_{i=1}^m x_i S_i=S$, where $x_i$ are coefficients to be determined.
The invariant sets corresponding to $S$ can be obtained using the solved $x_i$.
This inhomogeneous linear equation can be solved by first finding a special solution and then adding all the general solutions to its homogeneous counterpart, i.e., the solutions of the corresponding homogeneous linear equation, $\sum_{i=1}^m x_i S_i=0$, which gives the invariant sets that have zero SIs.
Following this line, we have developed an algorithm that automatically computes all possible sets of invariants for a given set of indicators.
The code and the full results which have been entered into a large table may be downloaded from
\href{https://github.com/yjiang-iop/MSG_mapping}{https://github.com/yjiang-iop/MSG\_mapping}.
\subsection{Weyl semimetal indicators}
Unlike non-magnetic spinful SGs\cite{song2018quantitative} where all SIs indicate topological gapped states, we find many SIs in MSGs correspond to gapless Weyl semimetals.
As the full mapping from TCs to SIs has been found for each MSG, indicator values that do not belong to the image of this mapping necessarily correspond to gapless states.
Although these Weyl semimetals have Weyl points at generic momenta in the BZ, the symmetries of MSGs require their creation or annihilation to happen at HSPs, which makes it possible to detect them using SIs.
These Weyl semimetals can be classified into two types, with one being inter-plane Weyl points that lie between $k_i=0$ and $k_i=\pi$ plane, and the other being in-plane Weyl points that lie on $k_i=0$ or $\pi$ plane, where $i$ denotes the main rotation axis of the MSG.
We leave the full discussion of Weyl semimetals to Appendix.\ref{AppendixI}, and use one example of type-4 MSG 81.37 $P_C\overline{4}$ to illustrate the basic idea.
MSG 81.37 can host a multiple of 4 Weyl points on both $k_z=0/\pi$ plane, with Weyl points connected by $S_4$ symmetry having opposite chirality, as shown in Fig.\ref{81_37weyl}. This MSG has TCI classification $\mathbb{Z}_2$ and SI group $\mathbb{Z}_2\times\mathbb{Z}_2$.
The SI group is larger than the TCI group.
One of the $\mathbb{Z}_2$ indicator can be chosen as $z_{2,S_4}$, the odd value of which corresponds to the decoration with $S_4$ invariant $\delta(S_4)=1$, while the other $\mathbb{Z}_2$ indicator can be chosen as $z_{4C}/2$, the odd value of which represents the Weyl states.
Notice that the Weyl points are $S_4$-symmetric, and they can only be annihilated at HSPs on $k_z=0/\pi$ plane.
Moreover, $z_{4C,k_z=0}=z_{4C,k_z=\pi}=0$ or $2$ always holds, which means the Weyl points will appear simultaneously on $k_z=0,\pi$ planes.
The in-plane nature of this type of Weyl states is enforced by the $C_2\cdot T$ symmetry.
\begin{figure}[H]
\centering
\includegraphics[width=0.2\textwidth]{81_37weyl.pdf}
\caption{\label{81_37weyl}The Weyl point configuration of MSG 81.37, where the blue and orange dots represent Weyl points of opposite chiralities, the red line represents the $S_4$ axis, and two yellow planes represent $k_z=0,\pi$ planes.}
\end{figure}
|
\section{Introduction}
Using pre-trained contextual representations (e.g., BERT) \cite{devlin2018bert} have become the standard way to improve the performance on the downstream tasks in natural language processing. Transformer \cite{vaswani2017attention} is the basic building block for almost all pre-training methods \cite{liu2019roberta, clark2019electra, devlin2018bert}. A Transformer layer is composed of an efficient densely connected network operated on each position separately and a less-efficient self-attention module, which costs $O(n^2)$ ($n$ is sequence length) time and space. This quadratic cost becomes a bottleneck in Transformer, especially when $n$ is large. Many recent works \cite{wang2020linformer, beltagy2020longformer, zaheer2020big, kitaev2020reformer, choromanski2020rethinking} tried to reduce the $O(n^2)$ cost to $O(n\sqrt{n})$ or $O(n \text{log} n)$, by sparsifying or approximating the attention matrix.
\begin{figure}[t]
\centering
\includegraphics[width=0.55\textwidth]{arch.pdf}
\caption{The basic block in LazyFormer.}
\label{fig:arch}
\end{figure}
In this paper, different from previous works, we explore a simple idea to improve the model efficiency by computing the self-attention \emph{infrequently}. We call it \emph{LazyFormer}. More specifically, a LazyFormer consists of multiple basic blocks, and each basic block is composed of $m$ Transformer layers, as shown in Figure \ref{fig:arch}. In each block, we only compute Query-Key dot-product attention once in the first layer and then reuse it in the upper layers. In this way, LazyFormer only needs to calculate the attention once in $m$ Transformer layers, reducing the computational cost from $O(n^2)$ to $O(n^2/m)$.
We conduct extensive experiments to verify the efficiency and effectiveness of LazyFormer in language pre-training. From the results, we observe: 1) compared with a standard Transformer model of the same model capacity, LazyFormer is 1.3x faster without hurting any performance. 2) LazyFormer allows us to train the larger models effectively. Specifically, with the same pre-training cost, the larger LazyFormer can outperform the baseline by 1 point. Besides, it can achieve a better GLUE score by only using 50\% costs.
3) LazyFormer can better handle longer sequences. With $n$ increasing, the speed-up is more significant.
\paragraph{Related work}
LazyFormer is inspired by several recent works which investigate what self-attention module learns in each layer \cite{clark2019does, vig2019analyzing, vig2019multiscale, xiao2019sharing}. From these previous works, we can easily see that the distributions of self-attention outputs are similar between the adjacent Transformer layers. Such observations motivate us to reuse the self-attention outputs in the lower layers to the upper layers.
There are some works that leverage such observation for better training, like \cite{gong2019efficient} and \cite{lan2019albert}. However, these works share the parameters of self-attention in different layers but still need to compute self-attention in all layers. Therefore they cannot save the computational costs. Our work is also orthogonal to the works that modify the self-attention modules to reduce the cost, such as Reformer\cite{kitaev2020reformer} and Linformer\cite{wang2020linformer}. These works aim to reduce the computation in each layer, while ours is to reuse the self-attention outputs from lower layers. Both works could be combined and further reduce the cost of language pre-training.
\section{LazyFormer}
The attention module \citep{vaswani2017attention} in Transformer can be generally formulated as querying a dictionary with key-value pairs, e.g., $\text{Attention}(Q,K,V)=\text{softmax}(\frac{QK^T}{\sqrt{d}})V$, where $d$ is the dimensionality of the hidden representations. In the self-attention module, $Q$ (Query), $K$ (Key) and $V$ (Value) are parameterized from the input $x$, i.e., $\text{Attention}(xW^Q,xW^K,xW^V)$, where $W^Q$, $W^K$ and $W^V$ are the projection matrices. It is easy to see that the computational complexity of $\text{softmax}(\frac{QK^T}{\sqrt{d}})$ is $O(n^2)$, since it has to compute the pair-wise correlations for all input tokens, and produces an $n \times n$ matrix (attention matrix). When we use stacked $k$ Transformer layers, as all attention matrices need to be computed, the total cost of self-attention calculation is $O(kn^2)$.
Many recent works \cite{vig2019analyzing, vig2019multiscale} show the attention matrices are similar in different layers, especially in the adjacent layers. Therefore, we argue that the attention matrix maybe does not need to be computed in every layer. For example, we can only calculate the attentions in a layer and reuse it for multiple adjacent upper layers. Formally, we define a new Transformer variant called \emph{LazyFormer}, which composes of several ``lazy'' blocks. Each lazy block consists of $m$ Transformer layers, as shown in Figure \ref{fig:arch}. In each block, the attention matrix will be only computed in the first layer using the input to the block, and then reused by all $m-1$ upper layers. We can stack $k/m$ lazy blocks to construct a $k$-layer Transformer model. In this way, the cost of computing the attention could be reduced from $O(kn^2)$ to $O(kn^2/m)$.
Based on the design of lazy block, we further use the following two additional methods to improve LazyFormer:
\paragraph{Wider Layers.} In LazyFormer, the number of projection matrices $W^Q$ and $W^K$ are reduced. Therefore, the total parameters in LazyFormer are less than the stacked Transformers when using the same width and depth. For example, the BERT-based model has 12 layers. The embedding dimension is set to 768, and the hidden dimension is set to 3072. This configuration leads to a model size of about 110M parameters. If we use $m=3$ for LazyFormer in the same setting, the model only contains about 100M parameters.
As the model capacity plays an important role in language pre-training \cite{raffel2019exploring}, we can slightly increase the hidden/embedding dimension in LazyFormer, to match the same number of parameters. Note that increasing the hidden/embedding dimension only slightly affects the efficiency. Wider LazyFormer is still much faster than a Transformer of the same depth. Besides, we can even increase the width of LazyFormer until its forward/backward speed match the Transformer to achieve better performance when using the same pre-training cost.
\paragraph{Remove dropout in self-attention.} Dropout is used in self-attention by default. The cost of that dropout is also $O(n^2)$, as it is applied on the $n \times n$ attention matrix. Recent work \cite{lan2019albert} shows the dropout in self-attention can be safely removed, without hurting the performance. Therefore, for better efficiency, we also remove the dropout in self-attention in LazyFormer.
\section{Experiment}
To verify the performance of the proposed LazyFormer, we conduct extensive experiments and demonstrate the results in this section. We use BERT-Base (112M parameters) architecture for all experiments. Specifically, BERT-Base is consists of 12 Transformer layers, in which the embedding dimension is 768, the number of attention heads is 12, and the hidden dimension is 3072. Besides absolute positional encoding, we further use the relative positional encoding \cite{raffel2019exploring} in the self-attention module for better performance. We provide all the experimental details and results in the Appendix.
\subsection{Experimental Design\label{sec:exprimentdesign}}
Following \citet{devlin2018bert}, we use the 16GB corpus (English Wikipedia corpus and BookCorpus \citep{moviebook}) for pre-training. We set the vocabulary size (sub-word tokens) as 32,768, sequence length as 512, and batch size as 256. We use the GLUE (\textbf{G}eneral \textbf{L}anguage \textbf{U}nderstanding \textbf{E}valuation) dataset \citep{DBLP:journals/corr/abs-1804-07461} as the downstream tasks to evaluate the performance of the pre-trained models. All codes are implemented based on \emph{fairseq} \citep{ott2019fairseq} in \emph{PyTorch} \citep{paszke2017automatic}. All models are run on 8 NVIDIA Tesla V100 GPUs with mixed-precision \citep{micikevicius2017mixed}.
We use M$\beta$x$\gamma$ to denote LazyFormer structure, where $\beta$ is the number of layers in each lazy block, and $\gamma$ is number of total blocks. For example, M2x6 denotes the LazyFormer with 6 blocks, each with 2 Transformer layers.
\subsection{Overall Comparison}
\begin{table*}[t]
\centering
\caption{GLUE scores on dev set. All models are pre-trained by 16GB data. Both M2x6 and M2x6-S are the LazyFormer with 6 two-layered blocks, without dropout in self-attention. M2x6-S uses the same parameter size as BERT. For the ablation study, based on M2x6-S, M2x6-SD keeps the dropout in self-attention. M2x6 increases the parameter to match the same pre-training cost as BERT. M2x6$^{mid}$ is the intermediate 500k-step checkpoint M2x6. The details of models are shown in Table \ref{tab:overall_meta}.}
\label{tab:overall}
\begin{tabular}{lcccccccccccc}
\toprule
& Steps & MNLI\footnotesize{-m/mm} & QNLI & QQP & SST & CoLA & MRPC & RTE & STS & Avg. \\
\hline
BERT&1$M$ & \textbf{85.72/85.67} & \textbf{92.07} & 91.21 & 92.78 & \textbf{58.80} & \textbf{88.73} & 67.51 & 89.23 & 83.52 \\
M2x6-S &1$M$ & 85.69/85.46 & 91.14 & 91.13 & 92.66 & 57.86 & 87.25 & \textbf{72.20} & \textbf{89.79} & \textbf{83.69} \\
M2x6-SD &1$M$ & 85.34/85.50 & 91.65 & \textbf{91.22} & \textbf{93.23} & 57.85 & 86.76 & 70.40 & 89.69 & 83.52 \\
M2x6 &1$M$ & \textbf{86.34/86.31} & \textbf{91.82} & \textbf{91.38} & \textbf{93.00} & \textbf{60.77} & \textbf{87.75} & \textbf{73.65} & \textbf{90.00} & \textbf{84.56} \\
\midrule
M2x6$^{mid}$ &500$k$ & 85.81/85.58 & 91.62 & 91.31 & 92.66 & 58.81 & 87.75 & 71.48 & 88.90 & 83.77 \\
\bottomrule
\end{tabular}
\end{table*}
\begin{figure*}[b]
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=2.1in]{valid_all.pdf}
\caption{Validation loss in pre-training.}
\label{fig:exp-all-valid}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=2.1in]{mnli_all.pdf}
\caption{MNLI-m score.}
\label{fig:exp-all-mnli}
\end{subfigure}
\begin{subfigure}{.33\textwidth}
\centering
\includegraphics[width=2.1in]{glue_all.pdf}
\caption{GLUE average score.}
\label{fig:exp-all-glue}
\end{subfigure}
\caption{Both M2x6-S and M2x6 converge much faster than the baselines. Besides, M2x6 achieves better performance in downstream tasks while using much fewer pre-training costs.}
\label{fig:exp-overall}
\end{figure*}
\begin{table}[t]
\centering
\caption{Details of models in Table \ref{tab:overall}. $W$ is the hidden dimension in feed-forward layers, $H$ is the embedding dimension, and $N$ is the number of attention heads.}
\label{tab:overall_meta}
\begin{tabular}{lcccc}
\toprule
& Params & ($W$,$H$,$N$) & Time \protect\footnotemark & Speedup \\
\hline
BERT & 112$M$ & (3072,768,12) & 45 & 1.0x \\
M2x6-S& 112$M$ & (3456,768,12) & 35 & 1.3x \\
M2x6-SD & 112$M$ & (3456,768,12) & 38 & 1.2x \\
M2x6 & 157$M$ & (4480,896,14) & 45 & 1.0x \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[h]
\centering
\caption{Comparison for different LazyFormer layouts. To match the same parameter size as BERT, we increase the hidden dimension $W$ for different layouts. We keep the dropout in self-attention in all settings.}
\label{tab:layer}
\begin{tabular}{lccc}
\toprule
& $W$ & Time \footref{footnote1} & GLUE-Avg. \\
\hline
BERT (M1x12-SD)& 3072 & 45 & \textbf{83.52} \\
M2x6-SD & 3456 & 38 & \textbf{83.52} \\
M3x4-SD & 3584 & 35 & 82.94 \\
M4x3-SD & 3648 & 34 & 83.22 \\
M6x2-SD & 3712 & 33 & 82.47 \\
\midrule
M5M3M2M2-SD & 3584 & 35 & 82.89 \\
M2M2M3M5-SD & 3584 & 35 & 82.82 \\
\bottomrule
\end{tabular}
\end{table}
First, we set up models for the overall comparison. Besides the baseline BERT model, we set up two LazyFormer variants: 1) \textbf{M2x6-S}, which uses six lazy blocks with two Transformer layers in each block, and increases hidden dimension to 3456, to retain the same parameter size as BERT; 2) \textbf{M2x6}, which increases hidden dimension to 4480, embedding dimension to 896, and the number of attention heads to 14, to retain the same pre-training cost as BERT. Both M2x6 and M2x6-S remove dropout in the self-attention module. The detailed settings can be found in Table~\ref{tab:overall_meta}.
The results are shown in Table~\ref{tab:overall}. Firstly, M2x6-S achieves a slight improvement over BERT but is about 1.3x faster. This result indicates that LazyFormer is much more efficient, without hurting any performance.
Furthermore, M2x6 outperforms baselines by 1 point in terms of GLUE average score and is consistently better on almost all GLUE tasks. This demonstrates another important strength of LazyFormer: it allows us to increase the model capacity to achieve better performance while using the same computational cost. Besides, from Table 2, we can also see that the results of M2x6$^{mid}$ (intermediate 500k-step checkpoint of M2x6) are already competitive to that of BERT (trained for 1M steps). This suggests LazyFormer can use a significantly short time to learn a better model. As shown in Figure~~\ref{fig:exp-overall}, both M2x6 and M2x6-S converge much faster than BERT in terms of both pre-training validation loss and down-stream task performance.
\footnotetext{Time indicates the training wall time of 100 iterations by 8 V100 GPUs. \label{footnote1}}
\subsection{Ablation Studies}
\paragraph{Remove dropout in self-attention.} As aforementioned, dropout in self-attention brings additional costs. So we empirically study whether using dropout is essential. First, as shown in Table \ref{tab:overall}, compared M2x6-S with M2x6-SD (M2x6-S with dropout in self-attention), removing dropout in self-attention slightly improves the performance. Besides, as shown in Table \ref{tab:overall_meta}, removing dropout in self-attention can bring the 10\% speed-up. In short, removing dropout in self-attention can improve efficiency without hurting the performance.
\paragraph{Different layouts.} There are many design choices in LazyFormer. For example, one can set low-level blocks with more/fewer layers or set the whole model with more/fewer blocks. We study how different layouts perform and summarize the results in Table~\ref{tab:layer}.
First, we find that when the number of blocks decreases, the training is faster, but the acceleration rate is not significant. At the same time, the final performance gets worse. Therefore, we observe using six blocks is a trade-off choice for both efficiency and effectiveness.
Second, we find the number of blocks is the key factor for the performance. For all the 4-block models, setting each block with 3 layers (M3x4-SD) achieves similar performance to the models that set different numbers of layers in different blocks (M5M3M2M2-SD \footnote{M5M3M2M2 denotes the model with 4 sequential lazy blocks, each with 5, 3, 2, 2 layers respectively.}, M2M2M3M5-SD).
\begin{table}[t]
\centering
\caption{Comparison for different ($W$, $H$, $N$) settings under the same pre-training cost.}
\label{tab:vary}
\begin{tabular}{lcccc}
\toprule
& Params & ($W$,$H$,$N$) & MNLI\footnotesize{-m/mm} & GLUE-Avg. \\
\hline
M2x6-S& 112$M$ & (3456,768,12) & 85.72/85.67 & 83.69 \\
M2x6 & 157$M$ & (4480,896,14) & \textbf{86.34/86.31} & 84.56 \\
M2x6 & 161$M$ & (6144,768,12) & \textbf{86.08/86.57} & \textbf{84.77} \\
M2x7 & 148$M$ & (4480,768,12) & 86.25/86.26 & 84.75 \\
\bottomrule
\end{tabular}
\end{table}
\begin{figure}[t]
\centering
\includegraphics[width=0.60\textwidth]{time_fig3.pdf}
\hspace{+14pt}
\caption{Speedup ratio of \textit{LazyFormer} under different settings of lazy blocks. }
\label{speedup}
\end{figure}
\paragraph{Different model settings under the same computational cost.} As aforementioned, LazyFormer allows us to use a larger model capacity within the same computational cost. Therefore, we investigate different settings to increase model parameters and summarize the results in Table~\ref{tab:vary}. Specifically, we study the performance when we increase hidden dimension $W$, embedding dimension $H$, the number of attention heads $N$, or stack more blocks. From the results, we find all these settings can achieve much better performance than the baseline.
\paragraph{Speedup for longer sequences.} If considering the cost from densely connected layer, LazyFormer reduces the cost from $O(n^2 + p)$ to $O(n^2/m + p)$, where $p$ is the cost of densely connected layer. Therefore, with $n$ increasing, the cost is dominated by the $n^2$ term, and the speedup brought by LazyFormer will be also larger. To study this, we provide speedup ratios given different $n$, shown in Figure~\ref{speedup}. It is easy to find the LazyFormer can bring almost 2x speed-up for long sequences.
\section{Conclusion}
We propose LazyFormer, which lazily updates the self-attention module and lead to an efficient model architecture of the Transformer. Specifically, LazyFormer composes of multiple lazy blocks, each of which contains multiple Transformer layers. In each lazy block, the self-attention distribution is only computed once in the first layer and then is reused in all upper layers. Besides, LazyFormer further removes dropout in the self-attention module for better efficiency and increases model capacity for better performance. Extensive experimental results demonstrate both the efficiency and effectiveness of LazyFormer.
\clearpage
|
\section{Introduction}
\label{intro}
Time series are widely present in many domains such as industry, energy, meteorology, e-commerce, social networks or health. They represent the temporal evolving of systems and help us to understand their temporal dynamics and perform short-, medium- or long-term predictions. A major research line has been dedicated to time series analysis. In this line, exponential smoothing models \citep{gardner_2006_int-journ-forecasting_exponent-smooth-paramet-initial-val-select, bergmeir_hyndman_benitez_2016_intern-journ-of-forecasting_expon-smooth-stl-plus-box-cox}, Box and Jenkins models \citep{box_jenkins_reinsel_et_al_2015_book_time_series_analysis} and nonlinear autoregressive neural networks
\citep{yu_zhou_tan_2014_jour-plos-one_ARIMA-and-NARNN, wang_wang_yan_2019_journ-ieee_NARNN, noman_alkawsi_alkahtani_2020_journ-elsevier_NARNN-wind-time-series} are essentially devoted to forecasting. In addition to the forecasting goal, regime-switching autoregressive models \citep{ubilava_helmesr_2013_environ-modelling-softwa_smooth-transition-autoregres-model, hamilton_1990_journ-econometrics_time_series-regime-changes} also allow to discover hidden behaviors of such systems.
In the cases when the studied system is \textbf{stationary}, that is its behavior is time-independent, the Linear AutoRegressive (LAR) model is a framework widely used to capture the autoregressive dynamics of the corresponding time series \citep{wold_1954_book_stationary-time-series, degtyarev_Gankevich_2019_book_AR-model-of-irregular-waves}. The LAR model is a simple linear regression model in which predictors are lagged values of the current value in the time series. However, many real-life systems are subject to changes in behaviors: for instance in econometry, we distinguish between recession and expansionary phases; in meteorology, anticyclonic conditions alternate with low pressure conditions. These systems are commonly referred to as \textbf{regime-switching systems}, where each regime corresponds to a specific behavior. Each time-step is associated with some state, amongst those allowed for the system. Regime-switching system modelling is achieved in two steps: (i) the state process modelling that enables to capture how states are generated, and (ii) the modelling of the autoregressive dynamics of the time series within each regime. In the latter step, a simple autoregressive framework such as the LAR model can be used. Generally, in step (i), the state process is modelled by a \textbf{discrete-valued Markov process}. In the current state-of-the-art literature, two categories of models can be distinguished.
In {\bf Hidden Regime-Switching Autoregressive (HRSAR) models}, the state process is hidden and is modelled by a Hidden Markov Process (HMP). This category of models has been introduced by \cite*{hamilton_1989_econometrica_nonstationary-timeseries-markov-switching-ar} in the context of United States's Gross National Product time series analysis. Several variants and extensions were subsequently designed.
In {\bf Observed Regime-Switching Autoregressive (ORSAR) models}, the state process is either observed or derived \textit{a priori}. In the latter case, a clustering algorithm is used before fitting the model, to extract the regimes. The clustering may either rely on endogenous variables ({\it i.e.}, the variables whose dynamics is observed through the time series) or on exogenous variables supposed to drive regime-switching. The recent work of \cite*{bessac_ailliot_cattiaux_et_al_2016_advanc-stat-climat-met-ocean_hidd-obs-regime-switch-ar-model} illustrates the application of these models to wind time series.
When the state process is partially observed, which means that the system state is known at some random time-steps and unknown for the remaining time-steps, ORSAR models cannot be directly applied while HRSAR models are suboptimal in the sense that the observed states cannot be included.
To overcome these limitations, in this work, we propose a novel \textbf{regime-} \textbf{switching autoregressive} model that capitalizes on the observed states while the hidden states are inferred. We consider a special case of Markov process henceforth named \textbf{Markov Chain}. Our model is referred to as the Partially Hidden Markov Chain Linear AutoRegressive (PHMC-LAR) model. The PHMC-LAR model is a flexible parametric model that supplies a unification of HRSAR and ORSAR models when the state process is a Markov Chain. Thus, when the state process is fully observed, PHMC-LAR is reduced to ORSAR. Reversely, when the state process is fully hidden, PHMC-LAR instantiates as HRSAR. Beyond the unification aspect, we contribute to the machine learning literature through designing the underlying algorithmic machinery dedicated to effective and efficient PHMC-LAR model training.
The main contributions of this paper are as follows:
\begin{enumerate}
\item We propose a new regime-switching autoregressive model that integrates the states observed at some random time-steps. This model, referred to as PHMC-LAR, provides a unification of HRSAR and ORSAR models when the state process is modelled by a Markov Chain (MC).
\item We propose a variant of the \textbf{Expectation-Maximization} (EM) \textbf{algorithm} that allows to learn the parameters of our model.
\item Inference on hidden states is carried out by a variant of the \textbf{Viterbi algorithm}, adapted to take into account the observed states.
\item Regarding the time series forecasting task, a prediction function is proposed. We distinguish between the case where the system state is known at forecast horizons from the case where it is latent.
\end{enumerate}
The ability of our model to infer the hidden states and to make accurate predictions on time series, even when the observed states are unreliable, is investigated through experiments performed on synthetic data. Our work underlines the benefits of using partially observed states to decrease EM convergence times. This performance is obtained with no or practically no impact on the quality of hidden state inference, as from labelling percentages around 20$\%$-30$\%$; the prediction accuracy is also preserved above such percentage thresholds. For instance, for a training set of 100 sequences, with $70\%$ labelled states, the EM algorithm converges after 22 iterations on average against 62 on average for the unsupervised case. Moreover, performing fully supervised training with a proportion of ill-labelled states is also beneficial for EM convergence. For example, given a training set of size 100 annotated with a $70\%$-reliable labelling function, the EM algorithm converges after a single iteration against $67$ iterations for the unsupervised case. This offers promising prospects to enhance model selection for the PHMC-LAR model. Further experimentations also show the ability of our variant of the Viterbi algorithm to infer hidden states in partially-labelled sequences. In addition, while assessing the impact on predictions generated by incorporating labelled states in the training sequences, we also compared the situations where all states are unknown at forecast horizons to the situations where all states are known. Prediction errors are subdued at all horizons in the latter case (by 44\% on average), but contrasted horizons are still evidenced with low (respectively high) scores as in the former case. The constrast is kept constant whatever the percentage of observed states in the training set. Besides, we also point out the robustness of our model to labelling errors in inference task, when large training datasets and moderate labelling error rates are considered. Finally, the latter experiment highlights the remarkable robustness to error labelling in the prediction task, over the whole range of error rates.
This paper is organized as follows. Related work is reviewed in Section \ref{related.work}. Section \ref{proposed.model} describes the PHMC-LAR model. Then a learning algorithm is derived in Section \ref{learning.algorithm}, to estimate the model parameters. Inference of the hidden states is addressed in Section \ref{inference}. Section \ref{forecasting} presents the time series forecasting procedure. Section \ref{experiments} depicts the experimental protocol that drove our experimentations on synthetic data, and discusses the results obtained. Section \ref{conclusion} concludes this paper.
\section{Related work}
\label{related.work}
This section first highlights the links between our proposal, PHMC-LAR, and the most closely related contributions of the literature. The PHMC-LAR combines a variant of the Hidden Markov Model (HMM), namely the Partially Hidden Markov Chain (PHMC), with the Linear AutoRegressive (LAR) model. The rest of this section reviews the two main models that compose the hybrid model proposed.
As mentioned in the introduction, the PHMC-LAR model unifies the HRSAR and ORSAR frameworks. However, the common thread between these latter frameworks is the implication of dependencies that drive the local dynamics within each regime. Therefore, the contributions of the literature most closely related to PHMC-LAR are also characterized by various local dynamics.
Several models closely related to HRSAR were proposed in the literature. The MS-AR model (Markov-switching AutoRegressive model) designed by \cite*{hamilton_1989_econometrica_nonstationary-timeseries-markov-switching-ar} combines ARIMA (AutoRegressive Integrated Moving Average) models with an HMM, to characterize changes in the parameters of an autoregressive process. The targeted application motivating the MS-AR model was economic analysis: the switch between fast growth and slow growth is governed by the outcome of the Markov process.
Further, \cite*{filardo_1994_j-of-busin-econo-stat_markov-switching-autoregress-time-varying-transit} incorporated time-varying transition probabilities between regimes in the MS-AR model. For instance, the resulting model was subsequently used to reproduce the cyclic patterns existing in climatic variables \citep{cardenas-gallo_sanchez-silva_akhavan-tabatabaei_et_al_2016_natural-hazards_markov-switching-autoregress-time-varying-transit_climate}. In parallel, the Hamilton's MS-AR model was also extended into a general dynamic linear model combined with Markov-switching \citep{kim_1994-j-of-economet-markov-switching-general-dynamic-lin_model}. Finally, Michalek and co-authors'work focused on a HRSAR model that integrates HMM with Moving Average (MA) models \citep{michalek_wagner_timmer_2000_ieee-trans-on-signal-process_arma-hmm-approx-estim}. In the same work, the parameter estimation approximation thus derived was generalized to deal with AutoRegressive Moving Average (ARMA) hybridized with HMM. Simulations of electrophysiological recordings showed that the derived estimators allow to recover the true dynamics where standard HMM fails. The model generalized by Michalek and collaborators, to integrate HMM with ARMA, was also applied to model human activity as time signals for activity early recognition \citep{li_fu_2012_conf-pattern-reco_arma-hmm-human-activity-recogn}.
More recently, a nonhomogeneous HRSAR model was developed to model wind time series \citep{ailliot_2015_jour-statist-planning-inferen_non-homog-hidden-markov-switching-models}. The aim was to acknowledge that the probability of switching from cyclonic conditions to anticyclonic conditions between time-steps $t$ and $t+1$ depends on the wind conditions at time-step $t$ at some given location off the French Brittany coast. A nonhomogeneous MS-AR (NHMS-AR) model was thus designed for this purpose.
To our knowledge, the investigations around ORSAR models are limited to the recent work of \cite*{bessac_ailliot_cattiaux_et_al_2016_advanc-stat-climat-met-ocean_hidd-obs-regime-switch-ar-model} which was applied to wind time series. Therein, observed regimes are derived by running a clustering procedure on the variables under study or on extra variables. Thus are identified the states, all distinct from one other, in which the data are homogeneous. Besides comparing the ORSAR models derived from various clustering procedures, Bessac and collaborators also compare the respective merits of HRSAR and ORSAR models on real-world and simulated data.
\subsection{Partially Hidden Markov Chain - PHMC($K$)}
\indent Hidden Markov models (HMMs) have been successfully used in such domains as natural language processing \citep{morwal_jahan_chopra_2012_journ-natural-lang-comput_HMM-nat-language-process}, handwriting recognition \citep{mouhcine_mustapha_zouhir_2018_journ-elsevier_handwritten-recognition-with-HMM}, speech emotion recognition \citep{schuller_rigoll_lang_2003_ic-multimed-expo_hmm-speech-emot-recogni}, human action recognition \citep{berg_reckordt_richter_2018_procedia-cirp_human-action-recogni-hmm} or renewable power prediction \citep{ghasvarian-jahromi_gharavian_mahdiani_2020_soft-comput_solar-power_predict-hmm}, to name but a few.
HMM$(K$) is a flexible probabilistic \textit{framework} able to model complex hidden-regime-switching systems. It exactly possesses $K$ states where each state drives the specific behavior of an observed variable. This variable is itself modelled through a usual probability law such as a Gaussian law, for example. The system state process, which specifies the ongoing behavior of the latter observed variable at each time-step, is fully latent. Therefore, state inference is the main purpose of HMM models: the goal is to learn about the latent sequence of states from the observed behavior. This task is generally driven by Maximum A Posteriori (MAP) estimation implemented through the \textbf{Viterbi algorithm} \citep{forney_1973_journ-proceedings_viterbi-algo}. Importantly, the HMM framework satisfies the Markov property, which stipulates that the conditional probability distribution of the hidden state at time-step $t$, given the hidden states at previous time-steps $t' < t$, only depends on the hidden state at time-step $t-1$. Besides, the observed behavior at time-step $t$ solely depends on the hidden variable at time-step $t$.
When dealing with systems in which the state process is partially observed or known, applying HMM would result in an important information loss in the sense that the observed states are ignored. To overcome this limitation, \cite*{scheffer-wrobel_2001_ecml-pkdd-workshop_partially-hidden-markov-models} have introduced the Partially Hidden Markov Chain (PHMC), which integrates partially observed states in the modelling process. The authors have proposed an active learning algorithm in which the user is asked to label difficult observations identified during model learning. More recently, \cite*{ramasso-denoeux_2013_jour-ieee-transact-fuzzy-sys_partially-hmms} have proposed a model that makes use of partial knowledge on HMM states. These authors have modelled the partial knowledge by a \textit{belief function} that specifies the probability of each state at each time-step. The works carried out by \cite*{ramasso-denoeux_2013_jour-ieee-transact-fuzzy-sys_partially-hmms} have shown that the use of partial knowledge on states accelerates HMM model learning.
\subsection{Linear AutoRegressive model - LAR($p$)}
An observed time series is considered to be one realization of a stochastic process. Time series analysis and forecasting thus require that the underlying stochastic process be modelled. The linear autoregressive (LAR) model is a stochastic model widely used for this purpose. A LAR model of order $p$ is a linear model in which the regressors are the $p$ past values of the variable, hence the term autoregression. Although the LAR model is conceptually simple and easy to learn, it can only be applied to \textbf{stationary time series}. When this condition is violated, \textit{model misspecification} issues arise. Nonetheless, it is well known that if the autoregressive coefficients of a LAR process are all less than one in module, then the process will be stationary. This is a necessary and sufficient condition which is tested through \textbf{unit root tests} \citep{phillips_perron_1988_biometrika_test-unit-root-time-series-regres, dickey_fuller_1979_journ-americ-stat-assoc_-ar-times-series-unit-root, kwiatkowski_phillips_schmidt_et_al_1992_journ-econometr_unit-root-test}.
In the LAR($p$) model, the hyper-parameter $p$ denotes the number of past observations to include in the prediction at time-step $t$. Two alternative methods are generally used to fix the value of $p$. The first one relies on a well-known property of the \textit{partial autocorrelation function} of the LAR($p$) model: the autocorrelation becomes null from lag $p+1$. The second method, more general, tests a range of candidate values for $p$, then selects the value that minimizes a model selection criterion such as the Bayesian information criterion (BIC) or the Akaike's information criterion (AIC).
\section{The PHMC-LAR model}
\label{proposed.model}
In this section, we explain how we have created a new regime-switching model called PHMC-LAR, based on the PHMC and LAR models. The section first introduces some notations. Then Subsection \ref{second.level} describes our proposal to model the state process by a PHMC model. Subsection \ref{first.level} details how, within each regime, the dynamics of the observed variable is governed by a LAR model. Thus, the bivariate process follows a PHMC-LAR model.
To note, the fundamental difference between our model and the two other approaches identified in the same line \cite*{scheffer-wrobel_2001_ecml-pkdd-workshop_partially-hidden-markov-models,ramasso-denoeux_2013_jour-ieee-transact-fuzzy-sys_partially-hmms} is the autoregressive dynamics of our model (see Fig. \ref{phmc.phmclar.graphs}).
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.38\textwidth}
\includegraphics[width=\textwidth, clip]{fig1a_conditional_independence_graph_phmc.png}
\caption{}
\label{phmc.model}
\end{subfigure}
\hspace{5mm}
\begin{subfigure}[b]{0.55\textwidth}
\includegraphics[width=\textwidth, clip]{fig1b_conditional_independence_graph_phmc-lar_2_.png}
\caption{}
\label{phmclar.conditional.ind.graph}
\end{subfigure}
\caption{The conditional independence graphs of the Partially Hidden Markov Chain and of the Partially Hidden Markov Chain Linear Autoregressive (PHMC-LAR) model, when the LAR order $p$ is equal to $2$. (a) PHMC model. (b) PHMC-LAR model. Observed states are shown in dark shade whereas hidden states are colored in light shade. When a state is observed, $\sigma_t$ is reduced to a singleton.}
\label{phmc.phmclar.graphs}
\end{figure}
\subsection{Notations}
\hspace{4mm} $\bullet$ Symbol $:=$ stands for the {\it definition symbol}.
$\bullet$ $\bm{1}_A: \Omega \rightarrow \{0,1\}$ denotes the {\it indicator function} that indicates membership of an element in a subset $A$ of $\Omega$. As from now, $\bm{1}_A$ will be noted $\bm{1}_{\{ x \in A \}}$.
$\bullet$ $\{X_t\}_{t \in \mathbb{Z}}$ denotes a stochatic process. By convention, $X_{1-p}^0$ denotes the $p$ initial values of the time series $\{X_t\}$. For each $t \geq 1$, $X_{t-p}^{t-1}$ stands for the subseries $\{ X_{t-p}, X_{t-p+1} \cdots X_{t-1} \}$.
$\bullet$ $\mathbf{x} = x_{1}^T$ represents an observed time series with $\mathbf{x}_0=x_{1-p}^0$ the corresponding initial values.
$\bullet$ $\{ S_t \}_{t \in \mathbb{N}^*}$ denotes a state process depicting the temporal evolution of a regime-switching system where the set of states is $\mathbf{K} = \{1, 2, \dots, K\}$. In this paper, states are instantaneous, whereas a regime is a succession of identical states. We denote $\sigma_t$ the set of possible states at time-step $t$ with $\sigma_t = \mathbf{K}$ when $S_t$ is latent, and $\sigma_t = \{k\}$ when $S_t = k$, that is $k^{\text{th}}$ state is observed at time-step $t$.
$\bullet$ $\mathcal{M}_p(\mathbb{R})$ is the set of square matrices of order $p$ with real coefficients.
$\bullet$ Symbols in bold represent nonscalar variables ({\it e.g.}, vectors).
\subsection{Modelling the state process}
\label{second.level}
Let $\{ (S_t, \sigma_t) \}$ the state process which is supposed to be partially observed. Remind that if $S_t=k$, {\it i.e.} $k^{\text{th}}$ state has been observed at time-step $t$, then $\sigma_t = \{k\}$; otherwise $\sigma_t = \mathbf{K}$, {\it i.e.} $S_t$ is latent.
Let $\mathcal{R} = \{ k \in \mathbf{K} \,|\, \exists\, t \in \mathbb{N}^*, \sigma_t = \{k\} \}$, the set of states that have been observed at least once. We have $|\mathcal{R}| \le K$ where $K$ is the total number of states. Thus, $K - |\mathcal{R}|$ states are fully latent and depict the hidden dynamics of the system under study. It has to be underlined that it is difficult (it not sometimes impossible) to associate a physical interpretation to the hidden dynamics. Such an interpretation requires strong knowledge upon the studied system.
In the PHMC-LAR model, $\{(S_t, \sigma_t)\}$ is modelled by a $K$-state PHMC, parametrized by
transition probabilities
\noindent $a_{i,j} = P(S_t=j|S_{t-1}=i), \quad a_{i,j} \in [0, 1], \, \sum_{j=1}^K a_{i,j} = 1$
\noindent and stationary law $\pi_i = P(S_1 = i), \quad \pi_i \in [0, 1], \, \sum_{i=1}^K \pi_i = 1$.
Let $\bm{\theta}^{(S)} = ( (\pi_i)_{i=1,...,K}, (a_{i,j})_{i,j=1,...,K})$ denote the set of parameters associated with the PHMC.
\subsection{Modelling the dynamics under each state}
\label{first.level}
For each state $k \in \mathbf{K}$, $\{X_t\}$ is supposed to be \textbf{stationary} and modelled by a $p$-order LAR process defined as follows:
\begin{align}
\label{X.model}
X_t \,|\, X_{t-p}^{t-1}, S_t=k \,&:=\, \phi_{0,k} + \sum_{i=1}^{p} \phi_{i,k} X_{t-i} + h_k \, \epsilon_t \quad \text{for} \quad t=1, \dots, T,
\end{align}{}
with $p$ the number of past values of $X_t$ to be used in modelling, $k$ the state at time-step $t$, $\bm{\mu}_k = (\phi_{0,k}, \, \phi_{1,k}, ..., \phi_{p,k})$ the intercept and autoregressive parameters associated with $k^{th}$ state, $h_k$ the standard deviation associated with $k^{th}$ state and $\{\epsilon_t\}$ the error terms.
It is important to underline that Eq. \ref{X.model} is not defined for the $p$ initial values denoted by $X_{1-p}^0$. These initial values are modelled by the initial law $g_0(x_{1-p}^0; \bm{\psi})$ parametrized by $\bm{\psi}$. For instance, $g_0$ can be a multivariate normal distribution $\mathcal{N}_p(\mathbf{m}, \mathbf{V})$ where $\mathbf{m} \in \mathbb{R}^p$ is the mean and $\mathbf{V} \in \mathcal{M}_p(\mathbb{R})$ is the variance-covariance matrix.
The $\epsilon_t$ terms are independent and identically distributed with zero mean and unit variance. Note that the law of $\{\epsilon_t\}$ and the conditional distribution $P(X_t|X_{t-p}^{t-1},$ $S_t=k; \, \bm{\mu}_k, h_k)$ belong to the same family. Usually, Gaussian white noises are used. In this case, the conditional distribution is Gaussian too, with mean and variance respectively equal to $\phi_{0,k} + \sum_{i=1}^p \phi_{i,k} X_{t-i}$ and $h_k^2$.
Let $\bm{\theta}^{(X,k)} = (\bm{\mu}_{k}, h_{k})$ the parameters of the LAR($p$) process associated with $k^{th}$ state. The law of $\{X_t\}$ is fully parametrized by $\bm{\theta}^{(X)} = (\bm{\theta}^{(X,k)})_{k=1, ...,K}$ and $\bm{\psi}$.
To note, as in \citep{scheffer-wrobel_2001_ecml-pkdd-workshop_partially-hidden-markov-models} and \citep{ramasso-denoeux_2013_jour-ieee-transact-fuzzy-sys_partially-hmms}, the PHMC-LAR model assumes that the same order $p$ is shared by all $| \mathbf{K} |$ LAR processes associated with the states in $\mathbf{K}$.
It has also to be highlighted that the state $S_t = k$ conditioning a LAR process of order $p$ on $X_t$ does not impose that the $p$ lagged values $X_{t-1}^{t-p}$ be observed under same state $k$. That is, the PHMC-LAR model may perfectly switch from regime to regime, and even from state to state, meanwhile keeping memory of values determined by previous regimes or states.
\section{Learning algorithm}
\label{learning.algorithm}
This section is dedicated to the presentation of an instance of the Expectation-Maximization (EM) algorithm, to estimate the PHMC-LAR parameters.
As seen in previous subsections, the PHMC component and the LAR components of our model are respectively parametrized by $\bm{\theta}^{(S)}$ and $(\bm{\theta}^{(X)}, \bm{\psi})$. Then, the whole PHMC-LAR model is parametrized by $(\bm{\theta}, \bm{\psi})$ where $\bm{\theta} = (\bm{\theta}^{(S)}, \bm{\theta}^{(X)})$. Thus, PHMC-LAR learning consists in estimating $(\bm{\theta}, \bm{\psi})$ from a training dataset.
Thanks to good statistical properties such as asymptotic efficiency, a maximum likelihood estimator (MLE) is considered. However, for models with hidden variables like ours, MLE computation results in an untractable problem. To address this issue, the Expectation-Maximization (EM) algorithm is generally used, in order to approximate a set of parameters that locally maximizes the likelihood function. EM was introduced by \cite*{baum-petrie-soules_1970_jour-annals-math_em-for-hmc-models} to cope with Hidden Markov Model learning. This version was further extended by \cite{dempster-laird-rubin_1977_jour-royal-stat-society_EM-for-incomplete-data} into the versatile EM algorithm, to handle parameter estimation in a more general framework. EM has also been applied to autoregressive Markov-switching models \citep{hamilton_1990_journ-econometrics_time_series-regime-changes} and PHMC models \citep{scheffer-wrobel_2001_ecml-pkdd-workshop_partially-hidden-markov-models, ramasso-denoeux_2013_jour-ieee-transact-fuzzy-sys_partially-hmms}.
We propose to learn the PHMC-LAR model through a dedicated instance of the EM algorithm. To fix ideas, in subsection \ref{particular.case}, we first consider the case where the model is trained in a univariate context, that is considering a unique couple of data ($x_{t=1-p}^T, \, \sigma_{t=1}^T$), with $x_{t=1-p}^T$ a realization of $\{ X_t \}$ and $\sigma_t$ the set of possible states at time-step $t$. Then, the general multivariate case of $N$ independent couples of data $(\mathbf{x}^{(1)},\Sigma^{(1)})$, $\dots$, $(\mathbf{x}^{(N)},\Sigma^{(N)})$ will be presented in subsection \ref{general.case}.
\subsection{Particular case: univariate scheme}
\label{particular.case}
Let $\mathbf{x} = x_{1-p}^T$ the observed time series with $x_{1-p}^0$ the initial values of the autoregressive process. Let $\Sigma = \sigma_{t=1}^T$, further simplified into $\sigma_{1}^T$, where $\sigma_t$ stands for the set of possible states at time-step $t$. Let $(S_1^T, \Sigma)$ the state process (partially observed) of $\mathbf{x}$ with $\sigma_t = \mathbf{K}$ if $S_t$ is hidden, and $\sigma_t = \{k\}$ if state $k$ is observed at time-step $t$.
MLE is implemented by maximizing the expectation (with respect to the latent variables) of the \textbf{complete data likelihood}.
Complete data likelihood is further referred to as $\mathcal{L}^c$. $\mathcal{L}^c$ denotes the evidence/likelihood of the training data when latent/hidden variables are supposed to be known. $\mathcal{L}^c$ writes as follows:
\begin{align}
\mathcal{L}^c (\bm{\theta}, \bm{\psi}) &= P(X_{1-p}^T=x_{1-p}^T, S_1^T=s_1^T; \, \bm{\theta}, \bm{\psi} ) \nonumber \\
&= P(X_1^T=x_1^T, S_1^T=s_1^T \,|\, X_{1-p}^0=x_{1-p}^0; \, \bm{\theta}) \times P(X_{1-p}^0=x_{1-p}^0; \, \bm{\psi}) \nonumber \\
\label{complete.data.like}
&= \mathcal{L}^c_c (\bm{\theta}) \times g_0(x_{1-p}^0; \, \bm{\psi}),
\end{align}
with $\mathcal{L}_c^c$ the \textbf{conditional complete data likelihood} and $g_0$ the initial law of $X_t$.
When the expectation of $\mathcal{L}^c$ with respect to the partially hidden states is calculated, term $g_0(x_{1-p}^0; \, \bm{\psi})$ in Eq. \ref{complete.data.like} can be taken out of the expectation since it does not depend on the states:
{\small
\begin{equation}
\label{expect.complete.data.like}
\begin{split}
\mathbb{E}_{P(S_1^T \,|\, X_{1-p}^{T}=x_{1-p}^{T}, \Sigma; \, \bm{\theta})} [\mathcal{L}^c(\bm{\theta}, \bm{\psi})] &= \mathbb{E}_{P(S_1^T \,|\, X_{1-p}^{T}=x_{1-p}^{T}, \Sigma; \, \bm{\theta})} [\mathcal{L}^c_c (\bm{\theta})] \times g_0(x_{1-p}^0; \, \bm{\psi}),
\end{split}
\end{equation}
}
where $P(S_1^T \,| \,X_{1-p}^{T}=x_{1-p}^{T}, \Sigma; \, \bm{\bm{\theta}})$ is the \textit{posterior probability} of partially hidden states $(S_1^T, \Sigma)$.\\
Then, by considering the logarithmic scale, Eq. \ref{expect.complete.data.like} can be separately maximized with respect to $\bm{\theta}$ and $\bm{\psi}$:
\begin{align}
\label{psi.max}
\hat{\bm{\psi}} &= \underset{\bm{\psi}}{\arg\max} \, \ln \left( g_0(x_{1-p}^0; \, \bm{\psi}) \right), \\
\label{theta.max}
\hat{\bm{\theta}} &= \underset{\bm{\theta}}{\arg\max} \, \ln \left( \mathbb{E}_{P(S_1^T \,|\, X_{1-p}^{T}=x_{1-p}^{T}, \Sigma; \, \bm{\theta})} [\mathcal{L}_c^c(\bm{\theta})] \right).
\end{align}
It has to be noted that Eq. \ref{psi.max} is a simple probability observation problem. In contrast, because of the hidden states, maximization with respect to $\bm{\theta}$ (Eq. \ref{theta.max}) is carried out by an instance of the EM algorithm.
\indent EM is an iterative algorithm that alternates between E(xpectation) step and M(aximization) step. At iteration $n$, we obtain:
\begin{align}
\label{e.step}
\text{E-step} & \quad Q(\bm{\theta}, \hat{\bm{\theta}}_{n-1}) = \mathbb{E}_{P(S_1^T \,|\, X_{1-p}^{T}=x_{1-p}^{T}, \Sigma; \, \hat{\bm{\theta}}_{n-1})} [\ln \mathcal{L}_c^c (\bm{\theta})], \\
\label{m.step}
\text{M-step} & \quad \hat{\bm{\theta}}_n = \underset{\bm{\theta}}{\arg\max} \,\ Q(\bm{\theta}, \hat{\bm{\theta}}_{n-1}),
\end{align}
with $P(S_1^T \,| \,X_{1-p}^{T}=x_{1-p}^{T}, \Sigma; \, \hat{\bm{\bm{\theta}}}_{n-1})$ the \textit{posterior probability} of partially hidden states $(S_1^T, \Sigma)$ at iteration $n-1$.
The rest of this Subsection details the two EM steps.
\subsubsection{Step E of EM}
In this step, the quantity $Q(\bm{\theta}, \hat{\bm{\theta}}_{n-1})$ (Eq. \ref{e.step}) is computed. Following the conditional independence graph of the PHMC-LAR model (see Fig. \ref{phmclar.conditional.ind.graph}), the conditional complete data likelihood writes:
\begin{equation}
\label{cond.comp.data.like}
\begin{split}
\mathcal{L}_c^c (\bm{\bm{\theta}})
&= P(X_1^T=x_1^T, S_1^T=s_1^T \,|\, X_{1-p}^0; \, \bm{\theta}) \\
&= P(S_1=s_1; \, \bm{\theta}^{(S)}) \prod_{t=2}^T P(S_t=s_t|S_{t-1}=s_{t-1}; \, \bm{\theta}^{(S)})\\
& \quad \prod_{t=1}^T P(X_t=x_t|X_{t-p}^{t-1}=x_{t-p}^{t-1}, S_t =s_t; \, \bm{\theta}^{(X,s_t)}),
\end{split}
\end{equation}
with $\bm{\theta}^{(X,k)}$ the parameters of the LAR process associated with $k^{th}$ state and $P(X_t=x_t \, | \, X_{t-p}^{t-1}, S_t =k; \, \bm{\theta}^{(X,k)})$ the conditional law of $X_t$ within $k$. \\
Notice that the terms in Eq. \ref{cond.comp.data.like} depend on either a single state $S_t$ or two consecutive states $S_t, S_{t-1}$. In this same equation, products are replaced by sums when considering the logarithm scale. Then $\ln \mathcal{L}_c^c (\bm{\bm{\theta}})$ is substituted in Eq. \ref{e.step} and the expectation with respect to the posterior probability of state process is developed. After some integrations, we find that $Q(\bm{\theta}, \hat{\bm{\theta}}_{n-1})$ only depends on the following probabilities:
\begin{equation}
\label{xi.definition}
\begin{split}
&\xi_t(k, \ell) = P(S_{t-1}=k, S_{t}=\ell \, | \, X_{1-p}^T = x_{1-p}^T, \Sigma; \, \hat{\bm{\theta}}_{n-1}),\\
&\text{for} \quad t=2, \dots, T, \quad 1 \le k, \,\ell \le K.\\
\end{split}
\end{equation}
\begin{equation}
\label{gamma.defintion}
\begin{split}
&\gamma_t(\ell) = P(S_{t}=\ell \, | \, X_{1-p}^T = x_{1-p}^T, \Sigma; \, \hat{\bm{\theta}}_{n-1}),\\
&\text{for} \quad t=2, \dots, T, \quad 1 \le \ell \le K.\\
\end{split}
\end{equation}
Therefore, the E-step is reduced to computing these probabilities. To this end, we have derived a \textit{backward-forward-backward} procedure as an extension of the forward-backward algorithm, one of the ingredients of the Baum-Welsh algorithm \citep{dempster-laird-rubin_1977_jour-royal-stat-society_EM-for-incomplete-data}. The backward-forward-backward algorithm was initially proposed by \cite*{scheffer-wrobel_2001_ecml-pkdd-workshop_partially-hidden-markov-models} for the purpose of PHMC model learning. We have adapted this algorithm to PHMC-LAR models by taking into consideration the autoregressive dynamics. The details about the adapted \textit{backward-forward-backward} algorithm are given in Appendix \ref{BFB}.
\subsubsection{Step M of EM}
At iteration $n$, this step consists in maximizing $Q(\bm{\theta}, \hat{\bm{\theta}}_{n-1})$ with respect to parameters $\bm{\theta} = (\bm{\theta}^{(S)}, \bm{\theta}^{(X)})$. It is straightforward to show that $Q(\bm{\theta}, \hat{\bm{\theta}}_{n-1})$ can be decomposed as follows:
$$Q(\bm{\theta}, \hat{\bm{\theta}}_{n-1}) = Q_S(\bm{\theta}^{(S)}, \hat{\bm{\theta}}_{n-1}) + Q_X(\bm{\theta}^{(X)}, \hat{\bm{\theta}}_{n-1}),$$
where $Q_S$ (respectively $Q_X$) only depends on parameters $\bm{\theta}_S$ (respectively $\bm{\theta}_X$). Therefore, $Q_S$ and $Q_X$ can be maximized apart:
\begin{equation}
\label{theta.S}
\hat{\bm{\theta}}_n^{(S)} = \underset{\bm{\theta}^{(S)}}{\arg\max} \,\ Q_S(\bm{\theta}^{(S)}, \hat{\bm{\theta}}_{n-1}).
\end{equation}
\begin{equation}
\label{theta.X}
\hat{\bm{\theta}}_n^{(X)} = \underset{\bm{\theta}^{(X)}}{\arg\max} \,\ Q_X(\bm{\theta}^{(X)}, \hat{\bm{\theta}}_{n-1}).
\end{equation}
The analytical expressions of $Q_S$ and $Q_X$ are given in Appendix \ref{decomposition.of.Q}. Cancelling the first derivative of $Q_S(\bm{\theta}^{(S)}, \hat{\bm{\theta}}_{n-1})$ provides the analytical expression of $\hat{\bm{\theta}}_n^{(S)}$. In contrast, it is not possible to derive the analytical expression for $\hat{\bm{\theta}}_n^{(X)}$. That is why $Q_X(\bm{\theta}^{(X)}, \hat{\bm{\theta}}_{n-1})$ has to be maximized relying on a numerical optimization method ({\it e.g.}, the quasi-Newton method).
\subsection{General case: multivariate scheme}
\label{general.case}
We now consider the general case in which PHMC-LAR model is learnt from $N$ independent couples of data $(\mathbf{x}^{(1)},\Sigma^{(1)})$, $\dots$, $(\mathbf{x}^{(N)},\Sigma^{(N)})$, with $\mathbf{x}_0^{(1)}, \dots, \mathbf{x}_0^{(N)}$ the associated initial values and $(\mathbf{S}^{(1)}, \Sigma^{(1)}), \dots, (\mathbf{S}^{(N)}, \Sigma^{(N)})$ the corresponding state processes. It has to be noted that time series $\mathbf{x}^{(i)}$'s can have different lengths while their respective initial vectors have a common size ($\mathbf{x}_0^{(i)} \in \mathbb{R}^p$, with $p$ the autoregressive order).
In this case, the MLE estimates of parameters ($\bm{\theta}, \bm{\psi}$) are defined as follows:
\begin{align}
\label{g.0.param.learning}
\hat{\bm{\psi}} &= \underset{\bm{\psi}}{\arg\max} \, \sum_{i=1}^N \ln \left( g_0(\mathbf{x}_0^{(i)}; \, \bm{\psi}) \right), \\
\label{theta.learning}
\hat{\bm{\theta}} &= \underset{\bm{\theta}}{\arg\max} \, \ln \left( \mathbb{E}_{P_S} [\mathcal{L}_c^c(\bm{\theta})] \right),
\end{align}
with $P_S = P(\mathbf{S}^{(1)}, \dots, \mathbf{S}^{(N)} \,|\, \mathbf{X}^{(1)}=\mathbf{x}^{(1)}, \dots, \mathbf{X}^{(N)}=\mathbf{x}^{(N)}, \mathbf{X}^{(1)}_0=\mathbf{x}^{(1)}_0, \dots,$ $\mathbf{X}^{(N)}_0=\mathbf{x}^{(N)}_0, \Sigma^{(1)}, \dots, \Sigma^{(N)}; \, \bm{\theta})$ the \textit{posterior distribution} of partially hidden states $(\mathbf{S}^{(1)}, \Sigma^{(1)}), \dots, (\mathbf{S}^{(N)}, \Sigma^{(N)})$.
Thus, the conditional complete data likelihood $\mathcal{L}_c^c (\bm{\bm{\theta}}) $ defined in Eq. \ref{cond.comp.data.like} becomes:
{\small
\begin{equation}
\begin{split}
\label{log.likelihood.several.seq}
\mathcal{L}_c^c (\bm{\bm{\theta}})
&= \prod_{i=1}^N P \left(\mathbf{X}^{(i)}=\mathbf{x}^{(i)}, \mathbf{S}^{(i)}=\mathbf{s}^{(i)} \,|\, \mathbf{X}_0^{(i)}; \, \bm{\theta} \right) \\
&= \prod_{i=1}^N \left[ P \left(S^{(i)}_1=s^{(i)}_1; \, \bm{\theta}^{(S)} \right) \prod_{t=2}^{T_i} P \left(S_t^{(i)}=s_t^{(i)} \,|\, S_{t-1}^{(i)}=s_{t-1}^{(i)}; \, \bm{\theta}^{(S)} \right) \, \right.\\
& \qquad \quad \times \left. \prod_{t=1}^{T_i} P \left(X_t^{(i)}=x_t^{(i)} \,|\, [X^{(i)}]_{t-p}^{t-1}=[x^{(i)}]_{t-p}^{t-1}, \, S_t^{(i)} =s_t^{(i)}; \, \bm{\theta}^{(X,s_t^{(i)})} \right) \right].
\end{split}
\end{equation}
}
As in Eq. \ref{psi.max}, Eq. \ref{g.0.param.learning} is a simple probability observation problem. When $g_0$ is a multivariate normal distribution $\mathcal{N}_p(\mathbf{m})$, with mean $\mathbf{m} \in \mathbb{R}^p$, variance-covariance matrix $\mathbf{V} \in \mathcal{M}_p(\mathbb{R})$ and $\bm{\psi} = (\mathbf{m}, \mathbf{V})$, we can show that
\begin{equation}
\hat{\mathbf{m}} = \frac{1}{N} \sum_{i=1}^N \mathbf{x}_0^{(i)}, \quad
\hat{\mathbf{V}} = \frac{1}{N} \sum_{i=1}^N \, (\mathbf{x}_0^{(i)} - \hat{\mathbf{m}}) \, (\mathbf{x}_0^{(i)} - \hat{\mathbf{m}})^{'},
\end{equation}
where ${'}$ stands for matrix transposition.
Equation \ref{theta.learning} is maximized using the instance of EM presented in Subsection \ref{particular.case}. At each iteration $n$, the E-step consists in computing probabilities $\xi_t^{(i)}$ (Eq. \ref{xi.definition}) by running the \textit{backward-forward-backward} algorithm presented in Appendix \ref{BFB}, on data $(\mathbf{x}^{(i)}, \Sigma^{(i)})$. Thus, the expectation $Q(\bm{\theta}, \hat{\bm{\theta}}_{n-1})$ (Eq. \ref{e.step}) can be computed.
Then, in the M-step, $Q(\bm{\theta}, \hat{\bm{\theta}}_{n-1})$ is maximized with respect to $\bm{\theta} = (\bm{\theta}^{(S)}, \bm{\theta}^{(X)})$ following Eq. \ref{theta.S}-\ref{theta.X}. We obtain the following formula for $\hat{\bm{\theta}}^{(S)}_n$:
\begin{equation}
\label{phmc.mle.1}
\hat{a}_{k,l}^{(n)} = \frac{ \sum_{i=1}^N \sum_{t=2}^{T_i} \xi_t^{(i)}(k,l) } { \sum_{i=1}^N \sum_{t=1}^{T_i} \gamma_t^{(i)} (k)}, \quad \text{for} \quad 1 \le k,l \le K.
\end{equation}
\begin{equation}
\label{phmc.mle.2}
\hat{\pi}_l^{(n)} = \frac{ \sum_{i=1}^N \gamma_1^{(i)}(l) }{N}, \quad \text{for} \quad 1 \le l \le K,
\end{equation}
\begin{equation}
\label{gamma_general}
\begin{split}
&\text{with} \quad \gamma_t^{(i)}(s) = P(S_t^{(i)}=s \,|\, [X^{(i)}]_{1-p}^{T_i} = x_{1-p}^{T_i}, \Sigma^{(i)}; \, \hat{\bm{\theta}}_{n-1}) = \sum_{j=1}^K \xi_t^{(i)}(j,s),\\
&\text{for} \quad t=2, \dots, T_i,
\end{split}
\end{equation}
\begin{equation}
\label{gamma_one}
\begin{split}
\text{and} \quad \gamma_1^{(i)}(s) &= P(S_1^{(i)}=s \,|\, [X^{(i)}]_{1-p}^{T_i} = x_{1-p}^{T_i}, \Sigma^{(i)}; \, \hat{\bm{\theta}}_{n-1}) = \sum_{j=1}^K \xi_2^{(i)}(s,j).
\end{split}
\end{equation}
Note that $\hat{\bm{\theta}}^{(X)}_n$ is computed through numerical optimization, for instance by using the quasi-Newton method.
Algorithm \ref{em.algorithm} sums up the instance of EM proposed for PHMC-LAR parameter learning.
\vspace{0.3cm}
\begin{algorithm}[H]
\begin{algorithmic}[1]
\STATE \textbf{Input}: \textit{data} $(\mathbf{x}^{(1)}, \Sigma^{(1)}), \dots, (\mathbf{x}^{(N)}, \Sigma^{(N)})$, precision $\kappa$, maximal number of iterations $max_{iter}$
\STATE Initialization: $\hat{\bm{\theta}}^{(0)}$ randomly chosen
\STATE $n \leftarrow 1$
\REPEAT
\STATE E-step
\STATE \hspace{0.4cm} For each couple $(\mathbf{x}^{(i)}, \Sigma^{(i)})$, $i= 1, \dots, N$
\STATE \hspace{0.4cm} Compute $\xi_t^{(i)}$ by running the \textit{backward-forward-backward} algorithm on $(\mathbf{x}^{(i)}, \Sigma^{(i)})$
\STATE M-step
\STATE \hspace{0.4cm} M-S : compute $\hat{\bm{\theta}}_n^{(S)}$ from Eq. \ref{phmc.mle.1}-\ref{phmc.mle.2}
\STATE \hspace{0.4cm} M-X : compute $\hat{\bm{\theta}}_n^{(X)}$ by numerical optimization of $Q_X(\bm{\theta}^{(X)}, \hat{\bm{\theta}}_{n-1})$
\STATE $incr(n)$
\UNTIL{($|\hat{\bm{\theta}}^{(n)} - \hat{\bm{\theta}}^{(n-1)}| < \kappa$) or ($n> max_{iter}$)}
\STATE /* parameters stay roughly stable between two successive iterations, */
\STATE /* or the maximum number of iterations is reached \verb^ ^*/
\end{algorithmic}
\caption{EM algorithm for PHMC-LAR model training}
\label{em.algorithm}
\end{algorithm}
\vspace{0.3cm}
It is well known that the EM algorithm is sensitive to the choice of the starting point $\hat{\bm{\theta}}^{(0)}$ as regards the risk of attraction in a local maximum. In practice, several initial values are tested and the model that provides the highest likelihood is chosen. In this work, the initialization procedure presented in Algorithm \ref{em.init} is used.
\vspace{0.3cm}
\begin{algorithm}[H]
\begin{algorithmic}[1]
\STATE \textbf{Input}: $L$, precision $\kappa$, maximum number of iterations $max_{iter}$
\STATE Let $\hat{\bm{\theta}}^{(0,1)}, \dots, \hat{\bm{\theta}}^{(0, L)}$ initial values randomly chosen.
\STATE For each $\hat{\bm{\theta}}^{(0,j)}$, EM is run with parameters $\kappa$ and $max_{iter}$.
\STATE Then, $\hat{\bm{\theta}}^{(0)}$ is fixed as the estimated parameters that provide the highest likelihood across the $L$ restarts.
\end{algorithmic}
\caption{EM initialization for PHMC-LAR model training}
\label{em.init}
\end{algorithm}
\section{Hidden state inference}
\label{inference}
In HMM modelling, after a model is learnt, inference consists in finding the state sequence that maximizes the likelihood of a given observed sequence. This is equivalent to solve a \textit{Maximum A Posteriori} (MAP) problem. The Greedy search method that enumerates all combinations of states requires $\mathcal{O}(K^T)$ operations, where $K$ is the number of states and $T$ is the sequence length. The \textbf{Viterbi algorithm} designed by \cite*{forney_1973_journ-proceedings_viterbi-algo} computes the optimal state sequence in $\mathcal{O}(TK^2)$ operations.
In this section, we propose a variant of the Viterbi algorithm that takes into account the observed states of the PHMC-LAR model. Thus, the hidden states are inferred given the observed states and the given observation sequence.
Let $\hat{\bm{\theta}}$ the MLE parameter estimates of the PHMC-LAR model trained on a given dataset. Let $\mathbf{x} = x_{1}^T$ an observed time series and $\mathbf{x}_0=x_{1-p}^0$ the corresponding initial values. Let $\Sigma = \sigma_{t=1}^T$ the possible states at each time-step with $\sigma_t = \{k\}$ if $k^{th}$ regime is observed at time-step $t$, and $\sigma_t = \mathbf{K}$ if the state process is latent at that time-step. Let $(\mathbf{S}, \Sigma)$ the partially hidden state process associated with this time series.
We search the optimal state sequence $\mathbf{z}^* = (z_1^*, \dots, z_T^*)$ that maximizes the posterior probability $P(\mathbf{S}=\mathbf{z} \,|\, \mathbf{X}=\mathbf{x}, \mathbf{X}_0=\mathbf{x}_0, \Sigma; \, \hat{\bm{\theta}})$. Thanks to Bayes' rule, maximizing this posterior probability is equivalent to maximizing the joint probability $P(\mathbf{S}=\mathbf{z}, \mathbf{X}=\mathbf{x} \,|\, \mathbf{X}_0=\mathbf{x}_0, \Sigma; \, \hat{\bm{\theta}})$:
\begin{equation}
\label{joint.prob}
P(\mathbf{S}=\mathbf{z} \,|\, \mathbf{X}=\mathbf{x}, \, \mathbf{X}_0=\mathbf{x}_0 ; \, \hat{\bm{\theta}}^{(X)}) = \frac{P(\mathbf{S}=\mathbf{z}, \mathbf{X}=\mathbf{x} \,|\, \mathbf{X}_0=\mathbf{x}_0 \, \Sigma; \, \hat{\bm{\theta}})}{P(\mathbf{X}=\mathbf{x} \,|\,\mathbf{X}_0=\mathbf{x}_0, \Sigma; \, \hat{\bm{\theta}}^{(S)})}.
\end{equation}
\begin{equation}
\label{z.opt}
\mathbf{z}^* = \underset{\mathbf{z} \, \in \, \mathbf{K}^T}{\arg\max} \, P(\mathbf{S}=\mathbf{z}, \mathbf{X}=\mathbf{x} \,|\, \mathbf{X}_0=\mathbf{x}_0, \Sigma; \, \hat{\bm{\theta}}),
\end{equation}
where $\mathbf{K} = \{1, 2, \dots, K\}$ is the set of possible states.
Note that the probability of a given state sequence is null if there is at least a time-step $t$ such that $z_t \notin \sigma_t$, that is if state $z_t$ is not allowed at time-step $t$. A consequence is that $\mathbf{z}^*$ must coincide with the observed states if there are any.
Following the dynamic programming paradigm, the Viterbi algorithm makes it possible to retrieve $\mathbf{z}^*$ by splitting the initial problem into subproblems and solving this set of smaller problems. Let $\delta_t(\ell; \, \hat{\bm{\theta}})$ the maximal probability of subsequence $(z_1, \dots, z_t = \ell)$ that ends within regime $\ell$:
\begin{equation}
\label{viterbi.delta}
\begin{split}
&\delta_t(\ell; \, \hat{\bm{\theta}}) = \underset{z_1, \dots, z_{t-1} \, \in \mathbf{K}^{t-1}}{\max} \, P(X_1^t=x_1^t, \, S_1^{t-1}=z_1^{t-1}, \, S_t=\ell \, | \, \mathbf{X}_0= \mathbf{x}_0, \, \sigma_1^t; \, \hat{\bm{\theta}}),\\
&\text{for} \quad t = 1, 2, \dots T.
\end{split}
\end{equation}
\indent These probabilities are iteratively computed as follows:
At first time-step,
\begin{equation}
\label{delta.1}
\delta_1(\ell; \, \hat{\bm{\theta}}) = P(X_1=x_1 \,|\, \mathbf{X}_0= \mathbf{x}_0, \, S_1=\ell; \, \bm{\theta}^{(X,\ell)}) \times P(S_1=\ell \,|\, \sigma_1; \, \hat{\bm{\theta}}^{(S)})
\end{equation}
where
\begin{equation*}
\label{delta.1}
\begin{split}
P(S_1=\ell \,|\, \sigma_1; \, \hat{\bm{\theta}}^{(S)}) &= \left\{
\begin{aligned}
\hat{\pi}_l \quad &\text{ if } \ell \in \sigma_1 \\
0 \quad &\text{ otherwise.}
\end{aligned}
\right.
\end{split}{}
\end{equation*}
For $t=2, \dots, T$ we have
\begin{equation}
\label{delta.t}
\begin{split}
\delta_t(\ell; \, \hat{\bm{\theta}}) &= \underset{k}{\max} \, \left[\delta_{t-1}(k; \, \hat{\bm{\theta}}) \, P(S_t=\ell \,|\, S_{t-1}=k, \, \sigma_t; \, \hat{\bm{\theta}}^{(S)}) \right]\\
&\quad \times \, P(X_t=x_t \,|\, X_1^{t-1}=x_1^{t-1}, \mathbf{X}_0 = \mathbf{x}_0, \, S_t=\ell; \, \bm{\theta}^{(X,\ell)}),
\end{split}
\end{equation}
with
\begin{equation*}
\label{}
\begin{split}
P(S_t=\ell \,|\, S_{t-1}=k, \sigma_t; \hat{\bm{\theta}}^{(S)}) &= \left\{
\begin{aligned}
\hat{a}_{k,\ell} \quad &\text{ if } \ell \in \sigma_t \text{ and } k \in \sigma_{t-1} \\
0 \quad &\text{ otherwise.}
\end{aligned}
\right.
\end{split}
\end{equation*}{}
Since the maximal probability of the complete state sequence, that is the maximum for the probability expressed in Eq. \ref{joint.prob}, also writes:
\begin{align}
P^* = \underset{\ell}{\max} \, \delta_T(\ell; \, \hat{\bm{\theta}}),
\end{align}
the optimal sequence $\mathbf{z}^*$, defined in Eq. \ref{z.opt} is retrieved by backtracking as follows:
\begin{align}
z_t^* &= \underset{\ell}{\arg\max}
\left\{
\begin{aligned}
\delta_T(\ell; \, \hat{\bm{\theta}}) \quad \quad \quad &\text{ for } t = T \\
\delta_t(\ell; \, \hat{\bm{\theta}}) \times \hat{a}_{\ell,z_{t+1}^*} \quad &\text{ for } t = T-1, \dots, 1.
\end{aligned}
\right.
\end{align}
\section{Forecasting}
\label{forecasting}
Forecasting for a time series consists in predicting future values based on past values. Let us consider a PHMC-LAR model trained on a sequence observed up to time-step $T$, and $\bm {\hat{\theta}}$ the corresponding parameters. Let $\sigma_{T+1}, \dots, \sigma_{T+h}$ the set of possible states from time-step $T+1$ to time-step $T+h$.
\indent The optimal prediction of $X_{T+h}$ (with respect to mean squared error) is the conditional mean $\mathbb{E}[X_{T+h} \, | \, X_{1-p}^{T}=x_{1-p}^{T}, \, \sigma_{T+1}^{T+h}; \, \bm {\hat{\theta}}]$, which writes as follows:
\begin{equation}
\label{forecasting.function}
\begin{split}
\hat{X}_{T+h}
&= \sum_{k=1}^K P(S_{T+h} = k \, | \, X_{1-p}^{T}=x_{1-p}^{T}, \, \sigma_{T+1}^{T+h}; \, \bm {\hat{\theta}})\\
& \qquad \quad \mathbb{E}[X_{T+h} \,|\, X_{T+h-p}^{T+h-1}=x_{T+h-p}^{T+h-1}, S_{T+h}=k; \, \bm {\hat{\theta}}]\\
&= \sum_{k=1}^K P(S_{T+h} = k \, | \, X_{1-p}^{T}=x_{1-p}^{T}, \, \sigma_{T+1}^{T+h}; \, \bm {\hat{\theta}}) \, \left(\mathbf{y}_{T+h} \, \hat{\bm{\mu}}_k^{'} \right),
\end{split}
\end{equation}
with $\mathbf{y}_{T+h} = (1, x_{T+h-1}, \dots, x_{T+h-p})$, $\hat{\bm{\mu}}_k = (\phi_{0,k}, \phi_{1,k}, \dots, \phi_{p,k})$ the intercept and autoregressive parameters associated with $k^{th}$ state, and $'$ denoting matrix transposition.
Equation \ref{forecasting.function} depends on \textbf{smoothed probabilities} $\bar{\gamma}(i, s) = P(S_{T+i}=s \,|\, X_{1-p}^{T}=x_{1-p}^{T}, \, \sigma_{T+1}^{T+i}; \, \bm{\hat{\theta}})$, which are recursively computed as follows:
\begin{equation}
\label{prediction.probs}
\left\{
\begin{aligned}
\bar{\gamma}(0, s) &= P(S_{T}=s \,|\, X_{1-p}^{T}=x_{1-p}^{T}; \, \bm{\hat{\theta}}) = \gamma_T(s), \\
\bar{\gamma}(i, s) &= \sum_{\ell=1}^K \, \hat{a}_{\ell,s} \, \bar{\gamma}(i-1, \, \ell) \quad \text{ if } \sigma_{T+i} = \mathbf{K}, \\
\bar{\gamma}(i, s) &= 1 \qquad \qquad \qquad \qquad \text{ if } \sigma_{T+i} = \{\ell\} \text{ and } s = \ell, \\
\bar{\gamma}(i, s) &= 0 \qquad \qquad \qquad \qquad \text{ if } \sigma_{T+i} = \{ \ell\} \text{ and } s \neq \ell,
\end{aligned}
\right.
\end{equation}
for $i=1, \dots, h$, $s \in \mathbf{K}$ and $\gamma_{T}(l)$ defined in Eq. \ref{gamma_general}.
From Eq. \ref{forecasting.function} and \ref{prediction.probs}, we can notice that if state $s$ is observed at time-step $T+h$ ({\it i.e.} $\sigma_{T+h} = \{s\}$), then prediction $\hat{X}_{T+h}$ equals the conditional mean of the LAR process associated with this state (since $\bar{\gamma}(h, k) = 0$ for $k \notin \sigma_{T+h}$). In contrast, if state process is latent at time-step $T+h$ ({\it i.e.,} $\sigma_{T+h} = \mathbf{K}$), $\hat{X}_{T+h}$ is computed as the weighted sum of the conditional means of all states, with probabilities $\bar{\gamma}(h, k)$ as weights.
Note that for $h=1$, the past values of the time series required in Eq. \ref{forecasting.function} are known. In contrast, for $h > 1$, the intermediate predictions $\hat{X}_{T+1}, \dots, \hat{X}_{T+h-1}$ are used in order to feed the autoregressive dynamics of the PHMC-LAR framework.
\section{Experiments}
\label{experiments}
The aim of this section is two-fold: (i) assess the ability of PHMC-LAR model to infer the hidden states, (ii) evaluate prediction accuracy. These evaluations were achieved on simulated data, following two experimental settings. On the one hand, we varied the percentage of observed states in training set, to evaluate its influence on hidden state recovery and prediction accuracy. On the other hand, we simulated unreliable observed states in training set, and evaluated the influence of uncertain labelling on hidden state inference and prediction accuracy.
This section starts with the description of the protocol used to simulate data in both experimental settings. Then, the section focuses on implementation aspects. We next present and discuss the results obtained in both experimental settings.
\subsection{Simulated datasets}
This subsection first focuses on the model used to generate data. Then we describe the precursor sets used to further generate the test-set and the training datasets.
\subsubsection{Generative model}
\label{generative_model}
These experiments were achieved on simulated data from a 4-state PHMC-LAR($2$) model whose transition matrix and initial probabilities are:
\begin{equation}
\label{4.regime.mc}
A = \begin{pmatrix}
0.5 & \, 0.2 & \, 0.1 & \, 0.2 \\
0.2 & \, 0.5 & \, 0.2 & \, 0.1 \\
0.1 & \, 0.2 & \, 0.5 & \, 0.2 \\
0.2 & \, 0.1 & \, 0.2 & \, 0.5
\end{pmatrix}, \quad
\pi = (0.25, \, 0.25, \, 0.25, \, 0.25).
\end{equation}
Within each state $k \in \{1, 2, 3, 4\}$, the autoregressive dynamics is a LAR($2$) process defined by parameters $\bm{\theta}^{(X,k)} = (\phi_{0,k}, \, \phi_{1,k}, \,\phi_{2,k}, \, h_k)$:
\begin{equation}
\label{lar.proc}
\begin{split}
\bm{\theta}^{(X,1)} &= (2, \, 0.5, \, 0.75, 0.2), \quad \bm{\theta}^{(X,2)} = (-2, -0.5, \, 0.75, \, 0.5), \\
\bm{\theta}^{(X,3)} &= (4, \, 0.5, -0.75, \, 0.7), \quad \bm{\theta}^{(X,4)} = (-4, -0.5, -0.75, \, 0.9).
\end{split}
\end{equation}
In the LAR(2) process associated with state $k$, stationarity is guaranteed by setting the following contraints: $\phi_{i,k} < 1,\ i \in \{1,2\}$.
Finally, the initial law $g_0$ is a bivariate Gaussian distribution\\
\begin{equation}
\label{init.law}
g_0 = \mathcal{N}_2 \, \left((3, 5), \, \begin{pmatrix}
1 & \, 0.1 \\ 0.1 & \, 1
\end{pmatrix} \right).
\end{equation}
Figure \ref{example.of.simulated.phmclar} shows an example of state process (Fig. \ref{phmclar.state.process}) and corresponding time series (Fig. \ref{phmclar.time.series}) that were simulated from the previously defined PHMC-LAR($2$).
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig2a_phmc-lar_state_process.png}
\caption{}
\label{phmclar.state.process}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig2b_phmc-lar_simulated_timeseries.png}
\caption{}
\label{phmclar.time.series}
\end{subfigure}
\caption{A simulation from the PHMC-LAR($2$) model defined by Eq. \ref{4.regime.mc}-\ref{init.law}: (a) state process, (b) the corresponding time series}
\label{example.of.simulated.phmclar}
\end{figure}
\subsubsection{Precursor sets for the test-set and training datasets}
\label{precursor_sets}
The training and test sets are common to both experimental settings (influence of the percentage of observed labels, influence of labelling error).
\paragraph{Inference}
The precursor set $\mathcal{P}_{infer\_test}$ of the test-set is composed of $M=100$ fully labelled observation sequences of length $\ell= 1000$. These sequences were generated from the PHMC-LAR(2) model described in Eq. \ref{4.regime.mc}-\ref{init.law}.
A protocol repeated for each $N \in \{1, 10, 100\}$ produced a precursor set $\mathcal{P}_{N\_infer\_train}$ consisting of $N$ fully labelled observation sequences of length $T=100$. The generative model in Eq. \ref{4.regime.mc}-\ref{init.law} was used for this purpose.
\paragraph{Forecasting}
In this case, training sets are each reduced to a single sequence. In each such sequence, the sequence's prefix of size $T=100$ is used for model training, whereas the subsequence $T+1, \cdots, T+10$ is used for testing prediction accuracy. The sequences of the unique precursor set denoted $\mathcal{P}_{N=1\_forecast\_train\_test}$ are generated using Eq. \ref{4.regime.mc}-\ref{init.law}.
\subsection{Implementation}
\label{implementation}
Our experiments required intensive computing resources from a Tier 2 data centre (Intel 2630v4, 2$\times$10 cores 2.2 Ghz, 20$\times$6 GB). We exploited data-driven parallelization to replicate our experiments on various training sets. On the other hand, code parallelization allowed us to process multiple sequences simultaneously in the step E of the EM algorithm. The software programs dedicated to model training, hidden state inference and forecasting were written in Python 3.6.9. We used the NumPy and Scipy Python libraries.
The models were learnt through the EM algorithm with precision $\kappa = 10^{-6}$ and initialization procedure parameters $(L, N_{iter}) = (5, 10)$.
\subsection{Influence of the percentage of observed states}
\label{exp.1}
To analyze the impact of observed states, we varied the percentage $P$ of labelled observations (equivalently the percentage of observed states) in the training sets. $P$ was varied from $0\%$ (fully unsupervised case) to $100\%$ (fully supervised case), with steps of $10\%$. The aim is to evaluate the performance of intermediate cases for different sizes of the training datasets.
\subsubsection{Hidden state inference}
\label{exp1.eval.inference}
The test-set $\mathcal{S}_{infer\_test}$ was generated by unlabelling all states from the precursor set $\mathcal{P}_{infer\_test}$ described in Subsection \ref{precursor_sets} ($M=100$ fully observed sequences of length $\ell=1000$).
To generate the training sets, the following protocol was repeated for each $N \in \{1, 10, 100\}$ and for each percentage $P$: (i) considering the appropriate precursor set $\mathcal{P}_{N\_infer\_train}$ ($N$ fully observed sequences of length $T=100$) depicted in Subsection \ref{precursor_sets}, only a proportion of $P$ observations was kept labelled while the rest was unlabelled; (ii) this process was repeated $15$ times, each time varying which observations are kept labelled. Thus were produced $15$ training datasets $\mathcal{S}_{N\_P\_infer\_train\_1}, \cdots, \mathcal{S}_{N\_P\_infer\_train\_15}$.
The PHMC-LAR(2) model with 4 states was trained on each training set $\mathcal{S}_{N,P,infer\_train\_i}$, $i=1, \cdots, 15$. For each trained model, state inference was achieved for the $M$ fully hidden sequences of test-set $\mathcal{S}_{infer\_test}$, which yielded $M$ sequences of predicted labels. Inference performance was evaluated by comparing the true state sequences with the inferred ones, using the \textbf{Mean Percentage Error} (MPE) score defined as follows:
\begin{equation}
\text{MPE} = \frac{1}{M} \sum_{i=1}^{M} \left[ \frac{1}{\ell} \sum_{j=1}^{\ell} \bm{1}_{s_j \neq \hat{s}_j} \right],
\end{equation}
where $s_j$'s and $\hat{s}_j$'s are respectively observed and inferred states. The MPE score varies between $0$ and $1$. The lower the value of the MPE score, the higher the inference performance.
Figure \ref{inference.exp1} displays $95\%$ confidence interval for the MPE score as a function of $P$.
As expected, the results show that inference ability increases with the number of training sequences denoted by $N$. Note that when the proportion of labelled observations is less than some threshold ($P=30\%$ for $N=1, 10$ and $P=20\%$ for $N=100$), inference performance is greatly impacted by the distribution of observed states since we obtain very large confidence intervals for the MPE score.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig3a_inference_exp1_p_t=100_n=1.png}
\caption{N = 1}
\label{inference.exp1.N.1}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig3b_inference_exp1_p_t=100_n=10.png}
\caption{N = 10}
\label{inference.exp1.N.10}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig3c_inference_exp1_p_t=100_n=100.png}
\caption{N = 100}
\label{inference.exp1.N.100}
\end{subfigure}
\caption{$95 \%$ confidence interval for mean percentage error (MPE) of hidden state inference, as a function of the percentage $P$ of labelled observations. Models were trained on datasets of $N$ sequences of length $100$, for each of $15$ replicates differing by the $P \%$ labelled observations. For each model, inference was performed for a test-set of $100$ unlabelled sequences of size $1000$. The $95 \%$ confidence interval of the MPE score was computed from the $15$ replicates. The dash (red) line indicates the MPE score obtained for the unsupervised learning case ($P = 0\%$). Mind the differences in scales between the three subfigures.}
\label{inference.exp1}
\end{figure}
For $N = 1$, the use of labelled observations makes it possible to outperform the fully unsupervised case ($P = 0\%$) (which translates into small MPE scores) when at least $30\%$ of observations are labelled (see Fig. \ref{inference.exp1.N.1}). In contrast, for $N=10, 100$, from some threshold value of $P$ (respectively $30\%$ and $20\%$), the use of larger proportions of labelled observations sustains inference performances equal to that of the fully unsupervised case (see Fig. \ref{inference.exp1.N.10} and \ref{inference.exp1.N.100}). Importantly, the results show that using large proportions of labelled observations considerably speeds up model training by decreasing the number of iterations of the EM algorithm (see Fig. \ref{nbiter.exp1}), and allows to better characterize the training data (which is reflected by a greater likelihood, see Fig. \ref{log.ll.exp1}). \cite*{ramasso-denoeux_2013_jour-ieee-transact-fuzzy-sys_partially-hmms} had already underlined the beneficial impact of partial knowledge integration on EM convergence in HPMCs. Our work confirms this advantage in the PHMC-LAR model, with a good preservation of inference performance.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig4a_exp1_p_t=100_n=1_nb_iters.png}
\caption{N = 1}
\label{nbiter.exp1.N.1}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig4b_exp1_p_t=100_n=10_nb_iters.png}
\caption{N = 10}
\label{nbiter.exp1.N.10}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig4c_exp1_p_t=100_n=100_nb_iters.png}
\caption{N = 100}
\label{nbiter.exp1.N.100}
\end{subfigure}
\caption{Number of EM iterations before convergence as a function of the percentage $P$ of labelled observations. For the description of the experimental protocol, see caption of Fig.\ref{inference.exp1}. The distribution of the number of EM iterations is studied across $15$ replicates. Dash (red) line and dot (green) line indicate the number of iterations for unsupervised and supersived learning cases respectively. Mind the differences in scales between the three subfigures.}
\label{nbiter.exp1}
\end{figure}
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 0cm 0cm, clip]{fig5a_exp1_p_t=100_n=1_log_ll.png}
\caption{N = 1}
\label{log.ll.exp1.N.1}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 0cm 0cm, clip]{fig5b_exp1_p_t=100_n=10_log_ll.png}
\caption{N = 10}
\label{log.ll.exp1.N.10}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 0cm 0cm, clip]{fig5c_exp1_p_t=100_n=100_log_ll.png}
\caption{N = 100}
\label{log.ll.exp1.N.100}
\end{subfigure}
\vspace{-2mm}
\caption{Log-likelihood as a function of the percentage $P$ of labelled observations. For the description of the experimental protocol, see caption of Fig.\ref{inference.exp1}. The distribution of the log-likelihood is studied across $15$ replicates. Dash (red) line and dot (green) line indicate the log-likelihoods for unsupervised and supervised learning cases respectively. Mind the differences in scales between the three subfigures.}
\label{log.ll.exp1}
\end{figure}
In order to evaluate the influence of observed states in recognition phase, we considered the case $P=10 \%$ which previously obtained the lowest inference performance. This time, we also kept labelled a proportion $Q$ of observations within the test-set $\mathcal{S}_{infer\_test}$. We assessed the inference performances for the models trained on $\mathcal{S}_{N,P=10\%,infer\_train\_i}$, $i=1, \cdots 15$. Figure \ref{inference.exp1.bis} presents MPEs as a function of $Q$ for $N = 1, 10$ and $100$. We observe that inference performances are improved by the presence of observed states.
More precisely, for $Q$ taking its values in $25 \%$, $50 \%$ and $75 \%$, respectively, MPE decreases by: (i) $19\%$, $42\%$ and $69\%$ for $N=1$ (Fig. \ref{inference.exp1.bis.N.1}); (ii) $27\%$, $52\%$ and $77\%$ for $N=10$ (Fig. \ref{inference.exp1.bis.N.10}); and (iii) $27\%$, $53\%$ and $77\%$ for $N=100$. (Fig. \ref{inference.exp1.bis.N.100}). These results show the ability of our variant of the Viterbi algorithm to infer partially-labelled sequences.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=0cm 0cm 2cm 1cm, clip]{fig6a_inference_exp1_bis_p_t=100_n=1.png}
\caption{N = 1}
\label{inference.exp1.bis.N.1}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=0cm 0cm 2cm 1cm, clip]{fig6b_inference_exp1_bis_p_t=100_n=10.png}
\caption{N = 10}
\label{inference.exp1.bis.N.10}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=0cm 0cm 2cm 1cm, clip]{fig6c_inference_exp1_bis_p_t=100_n=100.png}
\caption{N = 100}
\label{inference.exp1.bis.N.100}
\end{subfigure}
\caption{$95 \%$ confidence interval for mean percentage error (MPE) of hidden state inference, as a function of the percentage $Q$ of labelled observations within test-set, with $P = 10\%$ labelled observations in the training sets. Models were trained on datasets of $N$ sequences of length $100$ in which $P=10\%$ of observations have been labelled. Fifteen replicates differing by the $P=10 \%$ labelled observations were considered. For each model, inference was performed for a test-set of $100$ partially labelled ($Q\%$) sequences of size $1000$. {\bf The $95 \%$ confidence interval of the MPE score was computed from the $15$ replicates.} Mind the differences in scales between the three subfigures.}
\label{inference.exp1.bis}
\end{figure}
\subsubsection{Forecasting}
\label{exp1.eval.forecasting}
In this experiment, we consider models trained on a single sequence. This case corresponds to many real-world situations in which a unique time series is available ({\it e.g.}, the evolution of air pollution at some geographical location). Using the precursor set $\mathcal{P}_{N=1\_forecast\_train\_test}$ described in Subsection \ref{precursor_sets}, we generated datasets $\mathcal{S}_{N=1\_forecast\_train\_test\_i}$, $i=1, \cdots, 15$ each composed of a single sequence of size $110$. Again, the $15$ replicates differed by the $P\%$ labelled observations. In these sets, the sequence prefixes of length $T=100$ were used to train the models. \textit{Out-of-sample} forecasting was carried out at horizons $T+h$, $h=1, \dots, 10$, which means that prediction accuracy was assessed using subsequences $T+1, \cdots, T+h$. To note, the $P\%$ labelled observations were distributed in the sequence prefixes of length $T$.
Two experimental schemes were considered. First, the states at forecast horizons were supposed to be latent; that is, all states were unlabelled from $T+1$ to $T+h$, $h=1, \cdots, 10$. Then, we performed the prediction evaluation when states are observed at forecast horizons. The latter situation corresponds to performing the prediction conditional on some assumption on the regime. For instance, in econometrics, assuming we know which phase will be on (growth phase {\it versus} recession) might improve the forecasting performance of the Gross National Product (GNP). In this case, all states were kept labelled from $T+1$ to $T+h$, $h=1, \cdots, 10$.
Prediction performance is estimated by the \textbf{Root Mean Square Error} (RMSE) defined as follows:
\begin{equation}
\text{RMSE}_h = \sqrt{\frac{1}{N_{rep}} \sum_{i=1}^{N_{rep}} (X_{T+h}^{(i)} - \hat{X}_{T+h}^{(i)})^2},
\end{equation}
where $h$ is the forecast horizon and $N_{rep}=15$ is the number of replicates. Accurate predictions are characterized by low RMSEs.
Table \ref{exp1.prediction.perf} presents the RMSEs obtained when the states at forecast horizons are supposed to be latent. Fig. \ref{forecasting.exp1.dist.unknown} presents the mean, median and maximum of RMSEs, computed over all forecast horizons, as a function of $P$, the percentage of labelled observations in the training sets. Table \ref{exp1.prediction.perf} and Fig. \ref{forecasting.exp1.dist.unknown} show that as from some low $P$ threshold ($10\%$ or $20\%$), the prediction performance remains nearby constant across proportions.
\begin{table}[t!]
\scalefont{0.85}
\centering
\begin{tabular}{p{0.5cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm}}
\hline
\backslashbox{$P$}{h} & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\
\hline
0 & 1.860 & \bf{6.680} & 1.830 & 3.165 & 4.167 & 2.540 & 1.133 & 7.938 & 7.854 & 2.465 \\
10 & 2.035 & 8.273 & 1.829 & \bf{2.909} & 4.477 & 2.851 & \bf{0.957} & \bf{7.667} & \bf{7.583} & \bf{2.224} \\
20 & 1.934 & 7.612 & \bf{1.337} & 3.161 & 4.110 & 2.482 & 1.189 & 7.991 & 7.907 & 2.518 \\
30 & 1.323 & 7.450 & 1.373 & 3.168 & \bf{4.093} & \bf{2.469} & 1.201 & 8.005 & 7.921 & 2.532 \\
40 & 1.293 & 7.496 & 1.392 & 3.158 & 4.103 & 2.480 & 1.191 & 7.994 & 7.911 & 2.521 \\
50 & 1.308 & 7.525 & 1.402 & 3.135 & 4.122 & 2.496 & 1.174 & 7.978 & 7.894 & 2.505 \\
60 & 1.394 & 7.502 & 1.424 & 3.115 & 4.134 & 2.508 & 1.162 & 7.965 & 7.882 & 2.493 \\
70 & 1.363 & 7.560 & 1.431 & 3.094 & 4.155 & 2.527 & 1.142 & 7.946 & 7.862 & 2.473 \\
80 & 1.306 & 7.502 & 1.395 & 3.129 & 4.116 & 2.489 & 1.179 & 7.984 & 7.900 & 2.511 \\
90 & 1.294 & 7.569 & 1.444 & 3.088 & 4.155 & 2.526 & 1.142 & 7.947 & 7.863 & 2.473 \\
100 & \bf{1.267} & 7.613 & 1.447 & 3.076 & 4.164 & 2.535 & 1.132 & 7.937 & 7.854 & 2.464 \\
\hline
\end{tabular}
\caption{Root mean square error (RMSE) of prediction at horizon $h$ for different values of $P$, when the states are unknown throughout forecast horizons. $P$ is the percentage of labelled observations within the training datasets. The forecast horizons are time-steps $T+1$ to $T+h$, $T=100$. For a given value of $P$, models were each trained on a unique sequence: the sequence's prefix of length $T=100$ was used for training, for each of $15$ replicates differing by the $P \%$ labelled observations distributed in the prefix. Then, out-of-sample forecasting was carried out at time-steps $T+1, \dots, T+10$, for the same sequence. The figures in bold highlight the minimum RMSE obtained across all labelling percentages ($P$), at each horizon ($h$) considered}
\label{exp1.prediction.perf}
\scalefont{1.0}
\end{table}
\begin{figure}[]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig7a_forecasting_exp1_p_t=100_n=1.png}
\caption{}
\label{forecasting.exp1.dist.unknown}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig7b_forecasting_exp1_bis_p_t=100_n=1.png}
\caption{}
\label{forecasting.exp1.dist.known}
\end{subfigure}
\caption{Mean, median and maximum root mean square error (RMSE) of prediction at horizon $h$ as a function of $P$, the percentage of labelled observations in the training datasets. States at forecast time-steps $T+h,$ $h=1, \cdots 10$ are (a) hidden, (b) known. Models were trained on a single sequence, for each of $15$ replicates differing by the $P \%$ labelled observations. Model training was performed on subsequences of length $100$, whereas prediction was achieved for the $10$ subsequent time-steps. For each value of $P$, the statistics provided were computed across the $15$ replicates and all horizons.}
\label{forecasting.exp1.dist}
\end{figure}
In addition, Table \ref{exp1.prediction.perf} also highlights that the ability to predict depends on the forecast horizon under consideration. At any given labelling percentage $P$, high RMSE scores ({\it i.e.}, around $7$) alternate with low scores (around $1$) across horizons. The nonmonotonic error trend across horizons was observed empirically for MS-AR models and threshold autoregressive models when they are applied to US GNP time series \citep{clements-krolzig_1998_journ-econometrics_forecasting-msar-vs-threshold-ar}.
Finally, our experiments show that PHMC-LAR model's ability to better characterize the training data in presence of large proportions of labelled observations (characterized by greater likelihood, see Fig. \ref{log.ll.exp1.N.1}) does not translate into an improved forecast performance.
When states are known at forecast horizons, RMSEs (presented in Table \ref{exp1.prediction.perf.observed.states}) are reduced by $44\%$ on average. Moreover, Fig. \ref{forecasting.exp1.dist.known} shows that above percentage $P = 30 \%$, prediction performances are slightly greater than that of the unsupervised case ($P=0 \%$). Note that as in the case when the states are unknown at forecast horizons, the prediction ability depends on the forecast horizon. Again, for a given $P$, the RMSE score does not systematically increase with forecast horizon $h$, although previously predicted values are used as inputs when predicting at next horizons.
\begin{table}[ht]
\scalefont{0.85}
\centering
\begin{tabular}{p{0.5cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} }
\hline
\backslashbox{$P$}{h} & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\
\hline
0 & 0.083 & 0.325 & 0.870 & 1.577 & 1.509 & 1.171 & 2.220 & 1.216 & \bf{0.996} & \bf{1.104} \\
10 & 3.730 & 1.791 & 2.936 & 3.593 & 4.914 & 4.153 & 4.060 & 4.873 & 3.603 & 3.902 \\
20 & 1.831 & 0.510 & 1.806 & 1.939 & 2.171 & 1.250 & 2.581 & 1.438 & 1.239 & 1.458 \\
30 & 0.083 & \bf{0.321} & 0.854 & 1.542 & 1.477 & 1.158 & 2.167 & 1.137 & 1.109 & 1.289 \\
40 & 0.070 & 0.325 & 0.841 & 1.532 & 1.460 & 1.150 & 2.151 & 1.154 & 1.084 & 1.301 \\
50 & 0.065 & 0.324 & 0.832 & 1.540 & 1.459 & 1.154 & 2.161 & 1.130 & 1.078 & 1.229 \\
60 & 0.063 & 0.329 & 0.829 & 1.524 & 1.443 & 1.145 & 2.137 & 1.125 & 1.039 & 1.181 \\
70 & 0.057 & 0.329 & 0.810 & 1.531 & 1.431 & 1.143 & 2.143 & 1.118 & 1.036 & 1.263 \\
80 & 0.036 & 0.327 & 0.810 & 1.490 & 1.411 & 1.134 & 2.086 & 1.086 & 1.072 & 1.276 \\
90 & 0.036 & 0.325 & 0.788 & 1.479 & 1.386 & 1.124 & 2.065 & 1.067 & 1.023 & 1.161 \\
100 & \bf{0.001} & 0.326 & \bf{0.760} & \bf{1.473} & \bf{1.368} & \bf{1.121} & \bf{2.053} & \bf{1.065} & 1.002 & 1.133 \\
\hline
\end{tabular}
\caption{Root mean square error (RMSE) of prediction at horizon $h$ for different values of $P$, when the states are known throughout forecast horizons. $P$ is the percentage of labelled observations within the training datasets. The forecast horizons are time-steps $T+1$ to $T+h$, $T=100$. For the description of the experimental protocol, see caption of Table \ref{exp1.prediction.perf}. The states are known from $T+1$ to $T+10$ time-steps. The figures in bold highlight the minimum RMSE obtained across all labelling percentages ($P$), at each horizon ($h$) considered}
\label{exp1.prediction.perf.observed.states}
\scalefont{1.0}
\end{table}
\subsection{Influence of labelling error}
\label{exp.2}
In this experiment, the influence of labelling error is evaluated. To simulate unreliable labels, we proceeded as follows.
At each time-step $t$, an error probability $p_t$ was drawn randomly from a beta distribution with mean $\rho$ and variance $0.2$. With probability $p_t$, the observed state $s_t$ was replaced by a random state uniformly chosen from $\{1, 2, 3, 4\} \setminus \{s_t\}$. So, the unreliable labels $\tilde{s}_t$ were defined as follows:
\begin{equation}
\label{unreliable.labels}
\begin{split}
&p_t \sim \beta(0.2,\rho)\\
&\tilde{s}_t =
\left\{
\begin{aligned}
s_t \qquad \qquad \qquad \text{with probability} \quad 1 - p_t \\
\mathcal{U} \left(\{1, 2, 3, 4\} \setminus \{s_t\} \right) \quad \text{with probability} \quad p_t
\end{aligned}
\right.
\end{split}
\end{equation}
where $\mathcal{U}$ is the discrete-valued uniform distribution. Thus, on average a proportion $\rho$ of observations is assigned wrong labels.
\subsubsection{Inference of hidden states}
\label{eval.inference}
To assess inference performance in presence of labelling errors, we relied on the test-set $\mathcal{S}_{infer\_test}$ described in Subsection \ref{exp.1} ($M=100$ fully hidden sequences of length $\ell=1000$) corresponding to the fully labelled dataset $\mathcal{P}_{infer\_test}$.
To generate the training sets, for each N $\in \{1,\ 10,\ 100\}$, we considered the appropriate precursor set $\mathcal{P}_{N\_infer\_train}$ ($N$ fully observed sequences of length $T=100$) depicted in Subsection \ref{precursor_sets}.
We varied the mean labelling error probability $\rho$ in $\{0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7,$ $0.8, 0.9, 0.95 \}$. For $N \in \{1,\ 10,\ 100\}$, and each value of $\rho$, we generated $15$ replicates from dataset $\mathcal{P}_{N\_infer\_train}$, each time varying the distribution of the wrong labels amongst the observations. The PHMC-LAR(2) model with 4 states was trained on each of the training sets $\mathcal{S}_{N,\rho,infer\_train\_1}, \cdots, \mathcal{S}_{N\_\rho\_infer\_train\_15}$ thus obtained.
For each trained model, state inference was achieved, which yielded $M=100$ sequences of predicted labels of length $1000$, to be compared with the label sequences within $\mathcal{P}_{infer\_test}$ (see Subsection \ref{generative_model}).
Figure \ref{inference.exp2} presents $95\%$ confidence intervals for the MPE score as a function of $\rho$. Note that for all sizes $N \in \{1,\ 10,\ 100\}$ of training data, the average MPE gradually increases when $\rho$ tends to $1$. Moreover, confidence intervals become more and more tight when larger training data is considered. We also observe that up to $\rho = 0.7$, the robustness to labelling errors, translated into small MPE average and low dispersion, increases with $N$. However, from $\rho \ge 0.8$, this trend is reversed and inference performance slightly decreases when $N$ grows.
On the other hand, we underline that the fully unsupervised case outperforms supervised cases in presence of labelling errors. Up to relatively high labelling error rates ($\rho=70\%$), the trade-off between training time and inference performance becomes beneficial for large training datasets. For instance, for $N=100$, with a $70\%$-reliable labelling function ({\it i.e.} $\rho = 0.3$), the EM algorithm converges after a single iteration against $67$ iterations for the unsupervised case; and the resulting model has good inference abilities with an MPE score equal to $35\%$ on average (see Fig. \ref{inference.exp2.N.100}) against $5\%$ on average in the unsupervised case. Thus, when analyzing real-world data for which the number of states $K$ and auto-regressive order $p$ are unknown, model selection strategies can capitalize on such labelling functions in order to explore/prospect larger grids of values for the hyperparameters $K$ and $p$.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig8a_inference_rho_exp2_t=100_n=1.png}
\caption{N = 1}
\label{inference.exp2.N.1}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig8b_inference_rho_exp2_t=100_n=10.png}
\caption{N = 10}
\label{inference.exp2.N.10}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig8c_inference_rho_exp2_t=100_n=100.png}
\caption{N = 100}
\label{inference.exp2.N.100}
\end{subfigure}
\caption{$95 \%$ confidence interval for mean percentage error (MPE) of hidden state inference, as a function of the mean labelling error probability $\rho$. Models were trained on $N$ sequences, for each of $15$ replicates differing by the $\rho \%$ ill-labelled observations. The average MPE was computed from the $15$ replicates. The dash (red) line indicates the MPE score obtained for the unsupervised learning case. Mind the differences in scales between the three subfigures.}
\label{inference.exp2}
\end{figure}
\subsubsection{Forecasting}
\label{eval.forecasting}
As in Subsection \ref{exp1.eval.forecasting}, we considered models trained on a single sequence ($N = 1$). Again, for each value of the mean labelling error probability $\rho$, we used precursor set $\mathcal{P}_{N=1\_forecast\_train\_test}$ described in Subsection \ref{precursor_sets}, and we varied the distribution of wrong labels: $15$ replicates ({\it i.e.}, $15$ sequences of length $T=100$) were thus generated. Out-of-sample forecasting was carried out at horizons $T+h$, $h = 1,\cdots,10$.
Table \ref{exp2.prediction.perf} presents RMSE scores for different values of mean labelling error $\rho$ when states are unknown at forecast horizons $h = 1, \dots, 10$. The results show that at forecast horizons $h = 1, 2, 5, 6$, the best prediction accuracies are reached when $\rho$ is null, whereas at the remaining horizons, the highest accuracies are obtained when $\rho = 0.8$ or $0.9$. Figure \ref{exp2.prediction.perf.fig} presents the mean, median and maximum for the prediction errors computed over the whole forecast horizons as a function of $\rho$. We observe that the mean and median very slightly increase with $\rho$, whereas labelling errors exert a greater impact on the maximum values of RMSEs. Therefore, this second experiment also highlights the remarkable robustness to error labelling in the prediction task, over the whole range of error rates.
\begin{table}[t]
\scalefont{0.85}
\centering
\begin{tabular}{p{0.5cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm} p{0.6cm}}
\hline
\backslashbox{$\rho$}{h} & 1 & 2 & 3 & 4 & 5 & 6 & 7 & 8 & 9 & 10 \\
\hline
0 & \bf{1.267} & \bf{7.613} & 1.447 & 3.076 & \bf{4.164} & \bf{2.535} & 1.132 & 7.937 & 7.854 & 2.464 \\
0.1 & 1.814 & 8.992 & 1.393 & 3.193 & 4.334 & 2.625 & 1.117 & 7.865 & 7.780 & 2.407 \\
0.2 & 2.258 & 10.315 & 1.529 & 2.855 & 4.758 & 3.026 & 0.811 & 7.481 & 7.398 & 2.044 \\
0.3 & 2.793 & 10.458 & 1.575 & 2.911 & 4.689 & 3.004 & 0.801 & 7.512 & 7.426 & 2.062 \\
0.4 & 2.877 & 11.457 & 1.161 & 3.114 & 4.779 & 2.941 & 0.886 & 7.562 & 7.478 & 2.123 \\
0.5 & 2.655 & 11.104 & 1.396 & 2.953 & 4.812 & 3.008 & 0.843 & 7.488 & 7.410 & 2.062 \\
0.6 & 2.925 & 12.013 & 1.004 & 3.031 & 4.878 & 3.002 & 0.749 & 7.472 & 7.392 & 2.020 \\
0.7 & 3.088 & 11.643 & 1.271 & 2.969 & 4.901 & 3.082 & 0.706 & 7.409 & 7.321 & 1.954 \\
0.8 & 2.656 & 11.860 & \bf{0.905} & 3.001 & 4.848 & 2.954 & 0.768 & 7.498 & 7.422 & 2.046 \\
0.9 & 2.444 & 11.667 & 1.338 & \bf{2.786} & 5.011 & 3.164 & \bf{0.647} & \bf{7.310} & \bf{7.234} & \bf{1.875} \\
0.95 & 2.362 & 11.071 & 1.192 & 3.027 & 4.685 & 2.905 & 0.866 & 7.588 & 7.504 & 2.135 \\
\hline
\end{tabular}
\caption{Root mean square error (RMSE) of prediction at horizon $h$ for different values of the mean labelling error probability $\rho$, when the states are unknown throughout forecast horizons. The forecast horizons are time-steps $T+1$ to $T+h$, $T=100$. The states are unknown from $T+1$ to $T+10$ time-steps. For a given value of $\rho$, models were each trained on a unique sequence: the sequence's prefix of length $T=100$ was used for training, for each of $15$ replicates differing by the position of ill-labelled observations distributed in the prefix. Then, out-of-sample forecasting was carried out at time-steps $T+1, \dots, T+10$, for the same sequence. The figures in bold highlight the minimum RMSE obtained across all mean labelling error probabilities ($\rho$), at each horizon ($h$) considered}
\label{exp2.prediction.perf}
\scalefont{1.0}
\end{table}
\begin{figure}[h]
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig9a_forecasting_exp2_rho_t=100_n=1_mean_median.png}
\caption{}
\label{exp2.prediction.perf.mean}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth, trim=1cm 0cm 2cm 1cm, clip]{fig9b_forecasting_exp2_rho_t=100_n=1_max.png}
\caption{}
\label{exp2.prediction.perf.max}
\end{subfigure}
\caption{Descriptive statistics for the distribution of the root mean square error (RMSE) of prediction, as a function of $\rho$: (a) mean and median for prediction error, (b) maximum for prediction error. $\rho$ denotes the mean labelling error probability. The forecast horizons are time-steps $T+1$ to $T+h$, $T=100$. The statistics are computed over all horizons.}
\label{exp2.prediction.perf.fig}
\end{figure}
\newpage
\section{Conclusion}
\label{conclusion}
In this work, we have introduced the PHMC-LAR model to analyze time series subject to switches in regimes. Our model is a generalization of the well-known Hidden Regime-switching Autoregressive (HRSAR) and Observed Regime-switching Autoregressive (ORSAR) models when regime-switching is modelled by a Markov Chain. Our model allows to handle the intermediate case where the state process is partially observed.\\
\indent In the evaluation, we conducted our experiments on simulated data and considered both inference performance and prediction accuracy. The results show that the partially observed states (when they represent a reasonable proportion) allow a better characterization of training data (reflected by greater log-likelihood), in comparison with the unsupervised case. An interesting characteristics of the PHMC-LAR model is that the partially observed states allow faster convergence for the learning algorithm. This performance is obtained with no or practically no impact on the quality of hidden state inference, as from labelling percentages around $20\%$-$30\%$; the prediction accuracy is also preserved above such percentage thresholds. Furthermore, faster EM convergence is also verified in a fully supervised scheme where part of the observations is ill-labelled. Model selection strategies can therefore rely on an approximate labelling function (provided by an expert or by a supervised algorithm learnt on a small subset of data for which the true labels are known), to explore larger grids of hyperparameter values. In addition, complementary experimental studies have revealed the robustness of our model to labelling errors, particularly when large training datasets and moderate labelling error rates are considered. Finally, we showed the ability of our variant of the Viterbi algorithm to infer partially-labelled sequences.\\
\indent A natural extension of the PHMC-LAR model consists in putting uncertainty on partial knowledge: for instance instead of states observed with no doubt, a subset of possible states with various occurrence probabilities can be considered at each time-step.
On the other hand, it is more realistic to consider time-dependent state processes, especially when large time series are analyzed. These directions will be investigated in future work.
\section*{Acknowledgements}
The software development and the realization of the experiments were performed at the CCIPL (Centre
de Calcul Intensif des Pays de la Loire, Nantes, France).
\section*{Funding}
Fatoumata Dama is supported by a PhD scholarship granted by the French Ministery for Higher Education, Research and Innovation.
|
\section{Introduction}
In spite of an increasing variety of supernovae (SNe) sub-types, the majority of SNe are generally believed to release $0.5 - 2 \times 10^{51}$ erg, although superluminous SNe may release substantially more \citep{Howell2017}. Only a small percentage of this comes out as visible light, with most of a SN's energy initially carried away
in the form of kinetic energy. It is this enormous point deposition of energy that has a significant impact on the structure and energy content of a galaxy's interstellar medium through an expanding remnant that can last largely intact up to $\simeq$10$^{5}$ yr \citep{Cox1974,McKee1977,Shull1989,Blondin1998}.
Currently, there are some 300 confirmed Galactic supernova remnants (SNRs) catalogued with dozens of other suspected SNRs and more added every few years \citep{Safi2012,Green2019}. Most Galactic SNRs are less than a degree in angular size, more than 1 kpc distant and well evolved, with typical estimated ages between 10$^{4}$ and 10$^{5}$ yr.
\begin{deluxetable*}{lcccc}[ht]
\tablecolumns{5}
\tablecaption{Locations and Dimensions of SNRs }
\tablewidth{0pt}
\tablehead{
\colhead{SNR ID} & \colhead{Galactic Coordinates} & \colhead{Approximate Center (J2000)} & \colhead{Angular Size} & \colhead{Estimated Distance} }
\startdata
G249+24 & $l$ = 249.7 $b$ = $+24.7$ & RA = 09$^{\rm h}$33$^{\rm m}$ Dec = $-17\degr$00$'$ & $4.5\degr $ & $<$390 pc \\
Antlia & $l$ = 275.5 $b$ = $+18.4$ & RA = 10$^{\rm h}$38$^{\rm m}$ Dec = $-37\degr$20$'$ & $20\degr \times 26\degr$ & $\sim$250 pc \\
G354-33 & $l$ = 354.0 $b$ = $-33.5$ & RA = 20$^{\rm h}$16$^{\rm m}$ Dec = $-45\degr$50$'$ & $11\degr \times 14\degr$ & $\sim$500 pc \\
\enddata
\end{deluxetable*}
Out-sized attention relative to their population percentage has been directed to the handful of the Milky Way's young remnants with ages less than 5000 yr. These include the Crab Nebula, Cassiopeia A, the remnants of Tycho and Kepler, the Vela remnant, and Puppis A. This increased attention is due to their high expansion velocities, metal-rich ejecta, bright emission across the entire electromagnetic spectrum, and clearer connections to the various core-collapse and thermonuclear SN sub-types.
The majority of Galactic SNRs are first detected through radio observations due to their characteristic synchrotron nonthermal radio emission associated with shocked gas
\citep{Downes1971,Chevalier1977,Green1984,Green2004}. Such nonthermal emission leads to a power law flux density, S, with S $\varpropto$ $\nu^{-\alpha}$ where $\alpha$ is the emission spectral index with typical SNR values between -0.3 and -0.7. Rarer pulsar wind-driven SNRs exhibit a much flatter radio spectrum, with spectral indices around zero.
While multi-frequency radio surveys have been historically the dominant tool for finding SNRs,
X-ray studies have also led to the discovery of several additional SNRs. These include RX J1713.7-3946 (G347.3-0.5; \citealt{Pfeff1996}) and the young ``Vela Junior'' SNR (G266.2-1.2; \citealt{Aschenbach1998}) coincident with the much larger Vela SNR. Recent examples of X-ray confirmed or discovered remnants include several new SNRs in the LMC \citep{Maggi2014} and X-ray confirmation of the suspected Galactic remnant G53.4+0.0 \citep{Driessen2018}.
Despite the fact that less than 50\% of Galactic SNRs exhibit any appreciable associated optical emission, a remnant's optical emission can be useful in confirming the presence of high-velocity shocks and in defining a remnant's overall size and
morphology. Although discoveries of new SNRs in the optical is relatively rare, there have been several recently reported discoveries \citep{Stupar2007,Boumis2009,Stupar2012,Fesen2015,Neustadt2017,Stupar2018,How2018,Fesen2019} along with many proposed SNR candidates \citep{Stupar2008,Stupar2011,Boumis2009,Ali2012,Sabin2013}.
The main criterion for optical SNR identification is a line ratio of I([\ion{S}{2}])/I(H$\alpha$) $\geq 0.4$ which has proven useful in identifying the shocked emission of SNRs in both the Milky Way and nearby local group galaxies \citep{Blair1981,Dopita1984,Fesen1985,Leonidaki2013,Long2017}. Recently, the near infrared [\ion{Fe}{2}] line emissions at 1.27 and 1.64 $\mu$m have also been used to detect dust obscured SNRs in nearby galaxies (see review by \citealt{Long2017}).
One wavelength regime that has not yet been fully exploited to search for new Galactic SNRs is the ultraviolet (UV).
Here we present results of an initial study of SNRs located far away from the Galactic plane using wide field-of-view (FOV) UV images assembled from the Galaxy Evolution Explorer (GALEX) All-Sky survey \citep{Bianchi2009}. We began this research by investigating two unusually high latitude suspected supernova remnants including the exceptionally large Antlia remnant \citep{McCull2002} (see Table 1).
During this work, we also found a new apparent SNR.
However, after our initial submission, W. Becker kindly sent us a
preprint of their independent discovery of this new remnant based
on SRG/eRosita X-ray and Parks All-Sky Survey observations.
Since their paper has now been published \citep{Becker2021}, we have cited their
results in reference to our independent findings. UV and optical
images plus some follow-up optical spectra are described in
Section 2, with results presented in Section 3. We discuss the
general properties of these SNRs in Section 4, with our
conclusions and discussion of helpful follow-up observations
given in Section 5.
\begin{figure*}[ht]
\centerline{\includegraphics[angle=0,width=10.0cm]{GALEX_MAP_FIXED.jpg}}
\caption{GALEX FUV intensity map of half of the sky in Galactic coordinates, showing locations of two apparent SNRs, G249+24 and G354-33, and the exceptionally large Antlia SNR. White areas denote regions with no GALEX imaging data. \label{globe} }
\end{figure*}
\begin{deluxetable*}{llcccccc}[ht]
\tablecolumns{8}
\tablecaption{Optical Data on SNRs }
\tablehead{\colhead{SNR} & \colhead{Survey or ~~~~~~~~~~~ } & \colhead{Camera or} &
\multicolumn{3}{c}{\underline{~~~~~~~~~~~~~~~~~~~~~~~~Direct Images~~~~~~~~~~~~~~~~~~~~~~~~}} & \multicolumn{2}{c}{\underline{~~~~~~~~~~Spectra~~~~~~~~~~} } \\
\colhead{ID} & Observatory & Telescope & \colhead{Filters} & \colhead{FOV} & Pixel Scale &
\colhead{Wavelength} & \colhead{FWHM} }
\startdata
G249+24 & MDW H$\alpha$ Survey & 130 mm lens & H$\alpha$ &
$3.4\degr \times 3.4\degr$ & 3.17$''$ & \nodata & \nodata \\
& MDM Observatory & 2.4 m & H$\alpha$ + [\ion{N}{2}] &
$18' \times 18'$ & 0.55$''$ & $\lambda$4000 - $\lambda$6900 & 3.5 \AA \\
& MDM Observatory & 2.4 m & \nodata & \nodata & \nodata & $\lambda$5735 - $\lambda$6065 & 0.55 \AA \\
Antlia & SHASSA Survey & f/1.6 52 mm lens & H$\alpha$, $\lambda$6440, $\lambda$6770 &
$13\degr \times 13\degr$ & 47.6$''$ & \nodata & \nodata \\
& MDW H$\alpha$ Survey & 130 mm lens & H$\alpha$ &
$3.4\degr \times 3.4\degr$ & 3.17$''$ & \nodata & \nodata \\
& Ross Creek Obs. & f/2.0 200 mm lens &
H$\alpha$ + [\ion{N}{2}], [\ion{O}{3}], [\ion{S}{2}] &
$10.3\degr \times 6.9\degr$ & 3.87$''$ & \nodata & \nodata \\
& SALT Observatory & 10 m & \nodata & \nodata & \nodata & $\lambda$4050 - $\lambda$7120 & 5.0 \AA \\
G354-33 & SHASSA Survey & f/1.6 52 mm lens & H$\alpha$, $\lambda$6440, $\lambda$6770 &
$13\degr \times 13\degr$ & 47.6$''$ & \nodata & \nodata \\
& CHILESCOPE & 0.5 m & H$\alpha$ + [\ion{N}{2}], [\ion{O}{3}] &
$1.1\degr \times 1.1\degr$ & 0.96$''$ & \nodata & \nodata \\
\enddata
\label{Table_2}
\end{deluxetable*}
\section{Data and Observations}
\subsection{GALEX UV Images}
The GALEX satellite was a NASA science mission led by the California Institute of Technology who operated it from July 2003 through February 2012. The main instrument was a 50 cm diameter modified Ritchey-Chrétien telescope, a dichroic beam splitter and astigmatism corrector, and two microchannel plate detectors to simultaneously cover two wavelength bands with a 1.25 degree field of view with a resolution of $1.5''$ pixel$^{-1}$. Direct images were obtained in two broad bandpasses: a far-UV (FUV) channel sensitive to light in the 1344 to 1786 \AA\/ range, and a near-UV (NUV) channel covering 1771 to 2831 \AA \ \citep{Morrissey2007}. Resulting images are circular in shape with an image FWHM resolution of $\sim4.2''$ and $\sim5.3''$ in the FUV and NUV bands, respectively.
Being mainly a mission to study the UV properties of galaxies in the local universe, GALEX survey images largely avoided the complex and external galaxy poor Galactic plane. However, even its All-Sky Survey program focused away from the plane of the Milky Way did not cover all possible areas, leading to numerous gaps (see Fig.~\ref{globe}). Nonetheless, some 26,000 square degrees were imaged to a depth of m$_{\rm AB}$ = 20.5 \citep{Bianchi2009}. Using the on-line GALEX images, large FUV and NUV image mosaics were examined of several regions typically more than 10 degrees off the Galactic plane.
\subsection{Optical images}
Several different sources of narrow passband optical images were used in our investigation of the three SNR candidates.
Table 2 lists details of these optical images.
Wide-field, low-resolution H$\alpha$ images of regions around the suspected SNRs G249+24 and Antlia
obtained as part of the MDW Hydrogen-Alpha Survey\footnote{https://www.mdwskysurvey.org} were used to provide general optical size and emission features.
This survey uses a 130 mm f/2 camera at the New Mexico Skies Observatory, with a FLI ProLine 16803 CCD and 3 nm filter centered on H$\alpha$. This telescope-camera system produced field-of-view (FOV) of approximately $3.5\degr \times3.5\degr$ with a pixel size of $3\farcs17$. Each field was observed 12 times each with an exposure of 1200 s.
We have also made large mosaics from Southern H$\alpha$ Sky Survey Atlas (SHASSA; \citealt{Gaustad2001}). This wide-angle robotic survey covered $\delta$ = $+15\degr$ to $-90\degr$
with $13\degr$ square images with an angular resolution of $\simeq0.8'$. Narrow passband continuum filters on blue and red sides of H$\alpha$ centered at 6440 and 6770 \AA \ allow for stellar and background subtraction. The survey had a native sensitive level of $1.2 \times 10^{-17}$
erg cm$^{-2}$ s$^{-1}$ arcsec$^{-2}$ which could be significantly lowered with smoothing.
The western region of the Antlia remnant where it abuts the northeastern limb of the Gum Nebula was imaged using H$\alpha$ + [\ion{N}{2}],
[\ion{O}{3}] and [\ion{S}{2}] filters using a 200 mm f/2 lens and a $9576 \times 6388$ pixel ZWO ASI6200mm Pro
CCD. This system provided a FOV of $10.3\degr \times 6.9\degr$ with a $3.87''$ pixel$^{-1}$ image scale. Total exposure times for the three filters listed above were 6.9 hr, 13.3 hr, and 6.8 hr respectively.
Higher resolution images of the the northwestern limb of the suspected G354-33 remnant were obtained with a 0.5 m Newtonian telescope at the CHILESCOPE Observatory located 25 km south of the Gemini South telescope in
the Chilean Andes. A $4096 \times 4096$ FLI PROLINE 16803 CCD camera was used that provided a
FOV of $1.1\degr \times 1.1\degr$ with a
a pixel scale of 0\farcs963 pixel$^{-1}$. Several 10 minutes exposures were taken with H$\alpha$ + [\ion{N}{2}]
and [\ion{O}{3}] filters.
Images of optical filaments detected in MDW images associated with GALEX FUV filaments in G249+24 were obtained with a 2.4 m telescope at the MDM Observatory at Kitt Peak, Arizona using
using the Ohio State Multi-Object Spectrograph (OSMOS; \citealt{Martini2011}) in direct imaging mode. With a $4096 \times 4096$ CCD, this telescope/camera system yielded a clear FOV of $18' \times 18'$. With on-chip $2 \times 2$
or $4 \times 4$ pixel binning, spatial resolution was $0.55''$ and $1.10''$, respectively. Three exposures of 600 to 1200 s each using an H$\alpha$ + [\ion{N}{2}] filter were taken at four positions where long slit OSMOS spectra were also obtained (see below).
\begin{figure*}[htp]
\centerline{\includegraphics[angle=0,width=19.0cm]{Fig_FUV_J2000_v3.jpg}}
\caption{GALEX FUV intensity map of G354-33 showing a roughly spherical shell of UV emission filaments. Approximate shell center: 20$^{\rm h}$ 16.0$^{\rm m}$ 36$^{\rm s}$ $-45$\degr 40$'$. Note: Individual circular GALEX
images are $1.2\degr$ in diameter. \label{G1_FUV}
}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.49\textwidth]{G1_FUV_North_blowup_v1.jpg}
\includegraphics[angle=0,width=0.49\textwidth]{G1_FUV_West_blowup_v1.jpg}
\includegraphics[angle=0,width=0.49\textwidth]{G1_FUV_SE_blowup_v2.jpg}
\includegraphics[angle=0,width=0.49\textwidth]{G1_FUV_SW_blowup_v1.jpg}
\caption{Blow-ups of the GALEX FUV mosaic of G354-33 along the remnant's northern
(upper left), western (upper right), southeastern (lower left) and southwestern (lower right) limbs.
\label{G1_blow-ups}
}
\end{center}
\end{figure*}
\subsection{Optical Spectra}
Low-dispersion optical spectra of filaments in the northern suspected SNR, G249+24, were obtained with the MDM 2.4m Hiltner telescope using Ohio State Multi-Object Spectrograph (OSMOS; \citealt{Martini2011}). Employing a blue VPH grism (R = 1600) and a 1.2 arcsec wide slit, exposures of $2 \times 900$ s were taken covering 4000--6900 \AA \ with a spectral resolution of 1.2 \AA \ pixel$^{-1}$ and a FWHM = 3.5 \AA. Spectra were extracted from regions clean of appreciable emission along each of the $15'$ long slits.
As part of a program looking for high-velocity interstellar \ion{Na}{1} absorption features in background stars in the direction of Galactic SNRs, moderate-dispersion spectra of over a dozen early-type stars
with projected locations toward G249+24 and with estimated Gaia distances
between 300 and 2000 pc were also observed on with the MDM 2.4m Hiltner
The data were taken using a Boller \& Chivens
Spectrograph (CCDS) which employs a Loral 1200 $\times$ 800 CCD detector. Using a 1800 grooves mm$^{-1}$ grating blazed at 4700 \AA \ yielded a wavelength coverage of 330 \AA \ with a spectral scale of 0.275 \AA \ per pixel. Exposure times varied from $1-2 \times 400$ s to $3 \times 1200$ s depending on star brightness and sky conditions.
A $1.3''$ wide slit was used which resulted in a measured FWHM of 2.9 pixels, corresponding to a spectral resolution of 0.80 \AA \ providing an effective R $\simeq$7400 at the \ion{Na}{1} $\lambda\lambda$5890,5896 lines.
Although this spectral resolution is low
compared to conventional interstellar absorption studies where R values typically equal or exceed 30~000, this spectrograph provided sufficient velocity resolution at the \ion{Na}{1} lines for detecting $\geq$ 40 km s$^{-1}$ interstellar absorption features seen in SNRs \citep{Jenkins1976J,Danks1995,Welsh2003,Sallmen2004,Fesen2018,Ritchely2020}.
Wavelength calibration was done using Hg and Ne comparison lamps and have been corrected to local standard of rest (LSR) values. Deblended absorption feature velocities are believed accurate to $\pm$4 km s$^{-1}$.
For southern hemisphere Antlia nebula, spectra were taken of suspected SNR filaments using the Robert Stobie Spectrograph (RSS) on the 10~m SALT telescope in South Africa. Using a 900 lines per mm grating and a 1.5$''$ wide slit, spectra were obtained covering 4050 to 7120 \AA \ region with a FWHM resolution of 5 \AA \ and a dispersion of 1.0 Å pix$^{-1}$. Exposures ranged from $2 \times 300$s to $2 \times 600$s.
MDM spectra were reduced using using Pyraf software and OSMOS reduction pipelines\footnote{https://github.com/jrthorstensen/thorosmos} in Astropy. and PYRAF\footnote{PYRAF is a product of the Space Telescope Science Institute, which is operated by AURA for NASA.}. L.A. Cosmic \citep{vanDokkum2001} was used to remove cosmic rays and calibrated using a HgNe or Ar lamp and spectroscopic standard stars \citep{Oke1974,Massey1990}. SALT spectra were reduced using specific SALT reduction software.
\section{Results}
A few large and unusually high Galactic latitude nebulae have been recently proposed as previously unrecognized SNRs. These objects have angular diameters well in excess of the largest confirmed SNRs, such as the 8\degr \ diameter Vela SNR. If these objects were to be found to be true SNRs, they would expand the notion of how large an identifiable Galactic remnant can be and how far off the Galactic plane should SNR researchers be looking.
The proposed new SNRs include
the huge $\sim$24\degr\ diameter Antlia nebula (G275.5+18.4; \citealt{McCull2002}) and the $\sim$10\degr\ diameter radio remnant G354-33 \citep{Testori2008}. Due to limited data on it, Antlia is listed only as a possible SNR in the most recent catalogue of Galactic supernova remnants \citep{Green2019}, whereas no mention is made of G354-33 in either this 2019 SNR catalogue or in online SNR lists\footnote{http://www.mrao.cam.ac.uk/surveys/snrs/snrs.info.html
http://snrcat.physics.umanitoba.ca/index.php?}.
Below we discuss GALEX far UV (FUV) and optical properties of these two suspected SNRs, plus a third object, G249.2+24.4, which appears to be a new
SNR based both on our FUV and optical data, and X-ray and
radio data presented by \citet{Becker2021}. The presentation
order follows that of our research work.
\begin{figure*}
\begin{center}
\centerline{\includegraphics[width=18.0cm]{New_G1_Ha_coords_v7.jpg} }
\caption{Continuum subtracted SHASSA H$\alpha$ image of G354-33.
\label{G1_Ha}
}
\end{center}
\end{figure*}
\subsection{The Suspected Radio SNR G354.0-33.5}
Using a 1420 MHz linear polarization survey of the southern hemisphere sky, \citet{Testori2008} identified a large $\sim$10\degr\ depolarized shell at $\alpha$(J2000) = 20$^{\rm h}$ 25$^{\rm m}$, $\delta$(J2000) = $-47\degr$, roughly corresponding to Galactic coordinates $l$ = 353\degr \ $b = -34\degr$. This shell could also be seen in a 1.4 GHz polarization all-sky survey map \citep{Reich2009}.
Due to the object's spherical morphology and radio properties, \citet{Testori2008} suggested a SNR identification with a size suggesting a distance between 300 and 500 pc, a physical diameter of 17.4 pc $\times$ d$_{\rm 100 pc}$
and a z distance of 57.4 pc $\times$d$_{\rm 100 pc}$. \citet{Testori2008} noted that the object could be seen in 408 MHz data \citep{Haslam1982} and weakly in the SHASSA H$\alpha$ images \citep{Gaustad2001}. This object also shows up in gradients of linearly polarized synchrotron radio emission \citep{Iacob2014}.
More recently, \citet{Bracco2020} briefly commented on the presence of some GALEX FUV filaments coincident with this radio shell.
However, to our knowledge, there is no work showing this suspected SNR's FUV emissions or any associated optical emissions. Below we present large mosaics of GALEX FUV images along with a SHASSA H$\alpha$ image and compare these to the low resolution 1420 MHz radio images.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=0.49\textwidth]{New_FUV_West_coords_v1.jpg}
\includegraphics[angle=0,width=0.49\textwidth]{New_Ha_West_coords_v2.jpg}
\includegraphics[angle=0,width=0.49\textwidth]{New_FUV_NE_coords_v2.jpg}
\includegraphics[angle=0,width=0.49\textwidth]{New_Ha_NE_coords_v2.jpg}
\caption{Matched GALEX FUV and continuum subtracted SHASSA H$\alpha$ images for a bright filament along G354-33's western limb (upper panels) and northeastern limb (lower panels).
\label{G1_FUV_vs_Ha}
}
\end{center}
\end{figure*}
\subsubsection{Far UV and H$\alpha$ Images}
Figure~\ref{G1_FUV} shows a wide-field mosaic of GALEX FUV images of G354-34 smoothed by a nine point Gaussian. The image mosaic
reveals a large, coherent set of sharp UV emitting filaments arranged in a fairly circular and complete shell-like structure which is
positionally coincident with the 1420 MHz radio shell.\footnote{Because of distortions due to the large size of this emission structure, coordinates shown are only accurate to a few arc minutes.}
Multiple and overlapping filaments are common along the shell's northern and western limbs.
Because these filaments appear brightest in the GALEX FUV images compared to NUV images, we will concentrate mainly on the shell's far UV emission structure.
Although many of the FUV filaments are unresolved at GALEX's 4.6$''$ FWHM resolution, some filaments appear partially resolved in places, especially along its northernmost edge. It is not clear if this appearance is due to closely spaced multiple filaments or resolution of a single emission filament.
The strong shock-like morphology of these filaments and their arrangement in a nearly continuous shell lends support to \citet{Testori2008}'s suggestion for it to be a previously unrecognized Galactic supernova remnant, albeit with unusually large angular dimensions and distance from the Galactic plane.
\begin{figure*}
\begin{center}
\includegraphics[width=15.0cm]{Chile_Ha_DS9_trim_v1.jpg}
\includegraphics[angle=0,width=17.0cm]{FUV_Ha_O3_v4.jpg}
\caption{{\it{Top:}} CHILESCOPE H$\alpha$ + [\ion{N}{2}] image of G354-33's northwest limb revealing emission structure details lost in the low-resolution SHASSA image shown in Fig.\ 4. {\it{Bottom:}} Matched GALEX FUV and Chilescope H$\alpha$ and [\ion{O}{3}] images for the thin filaments seen in this NW region. The thin filament is clearly seen in the FUV and H$\alpha$ images but not in [\ion{O}{3}].
\label{G1_Chile}
}
\end{center}
\end{figure*}
Based on its FUV filaments, we estimate somewhat larger angular dimensions than the 10\degr \ cited by \citet{Testori2008}. Instead, we find dimensions of $\simeq$11\degr \ East-West and $\simeq$14\degr \ North-South, with a distinct indentation along its south-western limb. Its southern extent is poorly constrained in our GALEX FUV mosaic but FUV filaments extend south to at least a Declination of $-53.5\degr$.
While largely coincident in location with the 1420 MHZ radio emission map of \citet{Testori2008}, the $\sim$11\degr \ East-West diameter seen in the FUV images is centered around 20$^{\rm h}$ 16$^{\rm m}$, -45\degr\ 50$'$ ($l$ = 354.0\degr, $b$ = $-33.5$\degr). Hence, we will use the shorten name G354-33 based on the FUV images (see Table 1) which is slightly different from the G353-34 name cited by \citet {Testori2008}.
Due to its large size, many of the filamentary details visible in the full resolution FUV mosaic are lost in Figure~\ref{G1_FUV}. Consequently, we show in Figure~\ref{G1_blow-ups} blow-ups of four regions, along G354-33's northern, western, southeastern, and southwestern rims to give a better sense of the object's fine-scale UV emission structure and morphology.
Note that some individual filaments are two to three degrees in length, equivalent to the diameters of some of the largest confirmed Galactic SNRs.
A continuum subtracted SHASSA H$\alpha$ image mosaic of the G354-33 remnant is shown in
Figure~\ref{G1_Ha}. This image
reveals a thick diffuse shell of emission located within the boundaries of the FUV filaments. The H$\alpha$ emission structure is far more diffuse and broader than that seen in the GALEX FUV image with very few obvious thin filaments. We estimate an H$\alpha$ flux of around $1 - 2 \times 10^{-17}$ erg
cm$^{-2}$ s$^{-1}$ arcsec$^{-2}$ for most of the diffuse emission and about 3 times times this for the area of brighter emission along the nebula's northwestern limb.
While its H$\alpha$ emission's location matches that seen in the FUV image, some weak H$\alpha$ emission is seen to extend farther to the east and south than is readily visible in the GALEX FUV mosaic image. In general, however, the FUV filaments appear to mark the outer edges of the diffuse H$\alpha$ emission shell.
This is shown in Figure~\ref{G1_FUV_vs_Ha} where we compare FUV vs.\ H$\alpha$ images for two northeastern sections along its limb. The upper images show the presence of weak H$\alpha$ emission coincident with the bright main section of the FUV filament, with diffuse H$\alpha$ emission extending toward the east bordered by the long FUV filament. The lower panels show a section along the nebula's northeastern limb where again the UV filaments are seen to mark the extent of the diffuse H$\alpha$ emission. The SHASSA image also gives a hint of filamentary structure that resembles
that seen in the FUV image.
Higher resolution optical images give a better sense of the remnant's true emission structure.
Figure~\ref{G1_Chile} presents a higher resolution CHILISCOPE
H$\alpha$ + [\ion{N}{2}] image of a small portion of G354-33's
northwestern limb where Figure 4 revealed a particularly bright
emission patch. The upper panel of
this figure
shows this emission patch to be diffuse and situated mainly west of a $\sim45'$ long, thin H$\alpha$ filament which likely marks the western extent of the remnant in this region.
The three lower panels of Figure~\ref{G1_Chile}
compare the emission appearance of this H$\alpha$ filament in the
GALEX FUV image and in an equally deep [\ion{O}{3}] image. The similarity of the filament as seen in the FUV and H$\alpha$ images suggests G354-33's optical structure
would consist of numerous thin, unresolved filaments like that seen in the FUV mosaic image shown in Figure~\ref{G1_FUV}.
Based on the filament's absence of
[\ion{O}{3}] emission, its thin morphology and its brightness in the UV, it is likely to represent a non-radiative or a `Balmer dominated' emission filament. Given the sharpness of the other G354-33 filament's in the FUV, they are also probably Balmer dominated emission filaments marking
the location of the object's shock front and outer boundary.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=17.50cm]{WIDE_VIEW_radio_3.jpg}
\includegraphics[width=0.49\textwidth]{G1_FUV_Radio_OVERLAY.jpg}
\includegraphics[width=0.49\textwidth]{G1_Ha_radio_OVERLAY.jpg}
\caption{$Upper~Panel$ 1420 MHz Villa Elisa Stokes Q polarization intensity image covering a 60\degr $\times$ 25\degr region centered on G354-33. $Lower~ Panels:$ Contour overlays of the
Villa Elisa radio emission with the GALEX FUV (left) and SHASSA H$\alpha$ (right) images
of the G354-33 region showing excellent size and positional agreement of the 1420 MHz radio emission ring with the filamentary shell of far UV and H$\alpha$ emission filaments.
\label{Villa}
}
\end{center}
\end{figure*}
\subsubsection{Radio and X-rays}
As already noted, \citet{Testori2008} used 1420 MHz Villa Elisa radio data \citep{Testori2001} to identify the radio emission ring,
G354-33, as a possible SNR. The upper panel of Figure~\ref{Villa} shows a $60\degr \times 25\degr$ wide section of the Villa Elisa 1420 MHz polarization southern sky survey data. The emission ring appears distinct in shape and separate from other emission features along the southern edge of the Galactic plane (i.e., the right-hand side of the image).
The images in the lower panels of Figure~\ref{Villa} show the GALEX FUV image mosaic (left panel) and the SHASSA H$\alpha$ image (right panel) overlayed with the 1420 MHz Villa Elisa radio image contours. Overall, there is good agreement in both size and location for the FUV and H$\alpha$ emission shells and radio ring seen in these images. This includes the greater extent toward the southern limb area. We note that the brightest portion of the radio emission lies on the object's western limb, the side facing the Galactic plane.
Examination of ROSAT on-line data showed no obvious associated X-ray emission with this nebula. This is not surprising given its location 33.5\degr \ below the Galactic plane. Moreover, in view of the lack of any bright FUV emission nebulae along its limbs, this suggests that it lies in a region with few or none large-scale interstellar clouds which might generate significant X-ray flux.
\begin{figure*}[htp]
\centerline{\includegraphics[angle=0,width=16.0cm]{G2_FUV_v3.jpg}}
\caption{FUV intensity map of G249+24 showing an elliptical shaped shell of UV emission filaments.
\label{G2_FUV}
}
\end{figure*}
\subsection{SNR G249.7+24.7}
Using GALEX FUV mosaics, we discovered a region containing several long, thin, shock-like morphology filaments aligned mostly N-S and arranged in a roughly elliptical shape $\sim2.5 \times 4.0\degr$ in size.
These filaments appear centered at $\alpha$(J2000) = 9$^{\rm h}$ 30.5$^{\rm m}$ $\delta$(J2000) = $-16$\degr 40$'$. However, as discussed below, we believe the object to be larger and centered
$\alpha$(J2000) = 9$^{\rm h}$ 33$^{\rm m}$ $\delta$(J2000) = $-17$\degr 00$'$
which corresponds to Galactic coordinates $l$ = 249.7\degr, $b$ = +24.7\degr. We call this object G249+24 for short. Below, we present images of this object's UV and H$\alpha$ emission structure along with optical spectra of four of its filamentary regions.
\subsubsection{Far UV Mosaic Image}
Figure~\ref{G2_FUV} presents a $5.2\degr \times 5.2\degr$ mosaic of GALEX FUV images centered on G249+24's. The image shows several UV filaments along with a few smaller diffuse emission patches. Although GALEX imaging of this region is fairly incomplete, enough imaging exists to indicate G249+24's minimum size and extent\footnote{The bright feature at
9$^{\rm h}$30$^{\rm m}$, $-17$\degr 30$'$ is HD 82093, an Ap(EuSrCr) star; V = 7.08.}.
The nebula's brightest UV filaments are concentrated in its southern region where there is a bright, long gently curved filament
along its southwestern edge. This feature itself consists of several separate, closely aligned filaments. Although a smaller but a similarly bright filament caught on the western edge of a more northern GALEX image might give the impression that this filament's full length and extent is partially missing in this mosaic, this is not the case based on H$\alpha$ images (see below).
\subsubsection{H$\alpha$ Emission}
Whereas the object's FUV emission filaments are readily apparent in the GALEX image mosaic shown in Figure~\ref{G2_FUV}, its optical emissions are relatively faint. Although some of its brighter filaments are weakly visible on broad red passband Digital Sky Survey images, they are so faint and scattered over such a large area that it is not surprising that this nebula had not attracted prior attention.
The H$\alpha$ images of the MDW All-Sky survey reveal more fully G249+24's optical emissions. In Figure~\ref{G2_FUV_n_Ha}, we show a side-by-side comparison of G249+24's central emission structure in GALEX FUV and the MDW H$\alpha$ images. In general, there is a good agreement between the nebula's UV and H$\alpha$ emissions. The bright FUV and curved southwestern filament shows up strongly in H$\alpha$ as does a shorter filament a degree to its east.
Correlated UV -- H$\alpha$ emission is also seen along the upper left-hand portion of Figure 8, where diffuse and filament type of emission is seen in both.
(Note: Faint diffuse emission seen along the lower portion of the H$\alpha$ image extends several degrees to both the east and west of the region and hence does not seem to be connected to G249+24's other optical features (see also Figure~\ref{G2_slits} below).
However, little in the way of H$\alpha$ emission can be seen along the object's southeast limb, where in contrast one finds considerable FUV emission. A similar difference between FUV and H$\alpha$ fluxes is seen for the south-central area around
9$^{\rm h}$ 28$^{\rm m}$, $-18$\degr 15$'$.
Consequently, it is clear that this nebula's overall structure is more readily visible in the FUV than in even fairly deep H$\alpha$ images.
Higher resolution MDM H$\alpha$ images of selected regions of the nebula were also obtained and presented in Figure~\ref{G2_MDM_images}. Although these images were taken mainly in preparation for follow-up spectral observations, they showed a surprisingly amount filamentary detail to G249+24's optical emission than initially indicated in Figure~\ref{G2_FUV_n_Ha}.
Numerous long and delicate filaments are visible in several on these images, indicating a rich and beautiful optical shock emission structure largely hidden due to its relative faintness.
\begin{figure*}[htp]
\begin{center}
\includegraphics[angle=0,width=8.5cm]{G2_FUV_trim_v4.jpg}
\includegraphics[angle=0,width=8.5cm]{G2_MDW_trim_v4.jpg}
\caption{G249+24's FUV emissions (left) compared to its H$\alpha$ emission as seen in
MDW's All-Sky H$\alpha$ Survey images (right). \label{G2_FUV_n_Ha}
}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=12.71cm]{G2_Slit_Locations_v9.jpg}
\caption{MDW H$\alpha$ image of G249.7+24.7 showing
where low-dispersion optical spectra were taken. Also shown are the locations early type stars HD 82509 and HD 83636 which were found to exhibit high-velocity \ion{Na}{1} absorption features. The black circle (dia.\ = 4.5$\degr$) marks the approximate size of the remnant.
\label{G2_slits}
}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=8.4cm]{G2_P1_WCS_v5.jpg}
\includegraphics[angle=0,width=8.5cm]{G2_P2_WCS_v1.jpg}
\includegraphics[angle=0,width=8.5cm]{G2_P3_WCS_v1.jpg}
\includegraphics[angle=0,width=8.3cm]{G2_P4_WCS_v1.jpg}
\caption{MDM H$\alpha$ + [\ion{N}{2}] images of filaments showing locations of slit Positions 1 and 2. Slit lengths as shown are $1'$ in length.
Note the numerous long, fine filaments in these regions which are common characteristics of SNR shocks.
\label{G2_MDM_images}
}
\end{center}
\end{figure*}
\begin{figure*}[ht]
\begin{center}
\includegraphics[angle=0,width=15.0cm]{G2_spectra_plot.jpg}
\caption{Optical spectra of filamentary emission at positions in G249+24.
\label{G2_spectra}
}
\end{center}
\end{figure*}
\begin{figure*}
\centerline{\includegraphics[angle=0,width=9.0cm]{Pos1_O3_2d_v1.jpg}}
\centerline{\includegraphics[angle=0,width=9.0cm]{Pos1_Ha_2d_v1.jpg}}
\caption{Sections of the long slit spectra for Position 1. Both
[\ion{O}{3}] $\lambda$4959 and $\lambda$5007 lines are visible in the upper panel.
\label{G2_pos1}
}
\end{figure*}
\subsubsection{Low-Dispersion Optical Spectra}
Low-dispersion, exploratory slit spectra were taken at four positions (P1 through P4) in G249+24 as shown in Figures~\ref{G2_slits} and \ref{G2_MDM_images} in order to investigate the emission nature G249+24's optical filaments.
Since only the spectra taken at P1 exhibited any [\ion{O}{3}] emission, the four spectral plots shown in Figure~\ref{G2_spectra} cover only the wavelength region 6200 to 6900 \AA.
The resulting spectra reveal clear evidence for its filaments being shock heated ISM like that commonly seen in SNRs.
For example, spectra obtained at P2, P3 and P4 exhibited [\ion{S}{2}]/H$\alpha$ ratios of 1.26$\pm0.20$,
0.45$\pm0.04$, and 0.91$\pm0.15$, respectively,
thus well above the standard criteria ratio of 0.40 indicative of shock emission. Added support for shocks is the detection of [\ion{O}{1}] $\lambda$6300 emission in the spectra of both P3 and P4. The presence of [\ion{O}{1}] is a secondary
indicator for shock emission commonly observed in evolved SNRs \citep{Fesen1985,Kop2020}.
The [\ion{S}{2}]
$\lambda$6716/$\lambda$6731 line ratio can be used to estimate the electron
density in the S$^{+}$ recombination zone and is nearly independent of electron
temperature \citep{Osterbrock2006}. The $\lambda$6716/$\lambda$6731 ratio was found to be near the low density limit of 1.43 indicating n$_{\rm e}$ $<$100 cm$^{-3}$; $1.48\pm0.04$ for P2, and $1.42\pm0.03$ for P4.
A bit surprisingly, this ratio for P3 was observed to be $\simeq$1.15 indicating an n$_{\rm e}$ density much higher than the other locations of around 500 cm$^{-1}$. Interestingly, the spectrum seen at P3 also showed much weaker [\ion{N}{2}] $\lambda$6583 line emission compared to that seen at P2 and P4.
The spectrum taken at P1 deserves special attention due to significant line emission variations with respect to distance along the E-W aligned slit. This is shown in Figure~\ref{G2_pos1}, where line emission can be seen to start at the location of the mainly N-S aligned filament extending many arcseconds to the east (left).
The emission line ratio of H$\alpha$/[\ion{O}{3}] can be seen to vary considerably down stream from the shock front.
Figure~\ref{G2_pos1} also shows the 2D image of the background subtracted P1 spectrum for the areas around H$\alpha$, and [\ion{O}{3}]
$\lambda$4959, $\lambda$5007 and H$\beta$. This figure shows a clear separation of the filament's Balmer dominated emission followed by strong [\ion{O}{3}] emission resulting in large [\ion{O}{3}] $\lambda$5007/H$\alpha$ ratios. The arrows indicate the section where the P1 spectrum shown in Figure~\ref{G2_MDM_images} was taken.
Behind the leading edge of the shock front marked by start of line emissions, H$\alpha$ emission becomes strong with little or no emission in other lines. The intensity of H$\alpha$ then briefly drops in strength within a distance of a few arcseconds, then increases for a short distance followed by a long, gradual decline.
The [\ion{O}{3}] emission is barely detectable at the shock front where the H$\alpha$ starts but substantially increases some 10-20 arcseconds behind the shock front, to where it dominates the downstream spectrum. This is consistent with the picture of an extended post-shock cooling zone. The presence of such strong [\ion{O}{3}] emission behind the shock front suggests shock velocities at least 100 km s$^{-1}$ are present in certain filaments. In contrast, spectra at P2, P3, and P4 showed no appreciable [\ion{O}{3}] emission indicating much lower velocity shocks, less than about 70 km s$^{-1}$.
\begin{figure*}[ht]
\begin{center}
\includegraphics[angle=0,width=15.cm]{HD82509_deblend_plot.jpg}
\includegraphics[angle=0,width=15.0cm]{HD83636_deblend_plot.jpg}
\caption{Moderate dispersion CCDS spectra of HD~82509 (top) and HD 83636 (bottom) showing deblending of the \ion{Na}{1} lines into high-velocity blue and redshifted absorption components associated with G249+24. See Table 3.
\label{CCDS_spectra}
}
\end{center}
\end{figure*}
Similar 2D emission structures have been observed in the high latitude Galactic halo SNRs G70.0-21.5 \citep{Raymond2020} and G107.0+9.0 \citep{Fesen2020}. They can be interpreted as emission from a $\sim100$ km s$^{-1}$ shock in partially neutral gas. Neutral hydrogen atoms swept up by the shock are quickly ionized, but before that happens, some of them are excited to produce H$\alpha$. This emission region is very thin. It is followed by a thicker ionization zone where \ion{O}{2} is ionized to \ion{O}{3}. As the gas cools, the [\ion{O}{3}] fades and H$\alpha$ brightens. In the case of G107.0+9.0, the 10$"$ gap between H$\alpha$ and [\ion{O}{3}] and the $\sim$ 10$"$ wide diffuse patch of [\ion{O}{3}] emission are well matched by models of 100 km s$^{-1}$ shocks in a gas of density 0.1 cm$^{-3}$ at a distance of 1 kpc \citep{Fesen2020}. The wider patch of [\ion{O}{3}] diffuse emission at Position 1 of G249+24 suggests a longer cooling length, as would occur in a slightly faster shock, perhaps $\sim$120 km s$^{-1}$. These structures are spatially resolved thanks to the low density of the Galactic halo and the relatively small distances to these SNRs.
The flux level of H$\beta$ was either undetected or too faint at P2, P3, and P4 to accurately measure an H$\beta$/H$\alpha$ ratio which could be used to estimate extinction. Only the spectrum P1 showed a weak but measurable level of H$\beta$ flux. There the measured H$\alpha$/H$\beta$ ratio of $2.85\pm 0.20$ suggests a very low amount of extinction. Adopting a theoretical H$\alpha$/H$\beta$ ratio of 2.87 for 10$^{4}$ K we estimate a foreground $E(B-V) < 0.07$. However, in view of the weakness of the H$\beta$ detection, this estimate comes with a large uncertainty. Nonetheless, such a low interstellar extinction is not unexpected due to the remnant's high Galactic latitude of $24.4\degr$.
If this low extinction estimate is correct, then the lack of [\ion{O}{3}] emission at P2, P3, and P4 suggests a fairly low shock velocity, around 70 km s$^{-1}$ or less, in contrast to that indicated at P1. Consequently, it appears that a variety of shock speeds, from $<$70 to $\geq$120 km s$^{-1}$ appear to be present throughout the nebula's structure. This conclusion is supported by \ion{Na}{1} absorptions described below.
\begin{deluxetable*}{lcccccccccc}[ht]
\tablecolumns{10}
\tablecaption{Measured Interstellar Na I Velocity Components and Equivalent Widths}
\tablewidth{0pt}
\tablehead{
\colhead{Star} & \colhead{Distance} & \colhead{\ion{Na}{1}} & \colhead{$\mathrm{V}_0$} & \colhead{$\mathrm{EW}_0$} &
\colhead{$\mathrm{V_{R1}}$} & \colhead{$\mathrm{EW_{R1}}$} &
\colhead{$\mathrm{V_{B1}}$} & \colhead{$\mathrm{EW_{B1}}$} &
\colhead{$\mathrm{V_{B2}}$} & \colhead{$\mathrm{EW_{B2}}$} \\
\colhead{ID} & \colhead{(pc)} &\colhead{Line} & \colhead{(km s$^{-1}$)} & \colhead{(\AA)} &
\colhead{(km s$^{-1}$)} & \colhead{(\AA)} &
\colhead{(km s$^{-1}$)} & \colhead{(\AA)} &
\colhead{(km s$^{-1}$)} & \colhead{(\AA)} }
\startdata
HD 82509 & $825\pm25$ & $\lambda$5889.95 & $-3$ & 0.28 & +57 & 0.08 & $-63$ & 0.05 & \nodata & \nodata \\
& & $\lambda$5895.92 & $-2$ & 0.20 & +59 & 0.08 & $-56$ & 0.04 & \nodata & \nodata \\
HD 83636 & $386\pm4$ & $\lambda$5889.95 & $-9$ & 0.17 & +55 & 0.05 & $-73$ & 0.05 & $-122$ & 0.02 \\
& & $\lambda$5895.92 & $-6$ & 0.12 & +70 & 0.04 & $-70$ & 0.04 & $-123$ & 0.01 \\
\enddata
\label{CCDS_table}
\end{deluxetable*}
\bigskip
\subsubsection{High-Velocity Na I Absorption Lines}
Moderate-dispersion optical spectra (R $\simeq$ 7200) of several stars
with projected positions within the estimated radio boundary of
G249+24 (see \citet{Becker2021} and Section 3.2.5 below) were
obtained in a search for high-velocity Na I interstellar absorption components, which could yield both a direct measurement of its expansion velocity plus maximum remnant distance information. Of the stars examined, two showed evidence for the presence of high-velocity ($>$ 40 km s$^{-1}$) \ion{Na}{1} absorption features.
The two stars were: HD~82509 (B = 10.29, V = 10.4; A2 IV) with an estimated Gaia Early Release DR3 (EDR3) distance estimate of $825\pm25$ pc, and HD~83636 (B = 9.74, V = 9.57: A2 IV/V) with an EDR3 distance estimate of $386\pm4$ pc. Figure~\ref{G2_slits} shows the locations of these stars relative to the object's FUV and optical emissions.
Reduced and normalized spectra of both stars for the wavelength region around the \ion{Na}{1} D lines at $\lambda$5889.95 and $\lambda$5895.92 indicating the presence of high-velocity \ion{Na}{1} absorptions are shown in Figure~\ref{CCDS_spectra}. Using IRAF deblending software and allowing initially chosen absorption component wavelengths to vary but keeping the Gaussian line profiles and FWHMs all the same, we found red and blue absorption features in the observed \ion{Na}{1} profiles in both stars with LSR velocities around +55 to $-73$ km s$^{-1}$. In addition, to achieve a good fit to the Na absorption in the spectrum of HD~83636, an additional blue-shifted components with a radial velocity of around $-123$ km s$^{-1}$ was needed (see Table 3).
We note that the lower velocity components are consistent with the observed spectra seen for slit positions P2, P3, and P4 where no [\ion{O}{3}] was seen indicating shock velocities below $\simeq$ 70 km s$^{-1}$, while the presence of the higher velocity 123 km s$^{-1}$ component is consistent with the strong [\ion{O}{3}] seen at slit position P1.
The detection of these high-velocity \ion{Na}{1} absorptions in these two stars can be used to estimate a maximum distance to G249+24. HD~82509's Gaia EDR3 parallax derived distance sets a distance limit of less than 825 pc, while HD~83636 sets a maximum distance of less than 390 pc. Assuming an angular diameter of 4.5\degr (see below), this shorter maximum 390 pc distance implies a maximum physical size for G249+24 of $\sim$30 pc, similar to that estimated for the Cygnus Loop SNR \citep{Fesen2018}. \\
\subsubsection{Associated Radio and X-Ray Emissions}
Examination of several on-line radio maps including the NRAO VLA Sky Survey (NVSS; \citealt{Condon1998}) did not show obvious radio emission in the G249+24's region. However, the 1420 MHz Bonn survey \citep{Reich1982} did show some emission along G249+24's FUV and optical filaments near $\alpha$(J2000) = 9$^{\rm h}$ 27.5$^{\rm m}$, $\delta$(J2000) = $-16$\degr 10$'$.
\begin{figure*}[htp]
\begin{center}
\includegraphics[width=0.55\textwidth]{G2_408Mhz_v3.jpg}
\caption{A 408 MHz intensity image of the G249+24 region from the \citet{Haslam1982} All-Sky survey.
\label{G2_radio}
}
\end{center}
\end{figure*}
The 408 MHz all-sky survey \citep{Haslam1981,Haslam1982} also showed enhanced emission in a roughly spherical shape approximately 4.5\degr \ in diameter centered at 9$^{\rm h}$ 32.7$^{\rm m}$, $-16$\degr 45$'$, a position that is nearly the same indicated by FUV and optical images (see Fig.~\ref{G2_radio}).
A seemingly unrelated broad, vertical emission feature borders the emission shell
on its eastern side and complicates the assessment of this emission. Assuming
this emission shell is associated with the nebula,
it would suggest a far more spherical SNR than indicated by either the GALEX FUV
or MDW H$\alpha$ images.
Taking the object's largest N-S angular dimensions based of its
FUV filaments (see Figs.\ 8 and 9) and requiring that both stars, HD~82509 and HD~83636,
must lie inside the remnant's shell, we estimate G249+24 is $\simeq4.5\degr$ in diameter assuming a circular morphology. This size estimate is in
excellent agreement with that found by \citet{Becker2021}
using eRosita X-ray images and archival Parks All-Sky Survey radio data.
\citet{Becker2021} report the detection of diffuse X-ray emission filling almost the entire remnant with a gas temperture of order kT = 0.1 keV, along with radio data indicating a spectral index of $-0.69$ fully consistent with its SNR nature.
\subsection{The Exceptionally Large Suspected Antlia SNR}
\citet{McCull2002} reported a discovery of a
very large $\sim$24\degr \ diameter H$\alpha$ emission shell with interior ROSAT 0.25 keV X-ray emission located in the southern hemisphere and at a high Galactic latitude of +19\degr. They proposed it to be a previously unrecognized SNR and named it Antlia for the constellation it is found in. \citet{McCull2002} suggested it was an extremely old remnant with an estimated age of $\sim$1 Myr and located relatively nearby with an uncertain distance of between 60 pc and 340 pc.
The Antlia remnant does appears in the catalogue of high energy SNRs \citep{Safi2012} but not in the
\citet{Green2019} catalogue of confirmed Galactic SNRs,
which cited the need for further observations to confirm its nature and parameters. Except for a 2007 far UV study finding weak coincident \ion{C}{3} $\lambda$977 and \ion{C}{4} $\lambda\lambda$1548,1551 line emissions from the object's interior \citep{Shinn2007}, and a brief AAS abstract by
\citet{Orchard2015} who reported Wisconsin H-alpha Mapper data indicating
[\ion{S}{2}] emission from the Antlia Nebula consistent with a SNR interpretation,
there has been no detailed follow-up investigation of this exceptionally large suspected SNR. This situation plus its location at an unusually high Galactic latitude led us to include it in our high Galactic latitude GALEX FUV investigation of SNRs.
\subsubsection{GALEX FUV and MDW H$\alpha$ Images}
In the top panel of Figure~\ref{G3_SHASSA}, we present a mosaic of SHASSA H$\alpha$ images of the Antlia remnant at a higher resolution than the VTSS H$\alpha$ image \citep{Fink2003} which led \citet{McCull2002} to discover it. It shows a $\sim 20\degr \times 26\degr$ H$\alpha$ emission shell with a well determined boundary roughly centered at
$\alpha$(J2000) = 10$^{\rm h}$ 38$^{\rm m}$, $\delta$(J2000) = $-37\degr$ 18$'$ corresponding to Galactic coordinates $l = 275.5\degr$ $b = +18.4\degr$. (Note: These Galactic coordinates differ slightly from those given by \citealt{McCull2002}.)
A mosaic of FUV GALEX images covering just the remnant's northern limb is shown in the lower panel of Figure 13. This image reveals a long and nearly continuous set of thin, bright FUV emission filaments and filament clusters extending approximately 20 degrees along the remnant's northeastern limb.
The line of FUV filaments starts at RA = 10$^{\rm h}$ 10$^{\rm m}$ and Dec = $-24\degr$ and extends southward to the bottom left of the figure at
RA = 11$^{\rm h}$ 40$^{\rm m}$ and Dec = $-32\degr$.
Because the morphology of Antlia's FUV filaments is similar to that seen in the proposed SNRs G354-33 and G249+24, we examined MDW H$\alpha$ images of the FUV filaments and adjacent regions above the $-32\degr$ Declination limit of the MDW survey. Overall, we found excellent positional agreement between FUV and H$\alpha$ filaments.
To illustrate this agreement
Figure~\ref{G3_FUV_vs_MDW} shows a comparison of GALEX FUV and MDW survey H$\alpha$ images for three sections along the Antlia's northeastern boundary. While we find there is a close correlation of the outer FUV filaments with many H$\alpha$ filaments, there are also considerable differences in terms of the nebula's internal emissions. That is, far more H$\alpha$ emission is seen `behind', i.e., to the west, of the sharp, outlying filaments than seen in the FUV images. This is most striking in the lower panel images of Figure~\ref{G3_SHASSA} for the remnant's southeastern limb region.
While the SHASSA image shows considerable internal diffuse and filamentary emission, the brightest H$\alpha$ emissions are found along Antlia's northwestern, western and southeast limbs.
\begin{figure*}[htp]
\begin{center}
\centerline{\includegraphics[angle=0,width=17.12cm]{G3_SHASSA_v9.jpg} }
\centerline{\includegraphics[angle=0,width=17.12cm]{G3_FUV_North4.jpg}}
\caption{$Top:$ Continuum subtracted SHASSA H$\alpha$ image mosaic of the Antlia remnant. $Bottom:$ GALEX FUV mosaic showing a nearly unbroken line of UV emission filaments
along Antlia's northern and eastern boundary. \label{G3_SHASSA}
}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=17.0cm]{G3_Fig17_whole_new.jpg}
\caption{Comparison of GALEX FUV emission vs.\ MDW H$\alpha$ image mosaic of the northern, eastern, and southeastern portions of the Antlia SNR showing a close agreement of UV and H$\alpha$ emission filaments along its eastern boundary.
There is considerable H$\alpha$ emission but little FUV
emission in the nebula's interior. North is up, East to the left. \label{G3_FUV_vs_MDW}
}
\end{center}
\end{figure*}
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=16.0cm]{G3_SHASAA_n_Slits_NEW.jpg}
\includegraphics[angle=0,width=0.49\textwidth]{G3_spectra1.jpg}
\includegraphics[angle=0,width=0.49\textwidth]{G3_spectra2.jpg}
\caption{{\it{Top:}} Continuum subtracted SHASSA H$\alpha$ image mosaic of the Antlia remnant showing the locations of seven positions where SALT spectra were obtained.
{\it {Bottom:}} Observed optical spectra at these seven positions.
\label{G3_slits_spectra}
}
\end{center}
\end{figure*}
\begin{deluxetable*}{lccccccc}[htp]
\tiny
\tablecolumns{8}
\tablecaption{Observed Emission Line Fluxes for Antlia Filaments }
\tablehead{\colhead{Emission Line} &
\multicolumn{7}{c}{\underline{~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~Filament Position ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~}} \\
\colhead{(\AA)} & \colhead{P1} & \colhead{P2} & \colhead{P3} & \colhead{P4} & \colhead{P6} & \colhead{P7} & \colhead{P8} }
\startdata
H$\beta$ 4861 & 100 & 100 & 100 & 100 & 100 & 100 & 100 \\
$[$\ion{O}{3}$]$ 5007 & $< 10~~ $ & $< 10~~ $ & $< 10~~$ & $< 10~~$ & $< 10~~$ & $< 10~~$ & $< 10~~ $ \\
$[$\ion{N}{1}$]$ 5200 & \nodata & 23 & 29 & \nodata & \nodata & 20 & \nodata \\
$[$\ion{O}{1}$]$ 6300 & \nodata & 93 & 98 & \nodata & \nodata & 101 & 122 \\
$[$\ion{N}{2}$]$ 6548 & 57 & 89 & 97 & \nodata & \nodata & 137 & 181 \\
H$\alpha$ 6563 & 430 & 300 & 339 & 379 & 368 & 304 & 365 \\
$[$\ion{N}{2}$]$ 6583 & 167 & 247 & 275 & \nodata & \nodata & 403 & 542 \\
$[$\ion{S}{2}$]$ 6716 & 146 & 239 & 247 & \nodata & \nodata & 406 & 442 \\
$[$\ion{S}{2}$]$ 6731 & 91 & 167 & 184 & \nodata & \nodata & 283 & 307 \\
F($[$\ion{S}{2}$]$)/F(H$\alpha$) & 0.55 & 1.35 & 1.27 & \nodata & \nodata & 2.27 & 2.05 \\
$[$\ion{S}{2}$]$ 6716/6731 & 1.60 & 1.43 & 1.34 & \nodata & \nodata & 1.43 & 1.44 \\
$E(B-V)$ & 0.35 & 0.00 & 0.12 & 0.23 & 0.20 & 0.01 & 0.19 \\
F(H$\alpha$) erg cm$^{-2}$ s$^{-1}$ & 2.74E-15& 1.29E-14& 4.75E-15& 1.91E-15 & 2.06E-15 &2.16E-15& 6.06E-15 \\
\enddata
\label{Table_2}
\end{deluxetable*}
\subsubsection{Optical Spectra}
If \citet{McCull2002} is correct in their assessment that the Antlia H$\alpha$ nebula is a SNR, as seemingly supported by the GALEX FUV and our H$\alpha$ images described above, the remnant would then be both many times larger than any confirmed SNR and be located at an unusually high Galactic latitude. On the other hand, considering its size and location immediately adjacent to the enormous Gum Nebula with its extended emission, outer `filamentary wisps', and numerous H~I wind blown bubbles \citep{Brandt1976d,Chanot1983,Purcell2015}, there is some uncertainty about its true origin.
Optical spectra are a powerful tool for confirming the presence of shock emission and hence can be used to investigate the SNR origin of optical nebulosities.
Consequently, we obtained long-slit, low-dispersion spectra of seven filaments and emission clumps distributed across the whole of the Antlia nebula to explore the nature of its optical emissions.
Figure~\ref{G3_slits_spectra} shows the location of the seven filaments observed in the Antlia remnant, along with the resulting spectra. Table 2 lists the relative emission line strengths uncorrected for extinction. Listed relative line strengths are believed accurate to 10\%.
Based on the observed H$\alpha$/H$\beta$ ratios of 3.0 to 4.30, we find a variable degree of extinction across the Antlia nebula,
with $E(B - V)$ values ranging from 0.0 to 0.35
assuming an intrinsic ratio of 3.0 and an R value of 3.1. Here we have chosen an H$\alpha$/H$\beta$ value greater than the
theoretical value of 2.87 for 10$^{4}$ K due to the likelihood of significant
collisional excitation of the n = 3 level at postshock temperatures seen in SNRs. Such a range of extinction values is not unexpected in view of the nebula's large dimensions.
These optical spectra show clear evidence that the seven filaments observed distributed across the whole of the Antlia nebula exhibit line ratios indicative of shock emissions. In fact, emissions at Positions 4 and 6, located along Antlia's southeastern limb display non-radiative, pure Balmer line emission.\footnote{Note: The [\ion{O}{1}] emission seen in the spectrum for Position 4 is residual imperfect [\ion{O}{1}] sky emission.} Such shock spectra are usually seen in situations where the shock velocities are quite high ($>$ 500 km s$^{-1}$) as seen in young Type Ia SNRs like Tycho and SN~1006. However, similar Balmer dominated spectra has also been seen in much older remnants (G107.+9.0; \citealt{Fesen2020}). Relatively slow shocks can remain non-radiative if the ambient density is sufficiently low, like that expected in the Galactic halo. This requires that the cooling time of the shocked gas should be long enough to prevent the forbidden lines from becoming bright. Scaling the cooling times of \cite{Hartigan1987} with $n_0^{-1}$, the non-radiative shocks at the edge of a 100,000 year old remnant in a density of 0.1 $\rm cm^{-3}$ would be at least 170 $\rm km~s^{-1}$.
The other five filaments show optical spectra commonly seen in evolved SNRs like the Cygnus Loop and IC~443. In all these cases,
the I([\ion{S}{2}])/I(H$\alpha$) ratio exceeds the 0.4 criterion for shock emission. Indeed, filaments 7 and 8 display ratios above 2, which are among the highest values observed in SNRs. Interestingly, these five filaments lack appreciable [\ion{O}{3}] $\lambda\lambda$4959,5007 emission. Weak or absent [\ion{O}{3}] emission indicates shock velocities less than about 70 km s$^{-1}$. The high H$\alpha$/H$\beta$ ratio in Position 1 of the Antlia remnant suggests a substantial contribution of collisional excitation at fairly low temperatures to the Balmer lines, and this is borne out by the low [N II] and [S II] to H$\alpha$ ratios. That suggests shock speeds around 70 km s$^{-1}$.
Given these slow shock spectra along with pure Balmer emission spectra along the remnant's leading edges, there must be a considerable range of shock velocities present in the nebula. This is not too surprising given the large size of Antlia and a nearly 18\degr \ range in Galactic latitude.
\section{Discussion}
Based on the data presented above,
we find that two suspected, high latitude SNRs, namely G354-33 and Antlia, are likely bona fide SNRs. Moreover, their UV emission filaments appear best at marking these objects' forward shock front locations.
Below we discuss supporting evidence for our conclusions, followed by estimates regarding their physical dimensions and general properties.
We then briefly comment of the usefulness of far UV imaging for SNR detection particularly in areas far off the Galactic plane.
\subsection{G354.0-33.5}
There are several observations that support the
SNR identification of this large FUV, H$\alpha$, and radio emission nebula. At its Galactic latitude of $-33.5\degr$, there are no bright and nearby early type stars near the center of this UV shell that might have generated its UV emission shell through stellar winds. There is also no known nearby recurrent nova projected inside, and the emission shell is at least two orders of magnitude larger and does not have a similar appearance to known nova shells.
Then there is the large 1420 MHz radio emission shell roughly centered on and lying entirely within the borders of the FUV emission filaments. The FUV filaments also exhibit a shock-like morphology such as commonly seen in Galactic SNRs.
Furthermore, its nearly continuous ring of thin filaments surrounds a broad, diffuse H$\alpha$ emission shell. In summary, its shock-like filamentary appearance and the positional agreement between UV, H$\alpha$ and radio emissions for such a large object located far off the Galactic plane leaves few viable options other than a SNR origin.
While we have not yet obtained optical spectra of its H$\alpha$ and UV filaments, based on
the morphology of the H$\alpha$ emission filament along its northwestern limb and the lack of [\ion{O}{3}] emission, we suspect this remnant may resemble that of nonradiative filaments seen in a few remnants such as those that surround the northern and eastern limb of the Cygnus Loop
\cite{Fesen1982,Raymond1983,Hester1994,Sankrit2000,Medina2014}.
A nonradiative or `Balmer dominated' shock is one that heats plasma to a relatively high temperature and has encountered the plasma either so recently or is of such a low density that the gas has had insufficient time to radiatively cool and present optical line emissions. Such filaments are UV bright \citep{Shull1979,Raymond2003} a property consistent with GALEX FUV images for this object. Although the confirming test for the presence of nonradiative filaments in this remnant is spectra, the deep H$\alpha$ and
[\ion{O}{3}] images presented in Figure~\ref{G1_Chile} supports this conclusion.
Centered at a Galactic latitude around $-33.5\degr$, G354-33 would rank as the highest Galactic latitude SNR found yet. In addition, if not for the Antlia remnant, this object would also be among the largest angular dimensions among the roughly 300 SNRs in the \citet{Green2019} catalogue. Because of its $\sim11\degr \times 14\degr$ size and the fact that the largest known SNRs have physical dimensions $\sim$100 pc, we can make some crude estimates of its distance and properties.
If its diameter is taken to be 100 pc, then its distance is around 400 pc. Based on the presence of its stronger FUV than NUV emission, its shock velocity is likely greater than 100 km s$^{-1}$ but probably less than 150 km s$^{-1}$, according to Figure 4 of \citet{Bracco2020}. That shock speed is too high for this remnant to be in the snowplow phase of SNR evolution, but the presence of radiative shock waves around much of the rim indicates that it is no longer in the Sedov phase, at least in those places. That suggests that it is in the pressure-driven shell phase, when the recently shocked gas has cooled, but there is still relatively hot, high-pressure gas in the interior.
From the calculations of \cite{Cioffi1988}, this velocity range and a diameter of around 90 pc would be consistent with an age $\simeq$ 10$^{5}$ years and a preshock density of 0.1 $\rm cm^{-3}$ (see Fig.\ 14 of \citealt{Fesen2020}). One might expect detectable X-ray emission during the pressure-driven shell phase, but \citet{Shelton1998} has discussed halo SNRs in which the interior gas is too cool to produce X-rays, but still rich in high ionization states such as \ion{O}{6}.
\subsection{G249.7+24.7}
GALEX FUV, SHASSA, and MDM images show it displays
a highly filamentary morphology like that commonly seen in
SNRs. Moreover, optical spectra of its filaments reveal
emission-line ratios consistent with the presence of shocks.
The discovery of its X-ray emission along with associated radio
emission and archival Parks radio data \citep{Becker2021}
leave little doubt that this object is a new SNR.
With an angular dimension of 4.5\degr, the G249+24 remnant is
among the largest of known SNRs. Situated at Galactic latitude of over 24 degrees, G249+24 also lies farther from the Milky Way's plane than any other confirmed SNR -- that is, if it were not for G354-33.
The detection of high-velocity \ion{Na}{1} absorptions in the
spectrum of the star HD~83636 which has a Gaia EDR3 estimated
distance of 386 pc, sets a maximum limit on G249+24's distance
of less than 400 pc. This is a bit less than the lower distance range of 400--500 pc estimated by \citet{Becker2021}.
If high-velocity \ion{Na}{1} absorption lines are confimed by subsequent higher resolution echelle spectra, then the G249+24 remnant would then also rank among the closest SNRs known and with an especially robust distance limit.
The $\simeq$400 pc upper limit on the distance implies an upper limit on its diameter of about 30 pc. It also provides an estimate of the ambient density. \citet{Fesen2020} used the separation between the sharp H$\alpha$ filaments of G107.0+9.0 and the diffuse [\ion{O}{3}] emission to determine that remnant's
preshock density. The angular separation between the H$\alpha$ and [\ion{O}{3}] seen in Figure~\ref{G2_pos1} is similar to that seen in G107.0+9.0,
but here the distance upper limit is 2.5 times smaller than the 1 kpc distance adopted for G107.0+9.0. Therefore, the preshock density is at least 0.25 cm$^{-3}$.
Because the emission from G249+24 is so faint, the density cannot be much higher than 0.25 cm$^{-3}$, indicating that the distance is close to the upper limit. This is in accord with its morphology, which suggests that G247+24 is significantly above the Galactic disk and interacting with it on its southern side, indicating that it is more than 100 pc from the plane, and therefore more than 250 pc distant.
The pure Balmer line emission at P1 indicates the that remnant is still in the Sedov phase there, while the emission and absorption spectra indicate shock speeds $\sim$60 - 100 km s$^{-1}$ in most of the rest of the remnant. The faint X-ray emission and the 0.1 keV temperature found by \citet{Becker2021} also indicate that the remnant has left the Sedov phase and entered the pressure-driven shell phase. Assuming that the cooling occurred recently, when the shock speed fell below 400 $\rm km~s^{-1}$, we can combine the Sedov equations for SNR radius and shock speed to estimate and age around 15,000 years for a $10^{51}$ erg explosion. Thus the parameters of G247+24 are similar to those of the Cygnus Loop \citep{Fesen2018} but is much fainter because it lacks the dense clouds that make parts of the Cygnus Loop bright at optical and X-ray wavelengths.
\subsection{The Nature of the Antlia Nebula}
\citet{McCull2002} claimed this extraordinarily large emission nebula was a likely SNR based on its appearance on the deep H$\alpha$ image of the VTSS survey and on the presence of diffuse soft X-ray emission in its interior. However, this conclusion does not appear to be widely accepted, as measured by the remnant having attracted little subsequent attention.
This situation might in part be due to a reluctance by SNR researchers to accept its huge $20\degr \times 26\degr$ angular size, more than 3-5 times larger than the largest known confirmed SNRs, plus its location so close to the even larger 36\degr diameter Gum Nebula with its complex of large emission shells and wind-blown bubbles plus the embedded Vela SNR.
Consequently, except for a far UV study by \citet{Shinn2007} and a brief AAS abstract by \citet{Orchard2015}, the Antlia remnant has not be studied in any detail, leaving open its properties and nature.
Our GALEX FUV mosaics show a well-defined shell in H$\alpha$ with many individual and overlapping filaments that strongly resemble ISM shocks. In addition, the locations of sharp UV filaments along the boundary of the object's H$\alpha$ emission are consistent with a SNR where such UV filaments mark the location of the remnant's shock front.
In addition, results of our seven optical spectra of Antlia's filaments clearlu indicate shock emissions throughout the nebula and hence a SNR. These spectra include two textbook cases of non-radiative Balmer dominated spectra (Positions 4 \& 6), plus several other filaments exhibiting high [\ion{S}{2}]/H$\alpha$ line ratios well above the 0.4 value distinguishing shocked from photoionized nebulae. We conclude that if the Antlia remnant was not so large, it would be an easy case for supernova remnant classification.
\begin{figure*}[hbp]
\begin{center}
\includegraphics[angle=0,width=8.5cm]{VTTS_figure_v4.jpg}
\includegraphics[angle=0,width=8.0cm]{G3_SHASSA_Gum_v4.jpg}
\caption{$Left:$ VTSS H$\alpha$ image of the Galactic plane around the Gum Nebula with the location of the Antlia SNR marked. $Right:$ The continuum subtracted SHASSA H$\alpha$ image of the Antlia remnant showing the presence of bright filaments at the overlapping Antlia and Gum Nebula region suggestive of a physical interaction.
\label{Gum}
}
\end{center}
\end{figure*}
\begin{figure*}[ht]
\begin{center}
\includegraphics[angle=0,width=18.0cm]{Ludgate_color_composite_v9.jpg}
\caption{Wide-field images of the northwestern Antlia--Gum interaction
(approximate center: 9$^{\rm h}$ 42$^{\rm m}$, $-35\degr$ 47$'$).
$Left:$ Composite of H$\alpha$, [\ion{O}{3}], and [\ion{S}{2} images.
$Center:$ Composite of H$\alpha$ and [\ion{S}{2} images.
$Right:$ Ratio of H$\alpha$ and [\ion{S}{2} images.
\label{Ludgate}
}
\end{center}
\end{figure*}
If the FUV and optical images and spectra presented here were not enough to confirm the Antlia nebula as a SNR, the recently released eRosita All-Sky soft X-ray 0.3--2.3 keV image reveals a large emission shell at the site of the Antlia remnant \citep{Predehl2020,Becker2021}. Both the size and morphology of this X-ray bright limb shell matches that of the remnant's H$\alpha$ as shown in Figure~\ref{G3_SHASSA}. Although a full analysis of Antlia's X-ray emission properties is needed to assess its properties, the positional and morphological similaritiy of this X-ray shell to its FUV and H$\alpha$ emissions are strong additional indicators for a SNR classification.
The physical size of the Antlia SNR is uncertain due to its unknown distance. \citet{McCull2002} estimated a wide range of possible distances, from 60 pc to 320 pc, and believed the remnant to be extremely old, at least 1 Myr, and hence in the final snowplow phase of SNR evolution.
However, our optical spectra do not support such an old object or a distance less than 200 pc.
A 1 Myr old SNR is expected to have a a very low expansion velocity of around 10 to 20 km s$^{-1}$ \citep{Chevalier1977}.
In contrast, our optical spectra show line emissions indicating shock velocities between 50 and 150 km s$^{-1}$, making it far younger ($< 10^{5}$ yr) remnant in the pressure-driven shell phase.
Moreover, there are other data supporting a much younger remnant with relatively high-velocity gas inside.
Before the Antlia nebula was discovered, \citet{Bajaja1989} reported high-velocity clouds in the remnant's direction, with
\citet{Penprase1992} reporting high-velocity \ion{Ca}{2} absorption lines in the spectrum of the B9/A0 III/IV star HD 93721 which lies in the direction of the Antlia SNR (see Fig.\ 15). This star has a estimated Gaia Early Data Release 3 (EDR3) distance of $512\pm7$ pc. \citet{Penprase1992} found at least 10 \ion{Ca}{2} absorption components with v$_{\rm LSR}$ ranging from $-65$ km s$^{-1}$ to $+75$ km s$^{-1}$. Because a second star, HD 94724 (d = 210 pc) lying in the same general direction did not show any high-velocity components, they concluded that the high-velocity absorbing cloud's distance was between 200 and 500 pc. However, because not all stars behind a SNR display high-velocity absorption lines, a lack of high-velocity components does not provide a robust minimum distance estimate.
\begin{figure*}
\begin{center}
\includegraphics[angle=0,width=8.0cm]{G4_MDW.jpg}
\includegraphics[angle=0,width=8.0cm]{G4_FUV.jpg}
\caption{Comparison of MDW H$\alpha$ and GALEX FUV images of the high-latitude remnant G70.0-21.5.
North is up, East to the left.
\label{G70}
}
\end{center}
\end{figure*}
An alternative means of estimating the distance to the Antlia remnant is its apparent collision with the Gum Nebula \citep{Gum1952,Gum1955,Brandt1971,Sivan1974,Chanot1983}.
Although the positional coincidence or abutment of the remnant's southwestern rim with the Gum Nebula's bright northeastern emission cloud has been noted in some studies of the Gum Nebula \citep{Shinn2007,Iacob2014,Purcell2015}, no proposal has been made about there being evidence for a physical interaction between the two nebulae.
Nonetheless, the SHASSA imaging of this region strongly indicates such a physical collision has occurred between the Antlia remnant and the Gum Nebula. The left panel of Figure~\ref{Gum} presents a section of the VTSS H$\alpha$ image of the Galactic plane showing the Antlia emission shell, similar to the VTSS image presented in the \citet{McCull2002} Antlia discovery paper. This shows Antlia's position relative to the Gum Nebula. The remnant's bright southern limb coincides with the northern rim of the Gum Nebula.
The right panel of Figure~\ref{Gum} shows a low contrast version our SHASSA image of Antlia. An unusual line of bright filaments nearly 10\degr in length can be seen in the projected overlap region of both shells (see also top panel of Fig.~\ref{G3_slits_spectra}). The simplest explanation for such bright filaments situated only in the exact Antlia/Gum overlap region is that the Antlia remnant has collided with the outer northeastern rim of the Gum Nebula.
Support for an Antlia--Gum collision is shown in Figure~\ref{Ludgate} where wide-field
H$\alpha$ (+ \ion{N}{2}]), [\ion{O}{3}], and [\ion{S}{2}] images are presented for the northern section of the Antlia--Gum Nebula overlap region. The left panel is a color composite image revealing [\ion{O}{3}] bright emission filaments out ahead of the [\ion{S}{2}] bright emission filaments marking the location
of a local shock front. The center panel highlights the much stronger [\ion{S}{2}] emission filaments compared to the Gum Nebula's more diffuse emission. Because ISM shocks display stronger [\ion{S}{2}] compared to that
of H$\alpha$ as demonstrated in the Antlia spectra (see Table 4), the
observed strong [\ion{S}{2}] emission in the proposed collision filaments leads
support for a collision scenario. This conclusion is further indicated by the right panel which shows the observed ratio of F([\ion{S}{2}])/F(H$\alpha$ + [\ion{N}{2}]) with the stars removed. The sharp dark filaments in this panel image represent ratios between 1 and 1.3, values consistent with the measured spectral line strengths of Antlia's shock filaments (see Table 4).
If the Antlia remnant's expansion has led to a collision of it with an
outer portion of the Gum Nebula, then the Antlia's distance must be roughly the same as for the Gum Nebula. Unfortunately, the distance to the Gum Nebula is poorly known, with values ranging from $200 - 500$ pc \citep{Brandt1971,Zealey1983,Graham1986,Woer2001,Kim2005,Howarth2019}.
Nonetheless, distances much below 200 pc like suggested by \citet{McCull2002} would appear to be ruled out, as would the suggestion by \citet{Tetzlaff2013} that the 1.2 Myr old PSR J0630-2834 was formed in the Antlia remnant at a distance $\approx$135 pc.
If we adopt recent distances estimate to the Gum of 300 - 450 pc, which include estimates for some of the Gum Nebula's ionizing stars ($\zeta$ Pup; \citealt{Howarth2019}), then the Antlia remnant's physical diameter is $\sim120 - 180$ pc. However, given the Gum Nebula's 36\degr \ diameter, its outer parts extend over a distance of some 200 to 300 pc along the line of sight, this greatly complicates using a collision scenario to constrain the Antlia remnant's distance.
\subsection{The Power of FUV Emissions for Finding Interstellar Shocks}
A few large Galactic SNRs and superbubbles have been recently studied in terms of their far UV emissions. For example, studies of far UV emissions have been reported for the Vela SNR \citep{Nish2006}, the Cygnus Loop \citep{Seon2006}, the Lupus Loop \citep{Shinn2006}, and the Orion-Eridanus Superbubble \citep{Kregenow2006}. Many of these made use of the SPEAR imaging spectrograph \citep{Edel2006}.
However, there have been few papers reporting discoveries of large interstellar emission structures using FUV emissions. One such paper is that of \citet{Bracco2020} who reported finding a 30\degr \ long UV arc in Ursa Major using GALEX images. That work was a follow-up to an earlier detection of a much shorter 2.5\degr \ filament by \citet{McCull2001} using deep H$\alpha$ imaging. However, only through the GALEX's FUV images was the full extent of this faint, long interstellar filament finally revealed.
The FUV emission of moderate velocity shocks is dominated by the hydrogen 2-photon continuum, resonance lines of C IV and Mg II, and intercombination lines such as C III] and Si III]. \cite{Bracco2020} computed the GALEX FUV and NUV count rates for shocks from the \cite{Sutherland2017} MAPPINGS models. Shocks slower than about 100 $\rm km~s^{-1}$ are dominated by the 2-photon continuum, while faster shocks have strong contributions from C IV and other lines. The predicted ratios range from about 0.1 to 0.9. Based on our own model calculations, a reddening $E(B-V)$ $\sim$ 0.1 does not affect the ratios very much, though it reduces both the NUV and FUV count rates by a factor of 2. A major caveat, however, is that scattering in edge-on sheets of emitting gas, in particular SNR filaments, can strongly reduce the intensities of resonance lines such as C IV \citep{Cornett1992}, reducing the FUV/NUV ratio.
Above we have presented large GALEX FUV mosaic images of two large nebulae which appear to be true Galactic SNRs, but were either only suspected or missed in most radio studies which have concentrated their searches near the Galactic plane. The object G249+24 also appears to have been missed in wide FOV optical surveys due to its relatively weak H$\alpha$ emissions.
Both our investigations and that of \citet{Bracco2020} indicate that broad, far UV imaging can be an especially useful means for detecting and distinguishing interstellar shocks and, in some cases, is more sensitive compared to H$\alpha$ imaging. As noted by \citet{Bracco2020}, both line emissions and two photon continuum emissions contribute to FUV emission in shocks with velocities above $\simeq 50$ km s$^{-1}$.
To illustrate the usefulness of far UV emission imaging for detecting SNRs,
\citet{Bracco2020} noted the presence of networks of thin FUV filaments in both the Antlia SNR and the recently discovered Galactic remnant G70.0-21.5 \citep{Boumis2002,Fesen2015,Raymond2020}. In Figure~\ref{G70} we show a comparison of MDW's H$\alpha$ image and the GALEX FUV image of G70.0-21.5. Until the study of the SNRs discussed here, this remnant at $b = -21.5\degr$ had been the remnant with the highest Galactic latitude. This figure nicely demonstrates how that the FUV image makes the remnant easy to detect and helps to define its full dimensions despite the many missing individual GALEX images. Knowing about the existence of these GALEX FUV images could have helped \citet{Fesen2015} and \citet{Raymond2020} in their analyses of this remnant in regard to the remnant's true physical size. In summary, therefore, far UV images appear to be an especially useful tool for identifying interstellar shocks like those found in SNRs, albeit though best suited for high Galactic latitude searches.
\section{Conclusions}
We have investigated the nature of two large and suspected supernova remnants located at unusually high Galactic latitudes through GALEX far UV emission mosaics and optical images and spectra.
This research also has uncovered one new Galactic SNR. Our findings include:
1) The large radio emission shell, G354.0-33.5, seen in
1420 MHz and 1.4 GHz radio polarization maps is very likely a SNR. The remnant exhibits numerous sharp FUV emission filaments in a thin, unbroken shell with angular dimensions of $11\degr \times 14.0\degr$. It also exhibits a coincident H$\alpha$ emission shell.
2) A group of bright, sharp FUV emission filaments coincident with numerous but faint H$\alpha$ filaments appears to be a previously unrecognized SNR, G249.7+24.7. Optical spectra of several filaments show evidence for the presence of 50 - 150 km s$^{-1}$ shocks. Deep H$\alpha$ images reveal a highly filamentary morphology like that seen in evolved SNRs.
\citet{Becker2021} independently discovered this remnant as a circular $4.4\degr$ shell of diffuse, soft X-ray emission (0.1 keV) with an estimated radio spectral index of
$-0.69 \pm0.08$. Moderate dispersion optical spectra of two stars lying in the direction of this remnant have been found to exhibit high-velocity, red and blue shifted \ion{Na}{1} absorption lines in the range of $55 - 73$ km s$^{-1}$, with one star showing an added blue absorption feature at $-125$ km s$^{-1}$. The closest of these two stars lies at a Gaia estimated distance of 386 pc which, if confirmed by follow-up higher dispersion data, sets a strict maximum distance to the G249.7+24.7 SNR at less than 400 pc making it one of the closest Galactic SNRs known.
3) Despite its enormous angular dimensions ($20\degr \times 26\degr$), our GALEX FUV mosaic images, plus wide-angle H$\alpha$ images and optical spectra strongly support a SNR origin for the Antlia nebula. This conclusion is in line with a reported $\pm 70$ km s$^{-1}$ \ion{Ca}{2} absorption in one background star. We estimate an age $\sim$10$^{5}$ yr, which is an order of magnitude less than the earlier estimate $\sim$1 Myr. We also find the remnant is in likely in direct physical contact along its southwestern rim with the Gum Nebula.
4) Our investigation of suspected SNR located at unusually high Galactic latitudes ($> 15\degr$) highlights the value of UV images to detect interstellar shocks.
Follow-up work on these three objects could include optical spectra of the FUV and H$\alpha$ filaments of the G354-33 remnant, wide-field [\ion{O}{3}] $\lambda$5007 imaging of G249+24 exploring its [\ion{O}{3}] emission,
and high-dispersion echelle spectra of the two apparent background stars toward G249+24 to investigate our reported high-velocity \ion{Na}{1} absorption lines.
Higher resolution H$\alpha$ images and optical spectra of the series of the long and very bright filaments seen in southwestern limb of the Antlia remnant could also provide an additional test of our conclusion regarding Antlia's collision with the Gum Nebula.
Although we have found that all three nebula in our study are SNRs, only the G249.7+24.7 remnant has a distance or upper limit estimate, leaving us with only approximate physical parameters and evolutionary status. This problem can be addressed especially well for the other two objects, G354.0-33.5 and Antlia, given their enormous angular dimensions and hence offering numerous potential background stars to explore high-velocity ISM absorptions. Given the era of accurate Gaia parallaxes, obtaining high-dispersion spectra looking for high-velocity
\ion{Na}{1} $\lambda\lambda$5890,5896 and \ion{Ca}{2} $\lambda$3934 absorptions in background stars like that done for several SNRs could be quite fruitful
(e.g., Vela: \citealt{Jenkins1976J,Danks1995,Cha2000}; IC~443: \citealt{Welsh2003}; S147: \citealt{Sallmen2004}; Cygnus Loop: \citealt{Fesen2018}; W28: \citealt{Ritchely2020}).
Such observations might also help gauge the range of shock velocities across the very large and expansive remnants of G354.0-33.5 and Antlia.
Our findings that both the G354.0-33.5 and Antlia nebula are real SNRs was a bit of a surprise. Out of the 300 or so confirmed SNRs, there are only about a dozen SNRs larger than two degrees and less than half that number located more than ten degrees off the Galactic plane \citep{Green2015,Green2019}. Now with
the realization that some Galactic remnants can reach angular sizes several times larger than even Vela, it encourages the investigation of other large emission shells suspected as possible SNRs or previously viewed as unlikely SNR candidates. It also raises the value of SNR searches in the $\mid b \mid$ $> 10\degr$ range, generally viewed before as unproductive.
Finally, we note that although far UV imaging appears to be a sensitive new tool for uncovering the presence of interstellar shocks, it is most useful in searching uncomplicated regions, like the objects discussed here located in the Galactic halo. However, the number of similar but unrecognized high latitude Galactic remnants is probably pretty small. Nonetheless, given the dozens of unconfirmed but seemingly likely or suspected SNRs in the literature (see list in \citealt{Green2019}'s catalogue), many new SNR discoveries may be aided by the use of UV imaging.
\bigskip
We thank Eric Galayda and the entire MDM staff for making the optical observations possible, and the SALT Observatory and Resident Astronomer staff for obtaining the excellent RSS spectra despite disruptions due to COVID-19 restrictions. We also thank R.\ Benjamin for helpful discussions. This work made use of the Simbad database, NASA's Skyview online data archives, and the Max Planck Institute for Radio Astronomy Survey Sampler. This work is part of R.A.F's Archangel III Research Program at Dartmouth. D.M.\ acknowledges support from the National Science Foundation from grants PHY-1914448 and AST-2037297.
|
\section{Introduction}
\label{sec:introduction}
An important aspect in nanotechnology is the determination of
characteristics of the fundamental building blocks of matter, namely
atoms and molecules.
One of these characteristics are the electrostatic properties that
govern in many cases the functionality of nanoscale objects and systems.
This is particularly important for new materials and devices associated
with nanoscale electronics, such as semiconductors.
The investigation of electrostatics at the nanoscale becomes therefore
more and more important and is a vivid field of ongoing research
\cite{Matyba2009, Musumeci2014, Fuchs2016, Hu2016, Wagner2019a}.
\begin{figure}[tb]
\centering
\includegraphics[width=1.0\linewidth]{figs/ReferenceImage/STM_SQDM.png}
\caption{
Top: Scanning tunneling microscope image of a sample presented in
\cite{Wagner2019a}.
The image size is \unit[600$\times$600]{\AA}.
Bottom: 2D electrostatic potential\xspace image of the same sample at height $z =
\unit[20]{\AA}$.
The image was divided into 200$\times$200 pixels.
}
\label{fig:STM_SQDM}
\end{figure}
Scanning quantum dot microscopy (SQDM), introduced in
\cite{Wagner2015, Green2016} allows to measure the electrostatic potentials\xspace of
nanostructures with sub-nanometer resolution at large imaging
distances.
It generates 2D images of the electrostatic potential\xspace (Fig.~\ref{fig:STM_SQDM}).
It furthermore allows to separately image the electrostatic potential
and the surface topography.
Recently it was shown in \cite{Wagner2019a} that SQDM does
not only generate qualitative but also quantitative images of the
electrostatic potential of nanostructures.
The paper also demonstrated large-scale imaging
(Fig.~\ref{fig:STM_SQDM}), resolving both small (single atoms and
molecules) and large structures (an island composed of several hundreds
of molecules) in the same image, whereas in the previous publications of
SQDM \cite{Wagner2015, Green2016} only isolated atoms and molecules were
imaged independently of each other.
Thereby, SQDM became a mature widely applicable microscopy technique for
the area of nanotechnology.
This paper is the control engineering counterpart of \cite{Wagner2019a},
wherein the controller was only briefly mentioned.
We present in detail the tailored two-degree-of-freedom
control algorithm that is a key enabling element that turned SQDM into
a well applicable microscopy technique.
The controller is a variant of the one presented in \cite{Maiworm2018},
which included a Gaussian process as a feedforward signal generator.
However, that approach is not yet ready for the deployment in the
experiment because further research, in particular regarding the
computational issues for the online learning of the Gaussian process,
has to be conducted.
Henceforth, the objective of this work is to present the controller
version that was actually used in \cite{Wagner2019a}.
Opposed to the initial method of generating SQDM images based on
spectroscopy grids (see \cite{Wagner2015}), the controller now allows to
continuously scan the sample, which yields order-of-magnitude faster
image generation and eliminates the need for spectroscopy.
This puts it in line with other microscopy techniques like
scanning tunneling microscopy \cite{Binnig1982} and atomic force
microscopy \cite{Albrecht1991}.
Furthermore, the controller now also allows to scan images with highly
varying electrostatic potentials\xspace.
Previously, using spectroscopy grids one had to set up the spectrum
range, which increases with the electrostatic potential\xspace, before starting the grid.
This was a huge setback in speed if the electrostatic potential\xspace and with that the spectrum
range was large because the entire voltage range had to be scanned.
~\\
After this introduction, the paper is structured as follows.
Sec.~\ref{sec:scanning_quantum_dot_microscopy} provides the
fundamentals of scanning quantum dot microscopy\xspace.
The 2DOF control framework is presented in Sec.~\ref{sec:control} and
thoroughly analyzed in simulations in Sec.~\ref{sec:simulations}.
Further experimental results are presented in
Sec.~\ref{sec:experimental_results} and the paper is concluded in
Sec.~\ref{sec:conclusion}.
\section{Scanning Quantum Dot Microscopy}
\label{sec:scanning_quantum_dot_microscopy}
This section explains the working principle of scanning quantum dot microscopy\xspace and the associated
image generation process. The process is analyzed and a model derived
that will be used in Sec.~\ref{sec:simulation_results} for simulations.
\subsection{Working Principle}
\label{sec:working_principle}
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{figs/SQDM_aufbau}
\caption{Schematic of Scanning Quantum Dot Microscopy (left). The tip
of a frequency modulated non-contact AFM is decorated with a quantum
dot (QD) and a bias voltage source is connected between the microscope
tip and the sample.
Depending on the electrostatic potential\xspace of the nanostructure on the sample surface and the
bias voltage \ensuremath{V_\text{b}}\xspace, a single electron ($e^-$) tunnels back and forth
between the AFM tip and the quantum dot.
The tip together with the quantum dot is moved in a raster scanning
pattern (right).}
\label{fig:SQDM_schematic}
\end{figure}
SQDM is able to measure electric surface potentials and allows to
distinguish between topographical and electrostatic effects.
It utilizes a frequency modulated non-contact atomic force microscope
(NC-AFM) \cite{Albrecht1991}, operated in ultra-high
vacuum and at a temperature of \unit[5]{K}.
The atomically sharp tip (Fig.~\ref{fig:SQDM_schematic}) is mounted to a
tuning fork (qPlus sensor \cite{Giessibl2003}) that oscillates with a
frequency $f = f_0 + \ensuremath{\Delta f}\xspace$ of around \unit[30]{kHz}, where
$f_0$ is the free resonance frequency and \ensuremath{\Delta f}\xspace is the frequency change that
is caused by a vertical force gradient acting on the tip.
Typically, this force is the result of the tip-sample interaction.
The AFM tip is decorated (\cite{Toher2011, Fournier2011, Wagner2012,
Wagner2014, Findeisen2016}) with a quantum dot (QD)\footnote{Currently a
PTCDA (Perylenetetracarboxylic dianhydride) molecule serves as the QD.},
a nano-sized object whose energy levels can take only discrete values.
Changes of the electrostatic potential\xspace \ensuremath{\Phi_\text s}\xspace of the surface can change the
QD's charge state via gating as an electron tunnels from the tip into
the QD.
This leads to an abrupt change in the tip-sample force.
These tip-sample force changes are detected by the NC-AFM,
effectively transducing the information about the electrostatic
potential of, e.g.\xspace a nanostructure, into the measurable quantity \ensuremath{\Delta f}\xspace.
Monitoring the charging events of the QD while scanning the sample is
the basic working principle of SQDM.
To detect the charging events, a bias voltage source \ensuremath{V_\text{b}}\xspace is
connected to the sample while the tip is grounded.
The associated electrostatic potential\xspace $\Phi_\text{b}$ that is generated by \ensuremath{V_\text{b}}\xspace is
superimposed on the intrinsic electrostatic surface potential of the
sample \ensuremath{\Phi_\text s}\xspace.
Accordingly, a change in \ensuremath{V_\text{b}}\xspace then leads to a change of the effective
electric potential at the QD.
If this reaches a threshold value, a change in the QD's charge state
is triggered.
Charging of the QD leads to a change in the tip-sample force, whose
gradient is proportional to \ensuremath{\Delta f}\xspace for small amplitudes of the AFM tip
oscillation (see \cite{Giessibl1997}).
The changes of the tip-sample force generated by the charging events
appear in the so-called \emph{spectrum} $\ensuremath{\Delta f}\xspace(\ensuremath{V_\text{b}}\xspace)$ as
features that we denote as \emph{dips} (Fig.~\ref{fig:spectrum}).
The $\ensuremath{\Delta f}\xspace(\ensuremath{V_\text{b}}\xspace)$ spectrum is the superposition of a parabola (\cite{Gross2009})
and two dips, one at negative \ensuremath{V_\text{b}}\xspace values and one at positive \ensuremath{V_\text{b}}\xspace values.
The dips separate \ensuremath{V_\text{b}}\xspace intervals with different charge states of the QD.
\begin{figure}[tb]
\centering
\includegraphics[width=\linewidth]{figs/Spectrum/spectrumFitZoom_cut_edited}
\caption{The spectrum $\ensuremath{\Delta f}\xspace(\ensuremath{V_\text{b}}\xspace)$ describes how the tuning fork
oscillation frequency changes with the bias voltage \ensuremath{V_\text{b}}\xspace at one
particular position in space.
The overall parabolic shape of the curve is a consequence of the
tip-sample capacitance (see \cite{Gross2009}).
The SQDM specific dips result from the charging events of the quantum
dot.
The voltage values where the two dips reach their minimum are
indicated by \ensuremath{V^-}\xspace and \ensuremath{V^+}\xspace.
}
\label{fig:spectrum}
\end{figure}
We denote the voltage values at which the dips reach their minimum with
\ensuremath{V^-}\xspace and \ensuremath{V^+}\xspace, or short \ensuremath{V^\mp}\xspace.
These values characterize the dips' positions within the spectrum.
By means of
\begin{align}
\ensuremath{\Phi^*}\xspace(p) = \frac{\ensuremath{V^-}\xspace_0 \cdot \ensuremath{\Delta V}\xspace(p)}{\ensuremath{\Delta V}\xspace_0} - \ensuremath{V^-}\xspace(p) \ ,
\label{eq:SQDM}
\end{align}
the effective surface potential \ensuremath{\Phi^*}\xspace at the position of the tip $p = (x,y,z)$
can be calculated, where $\ensuremath{\Delta V}\xspace = \ensuremath{V^+}\xspace - \ensuremath{V^-}\xspace$ and $\ensuremath{V^-}\xspace_0, \ensuremath{\Delta V}\xspace_0$ are
reference points.
The actual surface potential \ensuremath{\Phi_\text s}\xspace can be recovered from \ensuremath{\Phi^*}\xspace
through deconvolution in post-processing \cite{Wagner2019b}.
In the rest of this paper we will deal with \ensuremath{\Phi^*}\xspace unless otherwise
indicated.
For more details see \cite{Wagner2019b}.
The \ensuremath{V^\mp(p)}\xspace maps together with $\ensuremath{\Phi^*}\xspace(p)$ of the sample shown in
Fig.~\ref{fig:STM_SQDM} are depicted in Fig.~\ref{fig:dipMaps}.
\begin{figure}[htpb]
\centering
\includegraphics[width=\linewidth]{figs/ReferenceImage/referenceImage.png}
\caption{
From left to right: the \ensuremath{V^-}\xspace map, the \ensuremath{V^+}\xspace map, and the
resulting electrostatic potential image \ensuremath{\Phi^*}\xspace of
Fig.~\ref{fig:STM_SQDM}.
}
\label{fig:dipMaps}
\end{figure}
\subsection{Original Image Generation Process}
\label{sec:original_image_generation_process}
The image generation process that had been previously used in
\cite{Wagner2015, Green2016} is as follows.
The sample (Fig.~\ref{fig:STM_SQDM}) is discretized in pixels
(Fig.~\ref{fig:SQDM_schematic}) and the tip with the QD is moved from
pixel to pixel.
At the first pixel, a complete spectrum (like Fig.~\ref{fig:spectrum})
is measured and the positions of the dips \ensuremath{V^\mp}\xspace are determined.
It has to be assumed that the interval in which \ensuremath{V^\mp(p)}\xspace will change while
scanning the sample is approximately known a priori.
For the following pixels, the bias voltage \ensuremath{V_\text{b}}\xspace is swept accordingly
within these two intervals (e.g.\xspace a voltage range of \unit[0.2]{V} instead
of \unit[6]{V} for the complete spectrum).
This results in the measurement of local dip spectra.
After obtaining the local dip spectra for all pixels, the \ensuremath{V^\mp(p)}\xspace
values are determined for each pixel and used in \eqref{eq:SQDM}
to generate \ensuremath{\Phi^*}\xspace.
The main limitation of this image generation process is the required
large total measurement time.
For instance, measuring the local dip spectra takes about \unit[3]{s} for
each dip and pixel for a certain \ensuremath{V_\text{b}}\xspace interval size.
Hence, the determination of the complete \ensuremath{V^\mp}\xspace maps in
Fig.~\ref{fig:dipMaps} would require \unit[66.7]{h}.\footnote{The data
shown in Fig.~\ref{fig:dipMaps} was generated by the proposed control
approach in this paper, which results in significantly smaller
measurement times.}
This severely limits the applicability of SQDM.
In particular,
\begin{itemize}
\item the microscope is blocked for several hours for the generation
of one image,
\item the longer the measurement time, the higher the probability of
failures, and
\item effects like drift increase and deteriorate image quality.
\end{itemize}
Furthermore, obtaining \ensuremath{\Phi^*}\xspace images from a grid of spectra limits SQDM
in several ways:
\begin{itemize}
\item Generation of fast and rough images for a first impression is
not possible.
\item Generation of images like the one presented in
Fig.~\ref{fig:STM_SQDM} and Fig.~\ref{fig:dipMaps}
become practically impossible.
\item The measurement times of the local spectra increase if the
aforementioned voltage intervals, wherein the dips move, increase.
This can occur, for instance, for other substrate-sample
combinations with stronger electrostatic variations and thus, reduce
the applicability even further.
\end{itemize}
\subsection{Simulation Model}
\label{sec:simulation_model}
The controllers, as outlined in Section~\ref{sec:control}, are going to
be evaluated in simulations in Section~\ref{sec:simulations}.
This requires a SQDM simulation model, which we derive in the following.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.7\linewidth]{figs/Block_diagrams/block_diagram_SQDM}
\caption{SQDM block diagram.
}
\label{fig:block_diagram_SQDM}
\end{figure}
Fig.~\ref{fig:block_diagram_SQDM} illustrates the SQDM system in a block
diagram.
In the experiment, the current bias voltage \ensuremath{V_\text{b}}\xspace and tip
position $p$ lead to an oscillation frequency $f$ of the tuning fork
different from the free resonance frequency $f_0$.
Hence, \ensuremath{V_\text{b}}\xspace and $p$ are the inputs to the system and the frequency change
\ensuremath{\Delta f}\xspace is the output.
The bias voltage \ensuremath{V_\text{b}}\xspace is the control input to the system and
can be chosen freely, whereas the tip position $p$ changes according
to the raster scanning pattern as shown in Fig.~\ref{fig:SQDM_schematic}.
As \ensuremath{V_\text{b}}\xspace changes, the frequency changes almost instantaneously according
to Fig.~\ref{fig:spectrum}.
Therefore, the first block in Fig.~\ref{fig:block_diagram_SQDM} can be
considered stationary.
The relatively small frequency change \ensuremath{\Delta f}\xspace (only up to a few Hz,
compared to \unit[30]{kHz} of the free resonance frequency
$f_0$) is determined from the oscillation signal using a phase-locked loop\xspace (PLL),
which is modeled as a first order system with bandwidth
\ensuremath{\omega_\text{PLL}}\xspace and whose output is corrupted by white Gaussian noise.
To simulate SQDM and the image generation process, the spectrum
$\ensuremath{\Delta f}\xspace(\ensuremath{V_\text{b}}\xspace)$ has to be available as an analytic function.
It consists of a parabola and the two dips, which we model as
Gaussian curves\footnote{
Note that in the actual experiment, the shape of the dips is somewhere
between a Gaussian curve and a half-circle, depending on the chosen tip
oscillation amplitude, the value of \ensuremath{V^\mp}\xspace, and the width of the
electronic level of the QD.
See also \cite{Kocic2015}.
}.
The ansatz for the spectrum is
\begin{align}
\ensuremath{\Delta f}\xspace(\ensuremath{V_\text{b}}\xspace) &= \ensuremath{\Delta f}\xspace_\text{para}(\ensuremath{V_\text{b}}\xspace) + \ensuremath{\Delta f}\xspace^-(\ensuremath{V_\text{b}}\xspace) + \ensuremath{\Delta f}\xspace^+(\ensuremath{V_\text{b}}\xspace)
\label{eq:spectrum_fit}
\end{align}
with the parabola function
\begin{align*}
\ensuremath{\Delta f}\xspace_\text{para}(\ensuremath{V_\text{b}}\xspace) = p_1 \ensuremath{V_\text{b}}\xspace^2 + p_2 \ensuremath{V_\text{b}}\xspace + p_3
\end{align*}
and the Gaussian curves
\begin{align}
\ensuremath{\Delta f}\xspace^-(\ensuremath{V_\text{b}}\xspace) &= d^- \cdot
\exp\!\left(
-\!\left(\frac{\ensuremath{V_\text{b}}\xspace-\ensuremath{V^-}\xspace}{w^-}\right)^2
\right) \label{eq:negPeak} \\
\ensuremath{\Delta f}\xspace^+(\ensuremath{V_\text{b}}\xspace) &= d^+ \cdot
\exp\!\left(
-g\!\left(\frac{\ensuremath{V_\text{b}}\xspace-\ensuremath{V^+}\xspace}{w^+}\right)
\right) \label{eq:posPeak}
\end{align}
for the dips, where $d^\mp, \ensuremath{V^\mp}\xspace, w^\mp$ are the respective depth,
position, and width of the dips.
The function $g(\cdot)$ in \eqref{eq:posPeak} is
the polynomial $g(x) = a_1 x^2 + a_2 x^4 + a_3 x^6$.
The parameters in \eqref{eq:spectrum_fit}, \eqref{eq:negPeak}, and
\eqref{eq:posPeak} are then fitted using experimental data.
The resulting fit in Fig.~\ref{fig:spectrum} and
Tab.~\ref{tab:spectrumFit} shows that the proposed
ansatz is well suited to model the experimentally acquired \ensuremath{\Delta f}\xspace spectrum.
\begin{table}[htpb]
\centering
\caption{\ensuremath{\Delta f}\xspace Spectrum Fit Parameters}
\begin{tabular}{llll}
\toprule
$c_1 = -1.3$ & $a_1 = 0.70$ & $d^- = -1.1$ & $d^+ = -4.6$ \\
$c_2 = 0.56$ & $a_2 = -0.61$ & $V^- = -1.3$ & $V^+ = 4.3$ \\
$c_3 = -0.76$ & $a_3 = 1.64$ & $w^- = 0.022$ & $w^+ = 0.087$ \\
\bottomrule
\end{tabular}
\label{tab:spectrumFit}
\end{table}
To simulate the changing tip position, the experimental reference
data of Fig.~\ref{fig:dipMaps} for \ensuremath{V^\mp}\xspace for every pixel
is fed to \eqref{eq:negPeak} and \eqref{eq:posPeak}.
Thus, at each new pixel in the simulation, the dips are
shifted according to the experimental reference.
\section{Controller Development}
\label{sec:control}
The effective electrostatic potential\xspace of the sample at the position of the tip $\ensuremath{\Phi^*}\xspace(p)$
changes in the three dimensional space.
However, since SQDM is based on AFM, only 2D images can be generated.
Therefore, we consider in the following $\ensuremath{\Phi^*}\xspace(x,y)$ that, according to
\eqref{eq:SQDM}, depends in turn on the quantities $\ensuremath{V^\mp}\xspace(x,y)$, i.e.,\xspace
$\ensuremath{\Phi^*}\xspace\big( \ensuremath{V^\mp}\xspace(x,y) \big)$.
The \ensuremath{V^\mp}\xspace values are a priori unknown, cannot be measured directly, and
change with the tip position $p = (x,y)$.
Therefore, SQDM with the changing \ensuremath{V^\mp}\xspace values can be regarded as
a parameter varying system and the objective of this paper is to design
a control framework that automatically determines and tracks the unknown
parameters while scanning the sample.
As outlined, \ensuremath{V^\mp}\xspace cannot be measured directly and a model of the
respective dynamics is unavailable because it depends, besides the scan
speed, on the electrostatic potential\xspace, which itself is unknown.
Thus, a potential control approach has to adapt \ensuremath{V_\text{b}}\xspace indirectly, based on
a quantity that can be measured, in this case the frequency change \ensuremath{\Delta f}\xspace.
Hence, we are looking for a control law $\ensuremath{V_\text{b}}\xspace(t) = \kappa(\ensuremath{\Delta f}\xspace)$.
In the following we present a two-degree-of-freedom (2DOF) control approach,
consisting of a feedback and a feedforward (FF) part
(Fig.~\ref{fig:block_diagram_general}).
We furthermore develop two different feedback controllers, leading to
two different versions of the 2DOF controller.
The first feedback controller is an extremum seeking controller (ESC)
and the second controller is called \emph{slope tracking controller}
(STC).
\green{The central idea to both controllers is, instead of measuring the dip
spectrum (time consuming, contains a lot of unnecessary data), to track
one specific reference point in each dip.
A major difference between the ESC and the STC is this reference point.
While the ESC tracks directly \ensuremath{V^\mp}\xspace, the STC tracks a point on the dip's
slope (thus the name).}
The individual 2DOF parts will be discussed in more detail in the remainder
of this section.
\begin{figure}[htpb]
\centering
\includegraphics[width=\linewidth]{figs/Block_diagrams/block_diagram_general_control_structure}
\caption{Block diagram of the closed loop: The bias voltage \ensuremath{V_\text{b}}\xspace is the
sum of the feedback and feedforward output. The feedforward part is
influenced by the current tip position $p$.}
\label{fig:block_diagram_general}
\end{figure}
\subsection{Extremum Seeking Control}
\label{sec:extremum_seeking_control}
The fact that \ensuremath{V^\mp}\xspace characterize the minima of the local convex
dips $\ensuremath{\Delta f}\xspace^\mp$ satisfying
\begin{align}
\ensuremath{V^-}\xspace(x,y) &= \argmin_{\ensuremath{V_\text{b}}\xspace} \ensuremath{\Delta f}\xspace^-(x,y,\ensuremath{V_\text{b}}\xspace) \label{eq:Vneg} \\
\ensuremath{V^+}\xspace(x,y) &= \argmin_{\ensuremath{V_\text{b}}\xspace} \ensuremath{\Delta f}\xspace^+(x,y,\ensuremath{V_\text{b}}\xspace) ,
\label{eq:Vpos}
\end{align}
can be exploited for the determination of \ensuremath{V^\mp}\xspace.
Since \ensuremath{\Delta f}\xspace is measured online by the phase-locked loop\xspace, the minima of $\ensuremath{\Delta f}\xspace^\mp$ can be
determined continuously by appropriately adapting the input \ensuremath{V_\text{b}}\xspace in
\eqref{eq:Vneg} and \eqref{eq:Vpos} during the scanning process.
This can be achieved by employing methods of extremum
seeking control \cite{Ariyur2003, Zhang2012}.
As the name indicates, these methods are designed to find the extremum,
i.e.,\xspace a minimum or a maximum, of the output of a given system.
The core principle includes a seeking element that continuously samples
the output signal at the current operating point to obtain some kind of
direction or gradient information.
If a measure of the gradient is detected, the current operating point is
changed accordingly.
If the optimum is reached, the gradient is zero and the operating point
is not changed anymore.
Thus, ESC approaches are closely related to optimization and are
therefore also called \emph{real-time optimization} methods
\cite{Ariyur2003}.
In principle, an ESC can be realized with many different optimization
methods, though they have to satisfy a variety of additional properties,
such as, for instance, low computational load or the ability to deal
with constraints.
Works on extremum seeking date back to as early as 1922
\cite{Leblanc1922} or 1951 \cite{Draper1951}.
Though many other works were published in the second half of the 21th
century (see \cite{Guay2003}), it wasn't until the paper of
\cite{Krstic2000b} in the year
2000 that provided the first general stability analysis of ESC.
Since then, interest has sparked again and found its realizations in
applications such as anti-lock-breaking systems
\cite{Yu2002, Zhang2007}, maximum-power-point-tracking
\cite{Leyva2006, Brunton2010}, source seeking \cite{Zhang2007,
Cochran2009}, beam control in particle accelerators \cite{Scheinker2017},
or the control of plasma in a Tokamak reactor \cite{Centioli2008}.
Theoretic extensions for discrete time systems \cite{Choi2002},
for multivariable systems \cite{Rotea2000}, systems with partial model
information \cite{Guay2003, Adetola2006}, and further results on
stability \cite{Tan2006, Scheinker2017} have been presented.
Generalizations of the ESC scheme such as a unifying framework
\cite{Nesic2013a} and with other optimization approaches, such as
newton like extremum seeking \cite{Moase2009, Ghaffari2012}, stochastic
\cite{Liu2012, Liu2014} and a non-gradient approach \cite{Nesic2013b}
have been developed too.
For extensive lists of further publications see \cite{Tan2010,
Zhang2012, Scheinker2017}.
In this work we employ an adapted version of the approach of
\cite{Krstic2000b} as shown in Fig.~\ref{fig:block_diagram_ESC_basic}.
In this setup we deal with a local convex function $h(u)$ for which
there exists a minimum at $u = u^*$.
At this minimum we have naturally $\frac{\diff h}{\diff u}
\big\rvert_{u^*} = 0$.
To search for $u^*$, a dither signal $d(t) = \ensuremath{a_d}\xspace \sin(\ensuremath{\omega_d}\xspace t)$ is used
to perturb the current $u$.
The resulting signal $y(t) = h(u(t))$ is passed through a high pass
filter $\frac{s}{s + \ensuremath{\omega_\text H}\xspace}$ to eliminate any constant offsets.
The filtered signal $\xi_1(t)$ is then multiplied by the phase shifted
dither signal\footnote{
The phase of $\xi_1(t)$ is shifted by the high pass filter by $\phi$.
In order to be in phase, the dither signal that is multiplied with
$\xi_1(t)$ is shifted also by $\phi$.}
and low pass filtered with $\frac{\ensuremath{\omega_\text L}\xspace}{s + \ensuremath{\omega_\text L}\xspace}$.
This leads to $\xi_2(t)$ and it can be shown that
\begin{align*}
\lim_{t \to \infty} \xi_2(t) = \frac{\ensuremath{a_d}\xspace^2}{2} \frac{\diff h}{\diff u}
\Big\rvert_{\hat u}
\end{align*}
at the current operating point $\hat u$.
Choosing $K := \nicefrac{-2}{\ensuremath{a_d}\xspace^2}$, one obtains
\begin{align*}
\lim_{t \to \infty} \xi_3(t) = - \frac{\diff h}{\diff u}
\Big\rvert_{\hat u}
\end{align*}
i.e.,\xspace $\xi_3(t)$ tends to the negative gradient of $h(\hat u)$.
Hence, $\xi_3(t)$ can be used subsequently to implement a gradient
descent approach, realized by the integration of $\xi_3(t)$, turning
$\hat u$ into an estimation of the minimizer $u^*$.
\begin{figure}[htpb]
\centering
\includegraphics[width=\linewidth]{figs/Block_diagrams/block_diagram_ESC_basic}
\caption{Block diagram of the extremum seeking control approach.}
\label{fig:block_diagram_ESC_basic}
\end{figure}
~\\
Applied to SQDM we have $h(u) = \ensuremath{\Delta f}\xspace(\ensuremath{V_\text{b}}\xspace)$ as a local convex function and
the ESC computes the
derivative $\ensuremath{\Delta f}\xspace' = \frac{\diff\ensuremath{\Delta f}\xspace}{\diff\ensuremath{V_\text{b}}\xspace}$ by modulating the dither
signal $d(t)$ onto the \ensuremath{V_\text{b}}\xspace signal.
Since $\ensuremath{\Delta f}\xspace' = 0$ characterizes the dips' minima, it also characterizes
exactly the value of \ensuremath{V^\mp}\xspace.
Thus, if a potential controller regulates $\ensuremath{\Delta f}\xspace'$ to zero\footnote{Note
that only one dip can be tracked at a time.}, it automatically yields
$\ensuremath{V_\text{b}}\xspace = \ensuremath{V^\mp}\xspace$.
Therefore, $\ensuremath{\Delta f}\xspace' = 0$ is used as the reference and the gradient descent
is achieved by using an integral controller
\begin{align*}
\ensuremath{V_\text{b,C}}\xspace(t) = \ensuremath{K_\text{ESC}}\xspace \int_0^t \ensuremath{e_d}\xspace(\tau) \diff\tau
\end{align*}
that minimizes the error $\ensuremath{e_d}\xspace(t) = \ensuremath{\Df'_\text{ref}}\xspace - \ensuremath{\Delta f}\xspace'(t)$ with $\ensuremath{\Df'_\text{ref}}\xspace =
0$ and $\ensuremath{K_\text{ESC}}\xspace > 0$.
The voltage applied to the AFM cantilever is
\begin{align*}
\ensuremath{V_\text{b,mod}}\xspace(t) &= \ensuremath{V_\text{b,C}}\xspace(t) + d(t) + \ensuremath{V_\text{b,FF}}\xspace(t) \ ,
\end{align*}
where \ensuremath{V_\text{b,FF}}\xspace is the feedforward signal computed as detailed in
Section~\ref{sec:feedforward}.
Note that the signal $\ensuremath{V_\text{b}}\xspace = \ensuremath{V_\text{b,C}}\xspace + \ensuremath{V_\text{b,FF}}\xspace$ without the dither signal is
used for image generation.
The corresponding block diagram is depicted in
Fig.~\ref{fig:block_diagram_ESC}.
\begin{figure}[htpb]
\centering
\includegraphics[width=\linewidth]{figs/Block_diagrams/block_diagram_ESC}
\caption{Block diagram of the closed loop with the ESC.}
\label{fig:block_diagram_ESC}
\end{figure}
~\\
The ESC parameters that need to be chosen are the dither signal
amplitude \ensuremath{a_d}\xspace and frequency \ensuremath{\omega_d}\xspace, the low and high pass filter cut-off
frequencies \ensuremath{\omega_\text L}\xspace and \ensuremath{\omega_\text H}\xspace, and the control gain \ensuremath{K_\text{ESC}}\xspace.
In the following we present some general guidelines for choosing these
parameters based on the characteristics of the respective dip and the
phase-locked loop\xspace.
\subsubsection{Dither Signal}
The interval in which the applied bias voltage varies locally is
$[\ensuremath{V_\text{b}}\xspace-\ensuremath{a_d}\xspace, \ensuremath{V_\text{b}}\xspace+\ensuremath{a_d}\xspace]$.
The gradient of the dip $\frac{\diff \ensuremath{\Delta f}\xspace}{\diff \ensuremath{V_\text{b}}\xspace}$ is then
approximated within this interval, i.e.,\xspace an average gradient around the
current bias voltage \ensuremath{V_\text{b}}\xspace is computed.
A natural upper limit is therefore the dip width $w^\mp$ itself.
On the other hand, the smaller \ensuremath{a_d}\xspace, the more accurate the gradient at
\ensuremath{V_\text{b}}\xspace.
However, the \ensuremath{\Delta f}\xspace measurements for the gradient computation are corrupted
by noise, which poses a lower limit on \ensuremath{a_d}\xspace.
Thus, we get
\begin{align*}
\ensuremath{a_d}\xspace \leq w^\mp \ ,
\end{align*}
where \ensuremath{a_d}\xspace should be chosen as small as possible.
Regarding the choice of the dither signal frequency \ensuremath{\omega_d}\xspace: the higher
the frequency, the faster the gradient estimate converges but also the higher
the variance of the estimate.
Additionally, if \ensuremath{\omega_d}\xspace is much larger than the system's bandwidth, the dither
signal is damped and shifted significantly and in consequence the
gradient computation deteriorates.
Hence, an upper bound depends on the maximum bandwidth of the
system dynamics.
In SQDM, the phase-locked loop\xspace is the limiting element with its bandwidth $\ensuremath{\omega_\text{PLL}}\xspace$.
We have found
\begin{align*}
2 \ensuremath{\omega_\text{PLL}}\xspace \leq \ensuremath{\omega_d}\xspace \leq 10 \ensuremath{\omega_\text{PLL}}\xspace
\end{align*}
to work well, where larger values decrease convergence time but increase
variance.
\subsubsection{Filter Cut-Off Frequencies}
The objective of the high pass filter is to remove the constant offset
of the \ensuremath{\Delta f}\xspace signal.
This can be sufficiently fast achieved for
\begin{align*}
\ensuremath{\omega_\text H}\xspace \geq 0.5\ensuremath{\omega_d}\xspace \ .
\end{align*}
The objective of the low pass filter is to smoothen the signal $\xi_2$.
The smaller \ensuremath{\omega_\text L}\xspace the stronger the smoothing.
The stronger the smoothing, the less oscillatory the gradient
approximation becomes but also the slower the convergence.
Hence, choosing \ensuremath{\omega_\text L}\xspace is a trade-off between convergence speed and
variance, similar to \ensuremath{\omega_d}\xspace.
We have found that
\begin{align*}
0.1 \ensuremath{\omega_d}\xspace \leq \ensuremath{\omega_\text L}\xspace \leq 0.5 \ensuremath{\omega_d}\xspace
\end{align*}
works well.
Accordingly, \ensuremath{\omega_\text L}\xspace can be chosen within this interval, depending on the
objective of fast convergence (fast scanning) or small variance.
\subsubsection{Phase Shift}
The phase-locked loop\xspace and the high pass filter introduce an additional phase shift
\begin{align*}
\phi = \text{arg}\big(G_\text{PLL}(\im\ensuremath{\omega_d}\xspace) G_\text{HP}(\im\ensuremath{\omega_d}\xspace)\big)
\end{align*}
w.r.t. the dither signal, where $G_\text{PLL}(\im\ensuremath{\omega_d}\xspace) = \frac{1}{\im\ensuremath{\omega_d}\xspace
+ \ensuremath{\omega_\text{PLL}}\xspace}$ is the PLL transfer function and $G_\text{HP}(\im\ensuremath{\omega_d}\xspace) =
\frac{\im\ensuremath{\omega_d}\xspace}{\im\ensuremath{\omega_d}\xspace+\ensuremath{\omega_\text H}\xspace}$ the high pass filter transfer function.
This can be accounted for by adding the same phase shift $\phi$ to the
dither signal that is multiplied with the high pass outcoming signal
(see Fig.~\ref{fig:block_diagram_ESC_basic}).
\subsubsection{Control Gain}
To facilitate gain tuning, we define \ensuremath{K_\text{ESC}}\xspace via
\begin{align}
\ensuremath{K_\text{ESC}}\xspace = \frac{k}{\abs{ G_\text{PLL}(\im\ensuremath{\omega_d}\xspace) G_\text{HP}(\im\ensuremath{\omega_d}\xspace) }} \ ,
\label{eq:compGain}
\end{align}
where $k > 0$ is the new tunable ESC gain.
This redefinition automatically compensates the amplitude change of
$\xi_3(t)$ introduced by the PLL and the high pass filter.
This way the PLL and the high pass filter can be changed
without influencing the amplitude.
\green{Note that the low pass filter $G_\text{LP}(\im\ensuremath{\omega_d}\xspace)$
is not included in \eqref{eq:compGain} because then one loses the
possibility of adjusting the convergence of $\xi_3(t)$ independently of
that of $\hat u(t)$.
}
The larger $k$, the faster the convergence to the minimum
but also the more oscillatory the estimated minimizer $\hat u(t) =
\ensuremath{V_\text{b}}\xspace(t)$.
Hence, the larger $k$, the faster we can scan the sample but at the
cost of less accurate tracking.
In particular, oscillations due to high $k$ values eventually appear
in the final image as noise.
Furthermore, if the oscillations in \ensuremath{V_\text{b}}\xspace become too large, the dip might
be lost.
For instance, when the negative dip is tracked the oscillations might cause
\ensuremath{V_\text{b}}\xspace to leave the dip towards the left side of the dip.
There the gradient descent then leads the controller to further decrease
the \ensuremath{V_\text{b}}\xspace value down the parabola, making it impossible to recover the
dip.
\subsection{Slope Tracking Control}
\label{sec:slope_tracking_control}
Another possibility to control SQDM is by tracking a point on the dips'
slope (Fig.~\ref{fig:spectrum_negDip_refPoints}), instead of tracking
the dips' minima.
The resulting \ensuremath{\Delta f}\xspace value is then set as the reference value \ensuremath{\Df_\text{ref}}\xspace for
the whole sample and the deviations $\ensuremath{e_f}\xspace(t) = \ensuremath{\Df_\text{ref}}\xspace - \ensuremath{\Delta f}\xspace(t)$
are used directly as an error to the integral controller
\begin{align}
\ensuremath{V_\text{b,C}}\xspace(t) = \ensuremath{K_\text{STC}}\xspace \int_0^t \ensuremath{e_f}\xspace(\tau) \diff\tau
\label{eq:STC}
\end{align}
that adapts \ensuremath{V_\text{b}}\xspace accordingly with $\ensuremath{K_\text{STC}}\xspace < 0$.
The resulting voltage applied to the AFM cantilever is
\begin{align*}
\ensuremath{V_\text{b}}\xspace(t) &= \ensuremath{V_\text{b,C}}\xspace(t) + \ensuremath{V_\text{b,FF}}\xspace(t)
\end{align*}
with \ensuremath{V_\text{b,FF}}\xspace computed as detailed in Section~\ref{sec:feedforward}.
The resulting block diagram is depicted in
Fig.~\ref{fig:block_diagram_STC}.
\begin{figure}[htpb]
\centering
\includegraphics[width=\linewidth]{figs/Spectrum_negPeak_refPoints/spectrum_negPeak_refPoints_edit.png}
\caption{Measured spectrum of the negative dip with controller reference
points for ESC and STC.
The dashed black line indicates that the \ensuremath{\Df_\text{ref}}\xspace point of the STC has
three crossings with the blue spectrum.
Hence, three equilibria of the closed-loop system exist for
the STC.
The one on the dip's inner slope (on the right hand
side for the negative dip) is the point we want to stabilize.
}
\label{fig:spectrum_negDip_refPoints}
\end{figure}
\begin{figure}[htpb]
\centering
\includegraphics[width=0.8\linewidth]{figs/Block_diagrams/block_diagram_STC}
\caption{Block diagram of the closed loop with the STC.}
\label{fig:block_diagram_STC}
\end{figure}
~\\
The STC parameters that need to be chosen are the exact position of the
reference point \ensuremath{\Df_\text{ref}}\xspace and the control gain \ensuremath{K_\text{STC}}\xspace.
\subsubsection{Reference Point}
The choice of the STC reference point \ensuremath{\Df_\text{ref}}\xspace is very important.
In particular because most of the \ensuremath{\Delta f}\xspace values of a dip appear three times
in the local spectrum around the respective dip (see
Fig.~\ref{fig:spectrum_negDip_refPoints}).
We choose \ensuremath{\Df_\text{ref}}\xspace to lie on the inner slope of the dips (i.e.,\xspace on the right
slope of the negative dip and on the left slope of the positive dip) and
detail the reasons in the following.
First note, that the inner slope is always larger than the outer slope
(compare Fig.~\ref{fig:spectrum} and
Fig.~\ref{fig:spectrum_negDip_refPoints}), which suggests better control
performance on the inner slope.
Second, if the reference point is located on the inner slope, according
to \eqref{eq:STC} the controller is able to drive \ensuremath{V_\text{b}}\xspace to \ensuremath{\Df_\text{ref}}\xspace even if
\ensuremath{V_\text{b}}\xspace has left the dips towards the vertex
of the parabola (the sign of \ensuremath{e_f}\xspace doesn't change).
This would not be the case if the reference point was chosen to lie on
the outer slope and \ensuremath{V_\text{b}}\xspace had left the dips towards the part of the
parabola where values decrease indefinitely.
In that case the controller would drive the bias voltage to even larger
absolute values until the corresponding \ensuremath{\Df_\text{ref}}\xspace value on the parabola is
reached (crossing of the dashed black line and the blue parabola on the
left side in Fig.~\ref{fig:spectrum_negDip_refPoints}).
In general, the STC won't be able to recover the dip in that case.
Regarding the exact position, \ensuremath{\Df_\text{ref}}\xspace should lie relatively far away from the
dip minimum at \ensuremath{V^\mp}\xspace because this is another critical point for the STC.
If \ensuremath{V_\text{b}}\xspace moves over this point (e.g.\xspace $\ensuremath{V_\text{b}}\xspace < \ensuremath{V^-}\xspace$ for the negative dip)
the control error \ensuremath{e_f}\xspace becomes smaller instead of larger, which
automatically has a deteriorating effect on the control performance and
increases the probability that the dip is left towards the part of the
parabola where values decrease indefinitely.
\subsubsection{Controller Gain}
\label{sec:STC_controller_gain}
Regarding the STC gain \ensuremath{K_\text{STC}}\xspace, the same holds as for the ESC gain \ensuremath{K_\text{ESC}}\xspace.
The larger \ensuremath{K_\text{STC}}\xspace the faster the convergence to \ensuremath{\Df_\text{ref}}\xspace but also the more
oscillatory.
Thus, the larger \ensuremath{K_\text{STC}}\xspace the faster the sample can be scanned but at the
cost of less accurate tracking.
~\\
The STC does not require the computation of the derivative and is
therefore faster.
However, it introduces a systematic error due to the difference between
\ensuremath{V^\mp}\xspace and the \ensuremath{V_\text{b}}\xspace value at the STC reference \ensuremath{\Df_\text{ref}}\xspace
(approximately\footnote{Since the positive dip is wider, this error is
larger for the positive dip.} $\unit[20]{mV}$ in
Fig.~\ref{fig:spectrum_negDip_refPoints}) that we denote by \ensuremath{e_\text{STC}}\xspace.
Additionally, this error is not constant and changes while scanning
because when the dip changes its position it slides the parabola up- or
downwards.
For further clarification imagine that the negative dip moves vertically
upwards (no horizontal movement).
In that case, \ensuremath{V^-}\xspace does not change but $\ensuremath{V_\text{b}}\xspace(\ensuremath{\Df_\text{ref}}\xspace)$ changes to more
negative values, moving closer to \ensuremath{V^-}\xspace.
Hence, \ensuremath{e_\text{STC}}\xspace decreases as the dip moves vertically upwards and decreases
as the dip moves downwards.
This effect also occurs when the dip moves along the parabola because
there is always a vertical motion component.
In particular, the effect is larger for the positive dip because it is
typically located at steeper parts of the parabola where the vertical
motion is more pronounced.
A comparison of the advantages and drawbacks of the ESC and STC is
provided in Tab.~\ref{tab:comparison_feedback}.
\begin{table}[htpb]
\centering
\caption{Comparison of feedback approaches
}
\begin{tabular}{c|cc}
\toprule
& Advantages & Drawbacks \\
\midrule
ESC & robust & slow, many parameters \\
STC & fast, few parameters & delicate, systematic error \\
\bottomrule
\end{tabular}
\label{tab:comparison_feedback}
\end{table}
\subsection{Feedforward}
\label{sec:feedforward}
2DOF control approaches are well known in scanning probe techniques like
scanning tunneling or atomic force microscopy.
The control objective is usually to steer the piezo stages
that govern the movement of the microscope tip in $x$, $y$, and $z$
direction.
The $z$-piezo is controlled according to the topography feedback signal
(e.g.\xspace tunneling current in scanning tunneling microscopy or force in
contact mode atomic force microscopy \cite{Schitter2003, Schitter2004}),
whereas the piezos in $x$ and $y$ direction are controlled such that the
tip follows specific reference trajectories in the $(x,y)$-plane that
implement the raster scanning pattern.
For the main scanning direction, this is usually a triangular signal
\cite{Pao2007}.
The controllers are often based on models of the respective piezo
stages.
The feedforward part of the 2DOF controllers is therefore
also usually model based and techniques like $H_\infty$, $l_1$-optimal,
model inversion, or iterative learning control are employed.
For good overviews on this topic see \cite{Pao2007, Devasia2007,
Clayton2009, Leang2009, Yong2012, Gu2016}.
As already detailed throughout this section, the objective is to track
the dips' minima (with the ESC) or a point on the dips' inner slope
(with the STC).
This objective is taken care of by the respective feedback controller.
However, as already mentioned in
Section~\ref{sec:slope_tracking_control}, it is possible (and has
occurred) that the dips change their position faster than the respective
controller can adapt the bias voltage, which eventually leads to
the controller ``losing the dip''.
A combination of a rapid change of the electrostatic potential\xspace, a high scan speed,
and the controller dynamics is usually the cause.
In that case, the scanning process has to be aborted and restarted from
the beginning.
This risk can be substantially reduced by the generation of an
appropriate feedforward\xspace signal, such that the initial value for each controller
at each tip position stays always within the dips' interval.
In that way, the feedforward\xspace has not only the potential for increased
performance, and with that eventually higher scan speeds, but is
even essential for correct operation of SQDM.
Since an a priori model of the electrostatic potential\xspace is not available, a natural choice
is an approach that is based on the previously scanned line.
Fig.~\ref{fig:block_diagram_FF} shows a block diagram of the feedforward\xspace
signal generator for SQDM, which comprises the following elements and
features.
\begin{figure}[htpb]
\centering
\includegraphics[width=0.7\linewidth]{figs/Block_diagrams/block_diagram_FF}
\caption{Feedforward block diagram. $p = (x,y)$ denotes the current tip
position.
Current \ensuremath{V_\text{b}}\xspace values are stored in a buffer, delayed by one line and
outputted through a mean filter at the next line as the feedforward\xspace signal
\ensuremath{V_\text{b,FF}}\xspace.
}
\label{fig:block_diagram_FF}
\end{figure}
\begin{itemize}
\item \textbf{Buffer:}
The $\ensuremath{V_\text{b}}\xspace(x,y)$ values of the currently scanned line $y$ are stored in a
buffer alongside the indexing $x$ values within the line.
Hence, $x$ is the fast scan direction.
At the same time, already stored $\ensuremath{V_\text{b}}\xspace(x,y-1)$ values of the
previously scanned line $y-1$ are used as a basis for the feedforward\xspace signal
$\ensuremath{V_\text{b,FF}}\xspace(x,y)$ of the current line.
%
\item \textbf{Filter:}
The measured and buffered \ensuremath{V_\text{b}}\xspace values are corrupted by noise and
small ripples caused by the ESC (if this is used for control), which
have a deteriorating effect on the control performance,
especially in regions where the electrostatic potential\xspace is relatively flat.
Therefore, while scanning the current line $y$, the previously
measured $\ensuremath{V_\text{b}}\xspace(x,y-1)$ values are smoothed to generate the feedforward\xspace signal
$\ensuremath{V_\text{b,FF}}\xspace(x,y)$ using the mean filter
\begin{align*}
\ensuremath{V_\text{b,FF}}\xspace(x,y) = \frac{1}{n} \sum_{i=1}^n \ensuremath{V_\text{b}}\xspace
\left( x-\frac{n}{2}+i, y-1 \right)
\end{align*}
with filter window length $n$.
%
\item \textbf{Scan speed adaptation:}
The previous line $y-1$ is indexed using the measured $x$-position
values.
In the current line scan, the active $x$-position is determined and
used for picking the right reference value.
This allows for varying scan speed within a line.
\end{itemize}
This approach to generate the feedforward\xspace signal \ensuremath{V_\text{b,FF}}\xspace is simple,
straightforward to implement, yet practically powerful.
By adding \ensuremath{V_\text{b,FF}}\xspace to \ensuremath{V_\text{b}}\xspace, the controller has only to correct the
difference between the current and the previous line.
This results in a decreased control error and therewith a decreased
probability that the ``dips are lost''.
Accordingly, the scan speed can be increased while maintaining the same
image quality.
\section{Simulations}
\label{sec:simulations}
In this section, we simulate the SQDM process together with the ESC and
the STC and evaluate their performances qualitatively and
quantitatively.
To this end, we use the simulation model of
Section~\ref{sec:simulation_model} implemented in MATLAB/Simulink.
We use a fixed-step \texttt{ode3} (Bogacki-Shampine) solver with a
sampling time of $T_\text{s} = \unit[5]{ms}$.
Experimentally acquired spectra (Fig.~\ref{fig:exp_dips}) and the \ensuremath{V^\mp}\xspace
maps of Fig.~\ref{fig:dipMaps} are used as reference data.
The cutoff frequency of the phase-locked loop\xspace was determined by fitting a step
response and was computed as $\ensuremath{\omega_\text{PLL}}\xspace = \unit[10]{s^{-1}}$.
The normally distributed white noise has a standard deviation of
$\sigma_\text n = \unit[0.03]{Hz}$.
The parameters of the two controllers are listed in
Tab.~\ref{tab:controller_parameters}.
\begin{figure}[htbp]
\centering
\includegraphics[width=1.08\linewidth]{figs/Dips/exp_dips.png}
\caption{Experimentally acquired negative dip (top) and positive dip
(bottom) with model fit.
}
\label{fig:exp_dips}
\end{figure}
\begin{table}[htpb]
\centering
\caption{Controller Parameters (if not given otherwise)
}
\begin{tabular}{c|c}
\toprule
Negative Dip & Positive Dip \\
\midrule
ESC & ESC \\
$\ensuremath{a_d}\xspace = \unit[1]{mV}$ & $\ensuremath{a_d}\xspace = \unit[1]{mV}$ \\
$\ensuremath{\omega_d}\xspace = 4 \ensuremath{\omega_\text{PLL}}\xspace$ & $\ensuremath{\omega_d}\xspace = 4 \ensuremath{\omega_\text{PLL}}\xspace$ \\
$\ensuremath{\omega_\text L}\xspace = 0.2 \ensuremath{\omega_d}\xspace$ & $\ensuremath{\omega_\text L}\xspace = 0.2 \ensuremath{\omega_d}\xspace$ \\
$\ensuremath{\omega_\text H}\xspace = 3 \ensuremath{\omega_d}\xspace$ & $\ensuremath{\omega_\text H}\xspace = 3 \ensuremath{\omega_d}\xspace$ \\
$k = -5\cdot 10^{-5}$ & $k = -6\cdot 10^{-5}$ \\
\\
STC & STC \\
$\Delta f_\text{ref} = 1 \cdot w^-$
& $\Delta f_\text{ref} = -0.9 \cdot w^+$ \\
$K_\text{STC} = 0.04$ & $K_\text{STC} = -0.003$ \\
\bottomrule
\end{tabular}
\label{tab:controller_parameters}
\end{table}
\subsection{Influence of the Dip Parameters}
\label{sec:influence_of_the_dip_parameters}
In Sec.~\ref{sec:control} we have established a connection, in
the form of constraints, between some of the controllers' parameters and
the parameters of the dips and the phase-locked loop\xspace.
Here we furthermore investigate how the controllers' performance is
influenced by the dips' depth $d^\mp$ and width $w^\mp$.
As can be seen in Fig.~\ref{fig:influenceDipDepth} and
Fig.~\ref{fig:influenceDipWidth}, the deeper and narrower,
i.e.,\xspace the sharper the negative dip, the better because the faster the
convergence.
The same holds for the positive dip, as well as using the STC because
sharper dips have steeper slopes, which improve tracking.
We omit the corresponding plots for the sake of brevity.
Hence practitioners of SQDM should aim for sharp dips.
This can be achieved, for instance, with small oscillation amplitudes of
the microscope cantilever \cite{Kocic2015}.
However, at the same time such amplitudes increase the noise, which
ultimately imposes a lower bound on the oscillation amplitude.
Thus, the selection of the oscillation amplitude is a trade-off.
\begin{figure}[htbp]
\centering
\includegraphics[width=1.08\linewidth]{Sim/Sim0_opt_ctrl_parameters/influenceDipDepth.png}
\caption{ESC simulations of the negative dip with gain $k = 10^{-5}$.
The simulations were performed for different dip depths $m \cdot d^-$
with the scaling factors $m$ as indicated in the legend.
At $t = \unit[10]{s}$ the current bias voltage \ensuremath{V_\text{b}}\xspace was shifted towards
the right slope of the dip such that the ESC had to regain the minimum
at \ensuremath{V^-}\xspace.
All lines converge to the same signal because each simulation used the
same noise realization for the sake of better comparison.
Similar results are obtained for the positive dip.
}
\label{fig:influenceDipDepth}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[width=1.08\linewidth]{Sim/Sim0_opt_ctrl_parameters/influenceDipWidth.png}
\caption{ESC simulations of the negative dip with gain $k = 10^{-5}$.
The simulations were performed for different dip widths $m \cdot w^-$
with the scaling factors $m$ as indicated in the legend.
At $t = \unit[10]{s}$ the current bias voltage \ensuremath{V_\text{b}}\xspace was shifted towards
the right slope of the dip such that the ESC had to regain the minimum
at \ensuremath{V^-}\xspace.
All lines converge to the same signal because each simulation used the
same noise realization for the sake of better comparison.
Similar results are obtained for the positive dip.
}
\label{fig:influenceDipWidth}
\end{figure}
\subsection{Simulation Results}
\label{sec:simulation_results}
We now compare the ESC and STC with and without feedforward using
the experimentally acquired reference data.
We start with an exemplary time evolution of the involved signals and
illustrate the influence of the feedforward.
Afterwards we turn our attention to the whole image generation and
discuss the final images qualitatively and quantitatively.
Using the same measures, we also quantify the influence of the
scan speed on the image quality.
\subsubsection{Time Evolution}
In Fig.~\ref{fig:Sim1_negDip} and Fig.~\ref{fig:Sim1_posDip} exemplary
time evolutions of the \ensuremath{V_\text{b}}\xspace signal for the ESC and STC without and with
feedforward are shown.
Both simulations were performed with a scanning time of
\red{$\ensuremath{T_\text{scan}}\xspace = \unit[2]{h}$} for the respective dip map.
To mimic the experiment, each line is scanned back and forth.
This results in approximately \red{
$\unit[18]{s}$ for each single line
scan, which in turn equals a scanning speed of approximately
\unit[33.3]{\AA/s}
with the given area of \unit[600$\times$600]{\AA}.
}
In case of the negative dip (Fig.~\ref{fig:Sim1_negDip}),
at the beginning of the depicted evolution all four controller instances
are able to adequately track the dip, though the STC with a constant
error due to the fact that it tracks a point on the dips' slope.
As the variations in the electric potential increase, both controllers
without the feedforward\xspace are unable to fully follow the reference.
The ESC is unable to follow the $V^-$ variations
(beginning around $t = \unit[160]{s}$) and the STC loses the reference
around $t = \unit[340]{s}$.
In case of the positive dip (Fig.~\ref{fig:Sim1_posDip}),
the STC without feedforward\xspace again loses the dip (around $t = \unit[565]{s}$) and
the ESC tracking results in larger errors for the shown plot as compared
to its version with feedforward\xspace.
Indeed at later times, which are not depicted for reasons of clarity,
the ESC without feedforward\xspace presents the same behavior as in
Fig.~\ref{fig:Sim1_negDip}, where it is unable to follow the \ensuremath{V^\mp}\xspace
variations.
Hence, we can conclude at this point that with feedforward\xspace the probability of
both controllers successfully tracking the dips is significantly higher
and the resulting tracking error is reduced.
\begin{figure*}[htbp]
\centering
\includegraphics[width=1.08\linewidth]{Sim/Sim1_timePlot/Sim1NegDip.png}
\caption{
The top figure shows the reference evolution ($V^-$) together with
the results of the different controllers.
The bottom figure shows the error.
}
\label{fig:Sim1_negDip}
\end{figure*}
\begin{figure*}[htbp]
\centering
\includegraphics[width=1.08\linewidth]{Sim/Sim1_timePlot/Sim1PosDip.png}
\caption{
The top figure shows the reference evolution ($V^+$) together with
the results of the different controllers.
The bottom figure shows the error.
}
\label{fig:Sim1_posDip}
\end{figure*}
\subsubsection{Image Plots}
Now we turn to the simulation of the complete scans of the negative
and positive dip maps $\ensuremath{V^\mp}\xspace(x,y)$ and the resulting electrostatic potential\xspace $\ensuremath{\Phi^*}\xspace(x,y)$
computed by \eqref{eq:SQDM}.
For the sake of brevity we provide only the final $\ensuremath{\Phi^*}\xspace(x,y)$ images.
The two resulting potential images of the ESC+FF
(Fig.~\ref{fig:image_esc_FF_elPot}) and STC+FF
(Fig.~\ref{fig:image_stc_FF_elPot}) look almost identical to the
reference in Fig.~\ref{fig:STM_SQDM} and \ref{fig:dipMaps}.
Hence, the dips are successfully tracked over the whole sample.
The respective error images (Fig.~\ref{fig:image_esc_FF_elPot_error}
and Fig.~\ref{fig:image_stc_FF_elPot_error}) reveal that the errors are
in the range of \red{$\unit[-4]{mV}$ to $\unit[4.5]{mV}$}.
As the total \ensuremath{\Phi^*}\xspace variation is $\unit[190.5]{mV}$, the relative errors
w.r.t. this variation are \red{$\unit[-2.1]{\%}$ to $\unit[2.4]{\%}$}.
Furthermore, it can be seen that the STC version of the 2DOF controller
performs slightly better with smaller errors on average and with less
noise.
On the other hand, above the molecular island (right part in
Fig.~\ref{fig:image_stc_FF_elPot_error}) it can be observed that the
average STC error is shifted roughly about \unit[1.5]{mV} towards more
negative values.
This is due to the change of the systematic STC error \ensuremath{e_\text{STC}}\xspace as discussed
in Section~\ref{sec:STC_controller_gain}.
Nevertheless, the STC performs slightly better than the ESC, as is also
confirmed in Tab.~\ref{tab:MSE_PSNR}, where we quantify the image
quality using the image mean-square error (MSE) and the image peak
signal-to-noise ratio (PSNR) per pixel respectively\footnote{These
measures are often used in imaging science as objective quality
measures. See, for instance, \cite{Avcibas2002, Sheikh2006}.}.
The total scan time for each of the results in
Fig.~\ref{fig:image_esc_FF_elPot} and Fig.~\ref{fig:image_stc_FF_elPot}
was \unit[4]{h}.
This is approximately 17 times faster than the original image generation
process based on spectroscopy grids that would have taken
\unit[66.7]{h} (see Sec.~\ref{sec:original_image_generation_process}).
\begin{figure}[htpb]
\centering
\includegraphics[width=\linewidth]{Sim/Sim2_images/image_esc_FF_elPot.png}
\caption{Resulting image with the 2DOF controller employing the ESC.}
\label{fig:image_esc_FF_elPot}
\end{figure}
\begin{figure}[htpb]
\centering
\includegraphics[width=\linewidth]{Sim/Sim2_images/image_stc_FF_elPot.png}
\caption{Resulting image with the 2DOF controller employing the STC.}
\label{fig:image_stc_FF_elPot}
\end{figure}
\begin{figure}[htpb]
\centering
\includegraphics[width=\linewidth]{Sim/Sim2_images/image_esc_FF_elPot_error.png}
\caption{Resulting image error with the 2DOF controller employing the ESC.}
\label{fig:image_esc_FF_elPot_error}
\end{figure}
\begin{figure}[htpb]
\centering
\includegraphics[width=\linewidth]{Sim/Sim2_images/image_stc_FF_elPot_error.png}
\caption{Resulting image error with the 2DOF controller employing the STC.}
\label{fig:image_stc_FF_elPot_error}
\end{figure}
\begin{table}[htpb]
\centering
\caption{MSE and PSNR per pixel for Fig.~\ref{fig:image_esc_FF_elPot} and
Fig.~\ref{fig:image_stc_FF_elPot}.
}
\begin{tabular}{c|c|c}
\toprule
& MSE & PSNR \\
& $[$mV$]$ & $[$dB$]$ \\
\midrule
ESC & 0.427 & 63.7 \\
STC & 0.346 & 64.6 \\
\bottomrule
\end{tabular}
\label{tab:MSE_PSNR}
\end{table}
\subsubsection{Influence of scan time}
Finally we investigate how the resulting image quality is influenced by
the scan time \ensuremath{T_\text{scan}}\xspace.
To this end we repeat the simulation with scan times from
\red{\unit[2]{h} to \unit[6]{h}.
This equals scan speeds of \unit[33.3]{\AA/s} to \unit[11.1]{\AA/s}.}
After every simulation we compute the MSE and PSNR of the resulting
electrostatic potential\xspace image and plot it over \ensuremath{T_\text{scan}}\xspace.
The results are depicted Fig.~\ref{fig:image_MSE_PSNR}.
The slower we scan, the slower the \ensuremath{V^\mp}\xspace variations
and thus, \red{the feedback controllers are better able to follow the
references, which is reflected in lower MSE and larger PSNR values}.
\begin{figure}[htpb]
\centering
\includegraphics[width=\linewidth]{Sim/Sim3_scanSpeed/image_MSE_PSNR}
\caption{Influence of increasing scan times on the mean square error
(MSE) and the peak signal-to-noise ratio (PSNR) of the resulting electrostatic potential\xspace
image.}
\label{fig:image_MSE_PSNR}
\end{figure}
~\\
\emph{Remark:}
Of course, if the variability in the electrostatic potential\xspace is large enough, there exists
always a scan time that is too small such that good tracking is not
possible, even with the feedforward\xspace.
Nevertheless, we could show in this section that the probability to
successfully track the reference is significantly higher if the proposed
feedforward\xspace signal is employed and that reasonable scan times can be reached for
large samples.
\section{Experimental Results}
\label{sec:experimental_results}
In this section, we present information on the implementation of the
2DOF controller in the experimental setup, discuss the SQDM operation
procedure, and present further experimentally generated images of
nanoscale samples with the presented control approach of this work.
\subsection{Implementation}
\label{sec:implementation}
The 2DOF controller is built in MATLAB/Simulink and automatically
converted into C-code and then loaded to and executed on a
ds1104 controller board by dSPACE.
The controller board is mounted into a standard desktop PC and connected
via analog-digital and digital-analog converters to a
Createc non-contact atomic force/scanning tunneling microscope
(STM/NC-AFM) that operates at $\unit[5]{K}$ and under ultra high
vacuum.
The microscope is equipped with a qPlus sensor (\cite{Giessibl2004})
tuning fork with resonance frequency $f_0 = \unit[31.2]{kHz}$, stiffness
$\kappa_0 = \unit[1800]{Nm^{-1}}$.
An amplitude of $A = \unit[0.2]{\AA}$ was used in the measurements.
\subsection{Image Generation Procedure}
\label{sec:procedure}
The procedure used to acquire a new image is as follows:
\begin{enumerate}
\item Move to first pixel and measure the local spectrum of the
respective dip $\Delta f^\mp(\ensuremath{V_\text{b}}\xspace)$.
\item Adjust \ensuremath{V_\text{b}}\xspace manually to the reference point (the selection of the
reference point depends on whether the STC or the ESC is used, see
Fig.~\ref{fig:spectrum_negDip_refPoints}).
\item Start the 2DOF controller.
\item Start the raster scanning protocol.
\item Enable the feedforward\xspace after one or more lines have been scanned.
\item Increase scanning speed desired.
\item Finish the scan for the current dip $\Delta f^\mp(\ensuremath{V_\text{b}}\xspace)$.
\item Goto 1) and repeat for the other dip $\Delta f^\pm(\ensuremath{V_\text{b}}\xspace)$.
\end{enumerate}
\subsection{Images}
\label{sec:images}
Here we show a so far unpublished result obtained with the described
control approach.
The 2DOF controller with the STC has been employed together with the
above described procedure to investigate the electrostatic potential\xspace of the sample
presented in Fig.~\ref{fig:experiment_STM_image}.
The resulting SQDM image is shown in
Fig.~\ref{fig:experiment_SDQM_image}.
The sample shows three features on a Ag(111) surface, namely a
PTCDA molecule (top left), a PTCDA-Ag$_2$ complex (top right),
and a single Ag adatom (bottom).
\begin{figure}[tb]
\centering
\includegraphics[width=\linewidth]{figs/Experiment/3_features/STM_image_161205_165417.png}
\caption{Scanning tunneling microscope image of a
sample with three distinct features.
Measured with a voltage of \unit[20]{mv} and a tunneling current of
\unit[50]{pA}.
}
\label{fig:experiment_STM_image}
\end{figure}
\begin{figure}[tb]
\centering
\includegraphics[width=\linewidth]{figs/Experiment/3_features/elPot.png}
\caption{SQDM image of the electrostatic potential\xspace \ensuremath{\Phi^*}\xspace of the three features in
Fig.~\ref{fig:experiment_STM_image}.
The image \unit[260$\times$210]{\AA} was divided into
\unit[100$\times$81]{pixels} pixels and measured at a height of $z =
\unit[22]{\AA}$ above the surface.
The scan speed was \unitfrac[14]{\AA}{s}, resulting in a scan time of
\unit[50]{min} per dip.
}
\label{fig:experiment_SDQM_image}
\end{figure}
\section{Conclusion}
\label{sec:conclusion}
In this paper we have presented a two-degree-of-freedom control approach
for scanning quantum dot microscopy\xspace.
The approach consists of a feedback part and a feedforward signal
generator, where the latter is based on the previous line scan.
For the feedback part we have presented two different controllers,
namely an extremum seeking control approach that directly tracks the dip
minimum and a controller that tracks a reference point on the dip slope.
We have discussed the individual working principles, respective
advantages and drawbacks, and provided guidelines for controller
parameterization tailored to scanning quantum dot microscopy\xspace.
In simulations we could show how the utilization of the feedforward\xspace decreases
the probability of losing the dips and decreases the resulting
image error at the same time.
This leads to a several times faster image generation process and
enables to scan larger images in reasonable time than before, which was
also verified in experiments.
In addition, the presented control approach now allows to continuously
scan a sample, which puts scanning quantum dot microscopy\xspace in line with other scanning
probe microscopy techniques like scanning tunneling or atomic force
microscopy.
Furthermore, we could also show that the sharpness of
the dips plays a central role in the control performance, which led to
the recommendation that one should aim for sharper dips if possible.
The next steps are going to include an automatic gain adaptation for
varying scan speeds and the improvement of the feedforward by using
previous lines to generate an accurate prediction of the current line
using a Gaussian process and implement it in the experimental setup
\cite{Maiworm2018}.
|
Subsets and Splits